From 8290e21fb5c96f8d9f18b419eb0c17ae99c05f23 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Thu, 17 Nov 2016 13:21:15 -0500
Subject: [PATCH 001/680] replace sizeof(Packet) with PacketSize else it breaks
 for ZVector.Packet4f

---
 Eigen/src/Core/arch/ZVector/Complex.h       | 230 +++++++++++-
 Eigen/src/Core/arch/ZVector/MathFunctions.h |  69 ++--
 Eigen/src/Core/arch/ZVector/PacketMath.h    | 386 +++++++++++++++++++-
 test/packetmath.cpp                         |   2 +-
 4 files changed, 646 insertions(+), 41 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index e9d83eca6..d39d2d105 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -2,6 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -24,13 +25,48 @@ struct Packet1cd
   Packet2d v;
 };
 
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  union {
+    Packet4f v;
+    Packet1cd cd[2];
+  };
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasBlend  = 1,
+    HasSetLinear = 0
+  };
+};
+
+
 template<> struct packet_traits<std::complex<double> >  : default_packet_traits
 {
   typedef Packet1cd type;
   typedef Packet1cd half;
   enum {
     Vectorizable = 1,
-    AlignedOnScalar = 0,
+    AlignedOnScalar = 1,
     size = 1,
     HasHalfPacket = 0,
 
@@ -47,20 +83,68 @@ template<> struct packet_traits<std::complex<double> >  : default_packet_traits
   };
 };
 
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float>  type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
 
+/* Forward declaration */
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel);
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
 template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
 
 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
 
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
+  res.cd[1] = res.cd[0];
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED)
+{
+  return pload<Packet1cd>(from);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED)
+{
+  pstore<std::complex<double> >(to, from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
+  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
@@ -79,43 +163,90 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
 
   return Packet1cd(v1 + v2);
 }
-
-template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
-
-template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from)
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  return pset1<Packet1cd>(*from);
+  Packet2cf res;
+  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
+  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
+  return res;
 }
 
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from) {  return pset1<Packet1cd>(*from); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
 {
-  std::complex<double> EIGEN_ALIGN16 res[2];
-  pstore<std::complex<double> >(res, a);
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<float> >(res, a);
 
   return res[0];
 }
 
 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.cd[0] = a.cd[1];
+  res.cd[1] = a.cd[0];
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
 {
   return vecs[0];
 }
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  PacketBlock<Packet2cf,2> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  ptranspose(transpose);
+
+  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
+} 
 
 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
 
 template<int Offset>
 struct palign_impl<Offset,Packet1cd>
@@ -127,6 +258,18 @@ struct palign_impl<Offset,Packet1cd>
   }
 };
 
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset == 1) {
+      first.cd[0] = first.cd[1];
+      first.cd[1] = second.cd[0];
+    }
+  }
+};
+
 template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
 {
   EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
@@ -160,6 +303,39 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
   }
 };
 
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
   // TODO optimize it for AltiVec
@@ -168,17 +344,49 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, con
   return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
 }
 
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res;
+  res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]);
+  res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]);
+  return res;
+}
+
 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
 {
   return Packet1cd(preverse(Packet2d(x.v)));
 }
 
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
+{
+  Packet2cf res;
+  res.cd[0] = pcplxflip(x.cd[0]);
+  res.cd[1] = pcplxflip(x.cd[1]);
+  return res;
+}
+
 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
 {
   Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
   kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
   kernel.packet[0].v = tmp;
 }
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet1cd tmp = kernel.packet[0].cd[1];
+  kernel.packet[0].cd[1] = kernel.packet[1].cd[0];
+  kernel.packet[1].cd[0] = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] };
+  result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v);
+  return result;
+}
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h
index 6fff8524e..5c7aa7256 100644
--- a/Eigen/src/Core/arch/ZVector/MathFunctions.h
+++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -3,6 +3,7 @@
 //
 // Copyright (C) 2007 Julien Pommier
 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -19,32 +20,32 @@ namespace Eigen {
 
 namespace internal {
 
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d pexp<Packet2d>(const Packet2d& _x)
 {
   Packet2d x = _x;
 
-  _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
-  _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
-  _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
-
-  _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
-  _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
-
   Packet2d tmp, fx;
   Packet2l emm0;
 
@@ -91,18 +92,44 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
                  isnumber_mask);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
+  res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d psqrt<Packet2d>(const Packet2d& x)
 {
   return  __builtin_s390_vfsqdb(x);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d prsqrt<Packet2d>(const Packet2d& x) {
   // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
   return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  Packet4f res;
+  res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 }  // end namespace internal
 
 }  // end namespace Eigen
diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index 5a7226be6..e2deb25c8 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -28,9 +28,8 @@ namespace internal {
 #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
 #endif
 
-// NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  16
 #endif
 
 typedef __vector int                 Packet4i;
@@ -42,6 +41,10 @@ typedef __vector double              Packet2d;
 typedef __vector unsigned long long  Packet2ul;
 typedef __vector long long           Packet2l;
 
+typedef struct {
+	Packet2d  v4f[2];
+} Packet4f;
+
 typedef union {
   int32_t   i[4];
   uint32_t ui[4];
@@ -88,6 +91,7 @@ static Packet2d p2d_ONE = { 1.0, 1.0 };
 static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
 
 static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
 
 static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
@@ -132,13 +136,11 @@ template<> struct packet_traits<int>    : default_packet_traits
   typedef Packet4i type;
   typedef Packet4i half;
   enum {
-    // FIXME check the Has*
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 4,
     HasHalfPacket = 0,
 
-    // FIXME check the Has*
     HasAdd  = 1,
     HasSub  = 1,
     HasMul  = 1,
@@ -147,6 +149,37 @@ template<> struct packet_traits<int>    : default_packet_traits
   };
 };
 
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
 template<> struct packet_traits<double> : default_packet_traits
 {
   typedef Packet2d type;
@@ -157,7 +190,6 @@ template<> struct packet_traits<double> : default_packet_traits
     size=2,
     HasHalfPacket = 1,
 
-    // FIXME check the Has*
     HasAdd  = 1,
     HasSub  = 1,
     HasMul  = 1,
@@ -180,8 +212,12 @@ template<> struct packet_traits<double> : default_packet_traits
 };
 
 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
 
+/* Forward declaration */
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel);
+ 
 inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
 {
   Packet vt;
@@ -222,6 +258,32 @@ inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
   return s;
 }
 
+/* Helper function to simulate a vec_splat_packet4f
+ */
+template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
+{
+  Packet4f splat;
+  switch (element) {
+  case 0:
+    splat.v4f[0] = vec_splat(from.v4f[0], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 1:
+    splat.v4f[0] = vec_splat(from.v4f[0], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 2:
+    splat.v4f[0] = vec_splat(from.v4f[1], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 3:
+    splat.v4f[0] = vec_splat(from.v4f[1], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  }
+  return splat;
+}
+
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
@@ -238,6 +300,31 @@ struct palign_impl<Offset,Packet4i>
   }
 };
 
+/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
+ */
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
+      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
+      break;
+    case 2:
+      first.v4f[0] = first.v4f[1];
+      first.v4f[1] = second.v4f[0];
+      break;
+    case 3:
+      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
+      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
+      break;
+    }
+  }
+};
+
+
 template<int Offset>
 struct palign_impl<Offset,Packet2d>
 {
@@ -257,6 +344,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
   return vfrom->v4i;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet4f vfrom;
+  vfrom.v4f[0] = vec_ld2f(&from[0]);
+  vfrom.v4f[1] = vec_ld2f(&from[2]);
+  return vfrom;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
 {
   // FIXME: No intrinsic yet
@@ -275,6 +372,15 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& f
   vto->v4i = from;
 }
 
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_st2f(from.v4f[0], &to[0]);
+  vec_st2f(from.v4f[1], &to[2]);
+}
+
+
 template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
 {
   // FIXME: No intrinsic yet
@@ -288,10 +394,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)
 {
   return vec_splats(from);
 }
-
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
   return vec_splats(from);
 }
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
+{
+  Packet4f to;
+  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
+  to.v4f[1] = to.v4f[0];
+  return to;
+}
 
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet4i>(const int *a,
@@ -304,6 +416,17 @@ pbroadcast4<Packet4i>(const int *a,
   a3 = vec_splat(a3, 3);
 }
 
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat_packet4f<0>(a3);
+  a1 = vec_splat_packet4f<1>(a3);
+  a2 = vec_splat_packet4f<2>(a3);
+  a3 = vec_splat_packet4f<3>(a3);
+}
+
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet2d>(const double *a,
                       Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
@@ -326,6 +449,16 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f
  return pload<Packet4i>(ai);
 }
 
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4f>(ai);
+}
+
 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -344,6 +477,16 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const
   to[3*stride] = ai[3];
 }
 
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  pstore<float>((float *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -353,52 +496,160 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to,
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); }
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] + b.v4f[0];
+  c.v4f[1] = a.v4f[1] + b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); }
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] - b.v4f[0];
+  c.v4f[1] = a.v4f[1] - b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] * b.v4f[0];
+  c.v4f[1] = a.v4f[1] * b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); }
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] / b.v4f[0];
+  c.v4f[1] = a.v4f[1] / b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); }
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  Packet4f c;
+  c.v4f[0] = -a.v4f[0];
+  c.v4f[1] = -a.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
+{
+  Packet4f res;
+  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
+  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
 
 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)    { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); }
 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_round(a.v4f[0]);
+  res.v4f[1] = vec_round(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_ceil(a.v4f[0]);
+  res.v4f[1] = vec_ceil(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_floor(a.v4f[0]);
+  res.v4f[1] = vec_floor(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*       from) { return pload<Packet4i>(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double*    from) { return pload<Packet2d>(from); }
 
 
@@ -408,6 +659,14 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  p.v4f[1] = vec_splat(p.v4f[0], 1);
+  p.v4f[0] = vec_splat(p.v4f[0], 0);
+  return p;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 {
   Packet2d p = pload<Packet2d>(from);
@@ -415,12 +674,15 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 }
 
 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*        to, const Packet4i& from) { pstore<int>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { pstore<double>(to, from); }
 
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int    EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
@@ -433,8 +695,23 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
   return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
 }
 
-template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
-template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  Packet4f rev;
+  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
+  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
+  return rev;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = pabs(a.v4f[0]);
+  res.v4f[1] = pabs(a.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
@@ -453,6 +730,13 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
   sum = padd<Packet2d>(a, b);
   return pfirst(sum);
 }
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet2d sum;
+  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
+  double first = predux<Packet2d>(sum);
+  return static_cast<float>(first);
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
 {
@@ -493,6 +777,21 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
   return sum;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  PacketBlock<Packet4f,4> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  transpose.packet[2] = vecs[2];
+  transpose.packet[3] = vecs[3];
+  ptranspose(transpose);
+
+  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
+  sum = padd(sum, transpose.packet[2]);
+  sum = padd(sum, transpose.packet[3]);
+  return sum;
+}
+
 // Other reduction functions:
 // mul
 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
@@ -507,6 +806,12 @@ template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
   return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  // Return predux_mul<Packet2d> of the subvectors product
+  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
+}
+
 // min
 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
 {
@@ -521,6 +826,14 @@ template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
   return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
 // max
 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 {
@@ -536,6 +849,14 @@ template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
   return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<Packet4i,4>& kernel) {
   Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
@@ -556,12 +877,61 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) {
   kernel.packet[1] = t1;
 }
 
+/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
+ */
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  PacketBlock<Packet2d,2> t0,t1,t2,t3;
+  // copy top-left 2x2 Packet2d block
+  t0.packet[0] = kernel.packet[0].v4f[0];
+  t0.packet[1] = kernel.packet[1].v4f[0];
+
+  // copy top-right 2x2 Packet2d block
+  t1.packet[0] = kernel.packet[0].v4f[1];
+  t1.packet[1] = kernel.packet[1].v4f[1];
+
+  // copy bottom-left 2x2 Packet2d block
+  t2.packet[0] = kernel.packet[2].v4f[0];
+  t2.packet[1] = kernel.packet[3].v4f[0];
+
+  // copy bottom-right 2x2 Packet2d block
+  t3.packet[0] = kernel.packet[2].v4f[1];
+  t3.packet[1] = kernel.packet[3].v4f[1];
+
+  // Transpose all 2x2 blocks
+  ptranspose(t0);
+  ptranspose(t1);
+  ptranspose(t2);
+  ptranspose(t3);
+
+  // Copy back transposed blocks, but exchange t1 and t2 due to transposition
+  kernel.packet[0].v4f[0] = t0.packet[0];
+  kernel.packet[0].v4f[1] = t2.packet[0];
+  kernel.packet[1].v4f[0] = t0.packet[1];
+  kernel.packet[1].v4f[1] = t2.packet[1];
+  kernel.packet[2].v4f[0] = t1.packet[0];
+  kernel.packet[2].v4f[1] = t3.packet[0];
+  kernel.packet[3].v4f[0] = t1.packet[1];
+  kernel.packet[3].v4f[1] = t3.packet[1];
+}
+
 template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
   Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
   Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
   return vec_sel(elsePacket, thenPacket, mask);
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] };
+  Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet4f result;
+  result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi);
+  result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
+  return result;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
   Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
   Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index c18d73496..7821a1738 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -591,7 +591,7 @@ template<typename Scalar> void packetmath_scatter_gather()
   int stride = internal::random<int>(1,20);
 
   EIGEN_ALIGN_MAX Scalar buffer[PacketSize*20];
-  memset(buffer, 0, 20*sizeof(Packet));
+  memset(buffer, 0, 20*PacketSize*sizeof(Scalar));
   Packet packet = internal::pload<Packet>(data1);
   internal::pscatter<Scalar, Packet>(buffer, packet, stride);
 

From bbd97b4095ff9cbe9898d68b3ab7bdff8125f3fb Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Jul 2017 01:02:51 +0200
Subject: [PATCH 002/680] Add a EIGEN_NO_CUDA option, and introduce
 EIGEN_CUDACC and EIGEN_CUDA_ARCH aliases

---
 Eigen/Core                                    | 18 +++++--
 Eigen/src/Core/GeneralProduct.h               |  4 +-
 Eigen/src/Core/GenericPacketMath.h            |  4 +-
 Eigen/src/Core/MathFunctions.h                | 48 +++++++++----------
 Eigen/src/Core/MatrixBase.h                   |  2 +-
 Eigen/src/Core/ProductEvaluators.h            |  4 +-
 Eigen/src/Core/arch/CUDA/Complex.h            |  2 +-
 Eigen/src/Core/arch/CUDA/Half.h               | 30 ++++++------
 Eigen/src/Core/arch/CUDA/MathFunctions.h      |  2 +-
 Eigen/src/Core/arch/CUDA/PacketMath.h         | 10 ++--
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h     | 28 +++++------
 Eigen/src/Core/arch/CUDA/TypeCasting.h        |  8 ++--
 Eigen/src/Core/functors/AssignmentFunctors.h  |  2 +-
 Eigen/src/Core/util/Macros.h                  |  6 +--
 Eigen/src/Core/util/Meta.h                    |  8 ++--
 Eigen/src/SVD/BDCSVD.h                        |  2 +-
 Eigen/src/SparseQR/SparseQR.h                 |  2 +-
 .../CXX11/src/Tensor/TensorContractionCuda.h  |  4 +-
 .../CXX11/src/Tensor/TensorConvolution.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorDeviceCuda.h | 12 ++---
 .../CXX11/src/Tensor/TensorDeviceDefault.h    | 10 ++--
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |  4 +-
 .../Eigen/CXX11/src/Tensor/TensorIntDiv.h     | 12 ++---
 .../Eigen/CXX11/src/Tensor/TensorMacros.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorMeta.h       |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorRandom.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  6 +--
 .../CXX11/src/Tensor/TensorReductionCuda.h    | 10 ++--
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |  4 +-
 .../Eigen/CXX11/src/util/EmulateArray.h       |  2 +-
 .../SpecialFunctions/SpecialFunctionsImpl.h   |  4 +-
 .../arch/CUDA/CudaSpecialFunctions.h          |  2 +-
 33 files changed, 134 insertions(+), 126 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index a16942e19..d2edbd2bc 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -14,8 +14,16 @@
 // first thing Eigen does: stop the compiler from committing suicide
 #include "src/Core/util/DisableStupidWarnings.h"
 
+#if defined(__CUDACC__) && !defined(EIGEN_NO_CUDA)
+  #define EIGEN_CUDACC __CUDACC__
+#endif
+
+#if defined(__CUDA_ARCH__) && !defined(EIGEN_NO_CUDA)
+  #define EIGEN_CUDA_ARCH __CUDA_ARCH__
+#endif
+
 // Handle NVCC/CUDA/SYCL
-#if defined(__CUDACC__) || defined(__SYCL_DEVICE_ONLY__)
+#if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__)
   // Do not try asserts on CUDA and SYCL!
   #ifndef EIGEN_NO_DEBUG
   #define EIGEN_NO_DEBUG
@@ -30,7 +38,7 @@
   #endif
 
   // All functions callable from CUDA code must be qualified with __device__
-  #ifdef __CUDACC__
+  #ifdef EIGEN_CUDACC
     // Do not try to vectorize on CUDA and SYCL!
     #ifndef EIGEN_DONT_VECTORIZE
     #define EIGEN_DONT_VECTORIZE
@@ -50,13 +58,13 @@
 // When compiling CUDA device code with NVCC, pull in math functions from the
 // global namespace.  In host mode, and when device doee with clang, use the
 // std versions.
-#if defined(__CUDA_ARCH__) && defined(__NVCC__)
+#if defined(EIGEN_CUDA_ARCH) && defined(__NVCC__)
   #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
 #else
   #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
 #endif
 
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL)
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL)
   #define EIGEN_EXCEPTIONS
 #endif
 
@@ -233,7 +241,7 @@
   #define EIGEN_HAS_FP16_C
 #endif
 
-#if defined __CUDACC__
+#if defined EIGEN_CUDACC
   #define EIGEN_VECTORIZE_CUDA
   #include <vector_types.h>
   #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index b206b0a7a..95dadfea8 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -379,7 +379,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
   *
   * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
   */
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
 
 template<typename Derived>
 template<typename OtherDerived>
@@ -412,7 +412,7 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   return Product<Derived, OtherDerived>(derived(), other.derived());
 }
 
-#endif // __CUDACC__
+#endif // EIGEN_CUDACC
 
 /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
   *
diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index d19d5bbd2..30878eda6 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -299,7 +299,7 @@ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu
 /** \internal tries to do cache prefetching of \a addr */
 template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
 {
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
 #if defined(__LP64__)
   // 64-bit pointer operand constraint for inlined asm
   asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
@@ -526,7 +526,7 @@ inline void palign(PacketType& first, const PacketType& second)
 ***************************************************************************/
 
 // Eigen+CUDA does not support complexes.
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
 
 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
 { return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 75f34aa91..a906b7629 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -96,7 +96,7 @@ struct real_default_impl<Scalar,true>
 
 template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
 
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
 template<typename T>
 struct real_impl<std::complex<T> >
 {
@@ -144,7 +144,7 @@ struct imag_default_impl<Scalar,true>
 
 template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
 
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
 template<typename T>
 struct imag_impl<std::complex<T> >
 {
@@ -778,7 +778,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isfinite_impl(const T& x)
 {
-  #ifdef __CUDA_ARCH__
+  #ifdef EIGEN_CUDA_ARCH
     return (::isfinite)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isfinite;
@@ -793,7 +793,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isinf_impl(const T& x)
 {
-  #ifdef __CUDA_ARCH__
+  #ifdef EIGEN_CUDA_ARCH
     return (::isinf)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isinf;
@@ -808,7 +808,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isnan_impl(const T& x)
 {
-  #ifdef __CUDA_ARCH__
+  #ifdef EIGEN_CUDA_ARCH
     return (::isnan)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isnan;
@@ -874,7 +874,7 @@ template<typename T> T generic_fast_tanh_float(const T& a_x);
 
 namespace numext {
 
-#if !defined(__CUDA_ARCH__) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
@@ -1088,7 +1088,7 @@ EIGEN_ALWAYS_INLINE float   log1p(float x) { return cl::sycl::log1p(x); }
 EIGEN_ALWAYS_INLINE double  log1p(double x) { return cl::sycl::log1p(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log1p(const float &x) { return ::log1pf(x); }
 
@@ -1146,7 +1146,7 @@ EIGEN_ALWAYS_INLINE float   floor(float x) { return cl::sycl::floor(x); }
 EIGEN_ALWAYS_INLINE double  floor(double x) { return cl::sycl::floor(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float floor(const float &x) { return ::floorf(x); }
 
@@ -1167,7 +1167,7 @@ EIGEN_ALWAYS_INLINE float   ceil(float x) { return cl::sycl::ceil(x); }
 EIGEN_ALWAYS_INLINE double  ceil(double x) { return cl::sycl::ceil(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float ceil(const float &x) { return ::ceilf(x); }
 
@@ -1225,7 +1225,7 @@ EIGEN_ALWAYS_INLINE double  log(double x) { return cl::sycl::log(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log(const float &x) { return ::logf(x); }
 
@@ -1253,7 +1253,7 @@ EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
 EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float abs(const float &x) { return ::fabsf(x); }
 
@@ -1283,7 +1283,7 @@ EIGEN_ALWAYS_INLINE float   exp(float x) { return cl::sycl::exp(x); }
 EIGEN_ALWAYS_INLINE double  exp(double x) { return cl::sycl::exp(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float exp(const float &x) { return ::expf(x); }
 
@@ -1303,7 +1303,7 @@ EIGEN_ALWAYS_INLINE float   expm1(float x) { return cl::sycl::expm1(x); }
 EIGEN_ALWAYS_INLINE double  expm1(double x) { return cl::sycl::expm1(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float expm1(const float &x) { return ::expm1f(x); }
 
@@ -1323,7 +1323,7 @@ EIGEN_ALWAYS_INLINE float   cos(float x) { return cl::sycl::cos(x); }
 EIGEN_ALWAYS_INLINE double  cos(double x) { return cl::sycl::cos(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cos(const float &x) { return ::cosf(x); }
 
@@ -1343,7 +1343,7 @@ EIGEN_ALWAYS_INLINE float   sin(float x) { return cl::sycl::sin(x); }
 EIGEN_ALWAYS_INLINE double  sin(double x) { return cl::sycl::sin(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sin(const float &x) { return ::sinf(x); }
 
@@ -1363,7 +1363,7 @@ EIGEN_ALWAYS_INLINE float   tan(float x) { return cl::sycl::tan(x); }
 EIGEN_ALWAYS_INLINE double  tan(double x) { return cl::sycl::tan(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tan(const float &x) { return ::tanf(x); }
 
@@ -1393,7 +1393,7 @@ EIGEN_ALWAYS_INLINE float   acosh(float x) { return cl::sycl::acosh(x); }
 EIGEN_ALWAYS_INLINE double  acosh(double x) { return cl::sycl::acosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float acos(const float &x) { return ::acosf(x); }
 
@@ -1422,7 +1422,7 @@ EIGEN_ALWAYS_INLINE float   asinh(float x) { return cl::sycl::asinh(x); }
 EIGEN_ALWAYS_INLINE double  asinh(double x) { return cl::sycl::asinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float asin(const float &x) { return ::asinf(x); }
 
@@ -1451,7 +1451,7 @@ EIGEN_ALWAYS_INLINE float   atanh(float x) { return cl::sycl::atanh(x); }
 EIGEN_ALWAYS_INLINE double  atanh(double x) { return cl::sycl::atanh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float atan(const float &x) { return ::atanf(x); }
 
@@ -1472,7 +1472,7 @@ EIGEN_ALWAYS_INLINE float   cosh(float x) { return cl::sycl::cosh(x); }
 EIGEN_ALWAYS_INLINE double  cosh(double x) { return cl::sycl::cosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cosh(const float &x) { return ::coshf(x); }
 
@@ -1492,7 +1492,7 @@ EIGEN_ALWAYS_INLINE float   sinh(float x) { return cl::sycl::sinh(x); }
 EIGEN_ALWAYS_INLINE double  sinh(double x) { return cl::sycl::sinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sinh(const float &x) { return ::sinhf(x); }
 
@@ -1510,12 +1510,12 @@ T tanh(const T &x) {
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   tanh(float x) { return cl::sycl::tanh(x); }
 EIGEN_ALWAYS_INLINE double  tanh(double x) { return cl::sycl::tanh(x); }
-#elif (!defined(__CUDACC__)) && EIGEN_FAST_MATH
+#elif (!defined(EIGEN_CUDACC)) && EIGEN_FAST_MATH
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(float x) { return internal::generic_fast_tanh_float(x); }
 #endif
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(const float &x) { return ::tanhf(x); }
 
@@ -1535,7 +1535,7 @@ EIGEN_ALWAYS_INLINE float   fmod(float x, float y) { return cl::sycl::fmod(x, y)
 EIGEN_ALWAYS_INLINE double  fmod(double x, double y) { return cl::sycl::fmod(x, y); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 template <>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float fmod(const float& a, const float& b) {
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 200e57741..908d35dfe 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -160,7 +160,7 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator-=(const MatrixBase<OtherDerived>& other);
 
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     const Product<Derived,OtherDerived,LazyProduct>
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index c42725dbd..86966abdb 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -851,7 +851,7 @@ struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalSha
     return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col);
   }
   
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
   template<int LoadMode,typename PacketType>
   EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
   {
@@ -895,7 +895,7 @@ struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape,
     return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col);
   }
   
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
   template<int LoadMode,typename PacketType>
   EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
   {
diff --git a/Eigen/src/Core/arch/CUDA/Complex.h b/Eigen/src/Core/arch/CUDA/Complex.h
index ca0aaed32..57d1201f4 100644
--- a/Eigen/src/Core/arch/CUDA/Complex.h
+++ b/Eigen/src/Core/arch/CUDA/Complex.h
@@ -16,7 +16,7 @@ namespace Eigen {
 
 namespace internal {
 
-#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
 
 // Many std::complex methods such as operator+, operator-, operator* and
 // operator/ are not constexpr. Due to this, clang does not treat them as device
diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index e4e639fcd..e99847e24 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -140,7 +140,7 @@ struct half : public half_impl::half_base {
 
 namespace half_impl {
 
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
 
 // Intrinsics for native fp16 support. Note that on current hardware,
 // these are no faster than fp32 arithmetic (you need to use the half2
@@ -281,7 +281,7 @@ union FP32 {
 };
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
   return __float2half(ff);
 
 #elif defined(EIGEN_HAS_FP16_C)
@@ -336,7 +336,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
 }
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
   return __half2float(h);
 
 #elif defined(EIGEN_HAS_FP16_C)
@@ -370,7 +370,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
   return (a.x & 0x7fff) == 0x7c00;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return __hisnan(a);
 #else
   return (a.x & 0x7fff) > 0x7c00;
@@ -386,7 +386,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
   return result;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
   return half(hexp(a));
 #else
    return half(::expf(float(a)));
@@ -396,7 +396,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
   return half(numext::expm1(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return half(::hlog(a));
 #else
   return half(::logf(float(a)));
@@ -409,7 +409,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
   return half(::log10f(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
   return half(hsqrt(a));
 #else
     return half(::sqrtf(float(a)));
@@ -431,14 +431,14 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
   return half(::tanhf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
   return half(hfloor(a));
 #else
   return half(::floorf(float(a)));
 #endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
   return half(hceil(a));
 #else
   return half(::ceilf(float(a)));
@@ -446,7 +446,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
 }
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return __hlt(b, a) ? b : a;
 #else
   const float f1 = static_cast<float>(a);
@@ -455,7 +455,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
 #endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return __hlt(a, b) ? b : a;
 #else
   const float f1 = static_cast<float>(a);
@@ -576,7 +576,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
   return Eigen::half(::expf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return Eigen::half(::hlog(a));
 #else
   return Eigen::half(::logf(float(a)));
@@ -610,14 +610,14 @@ struct hash<Eigen::half> {
 
 
 // Add the missing shfl_xor intrinsic
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
 }
 #endif
 
 // ldg() has an overload for __half, but we also need one for Eigen::half.
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
   return Eigen::half_impl::raw_uint16_to_half(
       __ldg(reinterpret_cast<const unsigned short*>(ptr)));
@@ -625,7 +625,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr)
 #endif
 
 
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
 namespace Eigen {
 namespace numext {
 
diff --git a/Eigen/src/Core/arch/CUDA/MathFunctions.h b/Eigen/src/Core/arch/CUDA/MathFunctions.h
index 987a5291c..ff6256ce0 100644
--- a/Eigen/src/Core/arch/CUDA/MathFunctions.h
+++ b/Eigen/src/Core/arch/CUDA/MathFunctions.h
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 plog<float4>(const float4& a)
 {
diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index 8c46af09b..97a8abe59 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
 template<> struct is_arithmetic<float4>  { enum { value = true }; };
 template<> struct is_arithmetic<double2> { enum { value = true }; };
 
@@ -196,7 +196,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to
 
 template<>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
   return __ldg((const float4*)from);
 #else
   return make_float4(from[0], from[1], from[2], from[3]);
@@ -204,7 +204,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const fl
 }
 template<>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
   return __ldg((const double2*)from);
 #else
   return make_double2(from[0], from[1]);
@@ -213,7 +213,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const
 
 template<>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
   return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3));
 #else
   return make_float4(from[0], from[1], from[2], from[3]);
@@ -221,7 +221,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const
 }
 template<>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
   return make_double2(__ldg(from+0), __ldg(from+1));
 #else
   return make_double2(from[0], from[1]);
diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index f4ae3c3c5..4a730a0e0 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -15,7 +15,7 @@ namespace Eigen {
 namespace internal {
 
 // Most of the following operations require arch >= 3.0
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
 
 template<> struct is_arithmetic<half2> { enum { value = true }; };
 
@@ -69,7 +69,7 @@ template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half*
 
 template<>
  __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
-#if __CUDA_ARCH__ >= 350
+#if EIGEN_CUDA_ARCH >= 350
    return __ldg((const half2*)from);
 #else
   return __halves2half2(*(from+0), *(from+1));
@@ -78,7 +78,7 @@ template<>
 
 template<>
 __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
-#if __CUDA_ARCH__ >= 350
+#if EIGEN_CUDA_ARCH >= 350
    return __halves2half2(__ldg(from+0), __ldg(from+1));
 #else
   return __halves2half2(*(from+0), *(from+1));
@@ -116,7 +116,7 @@ ptranspose(PacketBlock<half2,2>& kernel) {
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __halves2half2(a, __hadd(a, __float2half(1.0f)));
 #else
   float f = __half2float(a) + 1.0f;
@@ -125,7 +125,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half&
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hadd2(a, b);
 #else
   float a1 = __low2float(a);
@@ -139,7 +139,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, cons
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hsub2(a, b);
 #else
   float a1 = __low2float(a);
@@ -153,7 +153,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, cons
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hneg2(a);
 #else
   float a1 = __low2float(a);
@@ -165,7 +165,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
 template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hmul2(a, b);
 #else
   float a1 = __low2float(a);
@@ -179,7 +179,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, cons
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
    return __hfma2(a, b, c);
 #else
   float a1 = __low2float(a);
@@ -225,7 +225,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, cons
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hadd(__low2half(a), __high2half(a));
 #else
   float a1 = __low2float(a);
@@ -235,7 +235,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
   return __hgt(first, second) ? first : second;
@@ -247,7 +247,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const ha
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
   return __hlt(first, second) ? first : second;
@@ -259,7 +259,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const ha
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
-#if __CUDA_ARCH__ >= 530
+#if EIGEN_CUDA_ARCH >= 530
   return __hmul(__low2half(a), __high2half(a));
 #else
   float a1 = __low2float(a);
@@ -284,7 +284,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
 
 template<>  __device__ EIGEN_STRONG_INLINE
 half2 plog<half2>(const half2& a) {
diff --git a/Eigen/src/Core/arch/CUDA/TypeCasting.h b/Eigen/src/Core/arch/CUDA/TypeCasting.h
index aa5fbce8e..30f870c3d 100644
--- a/Eigen/src/Core/arch/CUDA/TypeCasting.h
+++ b/Eigen/src/Core/arch/CUDA/TypeCasting.h
@@ -19,7 +19,7 @@ struct scalar_cast_op<float, Eigen::half> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef Eigen::half result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
       return __float2half(a);
     #else
       return Eigen::half(a);
@@ -37,7 +37,7 @@ struct scalar_cast_op<int, Eigen::half> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef Eigen::half result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
       return __float2half(static_cast<float>(a));
     #else
       return Eigen::half(static_cast<float>(a));
@@ -55,7 +55,7 @@ struct scalar_cast_op<Eigen::half, float> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef float result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
       return __half2float(a);
     #else
       return static_cast<float>(a);
@@ -69,7 +69,7 @@ struct functor_traits<scalar_cast_op<Eigen::half, float> >
 
 
 
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
 
 template <>
 struct type_casting_traits<Eigen::half, float> {
diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h
index 4153b877c..1077d8eb0 100644
--- a/Eigen/src/Core/functors/AssignmentFunctors.h
+++ b/Eigen/src/Core/functors/AssignmentFunctors.h
@@ -144,7 +144,7 @@ template<typename Scalar> struct swap_assign_op {
   EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const
   {
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
     // FIXME is there some kind of cuda::swap?
     Scalar t=b; const_cast<Scalar&>(b)=a; a=t;
 #else
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 755646795..322e8d899 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -427,7 +427,7 @@
 // Does the compiler fully support const expressions? (as in c++14)
 #ifndef EIGEN_HAS_CONSTEXPR
 
-#if defined(__CUDACC__)
+#if defined(EIGEN_CUDACC)
 // Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
 #if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500))
   #define EIGEN_HAS_CONSTEXPR 1
@@ -669,7 +669,7 @@ namespace Eigen {
  * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
  * vectorized and non-vectorized code.
  */
-#if (defined __CUDACC__)
+#if (defined EIGEN_CUDACC)
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
 #elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
@@ -990,7 +990,7 @@ namespace Eigen {
 #  define EIGEN_TRY try
 #  define EIGEN_CATCH(X) catch (X)
 #else
-#  ifdef __CUDA_ARCH__
+#  ifdef EIGEN_CUDA_ARCH
 #    define EIGEN_THROW_X(X) asm("trap;")
 #    define EIGEN_THROW asm("trap;")
 #  else
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 8de605500..0fa818008 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -11,7 +11,7 @@
 #ifndef EIGEN_META_H
 #define EIGEN_META_H
 
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
 #include <cfloat>
 #include <math_constants.h>
 #endif
@@ -169,7 +169,7 @@ template<bool Condition, typename T=void> struct enable_if;
 template<typename T> struct enable_if<true,T>
 { typedef T type; };
 
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
 #if !defined(__FLT_EPSILON__)
 #define __FLT_EPSILON__ FLT_EPSILON
 #define __DBL_EPSILON__ DBL_EPSILON
@@ -523,13 +523,13 @@ template<typename T, typename U> struct scalar_product_traits
 
 namespace numext {
   
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
 template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
 #else
 template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
 #endif
 
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
 using internal::device::numeric_limits;
 #else
 using std::numeric_limits;
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index d7a4271cb..b8c41c560 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -1211,7 +1211,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
 #endif
 }//end deflation
 
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
 /** \svd_module
   *
   * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm
diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h
index 2d4498b03..f7111fe2e 100644
--- a/Eigen/src/SparseQR/SparseQR.h
+++ b/Eigen/src/SparseQR/SparseQR.h
@@ -327,7 +327,7 @@ void SparseQR<MatrixType,OrderingType>::analyzePattern(const MatrixType& mat)
   internal::coletree(matCpy, m_etree, m_firstRowElt, m_outputPerm_c.indices().data());
   m_isEtreeOk = true;
   
-  m_R.resize(m, n);
+  m_R.resize(diagSize, n);
   m_Q.resize(m, diagSize);
   
   // Allocate space for nonzero elements : rough estimation
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
index c04b784a4..428b18499 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
@@ -12,7 +12,7 @@
 #ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
 #define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
 
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
 
 namespace Eigen {
 
@@ -1382,5 +1382,5 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
 } // end namespace Eigen
 
-#endif // EIGEN_USE_GPU and __CUDACC__
+#endif // EIGEN_USE_GPU and EIGEN_CUDACC
 #endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 6fa51fd64..84d5be173 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -553,7 +553,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
 
 
 // Use an optimized implementation of the evaluation code for GPUs whenever possible.
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
 
 template <int StaticKernelSize>
 struct GetKernelSize {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
index be8d69386..ded7129da 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
@@ -211,7 +211,7 @@ struct GpuDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
-#ifndef __CUDA_ARCH__
+#ifndef EIGEN_CUDA_ARCH
     cudaError_t err = cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice,
                                       stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
@@ -239,7 +239,7 @@ struct GpuDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
-#ifndef __CUDA_ARCH__
+#ifndef EIGEN_CUDA_ARCH
     cudaError_t err = cudaMemsetAsync(buffer, c, n, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
     assert(err == cudaSuccess);
@@ -265,7 +265,7 @@ struct GpuDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void synchronize() const {
-#if defined(__CUDACC__) && !defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDACC) && !defined(EIGEN_CUDA_ARCH)
     cudaError_t err = cudaStreamSynchronize(stream_->stream());
     if (err != cudaSuccess) {
       std::cerr << "Error detected in CUDA stream: "
@@ -304,7 +304,7 @@ struct GpuDevice {
   // This function checks if the CUDA runtime recorded an error for the
   // underlying stream device.
   inline bool ok() const {
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
     cudaError_t error = cudaStreamQuery(stream_->stream());
     return (error == cudaSuccess) || (error == cudaErrorNotReady);
 #else
@@ -323,9 +323,9 @@ struct GpuDevice {
 
 
 // FIXME: Should be device and kernel specific.
-#ifdef __CUDACC__
+#ifdef EIGEN_CUDACC
 static EIGEN_DEVICE_FUNC inline void setCudaSharedMemConfig(cudaSharedMemConfig config) {
-#ifndef __CUDA_ARCH__
+#ifndef EIGEN_CUDA_ARCH
   cudaError_t status = cudaDeviceSetSharedMemConfig(config);
   EIGEN_UNUSED_VARIABLE(status)
   assert(status == cudaSuccess);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
index ccaaa6cb2..341889e88 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
@@ -35,7 +35,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const {
-#ifndef __CUDA_ARCH__
+#ifndef EIGEN_CUDA_ARCH
     // Running on the host CPU
     return 1;
 #else
@@ -45,7 +45,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
-#if !defined(__CUDA_ARCH__) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
     // Running on the host CPU
     return l1CacheSize();
 #else
@@ -55,7 +55,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
-#if !defined(__CUDA_ARCH__) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
     // Running single threaded on the host CPU
     return l3CacheSize();
 #else
@@ -65,13 +65,13 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
-#ifndef __CUDA_ARCH__
+#ifndef EIGEN_CUDA_ARCH
     // Running single threaded on the host CPU
     // Should return an enum that encodes the ISA supported by the CPU
     return 1;
 #else
     // Running on a CUDA device
-    return __CUDA_ARCH__ / 100;
+    return EIGEN_CUDA_ARCH / 100;
 #endif
   }
 };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index fcf330b10..2264be391 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -131,7 +131,7 @@ T loadConstant(const T* address) {
   return *address;
 }
 // Use the texture cache on CUDA devices whenever possible
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
 template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float loadConstant(const float* address) {
   return __ldg(address);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index f01d77c0a..0ffe68ab3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -201,7 +201,7 @@ class TensorExecutor<Expression, GpuDevice, Vectorizable> {
 };
 
 
-#if defined(__CUDACC__)
+#if defined(EIGEN_CUDACC)
 template <typename Evaluator, typename Index, bool Vectorizable>
 struct EigenMetaKernelEval {
   static __device__ EIGEN_ALWAYS_INLINE
@@ -264,7 +264,7 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
   evaluator.cleanup();
 }
 
-#endif  // __CUDACC__
+#endif  // EIGEN_CUDACC
 #endif  // EIGEN_USE_GPU
 
 // SYCL Executor policy
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
index ef1c9c42c..fb6454623 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
@@ -35,7 +35,7 @@ namespace {
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   typename internal::enable_if<sizeof(T)==4,int>::type count_leading_zeros(const T val)
   {
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
     return __clz(val);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::clz(val);
@@ -53,7 +53,7 @@ namespace {
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   typename internal::enable_if<sizeof(T)==8,int>::type count_leading_zeros(const T val)
   {
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
     return __clzll(val);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::clz(val);
@@ -90,7 +90,7 @@ namespace {
 
   template <typename T>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
     return __umulhi(a, b);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
@@ -101,7 +101,7 @@ namespace {
 
   template <typename T>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
-#if defined(__CUDA_ARCH__)
+#if defined(EIGEN_CUDA_ARCH)
     return __umul64hi(a, b);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
@@ -124,7 +124,7 @@ namespace {
   template <typename T>
   struct DividerHelper<64, T> {
     static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
-#if defined(__SIZEOF_INT128__) && !defined(__CUDA_ARCH__) && !defined(__SYCL_DEVICE_ONLY__)
+#if defined(__SIZEOF_INT128__) && !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
       return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
 #else
       const uint64_t shift = 1ULL << log_div;
@@ -203,7 +203,7 @@ class TensorIntDivisor<int32_t, true> {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
     return (__umulhi(magic, n) >> shift);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return (cl::sycl::mul_hi(static_cast<uint64_t>(magic), static_cast<uint64_t>(n)) >> shift);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
index f92e39d69..c9e61f359 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
@@ -27,7 +27,7 @@
  */
 
 // SFINAE requires variadic templates
-#ifndef __CUDACC__
+#ifndef EIGEN_CUDACC
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   // SFINAE doesn't work for gcc <= 4.7
   #ifdef EIGEN_COMP_GNUC
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index 77c9c6c6e..5431eb740 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -52,7 +52,7 @@ struct PacketType : internal::packet_traits<Scalar> {
 };
 
 // For CUDA packet types when using a GpuDevice
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) && defined(EIGEN_HAS_CUDA_FP16)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC) && defined(EIGEN_HAS_CUDA_FP16)
 template <>
 struct PacketType<half, GpuDevice> {
   typedef half2 type;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
index f108c349f..230915db2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
@@ -16,7 +16,7 @@ namespace internal {
 namespace {
 
 EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
-#ifdef __CUDA_ARCH__
+#ifdef EIGEN_CUDA_ARCH
   // We don't support 3d kernels since we currently only use 1 and
   // 2d kernels.
   assert(threadIdx.z == 0);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 69079805d..da0ffe728 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -334,7 +334,7 @@ struct OuterReducer {
 };
 
 
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
 template <int B, int N, typename S, typename R, typename I>
 __global__ void FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
 
@@ -694,7 +694,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 #ifdef EIGEN_USE_THREADS
   template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
 #endif
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
   template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
 #ifdef EIGEN_HAS_CUDA_FP16
   template <typename S, typename R, typename I> KERNEL_FRIEND void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
@@ -781,7 +781,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   Op m_reducer;
 
   // For full reductions
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
   static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
   static const bool RunningOnSycl = false;
 #elif defined(EIGEN_USE_SYCL)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index 24a55a3d5..974eb7deb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -14,7 +14,7 @@ namespace Eigen {
 namespace internal {
 
 
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
 // Full reducers for GPU, don't vectorize for now
 
 // Reducer function that enables multiple cuda thread to safely accumulate at the same
@@ -23,7 +23,7 @@ namespace internal {
 // updated the content of the output address it will try again.
 template <typename T, typename R>
 __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) {
-#if __CUDA_ARCH__ >= 300
+#if EIGEN_CUDA_ARCH >= 300
   if (sizeof(T) == 4)
   {
     unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
@@ -102,7 +102,7 @@ __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer
 
 template <>
 __device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
-#if __CUDA_ARCH__ >= 300
+#if EIGEN_CUDA_ARCH >= 300
   atomicAdd(output, accum);
 #else // __CUDA_ARCH__ >= 300
   assert(0 && "Shouldn't be called on unsupported device");
@@ -124,7 +124,7 @@ template <int BlockSize, int NumPerThread, typename Self,
           typename Reducer, typename Index>
 __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num_coeffs,
                                     typename Self::CoeffReturnType* output, unsigned int* semaphore) {
-#if __CUDA_ARCH__ >= 300
+#if EIGEN_CUDA_ARCH >= 300
   // Initialize the output value
   const Index first_index = blockIdx.x * BlockSize * NumPerThread + threadIdx.x;
   if (gridDim.x == 1) {
@@ -372,7 +372,7 @@ template <int NumPerThread, typename Self,
           typename Reducer, typename Index>
 __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
                                          typename Self::CoeffReturnType* output) {
-#if __CUDA_ARCH__ >= 300
+#if EIGEN_CUDA_ARCH >= 300
   typedef typename Self::CoeffReturnType Type;
   eigen_assert(blockDim.y == 1);
   eigen_assert(blockDim.z == 1);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 2a85ed840..1f545ef1a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -242,7 +242,7 @@ struct ScanLauncher {
   }
 };
 
-#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
 
 // GPU implementation of scan
 // TODO(ibab) This placeholder implementation performs multiple scans in
@@ -281,7 +281,7 @@ struct ScanLauncher<Self, Reducer, GpuDevice> {
      LAUNCH_CUDA_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
   }
 };
-#endif  // EIGEN_USE_GPU && __CUDACC__
+#endif  // EIGEN_USE_GPU && EIGEN_CUDACC
 
 }  // end namespace Eigen
 
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 573ca435a..96b3a8261 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -15,7 +15,7 @@
 // The array class is only available starting with cxx11. Emulate our own here
 // if needed. Beware, msvc still doesn't advertise itself as a c++11 compiler!
 // Moreover, CUDA doesn't support the STL containers, so we use our own instead.
-#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(__CUDACC__) || defined(EIGEN_AVOID_STL_ARRAY)
+#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_CUDACC) || defined(EIGEN_AVOID_STL_ARRAY)
 
 namespace Eigen {
 template <typename T, size_t n> class array {
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 369ad97b4..5d1b8fcc2 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -121,7 +121,7 @@ template <>
 struct lgamma_impl<float> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE float run(float x) {
-#if !defined(__CUDA_ARCH__) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
+#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
     int dummy;
     return ::lgammaf_r(x, &dummy);
 #else
@@ -134,7 +134,7 @@ template <>
 struct lgamma_impl<double> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE double run(double x) {
-#if !defined(__CUDA_ARCH__) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
+#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
     int dummy;
     return ::lgamma_r(x, &dummy);
 #else
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
index ec4fa8448..e0e3a8be6 100644
--- a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
+++ b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
 
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 plgamma<float4>(const float4& a)

From 9f8136ff747088b9a74c89ab0b3d89ac1081e83d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Jul 2017 10:43:18 +0200
Subject: [PATCH 003/680] disable nvcc boolean-expr-is-constant warning

---
 Eigen/src/Core/util/DisableStupidWarnings.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index b91d1d1af..8ef0f3594 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -55,6 +55,7 @@
 #endif
 
 #if defined __NVCC__
+  #pragma diag_suppress boolean_controlling_expr_is_constant
   // Disable the "statement is unreachable" message
   #pragma diag_suppress code_is_unreachable
   // Disable the "dynamic initialization in unreachable code" message

From 3182bdbae68b11aa8278107ca71c5df2de789e66 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Jul 2017 11:01:28 +0200
Subject: [PATCH 004/680] Disable vectorization when compiled by nvcc, even is
 EIGEN_NO_CUDA is defined

---
 Eigen/Core | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/Eigen/Core b/Eigen/Core
index d2edbd2bc..f613ee9aa 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -55,6 +55,10 @@
   #define EIGEN_DEVICE_FUNC
 #endif
 
+#ifdef __NVCC__
+#define EIGEN_DONT_VECTORIZE
+#endif
+
 // When compiling CUDA device code with NVCC, pull in math functions from the
 // global namespace.  In host mode, and when device doee with clang, use the
 // std versions.

From a74b9ba7cda08b8fbcb187aa5e96f0e99cf9b684 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Jul 2017 11:05:26 +0200
Subject: [PATCH 005/680] Update documentation for CUDA

---
 doc/PreprocessorDirectives.dox |  2 ++
 doc/UsingNVCC.dox              | 12 +++++-------
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox
index f01b39aec..0919d4190 100644
--- a/doc/PreprocessorDirectives.dox
+++ b/doc/PreprocessorDirectives.dox
@@ -120,6 +120,8 @@ run time. However, these assertions do cost time and can thus be turned off.
  - \b \c EIGEN_STACK_ALLOCATION_LIMIT - defines the maximum bytes for a buffer to be allocated on the stack. For internal
    temporary buffers, dynamic memory allocation is employed as a fall back. For fixed-size matrices or arrays, exceeding
    this threshold raises a compile time assertion. Use 0 to set no limit. Default is 128 KB.
+ - \b \c EIGEN_NO_CUDA - disables CUDA support when defined. Might be useful in .cu files for which Eigen is used on the host only,
+   and never called from device code.
 
 
  - \c EIGEN_DONT_ALIGN - Deprecated, it is a synonym for \c EIGEN_MAX_ALIGN_BYTES=0. It disables alignment completely. %Eigen will not try to align its objects and does not expect that any objects passed to it are aligned. This will turn off vectorization if \b EIGEN_UNALIGNED_VECTORIZE=1. Not defined by default.
diff --git a/doc/UsingNVCC.dox b/doc/UsingNVCC.dox
index f8e755b79..9bcdf0bfc 100644
--- a/doc/UsingNVCC.dox
+++ b/doc/UsingNVCC.dox
@@ -3,18 +3,16 @@ namespace Eigen {
 
 /** \page TopicCUDA Using Eigen in CUDA kernels
 
-\b Disclaimer: this page is about an \b experimental feature in %Eigen.
-
-Staring from CUDA 5.0, the CUDA compiler, \c nvcc, is able to properly parse %Eigen's code (almost).
-A few adaptations of the %Eigen's code already allows to use some parts of %Eigen in your own CUDA kernels.
-To this end you need the devel branch of %Eigen, CUDA 5.0 or greater with GCC.
+Staring from CUDA 5.5 and Eigen 3.3, it is possible to use Eigen's matrices, vectors, and arrays for fixed size within CUDA kernels. This is especially useful when working on numerous but small problems. By default, when Eigen's headers are included within a .cu file compiled by nvcc most Eigen's functions and methods are prefixed by the \c __device__ \c __host__ keywords making them callable from both host and device code.
+This support can be disabled by defining \c EIGEN_NO_CUDA before including any Eigen's header.
+This might be usefull to disable some warnings when a .cu file makes use of Eigen on the host side only.
+However, in both cases, host's SIMD vectorization has to be disabled in .cu files.
+It is thus \b strongly \b recommended to properly move all costly host computation from your .cu files to regular .cpp files.
 
 Known issues:
 
  - \c nvcc with MS Visual Studio does not work (patch welcome)
  
- - \c nvcc with \c clang does not work (patch welcome)
- 
  - \c nvcc 5.5 with gcc-4.7 (or greater) has issues with the standard \c \<limits\> header file. To workaround this, you can add the following before including any other files:
    \code
     // workaround issue between gcc >= 4.7 and cuda 5.5

From cda47c42c23035130488fd9a4437b2c1910d0bab Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Jul 2017 21:08:20 +0200
Subject: [PATCH 006/680] Fix compilation in c++98 mode.

---
 Eigen/src/Core/MathFunctions.h | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index a906b7629..5ba5293a0 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1378,13 +1378,14 @@ T acos(const T &x) {
   return acos(x);
 }
 
-
+#if EIGEN_HAS_CXX11_MATH
 template<typename T>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T acosh(const T &x) {
   EIGEN_USING_STD_MATH(acosh);
   return acosh(x);
 }
+#endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   acos(float x) { return cl::sycl::acos(x); }
@@ -1408,12 +1409,14 @@ T asin(const T &x) {
   return asin(x);
 }
 
+#if EIGEN_HAS_CXX11_MATH
 template<typename T>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T asinh(const T &x) {
   EIGEN_USING_STD_MATH(asinh);
   return asinh(x);
 }
+#endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   asin(float x) { return cl::sycl::asin(x); }
@@ -1437,12 +1440,14 @@ T atan(const T &x) {
   return atan(x);
 }
 
+#if EIGEN_HAS_CXX11_MATH
 template<typename T>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T atanh(const T &x) {
   EIGEN_USING_STD_MATH(atanh);
   return atanh(x);
 }
+#endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   atan(float x) { return cl::sycl::atan(x); }

From 55d7181557aebe43fb009ff8669f32741e15cfba Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Jul 2017 09:47:28 +0200
Subject: [PATCH 007/680] Fix lazyness of operator* with CUDA

---
 Eigen/src/Core/GeneralProduct.h | 21 +++++++++++----------
 Eigen/src/Core/MatrixBase.h     |  9 ---------
 2 files changed, 11 insertions(+), 19 deletions(-)

diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 95dadfea8..dec24848d 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -24,12 +24,17 @@ template<int Rows, int Cols, int Depth> struct product_type_selector;
 
 template<int Size, int MaxSize> struct product_size_category
 {
-  enum { is_large = MaxSize == Dynamic ||
-                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
-                    (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
-         value = is_large  ? Large
-               : Size == 1 ? 1
-                           : Small
+  enum {
+    #ifndef EIGEN_CUDA_ARCH
+    is_large = MaxSize == Dynamic ||
+               Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
+               (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
+    #else
+    is_large = 0,
+    #endif
+    value = is_large  ? Large
+          : Size == 1 ? 1
+                      : Small
   };
 };
 
@@ -379,8 +384,6 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
   *
   * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
   */
-#ifndef EIGEN_CUDACC
-
 template<typename Derived>
 template<typename OtherDerived>
 inline const Product<Derived, OtherDerived>
@@ -412,8 +415,6 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   return Product<Derived, OtherDerived>(derived(), other.derived());
 }
 
-#endif // EIGEN_CUDACC
-
 /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
   *
   * The returned product will behave like any other expressions: the coefficients of the product will be
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 908d35dfe..5a2146455 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -160,20 +160,11 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator-=(const MatrixBase<OtherDerived>& other);
 
-#ifdef EIGEN_CUDACC
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    const Product<Derived,OtherDerived,LazyProduct>
-    operator*(const MatrixBase<OtherDerived> &other) const
-    { return this->lazyProduct(other); }
-#else
-
-    template<typename OtherDerived>
     const Product<Derived,OtherDerived>
     operator*(const MatrixBase<OtherDerived> &other) const;
 
-#endif
-
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     const Product<Derived,OtherDerived,LazyProduct>

From d580a90c9ab5ed5521a79670f73bcea5ee755fe0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Jul 2017 10:03:54 +0200
Subject: [PATCH 008/680] Disable BDCSVD preallocation check.

---
 test/bdcsvd.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/test/bdcsvd.cpp b/test/bdcsvd.cpp
index f9f687aac..6c7b09696 100644
--- a/test/bdcsvd.cpp
+++ b/test/bdcsvd.cpp
@@ -104,7 +104,8 @@ void test_bdcsvd()
   CALL_SUBTEST_7( BDCSVD<MatrixXf>(10,10) );
 
   // Check that preallocation avoids subsequent mallocs
-  CALL_SUBTEST_9( svd_preallocate<void>() );
+  // Disbaled because not supported by BDCSVD
+  // CALL_SUBTEST_9( svd_preallocate<void>() );
 
   CALL_SUBTEST_2( svd_underoverflow<void>() );
 }

From 1f4b24d2df6200c1074c2fca45d3905a33c54a3b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Jul 2017 10:13:48 +0200
Subject: [PATCH 009/680] Do not preallocate more space than the matrix size
 (when the sparse matrix boils down to a vector

---
 Eigen/src/SparseCore/SparseAssign.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h
index 18352a847..113463258 100644
--- a/Eigen/src/SparseCore/SparseAssign.h
+++ b/Eigen/src/SparseCore/SparseAssign.h
@@ -83,7 +83,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src)
     // eval without temporary
     dst.resize(src.rows(), src.cols());
     dst.setZero();
-    dst.reserve((std::max)(src.rows(),src.cols())*2);
+    dst.reserve((std::min)(src.rows()*src.cols(), (std::max)(src.rows(),src.cols())*2));
     for (Index j=0; j<outerEvaluationSize; ++j)
     {
       dst.startVec(j);
@@ -107,7 +107,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src)
     
     DstXprType temp(src.rows(), src.cols());
 
-    temp.reserve((std::max)(src.rows(),src.cols())*2);
+    temp.reserve((std::min)(src.rows()*src.cols(), (std::max)(src.rows(),src.cols())*2));
     for (Index j=0; j<outerEvaluationSize; ++j)
     {
       temp.startVec(j);

From 687bedfcadc0e30cba63539a2d6b3c2315ea4f0a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 11:51:22 +0200
Subject: [PATCH 010/680] Make NoAlias and JacobiRotation compatible with CUDA.

---
 Eigen/src/Core/MatrixBase.h |  4 +++-
 Eigen/src/Core/NoAlias.h    |  1 +
 Eigen/src/Jacobi/Jacobi.h   | 19 +++++++++++++++----
 3 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 5a2146455..5786f76e0 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -296,7 +296,7 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase<OtherDerived>& other) const
     { return cwiseNotEqual(other).any(); }
 
-    NoAlias<Derived,Eigen::MatrixBase > noalias();
+    NoAlias<Derived,Eigen::MatrixBase > EIGEN_DEVICE_FUNC noalias();
 
     // TODO forceAlignedAccess is temporarily disabled
     // Need to find a nicer workaround.
@@ -428,8 +428,10 @@ template<typename Derived> class MatrixBase
 ///////// Jacobi module /////////
 
     template<typename OtherScalar>
+    EIGEN_DEVICE_FUNC
     void applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j);
     template<typename OtherScalar>
+    EIGEN_DEVICE_FUNC
     void applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j);
 
 ///////// SparseCore module /////////
diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h
index 33908010b..41fae5096 100644
--- a/Eigen/src/Core/NoAlias.h
+++ b/Eigen/src/Core/NoAlias.h
@@ -33,6 +33,7 @@ class NoAlias
   public:
     typedef typename ExpressionType::Scalar Scalar;
     
+    EIGEN_DEVICE_FUNC
     explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
     
     template<typename OtherDerived>
diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index c30326e1d..75595517a 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -37,17 +37,20 @@ template<typename Scalar> class JacobiRotation
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
     /** Default constructor without any initialization. */
+    EIGEN_DEVICE_FUNC
     JacobiRotation() {}
 
     /** Construct a planar rotation from a cosine-sine pair (\a c, \c s). */
+    EIGEN_DEVICE_FUNC
     JacobiRotation(const Scalar& c, const Scalar& s) : m_c(c), m_s(s) {}
 
-    Scalar& c() { return m_c; }
-    Scalar c() const { return m_c; }
-    Scalar& s() { return m_s; }
-    Scalar s() const { return m_s; }
+    EIGEN_DEVICE_FUNC Scalar& c() { return m_c; }
+    EIGEN_DEVICE_FUNC Scalar c() const { return m_c; }
+    EIGEN_DEVICE_FUNC Scalar& s() { return m_s; }
+    EIGEN_DEVICE_FUNC Scalar s() const { return m_s; }
 
     /** Concatenates two planar rotation */
+    EIGEN_DEVICE_FUNC
     JacobiRotation operator*(const JacobiRotation& other)
     {
       using numext::conj;
@@ -56,19 +59,26 @@ template<typename Scalar> class JacobiRotation
     }
 
     /** Returns the transposed transformation */
+    EIGEN_DEVICE_FUNC
     JacobiRotation transpose() const { using numext::conj; return JacobiRotation(m_c, -conj(m_s)); }
 
     /** Returns the adjoint transformation */
+    EIGEN_DEVICE_FUNC
     JacobiRotation adjoint() const { using numext::conj; return JacobiRotation(conj(m_c), -m_s); }
 
     template<typename Derived>
+    EIGEN_DEVICE_FUNC
     bool makeJacobi(const MatrixBase<Derived>&, Index p, Index q);
+    EIGEN_DEVICE_FUNC
     bool makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z);
 
+    EIGEN_DEVICE_FUNC
     void makeGivens(const Scalar& p, const Scalar& q, Scalar* z=0);
 
   protected:
+    EIGEN_DEVICE_FUNC
     void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::true_type);
+    EIGEN_DEVICE_FUNC
     void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::false_type);
 
     Scalar m_c, m_s;
@@ -264,6 +274,7 @@ namespace internal {
   * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
   */
 template<typename VectorX, typename VectorY, typename OtherScalar>
+EIGEN_DEVICE_FUNC
 void apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j);
 }
 

From 89c01a494aff2bd03b48a9858eed95a4a7ce9556 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 11:55:00 +0200
Subject: [PATCH 011/680] Add unit test for has_ReturnType

---
 test/meta.cpp | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/test/meta.cpp b/test/meta.cpp
index b8dea68e8..bd505762e 100644
--- a/test/meta.cpp
+++ b/test/meta.cpp
@@ -15,6 +15,10 @@ bool check_is_convertible(const From&, const To&)
   return internal::is_convertible<From,To>::value;
 }
 
+struct FooReturnType {
+  typedef int ReturnType;
+};
+
 void test_meta()
 {
   VERIFY((internal::conditional<(3<4),internal::true_type, internal::false_type>::type::value));
@@ -75,6 +79,11 @@ void test_meta()
     VERIFY((!check_is_convertible(A*B, f) ));
     VERIFY(( check_is_convertible(A*B, A) ));
   }
+
+  VERIFY((  internal::has_ReturnType<FooReturnType>::value ));
+  VERIFY((  internal::has_ReturnType<ScalarBinaryOpTraits<int,int> >::value ));
+  VERIFY(( !internal::has_ReturnType<MatrixXf>::value ));
+  VERIFY(( !internal::has_ReturnType<int>::value ));
   
   VERIFY(internal::meta_sqrt<1>::ret == 1);
   #define VERIFY_META_SQRT(X) VERIFY(internal::meta_sqrt<X>::ret == int(std::sqrt(double(X))))

From b95f92843c58a914c46ab091009146288b8b775c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 12:07:10 +0200
Subject: [PATCH 012/680] Fix support for MKL's BLAS when using
 MKL_DIRECT_CALL.

---
 .../GeneralMatrixMatrixTriangular_BLAS.h      |  8 +++-
 .../Core/products/GeneralMatrixMatrix_BLAS.h  | 19 +++++---
 .../Core/products/GeneralMatrixVector_BLAS.h  | 19 +++++---
 .../products/SelfadjointMatrixMatrix_BLAS.h   | 48 ++++++++++++-------
 .../products/SelfadjointMatrixVector_BLAS.h   |  9 +++-
 .../products/TriangularMatrixMatrix_BLAS.h    | 39 ++++++++++-----
 .../products/TriangularMatrixVector_BLAS.h    | 46 +++++++++++-------
 .../products/TriangularSolverMatrix_BLAS.h    | 40 ++++++++++------
 Eigen/src/Core/util/MKL_support.h             |  8 ++--
 9 files changed, 157 insertions(+), 79 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
index 41e18ff07..9176a1382 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
@@ -88,7 +88,7 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
    BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
    char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \
    EIGTYPE beta(1); \
-   BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \
+   BLASFUNC(&uplo, &trans, &n, &k, (const BLASTYPE*)&numext::real_ref(alpha), lhs, &lda, (const BLASTYPE*)&numext::real_ref(beta), res, &ldc); \
   } \
 };
 
@@ -125,9 +125,13 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
   } \
 };
 
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_RANKUPDATE_R(double, double, dsyrk)
+EIGEN_BLAS_RANKUPDATE_R(float,  float,  ssyrk)
+#else
 EIGEN_BLAS_RANKUPDATE_R(double, double, dsyrk_)
 EIGEN_BLAS_RANKUPDATE_R(float,  float,  ssyrk_)
+#endif
 
 // TODO hanlde complex cases
 // EIGEN_BLAS_RANKUPDATE_C(dcomplex, double, double, zherk_)
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h b/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
index 7a3bdbf20..b0f6b0d5b 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
@@ -46,7 +46,7 @@ namespace internal {
 
 // gemm specialization
 
-#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, BLASTYPE, BLASPREFIX) \
+#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, BLASTYPE, BLASFUNC) \
 template< \
   typename Index, \
   int LhsStorageOrder, bool ConjugateLhs, \
@@ -100,13 +100,20 @@ static void run(Index rows, Index cols, Index depth, \
     ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
   } else b = _rhs; \
 \
-  BLASPREFIX##gemm_(&transa, &transb, &m, &n, &k, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+  BLASFUNC(&transa, &transb, &m, &n, &k, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 }};
 
-GEMM_SPECIALIZATION(double,   d,  double, d)
-GEMM_SPECIALIZATION(float,    f,  float,  s)
-GEMM_SPECIALIZATION(dcomplex, cd, double, z)
-GEMM_SPECIALIZATION(scomplex, cf, float,  c)
+#ifdef EIGEN_USE_MKL
+GEMM_SPECIALIZATION(double,   d,  double, dgemm)
+GEMM_SPECIALIZATION(float,    f,  float,  sgemm)
+GEMM_SPECIALIZATION(dcomplex, cd, MKL_Complex16, zgemm)
+GEMM_SPECIALIZATION(scomplex, cf, MKL_Complex8,  cgemm)
+#else
+GEMM_SPECIALIZATION(double,   d,  double, dgemm_)
+GEMM_SPECIALIZATION(float,    f,  float,  sgemm_)
+GEMM_SPECIALIZATION(dcomplex, cd, double, zgemm_)
+GEMM_SPECIALIZATION(scomplex, cf, float,  cgemm_)
+#endif
 
 } // end namespase internal
 
diff --git a/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h b/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
index e3a5d5892..6e36c2b3c 100644
--- a/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
+++ b/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
@@ -85,7 +85,7 @@ EIGEN_BLAS_GEMV_SPECIALIZE(float)
 EIGEN_BLAS_GEMV_SPECIALIZE(dcomplex)
 EIGEN_BLAS_GEMV_SPECIALIZE(scomplex)
 
-#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASPREFIX) \
+#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \
 template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \
 struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \
 { \
@@ -113,14 +113,21 @@ static void run( \
     x_ptr=x_tmp.data(); \
     incx=1; \
   } else x_ptr=rhs; \
-  BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
+  BLASFUNC(&trans, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &incy); \
 }\
 };
 
-EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, d)
-EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  s)
-EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, z)
-EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float,  c)
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, dgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  sgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, MKL_Complex16, zgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, MKL_Complex8 , cgemv)
+#else
+EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, dgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  sgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, zgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float,  cgemv_)
+#endif
 
 } // end namespase internal
 
diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
index a45238d69..9a5318507 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
@@ -40,7 +40,7 @@ namespace internal {
 
 /* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */
 
-#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -81,13 +81,13 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
       ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } else b = _rhs; \
 \
-    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+    BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
 
-#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -144,20 +144,26 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
       ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } \
 \
-    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+    BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
-EIGEN_BLAS_SYMM_L(double, double, d, d)
-EIGEN_BLAS_SYMM_L(float, float, f, s)
-EIGEN_BLAS_HEMM_L(dcomplex, double, cd, z)
-EIGEN_BLAS_HEMM_L(scomplex, float, cf, c)
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_SYMM_L(double, double, d, dsymm)
+EIGEN_BLAS_SYMM_L(float, float, f, ssymm)
+EIGEN_BLAS_HEMM_L(dcomplex, MKL_Complex16, cd, zhemm)
+EIGEN_BLAS_HEMM_L(scomplex, MKL_Complex8, cf, chemm)
+#else
+EIGEN_BLAS_SYMM_L(double, double, d, dsymm_)
+EIGEN_BLAS_SYMM_L(float, float, f, ssymm_)
+EIGEN_BLAS_HEMM_L(dcomplex, double, cd, zhemm_)
+EIGEN_BLAS_HEMM_L(scomplex, float, cf, chemm_)
+#endif
 
 /* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */
 
-#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -197,13 +203,13 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
       ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } else b = _lhs; \
 \
-    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+    BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
 
-#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -259,15 +265,21 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
       ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } \
 \
-    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+    BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
   } \
 };
 
-EIGEN_BLAS_SYMM_R(double, double, d, d)
-EIGEN_BLAS_SYMM_R(float, float, f, s)
-EIGEN_BLAS_HEMM_R(dcomplex, double, cd, z)
-EIGEN_BLAS_HEMM_R(scomplex, float, cf, c)
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_SYMM_R(double, double, d, dsymm)
+EIGEN_BLAS_SYMM_R(float, float, f, ssymm)
+EIGEN_BLAS_HEMM_R(dcomplex, MKL_Complex16, cd, zhemm)
+EIGEN_BLAS_HEMM_R(scomplex, MKL_Complex8, cf, chemm)
+#else
+EIGEN_BLAS_SYMM_R(double, double, d, dsymm_)
+EIGEN_BLAS_SYMM_R(float, float, f, ssymm_)
+EIGEN_BLAS_HEMM_R(dcomplex, double, cd, zhemm_)
+EIGEN_BLAS_HEMM_R(scomplex, float, cf, chemm_)
+#endif
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h b/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
index 38f23accf..1238345e3 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
@@ -95,14 +95,21 @@ const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \
     x_tmp=map_x.conjugate(); \
     x_ptr=x_tmp.data(); \
   } else x_ptr=_rhs; \
-  BLASFUNC(&uplo, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
+  BLASFUNC(&uplo, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &incy); \
 }\
 };
 
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, MKL_Complex16, zhemv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, MKL_Complex8,  chemv)
+#else
 EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv_)
 EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv_)
 EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_)
 EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float,  chemv_)
+#endif
 
 } // end namespace internal
 
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h b/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
index aecded6bb..a25197ab0 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
@@ -75,7 +75,7 @@ EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true)
 EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false)
 
 // implements col-major += alpha * op(triangular) * op(general)
-#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, int Mode, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -172,7 +172,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
    } \
    /*std::cout << "TRMM_L: A is square! Go to BLAS TRMM implementation! \n";*/ \
 /* call ?trmm*/ \
-   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
 \
 /* Add op(a_triangular)*b into res*/ \
    Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
@@ -180,13 +180,20 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
   } \
 };
 
-EIGEN_BLAS_TRMM_L(double, double, d, d)
-EIGEN_BLAS_TRMM_L(dcomplex, double, cd, z)
-EIGEN_BLAS_TRMM_L(float, float, f, s)
-EIGEN_BLAS_TRMM_L(scomplex, float, cf, c)
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMM_L(double, double, d, dtrmm)
+EIGEN_BLAS_TRMM_L(dcomplex, MKL_Complex16, cd, ztrmm)
+EIGEN_BLAS_TRMM_L(float, float, f, strmm)
+EIGEN_BLAS_TRMM_L(scomplex, MKL_Complex8, cf, ctrmm)
+#else
+EIGEN_BLAS_TRMM_L(double, double, d, dtrmm_)
+EIGEN_BLAS_TRMM_L(dcomplex, double, cd, ztrmm_)
+EIGEN_BLAS_TRMM_L(float, float, f, strmm_)
+EIGEN_BLAS_TRMM_L(scomplex, float, cf, ctrmm_)
+#endif
 
 // implements col-major += alpha * op(general) * op(triangular)
-#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
 template <typename Index, int Mode, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -282,7 +289,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
    } \
    /*std::cout << "TRMM_R: A is square! Go to BLAS TRMM implementation! \n";*/ \
 /* call ?trmm*/ \
-   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
 \
 /* Add op(a_triangular)*b into res*/ \
    Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
@@ -290,11 +297,17 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
   } \
 };
 
-EIGEN_BLAS_TRMM_R(double, double, d, d)
-EIGEN_BLAS_TRMM_R(dcomplex, double, cd, z)
-EIGEN_BLAS_TRMM_R(float, float, f, s)
-EIGEN_BLAS_TRMM_R(scomplex, float, cf, c)
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMM_R(double, double, d, dtrmm)
+EIGEN_BLAS_TRMM_R(dcomplex, MKL_Complex16, cd, ztrmm)
+EIGEN_BLAS_TRMM_R(float, float, f, strmm)
+EIGEN_BLAS_TRMM_R(scomplex, MKL_Complex8, cf, ctrmm)
+#else
+EIGEN_BLAS_TRMM_R(double, double, d, dtrmm_)
+EIGEN_BLAS_TRMM_R(dcomplex, double, cd, ztrmm_)
+EIGEN_BLAS_TRMM_R(float, float, f, strmm_)
+EIGEN_BLAS_TRMM_R(scomplex, float, cf, ctrmm_)
+#endif
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h b/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
index 07bf26ce5..3d47a2b94 100644
--- a/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
+++ b/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
@@ -71,7 +71,7 @@ EIGEN_BLAS_TRMV_SPECIALIZE(dcomplex)
 EIGEN_BLAS_TRMV_SPECIALIZE(scomplex)
 
 // implements col-major: res += alpha * op(triangular) * vector
-#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX, BLASPOSTFIX) \
 template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
 struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor> { \
   enum { \
@@ -121,10 +121,10 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
    diag = IsUnitDiag ? 'U' : 'N'; \
 \
 /* call ?TRMV*/ \
-   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
+   BLASPREFIX##trmv##BLASPOSTFIX(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
 \
 /* Add op(a_tr)rhs into res*/ \
-   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+   BLASPREFIX##axpy##BLASPOSTFIX(&n, (const BLASTYPE*)&numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
 /* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
    if (size<(std::max)(rows,cols)) { \
      if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
@@ -142,18 +142,25 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
        m = convert_index<BlasIndex>(size); \
        n = convert_index<BlasIndex>(cols-size); \
      } \
-     BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
+     BLASPREFIX##gemv##BLASPOSTFIX(&trans, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)y, &incy); \
    } \
   } \
 };
 
-EIGEN_BLAS_TRMV_CM(double,   double, d,  d)
-EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z)
-EIGEN_BLAS_TRMV_CM(float,    float,  f,  s)
-EIGEN_BLAS_TRMV_CM(scomplex, float,  cf, c)
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMV_CM(double,   double, d,  d,)
+EIGEN_BLAS_TRMV_CM(dcomplex, MKL_Complex16, cd, z,)
+EIGEN_BLAS_TRMV_CM(float,    float,  f,  s,)
+EIGEN_BLAS_TRMV_CM(scomplex, MKL_Complex8,  cf, c,)
+#else
+EIGEN_BLAS_TRMV_CM(double,   double, d,  d, _)
+EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z, _)
+EIGEN_BLAS_TRMV_CM(float,    float,  f,  s, _)
+EIGEN_BLAS_TRMV_CM(scomplex, float,  cf, c, _)
+#endif
 
 // implements row-major: res += alpha * op(triangular) * vector
-#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX, BLASPOSTFIX) \
 template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
 struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor> { \
   enum { \
@@ -203,10 +210,10 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
    diag = IsUnitDiag ? 'U' : 'N'; \
 \
 /* call ?TRMV*/ \
-   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
+   BLASPREFIX##trmv##BLASPOSTFIX(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
 \
 /* Add op(a_tr)rhs into res*/ \
-   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+   BLASPREFIX##axpy##BLASPOSTFIX(&n, (const BLASTYPE*)&numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
 /* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
    if (size<(std::max)(rows,cols)) { \
      if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
@@ -224,15 +231,22 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
        m = convert_index<BlasIndex>(size); \
        n = convert_index<BlasIndex>(cols-size); \
      } \
-     BLASPREFIX##gemv_(&trans, &n, &m, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
+     BLASPREFIX##gemv##BLASPOSTFIX(&trans, &n, &m, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)y, &incy); \
    } \
   } \
 };
 
-EIGEN_BLAS_TRMV_RM(double,   double, d,  d)
-EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z)
-EIGEN_BLAS_TRMV_RM(float,    float,  f,  s)
-EIGEN_BLAS_TRMV_RM(scomplex, float,  cf, c)
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMV_RM(double,   double, d,  d,)
+EIGEN_BLAS_TRMV_RM(dcomplex, MKL_Complex16, cd, z,)
+EIGEN_BLAS_TRMV_RM(float,    float,  f,  s,)
+EIGEN_BLAS_TRMV_RM(scomplex, MKL_Complex8,  cf, c,)
+#else
+EIGEN_BLAS_TRMV_RM(double,   double, d,  d,_)
+EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z,_)
+EIGEN_BLAS_TRMV_RM(float,    float,  f,  s,_)
+EIGEN_BLAS_TRMV_RM(scomplex, float,  cf, c,_)
+#endif
 
 } // end namespase internal
 
diff --git a/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h b/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
index 88c0fb794..f0775116a 100644
--- a/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
+++ b/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
@@ -38,7 +38,7 @@ namespace Eigen {
 namespace internal {
 
 // implements LeftSide op(triangular)^-1 * general
-#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASPREFIX) \
+#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASFUNC) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
 struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
 { \
@@ -80,18 +80,24 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorage
    } \
    if (IsUnitDiag) diag='U'; \
 /* call ?trsm*/ \
-   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
  } \
 };
 
-EIGEN_BLAS_TRSM_L(double,   double, d)
-EIGEN_BLAS_TRSM_L(dcomplex, double, z)
-EIGEN_BLAS_TRSM_L(float,    float,  s)
-EIGEN_BLAS_TRSM_L(scomplex, float,  c)
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRSM_L(double,   double, dtrsm)
+EIGEN_BLAS_TRSM_L(dcomplex, MKL_Complex16, ztrsm)
+EIGEN_BLAS_TRSM_L(float,    float,  strsm)
+EIGEN_BLAS_TRSM_L(scomplex, MKL_Complex8, ctrsm)
+#else
+EIGEN_BLAS_TRSM_L(double,   double, dtrsm_)
+EIGEN_BLAS_TRSM_L(dcomplex, double, ztrsm_)
+EIGEN_BLAS_TRSM_L(float,    float,  strsm_)
+EIGEN_BLAS_TRSM_L(scomplex, float,  ctrsm_)
+#endif
 
 // implements RightSide general * op(triangular)^-1
-#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASPREFIX) \
+#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASFUNC) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
 struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
 { \
@@ -133,16 +139,22 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorag
    } \
    if (IsUnitDiag) diag='U'; \
 /* call ?trsm*/ \
-   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
    /*std::cout << "TRMS_L specialization!\n";*/ \
  } \
 };
 
-EIGEN_BLAS_TRSM_R(double,   double, d)
-EIGEN_BLAS_TRSM_R(dcomplex, double, z)
-EIGEN_BLAS_TRSM_R(float,    float,  s)
-EIGEN_BLAS_TRSM_R(scomplex, float,  c)
-
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRSM_R(double,   double, dtrsm)
+EIGEN_BLAS_TRSM_R(dcomplex, MKL_Complex16, ztrsm)
+EIGEN_BLAS_TRSM_R(float,    float,  strsm)
+EIGEN_BLAS_TRSM_R(scomplex, MKL_Complex8,  ctrsm)
+#else
+EIGEN_BLAS_TRSM_R(double,   double, dtrsm_)
+EIGEN_BLAS_TRSM_R(dcomplex, double, ztrsm_)
+EIGEN_BLAS_TRSM_R(float,    float,  strsm_)
+EIGEN_BLAS_TRSM_R(scomplex, float,  ctrsm_)
+#endif
 
 } // end namespace internal
 
diff --git a/Eigen/src/Core/util/MKL_support.h b/Eigen/src/Core/util/MKL_support.h
index 26b59669e..e656799bf 100755
--- a/Eigen/src/Core/util/MKL_support.h
+++ b/Eigen/src/Core/util/MKL_support.h
@@ -53,6 +53,7 @@
   #define EIGEN_USE_MKL
 #endif
 
+
 #if defined EIGEN_USE_MKL
 #   include <mkl.h> 
 /*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/
@@ -108,6 +109,10 @@
 #endif
 #endif
 
+#if defined(EIGEN_USE_BLAS) && !defined(EIGEN_USE_MKL)
+#include "../../misc/blas.h"
+#endif
+
 namespace Eigen {
 
 typedef std::complex<double> dcomplex;
@@ -121,8 +126,5 @@ typedef int BlasIndex;
 
 } // end namespace Eigen
 
-#if defined(EIGEN_USE_BLAS)
-#include "../../misc/blas.h"
-#endif
 
 #endif // EIGEN_MKL_SUPPORT_H

From 8c858bd8919936f250d2e7b090c0d17f00dbb85b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 12:17:45 +0200
Subject: [PATCH 013/680] Clarify doc regarding the usage of MKL_DIRECT_CALL

---
 Eigen/src/Core/util/MKL_support.h | 2 +-
 doc/UsingIntelMKL.dox             | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/MKL_support.h b/Eigen/src/Core/util/MKL_support.h
index e656799bf..b7d6ecc76 100755
--- a/Eigen/src/Core/util/MKL_support.h
+++ b/Eigen/src/Core/util/MKL_support.h
@@ -49,7 +49,7 @@
   #define EIGEN_USE_LAPACKE
 #endif
 
-#if defined(EIGEN_USE_MKL_VML)
+#if defined(EIGEN_USE_MKL_VML) && !defined(EIGEN_USE_MKL)
   #define EIGEN_USE_MKL
 #endif
 
diff --git a/doc/UsingIntelMKL.dox b/doc/UsingIntelMKL.dox
index a1a3a18f2..53e5de42c 100644
--- a/doc/UsingIntelMKL.dox
+++ b/doc/UsingIntelMKL.dox
@@ -63,6 +63,8 @@ In addition you can choose which parts will be substituted by defining one or mu
 <tr><td>\c EIGEN_USE_MKL_ALL </td><td>Defines \c EIGEN_USE_BLAS, \c EIGEN_USE_LAPACKE, and \c EIGEN_USE_MKL_VML </td></tr>
 </table>
 
+In order to be able to use \b MKL_DIRECT_CALL for BLAS level 2 and 3 routines, then the macro \c EIGEN_USE_MKL must also be defined in the case none of the other \c EIGEN_USE_MKL_* macros has been defined. This is needed to tell Eigen that the BLAS backend is the MKL and that the MKL interface must be used instead of the generic F77 one.
+
 Note that the BLAS and LAPACKE backends can be enabled for any F77 compatible BLAS and LAPACK libraries. See this \link TopicUsingBlasLapack page \endlink for the details.
 
 Finally, the PARDISO sparse solver shipped with Intel MKL can be used through the \ref PardisoLU, \ref PardisoLLT and \ref PardisoLDLT classes of the \ref PardisoSupport_Module.

From f727844658f8c9c01302b5cb08d81c62c572b82b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 21:58:39 +0200
Subject: [PATCH 014/680] use MKL's lapacke.h header when using MKL

---
 Eigen/Cholesky    | 4 ++++
 Eigen/Eigenvalues | 4 ++++
 Eigen/LU          | 4 ++++
 Eigen/QR          | 4 ++++
 Eigen/SVD         | 4 ++++
 5 files changed, 20 insertions(+)

diff --git a/Eigen/Cholesky b/Eigen/Cholesky
index 369d1f5ec..bb899a2ce 100644
--- a/Eigen/Cholesky
+++ b/Eigen/Cholesky
@@ -31,7 +31,11 @@
 #include "src/Cholesky/LLT.h"
 #include "src/Cholesky/LDLT.h"
 #ifdef EIGEN_USE_LAPACKE
+#ifdef EIGEN_USE_MKL
+#include "mkl_lapacke.h"
+#else
 #include "src/misc/lapacke.h"
+#endif
 #include "src/Cholesky/LLT_LAPACKE.h"
 #endif
 
diff --git a/Eigen/Eigenvalues b/Eigen/Eigenvalues
index 009e529e1..f3f661b07 100644
--- a/Eigen/Eigenvalues
+++ b/Eigen/Eigenvalues
@@ -45,7 +45,11 @@
 #include "src/Eigenvalues/GeneralizedEigenSolver.h"
 #include "src/Eigenvalues/MatrixBaseEigenvalues.h"
 #ifdef EIGEN_USE_LAPACKE
+#ifdef EIGEN_USE_MKL
+#include "mkl_lapacke.h"
+#else
 #include "src/misc/lapacke.h"
+#endif
 #include "src/Eigenvalues/RealSchur_LAPACKE.h"
 #include "src/Eigenvalues/ComplexSchur_LAPACKE.h"
 #include "src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h"
diff --git a/Eigen/LU b/Eigen/LU
index 6f6c55629..6418a86e1 100644
--- a/Eigen/LU
+++ b/Eigen/LU
@@ -28,7 +28,11 @@
 #include "src/LU/FullPivLU.h"
 #include "src/LU/PartialPivLU.h"
 #ifdef EIGEN_USE_LAPACKE
+#ifdef EIGEN_USE_MKL
+#include "mkl_lapacke.h"
+#else
 #include "src/misc/lapacke.h"
+#endif
 #include "src/LU/PartialPivLU_LAPACKE.h"
 #endif
 #include "src/LU/Determinant.h"
diff --git a/Eigen/QR b/Eigen/QR
index 80838e3bd..c7e914469 100644
--- a/Eigen/QR
+++ b/Eigen/QR
@@ -36,7 +36,11 @@
 #include "src/QR/ColPivHouseholderQR.h"
 #include "src/QR/CompleteOrthogonalDecomposition.h"
 #ifdef EIGEN_USE_LAPACKE
+#ifdef EIGEN_USE_MKL
+#include "mkl_lapacke.h"
+#else
 #include "src/misc/lapacke.h"
+#endif
 #include "src/QR/HouseholderQR_LAPACKE.h"
 #include "src/QR/ColPivHouseholderQR_LAPACKE.h"
 #endif
diff --git a/Eigen/SVD b/Eigen/SVD
index 86143c23d..5d0e75f7f 100644
--- a/Eigen/SVD
+++ b/Eigen/SVD
@@ -37,7 +37,11 @@
 #include "src/SVD/JacobiSVD.h"
 #include "src/SVD/BDCSVD.h"
 #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT)
+#ifdef EIGEN_USE_MKL
+#include "mkl_lapacke.h"
+#else
 #include "src/misc/lapacke.h"
+#endif
 #include "src/SVD/JacobiSVD_LAPACKE.h"
 #endif
 

From 2810ba194be85af0012f786e6c032b2bfe432be9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 17 Aug 2017 22:12:26 +0200
Subject: [PATCH 015/680] Clarify MKL_DIRECT_CALL doc.

---
 doc/UsingIntelMKL.dox | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/UsingIntelMKL.dox b/doc/UsingIntelMKL.dox
index 53e5de42c..6de14afad 100644
--- a/doc/UsingIntelMKL.dox
+++ b/doc/UsingIntelMKL.dox
@@ -63,7 +63,7 @@ In addition you can choose which parts will be substituted by defining one or mu
 <tr><td>\c EIGEN_USE_MKL_ALL </td><td>Defines \c EIGEN_USE_BLAS, \c EIGEN_USE_LAPACKE, and \c EIGEN_USE_MKL_VML </td></tr>
 </table>
 
-In order to be able to use \b MKL_DIRECT_CALL for BLAS level 2 and 3 routines, then the macro \c EIGEN_USE_MKL must also be defined in the case none of the other \c EIGEN_USE_MKL_* macros has been defined. This is needed to tell Eigen that the BLAS backend is the MKL and that the MKL interface must be used instead of the generic F77 one.
+The options can be combined with \b MKL_DIRECT_CALL to enable MKL direct call feature. This may help to increase performance of some MKL BLAS (?GEMM, ?GEMV, ?TRSM, ?AXPY and ?DOT) and LAPACK (LU, Cholesky and QR) routines for very small matrices. To make it work properly, the macro \c EIGEN_USE_MKL must also be defined in the case none of the other \c EIGEN_USE_MKL_* macros has been defined.
 
 Note that the BLAS and LAPACKE backends can be enabled for any F77 compatible BLAS and LAPACK libraries. See this \link TopicUsingBlasLapack page \endlink for the details.
 

From e6021cc8cc6298196026119e8486c67ea2604376 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 11:32:55 +0200
Subject: [PATCH 016/680] bug #1458: fix documentation of LLT and LDLT info()
 method.

---
 Eigen/src/Cholesky/LDLT.h | 2 +-
 Eigen/src/Cholesky/LLT.h  | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h
index 9b4fdb414..968427b3a 100644
--- a/Eigen/src/Cholesky/LDLT.h
+++ b/Eigen/src/Cholesky/LDLT.h
@@ -248,7 +248,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
     /** \brief Reports whether previous computation was successful.
       *
       * \returns \c Success if computation was succesful,
-      *          \c NumericalIssue if the matrix.appears to be negative.
+      *          \c NumericalIssue if the factorization failed because of a zero pivot.
       */
     ComputationInfo info() const
     {
diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index e6c02d803..3b0b9faf4 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -177,7 +177,7 @@ template<typename _MatrixType, int _UpLo> class LLT
     /** \brief Reports whether previous computation was successful.
       *
       * \returns \c Success if computation was succesful,
-      *          \c NumericalIssue if the matrix.appears to be negative.
+      *          \c NumericalIssue if the matrix.appears not to be positive definite.
       */
     ComputationInfo info() const
     {

From a6e7a41a553d3663cefc45a5d2b509494d8adb37 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 11:37:32 +0200
Subject: [PATCH 017/680] bug #1455: Cholesky module depends on Jacobi for
 rank-updates.

---
 Eigen/Cholesky | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/Cholesky b/Eigen/Cholesky
index bb899a2ce..1332b540d 100644
--- a/Eigen/Cholesky
+++ b/Eigen/Cholesky
@@ -9,6 +9,7 @@
 #define EIGEN_CHOLESKY_MODULE_H
 
 #include "Core"
+#include "Jacobi"
 
 #include "src/Core/util/DisableStupidWarnings.h"
 

From 2c3d70d915a939d0da33ca22742a26c271adcb82 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 12:04:09 +0200
Subject: [PATCH 018/680] Re-enable hidden doc in LLT

---
 Eigen/src/Cholesky/LLT.h | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 3b0b9faf4..3de94f726 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -24,7 +24,7 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   *
   * \tparam _MatrixType the type of the matrix of which we are computing the LL^T Cholesky decomposition
   * \tparam _UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
-  *             The other triangular part won't be read.
+  *               The other triangular part won't be read.
   *
   * This class performs a LL^T Cholesky decomposition of a symmetric, positive definite
   * matrix A such that A = LL^* = U^*U, where L is lower triangular.
@@ -43,12 +43,11 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   *
   * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
   *
+  * Note that during the decomposition, only the lower (or upper, as defined by _UpLo) triangular part of A is considered.
+  * Therefore, the strict lower part does not have to store correct values.
+  *
   * \sa MatrixBase::llt(), SelfAdjointView::llt(), class LDLT
   */
- /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH)
-  * Note that during the decomposition, only the upper triangular part of A is considered. Therefore,
-  * the strict lower part does not have to store correct values.
-  */
 template<typename _MatrixType, int _UpLo> class LLT
 {
   public:

From 21d0a0bcf5eef2fb89f1ca48b65d52ec03e97272 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 12:46:35 +0200
Subject: [PATCH 019/680] bug #1456: add perf recommendation for LLT and
 storage format

---
 Eigen/src/Cholesky/LLT.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 3de94f726..52d101aff 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -41,6 +41,11 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   * Example: \include LLT_example.cpp
   * Output: \verbinclude LLT_example.out
   *
+  * \b Performance: for best performance, it is recommended to use a column-major storage format
+  * with the Lower triangular part (the default), or, equivalently, a row-major storage format,
+  * with the Upper triangular part. Otherwise, you might get a 20% slowdown for the full factorization
+  * step, and rank-updates can be up to 3 times slower.
+  *
   * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
   *
   * Note that during the decomposition, only the lower (or upper, as defined by _UpLo) triangular part of A is considered.

From e27f17bf5c921dca73b4d2dc1a90863b36292fdc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 13:27:37 +0200
Subject: [PATCH 020/680] Gub 1453: fix Map with non-default inner-stride but
 no outer-stride.

---
 Eigen/src/Cholesky/LLT.h |  2 +-
 Eigen/src/Core/Map.h     | 15 ++++++++---
 test/mapstride.cpp       | 57 +++++++++++++++++++++++++++++++++++++++-
 3 files changed, 68 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 52d101aff..7f29dad9c 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -42,7 +42,7 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   * Output: \verbinclude LLT_example.out
   *
   * \b Performance: for best performance, it is recommended to use a column-major storage format
-  * with the Lower triangular part (the default), or, equivalently, a row-major storage format,
+  * with the Lower triangular part (the default), or, equivalently, a row-major storage format
   * with the Upper triangular part. Otherwise, you might get a 20% slowdown for the full factorization
   * step, and rank-updates can be up to 3 times slower.
   *
diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index 06d196702..7ca6a9280 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -20,11 +20,17 @@ struct traits<Map<PlainObjectType, MapOptions, StrideType> >
 {
   typedef traits<PlainObjectType> TraitsBase;
   enum {
+    PlainObjectTypeInnerSize = ((traits<PlainObjectType>::Flags&RowMajorBit)==RowMajorBit)
+                             ? PlainObjectType::ColsAtCompileTime
+                             : PlainObjectType::RowsAtCompileTime,
+
     InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
                              ? int(PlainObjectType::InnerStrideAtCompileTime)
                              : int(StrideType::InnerStrideAtCompileTime),
     OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
-                             ? int(PlainObjectType::OuterStrideAtCompileTime)
+                             ? (InnerStrideAtCompileTime==Dynamic || PlainObjectTypeInnerSize==Dynamic
+                                ? Dynamic
+                                : int(InnerStrideAtCompileTime) * int(PlainObjectTypeInnerSize))
                              : int(StrideType::OuterStrideAtCompileTime),
     Alignment = int(MapOptions)&int(AlignedMask),
     Flags0 = TraitsBase::Flags & (~NestByRefBit),
@@ -108,9 +114,10 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
-           : IsVectorAtCompileTime ? this->size()
-           : int(Flags)&RowMajorBit ? this->cols()
-           : this->rows();
+           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
+           : IsVectorAtCompileTime ? (this->size() * innerStride())
+           : int(Flags)&RowMajorBit ? (this->cols() * innerStride())
+           : (this->rows() * innerStride());
     }
 
     /** Constructor in the fixed-size case.
diff --git a/test/mapstride.cpp b/test/mapstride.cpp
index 4858f8fea..de77dc5de 100644
--- a/test/mapstride.cpp
+++ b/test/mapstride.cpp
@@ -58,7 +58,7 @@ template<int Alignment,typename MatrixType> void map_class_matrix(const MatrixTy
   MatrixType m = MatrixType::Random(rows,cols);
   Scalar s1 = internal::random<Scalar>();
 
-  Index arraysize = 2*(rows+4)*(cols+4);
+  Index arraysize = 4*(rows+4)*(cols+4);
 
   Scalar* a_array1 = internal::aligned_new<Scalar>(arraysize+1);
   Scalar* array1 = a_array1;
@@ -143,9 +143,62 @@ template<int Alignment,typename MatrixType> void map_class_matrix(const MatrixTy
     VERIFY_IS_APPROX(map,s1*m);
   }
 
+  // test inner stride and no outer stride
+  for(int k=0; k<2; ++k)
+  {
+    if(k==1 && (m.innerSize()*2)*m.outerSize() > maxsize2)
+      break;
+    Scalar* array = (k==0 ? array1 : array2);
+
+    Map<MatrixType, Alignment, InnerStride<Dynamic> > map(array, rows, cols, InnerStride<Dynamic>(2));
+    map = m;
+    VERIFY(map.outerStride() == map.innerSize()*2);
+    for(int i = 0; i < m.outerSize(); ++i)
+      for(int j = 0; j < m.innerSize(); ++j)
+      {
+        VERIFY(array[map.innerSize()*i*2+j*2] == m.coeffByOuterInner(i,j));
+        VERIFY(map.coeffByOuterInner(i,j) == m.coeffByOuterInner(i,j));
+      }
+    VERIFY_IS_APPROX(s1*map,s1*m);
+    map *= s1;
+    VERIFY_IS_APPROX(map,s1*m);
+  }
+
   internal::aligned_delete(a_array1, arraysize+1);
 }
 
+// Additional tests for inner-stride but no outer-stride
+template<int>
+void bug1453()
+{
+  const int data[] = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31};
+  typedef Matrix<int,Dynamic,Dynamic,RowMajor> RowMatrixXi;
+  typedef Matrix<int,2,3,ColMajor> ColMatrix23i;
+  typedef Matrix<int,3,2,ColMajor> ColMatrix32i;
+  typedef Matrix<int,2,3,RowMajor> RowMatrix23i;
+  typedef Matrix<int,3,2,RowMajor> RowMatrix32i;
+
+  VERIFY_IS_APPROX(MatrixXi::Map(data, 2, 3, InnerStride<2>()), MatrixXi::Map(data, 2, 3, Stride<4,2>()));
+  VERIFY_IS_APPROX(MatrixXi::Map(data, 2, 3, InnerStride<>(2)), MatrixXi::Map(data, 2, 3, Stride<4,2>()));
+  VERIFY_IS_APPROX(MatrixXi::Map(data, 3, 2, InnerStride<2>()), MatrixXi::Map(data, 3, 2, Stride<6,2>()));
+  VERIFY_IS_APPROX(MatrixXi::Map(data, 3, 2, InnerStride<>(2)), MatrixXi::Map(data, 3, 2, Stride<6,2>()));
+
+  VERIFY_IS_APPROX(RowMatrixXi::Map(data, 2, 3, InnerStride<2>()), RowMatrixXi::Map(data, 2, 3, Stride<6,2>()));
+  VERIFY_IS_APPROX(RowMatrixXi::Map(data, 2, 3, InnerStride<>(2)), RowMatrixXi::Map(data, 2, 3, Stride<6,2>()));
+  VERIFY_IS_APPROX(RowMatrixXi::Map(data, 3, 2, InnerStride<2>()), RowMatrixXi::Map(data, 3, 2, Stride<4,2>()));
+  VERIFY_IS_APPROX(RowMatrixXi::Map(data, 3, 2, InnerStride<>(2)), RowMatrixXi::Map(data, 3, 2, Stride<4,2>()));
+
+  VERIFY_IS_APPROX(ColMatrix23i::Map(data, InnerStride<2>()), MatrixXi::Map(data, 2, 3, Stride<4,2>()));
+  VERIFY_IS_APPROX(ColMatrix23i::Map(data, InnerStride<>(2)), MatrixXi::Map(data, 2, 3, Stride<4,2>()));
+  VERIFY_IS_APPROX(ColMatrix32i::Map(data, InnerStride<2>()), MatrixXi::Map(data, 3, 2, Stride<6,2>()));
+  VERIFY_IS_APPROX(ColMatrix32i::Map(data, InnerStride<>(2)), MatrixXi::Map(data, 3, 2, Stride<6,2>()));
+
+  VERIFY_IS_APPROX(RowMatrix23i::Map(data, InnerStride<2>()), RowMatrixXi::Map(data, 2, 3, Stride<6,2>()));
+  VERIFY_IS_APPROX(RowMatrix23i::Map(data, InnerStride<>(2)), RowMatrixXi::Map(data, 2, 3, Stride<6,2>()));
+  VERIFY_IS_APPROX(RowMatrix32i::Map(data, InnerStride<2>()), RowMatrixXi::Map(data, 3, 2, Stride<4,2>()));
+  VERIFY_IS_APPROX(RowMatrix32i::Map(data, InnerStride<>(2)), RowMatrixXi::Map(data, 3, 2, Stride<4,2>()));
+}
+
 void test_mapstride()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -175,6 +228,8 @@ void test_mapstride()
     CALL_SUBTEST_5( map_class_matrix<Unaligned>(MatrixXi(internal::random<int>(1,maxn),internal::random<int>(1,maxn))) );
     CALL_SUBTEST_6( map_class_matrix<Aligned>(MatrixXcd(internal::random<int>(1,maxn),internal::random<int>(1,maxn))) );
     CALL_SUBTEST_6( map_class_matrix<Unaligned>(MatrixXcd(internal::random<int>(1,maxn),internal::random<int>(1,maxn))) );
+
+    CALL_SUBTEST_5( bug1453<0>() );
     
     TEST_SET_BUT_UNUSED_VARIABLE(maxn);
   }

From be281e528967ed00ed52f50a476beef10ff0dec3 Mon Sep 17 00:00:00 2001
From: Jim Radford <radford@blackbean.org>
Date: Wed, 4 Jan 2017 14:43:56 -0800
Subject: [PATCH 021/680] LLT: avoid making a copy when decomposing in place

---
 Eigen/src/Cholesky/LLT.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 7f29dad9c..c7683232d 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -428,7 +428,8 @@ LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);
-  m_matrix = a.derived();
+  if (!internal::is_same_dense(m_matrix, a.derived()))
+    m_matrix = a.derived();
 
   // Compute matrix L1 norm = max abs column sum.
   m_l1_norm = RealScalar(0);

From 0c226644d8cf21d35cfcf46c60ce66d2183f530e Mon Sep 17 00:00:00 2001
From: Jim Radford <radford@blackbean.org>
Date: Wed, 4 Jan 2017 14:42:57 -0800
Subject: [PATCH 022/680] LLT: const the arg to solveInPlace() to allow passing
 .transpose(), .block(), etc.

---
 Eigen/src/Cholesky/LLT.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index c7683232d..152837bfd 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -150,7 +150,7 @@ template<typename _MatrixType, int _UpLo> class LLT
     }
 
     template<typename Derived>
-    void solveInPlace(MatrixBase<Derived> &bAndX) const;
+    void solveInPlace(const MatrixBase<Derived> &bAndX) const;
 
     template<typename InputType>
     LLT& compute(const EigenBase<InputType>& matrix);
@@ -493,7 +493,7 @@ void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
   */
 template<typename MatrixType, int _UpLo>
 template<typename Derived>
-void LLT<MatrixType,_UpLo>::solveInPlace(MatrixBase<Derived> &bAndX) const
+void LLT<MatrixType,_UpLo>::solveInPlace(const MatrixBase<Derived> &bAndX) const
 {
   eigen_assert(m_isInitialized && "LLT is not initialized.");
   eigen_assert(m_matrix.rows()==bAndX.rows());

From b1e312edd607bcfa99192d53f55b2ac974644c44 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 15 Feb 2017 10:13:01 +0000
Subject: [PATCH 023/680] Adding TensorPatch.h for sycl backend.

---
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |  11 +-
 unsupported/test/CMakeLists.txt               |   1 +
 unsupported/test/cxx11_tensor_patch_sycl.cpp  | 252 ++++++++++++++++++
 3 files changed, 262 insertions(+), 2 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_patch_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 886a254f6..cead2eac8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -99,11 +99,11 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
  };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device)
+      : m_impl(op.expression(), device), patch_dims(op.patch_dims())
   {
     Index num_patches = 1;
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
-    const PatchDim& patch_dims = op.patch_dims();
+  //  const PatchDim& patch_dims = op.patch_dims();
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int i = 0; i < NumDims-1; ++i) {
         m_dimensions[i] = patch_dims[i];
@@ -255,6 +255,11 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PatchDim& functor() const { return patch_dims; }
+
  protected:
   Dimensions m_dimensions;
   array<Index, NumDims> m_outputStrides;
@@ -262,6 +267,8 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
   array<Index, NumDims-1> m_patchStrides;
 
   TensorEvaluator<ArgType, Device> m_impl;
+  // required by sycl
+  const PatchDim& patch_dims;
 };
 
 } // end namespace Eigen
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 003c9de0b..d01233fb2 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -167,6 +167,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_convolution_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_striding_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_patch_sycl.cpp b/unsupported/test/cxx11_tensor_patch_sycl.cpp
new file mode 100644
index 000000000..b75219a5b
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_patch_sycl.cpp
@@ -0,0 +1,252 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+// Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_patch_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+template <typename DataType, int DataLayout, typename IndexType>
+static void test_simple_patch_sycl(const Eigen::SyclDevice& sycl_device){
+
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 5;
+  IndexType sizeDim4 = 7;
+  array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  array<IndexType, 5> patchTensorRange;
+  if (DataLayout == ColMajor) {
+   patchTensorRange = {{1, 1, 1, 1, sizeDim1*sizeDim2*sizeDim3*sizeDim4}};
+  }else{
+     patchTensorRange = {{sizeDim1*sizeDim2*sizeDim3*sizeDim4,1, 1, 1, 1}};
+  }
+
+  Tensor<DataType, 4, DataLayout,IndexType> tensor(tensorRange);
+  Tensor<DataType, 5, DataLayout,IndexType> no_patch(patchTensorRange);
+
+  tensor.setRandom();
+
+  array<ptrdiff_t, 4> patch_dims;
+  patch_dims[0] = 1;
+  patch_dims[1] = 1;
+  patch_dims[2] = 1;
+  patch_dims[3] = 1;
+
+  const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
+  size_t patchTensorBuffSize =no_patch.size()*sizeof(DataType);
+  DataType* gpu_data_tensor  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+  DataType* gpu_data_no_patch  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_no_patch(gpu_data_no_patch, patchTensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_tensor, tensor.data(), tensorBuffSize);
+  gpu_no_patch.device(sycl_device)=gpu_tensor.extract_patches(patch_dims);
+  sycl_device.memcpyDeviceToHost(no_patch.data(), gpu_data_no_patch, patchTensorBuffSize);
+
+  if (DataLayout == ColMajor) {
+    VERIFY_IS_EQUAL(no_patch.dimension(0), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(1), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(2), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(3), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(4), tensor.size());
+  } else {
+    VERIFY_IS_EQUAL(no_patch.dimension(0), tensor.size());
+    VERIFY_IS_EQUAL(no_patch.dimension(1), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(2), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(3), 1);
+    VERIFY_IS_EQUAL(no_patch.dimension(4), 1);
+  }
+
+  for (int i = 0; i < tensor.size(); ++i) {
+    VERIFY_IS_EQUAL(tensor.data()[i], no_patch.data()[i]);
+  }
+  patch_dims[0] = 2;
+  patch_dims[1] = 3;
+  patch_dims[2] = 5;
+  patch_dims[3] = 7;
+  if (DataLayout == ColMajor) {
+   patchTensorRange = {{sizeDim1,sizeDim2,sizeDim3,sizeDim4,1}};
+  }else{
+     patchTensorRange = {{1,sizeDim1,sizeDim2,sizeDim3,sizeDim4}};
+  }
+  Tensor<DataType, 5, DataLayout,IndexType> single_patch(patchTensorRange);
+  patchTensorBuffSize =single_patch.size()*sizeof(DataType);
+  DataType* gpu_data_single_patch  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_single_patch(gpu_data_single_patch, patchTensorRange);
+
+  gpu_single_patch.device(sycl_device)=gpu_tensor.extract_patches(patch_dims);
+  sycl_device.memcpyDeviceToHost(single_patch.data(), gpu_data_single_patch, patchTensorBuffSize);
+
+  if (DataLayout == ColMajor) {
+    VERIFY_IS_EQUAL(single_patch.dimension(0), 2);
+    VERIFY_IS_EQUAL(single_patch.dimension(1), 3);
+    VERIFY_IS_EQUAL(single_patch.dimension(2), 5);
+    VERIFY_IS_EQUAL(single_patch.dimension(3), 7);
+    VERIFY_IS_EQUAL(single_patch.dimension(4), 1);
+  } else {
+    VERIFY_IS_EQUAL(single_patch.dimension(0), 1);
+    VERIFY_IS_EQUAL(single_patch.dimension(1), 2);
+    VERIFY_IS_EQUAL(single_patch.dimension(2), 3);
+    VERIFY_IS_EQUAL(single_patch.dimension(3), 5);
+    VERIFY_IS_EQUAL(single_patch.dimension(4), 7);
+  }
+
+  for (int i = 0; i < tensor.size(); ++i) {
+    VERIFY_IS_EQUAL(tensor.data()[i], single_patch.data()[i]);
+  }
+
+
+
+
+
+  patch_dims[0] = 1;
+  patch_dims[1] = 2;
+  patch_dims[2] = 2;
+  patch_dims[3] = 1;
+  if (DataLayout == ColMajor) {
+   patchTensorRange = {{1,2,2,1,2*2*4*7}};
+  }else{
+     patchTensorRange = {{2*2*4*7, 1, 2,2,1}};
+  }
+  Tensor<DataType, 5, DataLayout,IndexType> twod_patch(patchTensorRange);
+  patchTensorBuffSize =twod_patch.size()*sizeof(DataType);
+  DataType* gpu_data_twod_patch  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_twod_patch(gpu_data_twod_patch, patchTensorRange);
+
+  gpu_twod_patch.device(sycl_device)=gpu_tensor.extract_patches(patch_dims);
+  sycl_device.memcpyDeviceToHost(twod_patch.data(), gpu_data_twod_patch, patchTensorBuffSize);
+
+  if (DataLayout == ColMajor) {
+    VERIFY_IS_EQUAL(twod_patch.dimension(0), 1);
+    VERIFY_IS_EQUAL(twod_patch.dimension(1), 2);
+    VERIFY_IS_EQUAL(twod_patch.dimension(2), 2);
+    VERIFY_IS_EQUAL(twod_patch.dimension(3), 1);
+    VERIFY_IS_EQUAL(twod_patch.dimension(4), 2*2*4*7);
+  } else {
+    VERIFY_IS_EQUAL(twod_patch.dimension(0), 2*2*4*7);
+    VERIFY_IS_EQUAL(twod_patch.dimension(1), 1);
+    VERIFY_IS_EQUAL(twod_patch.dimension(2), 2);
+    VERIFY_IS_EQUAL(twod_patch.dimension(3), 2);
+    VERIFY_IS_EQUAL(twod_patch.dimension(4), 1);
+  }
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 2; ++j) {
+      for (int k = 0; k < 4; ++k) {
+        for (int l = 0; l < 7; ++l) {
+          int patch_loc;
+          if (DataLayout == ColMajor) {
+            patch_loc = i + 2 * (j + 2 * (k + 4 * l));
+          } else {
+            patch_loc = l + 7 * (k + 4 * (j + 2 * i));
+          }
+          for (int x = 0; x < 2; ++x) {
+            for (int y = 0; y < 2; ++y) {
+              if (DataLayout == ColMajor) {
+                VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l), twod_patch(0,x,y,0,patch_loc));
+              } else {
+                VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l), twod_patch(patch_loc,0,x,y,0));
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  patch_dims[0] = 1;
+  patch_dims[1] = 2;
+  patch_dims[2] = 3;
+  patch_dims[3] = 5;
+
+  if (DataLayout == ColMajor) {
+   patchTensorRange = {{1,2,3,5,2*2*3*3}};
+  }else{
+     patchTensorRange = {{2*2*3*3, 1, 2,3,5}};
+  }
+  Tensor<DataType, 5, DataLayout,IndexType> threed_patch(patchTensorRange);
+  patchTensorBuffSize =threed_patch.size()*sizeof(DataType);
+  DataType* gpu_data_threed_patch  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_threed_patch(gpu_data_threed_patch, patchTensorRange);
+
+  gpu_threed_patch.device(sycl_device)=gpu_tensor.extract_patches(patch_dims);
+  sycl_device.memcpyDeviceToHost(threed_patch.data(), gpu_data_threed_patch, patchTensorBuffSize);
+
+  if (DataLayout == ColMajor) {
+    VERIFY_IS_EQUAL(threed_patch.dimension(0), 1);
+    VERIFY_IS_EQUAL(threed_patch.dimension(1), 2);
+    VERIFY_IS_EQUAL(threed_patch.dimension(2), 3);
+    VERIFY_IS_EQUAL(threed_patch.dimension(3), 5);
+    VERIFY_IS_EQUAL(threed_patch.dimension(4), 2*2*3*3);
+  } else {
+    VERIFY_IS_EQUAL(threed_patch.dimension(0), 2*2*3*3);
+    VERIFY_IS_EQUAL(threed_patch.dimension(1), 1);
+    VERIFY_IS_EQUAL(threed_patch.dimension(2), 2);
+    VERIFY_IS_EQUAL(threed_patch.dimension(3), 3);
+    VERIFY_IS_EQUAL(threed_patch.dimension(4), 5);
+  }
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 2; ++j) {
+      for (int k = 0; k < 3; ++k) {
+        for (int l = 0; l < 3; ++l) {
+          int patch_loc;
+          if (DataLayout == ColMajor) {
+            patch_loc = i + 2 * (j + 2 * (k + 3 * l));
+          } else {
+            patch_loc = l + 3 * (k + 3 * (j + 2 * i));
+          }
+          for (int x = 0; x < 2; ++x) {
+            for (int y = 0; y < 3; ++y) {
+              for (int z = 0; z < 5; ++z) {
+                if (DataLayout == ColMajor) {
+                  VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l+z), threed_patch(0,x,y,z,patch_loc));
+                } else {
+                  VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l+z), threed_patch(patch_loc,0,x,y,z));
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data_tensor);
+  sycl_device.deallocate(gpu_data_no_patch);
+  sycl_device.deallocate(gpu_data_single_patch);
+  sycl_device.deallocate(gpu_data_twod_patch);
+  sycl_device.deallocate(gpu_data_threed_patch);
+}
+
+template<typename DataType, typename dev_Selector> void sycl_tensor_patch_test_per_device(dev_Selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_simple_patch_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_simple_patch_sycl<DataType, ColMajor, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_patch_sycl()
+{
+  for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(sycl_tensor_patch_test_per_device<float>(device));
+  }
+}

From 91982b91c02deb5e1ce557bbc5c96fee19c636ed Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 15 Feb 2017 16:28:12 +0000
Subject: [PATCH 024/680] Adding TensorLayoutSwapOp for sycl.

---
 .../TensorSyclConvertToDeviceExpression.h     |  31 +++--
 .../src/Tensor/TensorSyclExprConstructor.h    |  21 ++-
 .../src/Tensor/TensorSyclExtractAccessor.h    |  15 +++
 .../src/Tensor/TensorSyclExtractFunctors.h    |  27 ++--
 .../CXX11/src/Tensor/TensorSyclLeafCount.h    |  14 +-
 .../src/Tensor/TensorSyclPlaceHolderExpr.h    |  17 ++-
 unsupported/test/CMakeLists.txt               |   1 +
 .../test/cxx11_tensor_layout_swap_sycl.cpp    | 126 ++++++++++++++++++
 unsupported/test/cxx11_tensor_patch_sycl.cpp  |   9 +-
 9 files changed, 217 insertions(+), 44 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_layout_swap_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
index ee8f3c9c2..ff5097141 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -91,28 +91,35 @@ ASSIGNCONVERT(, false)
 #undef ASSIGNCONVERT
 
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node
-/// type is either TensorForcedEvalOp or TensorEvalToOp
+/// type is  TensorEvalToOp
 #define KERNELBROKERCONVERT(CVQual, Res, ExprNode)\
 template <typename Expr>\
 struct ConvertToDeviceExpression<CVQual ExprNode<Expr> > \
 : DeviceConvertor<ExprNode, Res, Expr>{};
 
-/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorForcedEvalOp
-#define KERNELBROKERCONVERTFORCEDEVAL(CVQual)\
-template <typename Expr>\
-struct ConvertToDeviceExpression<CVQual TensorForcedEvalOp<Expr> > {\
-  typedef CVQual TensorForcedEvalOp< typename ConvertToDeviceExpression<Expr>::Type> Type;\
-};
-KERNELBROKERCONVERTFORCEDEVAL(const)
-KERNELBROKERCONVERTFORCEDEVAL()
-#undef KERNELBROKERCONVERTFORCEDEVAL
-
-
 
 KERNELBROKERCONVERT(const, true, TensorEvalToOp)
 KERNELBROKERCONVERT(, false, TensorEvalToOp)
 #undef KERNELBROKERCONVERT
 
+/// specialisation of the \ref ConvertToDeviceExpression struct when the node types are TensorForcedEvalOp and TensorLayoutSwapOp
+#define KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(CVQual, ExprNode)\
+template <typename Expr>\
+struct ConvertToDeviceExpression<CVQual ExprNode<Expr> > {\
+  typedef CVQual ExprNode< typename ConvertToDeviceExpression<Expr>::Type> Type;\
+};
+
+// TensorForcedEvalOp
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(const,TensorForcedEvalOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(,TensorForcedEvalOp)
+
+// TensorLayoutSwapOp
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(const,TensorLayoutSwapOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(,TensorLayoutSwapOp)
+#undef KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP
+
+
+
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorReductionOp
 #define KERNELBROKERCONVERTREDUCTION(CVQual)\
 template <typename OP, typename Dim, typename subExpr, template <class> class MakePointer_>\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index 3b83b1d2c..6b6093fa3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -223,7 +223,7 @@ struct ExprConstructor<CVQual TensorEvalToOp<OrigExpr, MakeGlobalPointer>, CVQua
   Type expr;\
   template <typename FuncDetector>\
   ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
-  : nestedExpression(funcD.rhsExpr, t), buffer(t), expr(buffer.expr, nestedExpression.expr) {}\
+  : nestedExpression(funcD.xprExpr, t), buffer(t), expr(buffer.expr, nestedExpression.expr) {}\
 };
 
 EVALTO(const)
@@ -386,6 +386,25 @@ SYCLTENSORCHIPPINGOPEXPR()
 #undef SYCLTENSORCHIPPINGOPEXPR
 
 
+
+// TensorLayoutSwapOp
+#define SYCLTENSORLAYOUTSWAPOPEXPR(CVQual)\
+template<typename OrigXprType, typename XprType, typename... Params>\
+struct ExprConstructor<CVQual TensorLayoutSwapOp <OrigXprType> , CVQual TensorLayoutSwapOp<XprType>, Params... >{\
+  typedef ExprConstructor<OrigXprType, XprType, Params...> my_xpr_type;\
+  typedef CVQual TensorLayoutSwapOp<typename my_xpr_type::Type> Type;\
+  my_xpr_type xprExpr;\
+  Type expr;\
+  template <typename FuncDetector>\
+  ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
+  : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr) {}\
+};
+
+SYCLTENSORLAYOUTSWAPOPEXPR(const)
+SYCLTENSORLAYOUTSWAPOPEXPR()
+#undef SYCLTENSORLAYOUTSWAPOPEXPR
+
+
 /// template deduction for \ref ExprConstructor struct
 template <typename OrigExpr, typename IndexExpr, typename FuncD, typename... Params>
 auto createDeviceExpression(FuncD &funcD, const utility::tuple::Tuple<Params...> &t)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index b512d43f6..213dd25ea 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -226,6 +226,21 @@ SYCLTENSORCHIPPINGOPEXTACC()
 #undef SYCLTENSORCHIPPINGOPEXTACC
 
 
+// specialisation of the \ref ExtractAccessor struct when the node type is
+/// TensorLayoutSwapOp.
+#define SYCLTENSORLAYOUTSWAPOPEXTACC(CVQual)\
+template<typename XprType, typename Dev>\
+struct ExtractAccessor<TensorEvaluator<CVQual TensorLayoutSwapOp<XprType>, Dev> >{\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorLayoutSwapOp<XprType>, Dev>& eval)\
+  RETURN_CPP11(AccessorConstructor::getTuple(cgh, eval.impl()))\
+};
+
+SYCLTENSORLAYOUTSWAPOPEXTACC(const)
+SYCLTENSORLAYOUTSWAPOPEXTACC()
+#undef SYCLTENSORLAYOUTSWAPOPEXTACC
+
+
+
 /// template deduction for \ref ExtractAccessor
 template <typename Evaluator>
 auto createTupleOfAccessors(cl::sycl::handler& cgh, const Evaluator& eval)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index ee020184b..1506e8189 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -39,7 +39,6 @@ template <typename Evaluator> struct FunctorExtractor{
   EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
   FunctorExtractor(const Evaluator& expr)
   : m_dimensions(expr.dimensions()) {}
-
 };
 
 /// specialisation of the \ref FunctorExtractor struct when the node type does not require anything
@@ -143,19 +142,23 @@ SYCLEXTRFUNCASSIGNOP(const)
 SYCLEXTRFUNCASSIGNOP()
 #undef SYCLEXTRFUNCASSIGNOP
 
-/// specialisation of the \ref FunctorExtractor struct when the node type is
-/// TensorEvalToOp, This is an specialisation without OP so it has to be separated.
-#define SYCLEXTRFUNCEVALTOOP(CVQual)\
-template <typename RHSExpr, typename Dev>\
-struct FunctorExtractor<TensorEvaluator<CVQual TensorEvalToOp<RHSExpr>, Dev> > {\
-  FunctorExtractor<TensorEvaluator<RHSExpr, Dev> > rhsExpr;\
-  FunctorExtractor(const TensorEvaluator<CVQual TensorEvalToOp<RHSExpr>, Dev>& expr)\
-  : rhsExpr(expr.impl()) {}\
+/// specialisation of the \ref FunctorExtractor struct when the node types are
+/// TensorEvalToOp, TensorLayoutSwapOp. This is an specialisation without OP so it has to be separated.
+#define SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(CVQual, ExprNode)\
+template <typename Expr, typename Dev>\
+struct FunctorExtractor<TensorEvaluator<CVQual ExprNode<Expr>, Dev> > {\
+  FunctorExtractor<TensorEvaluator<Expr, Dev> > xprExpr;\
+  FunctorExtractor(const TensorEvaluator<CVQual ExprNode<Expr>, Dev>& expr)\
+  : xprExpr(expr.impl()) {}\
 };
+//TensorEvalToOp
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(const, TensorEvalToOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(, TensorEvalToOp)
+// TensorLayoutSwapOp
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(const, TensorLayoutSwapOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(, TensorLayoutSwapOp)
 
-SYCLEXTRFUNCEVALTOOP(const)
-SYCLEXTRFUNCEVALTOOP()
-#undef SYCLEXTRFUNCEVALTOOP
+#undef SYCLEXTRFUNCEVALTOOPSWAPLAYOUT
 
 template<typename Dim, size_t NumOutputDim> struct DimConstr {
 template<typename InDim>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
index a1c112f4d..15729310d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -94,15 +94,17 @@ SYCLFORCEDEVALLEAFCOUNT()
 #undef SYCLFORCEDEVALLEAFCOUNT
 
 /// specialisation of the \ref LeafCount struct when the node type is TensorEvalToOp
-#define EVALTOLEAFCOUNT(CVQual)\
+#define EVALTOLAYOUTSWAPLEAFCOUNT(CVQual , ExprNode, Num)\
 template <typename Expr>\
-struct LeafCount<CVQual TensorEvalToOp<Expr> > {\
-  static const size_t Count = 1 + CategoryCount<Expr>::Count;\
+struct LeafCount<CVQual ExprNode<Expr> > {\
+  static const size_t Count = Num + CategoryCount<Expr>::Count;\
 };
 
-EVALTOLEAFCOUNT(const)
-EVALTOLEAFCOUNT()
-#undef EVALTOLEAFCOUNT
+EVALTOLAYOUTSWAPLEAFCOUNT(const, TensorEvalToOp, 1)
+EVALTOLAYOUTSWAPLEAFCOUNT(, TensorEvalToOp, 1)
+EVALTOLAYOUTSWAPLEAFCOUNT(const, TensorLayoutSwapOp, 0)
+EVALTOLAYOUTSWAPLEAFCOUNT(, TensorLayoutSwapOp, 0)
+#undef EVALTOLAYOUTSWAPLEAFCOUNT
 
 /// specialisation of the \ref LeafCount struct when the node type is const TensorReductionOp
 #define REDUCTIONLEAFCOUNT(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
index 74566dcee..ba0d17e0c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -144,16 +144,19 @@ FORCEDEVAL()
 #undef FORCEDEVAL
 
 /// specialisation of the \ref PlaceHolderExpression when the node is
-/// TensorEvalToOp
-#define EVALTO(CVQual)\
+/// TensorEvalToOp, TensorLayoutSwapOp
+#define EVALTOLAYOUTSWAP(CVQual, ExprNode)\
 template <typename Expr, size_t N>\
-struct PlaceHolderExpression<CVQual TensorEvalToOp<Expr>, N> {\
-  typedef CVQual TensorEvalToOp<typename CalculateIndex <N, Expr>::ArgType> Type;\
+struct PlaceHolderExpression<CVQual ExprNode<Expr>, N> {\
+  typedef CVQual ExprNode<typename CalculateIndex <N, Expr>::ArgType> Type;\
 };
 
-EVALTO(const)
-EVALTO()
-#undef EVALTO
+EVALTOLAYOUTSWAP(const, TensorEvalToOp)
+EVALTOLAYOUTSWAP(, TensorEvalToOp)
+EVALTOLAYOUTSWAP(const, TensorLayoutSwapOp)
+EVALTOLAYOUTSWAP(, TensorLayoutSwapOp)
+
+#undef EVALTOLAYOUTSWAP
 
 
 /// specialisation of the \ref PlaceHolderExpression when the node is
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index d01233fb2..57580f805 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -168,6 +168,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_striding_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_layout_swap_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp b/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp
new file mode 100644
index 000000000..9e8db8b4b
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp
@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+// Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_layout_swap_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+template <typename DataType, typename IndexType>
+static void test_simple_swap_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 7;
+  array<IndexType, 3> tensorColRange = {{sizeDim1, sizeDim2, sizeDim3}};
+  array<IndexType, 3> tensorRowRange = {{sizeDim3, sizeDim2, sizeDim1}};
+
+
+  Tensor<DataType, 3, ColMajor, IndexType> tensor1(tensorColRange);
+  Tensor<DataType, 3, RowMajor, IndexType> tensor2(tensorRowRange);
+  tensor1.setRandom();
+
+  DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(tensor1.size()*sizeof(DataType)));
+  DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(tensor2.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 3, ColMajor, IndexType>> gpu1(gpu_data1, tensorColRange);
+  TensorMap<Tensor<DataType, 3, RowMajor, IndexType>> gpu2(gpu_data2, tensorRowRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data1, tensor1.data(),(tensor1.size())*sizeof(DataType));
+  gpu2.device(sycl_device)=gpu1.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor2.data(), gpu_data2,(tensor2.size())*sizeof(DataType));
+
+
+//  Tensor<float, 3, ColMajor> tensor(2,3,7);
+  //tensor.setRandom();
+
+//  Tensor<float, 3, RowMajor> tensor2 = tensor.swap_layout();
+  VERIFY_IS_EQUAL(tensor1.dimension(0), tensor2.dimension(2));
+  VERIFY_IS_EQUAL(tensor1.dimension(1), tensor2.dimension(1));
+  VERIFY_IS_EQUAL(tensor1.dimension(2), tensor2.dimension(0));
+
+  for (IndexType i = 0; i < 2; ++i) {
+    for (IndexType j = 0; j < 3; ++j) {
+      for (IndexType k = 0; k < 7; ++k) {
+        VERIFY_IS_EQUAL(tensor1(i,j,k), tensor2(k,j,i));
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data1);
+  sycl_device.deallocate(gpu_data2);
+}
+
+template <typename DataType, typename IndexType>
+static void test_swap_as_lvalue_sycl(const Eigen::SyclDevice& sycl_device)
+{
+
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 7;
+  array<IndexType, 3> tensorColRange = {{sizeDim1, sizeDim2, sizeDim3}};
+  array<IndexType, 3> tensorRowRange = {{sizeDim3, sizeDim2, sizeDim1}};
+
+  Tensor<DataType, 3, ColMajor, IndexType> tensor1(tensorColRange);
+  Tensor<DataType, 3, RowMajor, IndexType> tensor2(tensorRowRange);
+  tensor1.setRandom();
+
+  DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(tensor1.size()*sizeof(DataType)));
+  DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(tensor2.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 3, ColMajor, IndexType>> gpu1(gpu_data1, tensorColRange);
+  TensorMap<Tensor<DataType, 3, RowMajor, IndexType>> gpu2(gpu_data2, tensorRowRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data1, tensor1.data(),(tensor1.size())*sizeof(DataType));
+  gpu2.swap_layout().device(sycl_device)=gpu1;
+  sycl_device.memcpyDeviceToHost(tensor2.data(), gpu_data2,(tensor2.size())*sizeof(DataType));
+
+
+//  Tensor<float, 3, ColMajor> tensor(2,3,7);
+//  tensor.setRandom();
+
+  //Tensor<float, 3, RowMajor> tensor2(7,3,2);
+//  tensor2.swap_layout() = tensor;
+  VERIFY_IS_EQUAL(tensor1.dimension(0), tensor2.dimension(2));
+  VERIFY_IS_EQUAL(tensor1.dimension(1), tensor2.dimension(1));
+  VERIFY_IS_EQUAL(tensor1.dimension(2), tensor2.dimension(0));
+
+  for (IndexType i = 0; i < 2; ++i) {
+    for (IndexType j = 0; j < 3; ++j) {
+      for (IndexType k = 0; k < 7; ++k) {
+        VERIFY_IS_EQUAL(tensor1(i,j,k), tensor2(k,j,i));
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data1);
+  sycl_device.deallocate(gpu_data2);
+}
+
+
+template<typename DataType, typename dev_Selector> void sycl_tensor_layout_swap_test_per_device(dev_Selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_simple_swap_sycl<DataType, int64_t>(sycl_device);
+  test_swap_as_lvalue_sycl<DataType, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_layout_swap_sycl()
+{
+  for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(sycl_tensor_layout_swap_test_per_device<float>(device));
+  }
+}
diff --git a/unsupported/test/cxx11_tensor_patch_sycl.cpp b/unsupported/test/cxx11_tensor_patch_sycl.cpp
index b75219a5b..88a29cb31 100644
--- a/unsupported/test/cxx11_tensor_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_patch_sycl.cpp
@@ -12,7 +12,6 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
 #define EIGEN_TEST_FUNC cxx11_tensor_patch_sycl
@@ -80,10 +79,12 @@ static void test_simple_patch_sycl(const Eigen::SyclDevice& sycl_device){
   for (int i = 0; i < tensor.size(); ++i) {
     VERIFY_IS_EQUAL(tensor.data()[i], no_patch.data()[i]);
   }
+
   patch_dims[0] = 2;
   patch_dims[1] = 3;
   patch_dims[2] = 5;
   patch_dims[3] = 7;
+
   if (DataLayout == ColMajor) {
    patchTensorRange = {{sizeDim1,sizeDim2,sizeDim3,sizeDim4,1}};
   }else{
@@ -114,15 +115,11 @@ static void test_simple_patch_sycl(const Eigen::SyclDevice& sycl_device){
   for (int i = 0; i < tensor.size(); ++i) {
     VERIFY_IS_EQUAL(tensor.data()[i], single_patch.data()[i]);
   }
-
-
-
-
-
   patch_dims[0] = 1;
   patch_dims[1] = 2;
   patch_dims[2] = 2;
   patch_dims[3] = 1;
+  
   if (DataLayout == ColMajor) {
    patchTensorRange = {{1,2,2,1,2*2*4*7}};
   }else{

From 79ebc8f76137f151c78b4f61cd99fae62bf6c34f Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 20 Feb 2017 12:11:05 +0000
Subject: [PATCH 025/680] Adding Sycl backend for TensorImagePatchOP.h; adding
 Sycl backend for TensorInflation.h.

---
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |   26 +-
 .../Eigen/CXX11/src/Tensor/TensorInflation.h  |    6 +
 .../TensorSyclConvertToDeviceExpression.h     |   14 +
 .../src/Tensor/TensorSyclExprConstructor.h    |   18 +
 .../src/Tensor/TensorSyclExtractAccessor.h    |   14 +
 .../src/Tensor/TensorSyclExtractFunctors.h    |   36 +-
 .../CXX11/src/Tensor/TensorSyclLeafCount.h    |   10 +
 .../src/Tensor/TensorSyclPlaceHolderExpr.h    |   14 +
 unsupported/test/CMakeLists.txt               |    2 +
 .../test/cxx11_tensor_image_patchOP_sycl.cpp  | 1092 +++++++++++++++++
 .../test/cxx11_tensor_inflation_sycl.cpp      |  136 ++
 11 files changed, 1356 insertions(+), 12 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
 create mode 100644 unsupported/test/cxx11_tensor_inflation_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 566856ed2..2fb6b84b9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -70,12 +70,8 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            PaddingType padding_type, Scalar padding_value)
-      : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+      : TensorImagePatchOp(expr, patch_rows, patch_cols, row_strides,col_strides, in_row_strides, in_col_strides, row_inflate_strides,
+         col_inflate_strides, 0,0,0,0,padding_value, padding_type, false ){}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex row_strides, DenseIndex col_strides,
@@ -83,14 +79,15 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
-                                                           Scalar padding_value)
+                                                           Scalar padding_value, PaddingType padding_type=PADDING_VALID,
+                                                           bool padding_explicit=true)
       : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+        m_padding_explicit(padding_explicit), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
         m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+        m_padding_type(padding_type), m_padding_value(padding_value) {}
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_rows() const { return m_patch_rows; }
@@ -172,7 +169,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device)
+      : m_impl(op.expression(), device), m_op(op)
   {
     EIGEN_STATIC_ASSERT((NumDims >= 4), YOU_MADE_A_PROGRAMMING_MISTAKE);
 
@@ -241,6 +238,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
           break;
         default:
           eigen_assert(false && "unexpected padding");
+          m_outputCols=0; // silence the uninitialised warnig;
+          m_outputRows=0; //// silence the uninitialised warnig;
       }
     }
     eigen_assert(m_outputRows > 0);
@@ -420,7 +419,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
-  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  // required by sycl
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
+
 
   Index rowPaddingTop() const { return m_rowPaddingTop; }
   Index colPaddingLeft() const { return m_colPaddingLeft; }
@@ -501,6 +503,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   Scalar m_paddingValue;
 
   TensorEvaluator<ArgType, Device> m_impl;
+  // required for sycl
+  const XprType& m_op;
 };
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
index f391fb9ee..b6bf05fed 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
@@ -215,6 +215,12 @@ struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE  const Strides& functor() const { return m_strides; }
+
+
  protected:
   Dimensions m_dimensions;
   array<Index, NumDims> m_outputStrides;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
index ff5097141..5b4a9af9f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -165,6 +165,20 @@ KERNELBROKERCONVERTCHIPPINGOP()
 
 
 
+
+/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorImagePatchOp
+#define KERNELBROKERCONVERTIMAGEPATCHOP(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>\
+struct ConvertToDeviceExpression<CVQual TensorImagePatchOp<Rows, Cols, XprType> >{\
+  typedef CVQual TensorImagePatchOp<Rows, Cols, typename ConvertToDeviceExpression<XprType>::Type> Type;\
+};
+KERNELBROKERCONVERTIMAGEPATCHOP(const)
+KERNELBROKERCONVERTIMAGEPATCHOP()
+#undef KERNELBROKERCONVERTIMAGEPATCHOP
+
+
+
+
 }  // namespace internal
 }  // namespace TensorSycl
 }  // namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index 6b6093fa3..57a10d06b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -385,6 +385,24 @@ SYCLTENSORCHIPPINGOPEXPR(const)
 SYCLTENSORCHIPPINGOPEXPR()
 #undef SYCLTENSORCHIPPINGOPEXPR
 
+// TensorImagePatchOp
+#define SYCLTENSORIMAGEPATCHOPEXPR(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename OrigXprType, typename XprType, typename... Params>\
+struct  ExprConstructor<CVQual TensorImagePatchOp<Rows, Cols, OrigXprType>, CVQual TensorImagePatchOp<Rows, Cols, XprType>, Params... > {\
+  typedef ExprConstructor<OrigXprType, XprType, Params...> my_xpr_type;\
+  typedef CVQual TensorImagePatchOp<Rows, Cols, typename my_xpr_type::Type> Type;\
+  my_xpr_type xprExpr;\
+  Type expr;\
+  template <typename FuncDetector>\
+  ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
+  : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.m_patch_rows, funcD.m_patch_cols, funcD.m_row_strides, funcD.m_col_strides,\
+    funcD.m_in_row_strides, funcD.m_in_col_strides, funcD.m_row_inflate_strides, funcD.m_col_inflate_strides, \
+    funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, funcD.m_padding_value, funcD.m_padding_type, funcD.m_padding_explicit){}\
+};
+
+SYCLTENSORIMAGEPATCHOPEXPR(const)
+SYCLTENSORIMAGEPATCHOPEXPR()
+#undef SYCLTENSORIMAGEPATCHOPEXPR
 
 
 // TensorLayoutSwapOp
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index 213dd25ea..2be6f3710 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -226,6 +226,20 @@ SYCLTENSORCHIPPINGOPEXTACC()
 #undef SYCLTENSORCHIPPINGOPEXTACC
 
 
+// specialisation of the \ref ExtractAccessor struct when the node type is
+/// TensorImagePatchOp.
+#define SYCLTENSORIMAGEPATCHOPEXTACC(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename XprType, typename Dev>\
+struct ExtractAccessor<TensorEvaluator<CVQual TensorImagePatchOp<Rows, Cols, XprType>, Dev> >{\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorImagePatchOp<Rows, Cols, XprType>, Dev>& eval)\
+  RETURN_CPP11(AccessorConstructor::getTuple(cgh, eval.impl()))\
+};
+
+SYCLTENSORIMAGEPATCHOPEXTACC(const)
+SYCLTENSORIMAGEPATCHOPEXTACC()
+#undef SYCLTENSORIMAGEPATCHOPEXTACC
+
+
 // specialisation of the \ref ExtractAccessor struct when the node type is
 /// TensorLayoutSwapOp.
 #define SYCLTENSORLAYOUTSWAPOPEXTACC(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index 1506e8189..dbac01138 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -296,7 +296,7 @@ SYCLEXTRFUNCCONTRACTCONCAT(TensorConcatenationOp, axis(),)
 //TensorChippingOp
 #define SYCLEXTRFUNCCHIPPINGOP(CVQual)\
 template<DenseIndex DimId, typename XprType, typename Device>\
-struct FunctorExtractor<TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Device>>{\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Device> >{\
   FunctorExtractor<Eigen::TensorEvaluator<XprType, Device> > xprExpr;\
   const DenseIndex m_dim;\
   const DenseIndex m_offset;\
@@ -310,6 +310,40 @@ SYCLEXTRFUNCCHIPPINGOP(const)
 SYCLEXTRFUNCCHIPPINGOP()
 #undef SYCLEXTRFUNCCHIPPINGOP
 
+#define SYCLEXTRFUNCIMAGEPATCHOP(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename XprType, typename Device>\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorImagePatchOp<Rows, Cols, XprType>, Device> >{\
+typedef CVQual TensorImagePatchOp<Rows, Cols, XprType> Self;\
+FunctorExtractor<Eigen::TensorEvaluator<XprType, Device> > xprExpr;\
+const DenseIndex m_patch_rows;\
+const DenseIndex m_patch_cols;\
+const DenseIndex m_row_strides;\
+const DenseIndex m_col_strides;\
+const DenseIndex m_in_row_strides;\
+const DenseIndex m_in_col_strides;\
+const DenseIndex m_row_inflate_strides;\
+const DenseIndex m_col_inflate_strides;\
+const bool m_padding_explicit;\
+const DenseIndex m_padding_top;\
+const DenseIndex m_padding_bottom;\
+const DenseIndex m_padding_left;\
+const DenseIndex m_padding_right;\
+const PaddingType m_padding_type;\
+const typename Self::Scalar m_padding_value;\
+FunctorExtractor(const TensorEvaluator<Self, Device>& expr)\
+: xprExpr(expr.impl()), m_patch_rows(expr.xpr().patch_rows()), m_patch_cols(expr.xpr().patch_cols()),\
+  m_row_strides(expr.xpr().row_strides()), m_col_strides(expr.xpr().col_strides()),\
+  m_in_row_strides(expr.xpr().in_row_strides()), m_in_col_strides(expr.xpr().in_col_strides()),\
+  m_row_inflate_strides(expr.xpr().row_inflate_strides()), m_col_inflate_strides(expr.xpr().col_inflate_strides()),\
+  m_padding_explicit(expr.xpr().padding_explicit()),m_padding_top(expr.xpr().padding_top()),\
+  m_padding_bottom(expr.xpr().padding_bottom()), m_padding_left(expr.xpr().padding_left()),\
+  m_padding_right(expr.xpr().padding_right()), m_padding_type(expr.xpr().padding_type()),\
+  m_padding_value(expr.xpr().padding_value()){}\
+};
+
+SYCLEXTRFUNCIMAGEPATCHOP(const)
+SYCLEXTRFUNCIMAGEPATCHOP()
+#undef SYCLEXTRFUNCIMAGEPATCHOP
 /// template deduction function for FunctorExtractor
 template <typename Evaluator>
 auto inline extractFunctors(const Evaluator& evaluator)-> FunctorExtractor<Evaluator> {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
index 15729310d..b8e658824 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -161,6 +161,16 @@ SLICESTRIDEOPLEAFCOUNT()
 #undef SLICESTRIDEOPLEAFCOUNT
 
 
+#define TENSORIMAGEPATCHOPLEAFCOUNT(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>\
+struct LeafCount<CVQual TensorImagePatchOp<Rows, Cols, XprType> >:CategoryCount<XprType>{};
+
+
+TENSORIMAGEPATCHOPLEAFCOUNT(const)
+TENSORIMAGEPATCHOPLEAFCOUNT()
+#undef TENSORIMAGEPATCHOPLEAFCOUNT
+
+
 } /// namespace TensorSycl
 } /// namespace internal
 } /// namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
index ba0d17e0c..ab97235ae 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -221,6 +221,20 @@ SYCLSLICESTRIDEOPPLH()
 #undef SYCLSLICESTRIDEOPPLH
 
 
+
+/// specialisation of the \ref PlaceHolderExpression when the node is
+/// TensorImagePatchOp
+#define SYCLTENSORIMAGEPATCHOP(CVQual)\
+template<DenseIndex Rows, DenseIndex Cols, typename XprType, size_t N>\
+struct PlaceHolderExpression<CVQual TensorImagePatchOp<Rows, Cols, XprType>, N> {\
+  typedef CVQual TensorImagePatchOp<Rows, Cols, typename CalculateIndex <N, XprType>::ArgType> Type;\
+};
+
+SYCLTENSORIMAGEPATCHOP(const)
+SYCLTENSORIMAGEPATCHOP()
+#undef SYCLTENSORIMAGEPATCHOP
+
+
 /// template deduction for \ref PlaceHolderExpression struct
 template <typename Expr>
 struct createPlaceHolderExpression {
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 57580f805..282f9eb55 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -169,6 +169,8 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_layout_swap_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_inflation_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp b/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
new file mode 100644
index 000000000..ba6b2f15a
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
@@ -0,0 +1,1092 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_image_patchOP_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+static const int DataLayout = ColMajor;
+
+template <typename DataType, typename IndexType>
+static void test_simple_image_patch_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 5;
+  IndexType sizeDim4 = 7;
+  array<IndexType, 4> tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  array<IndexType, 4> tensorRowMajorRange = {{sizeDim4, sizeDim3, sizeDim2, sizeDim1}};
+  Tensor<DataType, 4, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  Tensor<DataType, 4, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+  tensor_col_major.setRandom();
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 4, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_col_major.size())*sizeof(DataType));
+
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(3), tensor_row_major.dimension(0));
+
+  // Single pixel patch: ColMajor
+  array<IndexType, 5> patchColMajorTensorRange={{sizeDim1, 1, 1, sizeDim2*sizeDim3, sizeDim4}};
+  Tensor<DataType, 5, DataLayout,IndexType> single_patch_col_major(patchColMajorTensorRange);
+  size_t patchTensorBuffSize =single_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_single_patch_col_major(gpu_data_single_patch_col_major, patchColMajorTensorRange);
+  gpu_single_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(1, 1);
+  sycl_device.memcpyDeviceToHost(single_patch_col_major.data(), gpu_data_single_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(0), 2);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(4), 7);
+
+  // Single pixel patch: RowMajor
+  array<IndexType, 5> patchRowMajorTensorRange={{sizeDim4, sizeDim2*sizeDim3, 1, 1, sizeDim1}};
+  Tensor<DataType, 5, RowMajor,IndexType> single_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =single_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_single_patch_row_major(gpu_data_single_patch_row_major, patchRowMajorTensorRange);
+  gpu_single_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(1, 1);
+  sycl_device.memcpyDeviceToHost(single_patch_row_major.data(), gpu_data_single_patch_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(0), 7);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(1), 3*5);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(3), 1);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(4), 2);
+
+  for (IndexType i = 0; i < tensor_col_major.size(); ++i) {
+    // ColMajor
+    if (tensor_col_major.data()[i] != single_patch_col_major.data()[i]) {
+      std::cout << "Mismatch detected at index colmajor " << i << " : "
+           << tensor_col_major.data()[i] << " vs " << single_patch_col_major.data()[i]
+           << std::endl;
+    }
+    VERIFY_IS_EQUAL(single_patch_col_major.data()[i], tensor_col_major.data()[i]);
+    // RowMajor
+    if (tensor_row_major.data()[i] != single_patch_row_major.data()[i]) {
+      std::cout << "Mismatch detected at index row major" << i << " : "
+           << tensor_row_major.data()[i] << " vs "
+           << single_patch_row_major.data()[i] << std::endl;
+    }
+    VERIFY_IS_EQUAL(single_patch_row_major.data()[i],
+                    tensor_row_major.data()[i]);
+    VERIFY_IS_EQUAL(tensor_col_major.data()[i], tensor_row_major.data()[i]);
+    VERIFY_IS_EQUAL(single_patch_col_major.data()[i],
+                    single_patch_row_major.data()[i]);
+  }
+
+
+  // Entire image patch: ColMajor
+  patchColMajorTensorRange={{sizeDim1, sizeDim2, sizeDim3, sizeDim2*sizeDim3, sizeDim4}};
+  Tensor<DataType, 5, DataLayout,IndexType> entire_image_patch_col_major(patchColMajorTensorRange);
+  patchTensorBuffSize =entire_image_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_entire_image_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_entire_image_patch_col_major(gpu_data_entire_image_patch_col_major, patchColMajorTensorRange);
+  gpu_entire_image_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(3, 5);
+  sycl_device.memcpyDeviceToHost(entire_image_patch_col_major.data(), gpu_data_entire_image_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(0), 2);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(1), 3);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(2), 5);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(4), 7);
+
+  // Entire image patch: RowMajor
+  patchRowMajorTensorRange={{sizeDim4, sizeDim2*sizeDim3, sizeDim3, sizeDim2, sizeDim1}};
+  Tensor<DataType, 5, RowMajor,IndexType> entire_image_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =entire_image_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_entire_image_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_entire_image_patch_row_major(gpu_data_entire_image_patch_row_major, patchRowMajorTensorRange);
+  gpu_entire_image_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(3, 5);
+  sycl_device.memcpyDeviceToHost(entire_image_patch_row_major.data(), gpu_data_entire_image_patch_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(0), 7);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(1), 3*5);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(2), 5);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(3), 3);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(4), 2);
+
+  for (IndexType i = 0; i < 3; ++i) {
+    for (IndexType j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (IndexType r = 0; r < 3; ++r) {
+        for (IndexType c = 0; c < 5; ++c) {
+          for (IndexType d = 0; d < 2; ++d) {
+            for (IndexType b = 0; b < 7; ++b) {
+              DataType expected_col_major = 0.0f;
+              DataType expected_row_major = 0.0f;
+              if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) {
+                expected_col_major = tensor_col_major(d, r-1+i, c-2+j, b);
+                expected_row_major = tensor_row_major(b, c-2+j, r-1+i, d);
+              }
+              // ColMajor
+              if (entire_image_patch_col_major(d, r, c, patchId, b) != expected_col_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(entire_image_patch_col_major(d, r, c, patchId, b), expected_col_major);
+              // RowMajor
+              if (entire_image_patch_row_major(b, patchId, c, r, d) !=
+                  expected_row_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j
+                     << " r=" << r << " c=" << c << " d=" << d << " b=" << b
+                     << std::endl;
+              }
+              VERIFY_IS_EQUAL(entire_image_patch_row_major(b, patchId, c, r, d),
+                              expected_row_major);
+              // Check that ColMajor and RowMajor agree.
+              VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // 2D patch: ColMajor
+  patchColMajorTensorRange={{sizeDim1, 2, 2, sizeDim2*sizeDim3, sizeDim4}};
+  Tensor<DataType, 5, DataLayout,IndexType> twod_patch_col_major(patchColMajorTensorRange);
+  patchTensorBuffSize =twod_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_twod_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_twod_patch_col_major(gpu_data_twod_patch_col_major, patchColMajorTensorRange);
+  gpu_twod_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(2, 2);
+  sycl_device.memcpyDeviceToHost(twod_patch_col_major.data(), gpu_data_twod_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(0), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(4), 7);
+
+  // 2D patch: RowMajor
+  patchRowMajorTensorRange={{sizeDim4, sizeDim2*sizeDim3, 2, 2, sizeDim1}};
+  Tensor<DataType, 5, RowMajor,IndexType> twod_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =twod_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_twod_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_twod_patch_row_major(gpu_data_twod_patch_row_major, patchRowMajorTensorRange);
+  gpu_twod_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(2, 2);
+  sycl_device.memcpyDeviceToHost(twod_patch_row_major.data(), gpu_data_twod_patch_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(0), 7);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(1), 3*5);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(3), 2);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(4), 2);
+
+
+  // Based on the calculation described in TensorTraits.h, padding happens to be 0.
+  IndexType row_padding = 0;
+  IndexType col_padding = 0;
+  IndexType stride = 1;
+
+  for (IndexType i = 0; i < 3; ++i) {
+    for (IndexType j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (IndexType r = 0; r < 2; ++r) {
+        for (IndexType c = 0; c < 2; ++c) {
+          for (IndexType d = 0; d < 2; ++d) {
+            for (IndexType b = 0; b < 7; ++b) {
+              DataType expected_col_major = 0.0f;
+              DataType expected_row_major = 0.0f;
+              IndexType row_offset = r*stride + i - row_padding;
+              IndexType col_offset = c*stride + j - col_padding;
+              // ColMajor
+              if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor_col_major.dimension(1) && col_offset < tensor_col_major.dimension(2)) {
+                expected_col_major = tensor_col_major(d, row_offset, col_offset, b);
+              }
+              if (twod_patch_col_major(d, r, c, patchId, b) != expected_col_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(twod_patch_col_major(d, r, c, patchId, b), expected_col_major);
+
+              // RowMajor
+              if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor_row_major.dimension(2) && col_offset < tensor_row_major.dimension(1)) {
+                expected_row_major = tensor_row_major(b, col_offset, row_offset, d);
+
+              }
+              if (twod_patch_row_major(b, patchId, c, r, d) != expected_row_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(twod_patch_row_major(b, patchId, c, r, d), expected_row_major);
+              // Check that ColMajor and RowMajor agree.
+              VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  sycl_device.deallocate(gpu_data_col_major);
+  sycl_device.deallocate(gpu_data_row_major);
+  sycl_device.deallocate(gpu_data_single_patch_col_major);
+  sycl_device.deallocate(gpu_data_single_patch_row_major);
+  sycl_device.deallocate(gpu_data_entire_image_patch_col_major);
+  sycl_device.deallocate(gpu_data_entire_image_patch_row_major);
+  sycl_device.deallocate(gpu_data_twod_patch_col_major);
+  sycl_device.deallocate(gpu_data_twod_patch_row_major);
+
+}
+
+
+// Verifies VALID padding (no padding) with incrementing values.
+template <typename DataType, typename IndexType>
+static void test_patch_padding_valid_sycl(const Eigen::SyclDevice& sycl_device){
+  IndexType input_depth = 3;
+  IndexType input_rows = 3;
+  IndexType input_cols = 3;
+  IndexType input_batches = 1;
+  IndexType ksize = 2;  // Corresponds to the Rows and Cols for tensor.extract_image_patches<>.
+  IndexType stride = 2;  // Only same stride is supported.
+
+  array<IndexType, 4> tensorColMajorRange = {{input_depth, input_rows, input_cols, input_batches}};
+  array<IndexType, 4> tensorRowMajorRange = {{input_batches, input_cols, input_rows, input_depth}};
+  Tensor<DataType, 4, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  Tensor<DataType, 4, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 4, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_col_major.size())*sizeof(DataType));
+
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(3), tensor_row_major.dimension(0));
+
+  // Initializes tensor with incrementing numbers.
+  for (IndexType i = 0; i < tensor_col_major.size(); ++i) {
+    tensor_col_major.data()[i] = i + 1;
+  }
+  // ColMajor
+  array<IndexType, 5> patchColMajorTensorRange={{input_depth, ksize, ksize, 1, input_batches}};
+  Tensor<DataType, 5, DataLayout,IndexType> result_col_major(patchColMajorTensorRange);
+  size_t patchTensorBuffSize =result_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_result_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_result_col_major(gpu_data_result_col_major, patchColMajorTensorRange);
+  gpu_result_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(ksize, ksize, stride, stride, 1, 1, PADDING_VALID);
+  sycl_device.memcpyDeviceToHost(result_col_major.data(), gpu_data_result_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), input_depth);  // depth
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), ksize);  // kernel rows
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), ksize);  // kernel cols
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), 1);  // number of patches
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), input_batches);  // number of batches
+
+  // RowMajor
+  array<IndexType, 5> patchRowMajorTensorRange={{input_batches, 1, ksize, ksize, input_depth }};
+  Tensor<DataType, 5, RowMajor,IndexType> result_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =result_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_result_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_result_row_major(gpu_data_result_row_major, patchRowMajorTensorRange);
+  gpu_result_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(ksize, ksize, stride, stride, 1, 1, PADDING_VALID);
+  sycl_device.memcpyDeviceToHost(result_row_major.data(), gpu_data_result_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), result_row_major.dimension(4));
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), result_row_major.dimension(3));
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), result_row_major.dimension(2));
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), result_row_major.dimension(1));
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), result_row_major.dimension(0));
+
+  // No padding is carried out.
+  IndexType row_padding = 0;
+  IndexType col_padding = 0;
+
+  for (IndexType i = 0; (i+stride+ksize-1) < input_rows; i += stride) {  // input rows
+    for (IndexType j = 0; (j+stride+ksize-1) < input_cols; j += stride) {  // input cols
+      int patchId = i+input_rows*j;
+      for (IndexType r = 0; r < ksize; ++r) {  // patch rows
+        for (IndexType c = 0; c < ksize; ++c) {  // patch cols
+          for (IndexType d = 0; d < input_depth; ++d) {  // depth
+            for (IndexType b = 0; b < input_batches; ++b) {  // batch
+              DataType expected_col_major = 0.0f;
+              DataType expected_row_major = 0.0f;
+              IndexType row_offset = r + i - row_padding;
+              IndexType col_offset = c + j - col_padding;
+              if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) {
+                expected_col_major = tensor_col_major(d, row_offset, col_offset, b);
+                expected_row_major = tensor_row_major(b, col_offset, row_offset, d);
+              }
+              // ColMajor
+              if (result_col_major(d, r, c, patchId, b) != expected_col_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_col_major(d, r, c, patchId, b), expected_col_major);
+              // RowMajor
+              if (result_row_major(b, patchId, c, r, d) != expected_row_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_row_major(b, patchId, c, r, d), expected_row_major);
+              // Check that ColMajor and RowMajor agree.
+              VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+            }
+          }
+        }
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data_col_major);
+  sycl_device.deallocate(gpu_data_row_major);
+  sycl_device.deallocate(gpu_data_result_col_major);
+  sycl_device.deallocate(gpu_data_result_row_major);
+}
+
+// Verifies VALID padding (no padding) with the same value.
+template <typename DataType, typename IndexType>
+static void test_patch_padding_valid_same_value_sycl(const Eigen::SyclDevice& sycl_device){
+  IndexType input_depth = 1;
+  IndexType input_rows = 5;
+  IndexType input_cols = 5;
+  IndexType input_batches = 2;
+  IndexType ksize = 3;  // Corresponds to the Rows and Cols for tensor.extract_image_patches<>.
+  IndexType stride = 2;  // Only same stride is supported.
+  // ColMajor
+
+  array<IndexType, 4> tensorColMajorRange = {{input_depth, input_rows, input_cols, input_batches}};
+  array<IndexType, 4> tensorRowMajorRange = {{input_batches, input_cols, input_rows, input_depth}};
+  Tensor<DataType, 4, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  Tensor<DataType, 4, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 4, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+  gpu_col_major.device(sycl_device)=gpu_col_major.constant(11.0f);
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor_col_major.data(), gpu_data_col_major, (tensor_col_major.size())*sizeof(DataType));
+  sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_row_major.size())*sizeof(DataType));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(3), tensor_row_major.dimension(0));
+
+  array<IndexType, 5> patchColMajorTensorRange={{input_depth, ksize, ksize, 4, input_batches}};
+  Tensor<DataType, 5, DataLayout,IndexType> result_col_major(patchColMajorTensorRange);
+  size_t patchTensorBuffSize =result_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_result_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_result_col_major(gpu_data_result_col_major, patchColMajorTensorRange);
+  gpu_result_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(ksize, ksize, stride, stride, 1, 1, PADDING_VALID);
+  sycl_device.memcpyDeviceToHost(result_col_major.data(), gpu_data_result_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), input_depth);  // depth
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), ksize);  // kernel rows
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), ksize);  // kernel cols
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), 4);  // number of patches
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), input_batches);  // number of batches
+
+  // RowMajor
+  array<IndexType, 5> patchRowMajorTensorRange={{input_batches, 4, ksize, ksize, input_depth }};
+  Tensor<DataType, 5, RowMajor,IndexType> result_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =result_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_result_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_result_row_major(gpu_data_result_row_major, patchRowMajorTensorRange);
+  gpu_result_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(ksize, ksize, stride, stride, 1, 1, PADDING_VALID);
+  sycl_device.memcpyDeviceToHost(result_row_major.data(), gpu_data_result_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), result_row_major.dimension(4));
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), result_row_major.dimension(3));
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), result_row_major.dimension(2));
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), result_row_major.dimension(1));
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), result_row_major.dimension(0));
+
+  // No padding is carried out.
+  IndexType row_padding = 0;
+  IndexType col_padding = 0;
+
+  for (IndexType i = 0; (i+stride+ksize-1) <= input_rows; i += stride) {  // input rows
+    for (IndexType j = 0; (j+stride+ksize-1) <= input_cols; j += stride) {  // input cols
+      IndexType patchId = i+input_rows*j;
+      for (IndexType r = 0; r < ksize; ++r) {  // patch rows
+        for (IndexType c = 0; c < ksize; ++c) {  // patch cols
+          for (IndexType d = 0; d < input_depth; ++d) {  // depth
+            for (IndexType b = 0; b < input_batches; ++b) {  // batch
+              DataType expected_col_major = 0.0f;
+              DataType expected_row_major = 0.0f;
+              IndexType row_offset = r + i - row_padding;
+              IndexType col_offset = c + j - col_padding;
+              if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) {
+                expected_col_major = tensor_col_major(d, row_offset, col_offset, b);
+                expected_row_major = tensor_row_major(b, col_offset, row_offset, d);
+              }
+              // ColMajor
+              if (result_col_major(d, r, c, patchId, b) != expected_col_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_col_major(d, r, c, patchId, b), expected_col_major);
+              // RowMajor
+              if (result_row_major(b, patchId, c, r, d) != expected_row_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_row_major(b, patchId, c, r, d), expected_row_major);
+              // Check that ColMajor and RowMajor agree.
+              VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+// Verifies SAME padding.
+template <typename DataType, typename IndexType>
+static void test_patch_padding_same_sycl(const Eigen::SyclDevice& sycl_device){
+  IndexType input_depth = 3;
+  IndexType input_rows = 4;
+  IndexType input_cols = 2;
+  IndexType input_batches = 1;
+  IndexType ksize = 2;  // Corresponds to the Rows and Cols for tensor.extract_image_patches<>.
+  IndexType stride = 2;  // Only same stride is supported.
+
+  // ColMajor
+  array<IndexType, 4> tensorColMajorRange = {{input_depth, input_rows, input_cols, input_batches}};
+  array<IndexType, 4> tensorRowMajorRange = {{input_batches, input_cols, input_rows, input_depth}};
+  Tensor<DataType, 4, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  Tensor<DataType, 4, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 4, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_col_major.size())*sizeof(DataType));
+
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(3), tensor_row_major.dimension(0));
+
+  // Initializes tensor with incrementing numbers.
+  for (IndexType i = 0; i < tensor_col_major.size(); ++i) {
+    tensor_col_major.data()[i] = i + 1;
+  }
+
+array<IndexType, 5> patchColMajorTensorRange={{input_depth, ksize, ksize, 2, input_batches}};
+Tensor<DataType, 5, DataLayout,IndexType> result_col_major(patchColMajorTensorRange);
+size_t patchTensorBuffSize =result_col_major.size()*sizeof(DataType);
+DataType* gpu_data_result_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_result_col_major(gpu_data_result_col_major, patchColMajorTensorRange);
+gpu_result_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(ksize, ksize, stride, stride, PADDING_SAME);
+sycl_device.memcpyDeviceToHost(result_col_major.data(), gpu_data_result_col_major, patchTensorBuffSize);
+
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), input_depth);  // depth
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), ksize);  // kernel rows
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), ksize);  // kernel cols
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), 2);  // number of patches
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), input_batches);  // number of batches
+
+  // RowMajor
+
+  array<IndexType, 5> patchRowMajorTensorRange={{input_batches, 2, ksize, ksize, input_depth }};
+  Tensor<DataType, 5, RowMajor,IndexType> result_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =result_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_result_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_result_row_major(gpu_data_result_row_major, patchRowMajorTensorRange);
+  gpu_result_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(ksize, ksize, stride, stride, PADDING_SAME);
+  sycl_device.memcpyDeviceToHost(result_row_major.data(), gpu_data_result_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(result_col_major.dimension(0), result_row_major.dimension(4));
+  VERIFY_IS_EQUAL(result_col_major.dimension(1), result_row_major.dimension(3));
+  VERIFY_IS_EQUAL(result_col_major.dimension(2), result_row_major.dimension(2));
+  VERIFY_IS_EQUAL(result_col_major.dimension(3), result_row_major.dimension(1));
+  VERIFY_IS_EQUAL(result_col_major.dimension(4), result_row_major.dimension(0));
+
+  // Based on the calculation described in TensorTraits.h, padding happens to be 0.
+  IndexType row_padding = 0;
+  IndexType col_padding = 0;
+
+  for (IndexType i = 0; (i+stride+ksize-1) <= input_rows; i += stride) {  // input rows
+    for (IndexType j = 0; (j+stride+ksize-1) <= input_cols; j += stride) {  // input cols
+      int patchId = i+input_rows*j;
+      for (IndexType r = 0; r < ksize; ++r) {  // patch rows
+        for (IndexType c = 0; c < ksize; ++c) {  // patch cols
+          for (IndexType d = 0; d < input_depth; ++d) {  // depth
+            for (IndexType b = 0; b < input_batches; ++b) {  // batch
+              DataType expected_col_major = 0.0f;
+              DataType expected_row_major = 0.0f;
+              IndexType row_offset = r*stride + i - row_padding;
+              IndexType col_offset = c*stride + j - col_padding;
+              if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) {
+                expected_col_major = tensor_col_major(d, row_offset, col_offset, b);
+                expected_row_major = tensor_row_major(b, col_offset, row_offset, d);
+              }
+              // ColMajor
+              if (result_col_major(d, r, c, patchId, b) != expected_col_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_col_major(d, r, c, patchId, b), expected_col_major);
+              // RowMajor
+              if (result_row_major(b, patchId, c, r, d) != expected_row_major) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(result_row_major(b, patchId, c, r, d), expected_row_major);
+              // Check that ColMajor and RowMajor agree.
+              VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+
+template <typename DataType, typename IndexType>
+static void test_patch_no_extra_dim_sycl(const Eigen::SyclDevice& sycl_device){
+
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 5;
+
+  // ColMajor
+  array<IndexType, 3> tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3}};
+  array<IndexType, 3> tensorRowMajorRange = {{sizeDim3, sizeDim2, sizeDim1}};
+  Tensor<DataType, 3, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  tensor_col_major.setRandom();
+  Tensor<DataType, 3, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 3, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 3, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+  sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_row_major.size())*sizeof(DataType));
+
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(0), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(1), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor_col_major.dimension(2), tensor_row_major.dimension(0));
+
+
+  // Single pixel patch: ColMajor
+  array<IndexType, 4> patchColMajorTensorRange={{sizeDim1, 1, 1, sizeDim2*sizeDim3}};
+  Tensor<DataType, 4, DataLayout,IndexType> single_patch_col_major(patchColMajorTensorRange);
+  size_t patchTensorBuffSize =single_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_single_patch_col_major(gpu_data_single_patch_col_major, patchColMajorTensorRange);
+  gpu_single_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(1, 1);
+  sycl_device.memcpyDeviceToHost(single_patch_col_major.data(), gpu_data_single_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(0), sizeDim1);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_patch_col_major.dimension(3), sizeDim2*sizeDim3);
+
+  // Single pixel patch: RowMajor
+  array<IndexType, 4> patchRowMajorTensorRange={{sizeDim2*sizeDim3, 1, 1, sizeDim1}};
+  Tensor<DataType, 4, RowMajor,IndexType> single_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =single_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, RowMajor,IndexType>> gpu_single_patch_row_major(gpu_data_single_patch_row_major, patchRowMajorTensorRange);
+  gpu_single_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(1, 1);
+  sycl_device.memcpyDeviceToHost(single_patch_row_major.data(), gpu_data_single_patch_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(0), sizeDim2*sizeDim3);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_patch_row_major.dimension(3), sizeDim1);
+
+  for (IndexType i = 0; i < tensor_col_major.size(); ++i) {
+    // ColMajor
+    if (tensor_col_major.data()[i] != single_patch_col_major.data()[i]) {
+      std::cout << "Mismatch detected at index " << i << " : " << tensor_col_major.data()[i] << " vs " << single_patch_col_major.data()[i] << std::endl;
+    }
+    VERIFY_IS_EQUAL(single_patch_col_major.data()[i], tensor_col_major.data()[i]);
+    // RowMajor
+    if (tensor_row_major.data()[i] != single_patch_row_major.data()[i]) {
+      std::cout << "Mismatch detected at index " << i << " : "
+           << tensor_col_major.data()[i] << " vs "
+           << single_patch_row_major.data()[i] << std::endl;
+    }
+    VERIFY_IS_EQUAL(single_patch_row_major.data()[i],
+                    tensor_row_major.data()[i]);
+    VERIFY_IS_EQUAL(tensor_col_major.data()[i], tensor_row_major.data()[i]);
+    VERIFY_IS_EQUAL(single_patch_col_major.data()[i],
+                    single_patch_row_major.data()[i]);
+  }
+
+  // Entire image patch: ColMajor
+  patchColMajorTensorRange={{sizeDim1, sizeDim2, sizeDim3, sizeDim2*sizeDim3}};
+  Tensor<DataType, 4, DataLayout,IndexType> entire_image_patch_col_major(patchColMajorTensorRange);
+  patchTensorBuffSize =entire_image_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_entire_image_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_entire_image_patch_col_major(gpu_data_entire_image_patch_col_major, patchColMajorTensorRange);
+  gpu_entire_image_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(3, 5);
+  sycl_device.memcpyDeviceToHost(entire_image_patch_col_major.data(), gpu_data_entire_image_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(0), 2);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(1), 3);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(2), 5);
+  VERIFY_IS_EQUAL(entire_image_patch_col_major.dimension(3), 3*5);
+
+  // Entire image patch: RowMajor
+patchRowMajorTensorRange={{sizeDim2*sizeDim3, sizeDim3, sizeDim2, sizeDim1}};
+Tensor<DataType, 4, RowMajor,IndexType> entire_image_patch_row_major(patchRowMajorTensorRange);
+patchTensorBuffSize =entire_image_patch_row_major.size()*sizeof(DataType);
+DataType* gpu_data_entire_image_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+TensorMap<Tensor<DataType, 4, RowMajor,IndexType>> gpu_entire_image_patch_row_major(gpu_data_entire_image_patch_row_major, patchRowMajorTensorRange);
+gpu_entire_image_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(3, 5);
+sycl_device.memcpyDeviceToHost(entire_image_patch_row_major.data(), gpu_data_entire_image_patch_row_major, patchTensorBuffSize);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(0), 3*5);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(1), 5);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(2), 3);
+  VERIFY_IS_EQUAL(entire_image_patch_row_major.dimension(3), 2);
+
+  for (IndexType i = 0; i < 3; ++i) {
+    for (IndexType j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (IndexType r = 0; r < 3; ++r) {
+        for (IndexType c = 0; c < 5; ++c) {
+          for (IndexType d = 0; d < 2; ++d) {
+            DataType expected_col_major = 0.0f;
+            DataType expected_row_major = 0.0f;
+            if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) {
+              expected_col_major = tensor_col_major(d, r-1+i, c-2+j);
+              expected_row_major = tensor_row_major(c-2+j, r-1+i, d);
+            }
+            // ColMajor
+            if (entire_image_patch_col_major(d, r, c, patchId) != expected_col_major) {
+              std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl;
+            }
+            VERIFY_IS_EQUAL(entire_image_patch_col_major(d, r, c, patchId), expected_col_major);
+            // RowMajor
+            if (entire_image_patch_row_major(patchId, c, r, d) !=
+                expected_row_major) {
+              std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl;
+            }
+            VERIFY_IS_EQUAL(entire_image_patch_row_major(patchId, c, r, d),
+                            expected_row_major);
+            // Check that ColMajor and RowMajor agree.
+            VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+          }
+        }
+      }
+    }
+  }
+
+  // 2D patch: ColMajor
+  patchColMajorTensorRange={{sizeDim1, 2, 2, sizeDim2*sizeDim3}};
+  Tensor<DataType, 4, DataLayout,IndexType> twod_patch_col_major(patchColMajorTensorRange);
+  patchTensorBuffSize =twod_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_twod_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_twod_patch_col_major(gpu_data_twod_patch_col_major, patchColMajorTensorRange);
+  gpu_twod_patch_col_major.device(sycl_device)=gpu_col_major.extract_image_patches(2, 2);
+  sycl_device.memcpyDeviceToHost(twod_patch_col_major.data(), gpu_data_twod_patch_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(0), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch_col_major.dimension(3), 3*5);
+
+  // 2D patch: RowMajor
+  patchRowMajorTensorRange={{sizeDim2*sizeDim3, 2, 2, sizeDim1}};
+  Tensor<DataType, 4, RowMajor,IndexType> twod_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =twod_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_twod_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, RowMajor,IndexType>> gpu_twod_patch_row_major(gpu_data_twod_patch_row_major, patchRowMajorTensorRange);
+  gpu_twod_patch_row_major.device(sycl_device)=gpu_row_major.extract_image_patches(2, 2);
+  sycl_device.memcpyDeviceToHost(twod_patch_row_major.data(), gpu_data_twod_patch_row_major, patchTensorBuffSize);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(0), 3*5);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch_row_major.dimension(3), 2);
+
+  // Based on the calculation described in TensorTraits.h, padding happens to be 0.
+  IndexType row_padding = 0;
+  IndexType col_padding = 0;
+  IndexType stride = 1;
+
+  for (IndexType i = 0; i < 3; ++i) {
+    for (IndexType j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (IndexType r = 0; r < 2; ++r) {
+        for (IndexType c = 0; c < 2; ++c) {
+          for (IndexType d = 0; d < 2; ++d) {
+            DataType expected_col_major = 0.0f;
+            DataType expected_row_major = 0.0f;
+            IndexType row_offset = r*stride + i - row_padding;
+            IndexType col_offset = c*stride + j - col_padding;
+            // ColMajor
+            if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor_col_major.dimension(1) && col_offset < tensor_col_major.dimension(2)) {
+              expected_col_major = tensor_col_major(d, row_offset, col_offset);
+            }
+            if (twod_patch_col_major(d, r, c, patchId) != expected_col_major) {
+              std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl;
+            }
+            VERIFY_IS_EQUAL(twod_patch_col_major(d, r, c, patchId), expected_col_major);
+            // RowMajor
+            if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor_row_major.dimension(1) && col_offset < tensor_row_major.dimension(0)) {
+              expected_row_major = tensor_row_major(col_offset, row_offset, d);
+            }
+            if (twod_patch_row_major(patchId, c, r, d) != expected_row_major) {
+              std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl;
+            }
+            VERIFY_IS_EQUAL(twod_patch_row_major(patchId, c, r, d), expected_row_major);
+            // Check that ColMajor and RowMajor agree.
+            VERIFY_IS_EQUAL(expected_col_major, expected_row_major);
+          }
+        }
+      }
+    }
+  }
+
+  sycl_device.deallocate(gpu_data_col_major);
+  sycl_device.deallocate(gpu_data_row_major);
+  sycl_device.deallocate(gpu_data_single_patch_col_major);
+  sycl_device.deallocate(gpu_data_single_patch_row_major);
+  sycl_device.deallocate(gpu_data_entire_image_patch_col_major);
+  sycl_device.deallocate(gpu_data_entire_image_patch_row_major);
+  sycl_device.deallocate(gpu_data_twod_patch_col_major);
+  sycl_device.deallocate(gpu_data_twod_patch_row_major);
+}
+
+template <typename DataType, typename IndexType>
+static void test_imagenet_patches_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  // Test the code on typical configurations used by the 'imagenet' benchmarks at
+  // https://github.com/soumith/convnet-benchmarks
+  // ColMajor
+  IndexType sizeDim1 = 3;
+  IndexType sizeDim2 = 128;
+  IndexType sizeDim3 = 128;
+  IndexType sizeDim4 = 16;
+  array<IndexType, 4> tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  Tensor<DataType, 4, DataLayout,IndexType> l_in_col_major(tensorColMajorRange);
+  l_in_col_major.setRandom();
+
+  DataType* gpu_data_l_in_col_major  = static_cast<DataType*>(sycl_device.allocate(l_in_col_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>> gpu_l_in_col_major(gpu_data_l_in_col_major, tensorColMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_l_in_col_major, l_in_col_major.data(),(l_in_col_major.size())*sizeof(DataType));
+
+  array<IndexType, 5> patchTensorRange={{sizeDim1, 11, 11, sizeDim2*sizeDim3, sizeDim4}};
+  Tensor<DataType, 5, DataLayout,IndexType> l_out_col_major(patchTensorRange);
+  size_t patchTensorBuffSize =l_out_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_l_out_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_l_out_col_major(gpu_data_l_out_col_major, patchTensorRange);
+  gpu_l_out_col_major.device(sycl_device)=gpu_l_in_col_major.extract_image_patches(11, 11);
+  sycl_device.memcpyDeviceToHost(l_out_col_major.data(), gpu_data_l_out_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(0), sizeDim1);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(1), 11);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(2), 11);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(3), sizeDim2*sizeDim3);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(4), sizeDim4);
+
+  // RowMajor
+  patchTensorRange={{sizeDim4, sizeDim2*sizeDim3, 11, 11, sizeDim1}};
+  Tensor<DataType, 5, RowMajor,IndexType> l_out_row_major(patchTensorRange);
+  patchTensorBuffSize =l_out_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_l_out_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>> gpu_l_out_row_major(gpu_data_l_out_row_major, patchTensorRange);
+  gpu_l_out_row_major.device(sycl_device)=gpu_l_in_col_major.swap_layout().extract_image_patches(11, 11);
+  sycl_device.memcpyDeviceToHost(l_out_row_major.data(), gpu_data_l_out_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(0), sizeDim4);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(1), sizeDim2*sizeDim3);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(2), 11);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(3), 11);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(4), sizeDim1);
+
+  for (IndexType b = 0; b < 16; ++b) {
+    for (IndexType i = 0; i < 128; ++i) {
+      for (IndexType j = 0; j < 128; ++j) {
+        int patchId = i+128*j;
+        for (IndexType c = 0; c < 11; ++c) {
+          for (IndexType r = 0; r < 11; ++r) {
+            for (IndexType d = 0; d < 3; ++d) {
+              DataType expected = 0.0f;
+              if (r-5+i >= 0 && c-5+j >= 0 && r-5+i < 128 && c-5+j < 128) {
+                expected = l_in_col_major(d, r-5+i, c-5+j, b);
+              }
+              // ColMajor
+              if (l_out_col_major(d, r, c, patchId, b) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_col_major(d, r, c, patchId, b), expected);
+              // RowMajor
+              if (l_out_row_major(b, patchId, c, r, d) !=
+                  expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j
+                     << " r=" << r << " c=" << c << " d=" << d << " b=" << b
+                     << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_row_major(b, patchId, c, r, d),
+                              expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // ColMajor
+  sycl_device.deallocate(gpu_data_l_in_col_major);
+  sycl_device.deallocate(gpu_data_l_out_col_major);
+  sizeDim1 = 16;
+  sizeDim2 = 64;
+  sizeDim3 = 64;
+  sizeDim4 = 32;
+  tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  l_in_col_major.resize(tensorColMajorRange);
+  l_in_col_major.setRandom();
+  gpu_data_l_in_col_major  = static_cast<DataType*>(sycl_device.allocate(l_in_col_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>>gpu_l_in_col_major_resize1(gpu_data_l_in_col_major, tensorColMajorRange);
+
+  patchTensorRange={{sizeDim1, 9, 9, sizeDim2*sizeDim3, sizeDim4}};
+  l_out_col_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_col_major.size()*sizeof(DataType);
+  gpu_data_l_out_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>>gpu_l_out_col_major_resize1(gpu_data_l_out_col_major, patchTensorRange);
+  sycl_device.memcpyHostToDevice(gpu_data_l_in_col_major, l_in_col_major.data(),(l_in_col_major.size())*sizeof(DataType));
+  gpu_l_out_col_major_resize1.device(sycl_device)=gpu_l_in_col_major_resize1.extract_image_patches(9, 9);
+  sycl_device.memcpyDeviceToHost(l_out_col_major.data(), gpu_data_l_out_col_major, patchTensorBuffSize);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(0), 16);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(1), 9);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(2), 9);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(3), 64*64);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(4), 32);
+
+// RowMajor
+  sycl_device.deallocate(gpu_data_l_out_row_major);
+  patchTensorRange={{sizeDim4, sizeDim2*sizeDim3, 9, 9 ,sizeDim1}};
+  l_out_row_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_row_major.size()*sizeof(DataType);
+  gpu_data_l_out_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>>gpu_l_out_row_major_resize1(gpu_data_l_out_row_major, patchTensorRange);
+  gpu_l_out_row_major_resize1.device(sycl_device)=gpu_l_in_col_major_resize1.swap_layout().extract_image_patches(9, 9);
+  sycl_device.memcpyDeviceToHost(l_out_row_major.data(), gpu_data_l_out_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(0), 32);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(1), 64*64);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(2), 9);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(3), 9);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(4), 16);
+
+  for (IndexType b = 0; b < 32; ++b) {
+    for (IndexType i = 0; i < 64; ++i) {
+      for (IndexType j = 0; j < 64; ++j) {
+        int patchId = i+64*j;
+        for (IndexType c = 0; c < 9; ++c) {
+          for (IndexType r = 0; r < 9; ++r) {
+            for (IndexType d = 0; d < 16; ++d) {
+              DataType expected = 0.0f;
+              if (r-4+i >= 0 && c-4+j >= 0 && r-4+i < 64 && c-4+j < 64) {
+                expected = l_in_col_major(d, r-4+i, c-4+j, b);
+              }
+              // ColMajor
+              if (l_out_col_major(d, r, c, patchId, b) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_col_major(d, r, c, patchId, b), expected);
+              // RowMajor
+              if (l_out_row_major(b, patchId, c, r, d) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_row_major(b, patchId, c, r, d), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // ColMajor
+
+  sycl_device.deallocate(gpu_data_l_in_col_major);
+  sycl_device.deallocate(gpu_data_l_out_col_major);
+  sizeDim1 = 32;
+  sizeDim2 = 16;
+  sizeDim3 = 16;
+  sizeDim4 = 32;
+  tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  l_in_col_major.resize(tensorColMajorRange);
+  l_in_col_major.setRandom();
+  gpu_data_l_in_col_major  = static_cast<DataType*>(sycl_device.allocate(l_in_col_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>>gpu_l_in_col_major_resize2(gpu_data_l_in_col_major, tensorColMajorRange);
+
+  patchTensorRange={{sizeDim1, 7, 7, sizeDim2*sizeDim3, sizeDim4}};
+  l_out_col_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_col_major.size()*sizeof(DataType);
+  gpu_data_l_out_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>>gpu_l_out_col_major_resize2(gpu_data_l_out_col_major, patchTensorRange);
+  sycl_device.memcpyHostToDevice(gpu_data_l_in_col_major, l_in_col_major.data(),(l_in_col_major.size())*sizeof(DataType));
+  gpu_l_out_col_major_resize2.device(sycl_device)=gpu_l_in_col_major_resize2.extract_image_patches(7, 7);
+  sycl_device.memcpyDeviceToHost(l_out_col_major.data(), gpu_data_l_out_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(0), 32);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(1), 7);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(2), 7);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(3), 16*16);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(4), 32);
+
+  // RowMajor
+  sycl_device.deallocate(gpu_data_l_out_row_major);
+  patchTensorRange={{sizeDim4, sizeDim2*sizeDim3, 7, 7 ,sizeDim1}};
+  l_out_row_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_row_major.size()*sizeof(DataType);
+  gpu_data_l_out_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>>gpu_l_out_row_major_resize2(gpu_data_l_out_row_major, patchTensorRange);
+  gpu_l_out_row_major_resize2.device(sycl_device)=gpu_l_in_col_major_resize2.swap_layout().extract_image_patches(7, 7);
+  sycl_device.memcpyDeviceToHost(l_out_row_major.data(), gpu_data_l_out_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(0), 32);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(1), 16*16);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(2), 7);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(3), 7);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(4), 32);
+
+  for (IndexType b = 0; b < 32; ++b) {
+    for (IndexType i = 0; i < 16; ++i) {
+      for (IndexType j = 0; j < 16; ++j) {
+        int patchId = i+16*j;
+        for (IndexType c = 0; c < 7; ++c) {
+          for (IndexType r = 0; r < 7; ++r) {
+            for (IndexType d = 0; d < 32; ++d) {
+              DataType expected = 0.0f;
+              if (r-3+i >= 0 && c-3+j >= 0 && r-3+i < 16 && c-3+j < 16) {
+                expected = l_in_col_major(d, r-3+i, c-3+j, b);
+              }
+              // ColMajor
+              if (l_out_col_major(d, r, c, patchId, b) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_col_major(d, r, c, patchId, b), expected);
+              // RowMajor
+              if (l_out_row_major(b, patchId, c, r, d) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_row_major(b, patchId, c, r, d), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // ColMajor
+  sycl_device.deallocate(gpu_data_l_in_col_major);
+  sycl_device.deallocate(gpu_data_l_out_col_major);
+  sizeDim1 = 64;
+  sizeDim2 = 13;
+  sizeDim3 = 13;
+  sizeDim4 = 32;
+  tensorColMajorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  l_in_col_major.resize(tensorColMajorRange);
+  l_in_col_major.setRandom();
+  gpu_data_l_in_col_major  = static_cast<DataType*>(sycl_device.allocate(l_in_col_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 4, ColMajor, IndexType>>gpu_l_in_col_major_resize3(gpu_data_l_in_col_major, tensorColMajorRange);
+
+  patchTensorRange={{sizeDim1, 3, 3, sizeDim2*sizeDim3, sizeDim4}};
+  l_out_col_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_col_major.size()*sizeof(DataType);
+  gpu_data_l_out_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, DataLayout,IndexType>>gpu_l_out_col_major_resize3(gpu_data_l_out_col_major, patchTensorRange);
+  sycl_device.memcpyHostToDevice(gpu_data_l_in_col_major, l_in_col_major.data(),(l_in_col_major.size())*sizeof(DataType));
+  gpu_l_out_col_major_resize3.device(sycl_device)=gpu_l_in_col_major_resize3.extract_image_patches(3, 3);
+  sycl_device.memcpyDeviceToHost(l_out_col_major.data(), gpu_data_l_out_col_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(0), 64);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(1), 3);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(2), 3);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(3), 13*13);
+  VERIFY_IS_EQUAL(l_out_col_major.dimension(4), 32);
+
+  // RowMajor
+  sycl_device.deallocate(gpu_data_l_out_row_major);
+  patchTensorRange={{sizeDim4, sizeDim2*sizeDim3, 3, 3 ,sizeDim1}};
+  l_out_row_major.resize(patchTensorRange);
+  patchTensorBuffSize =l_out_row_major.size()*sizeof(DataType);
+  gpu_data_l_out_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 5, RowMajor,IndexType>>gpu_l_out_row_major_resize3(gpu_data_l_out_row_major, patchTensorRange);
+  gpu_l_out_row_major_resize3.device(sycl_device)=gpu_l_in_col_major_resize3.swap_layout().extract_image_patches(3, 3);
+  sycl_device.memcpyDeviceToHost(l_out_row_major.data(), gpu_data_l_out_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(0), 32);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(1), 13*13);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(2), 3);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(3), 3);
+  VERIFY_IS_EQUAL(l_out_row_major.dimension(4), 64);
+
+  for (IndexType b = 0; b < 32; ++b) {
+    for (IndexType i = 0; i < 13; ++i) {
+      for (IndexType j = 0; j < 13; ++j) {
+        int patchId = i+13*j;
+        for (IndexType c = 0; c < 3; ++c) {
+          for (IndexType r = 0; r < 3; ++r) {
+            for (IndexType d = 0; d < 64; ++d) {
+              DataType expected = 0.0f;
+              if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 13 && c-1+j < 13) {
+                expected = l_in_col_major(d, r-1+i, c-1+j, b);
+              }
+              // ColMajor
+              if (l_out_col_major(d, r, c, patchId, b) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_col_major(d, r, c, patchId, b), expected);
+              // RowMajor
+              if (l_out_row_major(b, patchId, c, r, d) != expected) {
+                std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl;
+              }
+              VERIFY_IS_EQUAL(l_out_row_major(b, patchId, c, r, d), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data_l_in_col_major);
+  sycl_device.deallocate(gpu_data_l_out_col_major);
+  sycl_device.deallocate(gpu_data_l_out_row_major);
+}
+
+
+template<typename DataType, typename dev_Selector> void sycl_tensor_image_patch_test_per_device(dev_Selector s){
+QueueInterface queueInterface(s);
+auto sycl_device = Eigen::SyclDevice(&queueInterface);
+test_simple_image_patch_sycl<DataType, int64_t>(sycl_device);
+test_patch_padding_valid_sycl<DataType, int64_t>(sycl_device);
+test_patch_padding_valid_same_value_sycl<DataType, int64_t>(sycl_device);
+test_patch_padding_same_sycl<DataType, int64_t>(sycl_device);
+test_patch_no_extra_dim_sycl<DataType, int64_t>(sycl_device);
+test_imagenet_patches_sycl<DataType, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_image_patchOP_sycl()
+{
+for (const auto& device :Eigen::get_sycl_supported_devices()) {
+  CALL_SUBTEST(sycl_tensor_image_patch_test_per_device<float>(device));
+}
+}
diff --git a/unsupported/test/cxx11_tensor_inflation_sycl.cpp b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
new file mode 100644
index 000000000..f2f87f7ed
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
@@ -0,0 +1,136 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_inflation_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+// Inflation Defenition for each dimention the inflated val would be
+//((dim-1)*strid[dim] +1)
+
+// for 1 dimnention vector of size 3 with value (4,4,4) with the inflated stride value of 3 would be changed to
+// tensor of size (2*3) +1 = 7 with the value of
+// (4, 0, 0, 4, 0, 0, 4).
+
+template <typename DataType, int DataLayout, typename IndexType>
+void test_simple_inflation_sycl(const Eigen::SyclDevice &sycl_device) {
+
+
+  IndexType sizeDim1 = 2;
+  IndexType sizeDim2 = 3;
+  IndexType sizeDim3 = 5;
+  IndexType sizeDim4 = 7;
+  array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+  Tensor<DataType, 4, DataLayout,IndexType> tensor(tensorRange);
+  Tensor<DataType, 4, DataLayout,IndexType> no_stride(tensorRange);
+  tensor.setRandom();
+
+  array<IndexType, 4> strides;
+  strides[0] = 1;
+  strides[1] = 1;
+  strides[2] = 1;
+  strides[3] = 1;
+
+
+  const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
+  DataType* gpu_data_tensor  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+  DataType* gpu_data_no_stride  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
+  TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_no_stride(gpu_data_no_stride, tensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_tensor, tensor.data(), tensorBuffSize);
+  gpu_no_stride.device(sycl_device)=gpu_tensor.inflate(strides);
+  sycl_device.memcpyDeviceToHost(no_stride.data(), gpu_data_no_stride, tensorBuffSize);
+
+  VERIFY_IS_EQUAL(no_stride.dimension(0), sizeDim1);
+  VERIFY_IS_EQUAL(no_stride.dimension(1), sizeDim2);
+  VERIFY_IS_EQUAL(no_stride.dimension(2), sizeDim3);
+  VERIFY_IS_EQUAL(no_stride.dimension(3), sizeDim4);
+
+  for (IndexType i = 0; i < 2; ++i) {
+    for (IndexType j = 0; j < 3; ++j) {
+      for (IndexType k = 0; k < 5; ++k) {
+        for (IndexType l = 0; l < 7; ++l) {
+          VERIFY_IS_EQUAL(tensor(i,j,k,l), no_stride(i,j,k,l));
+        }
+      }
+    }
+  }
+
+
+  strides[0] = 2;
+  strides[1] = 4;
+  strides[2] = 2;
+  strides[3] = 3;
+
+  IndexType inflatedSizeDim1 = 3;
+  IndexType inflatedSizeDim2 = 9;
+  IndexType inflatedSizeDim3 = 9;
+  IndexType inflatedSizeDim4 = 19;
+  array<IndexType, 4> inflatedTensorRange = {{inflatedSizeDim1, inflatedSizeDim2, inflatedSizeDim3, inflatedSizeDim4}};
+
+  Tensor<DataType, 4, DataLayout, IndexType> inflated(inflatedTensorRange);
+
+  const size_t inflatedTensorBuffSize =inflated.size()*sizeof(DataType);
+  DataType* gpu_data_inflated  = static_cast<DataType*>(sycl_device.allocate(inflatedTensorBuffSize));
+  TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_inflated(gpu_data_inflated, inflatedTensorRange);
+  gpu_inflated.device(sycl_device)=gpu_tensor.inflate(strides);
+  sycl_device.memcpyDeviceToHost(inflated.data(), gpu_data_inflated, inflatedTensorBuffSize);
+
+  VERIFY_IS_EQUAL(inflated.dimension(0), inflatedSizeDim1);
+  VERIFY_IS_EQUAL(inflated.dimension(1), inflatedSizeDim2);
+  VERIFY_IS_EQUAL(inflated.dimension(2), inflatedSizeDim3);
+  VERIFY_IS_EQUAL(inflated.dimension(3), inflatedSizeDim4);
+
+  for (IndexType i = 0; i < inflatedSizeDim1; ++i) {
+    for (IndexType j = 0; j < inflatedSizeDim2; ++j) {
+      for (IndexType k = 0; k < inflatedSizeDim3; ++k) {
+        for (IndexType l = 0; l < inflatedSizeDim4; ++l) {
+          if (i % strides[0] == 0 &&
+              j % strides[1] == 0 &&
+              k % strides[2] == 0 &&
+              l % strides[3] == 0) {
+            VERIFY_IS_EQUAL(inflated(i,j,k,l),
+                            tensor(i/strides[0], j/strides[1], k/strides[2], l/strides[3]));
+          } else {
+            VERIFY_IS_EQUAL(0, inflated(i,j,k,l));
+          }
+        }
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data_tensor);
+  sycl_device.deallocate(gpu_data_no_stride);
+  sycl_device.deallocate(gpu_data_inflated);
+}
+
+template<typename DataType, typename dev_Selector> void sycl_inflation_test_per_device(dev_Selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_simple_inflation_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_simple_inflation_sycl<DataType, ColMajor, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_inflation_sycl()
+{
+  for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(sycl_inflation_test_per_device<float>(device));
+  }
+}

From 4f07ac16b0722597c55e2783cee33606a1f5e390 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 21 Feb 2017 10:09:47 +0000
Subject: [PATCH 026/680] Reducing the number of warnings.

---
 .../test/cxx11_tensor_image_patchOP_sycl.cpp  | 20 +++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp b/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
index ba6b2f15a..e5ca4e388 100644
--- a/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
@@ -134,7 +134,7 @@ static void test_simple_image_patch_sycl(const Eigen::SyclDevice& sycl_device)
 
   for (IndexType i = 0; i < 3; ++i) {
     for (IndexType j = 0; j < 5; ++j) {
-      int patchId = i+3*j;
+      IndexType patchId = i+3*j;
       for (IndexType r = 0; r < 3; ++r) {
         for (IndexType c = 0; c < 5; ++c) {
           for (IndexType d = 0; d < 2; ++d) {
@@ -206,7 +206,7 @@ static void test_simple_image_patch_sycl(const Eigen::SyclDevice& sycl_device)
 
   for (IndexType i = 0; i < 3; ++i) {
     for (IndexType j = 0; j < 5; ++j) {
-      int patchId = i+3*j;
+      IndexType patchId = i+3*j;
       for (IndexType r = 0; r < 2; ++r) {
         for (IndexType c = 0; c < 2; ++c) {
           for (IndexType d = 0; d < 2; ++d) {
@@ -323,7 +323,7 @@ static void test_patch_padding_valid_sycl(const Eigen::SyclDevice& sycl_device){
 
   for (IndexType i = 0; (i+stride+ksize-1) < input_rows; i += stride) {  // input rows
     for (IndexType j = 0; (j+stride+ksize-1) < input_cols; j += stride) {  // input cols
-      int patchId = i+input_rows*j;
+      IndexType patchId = i+input_rows*j;
       for (IndexType r = 0; r < ksize; ++r) {  // patch rows
         for (IndexType c = 0; c < ksize; ++c) {  // patch cols
           for (IndexType d = 0; d < input_depth; ++d) {  // depth
@@ -529,7 +529,7 @@ sycl_device.memcpyDeviceToHost(result_col_major.data(), gpu_data_result_col_majo
 
   for (IndexType i = 0; (i+stride+ksize-1) <= input_rows; i += stride) {  // input rows
     for (IndexType j = 0; (j+stride+ksize-1) <= input_cols; j += stride) {  // input cols
-      int patchId = i+input_rows*j;
+      IndexType patchId = i+input_rows*j;
       for (IndexType r = 0; r < ksize; ++r) {  // patch rows
         for (IndexType c = 0; c < ksize; ++c) {  // patch cols
           for (IndexType d = 0; d < input_depth; ++d) {  // depth
@@ -667,7 +667,7 @@ sycl_device.memcpyDeviceToHost(entire_image_patch_row_major.data(), gpu_data_ent
 
   for (IndexType i = 0; i < 3; ++i) {
     for (IndexType j = 0; j < 5; ++j) {
-      int patchId = i+3*j;
+      IndexType patchId = i+3*j;
       for (IndexType r = 0; r < 3; ++r) {
         for (IndexType c = 0; c < 5; ++c) {
           for (IndexType d = 0; d < 2; ++d) {
@@ -731,7 +731,7 @@ sycl_device.memcpyDeviceToHost(entire_image_patch_row_major.data(), gpu_data_ent
 
   for (IndexType i = 0; i < 3; ++i) {
     for (IndexType j = 0; j < 5; ++j) {
-      int patchId = i+3*j;
+      IndexType patchId = i+3*j;
       for (IndexType r = 0; r < 2; ++r) {
         for (IndexType c = 0; c < 2; ++c) {
           for (IndexType d = 0; d < 2; ++d) {
@@ -824,7 +824,7 @@ static void test_imagenet_patches_sycl(const Eigen::SyclDevice& sycl_device)
   for (IndexType b = 0; b < 16; ++b) {
     for (IndexType i = 0; i < 128; ++i) {
       for (IndexType j = 0; j < 128; ++j) {
-        int patchId = i+128*j;
+        IndexType patchId = i+128*j;
         for (IndexType c = 0; c < 11; ++c) {
           for (IndexType r = 0; r < 11; ++r) {
             for (IndexType d = 0; d < 3; ++d) {
@@ -899,7 +899,7 @@ static void test_imagenet_patches_sycl(const Eigen::SyclDevice& sycl_device)
   for (IndexType b = 0; b < 32; ++b) {
     for (IndexType i = 0; i < 64; ++i) {
       for (IndexType j = 0; j < 64; ++j) {
-        int patchId = i+64*j;
+        IndexType patchId = i+64*j;
         for (IndexType c = 0; c < 9; ++c) {
           for (IndexType r = 0; r < 9; ++r) {
             for (IndexType d = 0; d < 16; ++d) {
@@ -972,7 +972,7 @@ static void test_imagenet_patches_sycl(const Eigen::SyclDevice& sycl_device)
   for (IndexType b = 0; b < 32; ++b) {
     for (IndexType i = 0; i < 16; ++i) {
       for (IndexType j = 0; j < 16; ++j) {
-        int patchId = i+16*j;
+        IndexType patchId = i+16*j;
         for (IndexType c = 0; c < 7; ++c) {
           for (IndexType r = 0; r < 7; ++r) {
             for (IndexType d = 0; d < 32; ++d) {
@@ -1044,7 +1044,7 @@ static void test_imagenet_patches_sycl(const Eigen::SyclDevice& sycl_device)
   for (IndexType b = 0; b < 32; ++b) {
     for (IndexType i = 0; i < 13; ++i) {
       for (IndexType j = 0; j < 13; ++j) {
-        int patchId = i+13*j;
+        IndexType patchId = i+13*j;
         for (IndexType c = 0; c < 3; ++c) {
           for (IndexType r = 0; r < 3; ++r) {
             for (IndexType d = 0; d < 64; ++d) {

From 89dfd51fae868393b66b1949638e03de2ba17c1f Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 22 Feb 2017 16:36:24 +0000
Subject: [PATCH 027/680] Adding Sycl Backend for TensorGenerator.h.

---
 .../Eigen/CXX11/src/Tensor/TensorGenerator.h  |  13 +-
 unsupported/test/CMakeLists.txt               |   1 +
 .../test/cxx11_tensor_generator_sycl.cpp      | 147 ++++++++++++++++++
 3 files changed, 158 insertions(+), 3 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_generator_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index eb1d4934e..ca87f493a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -97,10 +97,9 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_generator(op.generator())
+      : m_generator(op.generator()), m_argImpl(op.expression(), device)
   {
-    TensorEvaluator<ArgType, Device> impl(op.expression(), device);
-    m_dimensions = impl.dimensions();
+    m_dimensions = m_argImpl.dimensions();
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       m_strides[0] = 1;
@@ -155,6 +154,12 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_argImpl; }
+  /// required by sycl in order to extract the accessor
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Generator& functor() const { return m_generator; }
+
+
  protected:
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void extract_coordinates(Index index, array<Index, NumDims>& coords) const {
@@ -178,6 +183,8 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   Dimensions m_dimensions;
   array<Index, NumDims> m_strides;
   Generator m_generator;
+  // required by sycl
+  TensorEvaluator<ArgType, Device> m_argImpl;
 };
 
 } // end namespace Eigen
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 282f9eb55..69c892362 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -171,6 +171,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_layout_swap_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_inflation_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_generator_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_generator_sycl.cpp b/unsupported/test/cxx11_tensor_generator_sycl.cpp
new file mode 100644
index 000000000..f551c8d0c
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_generator_sycl.cpp
@@ -0,0 +1,147 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_generator_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+static const float error_threshold =1e-8f;
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+struct Generator1D {
+  Generator1D() { }
+
+  float operator()(const array<Eigen::DenseIndex, 1>& coordinates) const {
+    return coordinates[0];
+  }
+};
+
+template <typename DataType, int DataLayout, typename IndexType>
+static void test_1D_sycl(const Eigen::SyclDevice& sycl_device)
+{
+
+  IndexType sizeDim1 = 6;
+  array<IndexType, 1> tensorRange = {{sizeDim1}};
+  Tensor<DataType, 1, DataLayout,IndexType> vec(tensorRange);
+  Tensor<DataType, 1, DataLayout,IndexType> result(tensorRange);
+
+  const size_t tensorBuffSize =vec.size()*sizeof(DataType);
+  DataType* gpu_data_vec  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+  DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+
+  TensorMap<Tensor<DataType, 1, DataLayout,IndexType>> gpu_vec(gpu_data_vec, tensorRange);
+  TensorMap<Tensor<DataType, 1, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_vec, vec.data(), tensorBuffSize);
+  gpu_result.device(sycl_device)=gpu_vec.generate(Generator1D());
+  sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
+
+  for (IndexType i = 0; i < 6; ++i) {
+    VERIFY_IS_EQUAL(result(i), i);
+  }
+}
+
+
+struct Generator2D {
+  Generator2D() { }
+
+  float operator()(const array<Eigen::DenseIndex, 2>& coordinates) const {
+    return 3 * coordinates[0] + 11 * coordinates[1];
+  }
+};
+
+template <typename DataType, int DataLayout, typename IndexType>
+static void test_2D_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  IndexType sizeDim1 = 5;
+  IndexType sizeDim2 = 7;
+  array<IndexType, 2> tensorRange = {{sizeDim1, sizeDim2}};
+  Tensor<DataType, 2, DataLayout,IndexType> matrix(tensorRange);
+  Tensor<DataType, 2, DataLayout,IndexType> result(tensorRange);
+
+  const size_t tensorBuffSize =matrix.size()*sizeof(DataType);
+  DataType* gpu_data_matrix  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+  DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+
+  TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_matrix(gpu_data_matrix, tensorRange);
+  TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_matrix, matrix.data(), tensorBuffSize);
+  gpu_result.device(sycl_device)=gpu_matrix.generate(Generator2D());
+  sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
+
+  for (IndexType i = 0; i < 5; ++i) {
+    for (IndexType j = 0; j < 5; ++j) {
+      VERIFY_IS_EQUAL(result(i, j), 3*i + 11*j);
+    }
+  }
+}
+
+template <typename DataType, int DataLayout, typename IndexType>
+static void test_gaussian_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  IndexType rows = 32;
+  IndexType cols = 48;
+  array<DataType, 2> means;
+  means[0] = rows / 2.0f;
+  means[1] = cols / 2.0f;
+  array<DataType, 2> std_devs;
+  std_devs[0] = 3.14f;
+  std_devs[1] = 2.7f;
+  internal::GaussianGenerator<DataType, Eigen::DenseIndex, 2> gaussian_gen(means, std_devs);
+
+  array<IndexType, 2> tensorRange = {{rows, cols}};
+  Tensor<DataType, 2, DataLayout,IndexType> matrix(tensorRange);
+  Tensor<DataType, 2, DataLayout,IndexType> result(tensorRange);
+
+  const size_t tensorBuffSize =matrix.size()*sizeof(DataType);
+  DataType* gpu_data_matrix  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+  DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
+
+  TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_matrix(gpu_data_matrix, tensorRange);
+  TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_matrix, matrix.data(), tensorBuffSize);
+  gpu_result.device(sycl_device)=gpu_matrix.generate(gaussian_gen);
+  sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
+
+  for (IndexType i = 0; i < rows; ++i) {
+    for (IndexType j = 0; j < cols; ++j) {
+      DataType g_rows = powf(rows/2.0f - i, 2) / (3.14f * 3.14f) * 0.5f;
+      DataType g_cols = powf(cols/2.0f - j, 2) / (2.7f * 2.7f) * 0.5f;
+      DataType gaussian = expf(-g_rows - g_cols);
+      Eigen::internal::isApprox(result(i, j), gaussian, error_threshold);
+    }
+  }
+}
+
+template<typename DataType, typename dev_Selector> void sycl_generator_test_per_device(dev_Selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_1D_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_1D_sycl<DataType, ColMajor, int64_t>(sycl_device);
+  test_2D_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_2D_sycl<DataType, ColMajor, int64_t>(sycl_device);
+  test_gaussian_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_gaussian_sycl<DataType, ColMajor, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_generator_sycl()
+{
+  for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(sycl_generator_test_per_device<float>(device));
+  }
+}

From 0b7875f1376a0f3f22754837712ddd885ca3f4dd Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Fri, 24 Feb 2017 18:13:30 +0000
Subject: [PATCH 028/680] Converting fixed float type into template type for
 TensorContraction.

---
 .../Eigen/CXX11/src/Tensor/TensorContractionSycl.h        | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
index e87de0c57..abc7ba551 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
@@ -230,13 +230,13 @@ typename HostExpr::Index LocalThreadSizeM, typename HostExpr::Index LocalThreadS
       const Index nGroupId = itemID.get_group(1); // Work-group ID localCol
       const Index linearLocalThreadId = nLocalThreadId*LocalThreadSizeM + mLocalThreadId; // linear local thread ID
       // Allocate register space
-      float privateLhs;
-      float privateRhs[WorkLoadPerThreadN];
-      float privateRes[WorkLoadPerThreadM][WorkLoadPerThreadN];
+      LhsScalar privateLhs;
+      RhsScalar privateRhs[WorkLoadPerThreadN];
+      OutScalar privateRes[WorkLoadPerThreadM][WorkLoadPerThreadN];
       // Initialise the privateResumulation registers
       for (Index wLPTM=0; wLPTM<WorkLoadPerThreadM; wLPTM++) {
           for (Index wLPTN=0; wLPTN<WorkLoadPerThreadN; wLPTN++) {
-              privateRes[wLPTM][wLPTN] = 0.0f;
+              privateRes[wLPTM][wLPTN] = static_cast<OutScalar>(0);
           }
       }
 

From 2fa2b617a97ba254343c7c1635a9b6d617a100e8 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Fri, 24 Feb 2017 19:16:24 +0000
Subject: [PATCH 029/680] Adding TensorVolumePatchOP.h for sycl

---
 .../TensorSyclConvertToDeviceExpression.h     |  10 +
 .../src/Tensor/TensorSyclExprConstructor.h    |  22 ++
 .../src/Tensor/TensorSyclExtractAccessor.h    |  15 ++
 .../src/Tensor/TensorSyclExtractFunctors.h    |  44 ++++
 .../CXX11/src/Tensor/TensorSyclLeafCount.h    |  13 +-
 .../src/Tensor/TensorSyclPlaceHolderExpr.h    |  14 ++
 .../CXX11/src/Tensor/TensorVolumePatch.h      |  30 ++-
 unsupported/test/CMakeLists.txt               |   5 +-
 .../test/cxx11_tensor_volume_patchOP_sycl.cpp | 222 ++++++++++++++++++
 9 files changed, 359 insertions(+), 16 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
index 5b4a9af9f..dd63a2e2f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -177,6 +177,16 @@ KERNELBROKERCONVERTIMAGEPATCHOP()
 #undef KERNELBROKERCONVERTIMAGEPATCHOP
 
 
+/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorVolumePatchOp
+#define KERNELBROKERCONVERTVOLUMEPATCHOP(CVQual)\
+template<DenseIndex Plannes, DenseIndex Rows, DenseIndex Cols, typename XprType>\
+struct ConvertToDeviceExpression<CVQual TensorVolumePatchOp<Plannes, Rows, Cols, XprType> >{\
+  typedef CVQual TensorVolumePatchOp<Plannes, Rows, Cols, typename ConvertToDeviceExpression<XprType>::Type> Type;\
+};
+KERNELBROKERCONVERTVOLUMEPATCHOP(const)
+KERNELBROKERCONVERTVOLUMEPATCHOP()
+#undef KERNELBROKERCONVERTVOLUMEPATCHOP
+
 
 
 }  // namespace internal
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index 57a10d06b..117b368ec 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -404,6 +404,28 @@ SYCLTENSORIMAGEPATCHOPEXPR(const)
 SYCLTENSORIMAGEPATCHOPEXPR()
 #undef SYCLTENSORIMAGEPATCHOPEXPR
 
+// TensorVolumePatchOp
+#define SYCLTENSORVOLUMEPATCHOPEXPR(CVQual)\
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename OrigXprType, typename XprType, typename... Params>\
+struct  ExprConstructor<CVQual TensorVolumePatchOp<Planes, Rows, Cols, OrigXprType>, CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Params... > {\
+  typedef ExprConstructor<OrigXprType, XprType, Params...> my_xpr_type;\
+  typedef CVQual TensorVolumePatchOp<Planes, Rows, Cols, typename my_xpr_type::Type> Type;\
+  my_xpr_type xprExpr;\
+  Type expr;\
+  template <typename FuncDetector>\
+  ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
+  : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.m_patch_planes, funcD.m_patch_rows, funcD.m_patch_cols, funcD.m_plane_strides, funcD.m_row_strides, funcD.m_col_strides,\
+    funcD.m_in_plane_strides, funcD.m_in_row_strides, funcD.m_in_col_strides,funcD.m_plane_inflate_strides, funcD.m_row_inflate_strides, funcD.m_col_inflate_strides, \
+    funcD.m_padding_top_z, funcD.m_padding_bottom_z, funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, funcD.m_padding_value,\
+    funcD.m_padding_type, funcD.m_padding_explicit){\
+    }\
+};
+
+SYCLTENSORVOLUMEPATCHOPEXPR(const)
+SYCLTENSORVOLUMEPATCHOPEXPR()
+#undef SYCLTENSORVOLUMEPATCHOPEXPR
+
+
 
 // TensorLayoutSwapOp
 #define SYCLTENSORLAYOUTSWAPOPEXPR(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index 2be6f3710..4a6322d44 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -240,6 +240,21 @@ SYCLTENSORIMAGEPATCHOPEXTACC()
 #undef SYCLTENSORIMAGEPATCHOPEXTACC
 
 
+
+// specialisation of the \ref ExtractAccessor struct when the node type is
+/// TensorVolumePatchOp.
+#define SYCLTENSORVOLUMEPATCHOPEXTACC(CVQual)\
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType, typename Dev>\
+struct ExtractAccessor<TensorEvaluator<CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Dev> >{\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Dev>& eval)\
+  RETURN_CPP11(AccessorConstructor::getTuple(cgh, eval.impl()))\
+};
+
+SYCLTENSORVOLUMEPATCHOPEXTACC(const)
+SYCLTENSORVOLUMEPATCHOPEXTACC()
+#undef SYCLTENSORVOLUMEPATCHOPEXTACC
+
+
 // specialisation of the \ref ExtractAccessor struct when the node type is
 /// TensorLayoutSwapOp.
 #define SYCLTENSORLAYOUTSWAPOPEXTACC(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index dbac01138..8828a0495 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -344,6 +344,50 @@ FunctorExtractor(const TensorEvaluator<Self, Device>& expr)\
 SYCLEXTRFUNCIMAGEPATCHOP(const)
 SYCLEXTRFUNCIMAGEPATCHOP()
 #undef SYCLEXTRFUNCIMAGEPATCHOP
+
+/// TensorVolumePatchOp
+#define SYCLEXTRFUNCVOLUMEPATCHOP(CVQual)\
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType, typename Device>\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Device> >{\
+typedef CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType> Self;\
+FunctorExtractor<Eigen::TensorEvaluator<XprType, Device> > xprExpr;\
+const DenseIndex m_patch_planes;\
+const DenseIndex m_patch_rows;\
+const DenseIndex m_patch_cols;\
+const DenseIndex m_plane_strides;\
+const DenseIndex m_row_strides;\
+const DenseIndex m_col_strides;\
+const DenseIndex m_in_plane_strides;\
+const DenseIndex m_in_row_strides;\
+const DenseIndex m_in_col_strides;\
+const DenseIndex m_plane_inflate_strides;\
+const DenseIndex m_row_inflate_strides;\
+const DenseIndex m_col_inflate_strides;\
+const bool m_padding_explicit;\
+const DenseIndex m_padding_top_z;\
+const DenseIndex m_padding_bottom_z;\
+const DenseIndex m_padding_top;\
+const DenseIndex m_padding_bottom;\
+const DenseIndex m_padding_left;\
+const DenseIndex m_padding_right;\
+const PaddingType m_padding_type;\
+const typename Self::Scalar m_padding_value;\
+FunctorExtractor(const TensorEvaluator<Self, Device>& expr)\
+: xprExpr(expr.impl()), m_patch_planes(expr.xpr().patch_planes()), m_patch_rows(expr.xpr().patch_rows()), m_patch_cols(expr.xpr().patch_cols()),\
+  m_plane_strides(expr.xpr().plane_strides()), m_row_strides(expr.xpr().row_strides()), m_col_strides(expr.xpr().col_strides()),\
+  m_in_plane_strides(expr.xpr().in_plane_strides()), m_in_row_strides(expr.xpr().in_row_strides()), m_in_col_strides(expr.xpr().in_col_strides()),\
+  m_plane_inflate_strides(expr.xpr().plane_inflate_strides()),m_row_inflate_strides(expr.xpr().row_inflate_strides()),\
+  m_col_inflate_strides(expr.xpr().col_inflate_strides()), m_padding_explicit(expr.xpr().padding_explicit()),\
+  m_padding_top_z(expr.xpr().padding_top_z()), m_padding_bottom_z(expr.xpr().padding_bottom_z()), \
+  m_padding_top(expr.xpr().padding_top()), m_padding_bottom(expr.xpr().padding_bottom()), m_padding_left(expr.xpr().padding_left()),\
+  m_padding_right(expr.xpr().padding_right()), m_padding_type(expr.xpr().padding_type()),m_padding_value(expr.xpr().padding_value()){}\
+};
+SYCLEXTRFUNCVOLUMEPATCHOP(const)
+SYCLEXTRFUNCVOLUMEPATCHOP()
+#undef SYCLEXTRFUNCVOLUMEPATCHOP
+
+
+
 /// template deduction function for FunctorExtractor
 template <typename Evaluator>
 auto inline extractFunctors(const Evaluator& evaluator)-> FunctorExtractor<Evaluator> {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
index b8e658824..50f4595fc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -151,7 +151,7 @@ CHIPPINGOPLEAFCOUNT(const)
 CHIPPINGOPLEAFCOUNT()
 #undef CHIPPINGOPLEAFCOUNT
 
-
+///TensorStridingSlicingOp
 #define SLICESTRIDEOPLEAFCOUNT(CVQual)\
 template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>\
 struct LeafCount<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >:CategoryCount<XprType>{};
@@ -160,7 +160,7 @@ SLICESTRIDEOPLEAFCOUNT(const)
 SLICESTRIDEOPLEAFCOUNT()
 #undef SLICESTRIDEOPLEAFCOUNT
 
-
+//TensorImagePatchOp
 #define TENSORIMAGEPATCHOPLEAFCOUNT(CVQual)\
 template<DenseIndex Rows, DenseIndex Cols, typename XprType>\
 struct LeafCount<CVQual TensorImagePatchOp<Rows, Cols, XprType> >:CategoryCount<XprType>{};
@@ -170,6 +170,15 @@ TENSORIMAGEPATCHOPLEAFCOUNT(const)
 TENSORIMAGEPATCHOPLEAFCOUNT()
 #undef TENSORIMAGEPATCHOPLEAFCOUNT
 
+// TensorVolumePatchOp
+#define TENSORVOLUMEPATCHOPLEAFCOUNT(CVQual)\
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>\
+struct LeafCount<CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType> >:CategoryCount<XprType>{};
+
+
+TENSORVOLUMEPATCHOPLEAFCOUNT(const)
+TENSORVOLUMEPATCHOPLEAFCOUNT()
+#undef TENSORVOLUMEPATCHOPLEAFCOUNT
 
 } /// namespace TensorSycl
 } /// namespace internal
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
index ab97235ae..fcef0be04 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -235,6 +235,20 @@ SYCLTENSORIMAGEPATCHOP()
 #undef SYCLTENSORIMAGEPATCHOP
 
 
+
+/// specialisation of the \ref PlaceHolderExpression when the node is
+/// TensorVolumePatchOp
+#define SYCLTENSORVOLUMEPATCHOP(CVQual)\
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType, size_t N>\
+struct PlaceHolderExpression<CVQual TensorVolumePatchOp<Planes,Rows, Cols, XprType>, N> {\
+  typedef CVQual TensorVolumePatchOp<Planes,Rows, Cols, typename CalculateIndex <N, XprType>::ArgType> Type;\
+};
+
+SYCLTENSORVOLUMEPATCHOP(const)
+SYCLTENSORVOLUMEPATCHOP()
+#undef SYCLTENSORVOLUMEPATCHOP
+
+
 /// template deduction for \ref PlaceHolderExpression struct
 template <typename Expr>
 struct createPlaceHolderExpression {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index 0ca2cac84..64474ee80 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -65,12 +65,8 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                             DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                             DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                             PaddingType padding_type, Scalar padding_value)
-      : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(false), m_padding_top_z(0), m_padding_bottom_z(0), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+      : TensorVolumePatchOp(expr, patch_planes, patch_rows, patch_cols, plane_strides, row_strides, col_strides, in_plane_strides, in_row_strides, in_col_strides,
+      plane_inflate_strides, row_inflate_strides, col_inflate_strides, 0,0,0,0,0,0,padding_value, padding_type, false) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
@@ -79,14 +75,14 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                            DenseIndex padding_top_z, DenseIndex padding_bottom_z,
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
-                                                           Scalar padding_value)
+                                                           Scalar padding_value, PaddingType padding_type=PADDING_VALID, bool padding_explicit=true)
       : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(true), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+        m_padding_explicit(padding_explicit), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
         m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+        m_padding_type(padding_type), m_padding_value(padding_value) {}
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_planes() const { return m_patch_planes; }
@@ -183,9 +179,13 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
     CoordAccess = false,
     RawAccess = false
   };
+#ifdef __SYCL_DEVICE_ONLY__
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType op, const Device& device)
+#else
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
+#endif
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device)
+      : m_impl(op.expression(), device), m_op(op)
   {
     EIGEN_STATIC_ASSERT((NumDims >= 5), YOU_MADE_A_PROGRAMMING_MISTAKE);
 
@@ -321,7 +321,9 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
     m_outputPlanesRows = m_outputPlanes * m_outputRows;
 
     // Fast representations of different variables.
+  //  printf("THis is m_otherStride: %lu\n", m_otherStride );
     m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
+
     m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
     m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
     m_fastRowStride = internal::TensorIntDivisor<Index>(m_rowStride);
@@ -338,7 +340,6 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
       m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
     }
   }
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
@@ -352,6 +353,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
   {
+
     // Patch index corresponding to the passed in index.
     const Index patchIndex = index / m_fastPatchStride;
 
@@ -505,6 +507,8 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  // required by sycl
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
 
   Index planePaddingTop() const { return m_planePaddingTop; }
   Index rowPaddingTop() const { return m_rowPaddingTop; }
@@ -600,6 +604,8 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   Scalar m_paddingValue;
 
   TensorEvaluator<ArgType, Device> m_impl;
+// required by sycl
+  XprType m_op;
 };
 
 
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 69c892362..508f29446 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -167,11 +167,12 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_convolution_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_striding_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_layout_swap_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_inflation_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_generator_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_volume_patchOP_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp b/unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp
new file mode 100644
index 000000000..ddc9e0d46
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp
@@ -0,0 +1,222 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_volume_patchOP_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+static const int DataLayout = ColMajor;
+
+template <typename DataType, typename IndexType>
+static void test_single_voxel_patch_sycl(const Eigen::SyclDevice& sycl_device)
+{
+
+IndexType sizeDim0 = 4;
+IndexType sizeDim1 = 2;
+IndexType sizeDim2 = 3;
+IndexType sizeDim3 = 5;
+IndexType sizeDim4 = 7;
+array<IndexType, 5> tensorColMajorRange = {{sizeDim0, sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
+array<IndexType, 5> tensorRowMajorRange = {{sizeDim4, sizeDim3, sizeDim2, sizeDim1, sizeDim0}};
+Tensor<DataType, 5, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+Tensor<DataType, 5, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+tensor_col_major.setRandom();
+
+
+  DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+  DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+  TensorMap<Tensor<DataType, 5, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+  TensorMap<Tensor<DataType, 5, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+  gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+
+
+  // single volume patch: ColMajor
+  array<IndexType, 6> patchColMajorTensorRange={{sizeDim0,1, 1, 1, sizeDim1*sizeDim2*sizeDim3, sizeDim4}};
+  Tensor<DataType, 6, DataLayout,IndexType> single_voxel_patch_col_major(patchColMajorTensorRange);
+  size_t patchTensorBuffSize =single_voxel_patch_col_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_voxel_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 6, DataLayout,IndexType>> gpu_single_voxel_patch_col_major(gpu_data_single_voxel_patch_col_major, patchColMajorTensorRange);
+  gpu_single_voxel_patch_col_major.device(sycl_device)=gpu_col_major.extract_volume_patches(1, 1, 1);
+  sycl_device.memcpyDeviceToHost(single_voxel_patch_col_major.data(), gpu_data_single_voxel_patch_col_major, patchTensorBuffSize);
+
+
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(0), 4);
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(3), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(4), 2 * 3 * 5);
+  VERIFY_IS_EQUAL(single_voxel_patch_col_major.dimension(5), 7);
+
+  array<IndexType, 6> patchRowMajorTensorRange={{sizeDim4, sizeDim1*sizeDim2*sizeDim3, 1, 1, 1, sizeDim0}};
+  Tensor<DataType, 6, RowMajor,IndexType> single_voxel_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =single_voxel_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_single_voxel_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 6, RowMajor,IndexType>> gpu_single_voxel_patch_row_major(gpu_data_single_voxel_patch_row_major, patchRowMajorTensorRange);
+  gpu_single_voxel_patch_row_major.device(sycl_device)=gpu_row_major.extract_volume_patches(1, 1, 1);
+  sycl_device.memcpyDeviceToHost(single_voxel_patch_row_major.data(), gpu_data_single_voxel_patch_row_major, patchTensorBuffSize);
+
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(0), 7);
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(1), 2 * 3 * 5);
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(3), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(4), 1);
+  VERIFY_IS_EQUAL(single_voxel_patch_row_major.dimension(5), 4);
+
+ sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_col_major.size())*sizeof(DataType));
+ for (IndexType i = 0; i < tensor_col_major.size(); ++i) {
+       VERIFY_IS_EQUAL(tensor_col_major.data()[i], single_voxel_patch_col_major.data()[i]);
+    VERIFY_IS_EQUAL(tensor_row_major.data()[i], single_voxel_patch_row_major.data()[i]);
+    VERIFY_IS_EQUAL(tensor_col_major.data()[i], tensor_row_major.data()[i]);
+  }
+
+
+  sycl_device.deallocate(gpu_data_col_major);
+  sycl_device.deallocate(gpu_data_row_major);
+  sycl_device.deallocate(gpu_data_single_voxel_patch_col_major);
+  sycl_device.deallocate(gpu_data_single_voxel_patch_row_major);
+}
+
+template <typename DataType, typename IndexType>
+static void test_entire_volume_patch_sycl(const Eigen::SyclDevice& sycl_device)
+{
+  const int depth = 4;
+  const int patch_z = 2;
+  const int patch_y = 3;
+  const int patch_x = 5;
+  const int batch = 7;
+
+  array<IndexType, 5> tensorColMajorRange = {{depth, patch_z, patch_y, patch_x, batch}};
+  array<IndexType, 5> tensorRowMajorRange = {{batch, patch_x, patch_y, patch_z, depth}};
+  Tensor<DataType, 5, DataLayout,IndexType> tensor_col_major(tensorColMajorRange);
+  Tensor<DataType, 5, RowMajor,IndexType> tensor_row_major(tensorRowMajorRange);
+  tensor_col_major.setRandom();
+
+
+    DataType* gpu_data_col_major  = static_cast<DataType*>(sycl_device.allocate(tensor_col_major.size()*sizeof(DataType)));
+    DataType* gpu_data_row_major  = static_cast<DataType*>(sycl_device.allocate(tensor_row_major.size()*sizeof(DataType)));
+    TensorMap<Tensor<DataType, 5, ColMajor, IndexType>> gpu_col_major(gpu_data_col_major, tensorColMajorRange);
+    TensorMap<Tensor<DataType, 5, RowMajor, IndexType>> gpu_row_major(gpu_data_row_major, tensorRowMajorRange);
+
+    sycl_device.memcpyHostToDevice(gpu_data_col_major, tensor_col_major.data(),(tensor_col_major.size())*sizeof(DataType));
+    gpu_row_major.device(sycl_device)=gpu_col_major.swap_layout();
+    sycl_device.memcpyDeviceToHost(tensor_row_major.data(), gpu_data_row_major, (tensor_col_major.size())*sizeof(DataType));
+
+
+    // single volume patch: ColMajor
+    array<IndexType, 6> patchColMajorTensorRange={{depth,patch_z, patch_y, patch_x, patch_z*patch_y*patch_x, batch}};
+    Tensor<DataType, 6, DataLayout,IndexType> entire_volume_patch_col_major(patchColMajorTensorRange);
+    size_t patchTensorBuffSize =entire_volume_patch_col_major.size()*sizeof(DataType);
+    DataType* gpu_data_entire_volume_patch_col_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+    TensorMap<Tensor<DataType, 6, DataLayout,IndexType>> gpu_entire_volume_patch_col_major(gpu_data_entire_volume_patch_col_major, patchColMajorTensorRange);
+    gpu_entire_volume_patch_col_major.device(sycl_device)=gpu_col_major.extract_volume_patches(patch_z, patch_y, patch_x);
+    sycl_device.memcpyDeviceToHost(entire_volume_patch_col_major.data(), gpu_data_entire_volume_patch_col_major, patchTensorBuffSize);
+
+
+//  Tensor<float, 5> tensor(depth, patch_z, patch_y, patch_x, batch);
+//  tensor.setRandom();
+//  Tensor<float, 5, RowMajor> tensor_row_major = tensor.swap_layout();
+
+  //Tensor<float, 6> entire_volume_patch;
+  //entire_volume_patch = tensor.extract_volume_patches(patch_z, patch_y, patch_x);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(0), depth);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(1), patch_z);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(2), patch_y);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(3), patch_x);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(4), patch_z * patch_y * patch_x);
+  VERIFY_IS_EQUAL(entire_volume_patch_col_major.dimension(5), batch);
+
+//  Tensor<float, 6, RowMajor> entire_volume_patch_row_major;
+  //entire_volume_patch_row_major = tensor_row_major.extract_volume_patches(patch_z, patch_y, patch_x);
+
+  array<IndexType, 6> patchRowMajorTensorRange={{batch,patch_z*patch_y*patch_x, patch_x, patch_y, patch_z, depth}};
+  Tensor<DataType, 6, RowMajor,IndexType> entire_volume_patch_row_major(patchRowMajorTensorRange);
+  patchTensorBuffSize =entire_volume_patch_row_major.size()*sizeof(DataType);
+  DataType* gpu_data_entire_volume_patch_row_major  = static_cast<DataType*>(sycl_device.allocate(patchTensorBuffSize));
+  TensorMap<Tensor<DataType, 6, RowMajor,IndexType>> gpu_entire_volume_patch_row_major(gpu_data_entire_volume_patch_row_major, patchRowMajorTensorRange);
+  gpu_entire_volume_patch_row_major.device(sycl_device)=gpu_row_major.extract_volume_patches(patch_z, patch_y, patch_x);
+  sycl_device.memcpyDeviceToHost(entire_volume_patch_row_major.data(), gpu_data_entire_volume_patch_row_major, patchTensorBuffSize);
+
+
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(0), batch);
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(1), patch_z * patch_y * patch_x);
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(2), patch_x);
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(3), patch_y);
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(4), patch_z);
+  VERIFY_IS_EQUAL(entire_volume_patch_row_major.dimension(5), depth);
+
+  const int dz = patch_z - 1;
+  const int dy = patch_y - 1;
+  const int dx = patch_x - 1;
+
+  const int forward_pad_z = dz - dz / 2;
+  const int forward_pad_y = dy - dy / 2;
+  const int forward_pad_x = dx - dx / 2;
+
+  for (int pz = 0; pz < patch_z; pz++) {
+    for (int py = 0; py < patch_y; py++) {
+      for (int px = 0; px < patch_x; px++) {
+        const int patchId = pz + patch_z * (py + px * patch_y);
+        for (int z = 0; z < patch_z; z++) {
+          for (int y = 0; y < patch_y; y++) {
+            for (int x = 0; x < patch_x; x++) {
+              for (int b = 0; b < batch; b++) {
+                for (int d = 0; d < depth; d++) {
+                  float expected = 0.0f;
+                  float expected_row_major = 0.0f;
+                  const int eff_z = z - forward_pad_z + pz;
+                  const int eff_y = y - forward_pad_y + py;
+                  const int eff_x = x - forward_pad_x + px;
+                  if (eff_z >= 0 && eff_y >= 0 && eff_x >= 0 &&
+                      eff_z < patch_z && eff_y < patch_y && eff_x < patch_x) {
+                    expected = tensor_col_major(d, eff_z, eff_y, eff_x, b);
+                    expected_row_major = tensor_row_major(b, eff_x, eff_y, eff_z, d);
+                  }
+                  VERIFY_IS_EQUAL(entire_volume_patch_col_major(d, z, y, x, patchId, b), expected);
+                  VERIFY_IS_EQUAL(entire_volume_patch_row_major(b, patchId, x, y, z, d), expected_row_major);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  sycl_device.deallocate(gpu_data_col_major);
+  sycl_device.deallocate(gpu_data_row_major);
+  sycl_device.deallocate(gpu_data_entire_volume_patch_col_major);
+  sycl_device.deallocate(gpu_data_entire_volume_patch_row_major);
+}
+
+
+
+template<typename DataType, typename dev_Selector> void sycl_tensor_volume_patch_test_per_device(dev_Selector s){
+QueueInterface queueInterface(s);
+auto sycl_device = Eigen::SyclDevice(&queueInterface);
+std::cout << "Running on " << s.template get_info<cl::sycl::info::device::name>() << std::endl;
+test_single_voxel_patch_sycl<DataType, int64_t>(sycl_device);
+test_entire_volume_patch_sycl<DataType, int64_t>(sycl_device);
+}
+void test_cxx11_tensor_volume_patchOP_sycl()
+{
+for (const auto& device :Eigen::get_sycl_supported_devices()) {
+  CALL_SUBTEST(sycl_tensor_volume_patch_test_per_device<float>(device));
+}
+}

From 34d9fce93bd5f1521017402154a8ab915af8fcb8 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 16:33:33 -0800
Subject: [PATCH 030/680] Avoid unecessary float to double conversions.

---
 Eigen/src/Core/arch/CUDA/PacketMath.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index ad66399e0..4dda63188 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -291,7 +291,7 @@ template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
 
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<float4,4>& kernel) {
-  double tmp = kernel.packet[0].y;
+  float tmp = kernel.packet[0].y;
   kernel.packet[0].y = kernel.packet[1].x;
   kernel.packet[1].x = tmp;
 

From 554116bec1b19d417521e9bee767d6b57813492a Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 16:45:31 -0800
Subject: [PATCH 031/680] Added EIGEN_DEVICE_FUNC to make the prototype of the
 EigenBase override match that of DenseBase

---
 Eigen/src/Core/EigenBase.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h
index f76995af9..ccc122cfc 100644
--- a/Eigen/src/Core/EigenBase.h
+++ b/Eigen/src/Core/EigenBase.h
@@ -128,6 +128,7 @@ template<typename Derived> struct EigenBase
   */
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
 {
   call_assignment(derived(), other.derived());
@@ -136,6 +137,7 @@ Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
 {
   call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -144,6 +146,7 @@ Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
 {
   call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());

From b1fc7c9a09e2aee938036ddba14e870c9658c791 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 16:48:30 -0800
Subject: [PATCH 032/680] Added missing EIGEN_DEVICE_FUNC qualifiers.

---
 Eigen/src/Core/ArrayBase.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h
index af5fb2566..9da960f08 100644
--- a/Eigen/src/Core/ArrayBase.h
+++ b/Eigen/src/Core/ArrayBase.h
@@ -175,7 +175,7 @@ template<typename Derived> class ArrayBase
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
 {
   call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -188,7 +188,7 @@ ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
 {
   call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -201,7 +201,7 @@ ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
 {
   call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -214,7 +214,7 @@ ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other)
 {
   call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar,typename OtherDerived::Scalar>());

From ed4dc9d01aa46688e12ef1e0772145c1b222602c Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 16:57:01 -0800
Subject: [PATCH 033/680] Declared the plset, ploadt_ro, and ploaddup packet
 primitives as usable within a gpu kernel

---
 Eigen/src/Core/GenericPacketMath.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index ac5552d3e..d19d5bbd2 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -231,7 +231,7 @@ pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(
   * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]}
   * Currently, this function is only used for scalar * complex products.
   */
-template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
 ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
 /** \internal \returns a packet with elements of \a *from quadrupled.
@@ -279,7 +279,7 @@ inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
 }
 
 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
 plset(const typename unpacket_traits<Packet>::type& a) { return a; }
 
 /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
@@ -487,7 +487,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& fro
   * by the current computation.
   */
 template<typename Packet, int LoadMode>
-inline Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
 {
   return ploadt<Packet, LoadMode>(from);
 }

From 193939d6aaca2d8b4ee7cac9f0a89637596c692f Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 17:11:47 -0800
Subject: [PATCH 034/680] Added missing EIGEN_DEVICE_FUNC qualifiers to several
 nullary op methods.

---
 Eigen/src/Core/CwiseNullaryOp.h | 80 ++++++++++++++++-----------------
 1 file changed, 40 insertions(+), 40 deletions(-)

diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h
index dd498f758..ddd607e38 100644
--- a/Eigen/src/Core/CwiseNullaryOp.h
+++ b/Eigen/src/Core/CwiseNullaryOp.h
@@ -105,7 +105,7 @@ class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func)
 {
   return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func);
@@ -150,7 +150,7 @@ DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
 {
   return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func);
@@ -192,7 +192,7 @@ DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Constant(Index size, const Scalar& value)
 {
   return DenseBase<Derived>::NullaryExpr(size, internal::scalar_constant_op<Scalar>(value));
@@ -208,7 +208,7 @@ DenseBase<Derived>::Constant(Index size, const Scalar& value)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Constant(const Scalar& value)
 {
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
@@ -220,7 +220,7 @@ DenseBase<Derived>::Constant(const Scalar& value)
   * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
@@ -232,7 +232,7 @@ DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const
   * \sa LinSpaced(Scalar,Scalar)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
@@ -264,7 +264,7 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig
   * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
@@ -276,7 +276,7 @@ DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
   * Special version for fixed size types which does not require the size parameter.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
@@ -286,7 +286,7 @@ DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
 
 /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
 template<typename Derived>
-bool DenseBase<Derived>::isApproxToConstant
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApproxToConstant
 (const Scalar& val, const RealScalar& prec) const
 {
   typename internal::nested_eval<Derived,1>::type self(derived());
@@ -301,7 +301,7 @@ bool DenseBase<Derived>::isApproxToConstant
   *
   * \returns true if all coefficients in this matrix are approximately equal to \a value, to within precision \a prec */
 template<typename Derived>
-bool DenseBase<Derived>::isConstant
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isConstant
 (const Scalar& val, const RealScalar& prec) const
 {
   return isApproxToConstant(val, prec);
@@ -312,7 +312,7 @@ bool DenseBase<Derived>::isConstant
   * \sa setConstant(), Constant(), class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
 {
   setConstant(val);
 }
@@ -322,7 +322,7 @@ EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
   * \sa fill(), setConstant(Index,const Scalar&), setConstant(Index,Index,const Scalar&), setZero(), setOnes(), Constant(), class CwiseNullaryOp, setZero(), setOnes()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
 {
   return derived() = Constant(rows(), cols(), val);
 }
@@ -337,7 +337,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
   * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
 {
   resize(size);
@@ -356,7 +356,7 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
   * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
 {
   resize(rows, cols);
@@ -380,7 +380,7 @@ PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
   * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar>(low,high,newSize));
@@ -400,7 +400,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, con
   * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return setLinSpaced(size(), low, high);
@@ -423,7 +423,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low,
   * \sa Zero(), Zero(Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Zero(Index rows, Index cols)
 {
   return Constant(rows, cols, Scalar(0));
@@ -446,7 +446,7 @@ DenseBase<Derived>::Zero(Index rows, Index cols)
   * \sa Zero(), Zero(Index,Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Zero(Index size)
 {
   return Constant(size, Scalar(0));
@@ -463,7 +463,7 @@ DenseBase<Derived>::Zero(Index size)
   * \sa Zero(Index), Zero(Index,Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Zero()
 {
   return Constant(Scalar(0));
@@ -478,7 +478,7 @@ DenseBase<Derived>::Zero()
   * \sa class CwiseNullaryOp, Zero()
   */
 template<typename Derived>
-bool DenseBase<Derived>::isZero(const RealScalar& prec) const
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isZero(const RealScalar& prec) const
 {
   typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
@@ -496,7 +496,7 @@ bool DenseBase<Derived>::isZero(const RealScalar& prec) const
   * \sa class CwiseNullaryOp, Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
 {
   return setConstant(Scalar(0));
 }
@@ -511,7 +511,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
   * \sa DenseBase::setZero(), setZero(Index,Index), class CwiseNullaryOp, DenseBase::Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setZero(Index newSize)
 {
   resize(newSize);
@@ -529,7 +529,7 @@ PlainObjectBase<Derived>::setZero(Index newSize)
   * \sa DenseBase::setZero(), setZero(Index), class CwiseNullaryOp, DenseBase::Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setZero(Index rows, Index cols)
 {
   resize(rows, cols);
@@ -553,7 +553,7 @@ PlainObjectBase<Derived>::setZero(Index rows, Index cols)
   * \sa Ones(), Ones(Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Ones(Index rows, Index cols)
 {
   return Constant(rows, cols, Scalar(1));
@@ -576,7 +576,7 @@ DenseBase<Derived>::Ones(Index rows, Index cols)
   * \sa Ones(), Ones(Index,Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Ones(Index newSize)
 {
   return Constant(newSize, Scalar(1));
@@ -593,7 +593,7 @@ DenseBase<Derived>::Ones(Index newSize)
   * \sa Ones(Index), Ones(Index,Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Ones()
 {
   return Constant(Scalar(1));
@@ -608,7 +608,7 @@ DenseBase<Derived>::Ones()
   * \sa class CwiseNullaryOp, Ones()
   */
 template<typename Derived>
-bool DenseBase<Derived>::isOnes
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isOnes
 (const RealScalar& prec) const
 {
   return isApproxToConstant(Scalar(1), prec);
@@ -622,7 +622,7 @@ bool DenseBase<Derived>::isOnes
   * \sa class CwiseNullaryOp, Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
 {
   return setConstant(Scalar(1));
 }
@@ -637,7 +637,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
   * \sa MatrixBase::setOnes(), setOnes(Index,Index), class CwiseNullaryOp, MatrixBase::Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setOnes(Index newSize)
 {
   resize(newSize);
@@ -655,7 +655,7 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
   * \sa MatrixBase::setOnes(), setOnes(Index), class CwiseNullaryOp, MatrixBase::Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
 {
   resize(rows, cols);
@@ -679,7 +679,7 @@ PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
   * \sa Identity(), setIdentity(), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
 MatrixBase<Derived>::Identity(Index rows, Index cols)
 {
   return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>());
@@ -696,7 +696,7 @@ MatrixBase<Derived>::Identity(Index rows, Index cols)
   * \sa Identity(Index,Index), setIdentity(), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
 MatrixBase<Derived>::Identity()
 {
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
@@ -771,7 +771,7 @@ struct setIdentity_impl<Derived, true>
   * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
 {
   return internal::setIdentity_impl<Derived>::run(derived());
 }
@@ -787,7 +787,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
   * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
 {
   derived().resize(rows, cols);
   return setIdentity();
@@ -800,7 +800,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index
   * \sa MatrixBase::Unit(Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return BasisReturnType(SquareMatrixType::Identity(newSize,newSize), i);
@@ -815,7 +815,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return BasisReturnType(SquareMatrixType::Identity(),i);
@@ -828,7 +828,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX()
 { return Derived::Unit(0); }
 
 /** \returns an expression of the Y axis unit vector (0,1{,0}^*)
@@ -838,7 +838,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY()
 { return Derived::Unit(1); }
 
 /** \returns an expression of the Z axis unit vector (0,0,1{,0}^*)
@@ -848,7 +848,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ()
 { return Derived::Unit(2); }
 
 /** \returns an expression of the W axis unit vector (0,0,0,1)
@@ -858,7 +858,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
 { return Derived::Unit(3); }
 
 } // end namespace Eigen

From 889c606f8fd242b1cf5e3c8f967ac7dad004775d Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 27 Feb 2017 17:17:47 -0800
Subject: [PATCH 035/680] Added missing EIGEN_DEVICE_FUNC to the SelfCwise
 binary ops

---
 Eigen/src/Core/SelfCwiseBinaryOp.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h
index 719ed72a5..50099df82 100644
--- a/Eigen/src/Core/SelfCwiseBinaryOp.h
+++ b/Eigen/src/Core/SelfCwiseBinaryOp.h
@@ -15,7 +15,7 @@ namespace Eigen {
 // TODO generalize the scalar type of 'other'
 
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar,Scalar>());
@@ -23,7 +23,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 }
 
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar,Scalar>());
@@ -31,7 +31,7 @@ EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
 }
 
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar,Scalar>());
@@ -39,7 +39,7 @@ EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
 }
 
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar,Scalar>());

From 478a9f53be33c23ac5e22e0bb09cad7f719fedd4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 28 Feb 2017 09:32:45 +0100
Subject: [PATCH 036/680] Fix typo.

---
 unsupported/Eigen/CXX11/src/Tensor/README.md | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index fbb7f3bfc..38cdb9c69 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -75,16 +75,16 @@ large enough to hold all the data.
 
     // Map a tensor of ints on top of stack-allocated storage.
     int storage[128];  // 2 x 4 x 2 x 8 = 128
-    TensorMap<int, 4> t_4d(storage, 2, 4, 2, 8);
+    TensorMap<Tensor<int, 4>> t_4d(storage, 2, 4, 2, 8);
 
     // The same storage can be viewed as a different tensor.
     // You can also pass the sizes as an array.
-    TensorMap<int, 2> t_2d(storage, 16, 8);
+    TensorMap<Tensor<int, 2>> t_2d(storage, 16, 8);
 
     // You can also map fixed-size tensors.  Here we get a 1d view of
     // the 2d fixed-size tensor.
     Tensor<float, Sizes<4, 5>> t_4x3;
-    TensorMap<float, 1> t_12(t_4x3, 12);
+    TensorMap<Tensor<float, 1>> t_12(t_4x3, 12);
 
 
 #### Class TensorRef

From 4e98a7b2f0501408751e2da5f24d65f642371226 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 28 Feb 2017 09:47:38 +0100
Subject: [PATCH 037/680] bug #1396: add some missing EIGEN_DEVICE_FUNC

---
 Eigen/src/Core/DenseBase.h  | 6 +++---
 Eigen/src/Core/MatrixBase.h | 4 ++--
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index fc807577b..91a8511be 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -484,9 +484,9 @@ template<typename Derived> class DenseBase
       return derived().coeff(0,0);
     }
 
-    bool all() const;
-    bool any() const;
-    Index count() const;
+    EIGEN_DEVICE_FUNC bool all() const;
+    EIGEN_DEVICE_FUNC bool any() const;
+    EIGEN_DEVICE_FUNC Index count() const;
 
     typedef VectorwiseOp<Derived, Horizontal> RowwiseReturnType;
     typedef const VectorwiseOp<const Derived, Horizontal> ConstRowwiseReturnType;
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 675c94e12..200e57741 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -294,7 +294,7 @@ template<typename Derived> class MatrixBase
       *          fuzzy comparison such as isApprox()
       * \sa isApprox(), operator!= */
     template<typename OtherDerived>
-    inline bool operator==(const MatrixBase<OtherDerived>& other) const
+    EIGEN_DEVICE_FUNC inline bool operator==(const MatrixBase<OtherDerived>& other) const
     { return cwiseEqual(other).all(); }
 
     /** \returns true if at least one pair of coefficients of \c *this and \a other are not exactly equal to each other.
@@ -302,7 +302,7 @@ template<typename Derived> class MatrixBase
       *          fuzzy comparison such as isApprox()
       * \sa isApprox(), operator== */
     template<typename OtherDerived>
-    inline bool operator!=(const MatrixBase<OtherDerived>& other) const
+    EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase<OtherDerived>& other) const
     { return cwiseNotEqual(other).any(); }
 
     NoAlias<Derived,Eigen::MatrixBase > noalias();

From 8296b87d7bd98c19c6064241880691f164790ede Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 28 Feb 2017 17:16:14 +0000
Subject: [PATCH 038/680] Adding sycl backend for TensorCustomOp; fixing the
 partial lhs modification issue on sycl when the rhs is TensorContraction,
 reduction or convolution; Fixing the partial modification for memset when
 sycl backend is used.

---
 .../CXX11/src/Tensor/TensorContractionSycl.h  |  21 +--
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |  30 ++--
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   |   6 +
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h |  23 ++-
 .../CXX11/src/Tensor/TensorReductionSycl.h    |  16 +-
 .../TensorSyclConvertToDeviceExpression.h     |   1 +
 .../src/Tensor/TensorSyclExprConstructor.h    |  61 +++++--
 .../src/Tensor/TensorSyclExtractAccessor.h    |  27 +++
 .../src/Tensor/TensorSyclExtractFunctors.h    |  43 ++++-
 .../CXX11/src/Tensor/TensorSyclFunctors.h     |  21 ++-
 .../CXX11/src/Tensor/TensorSyclLeafCount.h    |  20 +++
 .../src/Tensor/TensorSyclPlaceHolderExpr.h    |  27 +++
 unsupported/test/CMakeLists.txt               |   1 +
 .../test/cxx11_tensor_custom_op_sycl.cpp      | 165 ++++++++++++++++++
 .../test/cxx11_tensor_forced_eval_sycl.cpp    |   2 +-
 15 files changed, 397 insertions(+), 67 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_custom_op_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
index abc7ba551..fcd7d4d00 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
@@ -84,7 +84,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
     this->m_leftImpl.evalSubExprsIfNeeded(NULL);
     this->m_rightImpl.evalSubExprsIfNeeded(NULL);
-    if (data) {
+   if (data) {
       evalTo(data);
       return false;
     } else {
@@ -173,6 +173,7 @@ typename HostExpr::Index LocalThreadSizeM, typename HostExpr::Index LocalThreadS
   LhsLocalAcc localLhs;
   RhsLocalAcc localRhs;
   OutAccessor out_res;
+  size_t out_offset;
   Index roundUpK, M, N, K;
   ContractT m_k_strides, m_left_contracting_strides, m_right_contracting_strides;
   LeftNocontractT m_i_strides, m_left_nocontract_strides;
@@ -182,11 +183,12 @@ typename HostExpr::Index LocalThreadSizeM, typename HostExpr::Index LocalThreadS
   Device dev;
 
 
-  KernelConstructor(LHSFunctorExpr lhs_functors_, RHSFunctorExpr rhs_functors_, LhsLocalAcc localLhs_, RhsLocalAcc localRhs_, OutAccessor out_res_,
+  KernelConstructor(LHSFunctorExpr lhs_functors_, RHSFunctorExpr rhs_functors_, LhsLocalAcc localLhs_, RhsLocalAcc localRhs_, OutAccessor out_res_, size_t out_offset_,
     Index roundUpK_, Index M_, Index N_, Index K_, ContractT m_k_strides_, ContractT m_left_contracting_strides_,
     ContractT m_right_contracting_strides_, LeftNocontractT m_i_strides_, RightNocontractT m_j_strides_,
     LeftNocontractT m_left_nocontract_strides_, RightNocontractT m_right_nocontract_strides_, LHSTupleType left_tuple_of_accessors_, RHSTupleType right_tuple_of_accessors_, Device dev_)
-    :lhs_functors(lhs_functors_), rhs_functors(rhs_functors_), localLhs(localLhs_), localRhs(localRhs_), out_res(out_res_), roundUpK(roundUpK_), M(M_), N(N_), K(K_),
+    :lhs_functors(lhs_functors_), rhs_functors(rhs_functors_), localLhs(localLhs_), localRhs(localRhs_), out_res(out_res_),
+    out_offset(out_offset_), roundUpK(roundUpK_), M(M_), N(N_), K(K_),
     m_k_strides(m_k_strides_), m_left_contracting_strides(m_left_contracting_strides_),
     m_right_contracting_strides(m_right_contracting_strides_),
     m_i_strides(m_i_strides_), m_left_nocontract_strides(m_left_nocontract_strides_),
@@ -316,7 +318,7 @@ typename HostExpr::Index LocalThreadSizeM, typename HostExpr::Index LocalThreadS
             for (Index wLPTN=0; wLPTN<WorkLoadPerThreadN; wLPTN++) {
                 Index globalCol = nGroupId*TileSizeDimN + nLocalThreadId + wLPTN*LocalThreadSizeN;
                 if(globalCol<N)
-                  out_ptr[globalCol*M + globalRow] = privateRes[wLPTM][wLPTN];
+                  out_ptr[globalCol*M + globalRow +ConvertToActualSyclOffset(OutScalar, out_offset)] = privateRes[wLPTM][wLPTN];
             }
           }
       }
@@ -356,12 +358,12 @@ template< typename Self, typename OutScalar, typename ContractT, typename LeftNo
     // extract lhs functor list
     LHSFunctorExpr lhs_functors = Eigen::TensorSycl::internal::extractFunctors(self.left_impl());
     // extract rhs functor list
-    RHSFunctorExpr rhs_functors = Eigen::TensorSycl::internal::extractFunctors(self.left_impl());
+    RHSFunctorExpr rhs_functors = Eigen::TensorSycl::internal::extractFunctors(self.right_impl());
 
     Index roundUpK = RoundUp(K, TileSizeDimK);
     Index roundUpM = RoundUp(M, TileSizeDimM);
     Index roundUpN = RoundUp(N, TileSizeDimN);
-
+    ptrdiff_t out_offset = self.device().get_offset(buffer);
     self.device().sycl_queue().submit([&](cl::sycl::handler &cgh) {
       /// work-around for gcc bug
       typedef decltype(Eigen::TensorSycl::internal::createTupleOfAccessors<OrigLHSExpr>(cgh, self.left_impl())) LHSTupleType;
@@ -379,17 +381,16 @@ template< typename Self, typename OutScalar, typename ContractT, typename LeftNo
       typedef cl::sycl::accessor<RhsScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local> RhsLocalAcc;
       RhsLocalAcc localRhs(cl::sycl::range<1>(2* TileSizeDimK * TileSizeDimN), cgh);
 
-      typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer> OutAccessor;
+      typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::global_buffer> OutAccessor;
       //OutScalar memory
-      OutAccessor out_res= self.device(). template get_sycl_accessor<cl::sycl::access::mode::write>(cgh, buffer);
-
+      OutAccessor out_res= self.device(). template get_sycl_accessor<cl::sycl::access::mode::read_write>(cgh, buffer);
       // sycl parallel for
       cgh.parallel_for(cl::sycl::nd_range<2>(cl::sycl::range<2>(roundUpM/WorkLoadPerThreadM, roundUpN/WorkLoadPerThreadN),
       cl::sycl::range<2>(LocalThreadSizeM, LocalThreadSizeN)),
        KernelConstructor<HostExpr, OutScalar, LhsScalar, RhsScalar, LHSFunctorExpr, RHSFunctorExpr, LhsLocalAcc, RhsLocalAcc, OutAccessor, Index, ContractT, LeftNocontractT,
        RightNocontractT, lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, TileSizeDimM, TileSizeDimN, TileSizeDimK,
        WorkLoadPerThreadM, WorkLoadPerThreadN, LocalThreadSizeM, LocalThreadSizeN, LoadPerThreadLhs, LoadPerThreadRhs, LHSTupleType, RHSTupleType, Eigen::DefaultDevice>(lhs_functors, rhs_functors,
-          localLhs, localRhs, out_res, roundUpK, M, N, K, m_k_strides, m_left_contracting_strides, m_right_contracting_strides,m_i_strides, m_j_strides,
+          localLhs, localRhs, out_res, out_offset, roundUpK, M, N, K, m_k_strides, m_left_contracting_strides, m_right_contracting_strides,m_i_strides, m_j_strides,
           m_left_nocontract_strides,m_right_nocontract_strides, left_tuple_of_accessors, right_tuple_of_accessors, Eigen::DefaultDevice()));
     });
     self.device().asynchronousExec();
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index 4247c1c4a..66ffd819f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -32,14 +32,15 @@ internal::IndexMapper<Index, InputDims, 1, Eigen::internal::traits<HostExpr>::La
 Kernel_accessor kernel_filter;
 const size_t kernelSize, range_x, range_y;
 Buffer_accessor buffer_acc;
+ptrdiff_t out_offset;
 Local_accessor local_acc;
 FunctorExpr functors;
 TupleType tuple_of_accessors;
 EigenConvolutionKernel1D(internal::IndexMapper<Index, InputDims, 1, Eigen::internal::traits<HostExpr>::Layout> indexMapper_,
   Kernel_accessor kernel_filter_,  const size_t kernelSize_, const size_t range_x_, const size_t range_y_,
-  Buffer_accessor buffer_acc_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
+  Buffer_accessor buffer_acc_, ptrdiff_t out_offset_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
   :indexMapper(indexMapper_), kernel_filter(kernel_filter_), kernelSize(kernelSize_), range_x(range_x_), range_y(range_y_),
-  buffer_acc(buffer_acc_), local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
+  buffer_acc(buffer_acc_), out_offset(out_offset_),local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
 
   void operator()(cl::sycl::nd_item<2> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
@@ -75,7 +76,7 @@ EigenConvolutionKernel1D(internal::IndexMapper<Index, InputDims, 1, Eigen::inter
       }
       const size_t tensor_index = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(itemID.get_global(1))
       +indexMapper.mapCudaOutputKernelToTensorOutputOffset(itemID.get_local(0) + first_output_start);
-      buffer_ptr[tensor_index] = result;
+      buffer_ptr[tensor_index+ConvertToActualSyclOffset(CoeffReturnType, out_offset)] = result;
     }
   }
 };
@@ -89,14 +90,15 @@ internal::IndexMapper<Index, InputDims, 2, Eigen::internal::traits<HostExpr>::La
 Kernel_accessor kernel_filter;
 const size_t kernelSize_x, kernelSize_y, range_x, range_y , range_z;
 Buffer_accessor buffer_acc;
+ptrdiff_t out_offset;
 Local_accessor local_acc;
 FunctorExpr functors;
 TupleType tuple_of_accessors;
 EigenConvolutionKernel2D(internal::IndexMapper<Index, InputDims, 2, Eigen::internal::traits<HostExpr>::Layout> indexMapper_,
   Kernel_accessor kernel_filter_,  const size_t kernelSize_x_, const size_t kernelSize_y_ ,const size_t range_x_, const size_t range_y_, const size_t range_z_,
-  Buffer_accessor buffer_acc_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
+  Buffer_accessor buffer_acc_, ptrdiff_t out_offset_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
   :indexMapper(indexMapper_), kernel_filter(kernel_filter_), kernelSize_x(kernelSize_x_), kernelSize_y(kernelSize_y_), range_x(range_x_), range_y(range_y_), range_z(range_z_),
-  buffer_acc(buffer_acc_), local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
+  buffer_acc(buffer_acc_), out_offset(out_offset_), local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
 
   void operator()(cl::sycl::nd_item<3> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
@@ -141,7 +143,7 @@ EigenConvolutionKernel2D(internal::IndexMapper<Index, InputDims, 2, Eigen::inter
       }
       const size_t tensor_index = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(itemID.get_global(2))
       +indexMapper.mapCudaOutputKernelToTensorOutputOffset(itemID.get_local(0) + fitst_x_output_start, itemID.get_local(1) + fitst_y_output_start);
-      buffer_ptr[tensor_index] = result;
+      buffer_ptr[tensor_index  +ConvertToActualSyclOffset(CoeffReturnType, out_offset)] = result;
     }
   }
 };
@@ -156,16 +158,17 @@ internal::IndexMapper<Index, InputDims, 3, Eigen::internal::traits<HostExpr>::La
 Kernel_accessor kernel_filter;
 const size_t kernelSize_x, kernelSize_y, kernelSize_z, range_x, range_y , range_z, numP;
 Buffer_accessor buffer_acc;
+ptrdiff_t out_offset;
 Local_accessor local_acc;
 FunctorExpr functors;
 TupleType tuple_of_accessors;
 EigenConvolutionKernel3D(internal::IndexMapper<Index, InputDims, 3, Eigen::internal::traits<HostExpr>::Layout> indexMapper_,
   Kernel_accessor kernel_filter_,  const size_t kernelSize_x_, const size_t kernelSize_y_ , const size_t kernelSize_z_ ,
   const size_t range_x_, const size_t range_y_, const size_t range_z_, const size_t numP_,
-  Buffer_accessor buffer_acc_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
+  Buffer_accessor buffer_acc_, ptrdiff_t out_offset_, Local_accessor local_acc_, FunctorExpr functors_, TupleType tuple_of_accessors_)
   :indexMapper(indexMapper_), kernel_filter(kernel_filter_), kernelSize_x(kernelSize_x_), kernelSize_y(kernelSize_y_),
   kernelSize_z(kernelSize_z_), range_x(range_x_), range_y(range_y_), range_z(range_z_), numP(numP_),
-  buffer_acc(buffer_acc_), local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
+  buffer_acc(buffer_acc_), out_offset(out_offset_), local_acc(local_acc_), functors(functors_), tuple_of_accessors(tuple_of_accessors_) {}
 
   void operator()(cl::sycl::nd_item<3> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
@@ -215,7 +218,7 @@ EigenConvolutionKernel3D(internal::IndexMapper<Index, InputDims, 3, Eigen::inter
         }
         const size_t tensor_index = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p)
         +indexMapper.mapCudaOutputKernelToTensorOutputOffset(itemID.get_local(0) + fitst_x_output_start, itemID.get_local(1) + fitst_y_output_start, itemID.get_local(2) + fitst_z_output_start );
-        buffer_ptr[tensor_index] = result;
+        buffer_ptr[tensor_index+ConvertToActualSyclOffset(CoeffReturnType, out_offset)] = result;
       }
 
       itemID.barrier(cl::sycl::access::fence_space::local_space);
@@ -307,7 +310,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
       m_kernel = in_place;
       m_local_kernel = false;
     } else {
-      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      ptrdiff_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
       Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
       typedef TensorEvalToOp<const KernelArgType> EvalTo;
       EvalTo evalToTmp(local, m_kernelArg);
@@ -325,6 +328,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
     typedef Eigen::TensorSycl::internal::FunctorExtractor<InputEvaluator> InputFunctorExpr;
     // extract input functor list
     InputFunctorExpr input_functors = Eigen::TensorSycl::internal::extractFunctors(m_inputImpl);
+    ptrdiff_t out_offset = m_device.get_offset(data);
 
 
     m_device.sycl_queue().submit([&](cl::sycl::handler &cgh) {
@@ -358,7 +362,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           cgh.parallel_for(cl::sycl::nd_range<2>(global_range, local_range),
           EigenConvolutionKernel1D<CoeffReturnType, Scalar, InputArgType, InputFunctorExpr, Index,
           InputDims, KernelAccessorType, OutputAccessorType, InputLocalAcc, InputTupleType>(
-          indexMapper,kernel_acc, kernel_size, numX, numP, out_res, local_acc, input_functors, tuple_of_accessors));
+          indexMapper,kernel_acc, kernel_size, numX, numP, out_res, out_offset, local_acc, input_functors, tuple_of_accessors));
           break;
         }
 
@@ -383,7 +387,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           cgh.parallel_for(cl::sycl::nd_range<3>(global_range, local_range),
           EigenConvolutionKernel2D<CoeffReturnType, Scalar, InputArgType, InputFunctorExpr, Index,
           InputDims, KernelAccessorType, OutputAccessorType, InputLocalAcc, InputTupleType>(
-          indexMapper,kernel_acc, kernel_size_x,  kernel_size_y, numX, numY, numP, out_res, local_acc, input_functors, tuple_of_accessors));
+          indexMapper,kernel_acc, kernel_size_x,  kernel_size_y, numX, numY, numP, out_res, out_offset, local_acc, input_functors, tuple_of_accessors));
           break;
         }
 
@@ -412,7 +416,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           EigenConvolutionKernel3D<CoeffReturnType, Scalar, InputArgType, InputFunctorExpr, Index,
           InputDims, KernelAccessorType, OutputAccessorType, InputLocalAcc, InputTupleType>(
           indexMapper,kernel_acc, kernel_size_x,  kernel_size_y, kernel_size_z, numX, numY,
-          numZ, numP, out_res, local_acc, input_functors, tuple_of_accessors));
+          numZ, numP, out_res, out_offset, local_acc, input_functors, tuple_of_accessors));
           break;
         }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index e020d076f..c72d79435 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -140,6 +140,9 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
 
   EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
 
+  /// used by sycl in order to build the sycl buffer
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const{return m_device;}
+
  protected:
   EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
     TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(
@@ -295,6 +298,9 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
 
   EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
 
+  /// used by sycl in order to build the sycl buffer
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const{return m_device;}
+
  protected:
   EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
     TensorMap<Tensor<Scalar, NumDims, Layout> > result(data, m_dimensions);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index e209799bb..964222a15 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -18,6 +18,8 @@
 namespace Eigen {
 
   #define ConvertToActualTypeSycl(Scalar, buf_acc) reinterpret_cast<typename cl::sycl::global_ptr<Scalar>::pointer_t>((&(*buf_acc.get_pointer())))
+  #define ConvertToActualSyclOffset(Scalar, offset) offset/sizeof(Scalar)
+
 
   template <typename Scalar, typename read_accessor, typename write_accessor> class MemCopyFunctor {
   public:
@@ -43,11 +45,12 @@ namespace Eigen {
   struct memsetkernelFunctor{
    typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> AccType;
    AccType m_acc;
+   const ptrdiff_t buff_offset;
    const size_t m_rng, m_c;
-   memsetkernelFunctor(AccType acc, const size_t rng, const size_t c):m_acc(acc), m_rng(rng), m_c(c){}
+   memsetkernelFunctor(AccType acc, const ptrdiff_t buff_offset_, const size_t rng, const size_t c):m_acc(acc),  buff_offset(buff_offset_), m_rng(rng), m_c(c){}
    void operator()(cl::sycl::nd_item<1> itemID) {
      auto globalid=itemID.get_global_linear_id();
-     if (globalid< m_rng) m_acc[globalid] = m_c;
+     if (globalid< m_rng) m_acc[globalid + buff_offset] = m_c;
    }
 
   };
@@ -305,6 +308,11 @@ struct SyclDevice {
     synchronize();
   }
 
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    auto it = m_queue_stream->find_buffer(ptr);
+    return (static_cast<const uint8_t*>(ptr))-it->first;
+
+  }
   /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
   /// pointer created as a key we find the sycl buffer and get the host accessor with discard_write mode
   /// on it. Using a discard_write accessor guarantees that we do not bring back the current value of the
@@ -343,20 +351,23 @@ struct SyclDevice {
   EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
     size_t rng, GRange, tileSize;
     parallel_for_setup(n, tileSize, rng, GRange);
-    sycl_queue().submit(memsetCghFunctor(get_sycl_buffer(static_cast<uint8_t*>(static_cast<void*>(data))),rng, GRange, tileSize, c ));
+    auto it1 = m_queue_stream->find_buffer(static_cast<const void*>(data));
+    ptrdiff_t buff_offset= (static_cast<const uint8_t*>(data)) - it1->first;
+    sycl_queue().submit(memsetCghFunctor(it1->second, buff_offset, rng, GRange, tileSize, c ));
     synchronize();
   }
 
   struct memsetCghFunctor{
     cl::sycl::buffer<uint8_t, 1>& m_buf;
+    const ptrdiff_t& buff_offset;
     const size_t& rng , GRange, tileSize;
     const int  &c;
-    memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
-    :m_buf(buff), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
+    memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
+    :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
 
     void operator()(cl::sycl::handler &cgh) const {
       auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor(buf_acc, rng, c));
+      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor(buf_acc, buff_offset, rng, c));
     }
   };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index c3ca129e2..c9c7acfdc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -27,9 +27,9 @@ namespace internal {
 
 template<typename OP, typename CoeffReturnType> struct syclGenericBufferReducer{
 template<typename BufferTOut, typename BufferTIn>
-static void run(OP op, BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
+static void run(OP op, BufferTOut& bufOut, ptrdiff_t out_offset, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
   do {
-          auto f = [length, local, op, &bufOut, &bufI](cl::sycl::handler& h) mutable {
+          auto f = [length, local, op, out_offset, &bufOut, &bufI](cl::sycl::handler& h) mutable {
             cl::sycl::nd_range<1> r{cl::sycl::range<1>{std::max(length, local)},
                                     cl::sycl::range<1>{std::min(length, local)}};
             /* Two accessors are used: one to the buffer that is being reduced,
@@ -43,7 +43,7 @@ static void run(OP op, BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDev
 
             /* The parallel_for invocation chosen is the variant with an nd_item
              * parameter, since the code requires barriers for correctness. */
-            h.parallel_for(r, TensorSycl::internal::GenericKernelReducer<CoeffReturnType, OP, OutputAccessor, InputAccessor, LocalAccessor>(op, aOut, aI, scratch,  length, local));
+            h.parallel_for(r, TensorSycl::internal::GenericKernelReducer<CoeffReturnType, OP, OutputAccessor, InputAccessor, LocalAccessor>(op, aOut, out_offset, aI, scratch,  length, local));
           };
             dev.sycl_queue().submit(f);
             dev.asynchronousExec();
@@ -60,9 +60,9 @@ static void run(OP op, BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDev
 
 template<typename CoeffReturnType> struct syclGenericBufferReducer<Eigen::internal::MeanReducer<CoeffReturnType>, CoeffReturnType>{
 template<typename BufferTOut, typename BufferTIn>
-static void run(Eigen::internal::MeanReducer<CoeffReturnType>, BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
+static void run(Eigen::internal::MeanReducer<CoeffReturnType>, BufferTOut& bufOut,ptrdiff_t out_offset, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
    syclGenericBufferReducer<Eigen::internal::SumReducer<CoeffReturnType>, CoeffReturnType>::run(Eigen::internal::SumReducer<CoeffReturnType>(),
-    bufOut, bufI, dev, length, local);
+    bufOut, out_offset, bufI, dev, length, local);
 }
 };
 
@@ -127,8 +127,9 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
 
     // getting final out buffer at the moment the created buffer is true because there is no need for assign
     auto out_buffer =dev.get_sycl_buffer(output);
+    ptrdiff_t out_offset = dev.get_offset(output);
     /// This is used to recursively reduce the tmp value to an element of 1;
-    syclGenericBufferReducer<Op, CoeffReturnType>::run(reducer, out_buffer, temp_global_buffer,dev, GRange,  outTileSize);
+    syclGenericBufferReducer<Op, CoeffReturnType>::run(reducer, out_buffer, out_offset, temp_global_buffer,dev, GRange,  outTileSize);
   }
 
 };
@@ -158,10 +159,11 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
       // create a tuple of accessors from Evaluator
       Tuple_of_Acc tuple_of_accessors =  TensorSycl::internal::createTupleOfAccessors(cgh, self.impl());
       auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write>(cgh, output);
+      ptrdiff_t out_offset = dev.get_offset(output);
       Index red_size = (num_values_to_reduce!=0)? num_values_to_reduce : static_cast<Index>(1);
       cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)),
       TensorSycl::internal::ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Op, typename Self::Index>
-      (output_accessor, functors, tuple_of_accessors, self.xprDims(), reducer, range, red_size));
+      (output_accessor, out_offset, functors, tuple_of_accessors, self.xprDims(), reducer, range, red_size));
 
     });
     dev.asynchronousExec();
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
index dd63a2e2f..9476c0ea8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -109,6 +109,7 @@ struct ConvertToDeviceExpression<CVQual ExprNode<Expr> > {\
   typedef CVQual ExprNode< typename ConvertToDeviceExpression<Expr>::Type> Type;\
 };
 
+
 // TensorForcedEvalOp
 KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(const,TensorForcedEvalOp)
 KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(,TensorForcedEvalOp)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index 117b368ec..af4eb5f13 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -236,8 +236,12 @@ EVALTO()
 template <typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
 struct ExprConstructor<CVQual TensorForcedEvalOp<OrigExpr>,\
 CVQual PlaceHolder<CVQual TensorForcedEvalOp<DevExpr>, N>, Params...> {\
-  typedef CVQual TensorMap<Tensor<typename TensorForcedEvalOp<DevExpr>::Scalar,\
-  TensorForcedEvalOp<DevExpr>::NumDimensions, Eigen::internal::traits<TensorForcedEvalOp<DevExpr>>::Layout, typename TensorForcedEvalOp<DevExpr>::Index>, Eigen::internal::traits<TensorForcedEvalOp<DevExpr>>::Layout, MakeGlobalPointer> Type;\
+  typedef  TensorForcedEvalOp<OrigExpr> XprType;\
+  typedef CVQual TensorMap<\
+                            Tensor<typename XprType::Scalar,XprType::NumDimensions, Eigen::internal::traits<XprType>::Layout,typename XprType::Index>,\
+                            Eigen::internal::traits<XprType>::Layout, \
+                            MakeGlobalPointer\
+                          > Type;\
   Type expr;\
   template <typename FuncDetector>\
   ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
@@ -248,6 +252,28 @@ FORCEDEVAL(const)
 FORCEDEVAL()
 #undef FORCEDEVAL
 
+
+
+#define TENSORCUSTOMUNARYOP(CVQual)\
+template <typename CustomUnaryFunc, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
+struct ExprConstructor<CVQual TensorCustomUnaryOp<CustomUnaryFunc, OrigExpr>,\
+CVQual PlaceHolder<CVQual TensorCustomUnaryOp<CustomUnaryFunc, DevExpr>, N>, Params...> {\
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, OrigExpr> XprType;\
+  typedef CVQual TensorMap<\
+                            Tensor<typename XprType::Scalar,XprType::NumDimensions, Eigen::internal::traits<XprType>::Layout,typename XprType::Index>,\
+                            Eigen::internal::traits<XprType>::Layout, \
+                            MakeGlobalPointer\
+                          > Type;\
+  Type expr;\
+  template <typename FuncDetector>\
+  ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
+  : expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())) {}\
+};
+
+TENSORCUSTOMUNARYOP(const)
+TENSORCUSTOMUNARYOP()
+#undef TENSORCUSTOMUNARYOP
+
 template <bool Conds,  size_t X , size_t Y > struct ValueCondition {
   static const size_t Res =X;
 };
@@ -260,7 +286,7 @@ template<size_t X, size_t Y> struct ValueCondition<false, X , Y> {
 template <typename OP, typename Dim, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
 struct ExprConstructor<CVQual TensorReductionOp<OP, Dim, OrigExpr, MakeGlobalPointer>,\
 CVQual PlaceHolder<CVQual TensorReductionOp<OP, Dim, DevExpr>, N>, Params...> {\
-  static const size_t NumIndices= ValueCondition< TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions==0,  1, TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions >::Res;\
+  static const auto NumIndices= ValueCondition< TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions==0,  1, TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions >::Res;\
   typedef CVQual TensorMap<Tensor<typename TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::Scalar,\
   NumIndices, Eigen::internal::traits<TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>>::Layout, typename TensorReductionOp<OP, Dim, DevExpr>::Index>, Eigen::internal::traits<TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>>::Layout, MakeGlobalPointer> Type;\
   Type expr;\
@@ -275,28 +301,31 @@ SYCLREDUCTIONEXPR()
 
 
 /// specialisation of the \ref ExprConstructor struct when the node type is
-/// TensorContractionOp
-#define SYCLCONTRACTIONCONVOLUTION(CVQual, ExprNode)\
+/// TensorContractionOp, TensorConvolutionOp TensorCustomBinaryOp
+#define SYCLCONTRACTCONVCUSBIOPS(CVQual, ExprNode)\
 template <typename Indices, typename OrigLhsXprType, typename OrigRhsXprType, typename LhsXprType, typename RhsXprType, size_t N, typename... Params>\
 struct ExprConstructor<CVQual ExprNode<Indices, OrigLhsXprType, OrigRhsXprType>,\
 CVQual PlaceHolder<CVQual ExprNode<Indices, LhsXprType,  RhsXprType>, N>, Params...> {\
-  static const size_t NumIndices= Eigen::internal::traits<ExprNode<Indices, OrigLhsXprType, OrigRhsXprType> >::NumDimensions;\
-  typedef CVQual TensorMap<Tensor<typename ExprNode<Indices, OrigLhsXprType, OrigRhsXprType>::Scalar,\
-  NumIndices, Eigen::internal::traits<ExprNode<Indices, OrigRhsXprType, OrigRhsXprType> >::Layout,\
-  typename ExprNode<Indices, OrigRhsXprType, OrigRhsXprType>::Index>,\
-  Eigen::internal::traits<ExprNode<Indices, OrigRhsXprType, OrigRhsXprType>>::Layout, MakeGlobalPointer> Type;\
+  typedef ExprNode<Indices, OrigLhsXprType, OrigRhsXprType> XprTyp;\
+  static const auto NumIndices= Eigen::internal::traits<XprTyp>::NumDimensions;\
+  typedef CVQual TensorMap<\
+                            Tensor<typename XprTyp::Scalar,NumIndices, Eigen::internal::traits<XprTyp>::Layout, typename XprTyp::Index>,\
+                            Eigen::internal::traits<XprTyp>::Layout, \
+                            MakeGlobalPointer\
+                          > Type;\
   Type expr;\
   template <typename FuncDetector>\
   ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
   :expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())) {}\
 };
 
-SYCLCONTRACTIONCONVOLUTION(const, TensorContractionOp)
-SYCLCONTRACTIONCONVOLUTION(, TensorContractionOp)
-SYCLCONTRACTIONCONVOLUTION(const, TensorConvolutionOp)
-SYCLCONTRACTIONCONVOLUTION(, TensorConvolutionOp)
-#undef SYCLCONTRACTIONCONVOLUTION
-
+SYCLCONTRACTCONVCUSBIOPS(const, TensorContractionOp)
+SYCLCONTRACTCONVCUSBIOPS(, TensorContractionOp)
+SYCLCONTRACTCONVCUSBIOPS(const, TensorConvolutionOp)
+SYCLCONTRACTCONVCUSBIOPS(, TensorConvolutionOp)
+SYCLCONTRACTCONVCUSBIOPS(const, TensorCustomBinaryOp)
+SYCLCONTRACTCONVCUSBIOPS(, TensorCustomBinaryOp)
+#undef SYCLCONTRACTCONVCUSBIOPS
 
 
 #define SYCLSLICEOPEXPR(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index 4a6322d44..5a6a8f4c5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -148,6 +148,33 @@ SYCLFORCEDEVALEXTACC()
 #undef SYCLFORCEDEVALEXTACC
 
 
+#define SYCLCUSTOMUNARYOPEXTACC(CVQual)\
+template <typename CustomUnaryFunc, typename XprType, typename Dev >\
+struct ExtractAccessor<TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Dev> > {\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Dev>& eval)\
+  RETURN_CPP11(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval))\
+};
+
+
+SYCLCUSTOMUNARYOPEXTACC(const)
+SYCLCUSTOMUNARYOPEXTACC()
+#undef SYCLCUSTOMUNARYOPEXTACC
+
+
+#define SYCLCUSTOMBINARYOPEXTACC(CVQual)\
+template <typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType , typename Dev>\
+struct ExtractAccessor<TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Dev> > {\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Dev>& eval)\
+  RETURN_CPP11(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval))\
+};
+
+SYCLCUSTOMBINARYOPEXTACC(const)
+SYCLCUSTOMBINARYOPEXTACC()
+#undef SYCLCUSTOMBIBARYOPEXTACC
+
+
+
+
 /// specialisation of the \ref ExtractAccessor struct when the node type is TensorEvalToOp
 #define SYCLEVALTOEXTACC(CVQual)\
 template <typename Expr, typename Dev>\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index 8828a0495..9fcac5ecb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -33,14 +33,17 @@ namespace internal {
 /// re-instantiate them on the device.
 /// We have to pass instantiated functors to the device.
 // This struct is used for leafNode (TensorMap) and nodes behaving like leafNode (TensorForcedEval).
+#define DEFALTACTION(Evaluator)\
+typedef typename Evaluator::Dimensions Dimensions;\
+const Dimensions m_dimensions;\
+EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }\
+FunctorExtractor(const Evaluator& expr): m_dimensions(expr.dimensions()) {}
+
 template <typename Evaluator> struct FunctorExtractor{
-  typedef typename Evaluator::Dimensions Dimensions;
-  const Dimensions m_dimensions;
-  EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
-  FunctorExtractor(const Evaluator& expr)
-  : m_dimensions(expr.dimensions()) {}
+  DEFALTACTION(Evaluator)
 };
 
+
 /// specialisation of the \ref FunctorExtractor struct when the node type does not require anything
 ///TensorConversionOp
 #define SYCLEXTRFUNCCONVERSION(ExprNode, CVQual)\
@@ -112,6 +115,36 @@ SYCLEXTRFUNCTERNARY(const)
 SYCLEXTRFUNCTERNARY()
 #undef SYCLEXTRFUNCTERNARY
 
+
+
+//TensorCustomOp must be specialised otherewise it will be captured by UnaryCategory while its action is different
+//from the UnaryCategory and it is similar to the general FunctorExtractor.
+/// specialisation of TensorCustomOp
+#define SYCLEXTRFUNCCUSTOMUNARYOP(CVQual)\
+template <typename CustomUnaryFunc, typename ArgType, typename Dev >\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, ArgType>, Dev> > {\
+  typedef TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, ArgType>, Dev> Evaluator;\
+  DEFALTACTION(Evaluator)\
+};
+
+SYCLEXTRFUNCCUSTOMUNARYOP(const)
+SYCLEXTRFUNCCUSTOMUNARYOP()
+#undef SYCLEXTRFUNCCUSTOMUNARYOP
+
+
+#define SYCLEXTRFUNCCUSTOMBIBARYOP(CVQual)\
+template <typename CustomBinaryFunc, typename ArgType1, typename ArgType2, typename Dev >\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, ArgType1, ArgType2>, Dev> > {\
+  typedef TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, ArgType1, ArgType2>, Dev> Evaluator;\
+  DEFALTACTION(Evaluator)\
+};
+
+SYCLEXTRFUNCCUSTOMBIBARYOP(const)
+SYCLEXTRFUNCCUSTOMBIBARYOP()
+#undef SYCLEXTRFUNCCUSTOMBIBARYOP
+
+
+
 /// specialisation of the \ref FunctorExtractor struct when the node type is
 /// TensorCwiseSelectOp. This is an specialisation without OP so it has to be separated.
 #define SYCLEXTRFUNCSELECTOP(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
index 2f7779036..12237bfab 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
@@ -21,11 +21,12 @@ namespace internal {
   template<typename CoeffReturnType, typename OP, typename OutputAccessor, typename InputAccessor, typename LocalAccessor> struct GenericKernelReducer{
     OP op;
     OutputAccessor aOut;
+    ptrdiff_t out_offset;
     InputAccessor aI;
     LocalAccessor scratch;
     size_t length, local;
-    GenericKernelReducer(OP op_, OutputAccessor aOut_, InputAccessor aI_, LocalAccessor scratch_, size_t length_, size_t local_)
-    : op(op_), aOut(aOut_), aI(aI_), scratch(scratch_), length(length_), local(local_){}
+    GenericKernelReducer(OP op_, OutputAccessor aOut_, ptrdiff_t out_offset_, InputAccessor aI_, LocalAccessor scratch_, size_t length_, size_t local_)
+    : op(op_), aOut(aOut_), out_offset(out_offset_), aI(aI_), scratch(scratch_), length(length_), local(local_){}
     void operator()(cl::sycl::nd_item<1> itemID) {
       size_t globalid = itemID.get_global(0);
       size_t localid = itemID.get_local(0);
@@ -59,7 +60,7 @@ namespace internal {
           aI[itemID.get_group(0)] = scratch[localid];
           if((length<=local) && globalid ==0){
             auto aOutPtr = ConvertToActualTypeSycl(CoeffReturnType, aOut);
-            aOutPtr[0]=scratch[0];
+            aOutPtr[0 + ConvertToActualSyclOffset(CoeffReturnType, out_offset)]=scratch[0];
           }
         }
       }
@@ -72,8 +73,8 @@ template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typen
  public:
   typedef  typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr;
   typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> write_accessor;
-  ReductionFunctor(write_accessor output_accessor_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_, Op functor_, Index range_, Index)
-  :output_accessor(output_accessor_), functors(functors_), tuple_of_accessors(tuple_of_accessors_), dims(dims_), functor(functor_), range(range_) {}
+  ReductionFunctor(write_accessor output_accessor_, ptrdiff_t out_offset_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_, Op functor_, Index range_, Index)
+  :output_accessor(output_accessor_), out_offset(out_offset_), functors(functors_), tuple_of_accessors(tuple_of_accessors_), dims(dims_), functor(functor_), range(range_) {}
   void operator()(cl::sycl::nd_item<1> itemID) {
 
     typedef typename ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
@@ -93,11 +94,12 @@ template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typen
       typename DeviceSelf::CoeffReturnType accum = functor.initialize();
       Eigen::internal::GenericDimReducer<DeviceSelf::NumReducedDims-1, DeviceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(static_cast<typename DevExpr::Index>(globalid)),const_cast<Op&>(functor), &accum);
       functor.finalize(accum);
-      output_accessor_ptr[globalid]= accum;
+      output_accessor_ptr[globalid + ConvertToActualSyclOffset(typename DeviceSelf::CoeffReturnType, out_offset)]= accum;
     }
   }
  private:
   write_accessor output_accessor;
+  ptrdiff_t out_offset;
   FunctorExpr functors;
   Tuple_of_Acc tuple_of_accessors;
   Dims dims;
@@ -111,9 +113,9 @@ class ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Eigen::interna
   typedef  typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr;
   typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> write_accessor;
   typedef Eigen::internal::SumReducer<typename HostExpr::CoeffReturnType> Op;
-  ReductionFunctor(write_accessor output_accessor_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_,
+  ReductionFunctor(write_accessor output_accessor_, ptrdiff_t out_offset_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_,
     Eigen::internal::MeanReducer<typename HostExpr::CoeffReturnType>,  Index range_, Index num_values_to_reduce_)
-  :output_accessor(output_accessor_), functors(functors_), tuple_of_accessors(tuple_of_accessors_), dims(dims_), functor(Op()), range(range_), num_values_to_reduce(num_values_to_reduce_) {}
+  :output_accessor(output_accessor_),  out_offset(out_offset_), functors(functors_), tuple_of_accessors(tuple_of_accessors_), dims(dims_), functor(Op()), range(range_), num_values_to_reduce(num_values_to_reduce_) {}
   void operator()(cl::sycl::nd_item<1> itemID) {
 
     typedef typename ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
@@ -133,11 +135,12 @@ class ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Eigen::interna
       typename DeviceSelf::CoeffReturnType accum = functor.initialize();
       Eigen::internal::GenericDimReducer<DeviceSelf::NumReducedDims-1, DeviceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(static_cast<typename DevExpr::Index>(globalid)),const_cast<Op&>(functor), &accum);
       functor.finalize(accum);
-      output_accessor_ptr[globalid]= accum/num_values_to_reduce;
+      output_accessor_ptr[globalid+ ConvertToActualSyclOffset(typename DeviceSelf::CoeffReturnType, out_offset)]= accum/num_values_to_reduce;
     }
   }
  private:
   write_accessor output_accessor;
+  ptrdiff_t out_offset;
   FunctorExpr functors;
   Tuple_of_Acc tuple_of_accessors;
   Dims dims;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
index 50f4595fc..330283b39 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -93,6 +93,26 @@ SYCLFORCEDEVALLEAFCOUNT(const)
 SYCLFORCEDEVALLEAFCOUNT()
 #undef SYCLFORCEDEVALLEAFCOUNT
 
+#define SYCLCUSTOMUNARYOPLEAFCOUNT(CVQual)\
+template <typename CustomUnaryFunc, typename XprType>\
+struct LeafCount<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType> > {\
+static const size_t Count =1;\
+};
+
+SYCLCUSTOMUNARYOPLEAFCOUNT(const)
+SYCLCUSTOMUNARYOPLEAFCOUNT()
+#undef SYCLCUSTOMUNARYOPLEAFCOUNT
+
+
+#define SYCLCUSTOMBINARYOPLEAFCOUNT(CVQual)\
+template <typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>\
+struct LeafCount<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> > {\
+static const size_t Count =1;\
+};
+SYCLCUSTOMBINARYOPLEAFCOUNT( const)
+SYCLCUSTOMBINARYOPLEAFCOUNT()
+#undef SYCLCUSTOMBINARYOPLEAFCOUNT
+
 /// specialisation of the \ref LeafCount struct when the node type is TensorEvalToOp
 #define EVALTOLAYOUTSWAPLEAFCOUNT(CVQual , ExprNode, Num)\
 template <typename Expr>\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
index fcef0be04..99d528963 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -143,6 +143,33 @@ FORCEDEVAL(const)
 FORCEDEVAL()
 #undef FORCEDEVAL
 
+
+/// specialisation of the \ref PlaceHolderExpression when the node is
+/// TensorForcedEvalOp
+#define CUSTOMUNARYOPEVAL(CVQual)\
+template <typename CustomUnaryFunc, typename XprType, size_t N>\
+struct PlaceHolderExpression<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType>, N> {\
+  typedef CVQual PlaceHolder<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType>, N> Type;\
+};
+
+CUSTOMUNARYOPEVAL(const)
+CUSTOMUNARYOPEVAL()
+#undef CUSTOMUNARYOPEVAL
+
+
+/// specialisation of the \ref PlaceHolderExpression when the node is
+/// TensorForcedEvalOp
+#define CUSTOMBINARYOPEVAL(CVQual)\
+template <typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, size_t N>\
+struct PlaceHolderExpression<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, N> {\
+  typedef CVQual PlaceHolder<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, N> Type;\
+};
+
+CUSTOMBINARYOPEVAL(const)
+CUSTOMBINARYOPEVAL()
+#undef CUSTOMBINARYOPEVAL
+
+
 /// specialisation of the \ref PlaceHolderExpression when the node is
 /// TensorEvalToOp, TensorLayoutSwapOp
 #define EVALTOLAYOUTSWAP(CVQual, ExprNode)\
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 508f29446..996178292 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -173,6 +173,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_volume_patchOP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_custom_op_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
   # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_custom_op_sycl.cpp b/unsupported/test/cxx11_tensor_custom_op_sycl.cpp
new file mode 100644
index 000000000..9ff287fff
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_custom_op_sycl.cpp
@@ -0,0 +1,165 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_custom_op_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+template<typename TensorType>
+struct InsertZeros {
+  DSizes<DenseIndex, 2> dimensions(const TensorType& input) const {
+    DSizes<DenseIndex, 2> result;
+    result[0] = input.dimension(0) * 2;
+    result[1] = input.dimension(1) * 2;
+    return result;
+  }
+
+  template <typename Output, typename Device>
+  void eval(const TensorType& input, Output& output, const Device& device) const
+  {
+    array<DenseIndex, 2> strides;
+    strides[0] = 2;
+    strides[1] = 2;
+    output.stride(strides).device(device) = input;
+
+    Eigen::DSizes<DenseIndex, 2> offsets(1,1);
+    Eigen::DSizes<DenseIndex, 2> extents(output.dimension(0)-1, output.dimension(1)-1);
+    output.slice(offsets, extents).stride(strides).device(device) = input.constant(0.0f);
+  }
+};
+
+template<typename DataType, int DataLayout, typename IndexType>
+static void test_custom_unary_op_sycl(const Eigen::SyclDevice &sycl_device)
+{
+  IndexType sizeDim1 = 3;
+  IndexType sizeDim2 = 5;
+  Eigen::array<IndexType, 2> tensorRange = {{sizeDim1, sizeDim2}};
+  Eigen::array<IndexType, 2> tensorResultRange = {{6, 10}};
+
+  Eigen::Tensor<DataType, 2, DataLayout, IndexType> in1(tensorRange);
+  Eigen::Tensor<DataType, 2, DataLayout, IndexType> out(tensorResultRange);
+
+  DataType * gpu_in1_data  = static_cast<DataType*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(DataType)));
+  DataType * gpu_out_data =  static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType)));
+
+  typedef Eigen::TensorMap<Eigen::Tensor<DataType, 2, DataLayout, IndexType> > TensorType;
+  TensorType gpu_in1(gpu_in1_data, tensorRange);
+  TensorType gpu_out(gpu_out_data, tensorResultRange);
+
+  in1.setRandom();
+  sycl_device.memcpyHostToDevice(gpu_in1_data, in1.data(),(in1.dimensions().TotalSize())*sizeof(DataType));
+  gpu_out.device(sycl_device) = gpu_in1.customOp(InsertZeros<TensorType>());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType));
+
+  VERIFY_IS_EQUAL(out.dimension(0), 6);
+  VERIFY_IS_EQUAL(out.dimension(1), 10);
+
+  for (int i = 0; i < 6; i+=2) {
+    for (int j = 0; j < 10; j+=2) {
+      VERIFY_IS_EQUAL(out(i, j), in1(i/2, j/2));
+    }
+  }
+  for (int i = 1; i < 6; i+=2) {
+    for (int j = 1; j < 10; j+=2) {
+      VERIFY_IS_EQUAL(out(i, j), 0);
+    }
+  }
+}
+
+template<typename TensorType>
+struct BatchMatMul {
+  DSizes<DenseIndex, 3> dimensions(const TensorType& input1, const TensorType& input2) const {
+    DSizes<DenseIndex, 3> result;
+    result[0] = input1.dimension(0);
+    result[1] = input2.dimension(1);
+    result[2] = input2.dimension(2);
+    return result;
+  }
+
+  template <typename Output, typename Device>
+  void eval(const TensorType& input1, const TensorType& input2,
+            Output& output, const Device& device) const
+  {
+    typedef typename TensorType::DimensionPair DimPair;
+    array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    for (int64_t i = 0; i < output.dimension(2); ++i) {
+      output.template chip<2>(i).device(device) = input1.template chip<2>(i).contract(input2.template chip<2>(i), dims);
+    }
+  }
+};
+
+template<typename DataType, int DataLayout, typename IndexType>
+static void test_custom_binary_op_sycl(const Eigen::SyclDevice &sycl_device)
+{
+
+  Eigen::array<IndexType, 3> tensorRange1 = {{2, 3, 5}};
+  Eigen::array<IndexType, 3> tensorRange2 = {{3,7,5}};
+  Eigen::array<IndexType, 3> tensorResultRange  = {{2, 7, 5}};
+
+  Eigen::Tensor<DataType, 3, DataLayout, IndexType> in1(tensorRange1);
+  Eigen::Tensor<DataType, 3, DataLayout, IndexType> in2(tensorRange2);
+  Eigen::Tensor<DataType, 3, DataLayout, IndexType> out(tensorResultRange);
+
+  DataType * gpu_in1_data  = static_cast<DataType*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(DataType)));
+  DataType * gpu_in2_data  = static_cast<DataType*>(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(DataType)));
+  DataType * gpu_out_data =  static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType)));
+
+  typedef Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType> > TensorType;
+  TensorType gpu_in1(gpu_in1_data, tensorRange1);
+  TensorType gpu_in2(gpu_in2_data, tensorRange2);
+  TensorType gpu_out(gpu_out_data, tensorResultRange);
+
+  in1.setRandom();
+  in2.setRandom();
+
+  sycl_device.memcpyHostToDevice(gpu_in1_data, in1.data(),(in1.dimensions().TotalSize())*sizeof(DataType));
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.dimensions().TotalSize())*sizeof(DataType));
+
+  gpu_out.device(sycl_device) = gpu_in1.customOp(gpu_in2, BatchMatMul<TensorType>());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType));
+
+  for (IndexType i = 0; i < 5; ++i) {
+    typedef typename Eigen::Tensor<DataType, 3, DataLayout, IndexType>::DimensionPair DimPair;
+    array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    Eigen::Tensor<DataType, 2, DataLayout, IndexType> reference = in1.template chip<2>(i).contract(in2.template chip<2>(i), dims);
+    TensorRef<Eigen::Tensor<DataType, 2, DataLayout, IndexType> > val = out.template chip<2>(i);
+    for (IndexType j = 0; j < 2; ++j) {
+      for (IndexType k = 0; k < 7; ++k) {
+        VERIFY_IS_APPROX(val(j, k), reference(j, k));
+      }
+    }
+  }
+}
+
+template <typename DataType, typename Dev_selector> void custom_op_perDevice(Dev_selector s){
+  QueueInterface queueInterface(s);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_custom_unary_op_sycl<DataType, RowMajor, int64_t>(sycl_device);
+  test_custom_unary_op_sycl<DataType, ColMajor, int64_t>(sycl_device);
+  test_custom_binary_op_sycl<DataType, ColMajor, int64_t>(sycl_device);
+  test_custom_binary_op_sycl<DataType, RowMajor, int64_t>(sycl_device);
+
+}
+void test_cxx11_tensor_custom_op_sycl() {
+  for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(custom_op_perDevice<float>(device));
+  }
+}
diff --git a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
index aca036cde..a21514d56 100644
--- a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
@@ -44,7 +44,7 @@ void test_forced_eval_sycl(const Eigen::SyclDevice &sycl_device) {
   Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_in2(gpu_in2_data, tensorRange);
   Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_out(gpu_out_data, tensorRange);
   sycl_device.memcpyHostToDevice(gpu_in1_data, in1.data(),(in1.dimensions().TotalSize())*sizeof(DataType));
-  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in1.dimensions().TotalSize())*sizeof(DataType));
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.dimensions().TotalSize())*sizeof(DataType));
   /// c=(a+b)*b
   gpu_out.device(sycl_device) =(gpu_in1 + gpu_in2).eval() * gpu_in2;
   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType));

From f3e9c42876de4b49bdb16b72b09a342d930ce6f0 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 09:46:30 -0800
Subject: [PATCH 039/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/Diagonal.h  | 10 ++++++----
 Eigen/src/Core/Transpose.h | 10 +++++-----
 2 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h
index 49e711257..c62f5ff21 100644
--- a/Eigen/src/Core/Diagonal.h
+++ b/Eigen/src/Core/Diagonal.h
@@ -184,7 +184,7 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
   *
   * \sa class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::DiagonalReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::DiagonalReturnType
 MatrixBase<Derived>::diagonal()
 {
   return DiagonalReturnType(derived());
@@ -192,7 +192,7 @@ MatrixBase<Derived>::diagonal()
 
 /** This is the const version of diagonal(). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::ConstDiagonalReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::ConstDiagonalReturnType
 MatrixBase<Derived>::diagonal() const
 {
   return ConstDiagonalReturnType(derived());
@@ -210,7 +210,7 @@ MatrixBase<Derived>::diagonal() const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index)
 {
   return DiagonalDynamicIndexReturnType(derived(), index);
@@ -218,7 +218,7 @@ MatrixBase<Derived>::diagonal(Index index)
 
 /** This is the const version of diagonal(Index). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index) const
 {
   return ConstDiagonalDynamicIndexReturnType(derived(), index);
@@ -237,6 +237,7 @@ MatrixBase<Derived>::diagonal(Index index) const
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
 template<int Index_>
+EIGEN_DEVICE_FUNC
 inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal()
 {
@@ -246,6 +247,7 @@ MatrixBase<Derived>::diagonal()
 /** This is the const version of diagonal<int>(). */
 template<typename Derived>
 template<int Index_>
+EIGEN_DEVICE_FUNC
 inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal() const
 {
diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h
index 79b767bcc..ba7d6e629 100644
--- a/Eigen/src/Core/Transpose.h
+++ b/Eigen/src/Core/Transpose.h
@@ -168,7 +168,7 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
   *
   * \sa transposeInPlace(), adjoint() */
 template<typename Derived>
-inline Transpose<Derived>
+EIGEN_DEVICE_FUNC inline Transpose<Derived>
 DenseBase<Derived>::transpose()
 {
   return TransposeReturnType(derived());
@@ -180,7 +180,7 @@ DenseBase<Derived>::transpose()
   *
   * \sa transposeInPlace(), adjoint() */
 template<typename Derived>
-inline typename DenseBase<Derived>::ConstTransposeReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ConstTransposeReturnType
 DenseBase<Derived>::transpose() const
 {
   return ConstTransposeReturnType(derived());
@@ -206,7 +206,7 @@ DenseBase<Derived>::transpose() const
   *
   * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */
 template<typename Derived>
-inline const typename MatrixBase<Derived>::AdjointReturnType
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::AdjointReturnType
 MatrixBase<Derived>::adjoint() const
 {
   return AdjointReturnType(this->transpose());
@@ -281,7 +281,7 @@ struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non squ
   *
   * \sa transpose(), adjoint(), adjointInPlace() */
 template<typename Derived>
-inline void DenseBase<Derived>::transposeInPlace()
+EIGEN_DEVICE_FUNC inline void DenseBase<Derived>::transposeInPlace()
 {
   eigen_assert((rows() == cols() || (RowsAtCompileTime == Dynamic && ColsAtCompileTime == Dynamic))
                && "transposeInPlace() called on a non-square non-resizable matrix");
@@ -312,7 +312,7 @@ inline void DenseBase<Derived>::transposeInPlace()
   *
   * \sa transpose(), adjoint(), transposeInPlace() */
 template<typename Derived>
-inline void MatrixBase<Derived>::adjointInPlace()
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::adjointInPlace()
 {
   derived() = adjoint().eval();
 }

From 33443ec2b0d116daa5210fc5e7982ca4a1598bc7 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 09:50:10 -0800
Subject: [PATCH 040/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/Redux.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index b6e8f8887..2b5b73bf7 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -407,7 +407,7 @@ protected:
   */
 template<typename Derived>
 template<typename Func>
-typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
   eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
@@ -422,7 +422,7 @@ DenseBase<Derived>::redux(const Func& func) const
   * \warning the result is undefined if \c *this contains NaN.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff() const
 {
   return derived().redux(Eigen::internal::scalar_min_op<Scalar,Scalar>());
@@ -432,7 +432,7 @@ DenseBase<Derived>::minCoeff() const
   * \warning the result is undefined if \c *this contains NaN.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff() const
 {
   return derived().redux(Eigen::internal::scalar_max_op<Scalar,Scalar>());
@@ -445,7 +445,7 @@ DenseBase<Derived>::maxCoeff() const
   * \sa trace(), prod(), mean()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::sum() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
@@ -458,7 +458,7 @@ DenseBase<Derived>::sum() const
 * \sa trace(), prod(), sum()
 */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::mean() const
 {
 #ifdef __INTEL_COMPILER
@@ -479,7 +479,7 @@ DenseBase<Derived>::mean() const
   * \sa sum(), mean(), trace()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::prod() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
@@ -494,7 +494,7 @@ DenseBase<Derived>::prod() const
   * \sa diagonal(), sum()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 MatrixBase<Derived>::trace() const
 {
   return derived().diagonal().sum();

From e993c94f07faef161076c8710e8cb434174dd250 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 09:56:45 -0800
Subject: [PATCH 041/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/CommaInitializer.h |  4 ++--
 Eigen/src/Core/CwiseBinaryOp.h    |  5 ++---
 Eigen/src/Core/CwiseNullaryOp.h   |  4 ++--
 Eigen/src/Core/Fuzzy.h            |  6 +++---
 Eigen/src/Core/NestByValue.h      | 10 +++++-----
 Eigen/src/Core/ReturnByValue.h    |  2 +-
 6 files changed, 15 insertions(+), 16 deletions(-)

diff --git a/Eigen/src/Core/CommaInitializer.h b/Eigen/src/Core/CommaInitializer.h
index d218e9814..35fdbb819 100644
--- a/Eigen/src/Core/CommaInitializer.h
+++ b/Eigen/src/Core/CommaInitializer.h
@@ -141,7 +141,7 @@ struct CommaInitializer
   * \sa CommaInitializer::finished(), class CommaInitializer
   */
 template<typename Derived>
-inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s)
+EIGEN_DEVICE_FUNC inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s)
 {
   return CommaInitializer<Derived>(*static_cast<Derived*>(this), s);
 }
@@ -149,7 +149,7 @@ inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s
 /** \sa operator<<(const Scalar&) */
 template<typename Derived>
 template<typename OtherDerived>
-inline CommaInitializer<Derived>
+EIGEN_DEVICE_FUNC inline CommaInitializer<Derived>
 DenseBase<Derived>::operator<<(const DenseBase<OtherDerived>& other)
 {
   return CommaInitializer<Derived>(*static_cast<Derived *>(this), other);
diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h
index a36765e39..bf2632d9e 100644
--- a/Eigen/src/Core/CwiseBinaryOp.h
+++ b/Eigen/src/Core/CwiseBinaryOp.h
@@ -158,7 +158,7 @@ public:
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
 {
   call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -171,7 +171,7 @@ MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
 {
   call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -181,4 +181,3 @@ MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
 } // end namespace Eigen
 
 #endif // EIGEN_CWISE_BINARY_OP_H
-
diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h
index ddd607e38..144608ec2 100644
--- a/Eigen/src/Core/CwiseNullaryOp.h
+++ b/Eigen/src/Core/CwiseNullaryOp.h
@@ -131,7 +131,7 @@ DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& f
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
@@ -170,7 +170,7 @@ DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
 {
   return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value));
diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h
index 3e403a09d..43aa49b2b 100644
--- a/Eigen/src/Core/Fuzzy.h
+++ b/Eigen/src/Core/Fuzzy.h
@@ -100,7 +100,7 @@ struct isMuchSmallerThan_scalar_selector<Derived, true>
   */
 template<typename Derived>
 template<typename OtherDerived>
-bool DenseBase<Derived>::isApprox(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApprox(
   const DenseBase<OtherDerived>& other,
   const RealScalar& prec
 ) const
@@ -122,7 +122,7 @@ bool DenseBase<Derived>::isApprox(
   * \sa isApprox(), isMuchSmallerThan(const DenseBase<OtherDerived>&, RealScalar) const
   */
 template<typename Derived>
-bool DenseBase<Derived>::isMuchSmallerThan(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isMuchSmallerThan(
   const typename NumTraits<Scalar>::Real& other,
   const RealScalar& prec
 ) const
@@ -142,7 +142,7 @@ bool DenseBase<Derived>::isMuchSmallerThan(
   */
 template<typename Derived>
 template<typename OtherDerived>
-bool DenseBase<Derived>::isMuchSmallerThan(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isMuchSmallerThan(
   const DenseBase<OtherDerived>& other,
   const RealScalar& prec
 ) const
diff --git a/Eigen/src/Core/NestByValue.h b/Eigen/src/Core/NestByValue.h
index 13adf070e..01cf192e9 100644
--- a/Eigen/src/Core/NestByValue.h
+++ b/Eigen/src/Core/NestByValue.h
@@ -67,25 +67,25 @@ template<typename ExpressionType> class NestByValue
     }
 
     template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index row, Index col) const
     {
       return m_expression.template packet<LoadMode>(row, col);
     }
 
     template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    EIGEN_DEVICE_FUNC inline void writePacket(Index row, Index col, const PacketScalar& x)
     {
       m_expression.const_cast_derived().template writePacket<LoadMode>(row, col, x);
     }
 
     template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
+    EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index index) const
     {
       return m_expression.template packet<LoadMode>(index);
     }
 
     template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
+    EIGEN_DEVICE_FUNC inline void writePacket(Index index, const PacketScalar& x)
     {
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
     }
@@ -99,7 +99,7 @@ template<typename ExpressionType> class NestByValue
 /** \returns an expression of the temporary version of *this.
   */
 template<typename Derived>
-inline const NestByValue<Derived>
+EIGEN_DEVICE_FUNC inline const NestByValue<Derived>
 DenseBase<Derived>::nestByValue() const
 {
   return NestByValue<Derived>(derived());
diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h
index c44b7673b..11dc86d07 100644
--- a/Eigen/src/Core/ReturnByValue.h
+++ b/Eigen/src/Core/ReturnByValue.h
@@ -79,7 +79,7 @@ template<typename Derived> class ReturnByValue
 
 template<typename Derived>
 template<typename OtherDerived>
-Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
+EIGEN_DEVICE_FUNC Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
   other.evalTo(derived());
   return derived();

From 765f4cc4b4774fc114ab794e998c5ab4f2d733f9 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 11:57:00 -0800
Subject: [PATCH 042/680] Deleted extra: EIGEN_DEVICE_FUNC: the QR and Cholesky
 code isn't ready to run on GPU yet.

---
 Eigen/src/Cholesky/LDLT.h                      | 1 -
 Eigen/src/Cholesky/LLT.h                       | 1 -
 Eigen/src/QR/ColPivHouseholderQR.h             | 1 -
 Eigen/src/QR/CompleteOrthogonalDecomposition.h | 2 +-
 Eigen/src/QR/FullPivHouseholderQR.h            | 9 ++++-----
 Eigen/src/QR/HouseholderQR.h                   | 9 ++++-----
 6 files changed, 9 insertions(+), 14 deletions(-)

diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h
index fcee7b2e3..9b4fdb414 100644
--- a/Eigen/src/Cholesky/LDLT.h
+++ b/Eigen/src/Cholesky/LDLT.h
@@ -258,7 +258,6 @@ template<typename _MatrixType, int _UpLo> class LDLT
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
     #endif
 
diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 87ca8d423..e6c02d803 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -200,7 +200,6 @@ template<typename _MatrixType, int _UpLo> class LLT
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
     #endif
 
diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index 0e47c8332..d35395d04 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -416,7 +416,6 @@ template<typename _MatrixType> class ColPivHouseholderQR
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
     #endif
 
diff --git a/Eigen/src/QR/CompleteOrthogonalDecomposition.h b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
index 34c637b70..13b61fcdb 100644
--- a/Eigen/src/QR/CompleteOrthogonalDecomposition.h
+++ b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
@@ -367,7 +367,7 @@ class CompleteOrthogonalDecomposition {
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
   template <typename RhsType, typename DstType>
-  EIGEN_DEVICE_FUNC void _solve_impl(const RhsType& rhs, DstType& dst) const;
+  void _solve_impl(const RhsType& rhs, DstType& dst) const;
 #endif
 
  protected:
diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h
index e489bddc2..c31e47cc4 100644
--- a/Eigen/src/QR/FullPivHouseholderQR.h
+++ b/Eigen/src/QR/FullPivHouseholderQR.h
@@ -392,22 +392,21 @@ template<typename _MatrixType> class FullPivHouseholderQR
       *          diagonal coefficient of U.
       */
     RealScalar maxPivot() const { return m_maxpivot; }
-    
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
     #endif
 
   protected:
-    
+
     static void check_template_parameters()
     {
       EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
     }
-    
+
     void computeInPlace();
-    
+
     MatrixType m_qr;
     HCoeffsType m_hCoeffs;
     IntDiagSizeVectorType m_rows_transpositions;
diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h
index 3513d995c..762b21c36 100644
--- a/Eigen/src/QR/HouseholderQR.h
+++ b/Eigen/src/QR/HouseholderQR.h
@@ -204,28 +204,27 @@ template<typename _MatrixType> class HouseholderQR
 
     inline Index rows() const { return m_qr.rows(); }
     inline Index cols() const { return m_qr.cols(); }
-    
+
     /** \returns a const reference to the vector of Householder coefficients used to represent the factor \c Q.
       * 
       * For advanced uses only.
       */
     const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
-    
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
     #endif
 
   protected:
-    
+
     static void check_template_parameters()
     {
       EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
     }
 
     void computeInPlace();
-    
+
     MatrixType m_qr;
     HCoeffsType m_hCoeffs;
     RowVectorType m_temp;

From de7b0fdea9db957d2135c32e850ad069b64b5f1e Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 13:52:22 -0800
Subject: [PATCH 043/680] Made the TensorStorage class compile with clang 3.9

---
 unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
index 2854a4a17..e6a666f78 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
@@ -31,12 +31,12 @@ namespace Eigen {
   *
   * \sa Tensor
   */
-template<typename T, typename Dimensions, int Options_> class TensorStorage;
+template<typename T, typename Dimensions, int Options> class TensorStorage;
 
 
 // Pure fixed-size storage
-template<typename T, int Options_, typename FixedDimensions>
-class TensorStorage<T, FixedDimensions, Options_>
+template<typename T, typename FixedDimensions, int Options_>
+class TensorStorage
 {
  private:
   static const std::size_t Size = FixedDimensions::total_size;
@@ -66,7 +66,7 @@ class TensorStorage<T, FixedDimensions, Options_>
 
 
 // pure dynamic
-template<typename T, int Options_, typename IndexType, int NumIndices_>
+template<typename T, typename IndexType, int NumIndices_, int Options_>
 class TensorStorage<T, DSizes<IndexType, NumIndices_>, Options_>
 {
   public:

From 4a7df114c883eb17251f25b8b975c0ddf266acd6 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 14:00:15 -0800
Subject: [PATCH 044/680] Added missing EIGEN_DEVICE_FUNC

---
 Eigen/src/Core/Assign.h |  2 +-
 Eigen/src/Core/Dot.h    | 16 ++++++++--------
 2 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h
index 53806ba33..655412efd 100644
--- a/Eigen/src/Core/Assign.h
+++ b/Eigen/src/Core/Assign.h
@@ -16,7 +16,7 @@ namespace Eigen {
 
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
   ::lazyAssign(const DenseBase<OtherDerived>& other)
 {
   enum{
diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h
index 06ef18b8b..bb8e3fecc 100644
--- a/Eigen/src/Core/Dot.h
+++ b/Eigen/src/Core/Dot.h
@@ -90,7 +90,7 @@ MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
   * \sa dot(), norm(), lpNorm()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const
 {
   return numext::real((*this).cwiseAbs2().sum());
 }
@@ -102,7 +102,7 @@ EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scala
   * \sa lpNorm(), dot(), squaredNorm()
   */
 template<typename Derived>
-inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
+EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
 {
   return numext::sqrt(squaredNorm());
 }
@@ -117,7 +117,7 @@ inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real Matr
   * \sa norm(), normalize()
   */
 template<typename Derived>
-inline const typename MatrixBase<Derived>::PlainObject
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::normalized() const
 {
   typedef typename internal::nested_eval<Derived,2>::type _Nested;
@@ -139,7 +139,7 @@ MatrixBase<Derived>::normalized() const
   * \sa norm(), normalized()
   */
 template<typename Derived>
-inline void MatrixBase<Derived>::normalize()
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::normalize()
 {
   RealScalar z = squaredNorm();
   // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
@@ -160,7 +160,7 @@ inline void MatrixBase<Derived>::normalize()
   * \sa stableNorm(), stableNormalize(), normalized()
   */
 template<typename Derived>
-inline const typename MatrixBase<Derived>::PlainObject
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::stableNormalized() const
 {
   typedef typename internal::nested_eval<Derived,3>::type _Nested;
@@ -185,7 +185,7 @@ MatrixBase<Derived>::stableNormalized() const
   * \sa stableNorm(), stableNormalized(), normalize()
   */
 template<typename Derived>
-inline void MatrixBase<Derived>::stableNormalize()
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::stableNormalize()
 {
   RealScalar w = cwiseAbs().maxCoeff();
   RealScalar z = (derived()/w).squaredNorm();
@@ -257,9 +257,9 @@ struct lpNorm_selector<Derived, Infinity>
 template<typename Derived>
 template<int p>
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 #else
-MatrixBase<Derived>::RealScalar
+EIGEN_DEVICE_FUNC MatrixBase<Derived>::RealScalar
 #endif
 MatrixBase<Derived>::lpNorm() const
 {

From c36bc2d445596d46c7f5a9271bfa69c79e2e1558 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 14:58:45 -0800
Subject: [PATCH 045/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/BooleanRedux.h    | 6 +++---
 Eigen/src/Core/DiagonalProduct.h | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h
index ed607d5d8..ccf519067 100644
--- a/Eigen/src/Core/BooleanRedux.h
+++ b/Eigen/src/Core/BooleanRedux.h
@@ -76,7 +76,7 @@ struct any_unroller<Derived, Dynamic, Rows>
   * \sa any(), Cwise::operator<()
   */
 template<typename Derived>
-inline bool DenseBase<Derived>::all() const
+EIGEN_DEVICE_FUNC inline bool DenseBase<Derived>::all() const
 {
   typedef internal::evaluator<Derived> Evaluator;
   enum {
@@ -100,7 +100,7 @@ inline bool DenseBase<Derived>::all() const
   * \sa all()
   */
 template<typename Derived>
-inline bool DenseBase<Derived>::any() const
+EIGEN_DEVICE_FUNC inline bool DenseBase<Derived>::any() const
 {
   typedef internal::evaluator<Derived> Evaluator;
   enum {
@@ -124,7 +124,7 @@ inline bool DenseBase<Derived>::any() const
   * \sa all(), any()
   */
 template<typename Derived>
-inline Eigen::Index DenseBase<Derived>::count() const
+EIGEN_DEVICE_FUNC inline Eigen::Index DenseBase<Derived>::count() const
 {
   return derived().template cast<bool>().template cast<Index>().sum();
 }
diff --git a/Eigen/src/Core/DiagonalProduct.h b/Eigen/src/Core/DiagonalProduct.h
index d372b938f..7911d1cd1 100644
--- a/Eigen/src/Core/DiagonalProduct.h
+++ b/Eigen/src/Core/DiagonalProduct.h
@@ -17,7 +17,7 @@ namespace Eigen {
   */
 template<typename Derived>
 template<typename DiagonalDerived>
-inline const Product<Derived, DiagonalDerived, LazyProduct>
+EIGEN_DEVICE_FUNC inline const Product<Derived, DiagonalDerived, LazyProduct>
 MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const
 {
   return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived());

From 857adbbd52bb1a36c913a828fb5f24b95deee965 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 16:42:00 -0800
Subject: [PATCH 046/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/Random.h    | 2 +-
 Eigen/src/Core/Replicate.h | 4 ++--
 Eigen/src/Core/Reverse.h   | 6 +++---
 3 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/Random.h b/Eigen/src/Core/Random.h
index 6faf789c7..486e9ed52 100644
--- a/Eigen/src/Core/Random.h
+++ b/Eigen/src/Core/Random.h
@@ -128,7 +128,7 @@ DenseBase<Derived>::Random()
   * \sa class CwiseNullaryOp, setRandom(Index), setRandom(Index,Index)
   */
 template<typename Derived>
-inline Derived& DenseBase<Derived>::setRandom()
+EIGEN_DEVICE_FUNC inline Derived& DenseBase<Derived>::setRandom()
 {
   return *this = Random(rows(), cols());
 }
diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h
index 9960ef884..0b2d6d743 100644
--- a/Eigen/src/Core/Replicate.h
+++ b/Eigen/src/Core/Replicate.h
@@ -115,7 +115,7 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
   */
 template<typename Derived>
 template<int RowFactor, int ColFactor>
-const Replicate<Derived,RowFactor,ColFactor>
+EIGEN_DEVICE_FUNC const Replicate<Derived,RowFactor,ColFactor>
 DenseBase<Derived>::replicate() const
 {
   return Replicate<Derived,RowFactor,ColFactor>(derived());
@@ -130,7 +130,7 @@ DenseBase<Derived>::replicate() const
   * \sa VectorwiseOp::replicate(), DenseBase::replicate(), class Replicate
   */
 template<typename ExpressionType, int Direction>
-const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
+EIGEN_DEVICE_FUNC const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
 VectorwiseOp<ExpressionType,Direction>::replicate(Index factor) const
 {
   return typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h
index 0640cda2a..8b6b3ab03 100644
--- a/Eigen/src/Core/Reverse.h
+++ b/Eigen/src/Core/Reverse.h
@@ -114,7 +114,7 @@ template<typename MatrixType, int Direction> class Reverse
   *
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::ReverseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ReverseReturnType
 DenseBase<Derived>::reverse()
 {
   return ReverseReturnType(derived());
@@ -136,7 +136,7 @@ DenseBase<Derived>::reverse()
   *
   * \sa VectorwiseOp::reverseInPlace(), reverse() */
 template<typename Derived>
-inline void DenseBase<Derived>::reverseInPlace()
+EIGEN_DEVICE_FUNC inline void DenseBase<Derived>::reverseInPlace()
 {
   if(cols()>rows())
   {
@@ -201,7 +201,7 @@ struct vectorwise_reverse_inplace_impl<Horizontal>
   *
   * \sa DenseBase::reverseInPlace(), reverse() */
 template<typename ExpressionType, int Direction>
-void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
+EIGEN_DEVICE_FUNC void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
 {
   internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
 }

From c92406d6137e4c55fc608c6916cc20899d00cff4 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 17:03:11 -0800
Subject: [PATCH 047/680] Silenced clang compilation warning.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
index e6cee11ef..be8d69386 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
@@ -217,7 +217,10 @@ struct GpuDevice {
     EIGEN_UNUSED_VARIABLE(err)
     assert(err == cudaSuccess);
 #else
-  eigen_assert(false && "The default device should be used instead to generate kernel code");
+    EIGEN_UNUSED_VARIABLE(dst);
+    EIGEN_UNUSED_VARIABLE(src);
+    EIGEN_UNUSED_VARIABLE(n);
+    eigen_assert(false && "The default device should be used instead to generate kernel code");
 #endif
   }
 

From 7b61944669f23a20f6c850b9c07d930c049c6ede Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 17:05:28 -0800
Subject: [PATCH 048/680] Made most of the packet math primitives usable within
 CUDA kernel when compiling with clang

---
 Eigen/src/Core/arch/CUDA/PacketMath.h | 4 ++--
 Eigen/src/Core/util/Macros.h          | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index 4dda63188..8c46af09b 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -167,10 +167,10 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu<double2>(const d
   return make_double2(from[0], from[1]);
 }
 
-template<> EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float*   from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float*   from) {
   return make_float4(from[0], from[0], from[1], from[1]);
 }
-template<> EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double*  from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double*  from) {
   return make_double2(from[0], from[0]);
 }
 
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 29c796647..14ec87da8 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -542,8 +542,8 @@
 //  - static is not very good because it prevents definitions from different object files to be merged.
 //           So static causes the resulting linked executable to be bloated with multiple copies of the same function.
 //  - inline is not perfect either as it unwantedly hints the compiler toward inlining the function.
-#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
-#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS inline
+#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_DEVICE_FUNC
+#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_DEVICE_FUNC inline
 
 #ifdef NDEBUG
 # ifndef EIGEN_NO_DEBUG

From 3a3f040baa602e13fe36d2949712d4af1b2db354 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 28 Feb 2017 17:06:15 -0800
Subject: [PATCH 049/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/VectorwiseOp.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h
index 4fe267e9f..893bc796f 100644
--- a/Eigen/src/Core/VectorwiseOp.h
+++ b/Eigen/src/Core/VectorwiseOp.h
@@ -670,7 +670,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
   * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::ColwiseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ColwiseReturnType
 DenseBase<Derived>::colwise()
 {
   return ColwiseReturnType(derived());
@@ -684,7 +684,7 @@ DenseBase<Derived>::colwise()
   * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::RowwiseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::RowwiseReturnType
 DenseBase<Derived>::rowwise()
 {
   return RowwiseReturnType(derived());

From c1d87ec110d9ee96da39d58c8f88481f5cb9d04c Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 1 Mar 2017 10:08:50 -0800
Subject: [PATCH 050/680] Added missing EIGEN_DEVICE_FUNC qualifiers

---
 Eigen/src/Core/CoreEvaluators.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 412f5a661..54276b836 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -139,14 +139,14 @@ public:
     EIGEN_ONLY_USED_FOR_DEBUG(outerStride);
     eigen_internal_assert(outerStride==OuterStride);
   }
-  Index outerStride() const { return OuterStride; }
+  EIGEN_DEVICE_FUNC Index outerStride() const { return OuterStride; }
   const Scalar *data;
 };
 
 template<typename Scalar> class plainobjectbase_evaluator_data<Scalar,Dynamic> {
 public:
   plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {}
-  Index outerStride() const { return m_outerStride; }
+  EIGEN_DEVICE_FUNC Index outerStride() const { return m_outerStride; }
   const Scalar *data;
 protected:
   Index m_outerStride;

From 1e2d046651102c57f5f4eca38ff7844e1b0ca6fd Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 1 Mar 2017 10:13:42 -0800
Subject: [PATCH 051/680] Silenced a couple of compilation warnings

---
 unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h | 1 +
 unsupported/Eigen/CXX11/src/util/EmulateArray.h    | 1 +
 2 files changed, 2 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h b/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
index 3523e7c94..d23f2e4c8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
@@ -23,6 +23,7 @@ struct static_val {
 
   template <typename T>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE static_val(const T& v) {
+    EIGEN_UNUSED_VARIABLE(v);
     eigen_assert(v == n);
   }
 };
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 03169d591..573ca435a 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -169,6 +169,7 @@ template <typename T> class array<T, 0> {
 
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   EIGEN_DEVICE_FUNC array(std::initializer_list<T> l) : dummy() {
+    EIGEN_UNUSED_VARIABLE(l);
     eigen_assert(l.size() == 0);
   }
 #endif

From 09ae0e6586b978ce1ea9960984e1228dfc8971b8 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 1 Mar 2017 11:47:47 -0800
Subject: [PATCH 052/680] Adjusted the EIGEN_DEVICE_FUNC qualifiers to make
 sure that:   * they're used consistently between the declaration and the
 definition of a function   * we avoid calling host only methods from host
 device methods.

---
 Eigen/src/Core/CoreEvaluators.h                   |  4 ++--
 Eigen/src/Core/DiagonalMatrix.h                   |  4 ++--
 Eigen/src/Core/GeneralProduct.h                   |  2 +-
 Eigen/src/Core/SelfAdjointView.h                  |  4 ++--
 Eigen/src/Core/SolveTriangular.h                  |  2 +-
 Eigen/src/Core/TriangularMatrix.h                 | 15 ++++++++-------
 .../Core/products/GeneralMatrixMatrixTriangular.h |  6 +++---
 Eigen/src/Core/products/SelfadjointProduct.h      |  2 +-
 Eigen/src/Core/products/SelfadjointRank2Update.h  |  2 +-
 Eigen/src/Core/util/IntegralConstant.h            |  4 ++--
 Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h     |  4 ++--
 Eigen/src/LU/FullPivLU.h                          |  2 --
 Eigen/src/SVD/SVDBase.h                           |  1 -
 13 files changed, 25 insertions(+), 27 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 54276b836..15b361b38 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -134,7 +134,7 @@ private:
 // this helper permits to completely eliminate m_outerStride if it is known at compiletime.
 template<typename Scalar,int OuterStride> class plainobjectbase_evaluator_data {
 public:
-  plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr)
+  EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr)
   {
     EIGEN_ONLY_USED_FOR_DEBUG(outerStride);
     eigen_internal_assert(outerStride==OuterStride);
@@ -145,7 +145,7 @@ public:
 
 template<typename Scalar> class plainobjectbase_evaluator_data<Scalar,Dynamic> {
 public:
-  plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {}
+  EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {}
   EIGEN_DEVICE_FUNC Index outerStride() const { return m_outerStride; }
   const Scalar *data;
 protected:
diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h
index ecfdce8ef..4e8297ee6 100644
--- a/Eigen/src/Core/DiagonalMatrix.h
+++ b/Eigen/src/Core/DiagonalMatrix.h
@@ -44,7 +44,7 @@ class DiagonalBase : public EigenBase<Derived>
 
     EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const { return derived(); }
-    
+
     EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
     EIGEN_DEVICE_FUNC
@@ -273,7 +273,7 @@ class DiagonalWrapper
   * \sa class DiagonalWrapper, class DiagonalMatrix, diagonal(), isDiagonal()
   **/
 template<typename Derived>
-inline const DiagonalWrapper<const Derived>
+EIGEN_DEVICE_FUNC inline const DiagonalWrapper<const Derived>
 MatrixBase<Derived>::asDiagonal() const
 {
   return DiagonalWrapper<const Derived>(derived());
diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 0f16cd8e3..b206b0a7a 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -428,7 +428,7 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 template<typename Derived>
 template<typename OtherDerived>
 const Product<Derived,OtherDerived,LazyProduct>
-MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
+EIGEN_DEVICE_FUNC MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
 {
   enum {
     ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h
index 504c98f0e..7e71fe3c0 100644
--- a/Eigen/src/Core/SelfAdjointView.h
+++ b/Eigen/src/Core/SelfAdjointView.h
@@ -322,7 +322,7 @@ public:
 /** This is the const version of MatrixBase::selfadjointView() */
 template<typename Derived>
 template<unsigned int UpLo>
-typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
+EIGEN_DEVICE_FUNC typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView() const
 {
   return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
@@ -339,7 +339,7 @@ MatrixBase<Derived>::selfadjointView() const
   */
 template<typename Derived>
 template<unsigned int UpLo>
-typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
+EIGEN_DEVICE_FUNC typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView()
 {
   return typename SelfAdjointViewReturnType<UpLo>::Type(derived());
diff --git a/Eigen/src/Core/SolveTriangular.h b/Eigen/src/Core/SolveTriangular.h
index 049890b25..a0011d4f9 100644
--- a/Eigen/src/Core/SolveTriangular.h
+++ b/Eigen/src/Core/SolveTriangular.h
@@ -164,7 +164,7 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
 #ifndef EIGEN_PARSED_BY_DOXYGEN
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
-void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
 {
   OtherDerived& other = _other.const_cast_derived();
   eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h
index 667ef09dc..ed80da36a 100644
--- a/Eigen/src/Core/TriangularMatrix.h
+++ b/Eigen/src/Core/TriangularMatrix.h
@@ -488,7 +488,6 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       * \sa TriangularView::solveInPlace()
       */
     template<int Side, typename Other>
-    EIGEN_DEVICE_FUNC
     inline const internal::triangular_solve_retval<Side,TriangularViewType, Other>
     solve(const MatrixBase<Other>& other) const;
 
@@ -554,7 +553,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
 // FIXME should we keep that possibility
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-inline TriangularView<MatrixType, Mode>&
+EIGEN_DEVICE_FUNC inline TriangularView<MatrixType, Mode>&
 TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const MatrixBase<OtherDerived>& other)
 {
   internal::call_assignment_no_alias(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
@@ -564,7 +563,7 @@ TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const MatrixBase<OtherDer
 // FIXME should we keep that possibility
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other)
 {
   internal::call_assignment_no_alias(derived(), other.template triangularView<Mode>());
 }
@@ -573,7 +572,7 @@ void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<Ot
 
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-inline TriangularView<MatrixType, Mode>&
+EIGEN_DEVICE_FUNC inline TriangularView<MatrixType, Mode>&
 TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const TriangularBase<OtherDerived>& other)
 {
   eigen_assert(Mode == int(OtherDerived::Mode));
@@ -583,7 +582,7 @@ TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const TriangularBase<Othe
 
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other)
 {
   eigen_assert(Mode == int(OtherDerived::Mode));
   internal::call_assignment_no_alias(derived(), other.derived());
@@ -598,7 +597,7 @@ void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBas
   * If the matrix is triangular, the opposite part is set to zero. */
 template<typename Derived>
 template<typename DenseDerived>
-void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
+EIGEN_DEVICE_FUNC void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
 {
   evalToLazy(other.derived());
 }
@@ -624,6 +623,7 @@ void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
   */
 template<typename Derived>
 template<unsigned int Mode>
+EIGEN_DEVICE_FUNC
 typename MatrixBase<Derived>::template TriangularViewReturnType<Mode>::Type
 MatrixBase<Derived>::triangularView()
 {
@@ -633,6 +633,7 @@ MatrixBase<Derived>::triangularView()
 /** This is the const version of MatrixBase::triangularView() */
 template<typename Derived>
 template<unsigned int Mode>
+EIGEN_DEVICE_FUNC
 typename MatrixBase<Derived>::template ConstTriangularViewReturnType<Mode>::Type
 MatrixBase<Derived>::triangularView() const
 {
@@ -930,7 +931,7 @@ struct triangular_assignment_loop<Kernel, Mode, Dynamic, SetOpposite>
   * If the matrix is triangular, the opposite part is set to zero. */
 template<typename Derived>
 template<typename DenseDerived>
-void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
+EIGEN_DEVICE_FUNC void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
 {
   other.derived().resize(this->rows(), this->cols());
   internal::call_triangular_assignment_loop<Derived::Mode,(Derived::Mode&SelfAdjoint)==0 /* SetOpposite */>(other.derived(), derived().nestedExpression());
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index 7122efa60..ad38bcf51 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -292,12 +292,12 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename ProductType>
-TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
+EIGEN_DEVICE_FUNC TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
   eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
-  
+
   general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
-  
+
   return derived();
 }
 
diff --git a/Eigen/src/Core/products/SelfadjointProduct.h b/Eigen/src/Core/products/SelfadjointProduct.h
index f038d686f..39c5b59ff 100644
--- a/Eigen/src/Core/products/SelfadjointProduct.h
+++ b/Eigen/src/Core/products/SelfadjointProduct.h
@@ -120,7 +120,7 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename DerivedU>
-SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+EIGEN_DEVICE_FUNC SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
 ::rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha)
 {
   selfadjoint_product_selector<MatrixType,DerivedU,UpLo>::run(_expression().const_cast_derived(), u.derived(), alpha);
diff --git a/Eigen/src/Core/products/SelfadjointRank2Update.h b/Eigen/src/Core/products/SelfadjointRank2Update.h
index 2ae364111..d395888e5 100644
--- a/Eigen/src/Core/products/SelfadjointRank2Update.h
+++ b/Eigen/src/Core/products/SelfadjointRank2Update.h
@@ -57,7 +57,7 @@ template<bool Cond, typename T> struct conj_expr_if
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename DerivedU, typename DerivedV>
-SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+EIGEN_DEVICE_FUNC SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
 ::rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha)
 {
   typedef internal::blas_traits<DerivedU> UBlasTraits;
diff --git a/Eigen/src/Core/util/IntegralConstant.h b/Eigen/src/Core/util/IntegralConstant.h
index ae41015bd..78a4705cd 100644
--- a/Eigen/src/Core/util/IntegralConstant.h
+++ b/Eigen/src/Core/util/IntegralConstant.h
@@ -151,9 +151,9 @@ struct get_fixed_value<variable_if_dynamic<T,N>,Default> {
   static const int value = N;
 };
 
-template<typename T> Index get_runtime_value(const T &x) { return x; }
+template<typename T> EIGEN_DEVICE_FUNC Index get_runtime_value(const T &x) { return x; }
 #if !EIGEN_HAS_CXX14
-template<int N> Index get_runtime_value(FixedInt<N> (*)()) { return N; }
+template<int N> EIGEN_DEVICE_FUNC Index get_runtime_value(FixedInt<N> (*)()) { return N; }
 #endif
 
 // Cleanup integer/FixedInt/VariableAndFixedInt/etc types:
diff --git a/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h b/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
index 4fec8af0a..dbbd4806a 100644
--- a/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
+++ b/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
@@ -85,7 +85,7 @@ MatrixBase<Derived>::eigenvalues() const
   * \sa SelfAdjointEigenSolver::eigenvalues(), MatrixBase::eigenvalues()
   */
 template<typename MatrixType, unsigned int UpLo> 
-inline typename SelfAdjointView<MatrixType, UpLo>::EigenvaluesReturnType
+EIGEN_DEVICE_FUNC inline typename SelfAdjointView<MatrixType, UpLo>::EigenvaluesReturnType
 SelfAdjointView<MatrixType, UpLo>::eigenvalues() const
 {
   typedef typename SelfAdjointView<MatrixType, UpLo>::PlainObject PlainObject;
@@ -149,7 +149,7 @@ MatrixBase<Derived>::operatorNorm() const
   * \sa eigenvalues(), MatrixBase::operatorNorm()
   */
 template<typename MatrixType, unsigned int UpLo>
-inline typename SelfAdjointView<MatrixType, UpLo>::RealScalar
+EIGEN_DEVICE_FUNC inline typename SelfAdjointView<MatrixType, UpLo>::RealScalar
 SelfAdjointView<MatrixType, UpLo>::operatorNorm() const
 {
   return eigenvalues().cwiseAbs().maxCoeff();
diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h
index 03b6af706..ec61086d5 100644
--- a/Eigen/src/LU/FullPivLU.h
+++ b/Eigen/src/LU/FullPivLU.h
@@ -411,11 +411,9 @@ template<typename _MatrixType> class FullPivLU
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
 
     template<bool Conjugate, typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl_transposed(const RhsType &rhs, DstType &dst) const;
     #endif
 
diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h
index cc90a3b75..429414797 100644
--- a/Eigen/src/SVD/SVDBase.h
+++ b/Eigen/src/SVD/SVDBase.h
@@ -212,7 +212,6 @@ public:
   
   #ifndef EIGEN_PARSED_BY_DOXYGEN
   template<typename RhsType, typename DstType>
-  EIGEN_DEVICE_FUNC
   void _solve_impl(const RhsType &rhs, DstType &dst) const;
   #endif
 

From a71943b9a432c8962f025b56313584f33111ace4 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 2 Mar 2017 10:47:29 -0800
Subject: [PATCH 053/680] Made the Tensor code compile with clang 3.9

---
 .../CXX11/src/Tensor/TensorContractionCuda.h  | 101 +++++++++---------
 .../CXX11/src/Tensor/TensorReductionCuda.h    |   1 -
 2 files changed, 48 insertions(+), 54 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
index d65dbb40f..c04b784a4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
@@ -529,7 +529,6 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
                        float2 rhs_shmem2[][8], const Index m_size,
                        const Index n_size, const Index k_size,
                        const Index base_m, const Index base_n) {
-  typedef float Scalar;
 
   // prefetch registers
   float4 lhs_pf0, rhs_pf0;
@@ -540,27 +539,27 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
   }
 
 
-#define prefetch_lhs(reg, row, col)                   \
-    if (!CHECK_LHS_BOUNDARY) {                        \
-      if (col < k_size) {                             \
-        reg =lhs.loadPacket<Unaligned>(row, col);     \
-      }                                               \
-    } else {                                          \
-      if (col < k_size) {                             \
-        if (row + 3 < m_size) {                       \
-          reg =lhs.loadPacket<Unaligned>(row, col);   \
-        } else if (row + 2 < m_size) {                \
-          reg.x =lhs(row + 0, col);                   \
-          reg.y =lhs(row + 1, col);                   \
-          reg.z =lhs(row + 2, col);                   \
-        } else if (row + 1 < m_size) {                \
-          reg.x =lhs(row + 0, col);                   \
-          reg.y =lhs(row + 1, col);                   \
-        } else if (row  < m_size) {                   \
-          reg.x =lhs(row + 0, col);                   \
-        }                                             \
-      }                                               \
-    }                                                 \
+#define prefetch_lhs(reg, row, col)                            \
+    if (!CHECK_LHS_BOUNDARY) {                                 \
+      if (col < k_size) {                                      \
+        reg =lhs.template loadPacket<Unaligned>(row, col);     \
+      }                                                        \
+    } else {                                                   \
+      if (col < k_size) {                                      \
+        if (row + 3 < m_size) {                                \
+          reg =lhs.template loadPacket<Unaligned>(row, col);   \
+        } else if (row + 2 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+          reg.z =lhs(row + 2, col);                            \
+        } else if (row + 1 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+        } else if (row  < m_size) {                            \
+          reg.x =lhs(row + 0, col);                            \
+        }                                                      \
+      }                                                        \
+    }                                                          \
 
 
   Index lhs_vert = base_m+threadIdx.x*4;
@@ -578,7 +577,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -593,7 +592,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     } else {
       if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
-          rhs_pf0 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
           rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
@@ -766,7 +765,6 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
                        float2 rhs_shmem2[][8], const Index m_size,
                        const Index n_size, const Index k_size,
                        const Index base_m, const Index base_n) {
-  typedef float Scalar;
 
   // prefetch registers
   float4 lhs_pf0, lhs_pf1, lhs_pf2, lhs_pf3;
@@ -790,37 +788,37 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 
      if (!CHECK_LHS_BOUNDARY) {
       if ((threadIdx.y/4+k+24) < k_size) {
-        lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-        lhs_pf3 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
       } else if ((threadIdx.y/4+k+16) < k_size) {
-        lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
       } else if ((threadIdx.y/4+k+8) < k_size) {
-        lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
       } else if ((threadIdx.y/4+k) < k_size) {
-        lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
       }
     } else {
       // just CHECK_LHS_BOUNDARY
       if (lhs_vert + 3 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
-          lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-          lhs_pf3 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
         } else if ((threadIdx.y/4+k+16) < k_size) {
-          lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
         } else if ((threadIdx.y/4+k+8) < k_size) {
-          lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         } else if ((threadIdx.y/4+k) < k_size) {
-          lhs_pf0 =lhs.loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         }
       } else if (lhs_vert + 2 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
@@ -909,8 +907,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-        rhs_pf1 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -932,8 +930,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       if (rhs_horiz1 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          rhs_pf1 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
         } else if (rhs_vert + 2 < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -954,7 +952,7 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       } else if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -1137,9 +1135,6 @@ EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
   typedef float2 LHS_MEM[64][32];
   typedef float2 RHS_MEM[128][8];
 
-  typedef float2 LHS_MEM16x16[32][16];
-  typedef float2 RHS_MEM16x16[64][8];
-
   const Index m_block_idx = blockIdx.x;
   const Index n_block_idx = blockIdx.y;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index 65638b6a8..edb0ab280 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -287,7 +287,6 @@ struct FullReductionLauncher<
     void>::type> {
   static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) {
     typedef typename Self::Index Index;
-    typedef typename Self::CoeffReturnType Scalar;
     const int block_size = 256;
     const int num_per_thread = 128;
     const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);

From bbe717fa2f9fe4828c845b27f4a7d4bb77fffdb2 Mon Sep 17 00:00:00 2001
From: Julian Kent <anonymous@invalid.net>
Date: Fri, 3 Mar 2017 12:58:51 +0100
Subject: [PATCH 054/680] Make scaling work with non-square matrices

---
 unsupported/Eigen/src/IterativeSolvers/Scaling.h | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/unsupported/Eigen/src/IterativeSolvers/Scaling.h b/unsupported/Eigen/src/IterativeSolvers/Scaling.h
index d113e6e90..9b3eb53e0 100644
--- a/unsupported/Eigen/src/IterativeSolvers/Scaling.h
+++ b/unsupported/Eigen/src/IterativeSolvers/Scaling.h
@@ -104,12 +104,18 @@ class IterScaling
         for (int i = 0; i < m; ++i) 
         {
           Dr(i) = std::sqrt(Dr(i));
+        }
+        for (int i = 0; i < n; ++i) 
+        {
           Dc(i) = std::sqrt(Dc(i));
         }
         // Save the scaling factors 
         for (int i = 0; i < m; ++i) 
         {
           m_left(i) /= Dr(i);
+        }
+        for (int i = 0; i < n; ++i) 
+        {
           m_right(i) /= Dc(i);
         }
         // Scale the rows and the columns of the matrix

From 1c03d43a5cc421da067f12fe02db9eedaa1d125d Mon Sep 17 00:00:00 2001
From: Ilya Biryukov <ibiryukov@google.com>
Date: Mon, 6 Mar 2017 12:01:12 +0100
Subject: [PATCH 055/680] Fixed compilation with cuda-clang

---
 Eigen/src/Core/MathFunctions.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 7a6b999af..5ec6c395e 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -512,7 +512,7 @@ namespace std_fallback {
 
 template<typename Scalar>
 struct expm1_impl {
-  static inline Scalar run(const Scalar& x)
+  EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
     #if EIGEN_HAS_CXX11_MATH
@@ -549,7 +549,7 @@ namespace std_fallback {
 
 template<typename Scalar>
 struct log1p_impl {
-  static inline Scalar run(const Scalar& x)
+  EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
     #if EIGEN_HAS_CXX11_MATH

From 659087b622e94f7e35a56b7ed2cb01b024c80a7b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 7 Mar 2017 10:02:34 +0100
Subject: [PATCH 056/680] bug #1400: fix stableNorm with
 EIGEN_DONT_ALIGN_STATICALLY

---
 Eigen/src/Core/StableNorm.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h
index d2fe1e199..be04ed44d 100644
--- a/Eigen/src/Core/StableNorm.h
+++ b/Eigen/src/Core/StableNorm.h
@@ -170,7 +170,8 @@ MatrixBase<Derived>::stableNorm() const
   enum {
     CanAlign = (   (int(DerivedCopyClean::Flags)&DirectAccessBit)
                 || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
-               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT) // ifwe cannot allocate on the stack, then let's not bother about this optimization
+               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
+                 && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
   };
   typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
                                                    typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper;

From d9677185250d780b0011202621c90530937049d2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 7 Mar 2017 10:16:39 +0100
Subject: [PATCH 057/680] do not include std header within extern C

---
 Eigen/src/misc/lapacke.h | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/misc/lapacke.h b/Eigen/src/misc/lapacke.h
index 8c7e79b03..dd6fd3b0a 100755
--- a/Eigen/src/misc/lapacke.h
+++ b/Eigen/src/misc/lapacke.h
@@ -43,10 +43,6 @@
 #include "lapacke_config.h"
 #endif
 
-#ifdef __cplusplus
-extern "C" {
-#endif /* __cplusplus */
-
 #include <stdlib.h>
 
 #ifndef lapack_int
@@ -81,7 +77,7 @@ extern "C" {
 #endif
 
 #ifndef lapack_complex_float_real
-#define lapack_complex_float_real(z)       (creal(z))
+#define lapack_complex_float_rea@l(z)       (creal(z))
 #endif
 
 #ifndef lapack_complex_float_imag
@@ -108,6 +104,11 @@ lapack_complex_double lapack_make_complex_double( double re, double im );
 
 #endif
 
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
 #ifndef LAPACKE_malloc
 #define LAPACKE_malloc( size ) malloc( size )
 #endif

From e958c2baac15ccf0cf7b7919ef729b118e43a6ed Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 7 Mar 2017 10:47:40 +0100
Subject: [PATCH 058/680] remove UTF8 symbols

---
 Eigen/src/Core/arch/CUDA/Half.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index db9878796..67518da9f 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -13,7 +13,7 @@
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted.
 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 // HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,

From 5694315fbb7a2abc157cad284852b5e3df2f9576 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 7 Mar 2017 10:53:47 +0100
Subject: [PATCH 059/680] remove UTF8 symbol

---
 Eigen/src/Core/DenseBase.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index 91a8511be..fd933eed4 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -296,7 +296,7 @@ template<typename Derived> class DenseBase
     EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& func);
 
-    /** \ínternal
+    /** \internal
       * Copies \a other into *this without evaluating other. \returns a reference to *this.
       * \deprecated */
     template<typename OtherDerived>

From e5156e4d253fcc8ac13b0131973de3f56e810ab5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 7 Mar 2017 11:25:58 +0100
Subject: [PATCH 060/680] fix typo

---
 Eigen/src/misc/lapacke.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/misc/lapacke.h b/Eigen/src/misc/lapacke.h
index dd6fd3b0a..3d8e24f5a 100755
--- a/Eigen/src/misc/lapacke.h
+++ b/Eigen/src/misc/lapacke.h
@@ -77,7 +77,7 @@
 #endif
 
 #ifndef lapack_complex_float_real
-#define lapack_complex_float_rea@l(z)       (creal(z))
+#define lapack_complex_float_real(z)       (creal(z))
 #endif
 
 #ifndef lapack_complex_float_imag

From f84963ed95ff277bf3abb2e2517b3017a25ccf3f Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 7 Mar 2017 14:27:10 +0000
Subject: [PATCH 061/680] Adding TensorIndexTuple and TensorTupleReduceOP
 backend (ArgMax/Min) for sycl; fixing the address space issue for const
 TensorMap; converting all discard_write to write due to data missmatch.

---
 .../Eigen/CXX11/src/Tensor/TensorArgMax.h     |  19 +-
 .../Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h | 146 ++++++++++
 .../CXX11/src/Tensor/TensorContractionSycl.h  |   6 +-
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |  10 +-
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 268 +++++++++++-------
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  |   5 +
 .../Eigen/CXX11/src/Tensor/TensorMeta.h       |   2 +
 .../CXX11/src/Tensor/TensorReductionSycl.h    |   4 +-
 .../Eigen/CXX11/src/Tensor/TensorSycl.h       |  12 +
 .../TensorSyclConvertToDeviceExpression.h     |  36 ++-
 .../src/Tensor/TensorSyclExprConstructor.h    |  77 +++--
 .../src/Tensor/TensorSyclExtractAccessor.h    |  48 ++--
 .../src/Tensor/TensorSyclExtractFunctors.h    |  66 +++--
 .../CXX11/src/Tensor/TensorSyclFunctors.h     |  16 +-
 .../CXX11/src/Tensor/TensorSyclLeafCount.h    |  34 ++-
 .../src/Tensor/TensorSyclPlaceHolderExpr.h    |  38 ++-
 .../Eigen/CXX11/src/Tensor/TensorSyclRun.h    |   3 +-
 unsupported/test/CMakeLists.txt               |   1 +
 unsupported/test/cxx11_tensor_argmax_sycl.cpp | 248 ++++++++++++++++
 19 files changed, 813 insertions(+), 226 deletions(-)
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
 create mode 100644 unsupported/test/cxx11_tensor_argmax_sycl.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
index d06f40cd8..e81001c6e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@@ -119,6 +119,12 @@ struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
+  // required by sycl
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const {
+    return m_impl;
+  }
+
+
  protected:
   TensorEvaluator<ArgType, Device> m_impl;
 };
@@ -222,7 +228,7 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_orig_impl(op.expression(), device),
         m_impl(op.expression().index_tuples().reduce(op.reduce_dims(), op.reduce_op()), device),
-        m_return_dim(op.return_dim()) {
+        m_return_dim(op.return_dim()), m_device(device) {
 
     gen_strides(m_orig_impl.dimensions(), m_strides);
     if (Layout == static_cast<int>(ColMajor)) {
@@ -252,7 +258,16 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
     return (m_return_dim < 0) ? v.first : (v.first % m_stride_mod) / m_stride_div;
   }
 
+  #ifndef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  #else // following functions are required by sycl
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TupleType* data() const { return m_impl.data(); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int return_dim() const {return m_return_dim;}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const StrideDims& strides() const {return m_strides;}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride_mod() const {return m_stride_mod;}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride_div() const {return m_stride_div;}
+  const Device& device() const{return m_device;}
+  #endif
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
   costPerCoeff(bool vectorized) const {
@@ -292,6 +307,8 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
   StrideDims m_strides;
   Index m_stride_mod;
   Index m_stride_div;
+ // required by sycl
+  const Device& m_device;
 };
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
new file mode 100644
index 000000000..90cbe004f
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
@@ -0,0 +1,146 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ *  TensorArgMaxSycl.h
+ * \brief:
+ *  TensorArgMaxSycl
+ *
+*****************************************************************/
+
+#ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_ARGMAX_SYCL_HPP
+#define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_ARGMAX_SYCL_HPP
+namespace Eigen {
+namespace internal {
+  template<typename Dims, typename XprType>
+  struct eval<TensorTupleReducerDeviceOp<Dims, XprType>, Eigen::Dense>
+  {
+    typedef const TensorTupleReducerDeviceOp<Dims, XprType>& type;
+  };
+
+  template<typename Dims, typename XprType>
+  struct nested<TensorTupleReducerDeviceOp<Dims, XprType>, 1,
+                typename eval<TensorTupleReducerDeviceOp<Dims, XprType> >::type>
+  {
+    typedef TensorTupleReducerDeviceOp<Dims, XprType> type;
+  };
+
+template<typename StrideDims, typename XprType>
+struct traits<TensorTupleReducerDeviceOp<StrideDims, XprType> > : public traits<XprType>
+{
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+};
+
+
+}// end namespace internal
+template<typename StrideDims, typename XprType>
+class TensorTupleReducerDeviceOp : public TensorBase<TensorTupleReducerDeviceOp<StrideDims, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorTupleReducerDeviceOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename Eigen::internal::nested<TensorTupleReducerDeviceOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorTupleReducerDeviceOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorTupleReducerDeviceOp>::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerDeviceOp(XprType expr,
+                                                              const int return_dim,
+                                                              const StrideDims& strides,
+                                                              const Index& stride_mod, const Index& stride_div)
+          :m_xpr(expr), m_return_dim(return_dim), m_strides(strides), m_stride_mod(stride_mod), m_stride_div(stride_div) {}
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type&
+  expression() const { return m_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  int return_dim() const { return m_return_dim; }
+
+  EIGEN_DEVICE_FUNC
+  const StrideDims& strides() const { return m_strides; }
+
+  EIGEN_DEVICE_FUNC
+  const Index& stride_mod() const { return m_stride_mod; }
+
+  EIGEN_DEVICE_FUNC
+  const Index& stride_div() const { return m_stride_div; }
+
+  protected:
+    typename Eigen::internal::remove_all<typename
+    XprType::Nested
+    >::type m_xpr;
+    const int m_return_dim;
+    const StrideDims& m_strides;
+    const Index m_stride_mod;
+    const Index m_stride_div;
+};
+
+
+// Eval as rvalue
+template<typename StrideDims, typename ArgType>
+struct TensorEvaluator<const TensorTupleReducerDeviceOp<StrideDims, ArgType>, SyclKernelDevice>
+{
+  typedef TensorTupleReducerDeviceOp<StrideDims, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Index Scalar;
+  typedef Index CoeffReturnType;
+  typedef typename XprType::CoeffReturnType TupleType;
+  typedef typename TensorEvaluator<ArgType, SyclKernelDevice>::Dimensions Dimensions;
+
+  enum {
+    IsAligned =  false,
+    PacketAccess = false,
+    BlockAccess = false,
+    Layout = TensorEvaluator<ArgType, SyclKernelDevice>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,  const SyclKernelDevice& device)
+      : m_impl(op.expression(), device), m_return_dim(op.return_dim()), m_strides(op.strides()), m_stride_mod(op.stride_mod()),
+      m_stride_div(op.stride_div()){}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_impl.dimensions();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    const TupleType v = m_impl.coeff(index);
+    return (m_return_dim < 0) ? v.first : (v.first % m_stride_mod) / m_stride_div;
+  }
+typedef typename MakeGlobalPointer<typename TensorEvaluator<ArgType , SyclKernelDevice>::CoeffReturnType >::Type ptr_Dev_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ptr_Dev_type  data() const { return const_cast<ptr_Dev_type>(m_impl.data()); }
+
+protected:
+ TensorEvaluator<ArgType , SyclKernelDevice> m_impl;
+ const int m_return_dim;
+ const StrideDims& m_strides;
+ const Index& m_stride_mod;
+ const Index& m_stride_div;
+};
+} // end namespace Eigen
+#endif //UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_ARGMAX_SYCL_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
index fcd7d4d00..5b4c3c5bd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
@@ -11,7 +11,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 /*****************************************************************
- * TensorSyclConvertToDeviceExpression.h
+ * TensorTensorContractionsycl.h
  *
  * \brief:
  *  TensorContractionsycl
@@ -389,9 +389,9 @@ template< typename Self, typename OutScalar, typename ContractT, typename LeftNo
       cl::sycl::range<2>(LocalThreadSizeM, LocalThreadSizeN)),
        KernelConstructor<HostExpr, OutScalar, LhsScalar, RhsScalar, LHSFunctorExpr, RHSFunctorExpr, LhsLocalAcc, RhsLocalAcc, OutAccessor, Index, ContractT, LeftNocontractT,
        RightNocontractT, lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, TileSizeDimM, TileSizeDimN, TileSizeDimK,
-       WorkLoadPerThreadM, WorkLoadPerThreadN, LocalThreadSizeM, LocalThreadSizeN, LoadPerThreadLhs, LoadPerThreadRhs, LHSTupleType, RHSTupleType, Eigen::DefaultDevice>(lhs_functors, rhs_functors,
+       WorkLoadPerThreadM, WorkLoadPerThreadN, LocalThreadSizeM, LocalThreadSizeN, LoadPerThreadLhs, LoadPerThreadRhs, LHSTupleType, RHSTupleType, Eigen::SyclKernelDevice>(lhs_functors, rhs_functors,
           localLhs, localRhs, out_res, out_offset, roundUpK, M, N, K, m_k_strides, m_left_contracting_strides, m_right_contracting_strides,m_i_strides, m_j_strides,
-          m_left_nocontract_strides,m_right_nocontract_strides, left_tuple_of_accessors, right_tuple_of_accessors, Eigen::DefaultDevice()));
+          m_left_nocontract_strides,m_right_nocontract_strides, left_tuple_of_accessors, right_tuple_of_accessors, Eigen::SyclKernelDevice()));
     });
     self.device().asynchronousExec();
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index 66ffd819f..5db16d559 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -45,7 +45,7 @@ EigenConvolutionKernel1D(internal::IndexMapper<Index, InputDims, 1, Eigen::inter
   void operator()(cl::sycl::nd_item<2> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
     auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
-    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
+    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
 
     auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
     auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -103,7 +103,7 @@ EigenConvolutionKernel2D(internal::IndexMapper<Index, InputDims, 2, Eigen::inter
   void operator()(cl::sycl::nd_item<3> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
     auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
-    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
+    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
 
     auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
     auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -173,7 +173,7 @@ EigenConvolutionKernel3D(internal::IndexMapper<Index, InputDims, 3, Eigen::inter
   void operator()(cl::sycl::nd_item<3> itemID) {
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
     auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
-    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
+    auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
 
     auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
     auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -339,8 +339,8 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
       // create input tuple of accessors
       InputTupleType tuple_of_accessors = Eigen::TensorSycl::internal::createTupleOfAccessors<InputEvaluator>(cgh, m_inputImpl);
 
-      typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> OutputAccessorType;
-      OutputAccessorType out_res= m_device. template get_sycl_accessor<cl::sycl::access::mode::discard_write>(cgh, data);
+      typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer> OutputAccessorType;
+      OutputAccessorType out_res= m_device. template get_sycl_accessor<cl::sycl::access::mode::write>(cgh, data);
       typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer> KernelAccessorType;
       KernelAccessorType kernel_acc= m_device. template get_sycl_accessor<cl::sycl::access::mode::read>(cgh, m_kernel);
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 964222a15..258218463 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -41,9 +41,8 @@ namespace Eigen {
     size_t m_i;
     size_t m_offset;
   };
-
+template<typename AccType>
   struct memsetkernelFunctor{
-   typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> AccType;
    AccType m_acc;
    const ptrdiff_t buff_offset;
    const size_t m_rng, m_c;
@@ -55,15 +54,19 @@ namespace Eigen {
 
   };
 
+  //get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
 EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device::get_devices()){
   auto devices = cl::sycl::device::get_devices();
   std::vector<cl::sycl::device>::iterator it =devices.begin();
   while(it!=devices.end()) {
-    /// get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU )
+    ///FIXME: Currently there is a bug in amd cpu OpenCL
     auto s=  (*it).template get_info<cl::sycl::info::device::vendor>();
     std::transform(s.begin(), s.end(), s.begin(), ::tolower);
     if((*it).is_cpu() && s.find("amd")!=std::string::npos && s.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
       it=devices.erase(it);
+      //FIXME: currently there is a bug in intel gpu driver regarding memory allignment issue.
+    }else if((*it).is_gpu() && s.find("intel")!=std::string::npos){
+      it=devices.erase(it);
     }
     else{
       ++it;
@@ -112,6 +115,154 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
 }))
 #endif
   {}
+//FIXME: currently we have to switch back to write as discard_write doesnot work in forloop
+template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
+      std::lock_guard<std::mutex> lock(mutex_);
+      auto host_acc= find_buffer(dst)->second. template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::host_buffer>();
+      ::memcpy(host_acc.get_pointer(), src, n);
+}
+
+template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  // Assuming that the dst is the start of the destination pointer
+auto it =find_buffer(src);
+auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
+offset/=sizeof(Index);
+size_t rng, GRange, tileSize;
+parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+  auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
+  m_queue.submit([&](cl::sycl::handler &cgh) {
+    auto src_acc= it->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+    auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
+    typedef decltype(src_acc) read_accessor;
+    typedef decltype(dst_acc) write_accessor;
+    cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
+  });
+  synchronize();
+
+}
+
+EIGEN_STRONG_INLINE void synchronize() const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  m_queue.wait_and_throw(); //pass
+}
+EIGEN_STRONG_INLINE void asynchronousExec() const {
+  ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
+  //sycl_queue().throw_asynchronous();// does not pass. Temporarily disabled
+  std::lock_guard<std::mutex> lock(mutex_);
+  m_queue.wait_and_throw(); //pass
+
+}
+
+template<typename Index>
+EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
+  tileSize =static_cast<Index>(m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
+  auto s=  m_queue.get_device().template get_info<cl::sycl::info::device::vendor>();
+  std::transform(s.begin(), s.end(), s.begin(), ::tolower);
+  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
+  }
+  rng = n;
+  if (rng==0) rng=static_cast<Index>(1);
+  GRange=rng;
+  if (tileSize>GRange) tileSize=GRange;
+  else if(GRange>tileSize){
+    Index xMode =  static_cast<Index>(GRange % tileSize);
+    if (xMode != 0) GRange += static_cast<Index>(tileSize - xMode);
+  }
+}
+
+/// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+template<typename Index>
+EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
+  Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
+  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
+  }
+  Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+  tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
+  rng1=dim1;
+  if (rng1==0 ) rng1=static_cast<Index>(1);
+  GRange1=rng1;
+  if (tileSize1>GRange1) tileSize1=GRange1;
+  else if(GRange1>tileSize1){
+    Index xMode =  static_cast<Index>(GRange1 % tileSize1);
+    if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
+  }
+  tileSize0 = static_cast<Index>(max_workgroup_Size/tileSize1);
+  rng0 = dim0;
+  if (rng0==0 ) rng0=static_cast<Index>(1);
+  GRange0=rng0;
+  if (tileSize0>GRange0) tileSize0=GRange0;
+  else if(GRange0>tileSize0){
+    Index xMode =  static_cast<Index>(GRange0 % tileSize0);
+    if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
+  }
+}
+
+
+
+/// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+template<typename Index>
+EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
+  Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
+  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
+  }
+  Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+  tileSize2 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/3)));
+  rng2=dim2;
+  if (rng2==0 ) rng1=static_cast<Index>(1);
+  GRange2=rng2;
+  if (tileSize2>GRange2) tileSize2=GRange2;
+  else if(GRange2>tileSize2){
+    Index xMode =  static_cast<Index>(GRange2 % tileSize2);
+    if (xMode != 0) GRange2 += static_cast<Index>(tileSize2 - xMode);
+  }
+  pow_of_2 = static_cast<Index>(std::log2(static_cast<Index>(max_workgroup_Size/tileSize2)));
+  tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
+  rng1=dim1;
+  if (rng1==0 ) rng1=static_cast<Index>(1);
+  GRange1=rng1;
+  if (tileSize1>GRange1) tileSize1=GRange1;
+  else if(GRange1>tileSize1){
+    Index xMode =  static_cast<Index>(GRange1 % tileSize1);
+    if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
+  }
+  tileSize0 = static_cast<Index>(max_workgroup_Size/(tileSize1*tileSize2));
+  rng0 = dim0;
+  if (rng0==0 ) rng0=static_cast<Index>(1);
+  GRange0=rng0;
+  if (tileSize0>GRange0) tileSize0=GRange0;
+  else if(GRange0>tileSize0){
+    Index xMode =  static_cast<Index>(GRange0 % tileSize0);
+    if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
+  }
+}
+
+
+EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return m_queue.get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
+//  return stream_->deviceProperties().multiProcessorCount;
+}
+EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+
+//  return stream_->deviceProperties().maxThreadsPerBlock;
+}
+EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  // OpenCL doesnot have such concept
+  return 2;//sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+//  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
+}
+EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return m_queue.get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
+//  return stream_->deviceProperties().sharedMemPerBlock;
+}
 
   /// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
   /// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
@@ -119,10 +270,10 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   /// use this pointer as a key in our buffer_map and we make sure that we dedicate only one buffer only for this pointer.
   /// The device pointer would be deleted by calling deallocate function.
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    std::lock_guard<std::mutex> lock(mutex_);
     auto buf = cl::sycl::buffer<uint8_t,1>(cl::sycl::range<1>(num_bytes));
     auto ptr =buf.get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>().get_pointer();
     buf.set_final_data(nullptr);
-    std::lock_guard<std::mutex> lock(mutex_);
     buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(static_cast<const uint8_t*>(ptr),buf));
     return static_cast<void*>(ptr);
   }
@@ -193,48 +344,13 @@ struct SyclDevice {
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
-    tileSize =static_cast<Index>(sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
-    auto s=  sycl_queue().get_device().template get_info<cl::sycl::info::device::vendor>();
-    std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-      tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
-    }
-    rng = n;
-    if (rng==0) rng=static_cast<Index>(1);
-    GRange=rng;
-    if (tileSize>GRange) tileSize=GRange;
-    else if(GRange>tileSize){
-      Index xMode =  static_cast<Index>(GRange % tileSize);
-      if (xMode != 0) GRange += static_cast<Index>(tileSize - xMode);
-    }
+  m_queue_stream->parallel_for_setup(n, tileSize, rng, GRange);
   }
 
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
-    Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-      max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
-    }
-    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
-    tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
-    rng1=dim1;
-    if (rng1==0 ) rng1=static_cast<Index>(1);
-    GRange1=rng1;
-    if (tileSize1>GRange1) tileSize1=GRange1;
-    else if(GRange1>tileSize1){
-      Index xMode =  static_cast<Index>(GRange1 % tileSize1);
-      if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
-    }
-    tileSize0 = static_cast<Index>(max_workgroup_Size/tileSize1);
-    rng0 = dim0;
-    if (rng0==0 ) rng0=static_cast<Index>(1);
-    GRange0=rng0;
-    if (tileSize0>GRange0) tileSize0=GRange0;
-    else if(GRange0>tileSize0){
-      Index xMode =  static_cast<Index>(GRange0 % tileSize0);
-      if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
-    }
+    m_queue_stream->parallel_for_setup(dim0, dim1, tileSize0, tileSize1, rng0, rng1, GRange0, GRange1);
   }
 
 
@@ -242,39 +358,8 @@ struct SyclDevice {
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
-    Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-      max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
-    }
-    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
-    tileSize2 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/3)));
-    rng2=dim2;
-    if (rng2==0 ) rng1=static_cast<Index>(1);
-    GRange2=rng2;
-    if (tileSize2>GRange2) tileSize2=GRange2;
-    else if(GRange2>tileSize2){
-      Index xMode =  static_cast<Index>(GRange2 % tileSize2);
-      if (xMode != 0) GRange2 += static_cast<Index>(tileSize2 - xMode);
-    }
-    pow_of_2 = static_cast<Index>(std::log2(static_cast<Index>(max_workgroup_Size/tileSize2)));
-    tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
-    rng1=dim1;
-    if (rng1==0 ) rng1=static_cast<Index>(1);
-    GRange1=rng1;
-    if (tileSize1>GRange1) tileSize1=GRange1;
-    else if(GRange1>tileSize1){
-      Index xMode =  static_cast<Index>(GRange1 % tileSize1);
-      if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
-    }
-    tileSize0 = static_cast<Index>(max_workgroup_Size/(tileSize1*tileSize2));
-    rng0 = dim0;
-    if (rng0==0 ) rng0=static_cast<Index>(1);
-    GRange0=rng0;
-    if (tileSize0>GRange0) tileSize0=GRange0;
-    else if(GRange0>tileSize0){
-      Index xMode =  static_cast<Index>(GRange0 % tileSize0);
-      if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
-    }
+    m_queue_stream->parallel_for_setup(dim0, dim1, dim2, tileSize0, tileSize1, tileSize2, rng0, rng1, rng2, GRange0, GRange1, GRange2);
+
   }
   /// allocate device memory
   EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
@@ -319,8 +404,7 @@ struct SyclDevice {
   /// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
   /// this buffer is accessed, the data will be copied to the device.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
-    auto host_acc= get_sycl_buffer(dst). template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>();
-    ::memcpy(host_acc.get_pointer(), src, n);
+     m_queue_stream->memcpyHostToDevice(dst,src,n);
   }
   /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
   /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
@@ -329,21 +413,7 @@ struct SyclDevice {
   /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
   /// to the cpu only once per function call.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
-    auto it = m_queue_stream->find_buffer(src);
-    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
-    offset/=sizeof(Index);
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
-    // Assuming that the dst is the start of the destination pointer
-    auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
-    sycl_queue().submit([&](cl::sycl::handler &cgh) {
-      auto src_acc= it->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
-      auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
-      typedef decltype(src_acc) read_accessor;
-      typedef decltype(dst_acc) write_accessor;
-      cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
-    });
-    synchronize();
+    m_queue_stream->memcpyDeviceToHost(dst,src,n);
   }
   /// returning the sycl queue
   EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->m_queue;}
@@ -366,8 +436,9 @@ struct SyclDevice {
     :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
 
     void operator()(cl::sycl::handler &cgh) const {
-      auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor(buf_acc, buff_offset, rng, c));
+      auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+      typedef decltype(buf_acc) AccType;
+      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor<AccType>(buf_acc, buff_offset, rng, c));
     }
   };
 
@@ -403,14 +474,13 @@ struct SyclDevice {
   EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
 
   EIGEN_STRONG_INLINE void synchronize() const {
-    sycl_queue().wait_and_throw(); //pass
+    m_queue_stream->synchronize(); //pass
   }
 
   EIGEN_STRONG_INLINE void asynchronousExec() const {
     ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
     //sycl_queue().throw_asynchronous();// does not pass. Temporarily disabled
-    sycl_queue().wait_and_throw(); //pass
-
+    m_queue_stream->asynchronousExec();
   }
   // This function checks if the runtime recorded an error for the
   // underlying stream device.
@@ -418,8 +488,10 @@ struct SyclDevice {
     return m_queue_stream->ok();
   }
 };
-
-
+// This is used as a distingushable device inside the kernel as the sycl device class is not Standard layout.
+// This is internal and must not be used by user. This dummy device allow us to specialise the tensor evaluator
+// inside the kenrel. So we can have two types of eval for host and device. This is required for TensorArgMax operation
+struct SyclKernelDevice:DefaultDevice{};
 
 }  // end namespace Eigen
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index d6415817b..8516b37b3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -193,7 +193,12 @@ struct TensorEvaluator<const Derived, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
     eigen_assert(m_data);
+#ifndef __SYCL_DEVICE_ONLY__
     return loadConstant(m_data+index);
+#else
+    CoeffReturnType tmp = m_data[index];
+    return tmp;
+#endif
   }
 
   template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index b5ef31d55..77c9c6c6e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -124,7 +124,9 @@ template <typename U, typename V> struct Tuple {
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   Tuple& operator= (const Tuple& rhs) {
+  #ifndef __SYCL_DEVICE_ONLY__
     if (&rhs == this) return *this;
+  #endif
     first = rhs.first;
     second = rhs.second;
     return *this;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index c9c7acfdc..94899252b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -35,7 +35,7 @@ static void run(OP op, BufferTOut& bufOut, ptrdiff_t out_offset, BufferTIn& bufI
             /* Two accessors are used: one to the buffer that is being reduced,
              * and a second to local memory, used to store intermediate data. */
             auto aI =bufI.template get_access<cl::sycl::access::mode::read_write>(h);
-            auto aOut =bufOut.template get_access<cl::sycl::access::mode::discard_write>(h);
+            auto aOut =bufOut.template get_access<cl::sycl::access::mode::write>(h);
             typedef decltype(aI) InputAccessor;
             typedef decltype(aOut) OutputAccessor;
             typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write,cl::sycl::access::target::local> LocalAccessor;
@@ -158,7 +158,7 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
       typedef decltype(TensorSycl::internal::createTupleOfAccessors(cgh, self.impl())) Tuple_of_Acc;
       // create a tuple of accessors from Evaluator
       Tuple_of_Acc tuple_of_accessors =  TensorSycl::internal::createTupleOfAccessors(cgh, self.impl());
-      auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write>(cgh, output);
+      auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::write>(cgh, output);
       ptrdiff_t out_offset = dev.get_offset(output);
       Index red_size = (num_values_to_reduce!=0)? num_values_to_reduce : static_cast<Index>(1);
       cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)),
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
index 9d5a6d4c1..3d6270614 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
@@ -32,6 +32,8 @@ struct MakeLocalPointer {
 
 
 namespace Eigen {
+  template<typename StrideDims, typename XprType> class TensorTupleReducerDeviceOp;
+  template<typename StrideDims, typename ArgType> struct TensorEvaluator<const TensorTupleReducerDeviceOp<StrideDims, ArgType>, SyclKernelDevice>;
 namespace TensorSycl {
 namespace internal {
 
@@ -48,6 +50,13 @@ template<typename T> struct GetType<false, T>{
   typedef T Type;
 };
 
+template <bool Conds,  size_t X , size_t Y > struct ValueCondition {
+  static const size_t Res =X;
+};
+template<size_t X, size_t Y> struct ValueCondition<false, X , Y> {
+  static const size_t Res =Y;
+};
+
 }
 }
 }
@@ -80,6 +89,9 @@ template<typename T> struct GetType<false, T>{
 /// this is used for extracting tensor reduction
 #include "TensorReductionSycl.h"
 
+// TensorArgMaxSycl.h
+#include "TensorArgMaxSycl.h"
+
 /// this is used for extracting tensor convolution
 #include "TensorConvolutionSycl.h"
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
index 9476c0ea8..d6ac7b91f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -103,7 +103,7 @@ KERNELBROKERCONVERT(, false, TensorEvalToOp)
 #undef KERNELBROKERCONVERT
 
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node types are TensorForcedEvalOp and TensorLayoutSwapOp
-#define KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(CVQual, ExprNode)\
+#define KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(CVQual, ExprNode)\
 template <typename Expr>\
 struct ConvertToDeviceExpression<CVQual ExprNode<Expr> > {\
   typedef CVQual ExprNode< typename ConvertToDeviceExpression<Expr>::Type> Type;\
@@ -111,15 +111,17 @@ struct ConvertToDeviceExpression<CVQual ExprNode<Expr> > {\
 
 
 // TensorForcedEvalOp
-KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(const,TensorForcedEvalOp)
-KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(,TensorForcedEvalOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(const,TensorForcedEvalOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(,TensorForcedEvalOp)
 
 // TensorLayoutSwapOp
-KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(const,TensorLayoutSwapOp)
-KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP(,TensorLayoutSwapOp)
-#undef KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAP
-
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(const,TensorLayoutSwapOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(,TensorLayoutSwapOp)
 
+//TensorIndexTupleOp
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(const,TensorIndexTupleOp)
+KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP(,TensorIndexTupleOp)
+#undef KERNELBROKERCONVERTFORCEDEVALLAYOUTSWAPINDEXTUPLEOP
 
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorReductionOp
 #define KERNELBROKERCONVERTREDUCTION(CVQual)\
@@ -132,6 +134,18 @@ KERNELBROKERCONVERTREDUCTION(const)
 KERNELBROKERCONVERTREDUCTION()
 #undef KERNELBROKERCONVERTREDUCTION
 
+/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorReductionOp
+#define KERNELBROKERCONVERTTUPLEREDUCTION(CVQual)\
+template <typename OP, typename Dim, typename subExpr>\
+struct ConvertToDeviceExpression<CVQual TensorTupleReducerOp<OP, Dim, subExpr> > {\
+  typedef CVQual TensorTupleReducerOp<OP, Dim, typename ConvertToDeviceExpression<subExpr>::Type> Type;\
+};
+
+KERNELBROKERCONVERTTUPLEREDUCTION(const)
+KERNELBROKERCONVERTTUPLEREDUCTION()
+#undef KERNELBROKERCONVERTTUPLEREDUCTION
+
+//TensorSlicingOp
 #define KERNELBROKERCONVERTSLICEOP(CVQual)\
 template<typename StartIndices, typename Sizes, typename XprType>\
 struct ConvertToDeviceExpression<CVQual TensorSlicingOp <StartIndices, Sizes, XprType> >{\
@@ -142,7 +156,7 @@ KERNELBROKERCONVERTSLICEOP(const)
 KERNELBROKERCONVERTSLICEOP()
 #undef KERNELBROKERCONVERTSLICEOP
 
-
+//TensorStridingSlicingOp
 #define KERNELBROKERCONVERTERSLICESTRIDEOP(CVQual)\
 template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>\
 struct ConvertToDeviceExpression<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >{\
@@ -153,7 +167,6 @@ KERNELBROKERCONVERTERSLICESTRIDEOP(const)
 KERNELBROKERCONVERTERSLICESTRIDEOP()
 #undef KERNELBROKERCONVERTERSLICESTRIDEOP
 
-
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorChippingOp
 #define KERNELBROKERCONVERTCHIPPINGOP(CVQual)\
 template <DenseIndex DimId, typename Expr>\
@@ -164,9 +177,6 @@ KERNELBROKERCONVERTCHIPPINGOP(const)
 KERNELBROKERCONVERTCHIPPINGOP()
 #undef KERNELBROKERCONVERTCHIPPINGOP
 
-
-
-
 /// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorImagePatchOp
 #define KERNELBROKERCONVERTIMAGEPATCHOP(CVQual)\
 template<DenseIndex Rows, DenseIndex Cols, typename XprType>\
@@ -188,8 +198,6 @@ KERNELBROKERCONVERTVOLUMEPATCHOP(const)
 KERNELBROKERCONVERTVOLUMEPATCHOP()
 #undef KERNELBROKERCONVERTVOLUMEPATCHOP
 
-
-
 }  // namespace internal
 }  // namespace TensorSycl
 }  // namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index af4eb5f13..24cc23f45 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -65,7 +65,6 @@ CVQual PlaceHolder<CVQual TensorMap<T, Options_, MakePointer_>, N>, Params...>{\
   : expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())){}\
 };
 
-
 TENSORMAP(const)
 TENSORMAP()
 #undef TENSORMAP
@@ -83,6 +82,7 @@ CVQual PlaceHolder<CVQual TensorMap<TensorFixedSize<Scalar_, Dimensions_, Option
   ExprConstructor(FuncDetector &, const utility::tuple::Tuple<Params...> &t)\
   : expr(DeviceFixedSizeTensor<Type,Dimensions_>::instantiate(utility::tuple::get<N>(t))){}\
 };
+
 TENSORMAPFIXEDSIZE(const)
 TENSORMAPFIXEDSIZE()
 #undef TENSORMAPFIXEDSIZE
@@ -189,9 +189,6 @@ struct ExprConstructor<CVQual TensorAssignOp<OrigLHSExpr, OrigRHSExpr>,  CVQual
  ASSIGN()
  #undef ASSIGN
 
-
-
-
  /// specialisation of the \ref ExprConstructor struct when the node type is
  /// const TensorAssignOp
  #define CONVERSIONEXPRCONST(CVQual)\
@@ -252,8 +249,6 @@ FORCEDEVAL(const)
 FORCEDEVAL()
 #undef FORCEDEVAL
 
-
-
 #define TENSORCUSTOMUNARYOP(CVQual)\
 template <typename CustomUnaryFunc, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
 struct ExprConstructor<CVQual TensorCustomUnaryOp<CustomUnaryFunc, OrigExpr>,\
@@ -274,13 +269,6 @@ TENSORCUSTOMUNARYOP(const)
 TENSORCUSTOMUNARYOP()
 #undef TENSORCUSTOMUNARYOP
 
-template <bool Conds,  size_t X , size_t Y > struct ValueCondition {
-  static const size_t Res =X;
-};
-template<size_t X, size_t Y> struct ValueCondition<false, X , Y> {
-  static const size_t Res =Y;
-};
-
 /// specialisation of the \ref ExprConstructor struct when the node type is TensorReductionOp
 #define SYCLREDUCTIONEXPR(CVQual)\
 template <typename OP, typename Dim, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
@@ -299,6 +287,35 @@ SYCLREDUCTIONEXPR(const)
 SYCLREDUCTIONEXPR()
 #undef SYCLREDUCTIONEXPR
 
+/// specialisation of the \ref ExprConstructor struct when the node type is TensorTupleReducerOp
+/// use reductionOp instead of the TensorTupleReducerOp in order to build the tensor map. Because the tensorMap is the output of Tensor ReductionOP.
+#define SYCLTUPLEREDUCTIONEXPR(CVQual)\
+template <typename OP, typename Dim, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
+struct ExprConstructor<CVQual TensorTupleReducerOp<OP, Dim, OrigExpr>,\
+CVQual PlaceHolder<CVQual TensorTupleReducerOp<OP, Dim, DevExpr>, N>, Params...> {\
+  static const auto NumRedDims= TensorReductionOp<OP, Dim, const TensorIndexTupleOp<OrigExpr> , MakeGlobalPointer>::NumDimensions;\
+  static const auto NumIndices= ValueCondition<NumRedDims==0, 1, NumRedDims>::Res;\
+static const int Layout =static_cast<int>(Eigen::internal::traits<TensorReductionOp<OP, Dim, const TensorIndexTupleOp<OrigExpr>, MakeGlobalPointer>>::Layout);\
+  typedef CVQual TensorMap<\
+                          Tensor<typename TensorIndexTupleOp<OrigExpr>::CoeffReturnType,NumIndices, Layout, typename TensorTupleReducerOp<OP, Dim, OrigExpr>::Index>,\
+                          Layout,\
+                          MakeGlobalPointer\
+                          > XprType;\
+  typedef typename TensorEvaluator<const TensorIndexTupleOp<OrigExpr> , SyclKernelDevice>::Dimensions InputDimensions;\
+  static const int NumDims = Eigen::internal::array_size<InputDimensions>::value;\
+  typedef array<Index, NumDims> StrideDims;\
+  typedef const TensorTupleReducerDeviceOp<StrideDims, XprType> Type;\
+  Type expr;\
+  template <typename FuncDetector>\
+  ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
+  :expr(Type(XprType(ConvertToActualTypeSycl(typename XprType::CoeffReturnType, utility::tuple::get<N>(t)), fd.dimensions()),\
+    fd.return_dim(), fd.strides(), fd.stride_mod(), fd.stride_div())) {\
+    }\
+};
+
+SYCLTUPLEREDUCTIONEXPR(const)
+SYCLTUPLEREDUCTIONEXPR()
+#undef SYCLTUPLEREDUCTIONEXPR
 
 /// specialisation of the \ref ExprConstructor struct when the node type is
 /// TensorContractionOp, TensorConvolutionOp TensorCustomBinaryOp
@@ -319,15 +336,18 @@ CVQual PlaceHolder<CVQual ExprNode<Indices, LhsXprType,  RhsXprType>, N>, Params
   :expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())) {}\
 };
 
+//TensorContractionOp
 SYCLCONTRACTCONVCUSBIOPS(const, TensorContractionOp)
 SYCLCONTRACTCONVCUSBIOPS(, TensorContractionOp)
+//TensorConvolutionOp
 SYCLCONTRACTCONVCUSBIOPS(const, TensorConvolutionOp)
 SYCLCONTRACTCONVCUSBIOPS(, TensorConvolutionOp)
+//TensorCustomBinaryOp
 SYCLCONTRACTCONVCUSBIOPS(const, TensorCustomBinaryOp)
 SYCLCONTRACTCONVCUSBIOPS(, TensorCustomBinaryOp)
 #undef SYCLCONTRACTCONVCUSBIOPS
 
-
+//TensorSlicingOp
 #define SYCLSLICEOPEXPR(CVQual)\
 template<typename StartIndices, typename Sizes, typename OrigXprType, typename XprType, typename... Params>\
 struct ExprConstructor<CVQual TensorSlicingOp <StartIndices, Sizes, OrigXprType> , CVQual TensorSlicingOp<StartIndices, Sizes, XprType>, Params... >{\
@@ -344,7 +364,7 @@ SYCLSLICEOPEXPR(const)
 SYCLSLICEOPEXPR()
 #undef SYCLSLICEOPEXPR
 
-
+//TensorStridingSlicingOp
 #define SYCLSLICESTRIDEOPEXPR(CVQual)\
 template<typename StartIndices, typename StopIndices, typename Strides, typename OrigXprType, typename XprType, typename... Params>\
 struct ExprConstructor<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, OrigXprType>, CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, Params... >{\
@@ -361,6 +381,7 @@ SYCLSLICESTRIDEOPEXPR(const)
 SYCLSLICESTRIDEOPEXPR()
 #undef SYCLSLICESTRIDEOPEXPR
 
+//TensorReshapingOp and TensorShufflingOp
 #define SYCLRESHAPEANDSHUFFLEOPEXPRCONST(OPEXPR, CVQual)\
 template<typename Param, typename OrigXprType, typename XprType, typename... Params>\
 struct ExprConstructor<CVQual OPEXPR <Param, OrigXprType> , CVQual OPEXPR <Param, XprType>, Params... >{\
@@ -373,13 +394,15 @@ struct ExprConstructor<CVQual OPEXPR <Param, OrigXprType> , CVQual OPEXPR <Param
   : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.param()) {}\
 };
 
+// TensorReshapingOp
 SYCLRESHAPEANDSHUFFLEOPEXPRCONST(TensorReshapingOp, const)
 SYCLRESHAPEANDSHUFFLEOPEXPRCONST(TensorReshapingOp, )
-
+// TensorShufflingOp
 SYCLRESHAPEANDSHUFFLEOPEXPRCONST(TensorShufflingOp, const)
 SYCLRESHAPEANDSHUFFLEOPEXPRCONST(TensorShufflingOp, )
 #undef SYCLRESHAPEANDSHUFFLEOPEXPRCONST
 
+//TensorPaddingOp
 #define SYCLPADDINGOPEXPRCONST(OPEXPR, CVQual)\
 template<typename Param, typename OrigXprType, typename XprType, typename... Params>\
 struct ExprConstructor<CVQual OPEXPR <Param, OrigXprType> , CVQual OPEXPR <Param, XprType>, Params... >{\
@@ -392,11 +415,11 @@ struct ExprConstructor<CVQual OPEXPR <Param, OrigXprType> , CVQual OPEXPR <Param
   : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.param() , funcD.scalar_param()) {}\
 };
 
+//TensorPaddingOp
 SYCLPADDINGOPEXPRCONST(TensorPaddingOp, const)
 SYCLPADDINGOPEXPRCONST(TensorPaddingOp, )
 #undef SYCLPADDINGOPEXPRCONST
 
-
 // TensorChippingOp
 #define SYCLTENSORCHIPPINGOPEXPR(CVQual)\
 template<DenseIndex DimId, typename OrigXprType, typename XprType, typename... Params>\
@@ -454,14 +477,12 @@ SYCLTENSORVOLUMEPATCHOPEXPR(const)
 SYCLTENSORVOLUMEPATCHOPEXPR()
 #undef SYCLTENSORVOLUMEPATCHOPEXPR
 
-
-
-// TensorLayoutSwapOp
-#define SYCLTENSORLAYOUTSWAPOPEXPR(CVQual)\
+// TensorLayoutSwapOp and TensorIndexTupleOp
+#define SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR(CVQual, ExprNode)\
 template<typename OrigXprType, typename XprType, typename... Params>\
-struct ExprConstructor<CVQual TensorLayoutSwapOp <OrigXprType> , CVQual TensorLayoutSwapOp<XprType>, Params... >{\
+struct ExprConstructor<CVQual ExprNode <OrigXprType> , CVQual ExprNode<XprType>, Params... >{\
   typedef ExprConstructor<OrigXprType, XprType, Params...> my_xpr_type;\
-  typedef CVQual TensorLayoutSwapOp<typename my_xpr_type::Type> Type;\
+  typedef CVQual ExprNode<typename my_xpr_type::Type> Type;\
   my_xpr_type xprExpr;\
   Type expr;\
   template <typename FuncDetector>\
@@ -469,10 +490,14 @@ struct ExprConstructor<CVQual TensorLayoutSwapOp <OrigXprType> , CVQual TensorLa
   : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr) {}\
 };
 
-SYCLTENSORLAYOUTSWAPOPEXPR(const)
-SYCLTENSORLAYOUTSWAPOPEXPR()
-#undef SYCLTENSORLAYOUTSWAPOPEXPR
+//TensorLayoutSwapOp
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR(const, TensorLayoutSwapOp)
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR(, TensorLayoutSwapOp)
+//TensorIndexTupleOp
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR(const, TensorIndexTupleOp)
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR(, TensorIndexTupleOp)
 
+#undef SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXPR
 
 /// template deduction for \ref ExprConstructor struct
 template <typename OrigExpr, typename IndexExpr, typename FuncD, typename... Params>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index 5a6a8f4c5..fb95af59e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -147,7 +147,7 @@ SYCLFORCEDEVALEXTACC(const)
 SYCLFORCEDEVALEXTACC()
 #undef SYCLFORCEDEVALEXTACC
 
-
+//TensorCustomUnaryOp
 #define SYCLCUSTOMUNARYOPEXTACC(CVQual)\
 template <typename CustomUnaryFunc, typename XprType, typename Dev >\
 struct ExtractAccessor<TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Dev> > {\
@@ -160,7 +160,7 @@ SYCLCUSTOMUNARYOPEXTACC(const)
 SYCLCUSTOMUNARYOPEXTACC()
 #undef SYCLCUSTOMUNARYOPEXTACC
 
-
+//TensorCustomBinaryOp
 #define SYCLCUSTOMBINARYOPEXTACC(CVQual)\
 template <typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType , typename Dev>\
 struct ExtractAccessor<TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Dev> > {\
@@ -172,9 +172,6 @@ SYCLCUSTOMBINARYOPEXTACC(const)
 SYCLCUSTOMBINARYOPEXTACC()
 #undef SYCLCUSTOMBIBARYOPEXTACC
 
-
-
-
 /// specialisation of the \ref ExtractAccessor struct when the node type is TensorEvalToOp
 #define SYCLEVALTOEXTACC(CVQual)\
 template <typename Expr, typename Dev>\
@@ -188,15 +185,19 @@ SYCLEVALTOEXTACC()
 #undef SYCLEVALTOEXTACC
 
 /// specialisation of the \ref ExtractAccessor struct when the node type is TensorReductionOp
-#define SYCLREDUCTIONEXTACC(CVQual)\
+#define SYCLREDUCTIONEXTACC(CVQual, ExprNode)\
 template <typename OP, typename Dim, typename Expr, typename Dev>\
-struct ExtractAccessor<TensorEvaluator<CVQual TensorReductionOp<OP, Dim, Expr>, Dev> > {\
-  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorReductionOp<OP, Dim, Expr>, Dev>& eval)\
+struct ExtractAccessor<TensorEvaluator<CVQual ExprNode<OP, Dim, Expr>, Dev> > {\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual ExprNode<OP, Dim, Expr>, Dev>& eval)\
   RETURN_CPP11(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval))\
 };
+// TensorReductionOp
+SYCLREDUCTIONEXTACC(const,TensorReductionOp)
+SYCLREDUCTIONEXTACC(,TensorReductionOp)
 
-SYCLREDUCTIONEXTACC(const)
-SYCLREDUCTIONEXTACC()
+// TensorTupleReducerOp
+SYCLREDUCTIONEXTACC(const,TensorTupleReducerOp)
+SYCLREDUCTIONEXTACC(,TensorTupleReducerOp)
 #undef SYCLREDUCTIONEXTACC
 
 /// specialisation of the \ref ExtractAccessor struct when the node type is TensorContractionOp and TensorConvolutionOp
@@ -206,14 +207,14 @@ template<typename Indices, typename LhsXprType, typename RhsXprType, typename De
   static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual ExprNode<Indices, LhsXprType, RhsXprType>, Dev>& eval)\
   RETURN_CPP11(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval))\
 };
-
+//TensorContractionOp
 SYCLCONTRACTIONCONVOLUTIONEXTACC(const,TensorContractionOp)
 SYCLCONTRACTIONCONVOLUTIONEXTACC(,TensorContractionOp)
+//TensorConvolutionOp
 SYCLCONTRACTIONCONVOLUTIONEXTACC(const,TensorConvolutionOp)
 SYCLCONTRACTIONCONVOLUTIONEXTACC(,TensorConvolutionOp)
 #undef SYCLCONTRACTIONCONVOLUTIONEXTACC
 
-
 /// specialisation of the \ref ExtractAccessor struct when the node type is
 /// const TensorSlicingOp.
 #define SYCLSLICEOPEXTACC(CVQual)\
@@ -252,7 +253,6 @@ SYCLTENSORCHIPPINGOPEXTACC(const)
 SYCLTENSORCHIPPINGOPEXTACC()
 #undef SYCLTENSORCHIPPINGOPEXTACC
 
-
 // specialisation of the \ref ExtractAccessor struct when the node type is
 /// TensorImagePatchOp.
 #define SYCLTENSORIMAGEPATCHOPEXTACC(CVQual)\
@@ -266,8 +266,6 @@ SYCLTENSORIMAGEPATCHOPEXTACC(const)
 SYCLTENSORIMAGEPATCHOPEXTACC()
 #undef SYCLTENSORIMAGEPATCHOPEXTACC
 
-
-
 // specialisation of the \ref ExtractAccessor struct when the node type is
 /// TensorVolumePatchOp.
 #define SYCLTENSORVOLUMEPATCHOPEXTACC(CVQual)\
@@ -281,21 +279,23 @@ SYCLTENSORVOLUMEPATCHOPEXTACC(const)
 SYCLTENSORVOLUMEPATCHOPEXTACC()
 #undef SYCLTENSORVOLUMEPATCHOPEXTACC
 
-
 // specialisation of the \ref ExtractAccessor struct when the node type is
-/// TensorLayoutSwapOp.
-#define SYCLTENSORLAYOUTSWAPOPEXTACC(CVQual)\
+/// TensorLayoutSwapOp, TensorIndexTupleOp
+#define SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC(CVQual, ExprNode)\
 template<typename XprType, typename Dev>\
-struct ExtractAccessor<TensorEvaluator<CVQual TensorLayoutSwapOp<XprType>, Dev> >{\
-  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorLayoutSwapOp<XprType>, Dev>& eval)\
+struct ExtractAccessor<TensorEvaluator<CVQual ExprNode<XprType>, Dev> >{\
+  static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual ExprNode<XprType>, Dev>& eval)\
   RETURN_CPP11(AccessorConstructor::getTuple(cgh, eval.impl()))\
 };
 
-SYCLTENSORLAYOUTSWAPOPEXTACC(const)
-SYCLTENSORLAYOUTSWAPOPEXTACC()
-#undef SYCLTENSORLAYOUTSWAPOPEXTACC
-
+// TensorLayoutSwapOp
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC(const,TensorLayoutSwapOp)
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC(,TensorLayoutSwapOp)
+//TensorIndexTupleOp
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC(const,TensorIndexTupleOp)
+SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC(,TensorIndexTupleOp)
 
+#undef SYCLTENSORLAYOUTSWAPINDEXTUPLEOPEXTACC
 
 /// template deduction for \ref ExtractAccessor
 template <typename Evaluator>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index 9fcac5ecb..942e9d307 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -126,19 +126,19 @@ struct FunctorExtractor<TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFu
   typedef TensorEvaluator<CVQual TensorCustomUnaryOp<CustomUnaryFunc, ArgType>, Dev> Evaluator;\
   DEFALTACTION(Evaluator)\
 };
-
+//TensorCustomUnaryOp
 SYCLEXTRFUNCCUSTOMUNARYOP(const)
 SYCLEXTRFUNCCUSTOMUNARYOP()
 #undef SYCLEXTRFUNCCUSTOMUNARYOP
 
-
+//TensorCustomBinaryOp
 #define SYCLEXTRFUNCCUSTOMBIBARYOP(CVQual)\
 template <typename CustomBinaryFunc, typename ArgType1, typename ArgType2, typename Dev >\
 struct FunctorExtractor<TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, ArgType1, ArgType2>, Dev> > {\
   typedef TensorEvaluator<CVQual TensorCustomBinaryOp<CustomBinaryFunc, ArgType1, ArgType2>, Dev> Evaluator;\
   DEFALTACTION(Evaluator)\
 };
-
+//TensorCustomBinaryOp
 SYCLEXTRFUNCCUSTOMBIBARYOP(const)
 SYCLEXTRFUNCCUSTOMBIBARYOP()
 #undef SYCLEXTRFUNCCUSTOMBIBARYOP
@@ -177,7 +177,7 @@ SYCLEXTRFUNCASSIGNOP()
 
 /// specialisation of the \ref FunctorExtractor struct when the node types are
 /// TensorEvalToOp, TensorLayoutSwapOp. This is an specialisation without OP so it has to be separated.
-#define SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(CVQual, ExprNode)\
+#define SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(CVQual, ExprNode)\
 template <typename Expr, typename Dev>\
 struct FunctorExtractor<TensorEvaluator<CVQual ExprNode<Expr>, Dev> > {\
   FunctorExtractor<TensorEvaluator<Expr, Dev> > xprExpr;\
@@ -185,13 +185,16 @@ struct FunctorExtractor<TensorEvaluator<CVQual ExprNode<Expr>, Dev> > {\
   : xprExpr(expr.impl()) {}\
 };
 //TensorEvalToOp
-SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(const, TensorEvalToOp)
-SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(, TensorEvalToOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(const, TensorEvalToOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(, TensorEvalToOp)
 // TensorLayoutSwapOp
-SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(const, TensorLayoutSwapOp)
-SYCLEXTRFUNCEVALTOOPSWAPLAYOUT(, TensorLayoutSwapOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(const, TensorLayoutSwapOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(, TensorLayoutSwapOp)
+// TensorIndexTupleOp
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(const, TensorIndexTupleOp)
+SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE(, TensorIndexTupleOp)
 
-#undef SYCLEXTRFUNCEVALTOOPSWAPLAYOUT
+#undef SYCLEXTRFUNCEVALTOOPSWAPLAYOUTINDEXTUPLE
 
 template<typename Dim, size_t NumOutputDim> struct DimConstr {
 template<typename InDim>
@@ -202,10 +205,10 @@ template<typename Dim> struct DimConstr<Dim, 0> {
   template<typename InDim>
     static EIGEN_STRONG_INLINE Dim getDim(InDim dims ) {return Dim(static_cast<Dim>(dims.TotalSize()));}
 };
-
+//TensorReductionOp
 #define SYCLEXTRFUNCREDUCTIONOP(CVQual)\
 template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>\
-struct FunctorExtractor<TensorEvaluator<CVQual TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>>{\
+struct FunctorExtractor<TensorEvaluator<CVQual TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> >{\
   typedef TensorEvaluator<CVQual TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> Evaluator;\
   typedef typename Eigen::internal::conditional<Evaluator::NumOutputDims==0, DSizes<typename Evaluator::Index, 1>, typename Evaluator::Dimensions >::type Dimensions;\
   const Dimensions m_dimensions;\
@@ -213,12 +216,39 @@ struct FunctorExtractor<TensorEvaluator<CVQual TensorReductionOp<Op, Dims, ArgTy
   FunctorExtractor(const TensorEvaluator<CVQual TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>& expr)\
   : m_dimensions(DimConstr<Dimensions, Evaluator::NumOutputDims>::getDim(expr.dimensions())) {}\
 };
-
-
 SYCLEXTRFUNCREDUCTIONOP(const)
 SYCLEXTRFUNCREDUCTIONOP()
 #undef SYCLEXTRFUNCREDUCTIONOP
 
+//TensorTupleReducerOp
+#define SYCLEXTRFUNCTUPLEREDUCTIONOP(CVQual)\
+template<typename ReduceOp, typename Dims, typename ArgType, typename Device>\
+ struct FunctorExtractor<TensorEvaluator<CVQual TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device> >{\
+ typedef TensorEvaluator<CVQual TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device> Evaluator;\
+ static const int  NumOutputDims= Eigen::internal::traits<TensorTupleReducerOp<ReduceOp, Dims, ArgType> >::NumDimensions;\
+ typedef typename Evaluator::StrideDims StrideDims;\
+ typedef typename Evaluator::Index Index;\
+ typedef typename Eigen::internal::conditional<NumOutputDims==0, DSizes<Index, 1>, typename Evaluator::Dimensions >::type Dimensions;\
+ const Dimensions m_dimensions;\
+ const int m_return_dim;\
+ const StrideDims m_strides;\
+ const Index m_stride_mod;\
+ const Index m_stride_div;\
+ EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }\
+ EIGEN_STRONG_INLINE  int return_dim() const {return m_return_dim;}\
+ EIGEN_STRONG_INLINE const StrideDims& strides() const {return m_strides;}\
+ EIGEN_STRONG_INLINE const Index& stride_mod() const {return m_stride_mod;}\
+ EIGEN_STRONG_INLINE const Index& stride_div() const {return m_stride_div;}\
+ FunctorExtractor(const TensorEvaluator<CVQual TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device>& expr)\
+ : m_dimensions(DimConstr<Dimensions, NumOutputDims>::getDim(expr.dimensions())), m_return_dim(expr.return_dim()),\
+   m_strides(expr.strides()), m_stride_mod(expr.stride_mod()), m_stride_div(expr.stride_div()){}\
+};
+
+SYCLEXTRFUNCTUPLEREDUCTIONOP(const)
+SYCLEXTRFUNCTUPLEREDUCTIONOP()
+#undef SYCLEXTRFUNCTUPLEREDUCTIONOP
+
+//TensorContractionOp and TensorConvolutionOp
 #define SYCLEXTRFUNCCONTRACTCONVOLUTIONOP(CVQual, ExprNode)\
 template<typename Indices, typename LhsXprType, typename RhsXprType, typename Device>\
 struct FunctorExtractor<TensorEvaluator<CVQual ExprNode<Indices, LhsXprType, RhsXprType>, Device>>{\
@@ -230,9 +260,10 @@ struct FunctorExtractor<TensorEvaluator<CVQual ExprNode<Indices, LhsXprType, Rhs
   : m_dimensions(expr.dimensions()) {}\
 };
 
-
+//TensorContractionOp
 SYCLEXTRFUNCCONTRACTCONVOLUTIONOP(const,TensorContractionOp)
 SYCLEXTRFUNCCONTRACTCONVOLUTIONOP(,TensorContractionOp)
+//TensorConvolutionOp
 SYCLEXTRFUNCCONTRACTCONVOLUTIONOP(const,TensorConvolutionOp)
 SYCLEXTRFUNCCONTRACTCONVOLUTIONOP(,TensorConvolutionOp)
 #undef SYCLEXTRFUNCCONTRACTCONVOLUTIONOP
@@ -255,6 +286,7 @@ SYCLEXTRFUNCTSLICEOP(const)
 SYCLEXTRFUNCTSLICEOP()
 #undef SYCLEXTRFUNCTSLICEOP
 
+//TensorStridingSlicingOp
 #define SYCLEXTRFUNCTSLICESTRIDEOP(CVQual)\
 template<typename StartIndices, typename StopIndices, typename Strides, typename XprType, typename Dev>\
 struct FunctorExtractor<TensorEvaluator<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, Dev> >{\
@@ -273,7 +305,7 @@ SYCLEXTRFUNCTSLICESTRIDEOP(const)
 SYCLEXTRFUNCTSLICESTRIDEOP()
 #undef SYCLEXTRFUNCTSLICESTRIDEOP
 
-// Had to separate reshapeOP otherwise it will be mistaken by UnaryCategory
+// Had to separate TensorReshapingOp and TensorShufflingOp. Otherwise it will be mistaken by UnaryCategory
 #define SYCLRESHAPEANDSHUFFLEOPFUNCEXT(OPEXPR, FUNCCALL, CVQual)\
 template<typename Param, typename XprType, typename Dev>\
 struct FunctorExtractor<Eigen::TensorEvaluator<CVQual Eigen::OPEXPR<Param, XprType>, Dev> > {\
@@ -284,9 +316,11 @@ struct FunctorExtractor<Eigen::TensorEvaluator<CVQual Eigen::OPEXPR<Param, XprTy
   : xprExpr(expr.impl()), m_param(expr.FUNCCALL) {}\
 };
 
+//TensorReshapingOp
 SYCLRESHAPEANDSHUFFLEOPFUNCEXT(TensorReshapingOp, dimensions(), const)
 SYCLRESHAPEANDSHUFFLEOPFUNCEXT(TensorReshapingOp, dimensions(), )
 
+//TensorShufflingOp
 SYCLRESHAPEANDSHUFFLEOPFUNCEXT(TensorShufflingOp, shufflePermutation(), const)
 SYCLRESHAPEANDSHUFFLEOPFUNCEXT(TensorShufflingOp, shufflePermutation(), )
 #undef SYCLRESHAPEANDSHUFFLEOPFUNCEXT
@@ -343,6 +377,7 @@ SYCLEXTRFUNCCHIPPINGOP(const)
 SYCLEXTRFUNCCHIPPINGOP()
 #undef SYCLEXTRFUNCCHIPPINGOP
 
+//TensorImagePatchOp
 #define SYCLEXTRFUNCIMAGEPATCHOP(CVQual)\
 template<DenseIndex Rows, DenseIndex Cols, typename XprType, typename Device>\
 struct FunctorExtractor<TensorEvaluator<CVQual TensorImagePatchOp<Rows, Cols, XprType>, Device> >{\
@@ -420,7 +455,6 @@ SYCLEXTRFUNCVOLUMEPATCHOP()
 #undef SYCLEXTRFUNCVOLUMEPATCHOP
 
 
-
 /// template deduction function for FunctorExtractor
 template <typename Evaluator>
 auto inline extractFunctors(const Evaluator& evaluator)-> FunctorExtractor<Evaluator> {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
index 12237bfab..e5b892f2e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
@@ -72,7 +72,7 @@ namespace internal {
 template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typename Dims, typename Op, typename Index> class ReductionFunctor {
  public:
   typedef  typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr;
-  typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> write_accessor;
+  typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer> write_accessor;
   ReductionFunctor(write_accessor output_accessor_, ptrdiff_t out_offset_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_, Op functor_, Index range_, Index)
   :output_accessor(output_accessor_), out_offset(out_offset_), functors(functors_), tuple_of_accessors(tuple_of_accessors_), dims(dims_), functor(functor_), range(range_) {}
   void operator()(cl::sycl::nd_item<1> itemID) {
@@ -85,8 +85,8 @@ template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typen
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
     /// the device_evaluator is detectable and recognisable on the device.
-    typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice> DeviceSelf;
-    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
+    typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice> DeviceSelf;
+    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice>(device_self_expr, Eigen::SyclKernelDevice());
     auto output_accessor_ptr =ConvertToActualTypeSycl(typename DeviceSelf::CoeffReturnType, output_accessor);
     /// const cast added as a naive solution to solve the qualifier drop error
     auto globalid=static_cast<Index>(itemID.get_global_linear_id());
@@ -111,7 +111,7 @@ template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typen
 class ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Eigen::internal::MeanReducer<typename HostExpr::CoeffReturnType>, Index> {
  public:
   typedef  typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr;
-  typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> write_accessor;
+  typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer> write_accessor;
   typedef Eigen::internal::SumReducer<typename HostExpr::CoeffReturnType> Op;
   ReductionFunctor(write_accessor output_accessor_, ptrdiff_t out_offset_, FunctorExpr functors_, Tuple_of_Acc tuple_of_accessors_,Dims dims_,
     Eigen::internal::MeanReducer<typename HostExpr::CoeffReturnType>,  Index range_, Index num_values_to_reduce_)
@@ -126,8 +126,8 @@ class ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Eigen::interna
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
     /// the device_evaluator is detectable and recognisable on the device.
-    typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice> DeviceSelf;
-    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
+    typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice> DeviceSelf;
+    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice>(device_self_expr, Eigen::SyclKernelDevice());
     auto output_accessor_ptr =ConvertToActualTypeSycl(typename DeviceSelf::CoeffReturnType, output_accessor);
     /// const cast added as a naive solution to solve the qualifier drop error
     auto globalid=static_cast<Index>(itemID.get_global_linear_id());
@@ -173,7 +173,7 @@ public:
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, op);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
     /// the device_evaluator is detectable and recognisable on the device.
-    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
+    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice>(device_self_expr, Eigen::SyclKernelDevice());
     /// const cast added as a naive solution to solve the qualifier drop error
     auto globalid=itemID.get_global_linear_id();
 
@@ -220,7 +220,7 @@ public:
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, op);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
     /// the device_evaluator is detectable and recognisable on the device.
-    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
+    auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::SyclKernelDevice>(device_self_expr, Eigen::SyclKernelDevice());
     /// const cast added as a naive solution to solve the qualifier drop error
     auto globalid=itemID.get_global_linear_id();
     auto scale = (rng*red_factor) + remaining;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
index 330283b39..234580c7c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -114,27 +114,37 @@ SYCLCUSTOMBINARYOPLEAFCOUNT()
 #undef SYCLCUSTOMBINARYOPLEAFCOUNT
 
 /// specialisation of the \ref LeafCount struct when the node type is TensorEvalToOp
-#define EVALTOLAYOUTSWAPLEAFCOUNT(CVQual , ExprNode, Num)\
+#define EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(CVQual , ExprNode, Num)\
 template <typename Expr>\
 struct LeafCount<CVQual ExprNode<Expr> > {\
   static const size_t Count = Num + CategoryCount<Expr>::Count;\
 };
 
-EVALTOLAYOUTSWAPLEAFCOUNT(const, TensorEvalToOp, 1)
-EVALTOLAYOUTSWAPLEAFCOUNT(, TensorEvalToOp, 1)
-EVALTOLAYOUTSWAPLEAFCOUNT(const, TensorLayoutSwapOp, 0)
-EVALTOLAYOUTSWAPLEAFCOUNT(, TensorLayoutSwapOp, 0)
-#undef EVALTOLAYOUTSWAPLEAFCOUNT
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(const, TensorEvalToOp, 1)
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(, TensorEvalToOp, 1)
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(const, TensorLayoutSwapOp, 0)
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(, TensorLayoutSwapOp, 0)
+
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(const, TensorIndexTupleOp, 0)
+EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT(, TensorIndexTupleOp, 0)
+
+#undef EVALTOLAYOUTSWAPINDEXTUPLELEAFCOUNT
 
 /// specialisation of the \ref LeafCount struct when the node type is const TensorReductionOp
-#define REDUCTIONLEAFCOUNT(CVQual)\
+#define REDUCTIONLEAFCOUNT(CVQual, ExprNode)\
 template <typename OP, typename Dim, typename Expr>\
-struct LeafCount<CVQual TensorReductionOp<OP, Dim, Expr> > {\
+struct LeafCount<CVQual ExprNode<OP, Dim, Expr> > {\
     static const size_t Count =1;\
 };
 
-REDUCTIONLEAFCOUNT(const)
-REDUCTIONLEAFCOUNT()
+// TensorReductionOp
+REDUCTIONLEAFCOUNT(const,TensorReductionOp)
+REDUCTIONLEAFCOUNT(,TensorReductionOp)
+
+// tensor Argmax -TensorTupleReducerOp
+REDUCTIONLEAFCOUNT(const, TensorTupleReducerOp)
+REDUCTIONLEAFCOUNT(, TensorTupleReducerOp)
+
 #undef REDUCTIONLEAFCOUNT
 
 /// specialisation of the \ref LeafCount struct when the node type is const TensorContractionOp
@@ -150,8 +160,6 @@ CONTRACTIONCONVOLUTIONLEAFCOUNT(const,TensorConvolutionOp)
 CONTRACTIONCONVOLUTIONLEAFCOUNT(,TensorConvolutionOp)
 #undef CONTRACTIONCONVOLUTIONLEAFCOUNT
 
-
-
 /// specialisation of the \ref LeafCount struct when the node type is  TensorSlicingOp
 #define SLICEOPLEAFCOUNT(CVQual)\
 template <typename StartIndices, typename Sizes, typename XprType>\
@@ -161,7 +169,6 @@ SLICEOPLEAFCOUNT(const)
 SLICEOPLEAFCOUNT()
 #undef SLICEOPLEAFCOUNT
 
-
 /// specialisation of the \ref LeafCount struct when the node type is  TensorChippingOp
 #define CHIPPINGOPLEAFCOUNT(CVQual)\
 template <DenseIndex DimId, typename XprType>\
@@ -195,7 +202,6 @@ TENSORIMAGEPATCHOPLEAFCOUNT()
 template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>\
 struct LeafCount<CVQual TensorVolumePatchOp<Planes, Rows, Cols, XprType> >:CategoryCount<XprType>{};
 
-
 TENSORVOLUMEPATCHOPLEAFCOUNT(const)
 TENSORVOLUMEPATCHOPLEAFCOUNT()
 #undef TENSORVOLUMEPATCHOPLEAFCOUNT
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
index 99d528963..9d5708fc5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -171,19 +171,24 @@ CUSTOMBINARYOPEVAL()
 
 
 /// specialisation of the \ref PlaceHolderExpression when the node is
-/// TensorEvalToOp, TensorLayoutSwapOp
-#define EVALTOLAYOUTSWAP(CVQual, ExprNode)\
+/// TensoroOp, TensorLayoutSwapOp, and TensorIndexTupleOp
+#define EVALTOLAYOUTSWAPINDEXTUPLE(CVQual, ExprNode)\
 template <typename Expr, size_t N>\
 struct PlaceHolderExpression<CVQual ExprNode<Expr>, N> {\
   typedef CVQual ExprNode<typename CalculateIndex <N, Expr>::ArgType> Type;\
 };
 
-EVALTOLAYOUTSWAP(const, TensorEvalToOp)
-EVALTOLAYOUTSWAP(, TensorEvalToOp)
-EVALTOLAYOUTSWAP(const, TensorLayoutSwapOp)
-EVALTOLAYOUTSWAP(, TensorLayoutSwapOp)
+// TensorEvalToOp
+EVALTOLAYOUTSWAPINDEXTUPLE(const, TensorEvalToOp)
+EVALTOLAYOUTSWAPINDEXTUPLE(, TensorEvalToOp)
+//TensorLayoutSwapOp
+EVALTOLAYOUTSWAPINDEXTUPLE(const, TensorLayoutSwapOp)
+EVALTOLAYOUTSWAPINDEXTUPLE(, TensorLayoutSwapOp)
+//TensorIndexTupleOp
+EVALTOLAYOUTSWAPINDEXTUPLE(const, TensorIndexTupleOp)
+EVALTOLAYOUTSWAPINDEXTUPLE(, TensorIndexTupleOp)
 
-#undef EVALTOLAYOUTSWAP
+#undef EVALTOLAYOUTSWAPINDEXTUPLE
 
 
 /// specialisation of the \ref PlaceHolderExpression when the node is
@@ -199,17 +204,24 @@ CHIPPINGOP()
 #undef CHIPPINGOP
 
 /// specialisation of the \ref PlaceHolderExpression when the node is
-/// TensorReductionOp
-#define SYCLREDUCTION(CVQual)\
+/// TensorReductionOp and TensorTupleReducerOp (Argmax)
+#define SYCLREDUCTION(CVQual, ExprNode)\
 template <typename OP, typename Dims, typename Expr, size_t N>\
-struct PlaceHolderExpression<CVQual TensorReductionOp<OP, Dims, Expr>, N>{\
-  typedef CVQual PlaceHolder<CVQual TensorReductionOp<OP, Dims,Expr>, N> Type;\
+struct PlaceHolderExpression<CVQual ExprNode<OP, Dims, Expr>, N>{\
+  typedef CVQual PlaceHolder<CVQual ExprNode<OP, Dims,Expr>, N> Type;\
 };
-SYCLREDUCTION(const)
-SYCLREDUCTION()
+
+// tensor reduction
+SYCLREDUCTION(const, TensorReductionOp)
+SYCLREDUCTION(, TensorReductionOp)
+
+// tensor Argmax -TensorTupleReducerOp
+SYCLREDUCTION(const, TensorTupleReducerOp)
+SYCLREDUCTION(, TensorTupleReducerOp)
 #undef SYCLREDUCTION
 
 
+
 /// specialisation of the \ref PlaceHolderExpression when the node is
 /// TensorReductionOp
 #define SYCLCONTRACTIONCONVOLUTIONPLH(CVQual, ExprNode)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
index cac785540..29c78184d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
@@ -25,7 +25,6 @@
 
 namespace Eigen {
 namespace TensorSycl {
-
 template<typename Expr, typename FunctorExpr, typename TupleType > struct ExecExprFunctorKernel{
   typedef  typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr;
 
@@ -38,7 +37,7 @@ template<typename Expr, typename FunctorExpr, typename TupleType > struct ExecEx
   void operator()(cl::sycl::nd_item<1> itemID) {
     typedef  typename internal::ConvertToDeviceExpression<Expr>::Type DevExpr;
     auto device_expr =internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
-    auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
+    auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
     typename DevExpr::Index gId = static_cast<typename DevExpr::Index>(itemID.get_global_linear_id());
     if (gId < range)
       device_evaluator.evalScalar(gId);
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 996178292..4a558f856 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -173,6 +173,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_volume_patchOP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_argmax_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_custom_op_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
   # It should be safe to always run these tests as there is some fallback code for
diff --git a/unsupported/test/cxx11_tensor_argmax_sycl.cpp b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
new file mode 100644
index 000000000..9b22f1eca
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
@@ -0,0 +1,248 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_argmax_sycl
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
+#define EIGEN_USE_SYCL
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::array;
+using Eigen::SyclDevice;
+using Eigen::Tensor;
+using Eigen::TensorMap;
+
+template <typename DataType, int Layout, typename DenseIndex>
+static void test_sycl_simple_argmax(const Eigen::SyclDevice &sycl_device){
+
+  Tensor<DataType, 3, Layout, DenseIndex> in(Eigen::array<DenseIndex, 3>{{2,2,2}});
+  Tensor<DenseIndex, 0, Layout, DenseIndex> out_max;
+  Tensor<DenseIndex, 0, Layout, DenseIndex> out_min;
+  in.setRandom();
+  in *= in.constant(100.0);
+  in(0, 0, 0) = -1000.0;
+  in(1, 1, 1) = 1000.0;
+
+  std::size_t in_bytes = in.size() * sizeof(DataType);
+  std::size_t out_bytes = out_max.size() * sizeof(DenseIndex);
+
+  DataType * d_in       = static_cast<DataType*>(sycl_device.allocate(in_bytes));
+  DenseIndex* d_out_max = static_cast<DenseIndex*>(sycl_device.allocate(out_bytes));
+  DenseIndex* d_out_min = static_cast<DenseIndex*>(sycl_device.allocate(out_bytes));
+
+  Eigen::TensorMap<Eigen::Tensor<DataType, 3, Layout, DenseIndex> > gpu_in(d_in, Eigen::array<DenseIndex, 3>{{2,2,2}});
+  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 0, Layout, DenseIndex> > gpu_out_max(d_out_max);
+  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 0, Layout, DenseIndex> > gpu_out_min(d_out_min);
+  sycl_device.memcpyHostToDevice(d_in, in.data(),in_bytes);
+
+  gpu_out_max.device(sycl_device) = gpu_in.argmax();
+  gpu_out_min.device(sycl_device) = gpu_in.argmin();
+
+  sycl_device.memcpyDeviceToHost(out_max.data(), d_out_max, out_bytes);
+  sycl_device.memcpyDeviceToHost(out_min.data(), d_out_min, out_bytes);
+
+  VERIFY_IS_EQUAL(out_max(), 2*2*2 - 1);
+  VERIFY_IS_EQUAL(out_min(), 0);
+
+  sycl_device.deallocate(d_in);
+  sycl_device.deallocate(d_out_max);
+  sycl_device.deallocate(d_out_min);
+}
+
+
+template <typename DataType, int DataLayout, typename DenseIndex>
+static void test_sycl_argmax_dim(const Eigen::SyclDevice &sycl_device)
+{
+  DenseIndex sizeDim0=9;
+  DenseIndex sizeDim1=3;
+  DenseIndex sizeDim2=5;
+  DenseIndex sizeDim3=7;
+  Tensor<DataType, 4, DataLayout, DenseIndex> tensor(sizeDim0,sizeDim1,sizeDim2,sizeDim3);
+
+  std::vector<DenseIndex> dims;
+  dims.push_back(sizeDim0); dims.push_back(sizeDim1); dims.push_back(sizeDim2); dims.push_back(sizeDim3);
+  for (DenseIndex dim = 0; dim < 4; ++dim) {
+
+    array<DenseIndex, 3> out_shape;
+    for (DenseIndex d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
+
+    Tensor<DenseIndex, 3, DataLayout, DenseIndex> tensor_arg(out_shape);
+
+    array<DenseIndex, 4> ix;
+    for (DenseIndex i = 0; i < sizeDim0; ++i) {
+      for (DenseIndex j = 0; j < sizeDim1; ++j) {
+        for (DenseIndex k = 0; k < sizeDim2; ++k) {
+          for (DenseIndex l = 0; l < sizeDim3; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
+            tensor(ix)=(ix[dim] != 0)?-1.0:10.0;
+          }
+        }
+      }
+    }
+
+    std::size_t in_bytes = tensor.size() * sizeof(DataType);
+    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
+
+
+    DataType * d_in       = static_cast<DataType*>(sycl_device.allocate(in_bytes));
+    DenseIndex* d_out= static_cast<DenseIndex*>(sycl_device.allocate(out_bytes));
+
+    Eigen::TensorMap<Eigen::Tensor<DataType, 4, DataLayout, DenseIndex> > gpu_in(d_in, Eigen::array<DenseIndex, 4>{{sizeDim0,sizeDim1,sizeDim2,sizeDim3}});
+    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout, DenseIndex> > gpu_out(d_out, out_shape);
+
+    sycl_device.memcpyHostToDevice(d_in, tensor.data(),in_bytes);
+    gpu_out.device(sycl_device) = gpu_in.argmax(dim);
+    sycl_device.memcpyDeviceToHost(tensor_arg.data(), d_out, out_bytes);
+
+    VERIFY_IS_EQUAL(static_cast<size_t>(tensor_arg.size()),
+                    size_t(sizeDim0*sizeDim1*sizeDim2*sizeDim3 / tensor.dimension(dim)));
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the first index of the reduced dimension
+       VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
+    }
+
+    sycl_device.synchronize();
+
+    for (DenseIndex i = 0; i < sizeDim0; ++i) {
+      for (DenseIndex j = 0; j < sizeDim1; ++j) {
+        for (DenseIndex k = 0; k < sizeDim2; ++k) {
+          for (DenseIndex l = 0; l < sizeDim3; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
+            tensor(ix)=(ix[dim] != tensor.dimension(dim) - 1)?-1.0:20.0;
+          }
+        }
+      }
+    }
+
+    sycl_device.memcpyHostToDevice(d_in, tensor.data(),in_bytes);
+    gpu_out.device(sycl_device) = gpu_in.argmax(dim);
+    sycl_device.memcpyDeviceToHost(tensor_arg.data(), d_out, out_bytes);
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the last index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
+    }
+    sycl_device.deallocate(d_in);
+    sycl_device.deallocate(d_out);
+  }
+}
+
+
+
+
+
+template <typename DataType, int DataLayout, typename DenseIndex>
+static void test_sycl_argmin_dim(const Eigen::SyclDevice &sycl_device)
+{
+  DenseIndex sizeDim0=9;
+  DenseIndex sizeDim1=3;
+  DenseIndex sizeDim2=5;
+  DenseIndex sizeDim3=7;
+  Tensor<DataType, 4, DataLayout, DenseIndex> tensor(sizeDim0,sizeDim1,sizeDim2,sizeDim3);
+
+  std::vector<DenseIndex> dims;
+  dims.push_back(sizeDim0); dims.push_back(sizeDim1); dims.push_back(sizeDim2); dims.push_back(sizeDim3);
+  for (DenseIndex dim = 0; dim < 4; ++dim) {
+
+    array<DenseIndex, 3> out_shape;
+    for (DenseIndex d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
+
+    Tensor<DenseIndex, 3, DataLayout, DenseIndex> tensor_arg(out_shape);
+
+    array<DenseIndex, 4> ix;
+    for (DenseIndex i = 0; i < sizeDim0; ++i) {
+      for (DenseIndex j = 0; j < sizeDim1; ++j) {
+        for (DenseIndex k = 0; k < sizeDim2; ++k) {
+          for (DenseIndex l = 0; l < sizeDim3; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
+            tensor(ix)=(ix[dim] != 0)?1.0:-10.0;
+          }
+        }
+      }
+    }
+
+    std::size_t in_bytes = tensor.size() * sizeof(DataType);
+    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
+
+
+    DataType * d_in       = static_cast<DataType*>(sycl_device.allocate(in_bytes));
+    DenseIndex* d_out= static_cast<DenseIndex*>(sycl_device.allocate(out_bytes));
+
+    Eigen::TensorMap<Eigen::Tensor<DataType, 4, DataLayout, DenseIndex> > gpu_in(d_in, Eigen::array<DenseIndex, 4>{{sizeDim0,sizeDim1,sizeDim2,sizeDim3}});
+    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout, DenseIndex> > gpu_out(d_out, out_shape);
+
+    sycl_device.memcpyHostToDevice(d_in, tensor.data(),in_bytes);
+    gpu_out.device(sycl_device) = gpu_in.argmin(dim);
+    sycl_device.memcpyDeviceToHost(tensor_arg.data(), d_out, out_bytes);
+
+    VERIFY_IS_EQUAL(static_cast<size_t>(tensor_arg.size()),
+                    size_t(sizeDim0*sizeDim1*sizeDim2*sizeDim3 / tensor.dimension(dim)));
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the first index of the reduced dimension
+       VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
+    }
+
+    sycl_device.synchronize();
+
+    for (DenseIndex i = 0; i < sizeDim0; ++i) {
+      for (DenseIndex j = 0; j < sizeDim1; ++j) {
+        for (DenseIndex k = 0; k < sizeDim2; ++k) {
+          for (DenseIndex l = 0; l < sizeDim3; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
+            tensor(ix)=(ix[dim] != tensor.dimension(dim) - 1)?1.0:-20.0;
+          }
+        }
+      }
+    }
+
+    sycl_device.memcpyHostToDevice(d_in, tensor.data(),in_bytes);
+    gpu_out.device(sycl_device) = gpu_in.argmin(dim);
+    sycl_device.memcpyDeviceToHost(tensor_arg.data(), d_out, out_bytes);
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the last index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
+    }
+    sycl_device.deallocate(d_in);
+    sycl_device.deallocate(d_out);
+  }
+}
+
+
+
+
+template<typename DataType, typename Device_Selector> void sycl_argmax_test_per_device(const Device_Selector& d){
+  QueueInterface queueInterface(d);
+  auto sycl_device = Eigen::SyclDevice(&queueInterface);
+  test_sycl_simple_argmax<DataType, RowMajor, int64_t>(sycl_device);
+  test_sycl_simple_argmax<DataType, ColMajor, int64_t>(sycl_device);
+  test_sycl_argmax_dim<DataType, ColMajor, int64_t>(sycl_device);
+  test_sycl_argmax_dim<DataType, RowMajor, int64_t>(sycl_device);
+  test_sycl_argmin_dim<DataType, ColMajor, int64_t>(sycl_device);
+  test_sycl_argmin_dim<DataType, RowMajor, int64_t>(sycl_device);
+}
+
+void test_cxx11_tensor_argmax_sycl() {
+ for (const auto& device :Eigen::get_sycl_supported_devices()) {
+    CALL_SUBTEST(sycl_argmax_test_per_device<double>(device));
+  }
+}

From e2e3f785331cb90ae07b7ca7829be0ffecf6811b Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 7 Mar 2017 17:48:15 +0000
Subject: [PATCH 062/680]  Fixing potential race condition on sycl device.

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 508 +++++++++---------
 1 file changed, 262 insertions(+), 246 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 258218463..23297a0a7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -41,6 +41,7 @@ namespace Eigen {
     size_t m_i;
     size_t m_offset;
   };
+
 template<typename AccType>
   struct memsetkernelFunctor{
    AccType m_acc;
@@ -54,6 +55,21 @@ template<typename AccType>
 
   };
 
+struct memsetCghFunctor{
+  cl::sycl::buffer<uint8_t, 1>& m_buf;
+  const ptrdiff_t& buff_offset;
+  const size_t& rng , GRange, tileSize;
+  const int  &c;
+  memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
+  :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
+
+  void operator()(cl::sycl::handler &cgh) const {
+    auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+    typedef decltype(buf_acc) AccType;
+    cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor<AccType>(buf_acc, buff_offset, rng, c));
+  }
+};
+
   //get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
 EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device::get_devices()){
   auto devices = cl::sycl::device::get_devices();
@@ -75,18 +91,8 @@ EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device
   return devices;
 }
 
-struct QueueInterface {
-  /// class members:
-  bool exception_caught_ = false;
-
-  mutable std::mutex mutex_;
-
-  /// std::map is the container used to make sure that we create only one buffer
-  /// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice.
-  /// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it.
-  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
-  /// sycl queue
-  mutable cl::sycl::queue m_queue;
+class QueueInterface {
+public:
   /// creating device by using cl::sycl::selector or cl::sycl::device both are the same and can be captured through dev_Selector typename
   /// SyclStreamDevice is not owned. it is the caller's responsibility to destroy it.
   template<typename dev_Selector> explicit QueueInterface(const dev_Selector& s):
@@ -115,155 +121,6 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
 }))
 #endif
   {}
-//FIXME: currently we have to switch back to write as discard_write doesnot work in forloop
-template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
-      std::lock_guard<std::mutex> lock(mutex_);
-      auto host_acc= find_buffer(dst)->second. template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::host_buffer>();
-      ::memcpy(host_acc.get_pointer(), src, n);
-}
-
-template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  // Assuming that the dst is the start of the destination pointer
-auto it =find_buffer(src);
-auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
-offset/=sizeof(Index);
-size_t rng, GRange, tileSize;
-parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
-  auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
-  m_queue.submit([&](cl::sycl::handler &cgh) {
-    auto src_acc= it->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
-    auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
-    typedef decltype(src_acc) read_accessor;
-    typedef decltype(dst_acc) write_accessor;
-    cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
-  });
-  synchronize();
-
-}
-
-EIGEN_STRONG_INLINE void synchronize() const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  m_queue.wait_and_throw(); //pass
-}
-EIGEN_STRONG_INLINE void asynchronousExec() const {
-  ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
-  //sycl_queue().throw_asynchronous();// does not pass. Temporarily disabled
-  std::lock_guard<std::mutex> lock(mutex_);
-  m_queue.wait_and_throw(); //pass
-
-}
-
-template<typename Index>
-EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
-  tileSize =static_cast<Index>(m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
-  auto s=  m_queue.get_device().template get_info<cl::sycl::info::device::vendor>();
-  std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-    tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
-  }
-  rng = n;
-  if (rng==0) rng=static_cast<Index>(1);
-  GRange=rng;
-  if (tileSize>GRange) tileSize=GRange;
-  else if(GRange>tileSize){
-    Index xMode =  static_cast<Index>(GRange % tileSize);
-    if (xMode != 0) GRange += static_cast<Index>(tileSize - xMode);
-  }
-}
-
-/// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
-template<typename Index>
-EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
-  Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-    max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
-  }
-  Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
-  tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
-  rng1=dim1;
-  if (rng1==0 ) rng1=static_cast<Index>(1);
-  GRange1=rng1;
-  if (tileSize1>GRange1) tileSize1=GRange1;
-  else if(GRange1>tileSize1){
-    Index xMode =  static_cast<Index>(GRange1 % tileSize1);
-    if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
-  }
-  tileSize0 = static_cast<Index>(max_workgroup_Size/tileSize1);
-  rng0 = dim0;
-  if (rng0==0 ) rng0=static_cast<Index>(1);
-  GRange0=rng0;
-  if (tileSize0>GRange0) tileSize0=GRange0;
-  else if(GRange0>tileSize0){
-    Index xMode =  static_cast<Index>(GRange0 % tileSize0);
-    if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
-  }
-}
-
-
-
-/// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
-template<typename Index>
-EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
-  Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-  if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
-    max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
-  }
-  Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
-  tileSize2 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/3)));
-  rng2=dim2;
-  if (rng2==0 ) rng1=static_cast<Index>(1);
-  GRange2=rng2;
-  if (tileSize2>GRange2) tileSize2=GRange2;
-  else if(GRange2>tileSize2){
-    Index xMode =  static_cast<Index>(GRange2 % tileSize2);
-    if (xMode != 0) GRange2 += static_cast<Index>(tileSize2 - xMode);
-  }
-  pow_of_2 = static_cast<Index>(std::log2(static_cast<Index>(max_workgroup_Size/tileSize2)));
-  tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
-  rng1=dim1;
-  if (rng1==0 ) rng1=static_cast<Index>(1);
-  GRange1=rng1;
-  if (tileSize1>GRange1) tileSize1=GRange1;
-  else if(GRange1>tileSize1){
-    Index xMode =  static_cast<Index>(GRange1 % tileSize1);
-    if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
-  }
-  tileSize0 = static_cast<Index>(max_workgroup_Size/(tileSize1*tileSize2));
-  rng0 = dim0;
-  if (rng0==0 ) rng0=static_cast<Index>(1);
-  GRange0=rng0;
-  if (tileSize0>GRange0) tileSize0=GRange0;
-  else if(GRange0>tileSize0){
-    Index xMode =  static_cast<Index>(GRange0 % tileSize0);
-    if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
-  }
-}
-
-
-EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  return m_queue.get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
-//  return stream_->deviceProperties().multiProcessorCount;
-}
-EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  return m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
-
-//  return stream_->deviceProperties().maxThreadsPerBlock;
-}
-EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  // OpenCL doesnot have such concept
-  return 2;//sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
-//  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
-}
-EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
-  std::lock_guard<std::mutex> lock(mutex_);
-  return m_queue.get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
-//  return stream_->deviceProperties().sharedMemPerBlock;
-}
-
   /// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
   /// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
   /// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
@@ -292,9 +149,233 @@ EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
     std::lock_guard<std::mutex> lock(mutex_);
     buffer_map.clear();
   }
-
-  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
+  //FIXME: currently we have to switch back to write as discard_write doesnot work in forloop
+  /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
+  /// pointer created as a key we find the sycl buffer and get the host accessor with discard_write mode
+  /// on it. Using a discard_write accessor guarantees that we do not bring back the current value of the
+  /// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
+  /// this buffer is accessed, the data will be copied to the device.
+  template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
+        std::lock_guard<std::mutex> lock(mutex_);
+        auto host_acc= find_buffer(dst)->second. template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::host_buffer>();
+        ::memcpy(host_acc.get_pointer(), src, n);
+  }
+  /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
+  /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
+  /// lifespan of the memcpyDeviceToHost function. We create a kernel to copy the data, from the device- only source buffer to the destination
+  /// buffer with map_allocator on the gpu in parallel. At the end of the function call the destination buffer would be destroyed and the data
+  /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
+  /// to the cpu only once per function call.
+  template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
     std::lock_guard<std::mutex> lock(mutex_);
+    auto it =find_buffer(src);
+    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
+    offset/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+      auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
+      m_queue.submit([&](cl::sycl::handler &cgh) {
+        auto src_acc= it->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+        auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
+        typedef decltype(src_acc) read_accessor;
+        typedef decltype(dst_acc) write_accessor;
+        cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
+      });
+      synchronize();
+  }
+
+  /// the memcpy function
+  template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    auto it1 = find_buffer(static_cast<const void*>(src));
+    auto it2 = find_buffer(dst);
+    auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
+    auto i= (static_cast<const uint8_t*>(dst)) - it2->first;
+    offset/=sizeof(Index);
+    i/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+    m_queue.submit([&](cl::sycl::handler &cgh) {
+      auto src_acc =it1->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+      auto dst_acc =it2->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+      typedef decltype(src_acc) read_accessor;
+      typedef decltype(dst_acc) write_accessor;
+      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, i, offset));
+    });
+    synchronize();
+  }
+
+  EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n, tileSize, rng, GRange);
+    auto it1 = find_buffer(static_cast<const void*>(data));
+    ptrdiff_t buff_offset= (static_cast<const uint8_t*>(data)) - it1->first;
+    m_queue.submit(memsetCghFunctor(it1->second, buff_offset, rng, GRange, tileSize, c ));
+    synchronize();
+  }
+
+  /// Creation of sycl accessor for a buffer. This function first tries to find
+  /// the buffer in the buffer_map. If found it gets the accessor from it, if not,
+  /// the function then adds an entry by creating a sycl buffer for that particular pointer.
+  template <cl::sycl::access::mode AcMd> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
+  get_sycl_accessor(cl::sycl::handler &cgh, const void* ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return (find_buffer(ptr)->second.template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
+  }
+
+  /// Accessing the created sycl device buffer for the device pointer
+  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return find_buffer(ptr)->second;
+  }
+
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return (static_cast<const uint8_t*>(ptr))-(find_buffer(ptr)->first);
+  }
+
+  EIGEN_STRONG_INLINE void synchronize() const {
+    m_queue.wait_and_throw(); //pass
+  }
+
+  EIGEN_STRONG_INLINE void asynchronousExec() const {
+    ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
+    //sycl_queue().throw_asynchronous();// FIXME::does not pass. Temporarily disabled
+    m_queue.wait_and_throw(); //pass
+  }
+
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
+    tileSize =static_cast<Index>(m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
+    auto s=  m_queue.get_device().template get_info<cl::sycl::info::device::vendor>();
+    std::transform(s.begin(), s.end(), s.begin(), ::tolower);
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+      tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
+    }
+    rng = n;
+    if (rng==0) rng=static_cast<Index>(1);
+    GRange=rng;
+    if (tileSize>GRange) tileSize=GRange;
+    else if(GRange>tileSize){
+      Index xMode =  static_cast<Index>(GRange % tileSize);
+      if (xMode != 0) GRange += static_cast<Index>(tileSize - xMode);
+    }
+  }
+
+  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
+    Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+      max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
+    }
+    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+    tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
+    rng1=dim1;
+    if (rng1==0 ) rng1=static_cast<Index>(1);
+    GRange1=rng1;
+    if (tileSize1>GRange1) tileSize1=GRange1;
+    else if(GRange1>tileSize1){
+      Index xMode =  static_cast<Index>(GRange1 % tileSize1);
+      if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
+    }
+    tileSize0 = static_cast<Index>(max_workgroup_Size/tileSize1);
+    rng0 = dim0;
+    if (rng0==0 ) rng0=static_cast<Index>(1);
+    GRange0=rng0;
+    if (tileSize0>GRange0) tileSize0=GRange0;
+    else if(GRange0>tileSize0){
+      Index xMode =  static_cast<Index>(GRange0 % tileSize0);
+      if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
+    }
+  }
+
+  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
+    Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+      max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
+    }
+    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+    tileSize2 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/3)));
+    rng2=dim2;
+    if (rng2==0 ) rng1=static_cast<Index>(1);
+    GRange2=rng2;
+    if (tileSize2>GRange2) tileSize2=GRange2;
+    else if(GRange2>tileSize2){
+      Index xMode =  static_cast<Index>(GRange2 % tileSize2);
+      if (xMode != 0) GRange2 += static_cast<Index>(tileSize2 - xMode);
+    }
+    pow_of_2 = static_cast<Index>(std::log2(static_cast<Index>(max_workgroup_Size/tileSize2)));
+    tileSize1 =static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2/2)));
+    rng1=dim1;
+    if (rng1==0 ) rng1=static_cast<Index>(1);
+    GRange1=rng1;
+    if (tileSize1>GRange1) tileSize1=GRange1;
+    else if(GRange1>tileSize1){
+      Index xMode =  static_cast<Index>(GRange1 % tileSize1);
+      if (xMode != 0) GRange1 += static_cast<Index>(tileSize1 - xMode);
+    }
+    tileSize0 = static_cast<Index>(max_workgroup_Size/(tileSize1*tileSize2));
+    rng0 = dim0;
+    if (rng0==0 ) rng0=static_cast<Index>(1);
+    GRange0=rng0;
+    if (tileSize0>GRange0) tileSize0=GRange0;
+    else if(GRange0>tileSize0){
+      Index xMode =  static_cast<Index>(GRange0 % tileSize0);
+      if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
+    }
+  }
+
+  EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
+  }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+  }
+
+  /// No need for sycl it should act the same as CPU version
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
+    // OpenCL doesnot have such concept
+    return 2;
+  }
+
+  EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue;}
+
+  // This function checks if the runtime recorded an error for the
+  // underlying stream device.
+  EIGEN_STRONG_INLINE bool ok() const {
+    if (!exception_caught_) {
+      m_queue.wait_and_throw();
+    }
+    return !exception_caught_;
+  }
+
+  // destructor
+  ~QueueInterface() { buffer_map.clear(); }
+
+private:
+  /// class members:
+  bool exception_caught_ = false;
+
+  mutable std::mutex mutex_;
+
+  /// std::map is the container used to make sure that we create only one buffer
+  /// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice.
+  /// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it.
+  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
+  /// sycl queue
+  mutable cl::sycl::queue m_queue;
+  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
     auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
     if (it1 != buffer_map.end()){
       return it1;
@@ -308,37 +389,25 @@ EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
     std::cerr << "No sycl buffer found. Make sure that you have allocated memory for your buffer by calling malloc-ed function."<< std::endl;
     abort();
   }
-
-  // This function checks if the runtime recorded an error for the
-  // underlying stream device.
-  EIGEN_STRONG_INLINE bool ok() const {
-    if (!exception_caught_) {
-      m_queue.wait_and_throw();
-    }
-    return !exception_caught_;
-  }
-
-  // destructor
-  ~QueueInterface() { buffer_map.clear(); }
 };
 
+// Here is a sycl deviuce struct which accept the sycl queue interface
+// as an input
 struct SyclDevice {
   // class member.
   QueueInterface* m_queue_stream;
   /// QueueInterface is not owned. it is the caller's responsibility to destroy it.
   explicit SyclDevice(QueueInterface* queue_stream) : m_queue_stream(queue_stream){}
 
-  /// Creation of sycl accessor for a buffer. This function first tries to find
-  /// the buffer in the buffer_map. If found it gets the accessor from it, if not,
-  /// the function then adds an entry by creating a sycl buffer for that particular pointer.
+  // get sycl accessor
   template <cl::sycl::access::mode AcMd> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
   get_sycl_accessor(cl::sycl::handler &cgh, const void* ptr) const {
-    return (get_sycl_buffer(ptr).template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
+    return m_queue_stream->template get_sycl_accessor<AcMd>(cgh, ptr);
   }
 
   /// Accessing the created sycl device buffer for the device pointer
   EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
-    return m_queue_stream->find_buffer(ptr)->second;
+    return m_queue_stream->get_sycl_buffer(ptr);
   }
 
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
@@ -353,8 +422,6 @@ struct SyclDevice {
     m_queue_stream->parallel_for_setup(dim0, dim1, tileSize0, tileSize1, rng0, rng1, GRange0, GRange1);
   }
 
-
-
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
@@ -375,72 +442,27 @@ struct SyclDevice {
 
   /// the memcpy function
   template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
-    auto it1 = m_queue_stream->find_buffer(static_cast<const void*>(src));
-    auto it2 = m_queue_stream->find_buffer(dst);
-    auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
-    auto i= (static_cast<const uint8_t*>(dst)) - it2->first;
-    offset/=sizeof(Index);
-    i/=sizeof(Index);
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
-    sycl_queue().submit([&](cl::sycl::handler &cgh) {
-      auto src_acc =it1->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
-      auto dst_acc =it2->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
-      typedef decltype(src_acc) read_accessor;
-      typedef decltype(dst_acc) write_accessor;
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, i, offset));
-    });
-    synchronize();
+    m_queue_stream->memcpy(dst,src,n);
   }
 
   EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
-    auto it = m_queue_stream->find_buffer(ptr);
-    return (static_cast<const uint8_t*>(ptr))-it->first;
+    return m_queue_stream->get_offset(ptr);
 
   }
-  /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
-  /// pointer created as a key we find the sycl buffer and get the host accessor with discard_write mode
-  /// on it. Using a discard_write accessor guarantees that we do not bring back the current value of the
-  /// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
-  /// this buffer is accessed, the data will be copied to the device.
+// memcpyHostToDevice
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
      m_queue_stream->memcpyHostToDevice(dst,src,n);
   }
-  /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
-  /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
-  /// lifespan of the memcpyDeviceToHost function. We create a kernel to copy the data, from the device- only source buffer to the destination
-  /// buffer with map_allocator on the gpu in parallel. At the end of the function call the destination buffer would be destroyed and the data
-  /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
-  /// to the cpu only once per function call.
+/// here is the memcpyDeviceToHost
   template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
     m_queue_stream->memcpyDeviceToHost(dst,src,n);
   }
-  /// returning the sycl queue
-  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->m_queue;}
   /// Here is the implementation of memset function on sycl.
   EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n, tileSize, rng, GRange);
-    auto it1 = m_queue_stream->find_buffer(static_cast<const void*>(data));
-    ptrdiff_t buff_offset= (static_cast<const uint8_t*>(data)) - it1->first;
-    sycl_queue().submit(memsetCghFunctor(it1->second, buff_offset, rng, GRange, tileSize, c ));
-    synchronize();
+    m_queue_stream->memset(data,c,n);
   }
-
-  struct memsetCghFunctor{
-    cl::sycl::buffer<uint8_t, 1>& m_buf;
-    const ptrdiff_t& buff_offset;
-    const size_t& rng , GRange, tileSize;
-    const int  &c;
-    memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
-    :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
-
-    void operator()(cl::sycl::handler &cgh) const {
-      auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
-      typedef decltype(buf_acc) AccType;
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor<AccType>(buf_acc, buff_offset, rng, c));
-    }
-  };
+  /// returning the sycl queue
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->sycl_queue();}
 
   EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
     // FIXME
@@ -449,37 +471,31 @@ struct SyclDevice {
 
   EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
     // We won't try to take advantage of the l2 cache for the time being, and
-    // there is no l3 cache on cuda devices.
+    // there is no l3 cache on sycl devices.
     return firstLevelCacheSize();
   }
   EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
-  //  return stream_->deviceProperties().multiProcessorCount;
+    return m_queue_stream->getNumSyclMultiProcessors();
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
-
-  //  return stream_->deviceProperties().maxThreadsPerBlock;
+    return m_queue_stream->maxSyclThreadsPerBlock();
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
     // OpenCL doesnot have such concept
-    return 2;//sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+    return m_queue_stream->maxSyclThreadsPerMultiProcessor();
   //  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
   }
   EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
-  //  return stream_->deviceProperties().sharedMemPerBlock;
+    return m_queue_stream->sharedMemPerBlock();
   }
   /// No need for sycl it should act the same as CPU version
-  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return m_queue_stream->majorDeviceVersion(); }
 
   EIGEN_STRONG_INLINE void synchronize() const {
     m_queue_stream->synchronize(); //pass
   }
 
   EIGEN_STRONG_INLINE void asynchronousExec() const {
-    ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
-    //sycl_queue().throw_asynchronous();// does not pass. Temporarily disabled
     m_queue_stream->asynchronousExec();
   }
   // This function checks if the runtime recorded an error for the

From 5e9a1e7a7a7eccbb20a2c4eb44141727b0943f11 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Mar 2017 14:17:48 +0000
Subject: [PATCH 063/680] Adding sycl Benchmarks.

---
 bench/tensors/README                          |  10 +-
 bench/tensors/tensor_benchmarks.h             | 106 +++++++++++++++---
 bench/tensors/tensor_benchmarks_sycl.cc       |  73 ++++++++++--
 .../tensor_benchmarks_sycl_include_headers.cc |   2 +
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h |  33 +++---
 5 files changed, 184 insertions(+), 40 deletions(-)
 create mode 100644 bench/tensors/tensor_benchmarks_sycl_include_headers.cc

diff --git a/bench/tensors/README b/bench/tensors/README
index 3a5fdbe17..c4b742749 100644
--- a/bench/tensors/README
+++ b/bench/tensors/README
@@ -14,8 +14,12 @@ nvcc tensor_benchmarks_fp16_gpu.cu benchmark_main.cc -I ../../ -std=c++11 -O2 -D
 last but not least, we also provide a suite of benchmarks to measure the scalability of the contraction code on CPU. To compile these benchmarks, call
 g++ contraction_benchmarks_cpu.cc benchmark_main.cc -I ../../ -std=c++11 -O3 -DNDEBUG -pthread -mavx -o benchmarks_cpu
 
-To compile the benchmark for SYCL, using ComputeCpp you currently need 2 passes (only for translation units containing device code):
+To compile and run the benchmark for SYCL, using ComputeCpp you currently need following passes (only for translation units containing device code):
 1. The device compilation pass that generates the device code (SYCL kernels and referenced device functions) and glue code needed by the host compiler to reference the device code from host code.
-{ComputeCpp_ROOT}/bin/compute++ -I ../../ -I {ComputeCpp_ROOT}/include/ -std=c++11 -mllvm -inline-threshold=1000 -Wno-ignored-attributes -sycl -intelspirmetadata -emit-llvm -no-serial-memop -sycl-compress-name -DBUILD_PLATFORM_SPIR -DNDBUG -O3 -c tensor_benchmarks_sycl.cc
+{ComputeCpp_ROOT}/bin/compute++ -I ../../ -I {ComputeCpp_ROOT}/include/ -std=c++11 -mllvm -inline-threshold=1000 -Wno-ignored-attributes -sycl -intelspirmetadata -emit-llvm -no-serial-memop -sycl-compress-name -DBUILD_PLATFORM_SPIR -DNDBUG -O3 -c tensor_benchmarks_sycl.cc -DEIGEN_USE_SYCL=1
 2. The host compilation pass that generates the final host binary.
-clang++-3.7 -include tensor_benchmarks_sycl.sycl benchmark_main.cc tensor_benchmarks_sycl.cc -pthread -I ../../ -I {ComputeCpp_ROOT}/include/ -L {ComputeCpp_ROOT}/lib/ -lComputeCpp -lOpenCL -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11 -o tensor_benchmark_sycl
+clang++ -c benchmark_main.cc -pthread -I ../../ -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 -o benchmark_main.o
+clang++ tensor_benchmarks_sycl_include_headers.cc -pthread -I ../../ -I  {ComputeCpp_ROOT}/include/ -L  {ComputeCpp_ROOT}/lib/ -lComputeCpp -lOpenCL -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 benchmark_main.o -o tensor_benchmark_sycl
+export LD_LIBRARY_PATH={ComputeCpp_ROOT}/lib
+3. Run the benchmark
+./tensor_benchmark_sycl
diff --git a/bench/tensors/tensor_benchmarks.h b/bench/tensors/tensor_benchmarks.h
index c2fb3dede..325026113 100644
--- a/bench/tensors/tensor_benchmarks.h
+++ b/bench/tensors/tensor_benchmarks.h
@@ -35,6 +35,11 @@ template <typename Device, typename T> class BenchmarkSuite {
 
   void memcpy(int num_iters) {
     eigen_assert(m_ == k_ && k_ == n_);
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      device_.memcpy(c_, a_, m_ * m_ * sizeof(T));
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       device_.memcpy(c_, a_, m_ * m_ * sizeof(T));
@@ -55,7 +60,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     }
     const TensorMap<Tensor<int, 2, 0, TensorIndex>, Eigen::Aligned> A((int*)a_, sizes);
     TensorMap<Tensor<T, 2, 0, TensorIndex>, Eigen::Aligned> B(b_, sizes);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      B.device(device_) = A.template cast<T>();
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       B.device(device_) = A.template cast<T>();
@@ -70,7 +79,6 @@ template <typename Device, typename T> class BenchmarkSuite {
     sizes[0] = m_;
     sizes[1] = m_;
     TensorMap<Tensor<T, 2>, Eigen::Aligned> C(c_, sizes);
-
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = C.random();
@@ -93,7 +101,18 @@ template <typename Device, typename T> class BenchmarkSuite {
     const Eigen::DSizes<TensorIndex, 2> second_quadrant(0, m_/2);
     const Eigen::DSizes<TensorIndex, 2> third_quadrant(m_/2, 0);
     const Eigen::DSizes<TensorIndex, 2> fourth_quadrant(m_/2, m_/2);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.slice(first_quadrant, quarter_sizes).device(device_) =
+          A.slice(first_quadrant, quarter_sizes);
+      C.slice(second_quadrant, quarter_sizes).device(device_) =
+          B.slice(second_quadrant, quarter_sizes);
+      C.slice(third_quadrant, quarter_sizes).device(device_) =
+          A.slice(third_quadrant, quarter_sizes);
+      C.slice(fourth_quadrant, quarter_sizes).device(device_) =
+          B.slice(fourth_quadrant, quarter_sizes);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.slice(first_quadrant, quarter_sizes).device(device_) =
@@ -118,7 +137,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::array<TensorIndex, 1> output_size;
     output_size[0] = n_;
     TensorMap<Tensor<T, 1, 0, TensorIndex>, Eigen::Aligned> C(c_, output_size);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = B.chip(iter % k_, 0);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = B.chip(iter % k_, 0);
@@ -135,7 +158,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::array<TensorIndex, 1> output_size;
     output_size[0] = n_;
     TensorMap<Tensor<T, 1, 0, TensorIndex>, Eigen::Aligned> C(c_, output_size);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = B.chip(iter % n_, 1);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = B.chip(iter % n_, 1);
@@ -158,7 +185,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::array<int, 2> shuffle;
     shuffle[0] = 1;
     shuffle[1] = 0;
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      B.device(device_) = A.shuffle(shuffle);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       B.device(device_) = A.shuffle(shuffle);
@@ -186,7 +217,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     paddings[0] = Eigen::IndexPair<TensorIndex>(0, 0);
     paddings[1] = Eigen::IndexPair<TensorIndex>(2, 1);
 #endif
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      B.device(device_) = A.pad(paddings);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       B.device(device_) = A.pad(paddings);
@@ -216,6 +251,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::IndexList<Eigen::type2index<1>, Eigen::type2index<2> > strides;
 #endif
 
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      B.device(device_) = A.stride(strides);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       B.device(device_) = A.stride(strides);
@@ -245,6 +285,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     broadcast.set(1, n_);
 #endif
 
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A.broadcast(broadcast);
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A.broadcast(broadcast);
@@ -261,7 +306,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     const TensorMap<Tensor<T, 2>, Eigen::Aligned> A(a_, sizes);
     const TensorMap<Tensor<T, 2>, Eigen::Aligned> B(b_, sizes);
     TensorMap<Tensor<T, 2>, Eigen::Aligned> C(c_, sizes);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A * A.constant(static_cast<T>(3.14)) + B * B.constant(static_cast<T>(2.7));
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A * A.constant(static_cast<T>(3.14)) + B * B.constant(static_cast<T>(2.7));
@@ -280,6 +329,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     const TensorMap<Tensor<T, 2>, Eigen::Aligned> B(b_, sizes);
     TensorMap<Tensor<T, 2>, Eigen::Aligned> C(c_, sizes);
 
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A.rsqrt() + B.sqrt() * B.square();
+}
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A.rsqrt() + B.sqrt() * B.square();
@@ -297,7 +351,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     const TensorMap<Tensor<T, 2>, Eigen::Aligned> A(a_, sizes);
     const TensorMap<Tensor<T, 2>, Eigen::Aligned> B(b_, sizes);
     TensorMap<Tensor<T, 2>, Eigen::Aligned> C(c_, sizes);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A.exp() + B.log();
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A.exp() + B.log();
@@ -325,7 +383,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     // optimize the code.
     Eigen::IndexList<Eigen::type2index<0>> sum_along_dim;
 #endif
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+  for (int iter = 0; iter < 10; ++iter) {
+    C.device(device_) = B.sum(sum_along_dim);
+  }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = B.sum(sum_along_dim);
@@ -355,7 +417,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     // optimize the code.
     Eigen::IndexList<Eigen::type2index<1>> sum_along_dim;
 #endif
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+  for (int iter = 0; iter < 10; ++iter) {
+    C.device(device_) = B.sum(sum_along_dim);
+  }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = B.sum(sum_along_dim);
@@ -375,7 +441,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::array<TensorIndex, 0> output_size;
     TensorMap<Tensor<T, 0, 0, TensorIndex>, Eigen::Aligned> C(
         c_, output_size);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = B.sum();
+    }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = B.sum();
@@ -404,7 +474,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     typedef typename Tensor<T, 2>::DimensionPair DimPair;
     Eigen::array<DimPair, 1> dims;
     dims[0] = DimPair(1, 0);
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A.contract(B, dims);
+     }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A.contract(B, dims);
@@ -430,7 +504,11 @@ template <typename Device, typename T> class BenchmarkSuite {
     Eigen::array<TensorIndex, 2> dims;
     dims[0] = 0;
     dims[1] = 1;
-
+#ifdef EIGEN_USE_SYCL // warmup for sycl
+    for (int iter = 0; iter < 10; ++iter) {
+      C.device(device_) = A.convolve(B, dims);
+     }
+#endif
     StartBenchmarkTiming();
     for (int iter = 0; iter < num_iters; ++iter) {
       C.device(device_) = A.convolve(B, dims);
diff --git a/bench/tensors/tensor_benchmarks_sycl.cc b/bench/tensors/tensor_benchmarks_sycl.cc
index 6df190869..cb6daac15 100644
--- a/bench/tensors/tensor_benchmarks_sycl.cc
+++ b/bench/tensors/tensor_benchmarks_sycl.cc
@@ -1,20 +1,73 @@
-#define EIGEN_USE_SYCL
+#ifdef EIGEN_USE_SYCL
 
 #include <SYCL/sycl.hpp>
 #include <iostream>
 
 #include "tensor_benchmarks.h"
 
-#define BM_FuncGPU(FUNC)                                       \
-  static void BM_##FUNC(int iters, int N) {                    \
-    StopBenchmarkTiming();                                     \
-    cl::sycl::gpu_selector selector; \
-    Eigen::QueueInterface queue(selector);     \
-    Eigen::SyclDevice device(&queue);                          \
-    BenchmarkSuite<Eigen::SyclDevice, float> suite(device, N); \
-    suite.FUNC(iters);                                         \
-  }                                                            \
+#define BM_FuncGPU(FUNC)                                                       \
+  static void BM_##FUNC(int iters, int N) {                                    \
+    StopBenchmarkTiming();                                                     \
+    cl::sycl::gpu_selector selector;                                           \
+    Eigen::QueueInterface queue(selector);                                     \
+    Eigen::SyclDevice device(&queue);                                          \
+    BenchmarkSuite<Eigen::SyclDevice, float> suite(device, N);                 \
+    suite.FUNC(iters);                                                         \
+  }                                                                            \
   BENCHMARK_RANGE(BM_##FUNC, 10, 5000);
 
+BM_FuncGPU(memcpy);
+BM_FuncGPU(typeCasting);
+BM_FuncGPU(slicing);
+BM_FuncGPU(rowChip);
+BM_FuncGPU(colChip);
+BM_FuncGPU(shuffling);
+BM_FuncGPU(padding);
+BM_FuncGPU(striding);
 BM_FuncGPU(broadcasting);
 BM_FuncGPU(coeffWiseOp);
+BM_FuncGPU(algebraicFunc);
+BM_FuncGPU(transcendentalFunc);
+BM_FuncGPU(rowReduction);
+BM_FuncGPU(colReduction);
+BM_FuncGPU(fullReduction);
+
+
+// Contractions
+#define BM_FuncWithInputDimsGPU(FUNC, D1, D2, D3)                              \
+  static void BM_##FUNC##_##D1##x##D2##x##D3(int iters, int N) {               \
+    StopBenchmarkTiming();                                                     \
+    cl::sycl::gpu_selector selector;                                           \
+    Eigen::QueueInterface queue(selector);                                     \
+    Eigen::SyclDevice device(&queue);                                          \
+    BenchmarkSuite<Eigen::SyclDevice, float> suite(device, D1, D2, D3);        \
+    suite.FUNC(iters);                                                         \
+  }                                                                            \
+  BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2##x##D3, 10, 5000);
+
+
+BM_FuncWithInputDimsGPU(contraction, N, N, N);
+BM_FuncWithInputDimsGPU(contraction, 64, N, N);
+BM_FuncWithInputDimsGPU(contraction, N, 64, N);
+BM_FuncWithInputDimsGPU(contraction, N, N, 64);
+
+
+// Convolutions
+#define BM_FuncWithKernelDimsGPU(FUNC, DIM1, DIM2)                             \
+  static void BM_##FUNC##_##DIM1##x##DIM2(int iters, int N) {                  \
+    StopBenchmarkTiming();                                                     \
+    cl::sycl::gpu_selector selector;                                           \
+    Eigen::QueueInterface queue(selector);                                     \
+    Eigen::SyclDevice device(&queue);                                          \
+    BenchmarkSuite<Eigen::SyclDevice, float> suite(device, N);                 \
+    suite.FUNC(iters, DIM1, DIM2);                                             \
+  }                                                                            \
+  BENCHMARK_RANGE(BM_##FUNC##_##DIM1##x##DIM2, 128, 5000);
+
+BM_FuncWithKernelDimsGPU(convolution, 7, 1);
+BM_FuncWithKernelDimsGPU(convolution, 1, 7);
+BM_FuncWithKernelDimsGPU(convolution, 7, 4);
+BM_FuncWithKernelDimsGPU(convolution, 4, 7);
+BM_FuncWithKernelDimsGPU(convolution, 7, 64);
+BM_FuncWithKernelDimsGPU(convolution, 64, 7);
+#endif
diff --git a/bench/tensors/tensor_benchmarks_sycl_include_headers.cc b/bench/tensors/tensor_benchmarks_sycl_include_headers.cc
new file mode 100644
index 000000000..4b3110b85
--- /dev/null
+++ b/bench/tensors/tensor_benchmarks_sycl_include_headers.cc
@@ -0,0 +1,2 @@
+#include "/home/mehdi/Projects/upstr_benoit/upstr_7MAR17/bench/tensors/tensor_benchmarks_sycl.cc"
+#include "/home/mehdi/Projects/upstr_benoit/upstr_7MAR17/bench/tensors/tensor_benchmarks_sycl.sycl"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 23297a0a7..5fa7423d0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -149,16 +149,27 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
     std::lock_guard<std::mutex> lock(mutex_);
     buffer_map.clear();
   }
-  //FIXME: currently we have to switch back to write as discard_write doesnot work in forloop
   /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
-  /// pointer created as a key we find the sycl buffer and get the host accessor with discard_write mode
-  /// on it. Using a discard_write accessor guarantees that we do not bring back the current value of the
-  /// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
+  /// pointer created as a key we find the sycl buffer and get the host accessor with write mode
+  /// on it. Then we use the memcpy to copy the data to the host accessor. The first time that
   /// this buffer is accessed, the data will be copied to the device.
+  /// In this case we can separate the kernel actual execution from data transfer which is required for benchmark
+  /// Also, this is faster as it uses the map_allocator instead of memcpy
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
-        std::lock_guard<std::mutex> lock(mutex_);
-        auto host_acc= find_buffer(dst)->second. template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::host_buffer>();
-        ::memcpy(host_acc.get_pointer(), src, n);
+    auto it =find_buffer(dst);
+    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(dst))- it->first;
+    offset/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+      auto src_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(static_cast<void*>(const_cast<Index*>(src))), cl::sycl::range<1>(n));
+      m_queue.submit([&](cl::sycl::handler &cgh) {
+        auto dst_acc= it->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+        auto src_acc =src_buf.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+        typedef decltype(src_acc) read_accessor;
+        typedef decltype(dst_acc) write_accessor;
+        cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, offset, 0));
+      });
+      synchronize();
   }
   /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
   /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
@@ -167,7 +178,6 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
   /// to the cpu only once per function call.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     auto it =find_buffer(src);
     auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
     offset/=sizeof(Index);
@@ -186,7 +196,6 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
 
   /// the memcpy function
   template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     auto it1 = find_buffer(static_cast<const void*>(src));
     auto it2 = find_buffer(dst);
     auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
@@ -206,7 +215,6 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   }
 
   EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     size_t rng, GRange, tileSize;
     parallel_for_setup(n, tileSize, rng, GRange);
     auto it1 = find_buffer(static_cast<const void*>(data));
@@ -220,18 +228,15 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   /// the function then adds an entry by creating a sycl buffer for that particular pointer.
   template <cl::sycl::access::mode AcMd> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
   get_sycl_accessor(cl::sycl::handler &cgh, const void* ptr) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     return (find_buffer(ptr)->second.template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
   }
 
   /// Accessing the created sycl device buffer for the device pointer
   EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     return find_buffer(ptr)->second;
   }
 
   EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
-    std::lock_guard<std::mutex> lock(mutex_);
     return (static_cast<const uint8_t*>(ptr))-(find_buffer(ptr)->first);
   }
 
@@ -375,7 +380,9 @@ private:
   mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
   /// sycl queue
   mutable cl::sycl::queue m_queue;
+
   EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
     auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
     if (it1 != buffer_map.end()){
       return it1;

From 970ff78294503896940fb760d948f1eed156250f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 8 Mar 2017 16:16:53 +0100
Subject: [PATCH 064/680] bug #1401: fix compilation of "cond ? x : -x" with x
 an AutoDiffScalar

---
 unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
index 50fedf6ac..d2808860c 100755
--- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
+++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
@@ -108,7 +108,9 @@ class AutoDiffScalar
     template<typename OtherDerType>
     AutoDiffScalar(const AutoDiffScalar<OtherDerType>& other
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-    , typename internal::enable_if<internal::is_same<Scalar, typename internal::traits<typename internal::remove_all<OtherDerType>::type>::Scalar>::value,void*>::type = 0
+    , typename internal::enable_if<
+            internal::is_same<Scalar, typename internal::traits<typename internal::remove_all<OtherDerType>::type>::Scalar>::value
+        &&  internal::is_convertible<OtherDerType,DerType>::value , void*>::type = 0
 #endif
     )
       : m_value(other.value()), m_derivatives(other.derivatives())

From aadb7405a7362ce0160d8ecb3843dc33a59e809a Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Mar 2017 18:20:06 +0000
Subject: [PATCH 065/680] Fixing typo in sycl Benchmark.

---
 bench/tensors/README                                    | 4 ++--
 bench/tensors/tensor_benchmarks_sycl_include_headers.cc | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/bench/tensors/README b/bench/tensors/README
index c4b742749..69342cc9c 100644
--- a/bench/tensors/README
+++ b/bench/tensors/README
@@ -18,8 +18,8 @@ To compile and run the benchmark for SYCL, using ComputeCpp you currently need f
 1. The device compilation pass that generates the device code (SYCL kernels and referenced device functions) and glue code needed by the host compiler to reference the device code from host code.
 {ComputeCpp_ROOT}/bin/compute++ -I ../../ -I {ComputeCpp_ROOT}/include/ -std=c++11 -mllvm -inline-threshold=1000 -Wno-ignored-attributes -sycl -intelspirmetadata -emit-llvm -no-serial-memop -sycl-compress-name -DBUILD_PLATFORM_SPIR -DNDBUG -O3 -c tensor_benchmarks_sycl.cc -DEIGEN_USE_SYCL=1
 2. The host compilation pass that generates the final host binary.
-clang++ -c benchmark_main.cc -pthread -I ../../ -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 -o benchmark_main.o
-clang++ tensor_benchmarks_sycl_include_headers.cc -pthread -I ../../ -I  {ComputeCpp_ROOT}/include/ -L  {ComputeCpp_ROOT}/lib/ -lComputeCpp -lOpenCL -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 benchmark_main.o -o tensor_benchmark_sycl
+clang++ -O3 -c benchmark_main.cc -pthread -I ../../ -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 -o benchmark_main.o
+clang++ -O3 tensor_benchmarks_sycl_include_headers.cc -pthread -I ../../ -I  {ComputeCpp_ROOT}/include/ -L  {ComputeCpp_ROOT}/lib/ -lComputeCpp -lOpenCL -D_GLIBCXX_USE_CXX11_ABI=0 -DEIGEN_USE_SYCL=1 -std=c++11 benchmark_main.o -o tensor_benchmark_sycl
 export LD_LIBRARY_PATH={ComputeCpp_ROOT}/lib
 3. Run the benchmark
 ./tensor_benchmark_sycl
diff --git a/bench/tensors/tensor_benchmarks_sycl_include_headers.cc b/bench/tensors/tensor_benchmarks_sycl_include_headers.cc
index 4b3110b85..bcc3c4c79 100644
--- a/bench/tensors/tensor_benchmarks_sycl_include_headers.cc
+++ b/bench/tensors/tensor_benchmarks_sycl_include_headers.cc
@@ -1,2 +1,2 @@
-#include "/home/mehdi/Projects/upstr_benoit/upstr_7MAR17/bench/tensors/tensor_benchmarks_sycl.cc"
-#include "/home/mehdi/Projects/upstr_benoit/upstr_7MAR17/bench/tensors/tensor_benchmarks_sycl.sycl"
+#include "tensor_benchmarks_sycl.cc"
+#include "tensor_benchmarks_sycl.sycl"

From 1b32a10053a942b1c6010afd719b44393b115d42 Mon Sep 17 00:00:00 2001
From: Luke Iwanski <luke@codeplay.com>
Date: Wed, 8 Mar 2017 18:26:34 +0000
Subject: [PATCH 066/680] Use name to distinguish name instead of the vendor

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 5fa7423d0..ed21d7b56 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -76,7 +76,7 @@ EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device
   std::vector<cl::sycl::device>::iterator it =devices.begin();
   while(it!=devices.end()) {
     ///FIXME: Currently there is a bug in amd cpu OpenCL
-    auto s=  (*it).template get_info<cl::sycl::info::device::vendor>();
+    auto s=  (*it).template get_info<cl::sycl::info::device::name>();
     std::transform(s.begin(), s.end(), s.begin(), ::tolower);
     if((*it).is_cpu() && s.find("amd")!=std::string::npos && s.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
       it=devices.erase(it);

From 344c2694a64494721e6d36d1197bde47c7d12af9 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 9 Mar 2017 15:41:03 -0800
Subject: [PATCH 067/680] Make the non-blocking threadpool more flexible and
 less wasteful of CPU cycles for high-latency use-cases.

* Adds a hint to ThreadPool allowing us to turn off spin waiting. Currently each reader and record yielder op in a graph creates a threadpool with a thread that spins for 1000 iterations through the work stealing loop before yielding. This is wasteful for such ops that process I/O.

* This also changes the number of iterations through the steal loop to be inversely proportional to the number of threads. Since the time of each iteration is proportional to the number of threads, this yields roughly a constant spin time.

* Implement a separate worker loop for the num_threads == 1 case since there is no point in going through the expensive steal loop. Moreover, since Steal() calls PopBack() on the victim queues it might reverse the order in which ops are executed, compared to the order in which they are scheduled, which is usually counter-productive for the types of I/O workloads the single thread pools tend to be used for.

* Store num_threads in a member variable for simplicity and to avoid a data race between the thread creation loop and worker threads calling threads_.size().
---
 .../src/ThreadPool/NonBlockingThreadPool.h    | 163 ++++++++++++------
 .../test/cxx11_non_blocking_thread_pool.cpp   |   7 +-
 2 files changed, 111 insertions(+), 59 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index ed1a761b6..e28afedb4 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -20,7 +20,9 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   typedef RunQueue<Task, 1024> Queue;
 
   NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
-      : env_(env),
+      : num_threads_(num_threads),
+        allow_spinning_(true),
+        env_(env),
         threads_(num_threads),
         queues_(num_threads),
         coprimes_(num_threads),
@@ -30,34 +32,24 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
         done_(false),
         cancelled_(false),
         ec_(waiters_) {
-    waiters_.resize(num_threads);
+    Init();
+  }
 
-    // Calculate coprimes of num_threads.
-    // Coprimes are used for a random walk over all threads in Steal
-    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
-    // a walk starting thread index t and calculate num_threads - 1 subsequent
-    // indices as (t + coprime) % num_threads, we will cover all threads without
-    // repetitions (effectively getting a presudo-random permutation of thread
-    // indices).
-    for (int i = 1; i <= num_threads; i++) {
-      unsigned a = i;
-      unsigned b = num_threads;
-      // If GCD(a, b) == 1, then a and b are coprimes.
-      while (b != 0) {
-        unsigned tmp = a;
-        a = b;
-        b = tmp % b;
-      }
-      if (a == 1) {
-        coprimes_.push_back(i);
-      }
-    }
-    for (int i = 0; i < num_threads; i++) {
-      queues_.push_back(new Queue());
-    }
-    for (int i = 0; i < num_threads; i++) {
-      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
-    }
+  NonBlockingThreadPoolTempl(int num_threads, bool allow_spinning,
+                             Environment env = Environment())
+      : num_threads_(num_threads),
+        allow_spinning_(allow_spinning),
+        env_(env),
+        threads_(num_threads),
+        queues_(num_threads),
+        coprimes_(num_threads),
+        waiters_(num_threads),
+        blocked_(0),
+        spinning_(0),
+        done_(false),
+        cancelled_(false),
+        ec_(waiters_) {
+    Init();
   }
 
   ~NonBlockingThreadPoolTempl() {
@@ -77,8 +69,8 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
     }
 
     // Join threads explicitly to avoid destruction order issues.
-    for (size_t i = 0; i < threads_.size(); i++) delete threads_[i];
-    for (size_t i = 0; i < threads_.size(); i++) delete queues_[i];
+    for (size_t i = 0; i < num_threads_; i++) delete threads_[i];
+    for (size_t i = 0; i < num_threads_; i++) delete queues_[i];
   }
 
   void Schedule(std::function<void()> fn) {
@@ -125,7 +117,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   }
 
   int NumThreads() const final {
-    return static_cast<int>(threads_.size());
+    return num_threads_;
   }
 
   int CurrentThreadId() const final {
@@ -149,6 +141,8 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   };
 
   Environment env_;
+  const int num_threads_;
+  const bool allow_spinning_;
   MaxSizeVector<Thread*> threads_;
   MaxSizeVector<Queue*> queues_;
   MaxSizeVector<unsigned> coprimes_;
@@ -159,6 +153,37 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   std::atomic<bool> cancelled_;
   EventCount ec_;
 
+  void Init() {
+    waiters_.resize(num_threads_);
+
+    // Calculate coprimes of num_threads_.
+    // Coprimes are used for a random walk over all threads in Steal
+    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
+    // a walk starting thread index t and calculate num_threads - 1 subsequent
+    // indices as (t + coprime) % num_threads, we will cover all threads without
+    // repetitions (effectively getting a presudo-random permutation of thread
+    // indices).
+    for (int i = 1; i <= num_threads_; i++) {
+      unsigned a = i;
+      unsigned b = num_threads_;
+      // If GCD(a, b) == 1, then a and b are coprimes.
+      while (b != 0) {
+        unsigned tmp = a;
+        a = b;
+        b = tmp % b;
+      }
+      if (a == 1) {
+        coprimes_.push_back(i);
+      }
+    }
+    for (int i = 0; i < num_threads_; i++) {
+      queues_.push_back(new Queue());
+    }
+    for (int i = 0; i < num_threads_; i++) {
+      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
+    }
+  }
+
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
     PerThread* pt = GetPerThread();
@@ -167,36 +192,62 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
     pt->thread_id = thread_id;
     Queue* q = queues_[thread_id];
     EventCount::Waiter* waiter = &waiters_[thread_id];
-    while (!cancelled_) {
-      Task t = q->PopFront();
-      if (!t.f) {
-        t = Steal();
+    // TODO(dvyukov,rmlarsen): The time spent in Steal() is proportional
+    // to num_threads_ and we assume that new work is scheduled at a
+    // constant rate, so we set spin_count to 5000 / num_threads_. The
+    // constant was picked based on a fair dice roll, tune it.
+    const int spin_count =
+        allow_spinning_ && num_threads_ > 0 ? 5000 / num_threads_ : 0;
+    if (num_threads_ == 1) {
+      // For num_threads_ == 1 there is no point in going through the expensive
+      // steal loop. Moreover, since Steal() calls PopBack() on the victim
+      // queues it might reverse the order in which ops are executed compared to
+      // the order in which they are scheduled, which tends to be
+      // counter-productive for the types of I/O workloads the single thread
+      // pools tend to be used for.
+      while (!cancelled_) {
+        Task t = q->PopFront();
+        for (int i = 0; i < spin_count && !t.f; i++) {
+          if (!cancelled_.load(std::memory_order_relaxed)) {
+            t = q->PopFront();
+          }
+        }
         if (!t.f) {
-          // Leave one thread spinning. This reduces latency.
-          // TODO(dvyukov): 1000 iterations is based on fair dice roll, tune it.
-          // Also, the time it takes to attempt to steal work 1000 times depends
-          // on the size of the thread pool. However the speed at which the user
-          // of the thread pool submit tasks is independent of the size of the
-          // pool. Consider a time based limit instead.
-          if (!spinning_ && !spinning_.exchange(true)) {
-            for (int i = 0; i < 1000 && !t.f; i++) {
-              if (!cancelled_.load(std::memory_order_relaxed)) {
-                t = Steal();
-              } else {
+          if (!WaitForWork(waiter, &t)) {
+            return;
+          }
+        }
+        if (t.f) {
+          env_.ExecuteTask(t);
+        }
+      }
+    } else {
+      while (!cancelled_) {
+        Task t = q->PopFront();
+        if (!t.f) {
+          t = Steal();
+          if (!t.f) {
+            // Leave one thread spinning. This reduces latency.
+            if (allow_spinning_ && !spinning_ && !spinning_.exchange(true)) {
+              for (int i = 0; i < spin_count && !t.f; i++) {
+                if (!cancelled_.load(std::memory_order_relaxed)) {
+                  t = Steal();
+                } else {
+                  return;
+                }
+              }
+              spinning_ = false;
+            }
+            if (!t.f) {
+              if (!WaitForWork(waiter, &t)) {
                 return;
               }
             }
-            spinning_ = false;
-          }
-          if (!t.f) {
-            if (!WaitForWork(waiter, &t)) {
-              return;
-            }
           }
         }
-      }
-      if (t.f) {
-        env_.ExecuteTask(t);
+        if (t.f) {
+          env_.ExecuteTask(t);
+        }
       }
     }
   }
@@ -244,7 +295,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // If we are shutting down and all worker threads blocked without work,
     // that's we are done.
     blocked_++;
-    if (done_ && blocked_ == threads_.size()) {
+    if (done_ && blocked_ == num_threads_) {
       ec_.CancelWait(waiter);
       // Almost done, but need to re-check queues.
       // Consider that all queues are empty and all worker threads are preempted
diff --git a/unsupported/test/cxx11_non_blocking_thread_pool.cpp b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
index 2c5765ce4..48cd2d4e4 100644
--- a/unsupported/test/cxx11_non_blocking_thread_pool.cpp
+++ b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
@@ -23,11 +23,11 @@ static void test_create_destroy_empty_pool()
 }
 
 
-static void test_parallelism()
+static void test_parallelism(bool allow_spinning)
 {
   // Test we never-ever fail to match available tasks with idle threads.
   const int kThreads = 16;  // code below expects that this is a multiple of 4
-  NonBlockingThreadPool tp(kThreads);
+  NonBlockingThreadPool tp(kThreads, allow_spinning);
   VERIFY_IS_EQUAL(tp.NumThreads(), kThreads);
   VERIFY_IS_EQUAL(tp.CurrentThreadId(), -1);
   for (int iter = 0; iter < 100; ++iter) {
@@ -119,6 +119,7 @@ static void test_cancel()
 void test_cxx11_non_blocking_thread_pool()
 {
   CALL_SUBTEST(test_create_destroy_empty_pool());
-  CALL_SUBTEST(test_parallelism());
+  CALL_SUBTEST(test_parallelism(true));
+  CALL_SUBTEST(test_parallelism(false));
   CALL_SUBTEST(test_cancel());
 }

From d56ab01094d30cf57150b454b8ebaf20ac7bc85e Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 10 Mar 2017 08:30:22 -0800
Subject: [PATCH 068/680] Use C++11 ctor forwarding to simplify code a bit.

---
 .../src/ThreadPool/NonBlockingThreadPool.h    | 19 +++----------------
 1 file changed, 3 insertions(+), 16 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index e28afedb4..4fbd3af1e 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -19,22 +19,6 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   typedef typename Environment::Task Task;
   typedef RunQueue<Task, 1024> Queue;
 
-  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
-      : num_threads_(num_threads),
-        allow_spinning_(true),
-        env_(env),
-        threads_(num_threads),
-        queues_(num_threads),
-        coprimes_(num_threads),
-        waiters_(num_threads),
-        blocked_(0),
-        spinning_(0),
-        done_(false),
-        cancelled_(false),
-        ec_(waiters_) {
-    Init();
-  }
-
   NonBlockingThreadPoolTempl(int num_threads, bool allow_spinning,
                              Environment env = Environment())
       : num_threads_(num_threads),
@@ -52,6 +36,9 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
     Init();
   }
 
+  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
+      : NonBlockingThreadPoolTempl(num_threads, true, env) {}
+
   ~NonBlockingThreadPoolTempl() {
     done_ = true;
 

From bfd7bf9c5b4ebc932a339200bb4a455eadcf6d28 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 10 Mar 2017 08:48:20 -0800
Subject: [PATCH 069/680] Get rid of Init().

---
 .../src/ThreadPool/NonBlockingThreadPool.h    | 66 +++++++++----------
 1 file changed, 31 insertions(+), 35 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 4fbd3af1e..9dcc9dab7 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -19,6 +19,9 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   typedef typename Environment::Task Task;
   typedef RunQueue<Task, 1024> Queue;
 
+  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
+      : NonBlockingThreadPoolTempl(num_threads, true, env) {}
+
   NonBlockingThreadPoolTempl(int num_threads, bool allow_spinning,
                              Environment env = Environment())
       : num_threads_(num_threads),
@@ -33,11 +36,35 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
         done_(false),
         cancelled_(false),
         ec_(waiters_) {
-    Init();
-  }
+    waiters_.resize(num_threads_);
 
-  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
-      : NonBlockingThreadPoolTempl(num_threads, true, env) {}
+    // Calculate coprimes of num_threads_.
+    // Coprimes are used for a random walk over all threads in Steal
+    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
+    // a walk starting thread index t and calculate num_threads - 1 subsequent
+    // indices as (t + coprime) % num_threads, we will cover all threads without
+    // repetitions (effectively getting a presudo-random permutation of thread
+    // indices).
+    for (int i = 1; i <= num_threads_; i++) {
+      unsigned a = i;
+      unsigned b = num_threads_;
+      // If GCD(a, b) == 1, then a and b are coprimes.
+      while (b != 0) {
+        unsigned tmp = a;
+        a = b;
+        b = tmp % b;
+      }
+      if (a == 1) {
+        coprimes_.push_back(i);
+      }
+    }
+    for (int i = 0; i < num_threads_; i++) {
+      queues_.push_back(new Queue());
+    }
+    for (int i = 0; i < num_threads_; i++) {
+      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
+    }
+  }
 
   ~NonBlockingThreadPoolTempl() {
     done_ = true;
@@ -140,37 +167,6 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   std::atomic<bool> cancelled_;
   EventCount ec_;
 
-  void Init() {
-    waiters_.resize(num_threads_);
-
-    // Calculate coprimes of num_threads_.
-    // Coprimes are used for a random walk over all threads in Steal
-    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
-    // a walk starting thread index t and calculate num_threads - 1 subsequent
-    // indices as (t + coprime) % num_threads, we will cover all threads without
-    // repetitions (effectively getting a presudo-random permutation of thread
-    // indices).
-    for (int i = 1; i <= num_threads_; i++) {
-      unsigned a = i;
-      unsigned b = num_threads_;
-      // If GCD(a, b) == 1, then a and b are coprimes.
-      while (b != 0) {
-        unsigned tmp = a;
-        a = b;
-        b = tmp % b;
-      }
-      if (a == 1) {
-        coprimes_.push_back(i);
-      }
-    }
-    for (int i = 0; i < num_threads_; i++) {
-      queues_.push_back(new Queue());
-    }
-    for (int i = 0; i < num_threads_; i++) {
-      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
-    }
-  }
-
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
     PerThread* pt = GetPerThread();

From f499fe9496e7c5e6f70d4bdcfb6ed9088795431a Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 13 Mar 2017 09:18:37 +0000
Subject: [PATCH 070/680] Adding synchronisation to convolution kernel for sycl
 backend.

---
 bench/tensors/tensor_benchmarks.h                          | 5 +++++
 unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h | 1 +
 2 files changed, 6 insertions(+)

diff --git a/bench/tensors/tensor_benchmarks.h b/bench/tensors/tensor_benchmarks.h
index 325026113..3a640ede4 100644
--- a/bench/tensors/tensor_benchmarks.h
+++ b/bench/tensors/tensor_benchmarks.h
@@ -539,6 +539,11 @@ for (int iter = 0; iter < 10; ++iter) {
     if (Eigen::internal::is_same<Device, Eigen::GpuDevice>::value) {
       device_.synchronize();
     }
+#elif defined(EIGEN_USE_SYCL)
+    if (Eigen::internal::is_same<Device, Eigen::SyclDevice>::value) {
+      device_.synchronize();
+    }
+
 #endif
     StopBenchmarkTiming();
     SetBenchmarkFlopsProcessed(num_items);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index 5db16d559..2e6021b1e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -425,6 +425,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
         }
       }
     });
+    m_device.asynchronousExec();
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const

From f0f35911181cc7e2089a3319d966dcbd3596461b Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 14 Mar 2017 14:16:53 -0700
Subject: [PATCH 071/680] Made the reduction code compile with cuda-clang

---
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 23 ++++++++++++++-----
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index c841786b8..e341e2e9b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -11,6 +11,17 @@
 #ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H
 #define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H
 
+// clang is incompatible with the CUDA syntax wrt making a kernel a class friend,
+// so we'll use a macro to make clang happy.
+#ifndef KERNEL_FRIEND
+#if defined(__clang__) && defined(__CUDA__)
+#define KERNEL_FRIEND friend __global__
+#else
+#define KERNEL_FRIEND friend
+#endif
+#endif
+
+
 namespace Eigen {
 
 
@@ -681,15 +692,15 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
 #endif
 #if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
-  template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
+  template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
 #ifdef EIGEN_HAS_CUDA_FP16
-  template <typename S, typename R, typename I> friend void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
-  template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
-  template <int NPT, typename S, typename R, typename I> friend void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
+  template <typename S, typename R, typename I> KERNEL_FRIEND void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
+  template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
+  template <int NPT, typename S, typename R, typename I> KERNEL_FRIEND void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
 #endif
-  template <int NPT, typename S, typename R, typename I> friend void internal::InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
+  template <int NPT, typename S, typename R, typename I> KERNEL_FRIEND void internal::InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
 
-  template <int NPT, typename S, typename R, typename I> friend void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
+  template <int NPT, typename S, typename R, typename I> KERNEL_FRIEND void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
 #endif
 
 #if defined(EIGEN_USE_SYCL)

From 61160a21d23880749bce3b0d630d9880f70af6e5 Mon Sep 17 00:00:00 2001
From: Benoit Jacob <benoitjacob@google.com>
Date: Wed, 15 Mar 2017 06:57:25 -0400
Subject: [PATCH 072/680] ARM prefetch fixes: Implement prefetch on ARM64. Do
 not clobber cc on ARM32.

---
 Eigen/src/Core/arch/NEON/PacketMath.h | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 84a56bdcc..aede4a6d5 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -51,14 +51,17 @@ typedef uint32x4_t  Packet4ui;
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   const Packet4i p4i_##NAME = pset1<Packet4i>(X)
 
-// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function
-// which available on LLVM and GCC (at least)
-#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+#if EIGEN_ARCH_ARM64
+  // __builtin_prefetch tends to do nothing on ARM64 compilers because the
+  // prefetch instructions there are too detailed for __builtin_prefetch to map
+  // meaningfully to them.
+  #define EIGEN_ARM_PREFETCH(ADDR)  __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : );
+#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
   #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR);
 #elif defined __pld
   #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR)
-#elif !EIGEN_ARCH_ARM64
-  #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( "   pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#elif EIGEN_ARCH_ARM32
+  #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : );
 #else
   // by default no explicit prefetching
   #define EIGEN_ARM_PREFETCH(ADDR)

From 89fd0c38812b024734eeacf9c23ed3714c8b0f93 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 15 Mar 2017 15:18:03 +0100
Subject: [PATCH 073/680] better check array index before using it

---
 Eigen/src/OrderingMethods/Eigen_Colamd.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/OrderingMethods/Eigen_Colamd.h b/Eigen/src/OrderingMethods/Eigen_Colamd.h
index 933cd564b..da85b4d6e 100644
--- a/Eigen/src/OrderingMethods/Eigen_Colamd.h
+++ b/Eigen/src/OrderingMethods/Eigen_Colamd.h
@@ -1004,7 +1004,7 @@ static IndexType find_ordering /* return the number of garbage collections */
     COLAMD_ASSERT (head [min_score] >= COLAMD_EMPTY) ;
 
     /* get pivot column from head of minimum degree list */
-    while (head [min_score] == COLAMD_EMPTY && min_score < n_col)
+    while (min_score < n_col && head [min_score] == COLAMD_EMPTY)
     {
       min_score++ ;
     }

From c06861d15e16a8f6241d4e409ef0136f3277350f Mon Sep 17 00:00:00 2001
From: Luke Iwanski <luke@codeplay.com>
Date: Wed, 15 Mar 2017 19:26:08 +0000
Subject: [PATCH 074/680] Fixes bug in get_sycl_supported_devices() that was
 reporting unsupported Intel CPU on AMD platform - causing timeouts in that
 configuration

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h     | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index ed21d7b56..e9c3dc0a0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -76,13 +76,16 @@ EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device
   std::vector<cl::sycl::device>::iterator it =devices.begin();
   while(it!=devices.end()) {
     ///FIXME: Currently there is a bug in amd cpu OpenCL
-    auto s=  (*it).template get_info<cl::sycl::info::device::name>();
-    std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-    if((*it).is_cpu() && s.find("amd")!=std::string::npos && s.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
-      it=devices.erase(it);
+    auto name = (*it).template get_info<cl::sycl::info::device::name>();
+    std::transform(name.begin(), name.end(), name.begin(), ::tolower);
+    auto vendor = (*it).template get_info<cl::sycl::info::device::vendor>();
+    std::transform(vendor.begin(), vendor.end(), vendor.begin(), ::tolower);
+
+    if((*it).is_cpu() && vendor.find("amd")!=std::string::npos && vendor.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
+      it = devices.erase(it);
       //FIXME: currently there is a bug in intel gpu driver regarding memory allignment issue.
-    }else if((*it).is_gpu() && s.find("intel")!=std::string::npos){
-      it=devices.erase(it);
+    }else if((*it).is_gpu() && name.find("intel")!=std::string::npos){
+      it = devices.erase(it);
     }
     else{
       ++it;

From 9597d6f6aba6091f986fbe2348106dc96a24e34e Mon Sep 17 00:00:00 2001
From: Luke Iwanski <luke@codeplay.com>
Date: Wed, 15 Mar 2017 19:28:09 +0000
Subject: [PATCH 075/680] Temporary: Disables cxx11_tensor_argmax_sycl test
 since it is causing zombie thread

---
 unsupported/test/cxx11_tensor_argmax_sycl.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_argmax_sycl.cpp b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
index 9b22f1eca..b80cb8f06 100644
--- a/unsupported/test/cxx11_tensor_argmax_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
@@ -242,7 +242,9 @@ template<typename DataType, typename Device_Selector> void sycl_argmax_test_per_
 }
 
 void test_cxx11_tensor_argmax_sycl() {
- for (const auto& device :Eigen::get_sycl_supported_devices()) {
+// TODO {lukei}: re-enable once fixed
+/* for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_argmax_test_per_device<double>(device));
   }
+*/
 }

From fd7db52f9b3b1227719c6d8281ff5e5198aaae82 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 15 Mar 2017 20:02:39 -0700
Subject: [PATCH 076/680] Silenced compilation warning

---
 unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index f335edf7d..c46a778b5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -50,6 +50,7 @@ template <DenseIndex DimId>
 struct DimensionId
 {
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) {
+    EIGEN_UNUSED_VARIABLE(dim);
     eigen_assert(dim == DimId);
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const {

From aae19c70ac273a2d40daf18a3cd15c0b0075662b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 17 Mar 2017 17:33:15 +0100
Subject: [PATCH 077/680] update has_ReturnType to be more consistent with
 other has_ helpers

---
 Eigen/src/Core/util/Meta.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 90eda6e70..8de605500 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -433,10 +433,10 @@ struct meta_no  { char a[2]; };
 template <typename T>
 struct has_ReturnType
 {
-  template <typename C> static meta_yes testFunctor(typename C::ReturnType const *);
-  template <typename C> static meta_no testFunctor(...);
+  template <typename C> static meta_yes testFunctor(C const *, typename C::ReturnType const * = 0);
+  template <typename C> static meta_no  testFunctor(...);
 
-  enum { value = sizeof(testFunctor<T>(0)) == sizeof(meta_yes) };
+  enum { value = sizeof(testFunctor<T>(static_cast<T*>(0))) == sizeof(meta_yes) };
 };
 
 template<typename T> const T* return_ptr();

From a91417a7a5a210f424b8cfec4b2bc1e00aa340be Mon Sep 17 00:00:00 2001
From: Luke Iwanski <luke@codeplay.com>
Date: Mon, 20 Mar 2017 14:48:54 +0000
Subject: [PATCH 078/680] Introduces align allocator for SYCL buffer

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 30 +++++++++++++------
 1 file changed, 21 insertions(+), 9 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index e9c3dc0a0..c5142b7c9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -15,6 +15,17 @@
 #if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H)
 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
 
+template <typename Scalar, size_t Align = EIGEN_MAX_ALIGN_BYTES, class Allocator = std::allocator<Scalar>>
+struct SyclAllocator {
+  typedef Scalar value_type;
+  typedef typename std::allocator_traits<Allocator>::pointer pointer;
+  typedef typename std::allocator_traits<Allocator>::size_type size_type;
+
+  SyclAllocator( ){};
+  Scalar* allocate(std::size_t elements) { return static_cast<Scalar*>(aligned_alloc(Align, elements)); }
+  void deallocate(Scalar * p, std::size_t size) { EIGEN_UNUSED_VARIABLE(size); free(p); }
+};
+
 namespace Eigen {
 
   #define ConvertToActualTypeSycl(Scalar, buf_acc) reinterpret_cast<typename cl::sycl::global_ptr<Scalar>::pointer_t>((&(*buf_acc.get_pointer())))
@@ -56,11 +67,11 @@ template<typename AccType>
   };
 
 struct memsetCghFunctor{
-  cl::sycl::buffer<uint8_t, 1>& m_buf;
+  cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& m_buf;
   const ptrdiff_t& buff_offset;
   const size_t& rng , GRange, tileSize;
   const int  &c;
-  memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
+  memsetCghFunctor(cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
   :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
 
   void operator()(cl::sycl::handler &cgh) const {
@@ -124,6 +135,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
 }))
 #endif
   {}
+
   /// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
   /// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
   /// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
@@ -131,10 +143,10 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   /// The device pointer would be deleted by calling deallocate function.
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
     std::lock_guard<std::mutex> lock(mutex_);
-    auto buf = cl::sycl::buffer<uint8_t,1>(cl::sycl::range<1>(num_bytes));
+    auto buf = cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >(cl::sycl::range<1>(num_bytes));
     auto ptr =buf.get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>().get_pointer();
     buf.set_final_data(nullptr);
-    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(static_cast<const uint8_t*>(ptr),buf));
+    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >(static_cast<const uint8_t*>(ptr),buf));
     return static_cast<void*>(ptr);
   }
 
@@ -235,7 +247,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   }
 
   /// Accessing the created sycl device buffer for the device pointer
-  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
+  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& get_sycl_buffer(const void * ptr) const {
     return find_buffer(ptr)->second;
   }
 
@@ -380,18 +392,18 @@ private:
   /// std::map is the container used to make sure that we create only one buffer
   /// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice.
   /// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it.
-  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
+  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > > buffer_map;
   /// sycl queue
   mutable cl::sycl::queue m_queue;
 
-  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
+  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >::iterator find_buffer(const void* ptr) const {
     std::lock_guard<std::mutex> lock(mutex_);
     auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
     if (it1 != buffer_map.end()){
       return it1;
     }
     else{
-      for(std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
+      for(std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
         auto size = it->second.get_size();
         if((it->first <  (static_cast<const uint8_t*>(ptr))) && ((static_cast<const uint8_t*>(ptr)) < (it->first + size)) ) return it;
       }
@@ -416,7 +428,7 @@ struct SyclDevice {
   }
 
   /// Accessing the created sycl device buffer for the device pointer
-  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
+  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& get_sycl_buffer(const void * ptr) const {
     return m_queue_stream->get_sycl_buffer(ptr);
   }
 

From 511810797e35471568091c44e947329f0060698e Mon Sep 17 00:00:00 2001
From: Simon Praetorius <simon.praetorius@tu-dresden.de>
Date: Fri, 24 Mar 2017 17:45:56 +0100
Subject: [PATCH 079/680] Issue with mpreal and std::numeric_limits, i.e.
 digits is not a constant. Added a digits() traits in NumTraits with fallback
 to static constant. Specialization for mpreal added in MPRealSupport.

---
 Eigen/src/Core/NumTraits.h      | 34 +++++++++++++++++++++++++++++++++
 Eigen/src/Core/StableNorm.h     |  2 +-
 unsupported/Eigen/MPRealSupport |  3 +++
 3 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h
index aebc0c259..4d896a098 100644
--- a/Eigen/src/Core/NumTraits.h
+++ b/Eigen/src/Core/NumTraits.h
@@ -41,6 +41,34 @@ struct default_digits10_impl<T,false,true> // Integer
   static int run() { return 0; }
 };
 
+
+// default implementation of digits(), based on numeric_limits if specialized,
+// 0 for integer types, and log2(epsilon()) otherwise.
+template< typename T,
+          bool use_numeric_limits = std::numeric_limits<T>::is_specialized,
+          bool is_integer = NumTraits<T>::IsInteger>
+struct default_digits_impl
+{
+  static int run() { return std::numeric_limits<T>::digits; }
+};
+
+template<typename T>
+struct default_digits_impl<T,false,false> // Floating point
+{
+  static int run() {
+    using std::log;
+    using std::ceil;
+    typedef typename NumTraits<T>::Real Real;
+    return int(ceil(-log(NumTraits<Real>::epsilon())/log(static_cast<Real>(2))));
+  }
+};
+
+template<typename T>
+struct default_digits_impl<T,false,true> // Integer
+{
+  static int run() { return 0; }
+};
+
 } // end namespace internal
 
 /** \class NumTraits
@@ -118,6 +146,12 @@ template<typename T> struct GenericNumTraits
     return internal::default_digits10_impl<T>::run();
   }
 
+  EIGEN_DEVICE_FUNC
+  static inline int digits()
+  {
+    return internal::default_digits_impl<T>::run();
+  }
+
   EIGEN_DEVICE_FUNC
   static inline Real dummy_precision()
   {
diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h
index be04ed44d..4ea598ba8 100644
--- a/Eigen/src/Core/StableNorm.h
+++ b/Eigen/src/Core/StableNorm.h
@@ -74,7 +74,7 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
     // are used. For any specific computer, each of the assignment
     // statements can be replaced
     ibeta = std::numeric_limits<RealScalar>::radix;                 // base for floating-point numbers
-    it    = std::numeric_limits<RealScalar>::digits;                // number of base-beta digits in mantissa
+    it    = NumTraits<RealScalar>::digits();                        // number of base-beta digits in mantissa
     iemin = std::numeric_limits<RealScalar>::min_exponent;          // minimum exponent
     iemax = std::numeric_limits<RealScalar>::max_exponent;          // maximum exponent
     rbig  = (std::numeric_limits<RealScalar>::max)();               // largest floating-point number
diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport
index 7f0b70c63..b770abc4a 100644
--- a/unsupported/Eigen/MPRealSupport
+++ b/unsupported/Eigen/MPRealSupport
@@ -90,6 +90,9 @@ int main()
 #ifdef MPREAL_HAVE_DYNAMIC_STD_NUMERIC_LIMITS
     static inline int digits10 (long Precision = mpfr::mpreal::get_default_prec())  { return std::numeric_limits<Real>::digits10(Precision); }
     static inline int digits10 (const Real& x)                                      { return std::numeric_limits<Real>::digits10(x); }
+    
+    static inline int digits ()               { return std::numeric_limits<Real>::digits(); }
+    static inline int digits (const Real& x)  { return std::numeric_limits<Real>::digits(x); }
 #endif
 
     static inline Real dummy_precision()

From bd64ee8555559ee13f02f2921594b4bd224f9d00 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 28 Mar 2017 16:50:34 +0100
Subject: [PATCH 080/680] Fixing TensorArgMaxSycl.h; Removing warning related
 to the hardcoded type of dims to be int in Argmax.

---
 .../Eigen/CXX11/src/Tensor/TensorArgMax.h     | 10 +++++-----
 .../Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h | 20 +++++++++----------
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  8 ++++----
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  2 +-
 .../src/Tensor/TensorSyclExtractFunctors.h    | 10 +++++-----
 unsupported/test/cxx11_tensor_argmax_sycl.cpp |  9 ++-------
 7 files changed, 28 insertions(+), 33 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
index e81001c6e..44b79e725 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@@ -178,7 +178,7 @@ class TensorTupleReducerOp : public TensorBase<TensorTupleReducerOp<ReduceOp, Di
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerOp(const XprType& expr,
                                                           const ReduceOp& reduce_op,
-                                                          const int return_dim,
+                                                          const Index return_dim,
                                                           const Dims& reduce_dims)
       : m_xpr(expr), m_reduce_op(reduce_op), m_return_dim(return_dim), m_reduce_dims(reduce_dims) {}
 
@@ -193,12 +193,12 @@ class TensorTupleReducerOp : public TensorBase<TensorTupleReducerOp<ReduceOp, Di
   const Dims& reduce_dims() const { return m_reduce_dims; }
 
   EIGEN_DEVICE_FUNC
-  int return_dim() const { return m_return_dim; }
+  Index return_dim() const { return m_return_dim; }
 
   protected:
     typename XprType::Nested m_xpr;
     const ReduceOp m_reduce_op;
-    const int m_return_dim;
+    const Index m_return_dim;
     const Dims m_reduce_dims;
 };
 
@@ -262,7 +262,7 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
   #else // following functions are required by sycl
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TupleType* data() const { return m_impl.data(); }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int return_dim() const {return m_return_dim;}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index return_dim() const {return m_return_dim;}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const StrideDims& strides() const {return m_strides;}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride_mod() const {return m_stride_mod;}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride_div() const {return m_stride_div;}
@@ -303,7 +303,7 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
  protected:
   TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device> m_orig_impl;
   TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device> m_impl;
-  const int m_return_dim;
+  const Index m_return_dim;
   StrideDims m_strides;
   Index m_stride_mod;
   Index m_stride_div;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
index 90cbe004f..8f76b8254 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
@@ -61,9 +61,9 @@ class TensorTupleReducerDeviceOp : public TensorBase<TensorTupleReducerDeviceOp<
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerDeviceOp(XprType expr,
-                                                              const int return_dim,
-                                                              const StrideDims& strides,
-                                                              const Index& stride_mod, const Index& stride_div)
+                                                              const Index return_dim,
+                                                              const StrideDims strides,
+                                                              const Index stride_mod, const Index stride_div)
           :m_xpr(expr), m_return_dim(return_dim), m_strides(strides), m_stride_mod(stride_mod), m_stride_div(stride_div) {}
 
   EIGEN_DEVICE_FUNC
@@ -71,7 +71,7 @@ class TensorTupleReducerDeviceOp : public TensorBase<TensorTupleReducerDeviceOp<
   expression() const { return m_xpr; }
 
   EIGEN_DEVICE_FUNC
-  int return_dim() const { return m_return_dim; }
+  Index return_dim() const { return m_return_dim; }
 
   EIGEN_DEVICE_FUNC
   const StrideDims& strides() const { return m_strides; }
@@ -86,8 +86,8 @@ class TensorTupleReducerDeviceOp : public TensorBase<TensorTupleReducerDeviceOp<
     typename Eigen::internal::remove_all<typename
     XprType::Nested
     >::type m_xpr;
-    const int m_return_dim;
-    const StrideDims& m_strides;
+    const Index m_return_dim;
+    const StrideDims m_strides;
     const Index m_stride_mod;
     const Index m_stride_div;
 };
@@ -137,10 +137,10 @@ typedef typename MakeGlobalPointer<typename TensorEvaluator<ArgType , SyclKernel
 
 protected:
  TensorEvaluator<ArgType , SyclKernelDevice> m_impl;
- const int m_return_dim;
- const StrideDims& m_strides;
- const Index& m_stride_mod;
- const Index& m_stride_div;
+ const Index m_return_dim;
+ const StrideDims m_strides;
+ const Index m_stride_mod;
+ const Index m_stride_div;
 };
 } // end namespace Eigen
 #endif //UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_ARGMAX_SYCL_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index fbe340820..5b1235826 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -619,7 +619,7 @@ class TensorBase<Derived, ReadOnlyAccessors>
       const array<Index, NumDimensions>, const Derived>
     argmax() const {
       array<Index, NumDimensions> in_dims;
-      for (int d = 0; d < NumDimensions; ++d) in_dims[d] = d;
+      for (Index d = 0; d < NumDimensions; ++d) in_dims[d] = d;
       return TensorTupleReducerOp<
         internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
         const array<Index, NumDimensions>,
@@ -632,7 +632,7 @@ class TensorBase<Derived, ReadOnlyAccessors>
       const array<Index, NumDimensions>, const Derived>
     argmin() const {
       array<Index, NumDimensions> in_dims;
-      for (int d = 0; d < NumDimensions; ++d) in_dims[d] = d;
+      for (Index d = 0; d < NumDimensions; ++d) in_dims[d] = d;
       return TensorTupleReducerOp<
         internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
         const array<Index, NumDimensions>,
@@ -643,7 +643,7 @@ class TensorBase<Derived, ReadOnlyAccessors>
     const TensorTupleReducerOp<
       internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
       const array<Index, 1>, const Derived>
-    argmax(const int return_dim) const {
+    argmax(const Index return_dim) const {
       array<Index, 1> in_dims;
       in_dims[0] = return_dim;
       return TensorTupleReducerOp<
@@ -656,7 +656,7 @@ class TensorBase<Derived, ReadOnlyAccessors>
     const TensorTupleReducerOp<
       internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
       const array<Index, 1>, const Derived>
-    argmin(const int return_dim) const {
+    argmin(const Index return_dim) const {
       array<Index, 1> in_dims;
       in_dims[0] = return_dim;
       return TensorTupleReducerOp<
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index cead2eac8..1a105165d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -268,7 +268,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   TensorEvaluator<ArgType, Device> m_impl;
   // required by sycl
-  const PatchDim& patch_dims;
+  const PatchDim patch_dims;
 };
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index e341e2e9b..597f3f9ae 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -791,7 +791,7 @@ static const bool RunningOnGPU = false;
   typename MakePointer_<CoeffReturnType>::Type m_result;
 
   const Device& m_device;
-  const Dims& m_xpr_dims;
+  const Dims m_xpr_dims;
 };
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index 942e9d307..a7905706d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -230,15 +230,15 @@ template<typename ReduceOp, typename Dims, typename ArgType, typename Device>\
  typedef typename Evaluator::Index Index;\
  typedef typename Eigen::internal::conditional<NumOutputDims==0, DSizes<Index, 1>, typename Evaluator::Dimensions >::type Dimensions;\
  const Dimensions m_dimensions;\
- const int m_return_dim;\
+ const Index m_return_dim;\
  const StrideDims m_strides;\
  const Index m_stride_mod;\
  const Index m_stride_div;\
  EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }\
- EIGEN_STRONG_INLINE  int return_dim() const {return m_return_dim;}\
- EIGEN_STRONG_INLINE const StrideDims& strides() const {return m_strides;}\
- EIGEN_STRONG_INLINE const Index& stride_mod() const {return m_stride_mod;}\
- EIGEN_STRONG_INLINE const Index& stride_div() const {return m_stride_div;}\
+ EIGEN_STRONG_INLINE  Index return_dim() const {return m_return_dim;}\
+ EIGEN_STRONG_INLINE const StrideDims strides() const {return m_strides;}\
+ EIGEN_STRONG_INLINE const Index stride_mod() const {return m_stride_mod;}\
+ EIGEN_STRONG_INLINE const Index stride_div() const {return m_stride_div;}\
  FunctorExtractor(const TensorEvaluator<CVQual TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device>& expr)\
  : m_dimensions(DimConstr<Dimensions, NumOutputDims>::getDim(expr.dimensions())), m_return_dim(expr.return_dim()),\
    m_strides(expr.strides()), m_stride_mod(expr.stride_mod()), m_stride_div(expr.stride_div()){}\
diff --git a/unsupported/test/cxx11_tensor_argmax_sycl.cpp b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
index b80cb8f06..521a7f82c 100644
--- a/unsupported/test/cxx11_tensor_argmax_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
@@ -143,10 +143,6 @@ static void test_sycl_argmax_dim(const Eigen::SyclDevice &sycl_device)
   }
 }
 
-
-
-
-
 template <typename DataType, int DataLayout, typename DenseIndex>
 static void test_sycl_argmin_dim(const Eigen::SyclDevice &sycl_device)
 {
@@ -242,9 +238,8 @@ template<typename DataType, typename Device_Selector> void sycl_argmax_test_per_
 }
 
 void test_cxx11_tensor_argmax_sycl() {
-// TODO {lukei}: re-enable once fixed
-/* for (const auto& device :Eigen::get_sycl_supported_devices()) {
+ for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_argmax_test_per_device<double>(device));
   }
-*/
+
 }

From 73fcaa319fcd12328e0577042862db471488fd5c Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 31 Mar 2017 08:22:25 -0700
Subject: [PATCH 081/680] Gate the sycl specific code under #ifdef sycl

---
 unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h    |  9 +++++++--
 unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h  | 10 ++++++----
 unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h |  8 ++++----
 3 files changed, 17 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
index 44b79e725..85facfb64 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@@ -228,7 +228,11 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_orig_impl(op.expression(), device),
         m_impl(op.expression().index_tuples().reduce(op.reduce_dims(), op.reduce_op()), device),
-        m_return_dim(op.return_dim()), m_device(device) {
+        m_return_dim(op.return_dim())
+#ifdef EIGEN_USE_SYCL
+       ,m_device(device)
+#endif
+  {
 
     gen_strides(m_orig_impl.dimensions(), m_strides);
     if (Layout == static_cast<int>(ColMajor)) {
@@ -307,8 +311,9 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
   StrideDims m_strides;
   Index m_stride_mod;
   Index m_stride_div;
- // required by sycl
+#ifdef EIGEN_USE_SYCL
   const Device& m_device;
+#endif
 };
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index c72d79435..7e4c129bb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -140,8 +140,9 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
 
   EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
 
-  /// used by sycl in order to build the sycl buffer
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const{return m_device;}
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
+#endif
 
  protected:
   EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
@@ -298,8 +299,9 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
 
   EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
 
-  /// used by sycl in order to build the sycl buffer
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const{return m_device;}
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
+#endif
 
  protected:
   EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index ca87f493a..220bd7ad0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -154,11 +154,10 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
-  /// required by sycl in order to extract the accessor
+#ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_argImpl; }
-  /// required by sycl in order to extract the accessor
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Generator& functor() const { return m_generator; }
-
+#endif
 
  protected:
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -183,8 +182,9 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   Dimensions m_dimensions;
   array<Index, NumDims> m_strides;
   Generator m_generator;
-  // required by sycl
+#ifdef EIGEN_USE_SYCL
   TensorEvaluator<ArgType, Device> m_argImpl;
+#endif
 };
 
 } // end namespace Eigen

From e2d5d4e7b3ac51152e26de2d00eabb26ee2a4454 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 31 Mar 2017 08:26:13 -0700
Subject: [PATCH 082/680] Restore the old constructors to retain compatibility
 with non c++11 compilers.

---
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h     | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 2fb6b84b9..0ac697087 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -70,8 +70,12 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            PaddingType padding_type, Scalar padding_value)
-      : TensorImagePatchOp(expr, patch_rows, patch_cols, row_strides,col_strides, in_row_strides, in_col_strides, row_inflate_strides,
-         col_inflate_strides, 0,0,0,0,padding_value, padding_type, false ){}
+        : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+        m_row_strides(row_strides), m_col_strides(col_strides),
+        m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+        m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+        m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+        m_padding_type(padding_type), m_padding_value(padding_value) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex row_strides, DenseIndex col_strides,
@@ -79,15 +83,14 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
-                                                           Scalar padding_value, PaddingType padding_type=PADDING_VALID,
-                                                           bool padding_explicit=true)
+                                                           Scalar padding_value)
       : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(padding_explicit), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+        m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
         m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_rows() const { return m_patch_rows; }

From c1b3d5ecb6c7ee2d52260507826955c8adee3bf6 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 31 Mar 2017 08:31:28 -0700
Subject: [PATCH 083/680] Restored code compatibility with compilers that dont
 support c++11 Gated more sycl code under #ifdef sycl

---
 .../Eigen/CXX11/src/Tensor/TensorInflation.h     |  9 ++++-----
 unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorVolumePatch.h   | 16 ++++++++++------
 3 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
index b6bf05fed..af6ecf5f4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
@@ -215,11 +215,10 @@ struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
-  /// required by sycl in order to extract the accessor
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  /// required by sycl in order to extract the accessor
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE  const Strides& functor() const { return m_strides; }
-
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Strides& functor() const { return m_strides; }
+#endif
 
  protected:
   Dimensions m_dimensions;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 1a105165d..415021510 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -255,10 +255,10 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
-  /// required by sycl in order to extract the accessor
+#ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  /// required by sycl in order to extract the accessor
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PatchDim& functor() const { return patch_dims; }
+#endif
 
  protected:
   Dimensions m_dimensions;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index 64474ee80..33eb1b297 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -64,9 +64,13 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                             DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
                                                             DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                             DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
-                                                            PaddingType padding_type, Scalar padding_value)
-      : TensorVolumePatchOp(expr, patch_planes, patch_rows, patch_cols, plane_strides, row_strides, col_strides, in_plane_strides, in_row_strides, in_col_strides,
-      plane_inflate_strides, row_inflate_strides, col_inflate_strides, 0,0,0,0,0,0,padding_value, padding_type, false) {}
+                                                            PaddingType padding_type, Scalar padding_value) :
+        m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+        m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+        m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+        m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+        m_padding_explicit(false), m_padding_top_z(0), m_padding_bottom_z(0), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+        m_padding_type(padding_type), m_padding_value(padding_value) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
@@ -75,14 +79,14 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                            DenseIndex padding_top_z, DenseIndex padding_bottom_z,
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
-                                                           Scalar padding_value, PaddingType padding_type=PADDING_VALID, bool padding_explicit=true)
+                                                           Scalar padding_value)
       : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(padding_explicit), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+        m_padding_explicit(true), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
         m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_planes() const { return m_patch_planes; }

From 4910630c968f9803bef5572bf7f17767c2ef09f1 Mon Sep 17 00:00:00 2001
From: Gagan Goel <gagan.nith@gmail.com>
Date: Fri, 31 Mar 2017 20:32:16 -0400
Subject: [PATCH 084/680] fix typos in the Tensor readme

---
 unsupported/Eigen/CXX11/src/Tensor/README.md   | 18 +++++++++---------
 .../Eigen/CXX11/src/Tensor/TensorFixedSize.h   |  2 +-
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index 38cdb9c69..8d989f13a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -83,8 +83,8 @@ large enough to hold all the data.
 
     // You can also map fixed-size tensors.  Here we get a 1d view of
     // the 2d fixed-size tensor.
-    Tensor<float, Sizes<4, 5>> t_4x3;
-    TensorMap<Tensor<float, 1>> t_12(t_4x3, 12);
+    TensorFixedSize<float, Sizes<4, 5>> t_4x3;
+    TensorMap<Tensor<float, 1>> t_12(t_4x3.data(), 12);
 
 
 #### Class TensorRef
@@ -272,7 +272,7 @@ Operation to a TensorFixedSize instead of a Tensor, which is a bit more
 efficient.
 
     // We know that the result is a 4x4x2 tensor!
-    TensorFixedSize<float, 4, 4, 2> result = t5;
+    TensorFixedSize<float, Sizes<4, 4, 2>> result = t5;
 
 Simiarly, assigning an expression to a TensorMap causes its evaluation.  Like
 tensors of type TensorFixedSize, TensorMaps cannot be resized so they have to
@@ -296,7 +296,7 @@ the expression in a temporary Tensor of the right size.  The code above in
 effect does:
 
     // .eval() knows the size!
-    TensorFixedSize<float, 4, 4, 2> tmp = t1 + t2;
+    TensorFixedSize<float, Sizes<4, 4, 2>> tmp = t1 + t2;
     Tensor<float, 3> result = (tmp * 0.2f).exp();
 
 Note that the return value of ```eval()``` is itself an Operation, so the
@@ -567,11 +567,11 @@ to the rank of the tensor. The content of the tensor is not initialized.
 
 ### TensorFixedSize
 
-Creates a tensor of the specified size. The number of arguments in the Size<>
+Creates a tensor of the specified size. The number of arguments in the Sizes<>
 template parameter determines the rank of the tensor. The content of the tensor
 is not initialized.
 
-    Eigen::TensorFixedSize<float, Size<3, 4>> a;
+    Eigen::TensorFixedSize<float, Sizes<3, 4>> a;
     cout << "Rank: " << a.rank() << endl;
     => Rank: 2
     cout << "NumRows: " << a.dimension(0) << " NumCols: " << a.dimension(1) << endl;
@@ -584,11 +584,11 @@ until the TensorMap is discarded, and the size of the data must be large enough
 to accomodate of the coefficients of the tensor.
 
     float data[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
-    Eigen::TensorMap<float, 2> a(data, 3, 4);
+    Eigen::TensorMap<Tensor<float, 2>> a(data, 3, 4);
     cout << "NumRows: " << a.dimension(0) << " NumCols: " << a.dimension(1) << endl;
     => NumRows: 3 NumCols: 4
     cout << "a(1, 2): " << a(1, 2) << endl;
-    => a(1, 2): 9
+    => a(1, 2): 7
 
 
 ## Contents Initialization
@@ -1314,7 +1314,7 @@ The previous example can be rewritten as follow:
     Eigen::Tensor<float, 2, Eigen::ColMajor> a(2, 3);
     a.setValues({{0.0f, 100.0f, 200.0f}, {300.0f, 400.0f, 500.0f}});
     Eigen::array<Eigen::DenseIndex, 2> two_dim({2, 3});
-    Eigen::Tensor<float, 1, Eigen::ColMajor> b;
+    Eigen::Tensor<float, 1, Eigen::ColMajor> b(6);
     b.reshape(two_dim) = a;
     cout << "b" << endl << b << endl;
     =>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
index fcee5f60d..e943757ad 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
@@ -20,7 +20,7 @@ namespace Eigen {
   * The fixed sized equivalent of
   * Eigen::Tensor<float, 3> t(3, 5, 7);
   * is
-  * Eigen::TensorFixedSize<float, Size<3,5,7>> t;
+  * Eigen::TensorFixedSize<float, Sizes<3,5,7>> t;
   */
 
 template<typename Scalar_, typename Dimensions_, int Options_, typename IndexType>

From bc050ea9f02f54c250858d5b6d24b226a3359021 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 09:47:04 -0700
Subject: [PATCH 085/680] Fixed compilation error when sycl is enabled.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index 220bd7ad0..11ae21be9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -97,9 +97,13 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_generator(op.generator()), m_argImpl(op.expression(), device)
+      : m_generator(op.generator())
+#ifdef EIGEN_USE_SYCL
+      , m_argImpl(op.expression(), device)
+#endif
   {
-    m_dimensions = m_argImpl.dimensions();
+    TensorEvaluator<ArgType, Device> argImpl(op.expression(), device);
+    m_dimensions = argImpl.dimensions();
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       m_strides[0] = 1;

From 63840d4666f5f92fd235ef30862db06706e928b4 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 09:54:31 -0700
Subject: [PATCH 086/680] iGate the sycl specific code under a EIGEN_USE_SYCL
 define

---
 .../Eigen/CXX11/src/Tensor/TensorArgMax.h        |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h    | 16 +++++++++++-----
 2 files changed, 13 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
index 85facfb64..c0f33ba2d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@@ -119,11 +119,11 @@ struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
 
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
-  // required by sycl
+#ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const {
     return m_impl;
   }
-
+#endif
 
  protected:
   TensorEvaluator<ArgType, Device> m_impl;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 0ac697087..e1d5541bc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -70,7 +70,7 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            PaddingType padding_type, Scalar padding_value)
-        : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+      : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
@@ -172,7 +172,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device), m_op(op)
+      : m_impl(op.expression(), device)
+#ifdef EIGEN_USE_SYCL
+      , m_op(op)
+#endif
   {
     EIGEN_STATIC_ASSERT((NumDims >= 4), YOU_MADE_A_PROGRAMMING_MISTAKE);
 
@@ -423,9 +426,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  // required by sycl
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
 
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
+#endif
 
   Index rowPaddingTop() const { return m_rowPaddingTop; }
   Index colPaddingLeft() const { return m_colPaddingLeft; }
@@ -506,8 +510,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   Scalar m_paddingValue;
 
   TensorEvaluator<ArgType, Device> m_impl;
-  // required for sycl
+
+#ifdef EIGEN_USE_SYCL
   const XprType& m_op;
+#endif
 };
 
 

From e3e343390adf05e091b605bf4975c6e9f495e8a1 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 09:56:33 -0700
Subject: [PATCH 087/680] Guard the sycl specific code with a #ifdef
 EIGEN_USE_SYCL

---
 unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 415021510..6e8e7885b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -99,11 +99,14 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
  };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device), patch_dims(op.patch_dims())
+      : m_impl(op.expression(), device)
+#ifdef EIGEN_USE_SYCL
+      , m_patch_dims(op.patch_dims())
+#endif
   {
     Index num_patches = 1;
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
-  //  const PatchDim& patch_dims = op.patch_dims();
+    const PatchDim& patch_dims = op.patch_dims();
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int i = 0; i < NumDims-1; ++i) {
         m_dimensions[i] = patch_dims[i];
@@ -257,7 +260,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PatchDim& functor() const { return patch_dims; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PatchDim& functor() const { return m_patch_dims; }
 #endif
 
  protected:
@@ -267,8 +270,10 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
   array<Index, NumDims-1> m_patchStrides;
 
   TensorEvaluator<ArgType, Device> m_impl;
-  // required by sycl
-  const PatchDim patch_dims;
+
+#ifdef EIGEN_USE_SYCL
+  const PatchDim m_patch_dims;
+#endif
 };
 
 } // end namespace Eigen

From 66c63826bdd550da08e86efd31c7cf0b793b1526 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 09:59:09 -0700
Subject: [PATCH 088/680] Guard the sycl specific code with EIGEN_USE_SYCL

---
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 597f3f9ae..7356334e1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -421,7 +421,10 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   static const bool RunningFullReduction = (NumOutputDims==0);
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device), m_reducer(op.reducer()), m_result(NULL), m_device(device), m_xpr_dims(op.dims())
+      : m_impl(op.expression(), device), m_reducer(op.reducer()), m_result(NULL), m_device(device)
+#if defined(EIGEN_USE_SYCL)
+      , m_xpr_dims(op.dims())
+#endif
   {
     EIGEN_STATIC_ASSERT((NumInputDims >= NumReducedDims), YOU_MADE_A_PROGRAMMING_MISTAKE);
     EIGEN_STATIC_ASSERT((!ReducingInnerMostDims | !PreservingInnerMostDims | (NumReducedDims == NumInputDims)),
@@ -675,13 +678,12 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   }
 
   EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const { return m_result; }
-  /// required by sycl in order to extract the accessor
-  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  /// added for sycl in order to construct the buffer from the sycl device
-  const Device& device() const{return m_device;}
-  /// added for sycl in order to re-construct the reduction eval on the device for the sub-kernel
-  const Dims& xprDims() const {return m_xpr_dims;}
 
+#if defined(EIGEN_USE_SYCL)
+  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  const Device& device() const { return m_device; }
+  const Dims& xprDims() const { return m_xpr_dims; }
+#endif
 
   private:
   template <int, typename, typename> friend struct internal::GenericDimReducer;
@@ -791,7 +793,10 @@ static const bool RunningOnGPU = false;
   typename MakePointer_<CoeffReturnType>::Type m_result;
 
   const Device& m_device;
+
+#if defined(EIGEN_USE_SYCL)
   const Dims m_xpr_dims;
+#endif
 };
 
 } // end namespace Eigen

From a1304b95b7b3d667a7f04f1a75bd9427b2a59474 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 10:00:46 -0700
Subject: [PATCH 089/680] Code cleanup

---
 unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index 33eb1b297..ea404242d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -325,7 +325,6 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
     m_outputPlanesRows = m_outputPlanes * m_outputRows;
 
     // Fast representations of different variables.
-  //  printf("THis is m_otherStride: %lu\n", m_otherStride );
     m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
 
     m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);

From a5a0c8fac1d86b61501199c79d4cde36f9e22305 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 10:03:21 -0700
Subject: [PATCH 090/680] Guard sycl specific code under a EIGEN_USE_SYCL ifdef

---
 .../Eigen/CXX11/src/Tensor/TensorVolumePatch.h  | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index ea404242d..f7b28cf66 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -64,8 +64,8 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                             DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
                                                             DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                             DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
-                                                            PaddingType padding_type, Scalar padding_value) :
-        m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                            PaddingType padding_type, Scalar padding_value)
+      : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
         m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
         m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
         m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
@@ -189,7 +189,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
 #endif
 
-      : m_impl(op.expression(), device), m_op(op)
+      : m_impl(op.expression(), device)
+#ifdef EIGEN_USE_SYCL
+      , m_op(op)
+#endif
   {
     EIGEN_STATIC_ASSERT((NumDims >= 5), YOU_MADE_A_PROGRAMMING_MISTAKE);
 
@@ -510,8 +513,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  // required by sycl
+
+#ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
+#endif
 
   Index planePaddingTop() const { return m_planePaddingTop; }
   Index rowPaddingTop() const { return m_rowPaddingTop; }
@@ -607,8 +612,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   Scalar m_paddingValue;
 
   TensorEvaluator<ArgType, Device> m_impl;
-// required by sycl
+
+#ifdef EIGEN_USE_SYCL
   XprType m_op;
+#endif
 };
 
 

From c302ea7bc417ef479626266e15bff59a805e305f Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 10:05:16 -0700
Subject: [PATCH 091/680] Deleted empty line of code

---
 unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index f7b28cf66..b0f970ae6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -359,7 +359,6 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
   {
-
     // Patch index corresponding to the passed in index.
     const Index patchIndex = index / m_fastPatchStride;
 

From 068cc0970890b534d65dbc99e6b5795acbaaa801 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Apr 2017 10:09:10 -0700
Subject: [PATCH 092/680] Preserve file naming conventions

---
 unsupported/test/CMakeLists.txt                               | 4 ++--
 ...age_patchOP_sycl.cpp => cxx11_tensor_image_patch_sycl.cpp} | 0
 ...me_patchOP_sycl.cpp => cxx11_tensor_volume_patch_sycl.cpp} | 0
 3 files changed, 2 insertions(+), 2 deletions(-)
 rename unsupported/test/{cxx11_tensor_image_patchOP_sycl.cpp => cxx11_tensor_image_patch_sycl.cpp} (100%)
 rename unsupported/test/{cxx11_tensor_volume_patchOP_sycl.cpp => cxx11_tensor_volume_patch_sycl.cpp} (100%)

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 4a558f856..cdf151f15 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -171,8 +171,8 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_inflation_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_generator_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_image_patchOP_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_volume_patchOP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_image_patch_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_volume_patcP_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_argmax_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_custom_op_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)
diff --git a/unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp b/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
similarity index 100%
rename from unsupported/test/cxx11_tensor_image_patchOP_sycl.cpp
rename to unsupported/test/cxx11_tensor_image_patch_sycl.cpp
diff --git a/unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp b/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
similarity index 100%
rename from unsupported/test/cxx11_tensor_volume_patchOP_sycl.cpp
rename to unsupported/test/cxx11_tensor_volume_patch_sycl.cpp

From 0f83aeb6b22840c21c3fc2b90d3af18a65a0798a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 14 Apr 2017 10:22:12 +0200
Subject: [PATCH 093/680] Improve cmake scripts for Pastix and BLAS detection.

---
 bench/spbench/CMakeLists.txt |   27 +-
 cmake/FindBLAS.cmake         | 1503 ++++++++++++++++++++++++++++------
 cmake/FindBLASEXT.cmake      |  380 +++++++++
 cmake/FindHWLOC.cmake        |  331 ++++++++
 cmake/FindMetis.cmake        |  299 +++++--
 cmake/FindPTSCOTCH.cmake     |  423 ++++++++++
 cmake/FindPastix.cmake       |  717 +++++++++++++++-
 cmake/FindScotch.cmake       |  379 ++++++++-
 test/CMakeLists.txt          |   29 +-
 9 files changed, 3726 insertions(+), 362 deletions(-)
 create mode 100644 cmake/FindBLASEXT.cmake
 create mode 100644 cmake/FindHWLOC.cmake
 create mode 100644 cmake/FindPTSCOTCH.cmake

diff --git a/bench/spbench/CMakeLists.txt b/bench/spbench/CMakeLists.txt
index 8d53f4ae2..932735698 100644
--- a/bench/spbench/CMakeLists.txt
+++ b/bench/spbench/CMakeLists.txt
@@ -38,25 +38,32 @@ if(SUPERLU_FOUND AND BLAS_FOUND)
 endif()
 
 
-find_package(Pastix)
-find_package(Scotch)
-find_package(Metis)
-if(PASTIX_FOUND AND BLAS_FOUND)
+find_package(PASTIX QUIET COMPONENTS METIS SCOTCH)
+# check that the PASTIX found is a version without MPI
+find_path(PASTIX_pastix_nompi.h_INCLUDE_DIRS
+  NAMES pastix_nompi.h
+  HINTS ${PASTIX_INCLUDE_DIRS}
+)
+if (NOT PASTIX_pastix_nompi.h_INCLUDE_DIRS)
+  message(STATUS "A version of Pastix has been found but pastix_nompi.h does not exist in the include directory."
+                 " Because Eigen tests require a version without MPI, we disable the Pastix backend.")
+endif()
+if(PASTIX_FOUND AND PASTIX_pastix_nompi.h_INCLUDE_DIRS AND BLAS_FOUND)
   add_definitions("-DEIGEN_PASTIX_SUPPORT")
-  include_directories(${PASTIX_INCLUDES})
+  include_directories(${PASTIX_INCLUDE_DIRS_DEP})
   if(SCOTCH_FOUND)
-    include_directories(${SCOTCH_INCLUDES})
+    include_directories(${SCOTCH_INCLUDE_DIRS})
     set(PASTIX_LIBRARIES ${PASTIX_LIBRARIES} ${SCOTCH_LIBRARIES})
   elseif(METIS_FOUND)
-    include_directories(${METIS_INCLUDES})
+    include_directories(${METIS_INCLUDE_DIRS})
     set(PASTIX_LIBRARIES ${PASTIX_LIBRARIES} ${METIS_LIBRARIES})  
   endif(SCOTCH_FOUND)
-  set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES} ${ORDERING_LIBRARIES} ${BLAS_LIBRARIES})
-  set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES} ${BLAS_LIBRARIES})
+  set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES_DEP} ${ORDERING_LIBRARIES})
+  set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES_DEP})
 endif(PASTIX_FOUND AND BLAS_FOUND)
 
 if(METIS_FOUND)
-  include_directories(${METIS_INCLUDES})
+  include_directories(${METIS_INCLUDE_DIRS})
   set (SPARSE_LIBS ${SPARSE_LIBS} ${METIS_LIBRARIES})
   add_definitions("-DEIGEN_METIS_SUPPORT")
 endif(METIS_FOUND)
diff --git a/cmake/FindBLAS.cmake b/cmake/FindBLAS.cmake
index 68c4e0724..9f74b07fe 100644
--- a/cmake/FindBLAS.cmake
+++ b/cmake/FindBLAS.cmake
@@ -1,385 +1,1363 @@
-# Find BLAS library
+###
 #
-# This module finds an installed library that implements the BLAS
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2016 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find BLAS library
+# This module finds an installed fortran library that implements the BLAS
 # linear-algebra interface (see http://www.netlib.org/blas/).
-# The list of libraries searched for is mainly taken
+# The list of libraries searched for is taken
 # from the autoconf macro file, acx_blas.m4 (distributed at
 # http://ac-archive.sourceforge.net/ac-archive/acx_blas.html).
 #
 # This module sets the following variables:
 #  BLAS_FOUND - set to true if a library implementing the BLAS interface
 #    is found
-#  BLAS_INCLUDE_DIR - Directories containing the BLAS header files
-#  BLAS_DEFINITIONS - Compilation options to use BLAS
-#  BLAS_LINKER_FLAGS - Linker flags to use BLAS (excluding -l
+#  BLAS_LINKER_FLAGS - uncached list of required linker flags (excluding -l
 #    and -L).
-#  BLAS_LIBRARIES_DIR - Directories containing the BLAS libraries.
-#     May be null if BLAS_LIBRARIES contains libraries name using full path.
-#  BLAS_LIBRARIES - List of libraries to link against BLAS interface.
-#     May be null if the compiler supports auto-link (e.g. VC++).
-#  BLAS_USE_FILE - The name of the cmake module to include to compile
-#     applications or libraries using BLAS.
+#  BLAS_COMPILER_FLAGS - uncached list of required compiler flags (including -I for mkl headers).
+#  BLAS_LIBRARIES - uncached list of libraries (using full path name) to
+#    link against to use BLAS
+#  BLAS95_LIBRARIES - uncached list of libraries (using full path name)
+#    to link against to use BLAS95 interface
+#  BLAS95_FOUND - set to true if a library implementing the BLAS f95 interface
+#    is found
+#  BLA_STATIC  if set on this determines what kind of linkage we do (static)
+#  BLA_VENDOR  if set checks only the specified vendor, if not set checks
+#     all the possibilities
+#  BLAS_VENDOR_FOUND stores the BLAS vendor found 
+#  BLA_F95     if set on tries to find the f95 interfaces for BLAS/LAPACK
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DBLAS_DIR=path/to/blas):
+#  BLAS_DIR            - Where to find the base directory of blas
+#  BLAS_INCDIR         - Where to find the header files
+#  BLAS_LIBDIR         - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: BLAS_DIR, BLAS_INCDIR, BLAS_LIBDIR
+# For MKL case and if no paths are given as hints, we will try to use the MKLROOT
+# environment variable
+#  BLAS_VERBOSE Print some additional information during BLAS libraries detection
+##########
+### List of vendors (BLA_VENDOR) valid in this module
+########## List of vendors (BLA_VENDOR) valid in this module
+##  Open (for OpenBlas), Eigen (for EigenBlas), Goto, ATLAS PhiPACK,
+##  CXML, DXML, SunPerf, SCSL, SGIMATH, IBMESSL, IBMESSLMT
+##  Intel10_32 (intel mkl v10 32 bit), Intel10_64lp (intel mkl v10 64 bit,lp thread model, lp64 model),
+##  Intel10_64lp_seq (intel mkl v10 64 bit,sequential code, lp64 model),
+##  Intel( older versions of mkl 32 and 64 bit),
+##  ACML, ACML_MP, ACML_GPU, Apple, NAS, Generic
+# C/CXX should be enabled to use Intel mkl
+###
+# We handle different modes to find the dependency
 #
-# This module was modified by CGAL team:
-# - find libraries for a C++ compiler, instead of Fortran
-# - added BLAS_INCLUDE_DIR, BLAS_DEFINITIONS and BLAS_LIBRARIES_DIR
-# - removed BLAS95_LIBRARIES
+# - Detection if already installed on the system
+#   - BLAS libraries can be detected from different ways
+#     Here is the order of precedence:
+#     1) we look in cmake variable BLAS_LIBDIR or BLAS_DIR (we guess the libdirs) if defined
+#     2) we look in environment variable BLAS_LIBDIR or BLAS_DIR (we guess the libdirs) if defined
+#     3) we look in common environnment variables depending on the system (INCLUDE, C_INCLUDE_PATH, CPATH - LIB, DYLD_LIBRARY_PATH, LD_LIBRARY_PATH)
+#     4) we look in common system paths depending on the system, see for example paths contained in the following cmake variables:
+#       - CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES, CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES
+#       - CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES, CMAKE_C_IMPLICIT_LINK_DIRECTORIES
+#
+
+#=============================================================================
+# Copyright 2007-2009 Kitware, Inc.
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of CMake, substitute the full
+#  License text for the above reference.)
+
+## Some macros to print status when search for headers and libs
+# This macro informs why the _lib_to_find file has not been found
+macro(Print_Find_Library_Blas_Status _libname _lib_to_find)
+
+  # save _libname upper/lower case
+  string(TOUPPER ${_libname} LIBNAME)
+  string(TOLOWER ${_libname} libname)
+
+  # print status
+  #message(" ")
+  if(${LIBNAME}_LIBDIR)
+    message("${Yellow}${LIBNAME}_LIBDIR is defined but ${_lib_to_find}"
+      "has not been found in ${ARGN}${ColourReset}")
+  else()
+    if(${LIBNAME}_DIR)
+      message("${Yellow}${LIBNAME}_DIR is defined but ${_lib_to_find}"
+	"has not been found in ${ARGN}${ColourReset}")
+    else()
+      message("${Yellow}${_lib_to_find} not found."
+	"Nor ${LIBNAME}_DIR neither ${LIBNAME}_LIBDIR"
+	"are defined so that we look for ${_lib_to_find} in"
+	"system paths (Linux: LD_LIBRARY_PATH, Windows: LIB,"
+	"Mac: DYLD_LIBRARY_PATH,"
+	"CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES,"
+	"CMAKE_C_IMPLICIT_LINK_DIRECTORIES)${ColourReset}")
+      if(_lib_env)
+	message("${Yellow}${_lib_to_find} has not been found in"
+	  "${_lib_env}${ColourReset}")
+      endif()
+    endif()
+  endif()
+  message("${BoldYellow}Please indicate where to find ${_lib_to_find}. You have three options:\n"
+    "- Option 1: Provide the Installation directory of BLAS library with cmake option: -D${LIBNAME}_DIR=your/path/to/${libname}/\n"
+    "- Option 2: Provide the directory where to find the library with cmake option: -D${LIBNAME}_LIBDIR=your/path/to/${libname}/lib/\n"
+    "- Option 3: Update your environment variable (Linux: LD_LIBRARY_PATH, Windows: LIB, Mac: DYLD_LIBRARY_PATH)\n"
+    "- Option 4: If your library provides a PkgConfig file, make sure pkg-config finds your library${ColourReset}")
+
+endmacro()
+
+# This macro informs why the _lib_to_find file has not been found
+macro(Print_Find_Library_Blas_CheckFunc_Status _name)
+
+  # save _libname upper/lower case
+  string(TOUPPER ${_name} FUNCNAME)
+  string(TOLOWER ${_name} funcname)
+
+  # print status
+  #message(" ")
+  message("${Red}Libs have been found but check of symbol ${_name} failed "
+    "with following libraries ${ARGN}${ColourReset}")
+  message("${BoldRed}Please open your error file CMakeFiles/CMakeError.log"
+    "to figure out why it fails${ColourReset}")
+  #message(" ")
+
+endmacro()
+
+if (NOT BLAS_FOUND)
+  set(BLAS_DIR "" CACHE PATH "Installation directory of BLAS library")
+  if (NOT BLAS_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely BLAS_DIR, has been set to specify the install directory of BLAS")
+  endif()
+endif()
+
+option(BLAS_VERBOSE "Print some additional information during BLAS libraries detection" OFF)
+mark_as_advanced(BLAS_VERBOSE)
 
 include(CheckFunctionExists)
+include(CheckFortranFunctionExists)
 
+set(_blas_ORIG_CMAKE_FIND_LIBRARY_SUFFIXES ${CMAKE_FIND_LIBRARY_SUFFIXES})
 
-# This macro checks for the existence of the combination of fortran libraries
-# given by _list.  If the combination is found, this macro checks (using the
-# check_function_exists macro) whether can link against that library
-# combination using the name of a routine given by _name using the linker
-# flags given by _flags.  If the combination of libraries is found and passes
-# the link test, LIBRARIES is set to the list of complete library paths that
-# have been found and DEFINITIONS to the required definitions.
-# Otherwise, LIBRARIES is set to FALSE.
-# N.B. _prefix is the prefix applied to the names of all cached variables that
-# are generated internally and marked advanced by this macro.
-macro(check_fortran_libraries DEFINITIONS LIBRARIES _prefix _name _flags _list _path)
-  #message("DEBUG: check_fortran_libraries(${_list} in ${_path})")
+# Check the language being used
+get_property( _LANGUAGES_ GLOBAL PROPERTY ENABLED_LANGUAGES )
+if( _LANGUAGES_ MATCHES Fortran )
+  set( _CHECK_FORTRAN TRUE )
+elseif( (_LANGUAGES_ MATCHES C) OR (_LANGUAGES_ MATCHES CXX) )
+  set( _CHECK_FORTRAN FALSE )
+else()
+  if(BLAS_FIND_REQUIRED)
+    message(FATAL_ERROR "FindBLAS requires Fortran, C, or C++ to be enabled.")
+  else()
+    message(STATUS "Looking for BLAS... - NOT found (Unsupported languages)")
+    return()
+  endif()
+endif()
 
-  # Check for the existence of the libraries given by _list
-  set(_libraries_found TRUE)
-  set(_libraries_work FALSE)
-  set(${DEFINITIONS} "")
-  set(${LIBRARIES} "")
+macro(Check_Fortran_Libraries LIBRARIES _prefix _name _flags _list _thread)
+  # This macro checks for the existence of the combination of fortran libraries
+  # given by _list.  If the combination is found, this macro checks (using the
+  # Check_Fortran_Function_Exists macro) whether can link against that library
+  # combination using the name of a routine given by _name using the linker
+  # flags given by _flags.  If the combination of libraries is found and passes
+  # the link test, LIBRARIES is set to the list of complete library paths that
+  # have been found.  Otherwise, LIBRARIES is set to FALSE.
+
+  # N.B. _prefix is the prefix applied to the names of all cached variables that
+  # are generated internally and marked advanced by this macro.
+
+  set(_libdir ${ARGN})
+
+  set(_libraries_work TRUE)
+  set(${LIBRARIES})
   set(_combined_name)
+  set(ENV_MKLROOT "$ENV{MKLROOT}")
+  set(ENV_BLAS_DIR "$ENV{BLAS_DIR}")
+  set(ENV_BLAS_LIBDIR "$ENV{BLAS_LIBDIR}")
+  if (NOT _libdir)
+    if (BLAS_LIBDIR)
+      list(APPEND _libdir "${BLAS_LIBDIR}")
+    elseif (BLAS_DIR)
+      list(APPEND _libdir "${BLAS_DIR}")
+      list(APPEND _libdir "${BLAS_DIR}/lib")
+      if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
+	list(APPEND _libdir "${BLAS_DIR}/lib64")
+	list(APPEND _libdir "${BLAS_DIR}/lib/intel64")
+      else()
+	list(APPEND _libdir "${BLAS_DIR}/lib32")
+	list(APPEND _libdir "${BLAS_DIR}/lib/ia32")
+      endif()
+    elseif(ENV_BLAS_LIBDIR)
+      list(APPEND _libdir "${ENV_BLAS_LIBDIR}")
+    elseif(ENV_BLAS_DIR)
+      list(APPEND _libdir "${ENV_BLAS_DIR}")
+      list(APPEND _libdir "${ENV_BLAS_DIR}/lib")
+      if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
+	list(APPEND _libdir "${ENV_BLAS_DIR}/lib64")
+	list(APPEND _libdir "${ENV_BLAS_DIR}/lib/intel64")
+      else()
+	list(APPEND _libdir "${ENV_BLAS_DIR}/lib32")
+	list(APPEND _libdir "${ENV_BLAS_DIR}/lib/ia32")
+      endif()
+    else()
+      if (ENV_MKLROOT)
+	list(APPEND _libdir "${ENV_MKLROOT}/lib")
+	if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
+	  list(APPEND _libdir "${ENV_MKLROOT}/lib64")
+	  list(APPEND _libdir "${ENV_MKLROOT}/lib/intel64")
+	else()
+	  list(APPEND _libdir "${ENV_MKLROOT}/lib32")
+	  list(APPEND _libdir "${ENV_MKLROOT}/lib/ia32")
+	endif()
+      endif()
+      if (WIN32)
+	string(REPLACE ":" ";" _libdir2 "$ENV{LIB}")
+      elseif (APPLE)
+	string(REPLACE ":" ";" _libdir2 "$ENV{DYLD_LIBRARY_PATH}")
+      else ()
+	string(REPLACE ":" ";" _libdir2 "$ENV{LD_LIBRARY_PATH}")
+      endif ()
+      list(APPEND _libdir "${_libdir2}")
+      list(APPEND _libdir "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+      list(APPEND _libdir "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+    endif()
+  endif ()
+
+  if (BLAS_VERBOSE)
+    message("${Cyan}Try to find BLAS libraries: ${_list}")
+  endif ()
+
   foreach(_library ${_list})
     set(_combined_name ${_combined_name}_${_library})
 
-    if(_libraries_found)
-      # search first in ${_path}
-      find_library(${_prefix}_${_library}_LIBRARY
-                  NAMES ${_library}
-                  PATHS ${_path} NO_DEFAULT_PATH
-                  )
-      # if not found, search in environment variables and system
-      if ( WIN32 )
-        find_library(${_prefix}_${_library}_LIBRARY
-                    NAMES ${_library}
-                    PATHS ENV LIB
-                    )
-      elseif ( APPLE )
-        find_library(${_prefix}_${_library}_LIBRARY
-                    NAMES ${_library}
-                    PATHS /usr/local/lib /usr/lib /usr/local/lib64 /usr/lib64 ENV DYLD_LIBRARY_PATH
-                    )
+    if(_libraries_work)
+      if (BLA_STATIC)
+	if (WIN32)
+	  set(CMAKE_FIND_LIBRARY_SUFFIXES .lib ${CMAKE_FIND_LIBRARY_SUFFIXES})
+	endif ()
+	if (APPLE)
+	  set(CMAKE_FIND_LIBRARY_SUFFIXES .lib ${CMAKE_FIND_LIBRARY_SUFFIXES})
+	else ()
+	  set(CMAKE_FIND_LIBRARY_SUFFIXES .a ${CMAKE_FIND_LIBRARY_SUFFIXES})
+	endif ()
       else ()
-        find_library(${_prefix}_${_library}_LIBRARY
-                    NAMES ${_library}
-                    PATHS /usr/local/lib /usr/lib /usr/local/lib64 /usr/lib64 ENV LD_LIBRARY_PATH
-                    )
-      endif()
+	if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
+	  # for ubuntu's libblas3gf and liblapack3gf packages
+	  set(CMAKE_FIND_LIBRARY_SUFFIXES ${CMAKE_FIND_LIBRARY_SUFFIXES} .so.3gf)
+	endif ()
+      endif ()
+      find_library(${_prefix}_${_library}_LIBRARY
+	NAMES ${_library}
+	HINTS ${_libdir}
+	NO_DEFAULT_PATH
+	)
       mark_as_advanced(${_prefix}_${_library}_LIBRARY)
+      # Print status if not found
+      # -------------------------
+      if (NOT ${_prefix}_${_library}_LIBRARY AND NOT BLAS_FIND_QUIETLY AND BLAS_VERBOSE)
+	Print_Find_Library_Blas_Status(blas ${_library} ${_libdir})
+      endif ()
       set(${LIBRARIES} ${${LIBRARIES}} ${${_prefix}_${_library}_LIBRARY})
-      set(_libraries_found ${${_prefix}_${_library}_LIBRARY})
-    endif(_libraries_found)
+      set(_libraries_work ${${_prefix}_${_library}_LIBRARY})
+    endif(_libraries_work)
   endforeach(_library ${_list})
-  if(_libraries_found)
-    set(_libraries_found ${${LIBRARIES}})
-  endif()
 
-  # Test this combination of libraries with the Fortran/f2c interface.
-  # We test the Fortran interface first as it is well standardized.
-  if(_libraries_found AND NOT _libraries_work)
-    set(${DEFINITIONS}  "-D${_prefix}_USE_F2C")
-    set(${LIBRARIES}    ${_libraries_found})
-    # Some C++ linkers require the f2c library to link with Fortran libraries.
-    # I do not know which ones, thus I just add the f2c library if it is available.
-    find_package( F2C QUIET )
-    if ( F2C_FOUND )
-      set(${DEFINITIONS}  ${${DEFINITIONS}} ${F2C_DEFINITIONS})
-      set(${LIBRARIES}    ${${LIBRARIES}} ${F2C_LIBRARIES})
+  if(_libraries_work)
+    # Test this combination of libraries.
+    if (CMAKE_SYSTEM_NAME STREQUAL "Linux" AND BLA_STATIC)
+      list(INSERT ${LIBRARIES} 0 "-Wl,--start-group")
+      list(APPEND ${LIBRARIES} "-Wl,--end-group")
     endif()
-    set(CMAKE_REQUIRED_DEFINITIONS  ${${DEFINITIONS}})
-    set(CMAKE_REQUIRED_LIBRARIES    ${_flags} ${${LIBRARIES}})
-    #message("DEBUG: CMAKE_REQUIRED_DEFINITIONS = ${CMAKE_REQUIRED_DEFINITIONS}")
-    #message("DEBUG: CMAKE_REQUIRED_LIBRARIES = ${CMAKE_REQUIRED_LIBRARIES}")
-    # Check if function exists with f2c calling convention (ie a trailing underscore)
-    check_function_exists(${_name}_ ${_prefix}_${_name}_${_combined_name}_f2c_WORKS)
-    set(CMAKE_REQUIRED_DEFINITIONS} "")
-    set(CMAKE_REQUIRED_LIBRARIES    "")
-    mark_as_advanced(${_prefix}_${_name}_${_combined_name}_f2c_WORKS)
-    set(_libraries_work ${${_prefix}_${_name}_${_combined_name}_f2c_WORKS})
-  endif(_libraries_found AND NOT _libraries_work)
-
-  # If not found, test this combination of libraries with a C interface.
-  # A few implementations (ie ACML) provide a C interface. Unfortunately, there is no standard.
-  if(_libraries_found AND NOT _libraries_work)
-    set(${DEFINITIONS} "")
-    set(${LIBRARIES}   ${_libraries_found})
-    set(CMAKE_REQUIRED_DEFINITIONS "")
-    set(CMAKE_REQUIRED_LIBRARIES   ${_flags} ${${LIBRARIES}})
-    #message("DEBUG: CMAKE_REQUIRED_LIBRARIES = ${CMAKE_REQUIRED_LIBRARIES}")
-    check_function_exists(${_name} ${_prefix}_${_name}${_combined_name}_WORKS)
-    set(CMAKE_REQUIRED_LIBRARIES "")
-    mark_as_advanced(${_prefix}_${_name}${_combined_name}_WORKS)
-    set(_libraries_work ${${_prefix}_${_name}${_combined_name}_WORKS})
-  endif(_libraries_found AND NOT _libraries_work)
-
-  # on failure
-  if(NOT _libraries_work)
-    set(${DEFINITIONS} "")
-    set(${LIBRARIES}   FALSE)
+    set(CMAKE_REQUIRED_LIBRARIES "${_flags};${${LIBRARIES}};${_thread}")
+    set(CMAKE_REQUIRED_FLAGS "${BLAS_COMPILER_FLAGS}")
+    if (BLAS_VERBOSE)
+      message("${Cyan}BLAS libs found for BLA_VENDOR ${BLA_VENDOR}."
+	"Try to compile symbol ${_name} with following libraries:"
+	"${CMAKE_REQUIRED_LIBRARIES}")
+    endif ()
+    if(NOT BLAS_FOUND)
+      unset(${_prefix}${_combined_name}_WORKS CACHE)
+    endif()
+    if (_CHECK_FORTRAN)
+      if (CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
+	string(REPLACE "mkl_intel_lp64" "mkl_gf_lp64" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+	string(REPLACE "mkl_intel_ilp64" "mkl_gf_ilp64" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+      endif()
+      check_fortran_function_exists("${_name}" ${_prefix}${_combined_name}_WORKS)
+    else()
+      check_function_exists("${_name}_" ${_prefix}${_combined_name}_WORKS)
+    endif()
+    mark_as_advanced(${_prefix}${_combined_name}_WORKS)
+    set(_libraries_work ${${_prefix}${_combined_name}_WORKS})
+    # Print status if not found
+    # -------------------------
+    if (NOT _libraries_work AND NOT BLAS_FIND_QUIETLY AND BLAS_VERBOSE)
+      Print_Find_Library_Blas_CheckFunc_Status(${_name} ${CMAKE_REQUIRED_LIBRARIES})
+    endif ()
+    set(CMAKE_REQUIRED_LIBRARIES)
   endif()
-  #message("DEBUG: ${DEFINITIONS} = ${${DEFINITIONS}}")
-  #message("DEBUG: ${LIBRARIES} = ${${LIBRARIES}}")
-endmacro(check_fortran_libraries)
+
+  if(_libraries_work)
+    set(${LIBRARIES} ${${LIBRARIES}} ${_thread})
+  else(_libraries_work)
+    set(${LIBRARIES} FALSE)
+  endif(_libraries_work)
+
+endmacro(Check_Fortran_Libraries)
 
 
-#
-# main
-#
+set(BLAS_LINKER_FLAGS)
+set(BLAS_LIBRARIES)
+set(BLAS95_LIBRARIES)
+if ($ENV{BLA_VENDOR} MATCHES ".+")
+  set(BLA_VENDOR $ENV{BLA_VENDOR})
+else ()
+  if(NOT BLA_VENDOR)
+    set(BLA_VENDOR "All")
+  endif()
+endif ()
 
-# Is it already configured?
-if (BLAS_LIBRARIES_DIR OR BLAS_LIBRARIES)
+#BLAS in intel mkl 10 library? (em64t 64bit)
+if (BLA_VENDOR MATCHES "Intel*" OR BLA_VENDOR STREQUAL "All")
 
-  set(BLAS_FOUND TRUE)
+  if(NOT BLAS_LIBRARIES OR BLA_VENDOR MATCHES "Intel*")
+    # Looking for include
+    # -------------------
 
-else()
+    # Add system include paths to search include
+    # ------------------------------------------
+    unset(_inc_env)
+    set(ENV_MKLROOT "$ENV{MKLROOT}")
+    set(ENV_BLAS_DIR "$ENV{BLAS_DIR}")
+    set(ENV_BLAS_INCDIR "$ENV{BLAS_INCDIR}")
+    if(ENV_BLAS_INCDIR)
+      list(APPEND _inc_env "${ENV_BLAS_INCDIR}")
+    elseif(ENV_BLAS_DIR)
+      list(APPEND _inc_env "${ENV_BLAS_DIR}")
+      list(APPEND _inc_env "${ENV_BLAS_DIR}/include")
+    else()
+      if (ENV_MKLROOT)
+	list(APPEND _inc_env "${ENV_MKLROOT}/include")
+      endif()
+      # system variables
+      if(WIN32)
+	string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+	list(APPEND _inc_env "${_path_env}")
+      else()
+	string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+	list(APPEND _inc_env "${_path_env}")
+	string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+	list(APPEND _inc_env "${_path_env}")
+	string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+	list(APPEND _inc_env "${_path_env}")
+	string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+	list(APPEND _inc_env "${_path_env}")
+      endif()
+    endif()
+    list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+    list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+    list(REMOVE_DUPLICATES _inc_env)
 
-  # reset variables
-  set( BLAS_INCLUDE_DIR "" )
-  set( BLAS_DEFINITIONS "" )
-  set( BLAS_LINKER_FLAGS "" )
-  set( BLAS_LIBRARIES "" )
-  set( BLAS_LIBRARIES_DIR "" )
+    # set paths where to look for
+    set(PATH_TO_LOOK_FOR "${_inc_env}")
 
-    #
-    # If Unix, search for BLAS function in possible libraries
-    #
+    # Try to find the fftw header in the given paths
+    # -------------------------------------------------
+    # call cmake macro to find the header path
+    if(BLAS_INCDIR)
+      set(BLAS_mkl.h_DIRS "BLAS_mkl.h_DIRS-NOTFOUND")
+      find_path(BLAS_mkl.h_DIRS
+	NAMES mkl.h
+	HINTS ${BLAS_INCDIR})
+    else()
+      if(BLAS_DIR)
+	set(BLAS_mkl.h_DIRS "BLAS_mkl.h_DIRS-NOTFOUND")
+	find_path(BLAS_mkl.h_DIRS
+	  NAMES mkl.h
+	  HINTS ${BLAS_DIR}
+	  PATH_SUFFIXES "include")
+      else()
+	set(BLAS_mkl.h_DIRS "BLAS_mkl.h_DIRS-NOTFOUND")
+	find_path(BLAS_mkl.h_DIRS
+	  NAMES mkl.h
+	  HINTS ${PATH_TO_LOOK_FOR})
+      endif()
+    endif()
+    mark_as_advanced(BLAS_mkl.h_DIRS)
 
-    # BLAS in ATLAS library? (http://math-atlas.sourceforge.net/)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+    # If found, add path to cmake variable
+    # ------------------------------------
+    if (BLAS_mkl.h_DIRS)
+      set(BLAS_INCLUDE_DIRS "${BLAS_mkl.h_DIRS}")
+    else ()
+      set(BLAS_INCLUDE_DIRS "BLAS_INCLUDE_DIRS-NOTFOUND")
+      if(NOT BLAS_FIND_QUIETLY)
+	message(STATUS "Looking for BLAS -- mkl.h not found")
+      endif()
+    endif()
+
+    if (WIN32)
+      string(REPLACE ":" ";" _libdir "$ENV{LIB}")
+    elseif (APPLE)
+      string(REPLACE ":" ";" _libdir "$ENV{DYLD_LIBRARY_PATH}")
+    else ()
+      string(REPLACE ":" ";" _libdir "$ENV{LD_LIBRARY_PATH}")
+    endif ()
+    list(APPEND _libdir "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+    list(APPEND _libdir "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+    # libiomp5
+    # --------
+    set(OMP_iomp5_LIBRARY "OMP_iomp5_LIBRARY-NOTFOUND")
+    find_library(OMP_iomp5_LIBRARY
+      NAMES iomp5
+      HINTS ${_libdir}
+      )
+    mark_as_advanced(OMP_iomp5_LIBRARY)
+    set(OMP_LIB "")
+    # libgomp
+    # -------
+    set(OMP_gomp_LIBRARY "OMP_gomp_LIBRARY-NOTFOUND")
+    find_library(OMP_gomp_LIBRARY
+      NAMES gomp
+      HINTS ${_libdir}
+      )
+    mark_as_advanced(OMP_gomp_LIBRARY)
+    # choose one or another depending on the compilo
+    if (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+      if (OMP_gomp_LIBRARY)
+	set(OMP_LIB "${OMP_gomp_LIBRARY}")
+      endif()
+    else(CMAKE_C_COMPILER_ID STREQUAL "Intel")
+      if (OMP_iomp5_LIBRARY)
+	set(OMP_LIB "${OMP_iomp5_LIBRARY}")
+      endif()
+    endif()
+
+    if (UNIX AND NOT WIN32)
+      # m
+      find_library(M_LIBRARY
+	NAMES m
+	HINTS ${_libdir})
+      mark_as_advanced(M_LIBRARY)
+      if(M_LIBRARY)
+	set(LM "-lm")
+      else()
+	set(LM "")
+      endif()
+      # Fortran
+      set(LGFORTRAN "")
+      if (CMAKE_C_COMPILER_ID MATCHES "GNU")
+	find_library(
+	  FORTRAN_gfortran_LIBRARY
+	  NAMES gfortran
+	  HINTS ${_libdir}
+	  )
+	mark_as_advanced(FORTRAN_gfortran_LIBRARY)
+	if (FORTRAN_gfortran_LIBRARY)
+	  set(LGFORTRAN "${FORTRAN_gfortran_LIBRARY}")
+	endif()
+      elseif (CMAKE_C_COMPILER_ID MATCHES "Intel")
+	find_library(
+	  FORTRAN_ifcore_LIBRARY
+	  NAMES ifcore
+	  HINTS ${_libdir}
+	  )
+	mark_as_advanced(FORTRAN_ifcore_LIBRARY)
+	if (FORTRAN_ifcore_LIBRARY)
+	  set(LGFORTRAN "{FORTRAN_ifcore_LIBRARY}")
+	endif()
+      endif()
+      set(BLAS_COMPILER_FLAGS "")
+      if (NOT BLA_VENDOR STREQUAL "Intel10_64lp_seq")
+	if (CMAKE_C_COMPILER_ID STREQUAL "Intel")
+	  list(APPEND BLAS_COMPILER_FLAGS "-openmp")
+	endif()
+	if (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+	  list(APPEND BLAS_COMPILER_FLAGS "-fopenmp")
+	endif()
+      endif()
+      if (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+	if (BLA_VENDOR STREQUAL "Intel10_32")
+	  list(APPEND BLAS_COMPILER_FLAGS "-m32")
+	else()
+	  list(APPEND BLAS_COMPILER_FLAGS "-m64")
+	endif()
+	if (NOT BLA_VENDOR STREQUAL "Intel10_64lp_seq")
+	  list(APPEND OMP_LIB "-ldl")
+	endif()
+	if (ENV_MKLROOT)
+	  list(APPEND BLAS_COMPILER_FLAGS "-I${ENV_MKLROOT}/include")
+	endif()
+      endif()
+
+      set(additional_flags "")
+      if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND CMAKE_SYSTEM_NAME STREQUAL "Linux")
+	set(additional_flags "-Wl,--no-as-needed")
+      endif()
+    endif ()
+
+    if (_LANGUAGES_ MATCHES C OR _LANGUAGES_ MATCHES CXX)
+      if(BLAS_FIND_QUIETLY OR NOT BLAS_FIND_REQUIRED)
+	find_package(Threads)
+      else()
+	find_package(Threads REQUIRED)
+      endif()
+
+      set(BLAS_SEARCH_LIBS "")
+
+      if(BLA_F95)
+
+	set(BLAS_mkl_SEARCH_SYMBOL SGEMM)
+	set(_LIBRARIES BLAS95_LIBRARIES)
+	if (WIN32)
+	  if (BLA_STATIC)
+	    set(BLAS_mkl_DLL_SUFFIX "")
+	  else()
+	    set(BLAS_mkl_DLL_SUFFIX "_dll")
+	  endif()
+
+	  # Find the main file (32-bit or 64-bit)
+	  set(BLAS_SEARCH_LIBS_WIN_MAIN "")
+	  if (BLA_VENDOR STREQUAL "Intel10_32" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_MAIN
+	      "mkl_blas95${BLAS_mkl_DLL_SUFFIX} mkl_intel_c${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp*" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_MAIN
+	      "mkl_blas95_lp64${BLAS_mkl_DLL_SUFFIX} mkl_intel_lp64${BLAS_mkl_DLL_SUFFIX}")
+	  endif ()
+
+	  # Add threading/sequential libs
+	  set(BLAS_SEARCH_LIBS_WIN_THREAD "")
+	  if (BLA_VENDOR STREQUAL "*_seq" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "mkl_sequential${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+	  if (NOT BLA_VENDOR STREQUAL "*_seq" OR BLA_VENDOR STREQUAL "All")
+	    # old version
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "libguide40 mkl_intel_thread${BLAS_mkl_DLL_SUFFIX}")
+	    # mkl >= 10.3
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "libiomp5md mkl_intel_thread${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+
+	  # Cartesian product of the above
+	  foreach (MAIN ${BLAS_SEARCH_LIBS_WIN_MAIN})
+	    foreach (THREAD ${BLAS_SEARCH_LIBS_WIN_THREAD})
+	      list(APPEND BLAS_SEARCH_LIBS
+		"${MAIN} ${THREAD} mkl_core${BLAS_mkl_DLL_SUFFIX}")
+	    endforeach()
+	  endforeach()
+	else (WIN32)
+	  if (BLA_VENDOR STREQUAL "Intel10_32" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_blas95 mkl_intel mkl_intel_thread mkl_core guide")
+	  endif ()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp" OR BLA_VENDOR STREQUAL "All")
+	    # old version
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_blas95 mkl_intel_lp64 mkl_intel_thread mkl_core guide")
+	    # mkl >= 10.3
+	    if (CMAKE_C_COMPILER_ID STREQUAL "Intel")
+	      list(APPEND BLAS_SEARCH_LIBS
+		"mkl_blas95_lp64 mkl_intel_lp64 mkl_intel_thread mkl_core")
+	    endif()
+	    if (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+	      list(APPEND BLAS_SEARCH_LIBS
+		"mkl_blas95_lp64 mkl_intel_lp64 mkl_gnu_thread mkl_core")
+	    endif()
+	  endif ()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp_seq" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_intel_lp64 mkl_sequential mkl_core")
+	    if (BLA_VENDOR STREQUAL "Intel10_64lp_seq")
+	      set(OMP_LIB "")
+	    endif()
+	  endif ()
+	endif (WIN32)
+
+      else (BLA_F95)
+
+	set(BLAS_mkl_SEARCH_SYMBOL sgemm)
+	set(_LIBRARIES BLAS_LIBRARIES)
+	if (WIN32)
+	  if (BLA_STATIC)
+	    set(BLAS_mkl_DLL_SUFFIX "")
+	  else()
+	    set(BLAS_mkl_DLL_SUFFIX "_dll")
+	  endif()
+
+	  # Find the main file (32-bit or 64-bit)
+	  set(BLAS_SEARCH_LIBS_WIN_MAIN "")
+	  if (BLA_VENDOR STREQUAL "Intel10_32" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_MAIN
+	      "mkl_intel_c${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp*" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_MAIN
+	      "mkl_intel_lp64${BLAS_mkl_DLL_SUFFIX}")
+	  endif ()
+
+	  # Add threading/sequential libs
+	  set(BLAS_SEARCH_LIBS_WIN_THREAD "")
+	  if (NOT BLA_VENDOR STREQUAL "*_seq" OR BLA_VENDOR STREQUAL "All")
+	    # old version
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "libguide40 mkl_intel_thread${BLAS_mkl_DLL_SUFFIX}")
+	    # mkl >= 10.3
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "libiomp5md mkl_intel_thread${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+	  if (BLA_VENDOR STREQUAL "*_seq" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS_WIN_THREAD
+	      "mkl_sequential${BLAS_mkl_DLL_SUFFIX}")
+	  endif()
+
+	  # Cartesian product of the above
+	  foreach (MAIN ${BLAS_SEARCH_LIBS_WIN_MAIN})
+	    foreach (THREAD ${BLAS_SEARCH_LIBS_WIN_THREAD})
+	      list(APPEND BLAS_SEARCH_LIBS
+		"${MAIN} ${THREAD} mkl_core${BLAS_mkl_DLL_SUFFIX}")
+	    endforeach()
+	  endforeach()
+	else (WIN32)
+	  if (BLA_VENDOR STREQUAL "Intel10_32" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_intel mkl_intel_thread mkl_core guide")
+	  endif ()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp" OR BLA_VENDOR STREQUAL "All")
+	    # old version
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_intel_lp64 mkl_intel_thread mkl_core guide")
+	    # mkl >= 10.3
+	    if (CMAKE_C_COMPILER_ID STREQUAL "Intel")
+	      list(APPEND BLAS_SEARCH_LIBS
+		"mkl_intel_lp64 mkl_intel_thread mkl_core")
+	    endif()
+	    if (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+	      list(APPEND BLAS_SEARCH_LIBS
+		"mkl_intel_lp64 mkl_gnu_thread mkl_core")
+	    endif()
+	  endif ()
+	  if (BLA_VENDOR STREQUAL "Intel10_64lp_seq" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_intel_lp64 mkl_sequential mkl_core")
+	    if (BLA_VENDOR STREQUAL "Intel10_64lp_seq")
+	      set(OMP_LIB "")
+	    endif()
+	  endif ()
+	  #older vesions of intel mkl libs
+	  if (BLA_VENDOR STREQUAL "Intel" OR BLA_VENDOR STREQUAL "All")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_ia32")
+	    list(APPEND BLAS_SEARCH_LIBS
+	      "mkl_em64t")
+	  endif ()
+	endif (WIN32)
+
+      endif (BLA_F95)
+
+      foreach (IT ${BLAS_SEARCH_LIBS})
+	string(REPLACE " " ";" SEARCH_LIBS ${IT})
+	if (${_LIBRARIES})
+	else ()
+	  check_fortran_libraries(
+	    ${_LIBRARIES}
+	    BLAS
+	    ${BLAS_mkl_SEARCH_SYMBOL}
+	    "${additional_flags}"
+	    "${SEARCH_LIBS}"
+	    "${OMP_LIB};${CMAKE_THREAD_LIBS_INIT};${LM}"
+	    )
+	  if(_LIBRARIES)
+	    set(BLAS_LINKER_FLAGS "${additional_flags}")
+	  endif()
+	endif()
+      endforeach ()
+      if(NOT BLAS_FIND_QUIETLY)
+        if(${_LIBRARIES})
+          message(STATUS "Looking for MKL BLAS: found")
+        else()
+          message(STATUS "Looking for MKL BLAS: not found")
+        endif()
+      endif()
+      if (${_LIBRARIES} AND NOT BLAS_VENDOR_FOUND)
+          set (BLAS_VENDOR_FOUND "Intel MKL")
+      endif()
+    endif (_LANGUAGES_ MATCHES C OR _LANGUAGES_ MATCHES CXX)
+  endif(NOT BLAS_LIBRARIES OR BLA_VENDOR MATCHES "Intel*")
+endif (BLA_VENDOR MATCHES "Intel*" OR BLA_VENDOR STREQUAL "All")
+
+
+if (BLA_VENDOR STREQUAL "Goto" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    # gotoblas (http://www.tacc.utexas.edu/tacc-projects/gotoblas2)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "cblas;f77blas;atlas"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      "goto2"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for Goto BLAS: found")
+      else()
+	message(STATUS "Looking for Goto BLAS: not found")
+      endif()
     endif()
+  endif()
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Goto")
+  endif()
 
-    # BLAS in PhiPACK libraries? (requires generic BLAS lib, too)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+endif (BLA_VENDOR STREQUAL "Goto" OR BLA_VENDOR STREQUAL "All")
+
+
+# OpenBlas
+if (BLA_VENDOR STREQUAL "Open" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    # openblas (http://www.openblas.net/)
+    check_fortran_libraries(
+      BLAS_LIBRARIES
+      BLAS
+      sgemm
+      ""
+      "openblas"
+      ""
+      )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for Open BLAS: found")
+      else()
+	message(STATUS "Looking for Open BLAS: not found")
+      endif()
+    endif()
+  endif()
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Openblas")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "Open" OR BLA_VENDOR STREQUAL "All")
+
+
+# EigenBlas
+if (BLA_VENDOR STREQUAL "Eigen" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    # eigenblas (http://eigen.tuxfamily.org/index.php?title=Main_Page)
+    check_fortran_libraries(
+      BLAS_LIBRARIES
+      BLAS
+      sgemm
+      ""
+      "eigen_blas"
+      ""
+      )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+	message(STATUS "Looking for Eigen BLAS: found")
+      else()
+	message(STATUS "Looking for Eigen BLAS: not found")
+      endif()
+    endif()
+  endif()
+
+  if(NOT BLAS_LIBRARIES)
+    # eigenblas (http://eigen.tuxfamily.org/index.php?title=Main_Page)
+    check_fortran_libraries(
+      BLAS_LIBRARIES
+      BLAS
+      sgemm
+      ""
+      "eigen_blas_static"
+      ""
+      )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for Eigen BLAS: found")
+      else()
+	message(STATUS "Looking for Eigen BLAS: not found")
+      endif()
+    endif()
+  endif()
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Eigen")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "Eigen" OR BLA_VENDOR STREQUAL "All")
+
+
+if (BLA_VENDOR STREQUAL "ATLAS" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    # BLAS in ATLAS library? (http://math-atlas.sourceforge.net/)
+    check_fortran_libraries(
+      BLAS_LIBRARIES
+      BLAS
+      dgemm
+      ""
+      "f77blas;atlas"
+      ""
+      )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for Atlas BLAS: found")
+      else()
+	message(STATUS "Looking for Atlas BLAS: not found")
+      endif()
+    endif()
+  endif()
+
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Atlas")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "ATLAS" OR BLA_VENDOR STREQUAL "All")
+
+
+# BLAS in PhiPACK libraries? (requires generic BLAS lib, too)
+if (BLA_VENDOR STREQUAL "PhiPACK" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
       "sgemm;dgemm;blas"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for PhiPACK BLAS: found")
+      else()
+	message(STATUS "Looking for PhiPACK BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # BLAS in Alpha CXML library?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "PhiPACK")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "PhiPACK" OR BLA_VENDOR STREQUAL "All")
+
+
+# BLAS in Alpha CXML library?
+if (BLA_VENDOR STREQUAL "CXML" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
       "cxml"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for CXML BLAS: found")
+      else()
+	message(STATUS "Looking for CXML BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # BLAS in Alpha DXML library? (now called CXML, see above)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "CXML")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "CXML" OR BLA_VENDOR STREQUAL "All")
+
+
+# BLAS in Alpha DXML library? (now called CXML, see above)
+if (BLA_VENDOR STREQUAL "DXML" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
       "dxml"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for DXML BLAS: found")
+      else()
+	message(STATUS "Looking for DXML BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # BLAS in Sun Performance library?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "DXML")
+  endif()
+  
+endif (BLA_VENDOR STREQUAL "DXML" OR BLA_VENDOR STREQUAL "All")
+
+
+# BLAS in Sun Performance library?
+if (BLA_VENDOR STREQUAL "SunPerf" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       "-xlic_lib=sunperf"
       "sunperf;sunmath"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(BLAS_LIBRARIES)
+      set(BLAS_LINKER_FLAGS "-xlic_lib=sunperf")
+    endif()
+    if(NOT BLAS_FIND_QUIETLY)
       if(BLAS_LIBRARIES)
-        # Extra linker flag
-        set(BLAS_LINKER_FLAGS "-xlic_lib=sunperf")
+	message(STATUS "Looking for SunPerf BLAS: found")
+      else()
+	message(STATUS "Looking for SunPerf BLAS: not found")
       endif()
     endif()
+  endif()
 
-    # BLAS in SCSL library?  (SGI/Cray Scientific Library)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "SunPerf")
+  endif()
+
+endif ()
+
+
+# BLAS in SCSL library?  (SGI/Cray Scientific Library)
+if (BLA_VENDOR STREQUAL "SCSL" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
       "scsl"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for SCSL BLAS: found")
+      else()
+	message(STATUS "Looking for SCSL BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # BLAS in SGIMATH library?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "SunPerf")
+  endif()
+
+endif ()
+
+
+# BLAS in SGIMATH library?
+if (BLA_VENDOR STREQUAL "SGIMATH" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
       "complib.sgimath"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for SGIMATH BLAS: found")
+      else()
+	message(STATUS "Looking for SGIMATH BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # BLAS in IBM ESSL library? (requires generic BLAS lib, too)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "SGIMATH")
+  endif()
+
+endif ()
+
+
+# BLAS in IBM ESSL library (requires generic BLAS lib, too)
+if (BLA_VENDOR STREQUAL "IBMESSL" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "essl;blas"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      "essl;xlfmath;xlf90_r;blas"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for IBM ESSL BLAS: found")
+      else()
+	message(STATUS "Looking for IBM ESSL BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    #BLAS in intel mkl 10 library? (em64t 64bit)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "IBM ESSL")
+  endif()
+
+endif ()
+
+# BLAS in IBM ESSL_MT library (requires generic BLAS lib, too)
+if (BLA_VENDOR STREQUAL "IBMESSLMT" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "mkl_intel_lp64;mkl_intel_thread;mkl_core;guide;pthread"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
-      )
-    endif()
-
-    ### windows version of intel mkl 10?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
-      BLAS_LIBRARIES
-      BLAS
-      SGEMM
+      "esslsmp;xlsmp;xlfmath;xlf90_r;blas"
       ""
-      "mkl_c_dll;mkl_intel_thread_dll;mkl_core_dll;libguide40"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for IBM ESSL MT BLAS: found")
+      else()
+	message(STATUS "Looking for IBM ESSL MT BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    #older versions of intel mkl libs
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "IBM ESSL MT")
+  endif()
 
-    # BLAS in intel mkl library? (shared)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+endif ()
+
+
+#BLAS in acml library?
+if (BLA_VENDOR MATCHES "ACML.*" OR BLA_VENDOR STREQUAL "All")
+
+  if( ((BLA_VENDOR STREQUAL "ACML") AND (NOT BLAS_ACML_LIB_DIRS)) OR
+      ((BLA_VENDOR STREQUAL "ACML_MP") AND (NOT BLAS_ACML_MP_LIB_DIRS)) OR
+      ((BLA_VENDOR STREQUAL "ACML_GPU") AND (NOT BLAS_ACML_GPU_LIB_DIRS)))
+
+    # try to find acml in "standard" paths
+    if( WIN32 )
+      file( GLOB _ACML_ROOT "C:/AMD/acml*/ACML-EULA.txt" )
+    else()
+      file( GLOB _ACML_ROOT "/opt/acml*/ACML-EULA.txt" )
+    endif()
+    if( WIN32 )
+      file( GLOB _ACML_GPU_ROOT "C:/AMD/acml*/GPGPUexamples" )
+    else()
+      file( GLOB _ACML_GPU_ROOT "/opt/acml*/GPGPUexamples" )
+    endif()
+    list(GET _ACML_ROOT 0 _ACML_ROOT)
+    list(GET _ACML_GPU_ROOT 0 _ACML_GPU_ROOT)
+
+    if( _ACML_ROOT )
+
+      get_filename_component( _ACML_ROOT ${_ACML_ROOT} PATH )
+      if( SIZEOF_INTEGER EQUAL 8 )
+	set( _ACML_PATH_SUFFIX "_int64" )
+      else()
+	set( _ACML_PATH_SUFFIX "" )
+      endif()
+      if( CMAKE_Fortran_COMPILER_ID STREQUAL "Intel" )
+	set( _ACML_COMPILER32 "ifort32" )
+	set( _ACML_COMPILER64 "ifort64" )
+      elseif( CMAKE_Fortran_COMPILER_ID STREQUAL "SunPro" )
+	set( _ACML_COMPILER32 "sun32" )
+	set( _ACML_COMPILER64 "sun64" )
+      elseif( CMAKE_Fortran_COMPILER_ID STREQUAL "PGI" )
+	set( _ACML_COMPILER32 "pgi32" )
+	if( WIN32 )
+	  set( _ACML_COMPILER64 "win64" )
+	else()
+	  set( _ACML_COMPILER64 "pgi64" )
+	endif()
+      elseif( CMAKE_Fortran_COMPILER_ID STREQUAL "Open64" )
+	# 32 bit builds not supported on Open64 but for code simplicity
+	# We'll just use the same directory twice
+	set( _ACML_COMPILER32 "open64_64" )
+	set( _ACML_COMPILER64 "open64_64" )
+      elseif( CMAKE_Fortran_COMPILER_ID STREQUAL "NAG" )
+	set( _ACML_COMPILER32 "nag32" )
+	set( _ACML_COMPILER64 "nag64" )
+      else()
+	set( _ACML_COMPILER32 "gfortran32" )
+	set( _ACML_COMPILER64 "gfortran64" )
+      endif()
+
+      if( BLA_VENDOR STREQUAL "ACML_MP" )
+	set(_ACML_MP_LIB_DIRS
+	  "${_ACML_ROOT}/${_ACML_COMPILER32}_mp${_ACML_PATH_SUFFIX}/lib"
+	  "${_ACML_ROOT}/${_ACML_COMPILER64}_mp${_ACML_PATH_SUFFIX}/lib" )
+      else()
+	set(_ACML_LIB_DIRS
+	  "${_ACML_ROOT}/${_ACML_COMPILER32}${_ACML_PATH_SUFFIX}/lib"
+	  "${_ACML_ROOT}/${_ACML_COMPILER64}${_ACML_PATH_SUFFIX}/lib" )
+      endif()
+
+    endif(_ACML_ROOT)
+
+  elseif(BLAS_${BLA_VENDOR}_LIB_DIRS)
+
+    set(_${BLA_VENDOR}_LIB_DIRS ${BLAS_${BLA_VENDOR}_LIB_DIRS})
+
+  endif()
+
+  if( BLA_VENDOR STREQUAL "ACML_MP" )
+    foreach( BLAS_ACML_MP_LIB_DIRS ${_ACML_MP_LIB_DIRS})
+      check_fortran_libraries (
+	BLAS_LIBRARIES
+	BLAS
+	sgemm
+	"" "acml_mp;acml_mv" "" ${BLAS_ACML_MP_LIB_DIRS}
+	)
+      if( BLAS_LIBRARIES )
+	break()
+      endif()
+    endforeach()
+  elseif( BLA_VENDOR STREQUAL "ACML_GPU" )
+    foreach( BLAS_ACML_GPU_LIB_DIRS ${_ACML_GPU_LIB_DIRS})
+      check_fortran_libraries (
+	BLAS_LIBRARIES
+	BLAS
+	sgemm
+	"" "acml;acml_mv;CALBLAS" "" ${BLAS_ACML_GPU_LIB_DIRS}
+	)
+      if( BLAS_LIBRARIES )
+	break()
+      endif()
+    endforeach()
+  else()
+    foreach( BLAS_ACML_LIB_DIRS ${_ACML_LIB_DIRS} )
+      check_fortran_libraries (
+	BLAS_LIBRARIES
+	BLAS
+	sgemm
+	"" "acml;acml_mv" "" ${BLAS_ACML_LIB_DIRS}
+	)
+      if( BLAS_LIBRARIES )
+	break()
+      endif()
+    endforeach()
+  endif()
+
+  # Either acml or acml_mp should be in LD_LIBRARY_PATH but not both
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "mkl;guide;pthread"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      "acml;acml_mv"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for ACML BLAS: found")
+      else()
+	message(STATUS "Looking for ACML BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    #BLAS in intel mkl library? (static, 32bit)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "mkl_ia32;guide;pthread"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      "acml_mp;acml_mv"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for ACML BLAS: found")
+      else()
+	message(STATUS "Looking for ACML BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    #BLAS in intel mkl library? (static, em64t 64bit)
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
       sgemm
       ""
-      "mkl_em64t;guide;pthread"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
-      )
-    endif()
-
-    #BLAS in acml library?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
-      BLAS_LIBRARIES
-      BLAS
-      sgemm
+      "acml;acml_mv;CALBLAS"
       ""
-      "acml"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for ACML BLAS: found")
+      else()
+	message(STATUS "Looking for ACML BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    # Apple BLAS library?
-    if(NOT BLAS_LIBRARIES)
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "ACML")
+  endif()
+
+endif (BLA_VENDOR MATCHES "ACML.*" OR BLA_VENDOR STREQUAL "All") # ACML
+
+
+# Apple BLAS library?
+if (BLA_VENDOR STREQUAL "Apple" OR BLA_VENDOR STREQUAL "All")
+
+  if(NOT BLAS_LIBRARIES)
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
-      sgemm
+      dgemm
       ""
       "Accelerate"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
+      ""
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for Apple BLAS: found")
+      else()
+	message(STATUS "Looking for Apple BLAS: not found")
+      endif()
     endif()
+  endif()
 
-    if ( NOT BLAS_LIBRARIES )
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Apple Accelerate")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "Apple" OR BLA_VENDOR STREQUAL "All")
+
+
+if (BLA_VENDOR STREQUAL "NAS" OR BLA_VENDOR STREQUAL "All")
+
+  if ( NOT BLAS_LIBRARIES )
+    check_fortran_libraries(
       BLAS_LIBRARIES
       BLAS
-      sgemm
+      dgemm
       ""
       "vecLib"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
-      )
-    endif ( NOT BLAS_LIBRARIES )
-
-    # Generic BLAS library?
-    # This configuration *must* be the last try as this library is notably slow.
-    if ( NOT BLAS_LIBRARIES )
-      check_fortran_libraries(
-      BLAS_DEFINITIONS
-      BLAS_LIBRARIES
-      BLAS
-      sgemm
       ""
-      "blas"
-      "${CGAL_TAUCS_LIBRARIES_DIR} ENV BLAS_LIB_DIR"
       )
+    if(NOT BLAS_FIND_QUIETLY)
+      if(BLAS_LIBRARIES)
+	message(STATUS "Looking for NAS BLAS: found")
+      else()
+	message(STATUS "Looking for NAS BLAS: not found")
+      endif()
     endif()
+  endif ()
 
-  if(BLAS_LIBRARIES_DIR OR BLAS_LIBRARIES)
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "NAS")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "NAS" OR BLA_VENDOR STREQUAL "All")
+
+
+# Generic BLAS library?
+if (BLA_VENDOR STREQUAL "Generic" OR BLA_VENDOR STREQUAL "All")
+
+  set(BLAS_SEARCH_LIBS "blas;blas_LINUX;blas_MAC;blas_WINDOWS;refblas")
+  foreach (SEARCH_LIB ${BLAS_SEARCH_LIBS})
+    if (BLAS_LIBRARIES)
+    else ()
+      check_fortran_libraries(
+	BLAS_LIBRARIES
+	BLAS
+	sgemm
+	""
+	"${SEARCH_LIB}"
+	"${LGFORTRAN}"
+	)
+      if(NOT BLAS_FIND_QUIETLY)
+	if(BLAS_LIBRARIES)
+	  message(STATUS "Looking for Generic BLAS: found")
+	else()
+	  message(STATUS "Looking for Generic BLAS: not found")
+	endif()
+      endif()
+    endif()
+  endforeach ()
+
+  if (BLAS_LIBRARIES AND NOT BLAS_VENDOR_FOUND)
+      set (BLAS_VENDOR_FOUND "Netlib or other Generic libblas")
+  endif()
+
+endif (BLA_VENDOR STREQUAL "Generic" OR BLA_VENDOR STREQUAL "All")
+
+
+if(BLA_F95)
+
+  if(BLAS95_LIBRARIES)
+    set(BLAS95_FOUND TRUE)
+  else()
+    set(BLAS95_FOUND FALSE)
+  endif()
+
+  if(NOT BLAS_FIND_QUIETLY)
+    if(BLAS95_FOUND)
+      message(STATUS "A library with BLAS95 API found.")
+      message(STATUS "BLAS_LIBRARIES ${BLAS_LIBRARIES}")
+    else(BLAS95_FOUND)
+      message(WARNING "BLA_VENDOR has been set to ${BLA_VENDOR} but blas 95 libraries could not be found or check of symbols failed."
+	"\nPlease indicate where to find blas libraries. You have three options:\n"
+	"- Option 1: Provide the installation directory of BLAS library with cmake option: -DBLAS_DIR=your/path/to/blas\n"
+	"- Option 2: Provide the directory where to find BLAS libraries with cmake option: -DBLAS_LIBDIR=your/path/to/blas/libs\n"
+	"- Option 3: Update your environment variable (Linux: LD_LIBRARY_PATH, Windows: LIB, Mac: DYLD_LIBRARY_PATH)\n"
+	"\nTo follow libraries detection more precisely you can activate a verbose mode with -DBLAS_VERBOSE=ON at cmake configure."
+	"\nYou could also specify a BLAS vendor to look for by setting -DBLA_VENDOR=blas_vendor_name."
+	"\nList of possible BLAS vendor: Goto, ATLAS PhiPACK, CXML, DXML, SunPerf, SCSL, SGIMATH, IBMESSL, Intel10_32 (intel mkl v10 32 bit),"
+	"Intel10_64lp (intel mkl v10 64 bit, lp thread model, lp64 model), Intel10_64lp_seq (intel mkl v10 64 bit, sequential code, lp64 model),"
+	"Intel( older versions of mkl 32 and 64 bit), ACML, ACML_MP, ACML_GPU, Apple, NAS, Generic")
+      if(BLAS_FIND_REQUIRED)
+	message(FATAL_ERROR
+	  "A required library with BLAS95 API not found. Please specify library location.")
+      else()
+	message(STATUS
+	  "A library with BLAS95 API not found. Please specify library location.")
+      endif()
+    endif(BLAS95_FOUND)
+  endif(NOT BLAS_FIND_QUIETLY)
+
+  set(BLAS_FOUND TRUE)
+  set(BLAS_LIBRARIES "${BLAS95_LIBRARIES}")
+
+else(BLA_F95)
+
+  if(BLAS_LIBRARIES)
     set(BLAS_FOUND TRUE)
   else()
     set(BLAS_FOUND FALSE)
@@ -388,32 +1366,41 @@ else()
   if(NOT BLAS_FIND_QUIETLY)
     if(BLAS_FOUND)
       message(STATUS "A library with BLAS API found.")
+      message(STATUS "BLAS_LIBRARIES ${BLAS_LIBRARIES}")
     else(BLAS_FOUND)
+      message(WARNING "BLA_VENDOR has been set to ${BLA_VENDOR} but blas libraries could not be found or check of symbols failed."
+	"\nPlease indicate where to find blas libraries. You have three options:\n"
+	"- Option 1: Provide the installation directory of BLAS library with cmake option: -DBLAS_DIR=your/path/to/blas\n"
+	"- Option 2: Provide the directory where to find BLAS libraries with cmake option: -DBLAS_LIBDIR=your/path/to/blas/libs\n"
+	"- Option 3: Update your environment variable (Linux: LD_LIBRARY_PATH, Windows: LIB, Mac: DYLD_LIBRARY_PATH)\n"
+	"\nTo follow libraries detection more precisely you can activate a verbose mode with -DBLAS_VERBOSE=ON at cmake configure."
+	"\nYou could also specify a BLAS vendor to look for by setting -DBLA_VENDOR=blas_vendor_name."
+	"\nList of possible BLAS vendor: Goto, ATLAS PhiPACK, CXML, DXML, SunPerf, SCSL, SGIMATH, IBMESSL, Intel10_32 (intel mkl v10 32 bit),"
+	"Intel10_64lp (intel mkl v10 64 bit, lp thread model, lp64 model), Intel10_64lp_seq (intel mkl v10 64 bit, sequential code, lp64 model),"
+	"Intel( older versions of mkl 32 and 64 bit), ACML, ACML_MP, ACML_GPU, Apple, NAS, Generic")
       if(BLAS_FIND_REQUIRED)
-        message(FATAL_ERROR "A required library with BLAS API not found. Please specify library location.")
+	message(FATAL_ERROR
+	  "A required library with BLAS API not found. Please specify library location.")
       else()
-        message(STATUS "A library with BLAS API not found. Please specify library location.")
+	message(STATUS
+	  "A library with BLAS API not found. Please specify library location.")
       endif()
     endif(BLAS_FOUND)
   endif(NOT BLAS_FIND_QUIETLY)
 
-  # Add variables to cache
-  set( BLAS_INCLUDE_DIR   "${BLAS_INCLUDE_DIR}"
-                          CACHE PATH "Directories containing the BLAS header files" FORCE )
-  set( BLAS_DEFINITIONS   "${BLAS_DEFINITIONS}"
-                          CACHE STRING "Compilation options to use BLAS" FORCE )
-  set( BLAS_LINKER_FLAGS  "${BLAS_LINKER_FLAGS}"
-                          CACHE STRING "Linker flags to use BLAS" FORCE )
-  set( BLAS_LIBRARIES     "${BLAS_LIBRARIES}"
-                          CACHE FILEPATH "BLAS libraries name" FORCE )
-  set( BLAS_LIBRARIES_DIR "${BLAS_LIBRARIES_DIR}"
-                          CACHE PATH "Directories containing the BLAS libraries" FORCE )
+endif(BLA_F95)
 
-  #message("DEBUG: BLAS_INCLUDE_DIR = ${BLAS_INCLUDE_DIR}")
-  #message("DEBUG: BLAS_DEFINITIONS = ${BLAS_DEFINITIONS}")
-  #message("DEBUG: BLAS_LINKER_FLAGS = ${BLAS_LINKER_FLAGS}")
-  #message("DEBUG: BLAS_LIBRARIES = ${BLAS_LIBRARIES}")
-  #message("DEBUG: BLAS_LIBRARIES_DIR = ${BLAS_LIBRARIES_DIR}")
-  #message("DEBUG: BLAS_FOUND = ${BLAS_FOUND}")
+set(CMAKE_FIND_LIBRARY_SUFFIXES ${_blas_ORIG_CMAKE_FIND_LIBRARY_SUFFIXES})
 
-endif(BLAS_LIBRARIES_DIR OR BLAS_LIBRARIES)
+if (BLAS_FOUND)
+  list(GET BLAS_LIBRARIES 0 first_lib)
+  get_filename_component(first_lib_path "${first_lib}" PATH)
+  if (${first_lib_path} MATCHES "(/lib(32|64)?$)|(/lib/intel64$|/lib/ia32$)")
+    string(REGEX REPLACE "(/lib(32|64)?$)|(/lib/intel64$|/lib/ia32$)" "" not_cached_dir "${first_lib_path}")
+    set(BLAS_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of BLAS library" FORCE)
+  else()
+    set(BLAS_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of BLAS library" FORCE)
+  endif()
+endif()
+mark_as_advanced(BLAS_DIR)
+mark_as_advanced(BLAS_DIR_FOUND)
diff --git a/cmake/FindBLASEXT.cmake b/cmake/FindBLASEXT.cmake
new file mode 100644
index 000000000..0fe7fb849
--- /dev/null
+++ b/cmake/FindBLASEXT.cmake
@@ -0,0 +1,380 @@
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2016 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find BLAS EXTENDED for MORSE projects: find include dirs and libraries
+#
+# This module allows to find BLAS libraries by calling the official FindBLAS module
+# and handles the creation of different library lists whether the user wishes to link
+# with a sequential BLAS or a multihreaded (BLAS_SEQ_LIBRARIES and BLAS_PAR_LIBRARIES).
+# BLAS is detected with a FindBLAS call then if the BLAS vendor is Intel10_64lp, ACML
+# or IBMESSLMT then the module attempts to find the corresponding multithreaded libraries.
+#
+# The following variables have been added to manage links with sequential or multithreaded
+# versions:
+#  BLAS_INCLUDE_DIRS  - BLAS include directories
+#  BLAS_LIBRARY_DIRS  - Link directories for BLAS libraries
+#  BLAS_SEQ_LIBRARIES - BLAS component libraries to be linked (sequential)
+#  BLAS_PAR_LIBRARIES - BLAS component libraries to be linked (multithreaded)
+
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013-2016 Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
+
+# macro to factorize this call
+macro(find_package_blas)
+  if(BLASEXT_FIND_REQUIRED)
+    if(BLASEXT_FIND_QUIETLY)
+      find_package(BLAS REQUIRED QUIET)
+    else()
+      find_package(BLAS REQUIRED)
+    endif()
+  else()
+    if(BLASEXT_FIND_QUIETLY)
+      find_package(BLAS QUIET)
+    else()
+      find_package(BLAS)
+    endif()
+  endif()
+endmacro()
+
+# add a cache variable to let the user specify the BLAS vendor
+set(BLA_VENDOR "" CACHE STRING "list of possible BLAS vendor:
+    Open, Eigen, Goto, ATLAS PhiPACK, CXML, DXML, SunPerf, SCSL, SGIMATH, IBMESSL, IBMESSLMT,
+    Intel10_32 (intel mkl v10 32 bit),
+    Intel10_64lp (intel mkl v10 64 bit, lp thread model, lp64 model),
+    Intel10_64lp_seq (intel mkl v10 64 bit, sequential code, lp64 model),
+    Intel( older versions of mkl 32 and 64 bit),
+    ACML, ACML_MP, ACML_GPU, Apple, NAS, Generic")
+
+if(NOT BLASEXT_FIND_QUIETLY)
+  message(STATUS "In FindBLASEXT")
+  message(STATUS "If you want to force the use of one specific library, "
+    "\n   please specify the BLAS vendor by setting -DBLA_VENDOR=blas_vendor_name"
+    "\n   at cmake configure.")
+  message(STATUS "List of possible BLAS vendor: Goto, ATLAS PhiPACK, CXML, "
+    "\n   DXML, SunPerf, SCSL, SGIMATH, IBMESSL, IBMESSLMT, Intel10_32 (intel mkl v10 32 bit),"
+    "\n   Intel10_64lp (intel mkl v10 64 bit, lp thread model, lp64 model),"
+    "\n   Intel10_64lp_seq (intel mkl v10 64 bit, sequential code, lp64 model),"
+    "\n   Intel( older versions of mkl 32 and 64 bit),"
+    "\n   ACML, ACML_MP, ACML_GPU, Apple, NAS, Generic")
+endif()
+
+if (NOT BLAS_FOUND)
+  # First try to detect two cases:
+  # 1: only SEQ libs are handled
+  # 2: both SEQ and PAR libs are handled
+  find_package_blas()
+endif ()
+
+# detect the cases where SEQ and PAR libs are handled
+if(BLA_VENDOR STREQUAL "All" AND
+    (BLAS_mkl_core_LIBRARY OR BLAS_mkl_core_dll_LIBRARY)
+    )
+  set(BLA_VENDOR "Intel")
+  if(BLAS_mkl_intel_LIBRARY)
+    set(BLA_VENDOR "Intel10_32")
+  endif()
+  if(BLAS_mkl_intel_lp64_LIBRARY)
+    set(BLA_VENDOR "Intel10_64lp")
+  endif()
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "A BLAS library has been found (${BLAS_LIBRARIES}) but we"
+      "\n   have also potentially detected some multithreaded BLAS libraries from the MKL."
+      "\n   We try to find both libraries lists (Sequential/Multithreaded).")
+  endif()
+  set(BLAS_FOUND "")
+elseif(BLA_VENDOR STREQUAL "All" AND BLAS_acml_LIBRARY)
+  set(BLA_VENDOR "ACML")
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "A BLAS library has been found (${BLAS_LIBRARIES}) but we"
+      "\n   have also potentially detected some multithreaded BLAS libraries from the ACML."
+      "\n   We try to find both libraries lists (Sequential/Multithreaded).")
+  endif()
+  set(BLAS_FOUND "")
+elseif(BLA_VENDOR STREQUAL "All" AND BLAS_essl_LIBRARY)
+  set(BLA_VENDOR "IBMESSL")
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "A BLAS library has been found (${BLAS_LIBRARIES}) but we"
+      "\n   have also potentially detected some multithreaded BLAS libraries from the ESSL."
+      "\n   We try to find both libraries lists (Sequential/Multithreaded).")
+  endif()
+  set(BLAS_FOUND "")
+endif()
+
+# Intel case
+if(BLA_VENDOR MATCHES "Intel*")
+
+  ###
+  # look for include path if the BLAS vendor is Intel
+  ###
+
+  # gather system include paths
+  unset(_inc_env)
+  if(WIN32)
+    string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
+  else()
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+  endif()
+  list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+  list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+  set(ENV_MKLROOT "$ENV{MKLROOT}")
+  if (ENV_MKLROOT)
+    list(APPEND _inc_env "${ENV_MKLROOT}/include")
+  endif()
+  list(REMOVE_DUPLICATES _inc_env)
+
+  # find mkl.h inside known include paths
+  set(BLAS_mkl.h_INCLUDE_DIRS "BLAS_mkl.h_INCLUDE_DIRS-NOTFOUND")
+  if(BLAS_INCDIR)
+    set(BLAS_mkl.h_INCLUDE_DIRS "BLAS_mkl.h_INCLUDE_DIRS-NOTFOUND")
+    find_path(BLAS_mkl.h_INCLUDE_DIRS
+      NAMES mkl.h
+      HINTS ${BLAS_INCDIR})
+  else()
+    if(BLAS_DIR)
+      set(BLAS_mkl.h_INCLUDE_DIRS "BLAS_mkl.h_INCLUDE_DIRS-NOTFOUND")
+      find_path(BLAS_mkl.h_INCLUDE_DIRS
+	NAMES mkl.h
+	HINTS ${BLAS_DIR}
+	PATH_SUFFIXES include)
+    else()
+      set(BLAS_mkl.h_INCLUDE_DIRS "BLAS_mkl.h_INCLUDE_DIRS-NOTFOUND")
+      find_path(BLAS_mkl.h_INCLUDE_DIRS
+	NAMES mkl.h
+	HINTS ${_inc_env})
+    endif()
+  endif()
+  mark_as_advanced(BLAS_mkl.h_INCLUDE_DIRS)
+  ## Print status if not found
+  ## -------------------------
+  #if (NOT BLAS_mkl.h_INCLUDE_DIRS AND MORSE_VERBOSE)
+  #    Print_Find_Header_Status(blas mkl.h)
+  #endif ()
+  set(BLAS_INCLUDE_DIRS "")
+  if(BLAS_mkl.h_INCLUDE_DIRS)
+    list(APPEND BLAS_INCLUDE_DIRS "${BLAS_mkl.h_INCLUDE_DIRS}" )
+  endif()
+
+  ###
+  # look for libs
+  ###
+  # if Intel 10 64 bit -> look for sequential and multithreaded versions
+  if(BLA_VENDOR MATCHES "Intel10_64lp*")
+
+    ## look for the sequential version
+    set(BLA_VENDOR "Intel10_64lp_seq")
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "Look for the sequential version Intel10_64lp_seq")
+    endif()
+    find_package_blas()
+    if(BLAS_FOUND)
+      set(BLAS_SEQ_LIBRARIES "${BLAS_LIBRARIES}")
+    else()
+      set(BLAS_SEQ_LIBRARIES "${BLAS_SEQ_LIBRARIES-NOTFOUND}")
+    endif()
+
+    ## look for the multithreaded version
+    set(BLA_VENDOR "Intel10_64lp")
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "Look for the multithreaded version Intel10_64lp")
+    endif()
+    find_package_blas()
+    if(BLAS_FOUND)
+      set(BLAS_PAR_LIBRARIES "${BLAS_LIBRARIES}")
+    else()
+      set(BLAS_PAR_LIBRARIES "${BLAS_PAR_LIBRARIES-NOTFOUND}")
+    endif()
+
+  else()
+
+    if(BLAS_FOUND)
+      set(BLAS_SEQ_LIBRARIES "${BLAS_LIBRARIES}")
+    else()
+      set(BLAS_SEQ_LIBRARIES "${BLAS_SEQ_LIBRARIES-NOTFOUND}")
+    endif()
+
+  endif()
+
+  # ACML case
+elseif(BLA_VENDOR MATCHES "ACML*")
+
+  ## look for the sequential version
+  set(BLA_VENDOR "ACML")
+  find_package_blas()
+  if(BLAS_FOUND)
+    set(BLAS_SEQ_LIBRARIES "${BLAS_LIBRARIES}")
+  else()
+    set(BLAS_SEQ_LIBRARIES "${BLAS_SEQ_LIBRARIES-NOTFOUND}")
+  endif()
+
+  ## look for the multithreaded version
+  set(BLA_VENDOR "ACML_MP")
+  find_package_blas()
+  if(BLAS_FOUND)
+    set(BLAS_PAR_LIBRARIES "${BLAS_LIBRARIES}")
+  else()
+    set(BLAS_PAR_LIBRARIES "${BLAS_PAR_LIBRARIES-NOTFOUND}")
+  endif()
+
+  # IBMESSL case
+elseif(BLA_VENDOR MATCHES "IBMESSL*")
+
+  ## look for the sequential version
+  set(BLA_VENDOR "IBMESSL")
+  find_package_blas()
+  if(BLAS_FOUND)
+    set(BLAS_SEQ_LIBRARIES "${BLAS_LIBRARIES}")
+  else()
+    set(BLAS_SEQ_LIBRARIES "${BLAS_SEQ_LIBRARIES-NOTFOUND}")
+  endif()
+
+  ## look for the multithreaded version
+  set(BLA_VENDOR "IBMESSLMT")
+  find_package_blas()
+  if(BLAS_FOUND)
+    set(BLAS_PAR_LIBRARIES "${BLAS_LIBRARIES}")
+  else()
+    set(BLAS_PAR_LIBRARIES "${BLAS_PAR_LIBRARIES-NOTFOUND}")
+  endif()
+
+else()
+
+  if(BLAS_FOUND)
+    # define the SEQ libs as the BLAS_LIBRARIES
+    set(BLAS_SEQ_LIBRARIES "${BLAS_LIBRARIES}")
+  else()
+    set(BLAS_SEQ_LIBRARIES "${BLAS_SEQ_LIBRARIES-NOTFOUND}")
+  endif()
+  set(BLAS_PAR_LIBRARIES "${BLAS_PAR_LIBRARIES-NOTFOUND}")
+
+endif()
+
+
+if(BLAS_SEQ_LIBRARIES)
+  set(BLAS_LIBRARIES "${BLAS_SEQ_LIBRARIES}")
+endif()
+
+# extract libs paths
+# remark: because it is not given by find_package(BLAS)
+set(BLAS_LIBRARY_DIRS "")
+string(REPLACE " " ";" BLAS_LIBRARIES "${BLAS_LIBRARIES}")
+foreach(blas_lib ${BLAS_LIBRARIES})
+  if (EXISTS "${blas_lib}")
+    get_filename_component(a_blas_lib_dir "${blas_lib}" PATH)
+    list(APPEND BLAS_LIBRARY_DIRS "${a_blas_lib_dir}" )
+  else()
+    string(REPLACE "-L" "" blas_lib "${blas_lib}")
+    if (EXISTS "${blas_lib}")
+      list(APPEND BLAS_LIBRARY_DIRS "${blas_lib}" )
+    else()
+      get_filename_component(a_blas_lib_dir "${blas_lib}" PATH)
+      if (EXISTS "${a_blas_lib_dir}")
+	list(APPEND BLAS_LIBRARY_DIRS "${a_blas_lib_dir}" )
+      endif()
+    endif()
+  endif()
+endforeach()
+if (BLAS_LIBRARY_DIRS)
+  list(REMOVE_DUPLICATES BLAS_LIBRARY_DIRS)
+endif ()
+
+# check that BLAS has been found
+# ---------------------------------
+include(FindPackageHandleStandardArgs)
+if(BLA_VENDOR MATCHES "Intel*")
+  if(BLA_VENDOR MATCHES "Intel10_64lp*")
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "BLAS found is Intel MKL:"
+	"\n   we manage two lists of libs, one sequential and one parallel if found"
+	"\n   (see BLAS_SEQ_LIBRARIES and BLAS_PAR_LIBRARIES)")
+      message(STATUS "BLAS sequential libraries stored in BLAS_SEQ_LIBRARIES")
+    endif()
+    find_package_handle_standard_args(BLAS DEFAULT_MSG
+      BLAS_SEQ_LIBRARIES
+      BLAS_LIBRARY_DIRS
+      BLAS_INCLUDE_DIRS)
+    if(BLAS_PAR_LIBRARIES)
+      if(NOT BLASEXT_FIND_QUIETLY)
+	message(STATUS "BLAS parallel libraries stored in BLAS_PAR_LIBRARIES")
+      endif()
+      find_package_handle_standard_args(BLAS DEFAULT_MSG
+	BLAS_PAR_LIBRARIES)
+    endif()
+  else()
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "BLAS sequential libraries stored in BLAS_SEQ_LIBRARIES")
+    endif()
+    find_package_handle_standard_args(BLAS DEFAULT_MSG
+      BLAS_SEQ_LIBRARIES
+      BLAS_LIBRARY_DIRS
+      BLAS_INCLUDE_DIRS)
+  endif()
+elseif(BLA_VENDOR MATCHES "ACML*")
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "BLAS found is ACML:"
+      "\n   we manage two lists of libs, one sequential and one parallel if found"
+      "\n   (see BLAS_SEQ_LIBRARIES and BLAS_PAR_LIBRARIES)")
+    message(STATUS "BLAS sequential libraries stored in BLAS_SEQ_LIBRARIES")
+  endif()
+  find_package_handle_standard_args(BLAS DEFAULT_MSG
+    BLAS_SEQ_LIBRARIES
+    BLAS_LIBRARY_DIRS)
+  if(BLAS_PAR_LIBRARIES)
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "BLAS parallel libraries stored in BLAS_PAR_LIBRARIES")
+    endif()
+    find_package_handle_standard_args(BLAS DEFAULT_MSG
+      BLAS_PAR_LIBRARIES)
+  endif()
+elseif(BLA_VENDOR MATCHES "IBMESSL*")
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "BLAS found is ESSL:"
+      "\n   we manage two lists of libs, one sequential and one parallel if found"
+      "\n   (see BLAS_SEQ_LIBRARIES and BLAS_PAR_LIBRARIES)")
+    message(STATUS "BLAS sequential libraries stored in BLAS_SEQ_LIBRARIES")
+  endif()
+  find_package_handle_standard_args(BLAS DEFAULT_MSG
+    BLAS_SEQ_LIBRARIES
+    BLAS_LIBRARY_DIRS)
+  if(BLAS_PAR_LIBRARIES)
+    if(NOT BLASEXT_FIND_QUIETLY)
+      message(STATUS "BLAS parallel libraries stored in BLAS_PAR_LIBRARIES")
+    endif()
+    find_package_handle_standard_args(BLAS DEFAULT_MSG
+      BLAS_PAR_LIBRARIES)
+  endif()
+else()
+  if(NOT BLASEXT_FIND_QUIETLY)
+    message(STATUS "BLAS sequential libraries stored in BLAS_SEQ_LIBRARIES")
+  endif()
+  find_package_handle_standard_args(BLAS DEFAULT_MSG
+    BLAS_SEQ_LIBRARIES
+    BLAS_LIBRARY_DIRS)
+endif()
diff --git a/cmake/FindHWLOC.cmake b/cmake/FindHWLOC.cmake
new file mode 100644
index 000000000..a831b5c72
--- /dev/null
+++ b/cmake/FindHWLOC.cmake
@@ -0,0 +1,331 @@
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2014 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find HWLOC include dirs and libraries
+# Use this module by invoking find_package with the form:
+#  find_package(HWLOC
+#               [REQUIRED]) # Fail with error if hwloc is not found
+#
+# This module finds headers and hwloc library.
+# Results are reported in variables:
+#  HWLOC_FOUND           - True if headers and requested libraries were found
+#  HWLOC_INCLUDE_DIRS    - hwloc include directories
+#  HWLOC_LIBRARY_DIRS    - Link directories for hwloc libraries
+#  HWLOC_LIBRARIES       - hwloc component libraries to be linked
+#
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DHWLOC_DIR=path/to/hwloc):
+#  HWLOC_DIR             - Where to find the base directory of hwloc
+#  HWLOC_INCDIR          - Where to find the header files
+#  HWLOC_LIBDIR          - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: HWLOC_DIR, HWLOC_INCDIR, HWLOC_LIBDIR
+
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013      Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
+
+include(CheckStructHasMember)
+include(CheckCSourceCompiles)
+
+if (NOT HWLOC_FOUND)
+  set(HWLOC_DIR "" CACHE PATH "Installation directory of HWLOC library")
+  if (NOT HWLOC_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely HWLOC_DIR, has been set to specify the install directory of HWLOC")
+  endif()
+endif()
+
+set(ENV_HWLOC_DIR "$ENV{HWLOC_DIR}")
+set(ENV_HWLOC_INCDIR "$ENV{HWLOC_INCDIR}")
+set(ENV_HWLOC_LIBDIR "$ENV{HWLOC_LIBDIR}")
+set(HWLOC_GIVEN_BY_USER "FALSE")
+if ( HWLOC_DIR OR ( HWLOC_INCDIR AND HWLOC_LIBDIR) OR ENV_HWLOC_DIR OR (ENV_HWLOC_INCDIR AND ENV_HWLOC_LIBDIR) )
+  set(HWLOC_GIVEN_BY_USER "TRUE")
+endif()
+
+# Optionally use pkg-config to detect include/library dirs (if pkg-config is available)
+# -------------------------------------------------------------------------------------
+include(FindPkgConfig)
+find_package(PkgConfig QUIET)
+if( PKG_CONFIG_EXECUTABLE AND NOT HWLOC_GIVEN_BY_USER )
+
+  pkg_search_module(HWLOC hwloc)
+  if (NOT HWLOC_FIND_QUIETLY)
+    if (HWLOC_FOUND AND HWLOC_LIBRARIES)
+      message(STATUS "Looking for HWLOC - found using PkgConfig")
+      #if(NOT HWLOC_INCLUDE_DIRS)
+      #    message("${Magenta}HWLOC_INCLUDE_DIRS is empty using PkgConfig."
+      #        "Perhaps the path to hwloc headers is already present in your"
+      #        "C(PLUS)_INCLUDE_PATH environment variable.${ColourReset}")
+      #endif()
+    else()
+      message(STATUS "${Magenta}Looking for HWLOC - not found using PkgConfig."
+	"\n   Perhaps you should add the directory containing hwloc.pc to"
+	"\n   the PKG_CONFIG_PATH environment variable.${ColourReset}")
+    endif()
+  endif()
+
+endif( PKG_CONFIG_EXECUTABLE AND NOT HWLOC_GIVEN_BY_USER )
+
+if( (NOT PKG_CONFIG_EXECUTABLE) OR (PKG_CONFIG_EXECUTABLE AND NOT HWLOC_FOUND) OR (HWLOC_GIVEN_BY_USER) )
+
+  if (NOT HWLOC_FIND_QUIETLY)
+    message(STATUS "Looking for HWLOC - PkgConfig not used")
+  endif()
+
+  # Looking for include
+  # -------------------
+
+  # Add system include paths to search include
+  # ------------------------------------------
+  unset(_inc_env)
+  if(ENV_HWLOC_INCDIR)
+    list(APPEND _inc_env "${ENV_HWLOC_INCDIR}")
+  elseif(ENV_HWLOC_DIR)
+    list(APPEND _inc_env "${ENV_HWLOC_DIR}")
+    list(APPEND _inc_env "${ENV_HWLOC_DIR}/include")
+    list(APPEND _inc_env "${ENV_HWLOC_DIR}/include/hwloc")
+  else()
+    if(WIN32)
+      string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
+    else()
+      string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+      list(APPEND _inc_env "${_path_env}")
+      string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+      list(APPEND _inc_env "${_path_env}")
+      string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+      list(APPEND _inc_env "${_path_env}")
+      string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+      list(APPEND _inc_env "${_path_env}")
+    endif()
+  endif()
+  list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+  list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+  list(REMOVE_DUPLICATES _inc_env)
+
+  # set paths where to look for
+  set(PATH_TO_LOOK_FOR "${_inc_env}")
+
+  # Try to find the hwloc header in the given paths
+  # -------------------------------------------------
+  # call cmake macro to find the header path
+  if(HWLOC_INCDIR)
+    set(HWLOC_hwloc.h_DIRS "HWLOC_hwloc.h_DIRS-NOTFOUND")
+    find_path(HWLOC_hwloc.h_DIRS
+      NAMES hwloc.h
+      HINTS ${HWLOC_INCDIR})
+  else()
+    if(HWLOC_DIR)
+      set(HWLOC_hwloc.h_DIRS "HWLOC_hwloc.h_DIRS-NOTFOUND")
+      find_path(HWLOC_hwloc.h_DIRS
+	NAMES hwloc.h
+	HINTS ${HWLOC_DIR}
+	PATH_SUFFIXES "include" "include/hwloc")
+    else()
+      set(HWLOC_hwloc.h_DIRS "HWLOC_hwloc.h_DIRS-NOTFOUND")
+      find_path(HWLOC_hwloc.h_DIRS
+	NAMES hwloc.h
+	HINTS ${PATH_TO_LOOK_FOR}
+	PATH_SUFFIXES "hwloc")
+    endif()
+  endif()
+  mark_as_advanced(HWLOC_hwloc.h_DIRS)
+
+  # Add path to cmake variable
+  # ------------------------------------
+  if (HWLOC_hwloc.h_DIRS)
+    set(HWLOC_INCLUDE_DIRS "${HWLOC_hwloc.h_DIRS}")
+  else ()
+    set(HWLOC_INCLUDE_DIRS "HWLOC_INCLUDE_DIRS-NOTFOUND")
+    if(NOT HWLOC_FIND_QUIETLY)
+      message(STATUS "Looking for hwloc -- hwloc.h not found")
+    endif()
+  endif ()
+
+  if (HWLOC_INCLUDE_DIRS)
+    list(REMOVE_DUPLICATES HWLOC_INCLUDE_DIRS)
+  endif ()
+
+
+  # Looking for lib
+  # ---------------
+
+  # Add system library paths to search lib
+  # --------------------------------------
+  unset(_lib_env)
+  if(ENV_HWLOC_LIBDIR)
+    list(APPEND _lib_env "${ENV_HWLOC_LIBDIR}")
+  elseif(ENV_HWLOC_DIR)
+    list(APPEND _lib_env "${ENV_HWLOC_DIR}")
+    list(APPEND _lib_env "${ENV_HWLOC_DIR}/lib")
+  else()
+    if(WIN32)
+      string(REPLACE ":" ";" _lib_env "$ENV{LIB}")
+    else()
+      if(APPLE)
+	string(REPLACE ":" ";" _lib_env "$ENV{DYLD_LIBRARY_PATH}")
+      else()
+	string(REPLACE ":" ";" _lib_env "$ENV{LD_LIBRARY_PATH}")
+      endif()
+      list(APPEND _lib_env "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+      list(APPEND _lib_env "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+    endif()
+  endif()
+  list(REMOVE_DUPLICATES _lib_env)
+
+  # set paths where to look for
+  set(PATH_TO_LOOK_FOR "${_lib_env}")
+
+  # Try to find the hwloc lib in the given paths
+  # ----------------------------------------------
+
+  # call cmake macro to find the lib path
+  if(HWLOC_LIBDIR)
+    set(HWLOC_hwloc_LIBRARY "HWLOC_hwloc_LIBRARY-NOTFOUND")
+    find_library(HWLOC_hwloc_LIBRARY
+      NAMES hwloc
+      HINTS ${HWLOC_LIBDIR})
+  else()
+    if(HWLOC_DIR)
+      set(HWLOC_hwloc_LIBRARY "HWLOC_hwloc_LIBRARY-NOTFOUND")
+      find_library(HWLOC_hwloc_LIBRARY
+	NAMES hwloc
+	HINTS ${HWLOC_DIR}
+	PATH_SUFFIXES lib lib32 lib64)
+    else()
+      set(HWLOC_hwloc_LIBRARY "HWLOC_hwloc_LIBRARY-NOTFOUND")
+      find_library(HWLOC_hwloc_LIBRARY
+	NAMES hwloc
+	HINTS ${PATH_TO_LOOK_FOR})
+    endif()
+  endif()
+  mark_as_advanced(HWLOC_hwloc_LIBRARY)
+
+  # If found, add path to cmake variable
+  # ------------------------------------
+  if (HWLOC_hwloc_LIBRARY)
+    get_filename_component(hwloc_lib_path ${HWLOC_hwloc_LIBRARY} PATH)
+    # set cmake variables (respects naming convention)
+    set(HWLOC_LIBRARIES    "${HWLOC_hwloc_LIBRARY}")
+    set(HWLOC_LIBRARY_DIRS "${hwloc_lib_path}")
+  else ()
+    set(HWLOC_LIBRARIES    "HWLOC_LIBRARIES-NOTFOUND")
+    set(HWLOC_LIBRARY_DIRS "HWLOC_LIBRARY_DIRS-NOTFOUND")
+    if(NOT HWLOC_FIND_QUIETLY)
+      message(STATUS "Looking for hwloc -- lib hwloc not found")
+    endif()
+  endif ()
+
+  if (HWLOC_LIBRARY_DIRS)
+    list(REMOVE_DUPLICATES HWLOC_LIBRARY_DIRS)
+  endif ()
+
+  # check a function to validate the find
+  if(HWLOC_LIBRARIES)
+
+    set(REQUIRED_INCDIRS)
+    set(REQUIRED_LIBDIRS)
+    set(REQUIRED_LIBS)
+
+    # HWLOC
+    if (HWLOC_INCLUDE_DIRS)
+      set(REQUIRED_INCDIRS "${HWLOC_INCLUDE_DIRS}")
+    endif()
+    if (HWLOC_LIBRARY_DIRS)
+      set(REQUIRED_LIBDIRS "${HWLOC_LIBRARY_DIRS}")
+    endif()
+    set(REQUIRED_LIBS "${HWLOC_LIBRARIES}")
+
+    # set required libraries for link
+    set(CMAKE_REQUIRED_INCLUDES "${REQUIRED_INCDIRS}")
+    set(CMAKE_REQUIRED_LIBRARIES)
+    foreach(lib_dir ${REQUIRED_LIBDIRS})
+      list(APPEND CMAKE_REQUIRED_LIBRARIES "-L${lib_dir}")
+    endforeach()
+    list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LIBS}")
+    string(REGEX REPLACE "^ -" "-" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+
+    # test link
+    unset(HWLOC_WORKS CACHE)
+    include(CheckFunctionExists)
+    check_function_exists(hwloc_topology_init HWLOC_WORKS)
+    mark_as_advanced(HWLOC_WORKS)
+
+    if(NOT HWLOC_WORKS)
+      if(NOT HWLOC_FIND_QUIETLY)
+	message(STATUS "Looking for hwloc : test of hwloc_topology_init with hwloc library fails")
+	message(STATUS "CMAKE_REQUIRED_LIBRARIES: ${CMAKE_REQUIRED_LIBRARIES}")
+	message(STATUS "CMAKE_REQUIRED_INCLUDES: ${CMAKE_REQUIRED_INCLUDES}")
+	message(STATUS "Check in CMakeFiles/CMakeError.log to figure out why it fails")
+      endif()
+    endif()
+    set(CMAKE_REQUIRED_INCLUDES)
+    set(CMAKE_REQUIRED_FLAGS)
+    set(CMAKE_REQUIRED_LIBRARIES)
+  endif(HWLOC_LIBRARIES)
+
+endif( (NOT PKG_CONFIG_EXECUTABLE) OR (PKG_CONFIG_EXECUTABLE AND NOT HWLOC_FOUND) OR (HWLOC_GIVEN_BY_USER) )
+
+if (HWLOC_LIBRARIES)
+  if (HWLOC_LIBRARY_DIRS)
+    list(GET HWLOC_LIBRARY_DIRS 0 first_lib_path)
+  else()
+    list(GET HWLOC_LIBRARIES 0 first_lib)
+    get_filename_component(first_lib_path "${first_lib}" PATH)
+  endif()
+  if (${first_lib_path} MATCHES "/lib(32|64)?$")
+    string(REGEX REPLACE "/lib(32|64)?$" "" not_cached_dir "${first_lib_path}")
+    set(HWLOC_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of HWLOC library" FORCE)
+  else()
+    set(HWLOC_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of HWLOC library" FORCE)
+  endif()
+endif()
+mark_as_advanced(HWLOC_DIR)
+mark_as_advanced(HWLOC_DIR_FOUND)
+
+# check that HWLOC has been found
+# -------------------------------
+include(FindPackageHandleStandardArgs)
+if (PKG_CONFIG_EXECUTABLE AND HWLOC_FOUND)
+  find_package_handle_standard_args(HWLOC DEFAULT_MSG
+    HWLOC_LIBRARIES)
+else()
+  find_package_handle_standard_args(HWLOC DEFAULT_MSG
+    HWLOC_LIBRARIES
+    HWLOC_WORKS)
+endif()
+
+if (HWLOC_FOUND)
+  set(HWLOC_SAVE_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
+  list(APPEND CMAKE_REQUIRED_INCLUDES ${HWLOC_INCLUDE_DIRS})
+
+  # test headers to guess the version
+  check_struct_has_member( "struct hwloc_obj" parent hwloc.h HAVE_HWLOC_PARENT_MEMBER )
+  check_struct_has_member( "struct hwloc_cache_attr_s" size hwloc.h HAVE_HWLOC_CACHE_ATTR )
+  check_c_source_compiles( "#include <hwloc.h>
+	    int main(void) { hwloc_obj_t o; o->type = HWLOC_OBJ_PU; return 0;}" HAVE_HWLOC_OBJ_PU)
+  include(CheckLibraryExists)
+  check_library_exists(${HWLOC_LIBRARIES} hwloc_bitmap_free "" HAVE_HWLOC_BITMAP)
+
+  set(CMAKE_REQUIRED_INCLUDES ${HWLOC_SAVE_CMAKE_REQUIRED_INCLUDES})
+endif()
diff --git a/cmake/FindMetis.cmake b/cmake/FindMetis.cmake
index 6a0ce790c..da2f1f1d7 100644
--- a/cmake/FindMetis.cmake
+++ b/cmake/FindMetis.cmake
@@ -1,59 +1,264 @@
-# Pastix requires METIS or METIS (partitioning and reordering tools)
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2014 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find METIS include dirs and libraries
+# Use this module by invoking find_package with the form:
+#  find_package(METIS
+#               [REQUIRED]             # Fail with error if metis is not found
+#              )
+#
+# This module finds headers and metis library.
+# Results are reported in variables:
+#  METIS_FOUND           - True if headers and requested libraries were found
+#  METIS_INCLUDE_DIRS    - metis include directories
+#  METIS_LIBRARY_DIRS    - Link directories for metis libraries
+#  METIS_LIBRARIES       - metis component libraries to be linked
+#
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DMETIS_DIR=path/to/metis):
+#  METIS_DIR             - Where to find the base directory of metis
+#  METIS_INCDIR          - Where to find the header files
+#  METIS_LIBDIR          - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: METIS_DIR, METIS_INCDIR, METIS_LIBDIR
 
-if (METIS_INCLUDES AND METIS_LIBRARIES)
-  set(METIS_FIND_QUIETLY TRUE)
-endif (METIS_INCLUDES AND METIS_LIBRARIES)
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013      Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
 
-find_path(METIS_INCLUDES 
-  NAMES 
-  metis.h 
-  PATHS 
-  $ENV{METISDIR} 
-  ${INCLUDE_INSTALL_DIR} 
-  PATH_SUFFIXES
-  .
-  metis
-  include
-)
-
-macro(_metis_check_version)
-  file(READ "${METIS_INCLUDES}/metis.h" _metis_version_header)
-
-  string(REGEX MATCH "define[ \t]+METIS_VER_MAJOR[ \t]+([0-9]+)" _metis_major_version_match "${_metis_version_header}")
-  set(METIS_MAJOR_VERSION "${CMAKE_MATCH_1}")
-  string(REGEX MATCH "define[ \t]+METIS_VER_MINOR[ \t]+([0-9]+)" _metis_minor_version_match "${_metis_version_header}")
-  set(METIS_MINOR_VERSION "${CMAKE_MATCH_1}")
-  string(REGEX MATCH "define[ \t]+METIS_VER_SUBMINOR[ \t]+([0-9]+)" _metis_subminor_version_match "${_metis_version_header}")
-  set(METIS_SUBMINOR_VERSION "${CMAKE_MATCH_1}")
-  if(NOT METIS_MAJOR_VERSION)
-    message(STATUS "Could not determine Metis version. Assuming version 4.0.0")
-    set(METIS_VERSION 4.0.0)
-  else()
-    set(METIS_VERSION ${METIS_MAJOR_VERSION}.${METIS_MINOR_VERSION}.${METIS_SUBMINOR_VERSION})
+if (NOT METIS_FOUND)
+  set(METIS_DIR "" CACHE PATH "Installation directory of METIS library")
+  if (NOT METIS_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely METIS_DIR, has been set to specify the install directory of METIS")
   endif()
-  if(${METIS_VERSION} VERSION_LESS ${Metis_FIND_VERSION})
-    set(METIS_VERSION_OK FALSE)
+endif()
+
+# Looking for include
+# -------------------
+
+# Add system include paths to search include
+# ------------------------------------------
+unset(_inc_env)
+set(ENV_METIS_DIR "$ENV{METIS_DIR}")
+set(ENV_METIS_INCDIR "$ENV{METIS_INCDIR}")
+if(ENV_METIS_INCDIR)
+  list(APPEND _inc_env "${ENV_METIS_INCDIR}")
+elseif(ENV_METIS_DIR)
+  list(APPEND _inc_env "${ENV_METIS_DIR}")
+  list(APPEND _inc_env "${ENV_METIS_DIR}/include")
+  list(APPEND _inc_env "${ENV_METIS_DIR}/include/metis")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
   else()
-    set(METIS_VERSION_OK TRUE)
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+  endif()
+endif()
+list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(REMOVE_DUPLICATES _inc_env)
+
+
+# Try to find the metis header in the given paths
+# -------------------------------------------------
+# call cmake macro to find the header path
+if(METIS_INCDIR)
+  set(METIS_metis.h_DIRS "METIS_metis.h_DIRS-NOTFOUND")
+  find_path(METIS_metis.h_DIRS
+    NAMES metis.h
+    HINTS ${METIS_INCDIR})
+else()
+  if(METIS_DIR)
+    set(METIS_metis.h_DIRS "METIS_metis.h_DIRS-NOTFOUND")
+    find_path(METIS_metis.h_DIRS
+      NAMES metis.h
+      HINTS ${METIS_DIR}
+      PATH_SUFFIXES "include" "include/metis")
+  else()
+    set(METIS_metis.h_DIRS "METIS_metis.h_DIRS-NOTFOUND")
+    find_path(METIS_metis.h_DIRS
+      NAMES metis.h
+      HINTS ${_inc_env})
+  endif()
+endif()
+mark_as_advanced(METIS_metis.h_DIRS)
+
+
+# If found, add path to cmake variable
+# ------------------------------------
+if (METIS_metis.h_DIRS)
+  set(METIS_INCLUDE_DIRS "${METIS_metis.h_DIRS}")
+else ()
+  set(METIS_INCLUDE_DIRS "METIS_INCLUDE_DIRS-NOTFOUND")
+  if(NOT METIS_FIND_QUIETLY)
+    message(STATUS "Looking for metis -- metis.h not found")
+  endif()
+endif()
+
+
+# Looking for lib
+# ---------------
+
+# Add system library paths to search lib
+# --------------------------------------
+unset(_lib_env)
+set(ENV_METIS_LIBDIR "$ENV{METIS_LIBDIR}")
+if(ENV_METIS_LIBDIR)
+  list(APPEND _lib_env "${ENV_METIS_LIBDIR}")
+elseif(ENV_METIS_DIR)
+  list(APPEND _lib_env "${ENV_METIS_DIR}")
+  list(APPEND _lib_env "${ENV_METIS_DIR}/lib")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _lib_env "$ENV{LIB}")
+  else()
+    if(APPLE)
+      string(REPLACE ":" ";" _lib_env "$ENV{DYLD_LIBRARY_PATH}")
+    else()
+      string(REPLACE ":" ";" _lib_env "$ENV{LD_LIBRARY_PATH}")
+    endif()
+    list(APPEND _lib_env "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+    list(APPEND _lib_env "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+  endif()
+endif()
+list(REMOVE_DUPLICATES _lib_env)
+
+# Try to find the metis lib in the given paths
+# ----------------------------------------------
+# call cmake macro to find the lib path
+if(METIS_LIBDIR)
+  set(METIS_metis_LIBRARY "METIS_metis_LIBRARY-NOTFOUND")
+  find_library(METIS_metis_LIBRARY
+    NAMES metis
+    HINTS ${METIS_LIBDIR})
+else()
+  if(METIS_DIR)
+    set(METIS_metis_LIBRARY "METIS_metis_LIBRARY-NOTFOUND")
+    find_library(METIS_metis_LIBRARY
+      NAMES metis
+      HINTS ${METIS_DIR}
+      PATH_SUFFIXES lib lib32 lib64)
+  else()
+    set(METIS_metis_LIBRARY "METIS_metis_LIBRARY-NOTFOUND")
+    find_library(METIS_metis_LIBRARY
+      NAMES metis
+      HINTS ${_lib_env})
+  endif()
+endif()
+mark_as_advanced(METIS_metis_LIBRARY)
+
+
+# If found, add path to cmake variable
+# ------------------------------------
+if (METIS_metis_LIBRARY)
+  get_filename_component(metis_lib_path "${METIS_metis_LIBRARY}" PATH)
+  # set cmake variables
+  set(METIS_LIBRARIES    "${METIS_metis_LIBRARY}")
+  set(METIS_LIBRARY_DIRS "${metis_lib_path}")
+else ()
+  set(METIS_LIBRARIES    "METIS_LIBRARIES-NOTFOUND")
+  set(METIS_LIBRARY_DIRS "METIS_LIBRARY_DIRS-NOTFOUND")
+  if(NOT METIS_FIND_QUIETLY)
+    message(STATUS "Looking for metis -- lib metis not found")
+  endif()
+endif ()
+
+# check a function to validate the find
+if(METIS_LIBRARIES)
+
+  set(REQUIRED_INCDIRS)
+  set(REQUIRED_LIBDIRS)
+  set(REQUIRED_LIBS)
+
+  # METIS
+  if (METIS_INCLUDE_DIRS)
+    set(REQUIRED_INCDIRS  "${METIS_INCLUDE_DIRS}")
+  endif()
+  if (METIS_LIBRARY_DIRS)
+    set(REQUIRED_LIBDIRS "${METIS_LIBRARY_DIRS}")
+  endif()
+  set(REQUIRED_LIBS "${METIS_LIBRARIES}")
+  # m
+  find_library(M_LIBRARY NAMES m)
+  mark_as_advanced(M_LIBRARY)
+  if(M_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lm")
   endif()
 
-  if(NOT METIS_VERSION_OK)
-    message(STATUS "Metis version ${METIS_VERSION} found in ${METIS_INCLUDES}, "
-                   "but at least version ${Metis_FIND_VERSION} is required")
-  endif(NOT METIS_VERSION_OK)
-endmacro(_metis_check_version)
+  # set required libraries for link
+  set(CMAKE_REQUIRED_INCLUDES "${REQUIRED_INCDIRS}")
+  set(CMAKE_REQUIRED_LIBRARIES)
+  foreach(lib_dir ${REQUIRED_LIBDIRS})
+    list(APPEND CMAKE_REQUIRED_LIBRARIES "-L${lib_dir}")
+  endforeach()
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LIBS}")
+  string(REGEX REPLACE "^ -" "-" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
 
-  if(METIS_INCLUDES AND Metis_FIND_VERSION)
-    _metis_check_version()
-  else()
-    set(METIS_VERSION_OK TRUE)
+  # test link
+  unset(METIS_WORKS CACHE)
+  include(CheckFunctionExists)
+  check_function_exists(METIS_NodeND METIS_WORKS)
+  mark_as_advanced(METIS_WORKS)
+
+  if(NOT METIS_WORKS)
+    if(NOT METIS_FIND_QUIETLY)
+      message(STATUS "Looking for METIS : test of METIS_NodeND with METIS library fails")
+      message(STATUS "CMAKE_REQUIRED_LIBRARIES: ${CMAKE_REQUIRED_LIBRARIES}")
+      message(STATUS "CMAKE_REQUIRED_INCLUDES: ${CMAKE_REQUIRED_INCLUDES}")
+      message(STATUS "Check in CMakeFiles/CMakeError.log to figure out why it fails")
+    endif()
   endif()
+  set(CMAKE_REQUIRED_INCLUDES)
+  set(CMAKE_REQUIRED_FLAGS)
+  set(CMAKE_REQUIRED_LIBRARIES)
+endif(METIS_LIBRARIES)
 
+if (METIS_LIBRARIES)
+  list(GET METIS_LIBRARIES 0 first_lib)
+  get_filename_component(first_lib_path "${first_lib}" PATH)
+  if (${first_lib_path} MATCHES "/lib(32|64)?$")
+    string(REGEX REPLACE "/lib(32|64)?$" "" not_cached_dir "${first_lib_path}")
+    set(METIS_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of METIS library" FORCE)
+  else()
+    set(METIS_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of METIS library" FORCE)
+  endif()
+endif()
+mark_as_advanced(METIS_DIR)
+mark_as_advanced(METIS_DIR_FOUND)
 
-find_library(METIS_LIBRARIES metis PATHS $ENV{METISDIR} ${LIB_INSTALL_DIR} PATH_SUFFIXES lib)
-
+# check that METIS has been found
+# ---------------------------------
 include(FindPackageHandleStandardArgs)
 find_package_handle_standard_args(METIS DEFAULT_MSG
-                                  METIS_INCLUDES METIS_LIBRARIES METIS_VERSION_OK)
-
-mark_as_advanced(METIS_INCLUDES METIS_LIBRARIES)
+  METIS_LIBRARIES
+  METIS_WORKS)
+#
+# TODO: Add possibility to check for specific functions in the library
+#
diff --git a/cmake/FindPTSCOTCH.cmake b/cmake/FindPTSCOTCH.cmake
new file mode 100644
index 000000000..1396d0582
--- /dev/null
+++ b/cmake/FindPTSCOTCH.cmake
@@ -0,0 +1,423 @@
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2016 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find PTSCOTCH include dirs and libraries
+# Use this module by invoking find_package with the form:
+#  find_package(PTSCOTCH
+#               [REQUIRED]             # Fail with error if ptscotch is not found
+#               [COMPONENTS <comp1> <comp2> ...] # dependencies
+#              )
+#
+#  PTSCOTCH depends on the following libraries:
+#   - Threads
+#   - MPI
+#
+#  COMPONENTS can be some of the following:
+#   - ESMUMPS: to activate detection of PT-Scotch with the esmumps interface
+#
+# This module finds headers and ptscotch library.
+# Results are reported in variables:
+#  PTSCOTCH_FOUND            - True if headers and requested libraries were found
+#  PTSCOTCH_LINKER_FLAGS     - list of required linker flags (excluding -l and -L)
+#  PTSCOTCH_INCLUDE_DIRS     - ptscotch include directories
+#  PTSCOTCH_LIBRARY_DIRS     - Link directories for ptscotch libraries
+#  PTSCOTCH_LIBRARIES        - ptscotch component libraries to be linked
+#  PTSCOTCH_INCLUDE_DIRS_DEP - ptscotch + dependencies include directories
+#  PTSCOTCH_LIBRARY_DIRS_DEP - ptscotch + dependencies link directories
+#  PTSCOTCH_LIBRARIES_DEP    - ptscotch libraries + dependencies
+#  PTSCOTCH_INTSIZE          - Number of octets occupied by a SCOTCH_Num
+#
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DPTSCOTCH=path/to/ptscotch):
+#  PTSCOTCH_DIR              - Where to find the base directory of ptscotch
+#  PTSCOTCH_INCDIR           - Where to find the header files
+#  PTSCOTCH_LIBDIR           - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: PTSCOTCH_DIR, PTSCOTCH_INCDIR, PTSCOTCH_LIBDIR
+
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013-2016 Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
+
+if (NOT PTSCOTCH_FOUND)
+  set(PTSCOTCH_DIR "" CACHE PATH "Installation directory of PTSCOTCH library")
+  if (NOT PTSCOTCH_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely PTSCOTCH_DIR, has been set to specify the install directory of PTSCOTCH")
+  endif()
+endif()
+
+# Set the version to find
+set(PTSCOTCH_LOOK_FOR_ESMUMPS OFF)
+
+if( PTSCOTCH_FIND_COMPONENTS )
+  foreach( component ${PTSCOTCH_FIND_COMPONENTS} )
+    if (${component} STREQUAL "ESMUMPS")
+      # means we look for esmumps library
+      set(PTSCOTCH_LOOK_FOR_ESMUMPS ON)
+    endif()
+  endforeach()
+endif()
+
+# PTSCOTCH depends on Threads, try to find it
+if (NOT THREADS_FOUND)
+  if (PTSCOTCH_FIND_REQUIRED)
+    find_package(Threads REQUIRED)
+  else()
+    find_package(Threads)
+  endif()
+endif()
+
+# PTSCOTCH depends on MPI, try to find it
+if (NOT MPI_FOUND)
+  if (PTSCOTCH_FIND_REQUIRED)
+    find_package(MPI REQUIRED)
+  else()
+    find_package(MPI)
+  endif()
+endif()
+
+# Looking for include
+# -------------------
+
+# Add system include paths to search include
+# ------------------------------------------
+unset(_inc_env)
+set(ENV_PTSCOTCH_DIR "$ENV{PTSCOTCH_DIR}")
+set(ENV_PTSCOTCH_INCDIR "$ENV{PTSCOTCH_INCDIR}")
+if(ENV_PTSCOTCH_INCDIR)
+  list(APPEND _inc_env "${ENV_PTSCOTCH_INCDIR}")
+elseif(ENV_PTSCOTCH_DIR)
+  list(APPEND _inc_env "${ENV_PTSCOTCH_DIR}")
+  list(APPEND _inc_env "${ENV_PTSCOTCH_DIR}/include")
+  list(APPEND _inc_env "${ENV_PTSCOTCH_DIR}/include/ptscotch")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
+  else()
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+  endif()
+endif()
+list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(REMOVE_DUPLICATES _inc_env)
+
+
+# Try to find the ptscotch header in the given paths
+# -------------------------------------------------
+
+set(PTSCOTCH_hdrs_to_find "ptscotch.h;scotch.h")
+
+# call cmake macro to find the header path
+if(PTSCOTCH_INCDIR)
+  foreach(ptscotch_hdr ${PTSCOTCH_hdrs_to_find})
+    set(PTSCOTCH_${ptscotch_hdr}_DIRS "PTSCOTCH_${ptscotch_hdr}_DIRS-NOTFOUND")
+    find_path(PTSCOTCH_${ptscotch_hdr}_DIRS
+      NAMES ${ptscotch_hdr}
+      HINTS ${PTSCOTCH_INCDIR})
+    mark_as_advanced(PTSCOTCH_${ptscotch_hdr}_DIRS)
+  endforeach()
+else()
+  if(PTSCOTCH_DIR)
+    foreach(ptscotch_hdr ${PTSCOTCH_hdrs_to_find})
+      set(PTSCOTCH_${ptscotch_hdr}_DIRS "PTSCOTCH_${ptscotch_hdr}_DIRS-NOTFOUND")
+      find_path(PTSCOTCH_${ptscotch_hdr}_DIRS
+	NAMES ${ptscotch_hdr}
+	HINTS ${PTSCOTCH_DIR}
+	PATH_SUFFIXES "include" "include/scotch")
+      mark_as_advanced(PTSCOTCH_${ptscotch_hdr}_DIRS)
+    endforeach()
+  else()
+    foreach(ptscotch_hdr ${PTSCOTCH_hdrs_to_find})
+      set(PTSCOTCH_${ptscotch_hdr}_DIRS "PTSCOTCH_${ptscotch_hdr}_DIRS-NOTFOUND")
+      find_path(PTSCOTCH_${ptscotch_hdr}_DIRS
+	NAMES ${ptscotch_hdr}
+	HINTS ${_inc_env}
+	PATH_SUFFIXES "scotch")
+      mark_as_advanced(PTSCOTCH_${ptscotch_hdr}_DIRS)
+    endforeach()
+  endif()
+endif()
+
+# If found, add path to cmake variable
+# ------------------------------------
+foreach(ptscotch_hdr ${PTSCOTCH_hdrs_to_find})
+  if (PTSCOTCH_${ptscotch_hdr}_DIRS)
+    list(APPEND PTSCOTCH_INCLUDE_DIRS "${PTSCOTCH_${ptscotch_hdr}_DIRS}")
+  else ()
+    set(PTSCOTCH_INCLUDE_DIRS "PTSCOTCH_INCLUDE_DIRS-NOTFOUND")
+    if (NOT PTSCOTCH_FIND_QUIETLY)
+      message(STATUS "Looking for ptscotch -- ${ptscotch_hdr} not found")
+    endif()
+  endif()
+endforeach()
+list(REMOVE_DUPLICATES PTSCOTCH_INCLUDE_DIRS)
+
+# Looking for lib
+# ---------------
+
+# Add system library paths to search lib
+# --------------------------------------
+unset(_lib_env)
+set(ENV_PTSCOTCH_LIBDIR "$ENV{PTSCOTCH_LIBDIR}")
+if(ENV_PTSCOTCH_LIBDIR)
+  list(APPEND _lib_env "${ENV_PTSCOTCH_LIBDIR}")
+elseif(ENV_PTSCOTCH_DIR)
+  list(APPEND _lib_env "${ENV_PTSCOTCH_DIR}")
+  list(APPEND _lib_env "${ENV_PTSCOTCH_DIR}/lib")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _lib_env "$ENV{LIB}")
+  else()
+    if(APPLE)
+      string(REPLACE ":" ";" _lib_env "$ENV{DYLD_LIBRARY_PATH}")
+    else()
+      string(REPLACE ":" ";" _lib_env "$ENV{LD_LIBRARY_PATH}")
+    endif()
+    list(APPEND _lib_env "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+    list(APPEND _lib_env "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+  endif()
+endif()
+list(REMOVE_DUPLICATES _lib_env)
+
+# Try to find the ptscotch lib in the given paths
+# ----------------------------------------------
+
+set(PTSCOTCH_libs_to_find "ptscotch;ptscotcherr")
+if (PTSCOTCH_LOOK_FOR_ESMUMPS)
+  list(INSERT PTSCOTCH_libs_to_find 0 "ptesmumps")
+  list(APPEND PTSCOTCH_libs_to_find   "esmumps"  )
+endif()
+list(APPEND PTSCOTCH_libs_to_find "scotch;scotcherr")
+
+# call cmake macro to find the lib path
+if(PTSCOTCH_LIBDIR)
+  foreach(ptscotch_lib ${PTSCOTCH_libs_to_find})
+    set(PTSCOTCH_${ptscotch_lib}_LIBRARY "PTSCOTCH_${ptscotch_lib}_LIBRARY-NOTFOUND")
+    find_library(PTSCOTCH_${ptscotch_lib}_LIBRARY
+      NAMES ${ptscotch_lib}
+      HINTS ${PTSCOTCH_LIBDIR})
+  endforeach()
+else()
+  if(PTSCOTCH_DIR)
+    foreach(ptscotch_lib ${PTSCOTCH_libs_to_find})
+      set(PTSCOTCH_${ptscotch_lib}_LIBRARY "PTSCOTCH_${ptscotch_lib}_LIBRARY-NOTFOUND")
+      find_library(PTSCOTCH_${ptscotch_lib}_LIBRARY
+	NAMES ${ptscotch_lib}
+	HINTS ${PTSCOTCH_DIR}
+	PATH_SUFFIXES lib lib32 lib64)
+    endforeach()
+  else()
+    foreach(ptscotch_lib ${PTSCOTCH_libs_to_find})
+      set(PTSCOTCH_${ptscotch_lib}_LIBRARY "PTSCOTCH_${ptscotch_lib}_LIBRARY-NOTFOUND")
+      find_library(PTSCOTCH_${ptscotch_lib}_LIBRARY
+	NAMES ${ptscotch_lib}
+	HINTS ${_lib_env})
+    endforeach()
+  endif()
+endif()
+
+set(PTSCOTCH_LIBRARIES "")
+set(PTSCOTCH_LIBRARY_DIRS "")
+# If found, add path to cmake variable
+# ------------------------------------
+foreach(ptscotch_lib ${PTSCOTCH_libs_to_find})
+
+  if (PTSCOTCH_${ptscotch_lib}_LIBRARY)
+    get_filename_component(${ptscotch_lib}_lib_path "${PTSCOTCH_${ptscotch_lib}_LIBRARY}" PATH)
+    # set cmake variables
+    list(APPEND PTSCOTCH_LIBRARIES "${PTSCOTCH_${ptscotch_lib}_LIBRARY}")
+    list(APPEND PTSCOTCH_LIBRARY_DIRS "${${ptscotch_lib}_lib_path}")
+  else ()
+    list(APPEND PTSCOTCH_LIBRARIES "${PTSCOTCH_${ptscotch_lib}_LIBRARY}")
+    if (NOT PTSCOTCH_FIND_QUIETLY)
+      message(STATUS "Looking for ptscotch -- lib ${ptscotch_lib} not found")
+    endif()
+  endif ()
+
+  mark_as_advanced(PTSCOTCH_${ptscotch_lib}_LIBRARY)
+
+endforeach()
+list(REMOVE_DUPLICATES PTSCOTCH_LIBRARY_DIRS)
+
+# check a function to validate the find
+if(PTSCOTCH_LIBRARIES)
+
+  set(REQUIRED_LDFLAGS)
+  set(REQUIRED_INCDIRS)
+  set(REQUIRED_LIBDIRS)
+  set(REQUIRED_LIBS)
+
+  # PTSCOTCH
+  if (PTSCOTCH_INCLUDE_DIRS)
+    set(REQUIRED_INCDIRS  "${PTSCOTCH_INCLUDE_DIRS}")
+  endif()
+  if (PTSCOTCH_LIBRARY_DIRS)
+    set(REQUIRED_LIBDIRS "${PTSCOTCH_LIBRARY_DIRS}")
+  endif()
+  set(REQUIRED_LIBS "${PTSCOTCH_LIBRARIES}")
+  # MPI
+  if (MPI_FOUND)
+    if (MPI_C_INCLUDE_PATH)
+      list(APPEND CMAKE_REQUIRED_INCLUDES "${MPI_C_INCLUDE_PATH}")
+    endif()
+    if (MPI_C_LINK_FLAGS)
+      if (${MPI_C_LINK_FLAGS} MATCHES "  -")
+	string(REGEX REPLACE " -" "-" MPI_C_LINK_FLAGS ${MPI_C_LINK_FLAGS})
+      endif()
+      list(APPEND REQUIRED_LDFLAGS "${MPI_C_LINK_FLAGS}")
+    endif()
+    list(APPEND REQUIRED_LIBS "${MPI_C_LIBRARIES}")
+  endif()
+  # THREADS
+  if(CMAKE_THREAD_LIBS_INIT)
+    list(APPEND REQUIRED_LIBS "${CMAKE_THREAD_LIBS_INIT}")
+  endif()
+  set(Z_LIBRARY "Z_LIBRARY-NOTFOUND")
+  find_library(Z_LIBRARY NAMES z)
+  mark_as_advanced(Z_LIBRARY)
+  if(Z_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lz")
+  endif()
+  set(M_LIBRARY "M_LIBRARY-NOTFOUND")
+  find_library(M_LIBRARY NAMES m)
+  mark_as_advanced(M_LIBRARY)
+  if(M_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lm")
+  endif()
+  set(RT_LIBRARY "RT_LIBRARY-NOTFOUND")
+  find_library(RT_LIBRARY NAMES rt)
+  mark_as_advanced(RT_LIBRARY)
+  if(RT_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lrt")
+  endif()
+
+  # set required libraries for link
+  set(CMAKE_REQUIRED_INCLUDES "${REQUIRED_INCDIRS}")
+  set(CMAKE_REQUIRED_LIBRARIES)
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LDFLAGS}")
+  foreach(lib_dir ${REQUIRED_LIBDIRS})
+    list(APPEND CMAKE_REQUIRED_LIBRARIES "-L${lib_dir}")
+  endforeach()
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LIBS}")
+  list(APPEND CMAKE_REQUIRED_FLAGS "${REQUIRED_FLAGS}")
+  string(REGEX REPLACE "^ -" "-" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+
+  # test link
+  unset(PTSCOTCH_WORKS CACHE)
+  include(CheckFunctionExists)
+  check_function_exists(SCOTCH_dgraphInit PTSCOTCH_WORKS)
+  mark_as_advanced(PTSCOTCH_WORKS)
+
+  if(PTSCOTCH_WORKS)
+    # save link with dependencies
+    set(PTSCOTCH_LIBRARIES_DEP "${REQUIRED_LIBS}")
+    set(PTSCOTCH_LIBRARY_DIRS_DEP "${REQUIRED_LIBDIRS}")
+    set(PTSCOTCH_INCLUDE_DIRS_DEP "${REQUIRED_INCDIRS}")
+    set(PTSCOTCH_LINKER_FLAGS "${REQUIRED_LDFLAGS}")
+    list(REMOVE_DUPLICATES PTSCOTCH_LIBRARY_DIRS_DEP)
+    list(REMOVE_DUPLICATES PTSCOTCH_INCLUDE_DIRS_DEP)
+    list(REMOVE_DUPLICATES PTSCOTCH_LINKER_FLAGS)
+  else()
+    if(NOT PTSCOTCH_FIND_QUIETLY)
+      message(STATUS "Looking for PTSCOTCH : test of SCOTCH_dgraphInit with PTSCOTCH library fails")
+      message(STATUS "CMAKE_REQUIRED_LIBRARIES: ${CMAKE_REQUIRED_LIBRARIES}")
+      message(STATUS "CMAKE_REQUIRED_INCLUDES: ${CMAKE_REQUIRED_INCLUDES}")
+      message(STATUS "Check in CMakeFiles/CMakeError.log to figure out why it fails")
+    endif()
+  endif()
+  set(CMAKE_REQUIRED_INCLUDES)
+  set(CMAKE_REQUIRED_FLAGS)
+  set(CMAKE_REQUIRED_LIBRARIES)
+endif(PTSCOTCH_LIBRARIES)
+
+if (PTSCOTCH_LIBRARIES)
+  list(GET PTSCOTCH_LIBRARIES 0 first_lib)
+  get_filename_component(first_lib_path "${first_lib}" PATH)
+  if (${first_lib_path} MATCHES "/lib(32|64)?$")
+    string(REGEX REPLACE "/lib(32|64)?$" "" not_cached_dir "${first_lib_path}")
+    set(PTSCOTCH_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of PTSCOTCH library" FORCE)
+  else()
+    set(PTSCOTCH_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of PTSCOTCH library" FORCE)
+  endif()
+endif()
+mark_as_advanced(PTSCOTCH_DIR)
+mark_as_advanced(PTSCOTCH_DIR_FOUND)
+
+# Check the size of SCOTCH_Num
+# ---------------------------------
+set(CMAKE_REQUIRED_INCLUDES ${PTSCOTCH_INCLUDE_DIRS})
+
+include(CheckCSourceRuns)
+#stdio.h and stdint.h should be included by scotch.h directly
+set(PTSCOTCH_C_TEST_SCOTCH_Num_4 "
+#include <stdio.h>
+#include <stdint.h>
+#include <ptscotch.h>
+int main(int argc, char **argv) {
+  if (sizeof(SCOTCH_Num) == 4)
+    return 0;
+  else
+    return 1;
+}
+")
+
+set(PTSCOTCH_C_TEST_SCOTCH_Num_8 "
+#include <stdio.h>
+#include <stdint.h>
+#include <ptscotch.h>
+int main(int argc, char **argv) {
+  if (sizeof(SCOTCH_Num) == 8)
+    return 0;
+  else
+    return 1;
+}
+")
+check_c_source_runs("${PTSCOTCH_C_TEST_SCOTCH_Num_4}" PTSCOTCH_Num_4)
+if(NOT PTSCOTCH_Num_4)
+  check_c_source_runs("${PTSCOTCH_C_TEST_SCOTCH_Num_8}" PTSCOTCH_Num_8)
+  if(NOT PTSCOTCH_Num_8)
+    set(PTSCOTCH_INTSIZE -1)
+  else()
+    set(PTSCOTCH_INTSIZE 8)
+  endif()
+else()
+  set(PTSCOTCH_INTSIZE 4)
+endif()
+set(CMAKE_REQUIRED_INCLUDES "")
+
+# check that PTSCOTCH has been found
+# ---------------------------------
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(PTSCOTCH DEFAULT_MSG
+  PTSCOTCH_LIBRARIES
+  PTSCOTCH_WORKS)
+#
+# TODO: Add possibility to check for specific functions in the library
+#
diff --git a/cmake/FindPastix.cmake b/cmake/FindPastix.cmake
index e2e6c810d..470477fdc 100644
--- a/cmake/FindPastix.cmake
+++ b/cmake/FindPastix.cmake
@@ -1,25 +1,704 @@
-# Pastix lib requires linking to a blas library.
-# It is up to the user of this module to find a BLAS and link to it.
-# Pastix requires SCOTCH or METIS (partitioning and reordering tools) as well
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2014 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find PASTIX include dirs and libraries
+# Use this module by invoking find_package with the form:
+#  find_package(PASTIX
+#               [REQUIRED] # Fail with error if pastix is not found
+#               [COMPONENTS <comp1> <comp2> ...] # dependencies
+#              )
+#
+#  PASTIX depends on the following libraries:
+#   - Threads, m, rt
+#   - MPI
+#   - HWLOC
+#   - BLAS
+#
+#  COMPONENTS are optional libraries PASTIX could be linked with,
+#  Use it to drive detection of a specific compilation chain
+#  COMPONENTS can be some of the following:
+#   - MPI: to activate detection of the parallel MPI version (default)
+#        it looks for Threads, HWLOC, BLAS, MPI and ScaLAPACK libraries
+#   - SEQ: to activate detection of the sequential version (exclude MPI version)
+#   - STARPU: to activate detection of StarPU version
+#   it looks for MPI version of StarPU (default behaviour)
+#   if SEQ and STARPU are given, it looks for a StarPU without MPI
+#   - STARPU_CUDA: to activate detection of StarPU with CUDA
+#   - STARPU_FXT: to activate detection of StarPU with FxT
+#   - SCOTCH: to activate detection of PASTIX linked with SCOTCH
+#   - PTSCOTCH: to activate detection of PASTIX linked with SCOTCH
+#   - METIS: to activate detection of PASTIX linked with SCOTCH
+#
+# This module finds headers and pastix library.
+# Results are reported in variables:
+#  PASTIX_FOUND            - True if headers and requested libraries were found
+#  PASTIX_LINKER_FLAGS     - list of required linker flags (excluding -l and -L)
+#  PASTIX_INCLUDE_DIRS     - pastix include directories
+#  PASTIX_LIBRARY_DIRS     - Link directories for pastix libraries
+#  PASTIX_LIBRARIES        - pastix libraries
+#  PASTIX_INCLUDE_DIRS_DEP - pastix + dependencies include directories
+#  PASTIX_LIBRARY_DIRS_DEP - pastix + dependencies link directories
+#  PASTIX_LIBRARIES_DEP    - pastix libraries + dependencies
+#
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DPASTIX_DIR=path/to/pastix):
+#  PASTIX_DIR              - Where to find the base directory of pastix
+#  PASTIX_INCDIR           - Where to find the header files
+#  PASTIX_LIBDIR           - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: PASTIX_DIR, PASTIX_INCDIR, PASTIX_LIBDIR
 
-if (PASTIX_INCLUDES AND PASTIX_LIBRARIES)
-  set(PASTIX_FIND_QUIETLY TRUE)
-endif (PASTIX_INCLUDES AND PASTIX_LIBRARIES)
-
-find_path(PASTIX_INCLUDES
-  NAMES
-  pastix_nompi.h
-  PATHS
-  $ENV{PASTIXDIR}
-  ${INCLUDE_INSTALL_DIR}
-)
-
-find_library(PASTIX_LIBRARIES pastix PATHS $ENV{PASTIXDIR} ${LIB_INSTALL_DIR})
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013      Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
 
 
+if (NOT PASTIX_FOUND)
+  set(PASTIX_DIR "" CACHE PATH "Installation directory of PASTIX library")
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely PASTIX_DIR, has been set to specify the install directory of PASTIX")
+  endif()
+endif()
 
+# Set the version to find
+set(PASTIX_LOOK_FOR_MPI ON)
+set(PASTIX_LOOK_FOR_SEQ OFF)
+set(PASTIX_LOOK_FOR_STARPU OFF)
+set(PASTIX_LOOK_FOR_STARPU_CUDA OFF)
+set(PASTIX_LOOK_FOR_STARPU_FXT OFF)
+set(PASTIX_LOOK_FOR_SCOTCH ON)
+set(PASTIX_LOOK_FOR_PTSCOTCH OFF)
+set(PASTIX_LOOK_FOR_METIS OFF)
+
+if( PASTIX_FIND_COMPONENTS )
+  foreach( component ${PASTIX_FIND_COMPONENTS} )
+    if (${component} STREQUAL "SEQ")
+      # means we look for the sequential version of PaStiX (without MPI)
+      set(PASTIX_LOOK_FOR_SEQ ON)
+      set(PASTIX_LOOK_FOR_MPI OFF)
+    endif()
+    if (${component} STREQUAL "MPI")
+      # means we look for the MPI version of PaStiX (default)
+      set(PASTIX_LOOK_FOR_SEQ OFF)
+      set(PASTIX_LOOK_FOR_MPI ON)
+    endif()
+    if (${component} STREQUAL "STARPU")
+      # means we look for PaStiX with StarPU
+      set(PASTIX_LOOK_FOR_STARPU ON)
+    endif()
+    if (${component} STREQUAL "STARPU_CUDA")
+      # means we look for PaStiX with StarPU + CUDA
+      set(PASTIX_LOOK_FOR_STARPU ON)
+      set(PASTIX_LOOK_FOR_STARPU_CUDA ON)
+    endif()
+    if (${component} STREQUAL "STARPU_FXT")
+      # means we look for PaStiX with StarPU + FxT
+      set(PASTIX_LOOK_FOR_STARPU_FXT ON)
+    endif()
+    if (${component} STREQUAL "SCOTCH")
+      set(PASTIX_LOOK_FOR_SCOTCH ON)
+    endif()
+    if (${component} STREQUAL "SCOTCH")
+      set(PASTIX_LOOK_FOR_PTSCOTCH ON)
+    endif()
+    if (${component} STREQUAL "METIS")
+      set(PASTIX_LOOK_FOR_METIS ON)
+    endif()
+  endforeach()
+endif()
+
+# Dependencies detection
+# ----------------------
+
+
+# Required dependencies
+# ---------------------
+
+if (NOT PASTIX_FIND_QUIETLY)
+  message(STATUS "Looking for PASTIX - Try to detect pthread")
+endif()
+if (PASTIX_FIND_REQUIRED)
+  find_package(Threads REQUIRED QUIET)
+else()
+  find_package(Threads QUIET)
+endif()
+set(PASTIX_EXTRA_LIBRARIES "")
+if( THREADS_FOUND )
+  list(APPEND PASTIX_EXTRA_LIBRARIES ${CMAKE_THREAD_LIBS_INIT})
+endif ()
+
+# Add math library to the list of extra
+# it normally exists on all common systems provided with a C compiler
+if (NOT PASTIX_FIND_QUIETLY)
+  message(STATUS "Looking for PASTIX - Try to detect libm")
+endif()
+set(PASTIX_M_LIBRARIES "")
+if(UNIX OR WIN32)
+  find_library(
+    PASTIX_M_m_LIBRARY
+    NAMES m
+    )
+  mark_as_advanced(PASTIX_M_m_LIBRARY)
+  if (PASTIX_M_m_LIBRARY)
+    list(APPEND PASTIX_M_LIBRARIES "${PASTIX_M_m_LIBRARY}")
+    list(APPEND PASTIX_EXTRA_LIBRARIES "${PASTIX_M_m_LIBRARY}")
+  else()
+    if (PASTIX_FIND_REQUIRED)
+      message(FATAL_ERROR "Could NOT find libm on your system."
+	"Are you sure to a have a C compiler installed?")
+    endif()
+  endif()
+endif()
+
+# Try to find librt (libposix4 - POSIX.1b Realtime Extensions library)
+# on Unix systems except Apple ones because it does not exist on it
+if (NOT PASTIX_FIND_QUIETLY)
+  message(STATUS "Looking for PASTIX - Try to detect librt")
+endif()
+set(PASTIX_RT_LIBRARIES "")
+if(UNIX AND NOT APPLE)
+  find_library(
+    PASTIX_RT_rt_LIBRARY
+    NAMES rt
+    )
+  mark_as_advanced(PASTIX_RT_rt_LIBRARY)
+  if (PASTIX_RT_rt_LIBRARY)
+    list(APPEND PASTIX_RT_LIBRARIES "${PASTIX_RT_rt_LIBRARY}")
+    list(APPEND PASTIX_EXTRA_LIBRARIES "${PASTIX_RT_rt_LIBRARY}")
+  else()
+    if (PASTIX_FIND_REQUIRED)
+      message(FATAL_ERROR "Could NOT find librt on your system")
+    endif()
+  endif()
+endif()
+
+# PASTIX depends on HWLOC
+#------------------------
+if (NOT PASTIX_FIND_QUIETLY)
+  message(STATUS "Looking for PASTIX - Try to detect HWLOC")
+endif()
+if (PASTIX_FIND_REQUIRED)
+  find_package(HWLOC REQUIRED QUIET)
+else()
+  find_package(HWLOC QUIET)
+endif()
+
+# PASTIX depends on BLAS
+#-----------------------
+if (NOT PASTIX_FIND_QUIETLY)
+  message(STATUS "Looking for PASTIX - Try to detect BLAS")
+endif()
+if (PASTIX_FIND_REQUIRED)
+  find_package(BLASEXT REQUIRED QUIET)
+else()
+  find_package(BLASEXT QUIET)
+endif()
+
+# Optional dependencies
+# ---------------------
+
+# PASTIX may depend on MPI
+#-------------------------
+if (NOT MPI_FOUND AND PASTIX_LOOK_FOR_MPI)
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for PASTIX - Try to detect MPI")
+  endif()
+  # allows to use an external mpi compilation by setting compilers with
+  # -DMPI_C_COMPILER=path/to/mpicc -DMPI_Fortran_COMPILER=path/to/mpif90
+  # at cmake configure
+  if(NOT MPI_C_COMPILER)
+    set(MPI_C_COMPILER mpicc)
+  endif()
+  if (PASTIX_FIND_REQUIRED AND PASTIX_FIND_REQUIRED_MPI)
+    find_package(MPI REQUIRED QUIET)
+  else()
+    find_package(MPI QUIET)
+  endif()
+  if (MPI_FOUND)
+    mark_as_advanced(MPI_LIBRARY)
+    mark_as_advanced(MPI_EXTRA_LIBRARY)
+  endif()
+endif (NOT MPI_FOUND AND PASTIX_LOOK_FOR_MPI)
+
+# PASTIX may depend on STARPU
+#----------------------------
+if( NOT STARPU_FOUND AND PASTIX_LOOK_FOR_STARPU)
+
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for PASTIX - Try to detect StarPU")
+  endif()
+
+  set(PASTIX_STARPU_VERSION "1.1" CACHE STRING "oldest STARPU version desired")
+
+  # create list of components in order to make a single call to find_package(starpu...)
+  # we explicitly need a StarPU version built with hwloc
+  set(STARPU_COMPONENT_LIST "HWLOC")
+
+  # StarPU may depend on MPI
+  # allows to use an external mpi compilation by setting compilers with
+  # -DMPI_C_COMPILER=path/to/mpicc -DMPI_Fortran_COMPILER=path/to/mpif90
+  # at cmake configure
+  if (PASTIX_LOOK_FOR_MPI)
+    if(NOT MPI_C_COMPILER)
+      set(MPI_C_COMPILER mpicc)
+    endif()
+    list(APPEND STARPU_COMPONENT_LIST "MPI")
+  endif()
+  if (PASTIX_LOOK_FOR_STARPU_CUDA)
+    list(APPEND STARPU_COMPONENT_LIST "CUDA")
+  endif()
+  if (PASTIX_LOOK_FOR_STARPU_FXT)
+    list(APPEND STARPU_COMPONENT_LIST "FXT")
+  endif()
+  # set the list of optional dependencies we may discover
+  if (PASTIX_FIND_REQUIRED AND PASTIX_FIND_REQUIRED_STARPU)
+    find_package(STARPU ${PASTIX_STARPU_VERSION} REQUIRED
+      COMPONENTS ${STARPU_COMPONENT_LIST})
+  else()
+    find_package(STARPU ${PASTIX_STARPU_VERSION}
+      COMPONENTS ${STARPU_COMPONENT_LIST})
+  endif()
+
+endif( NOT STARPU_FOUND AND PASTIX_LOOK_FOR_STARPU)
+
+# PASTIX may depends on SCOTCH
+#-----------------------------
+if (NOT SCOTCH_FOUND AND PASTIX_LOOK_FOR_SCOTCH)
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for PASTIX - Try to detect SCOTCH")
+  endif()
+  if (PASTIX_FIND_REQUIRED AND PASTIX_FIND_REQUIRED_SCOTCH)
+    find_package(SCOTCH REQUIRED QUIET)
+  else()
+    find_package(SCOTCH QUIET)
+  endif()
+endif()
+
+# PASTIX may depends on PTSCOTCH
+#-------------------------------
+if (NOT PTSCOTCH_FOUND AND PASTIX_LOOK_FOR_PTSCOTCH)
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for PASTIX - Try to detect PTSCOTCH")
+  endif()
+  if (PASTIX_FIND_REQUIRED AND PASTIX_FIND_REQUIRED_PTSCOTCH)
+    find_package(PTSCOTCH REQUIRED QUIET)
+  else()
+    find_package(PTSCOTCH QUIET)
+  endif()
+endif()
+
+# PASTIX may depends on METIS
+#----------------------------
+if (NOT METIS_FOUND AND PASTIX_LOOK_FOR_METIS)
+  if (NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for PASTIX - Try to detect METIS")
+  endif()
+  if (PASTIX_FIND_REQUIRED AND PASTIX_FIND_REQUIRED_METIS)
+    find_package(METIS REQUIRED QUIET)
+  else()
+    find_package(METIS QUIET)
+  endif()
+endif()
+
+# Error if pastix required and no partitioning lib found
+if (PASTIX_FIND_REQUIRED AND NOT SCOTCH_FOUND AND NOT PTSCOTCH_FOUND AND NOT METIS_FOUND)
+  message(FATAL_ERROR "Could NOT find any partitioning library on your system"
+    " (install scotch, ptscotch or metis)")
+endif()
+
+
+# Looking for PaStiX
+# ------------------
+
+# Looking for include
+# -------------------
+
+# Add system include paths to search include
+# ------------------------------------------
+unset(_inc_env)
+set(ENV_PASTIX_DIR "$ENV{PASTIX_DIR}")
+set(ENV_PASTIX_INCDIR "$ENV{PASTIX_INCDIR}")
+if(ENV_PASTIX_INCDIR)
+  list(APPEND _inc_env "${ENV_PASTIX_INCDIR}")
+elseif(ENV_PASTIX_DIR)
+  list(APPEND _inc_env "${ENV_PASTIX_DIR}")
+  list(APPEND _inc_env "${ENV_PASTIX_DIR}/include")
+  list(APPEND _inc_env "${ENV_PASTIX_DIR}/include/pastix")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
+  else()
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+  endif()
+endif()
+list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(REMOVE_DUPLICATES _inc_env)
+
+
+# Try to find the pastix header in the given paths
+# ---------------------------------------------------
+# call cmake macro to find the header path
+if(PASTIX_INCDIR)
+  set(PASTIX_pastix.h_DIRS "PASTIX_pastix.h_DIRS-NOTFOUND")
+  find_path(PASTIX_pastix.h_DIRS
+    NAMES pastix.h
+    HINTS ${PASTIX_INCDIR})
+else()
+  if(PASTIX_DIR)
+    set(PASTIX_pastix.h_DIRS "PASTIX_pastix.h_DIRS-NOTFOUND")
+    find_path(PASTIX_pastix.h_DIRS
+      NAMES pastix.h
+      HINTS ${PASTIX_DIR}
+      PATH_SUFFIXES "include" "include/pastix")
+  else()
+    set(PASTIX_pastix.h_DIRS "PASTIX_pastix.h_DIRS-NOTFOUND")
+    find_path(PASTIX_pastix.h_DIRS
+      NAMES pastix.h
+      HINTS ${_inc_env}
+      PATH_SUFFIXES "pastix")
+  endif()
+endif()
+mark_as_advanced(PASTIX_pastix.h_DIRS)
+
+# If found, add path to cmake variable
+# ------------------------------------
+if (PASTIX_pastix.h_DIRS)
+  set(PASTIX_INCLUDE_DIRS "${PASTIX_pastix.h_DIRS}")
+else ()
+  set(PASTIX_INCLUDE_DIRS "PASTIX_INCLUDE_DIRS-NOTFOUND")
+  if(NOT PASTIX_FIND_QUIETLY)
+    message(STATUS "Looking for pastix -- pastix.h not found")
+  endif()
+endif()
+
+
+# Looking for lib
+# ---------------
+
+# Add system library paths to search lib
+# --------------------------------------
+unset(_lib_env)
+set(ENV_PASTIX_LIBDIR "$ENV{PASTIX_LIBDIR}")
+if(ENV_PASTIX_LIBDIR)
+  list(APPEND _lib_env "${ENV_PASTIX_LIBDIR}")
+elseif(ENV_PASTIX_DIR)
+  list(APPEND _lib_env "${ENV_PASTIX_DIR}")
+  list(APPEND _lib_env "${ENV_PASTIX_DIR}/lib")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _lib_env "$ENV{LIB}")
+  else()
+    if(APPLE)
+      string(REPLACE ":" ";" _lib_env "$ENV{DYLD_LIBRARY_PATH}")
+    else()
+      string(REPLACE ":" ";" _lib_env "$ENV{LD_LIBRARY_PATH}")
+    endif()
+    list(APPEND _lib_env "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+    list(APPEND _lib_env "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+  endif()
+endif()
+list(REMOVE_DUPLICATES _lib_env)
+
+# Try to find the pastix lib in the given paths
+# ------------------------------------------------
+
+# create list of libs to find
+set(PASTIX_libs_to_find "pastix_murge;pastix")
+
+# call cmake macro to find the lib path
+if(PASTIX_LIBDIR)
+  foreach(pastix_lib ${PASTIX_libs_to_find})
+    set(PASTIX_${pastix_lib}_LIBRARY "PASTIX_${pastix_lib}_LIBRARY-NOTFOUND")
+    find_library(PASTIX_${pastix_lib}_LIBRARY
+      NAMES ${pastix_lib}
+      HINTS ${PASTIX_LIBDIR})
+  endforeach()
+else()
+  if(PASTIX_DIR)
+    foreach(pastix_lib ${PASTIX_libs_to_find})
+      set(PASTIX_${pastix_lib}_LIBRARY "PASTIX_${pastix_lib}_LIBRARY-NOTFOUND")
+      find_library(PASTIX_${pastix_lib}_LIBRARY
+	NAMES ${pastix_lib}
+	HINTS ${PASTIX_DIR}
+	PATH_SUFFIXES lib lib32 lib64)
+    endforeach()
+  else()
+    foreach(pastix_lib ${PASTIX_libs_to_find})
+      set(PASTIX_${pastix_lib}_LIBRARY "PASTIX_${pastix_lib}_LIBRARY-NOTFOUND")
+      find_library(PASTIX_${pastix_lib}_LIBRARY
+	NAMES ${pastix_lib}
+	HINTS ${_lib_env})
+    endforeach()
+  endif()
+endif()
+
+# If found, add path to cmake variable
+# ------------------------------------
+foreach(pastix_lib ${PASTIX_libs_to_find})
+
+  get_filename_component(${pastix_lib}_lib_path ${PASTIX_${pastix_lib}_LIBRARY} PATH)
+  # set cmake variables (respects naming convention)
+  if (PASTIX_LIBRARIES)
+    list(APPEND PASTIX_LIBRARIES "${PASTIX_${pastix_lib}_LIBRARY}")
+  else()
+    set(PASTIX_LIBRARIES "${PASTIX_${pastix_lib}_LIBRARY}")
+  endif()
+  if (PASTIX_LIBRARY_DIRS)
+    list(APPEND PASTIX_LIBRARY_DIRS "${${pastix_lib}_lib_path}")
+  else()
+    set(PASTIX_LIBRARY_DIRS "${${pastix_lib}_lib_path}")
+  endif()
+  mark_as_advanced(PASTIX_${pastix_lib}_LIBRARY)
+
+endforeach(pastix_lib ${PASTIX_libs_to_find})
+
+# check a function to validate the find
+if(PASTIX_LIBRARIES)
+
+  set(REQUIRED_LDFLAGS)
+  set(REQUIRED_INCDIRS)
+  set(REQUIRED_LIBDIRS)
+  set(REQUIRED_LIBS)
+
+  # PASTIX
+  if (PASTIX_INCLUDE_DIRS)
+    set(REQUIRED_INCDIRS "${PASTIX_INCLUDE_DIRS}")
+  endif()
+  foreach(libdir ${PASTIX_LIBRARY_DIRS})
+    if (libdir)
+      list(APPEND REQUIRED_LIBDIRS "${libdir}")
+    endif()
+  endforeach()
+  set(REQUIRED_LIBS "${PASTIX_LIBRARIES}")
+  # STARPU
+  if (PASTIX_LOOK_FOR_STARPU AND STARPU_FOUND)
+    if (STARPU_INCLUDE_DIRS_DEP)
+      list(APPEND REQUIRED_INCDIRS "${STARPU_INCLUDE_DIRS_DEP}")
+    elseif (STARPU_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${STARPU_INCLUDE_DIRS}")
+    endif()
+    if(STARPU_LIBRARY_DIRS_DEP)
+      list(APPEND REQUIRED_LIBDIRS "${STARPU_LIBRARY_DIRS_DEP}")
+    elseif(STARPU_LIBRARY_DIRS)
+      list(APPEND REQUIRED_LIBDIRS "${STARPU_LIBRARY_DIRS}")
+    endif()
+    if (STARPU_LIBRARIES_DEP)
+      list(APPEND REQUIRED_LIBS "${STARPU_LIBRARIES_DEP}")
+    elseif (STARPU_LIBRARIES)
+      foreach(lib ${STARPU_LIBRARIES})
+	if (EXISTS ${lib} OR ${lib} MATCHES "^-")
+	  list(APPEND REQUIRED_LIBS "${lib}")
+	else()
+	  list(APPEND REQUIRED_LIBS "-l${lib}")
+	endif()
+      endforeach()
+    endif()
+  endif()
+  # CUDA
+  if (PASTIX_LOOK_FOR_STARPU_CUDA AND CUDA_FOUND)
+    if (CUDA_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${CUDA_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${CUDA_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    list(APPEND REQUIRED_LIBS "${CUDA_CUBLAS_LIBRARIES};${CUDA_LIBRARIES}")
+  endif()
+  # MPI
+  if (PASTIX_LOOK_FOR_MPI AND MPI_FOUND)
+    if (MPI_C_INCLUDE_PATH)
+      list(APPEND REQUIRED_INCDIRS "${MPI_C_INCLUDE_PATH}")
+    endif()
+    if (MPI_C_LINK_FLAGS)
+      if (${MPI_C_LINK_FLAGS} MATCHES "  -")
+	string(REGEX REPLACE " -" "-" MPI_C_LINK_FLAGS ${MPI_C_LINK_FLAGS})
+      endif()
+      list(APPEND REQUIRED_LDFLAGS "${MPI_C_LINK_FLAGS}")
+    endif()
+    list(APPEND REQUIRED_LIBS "${MPI_C_LIBRARIES}")
+  endif()
+  # HWLOC
+  if (HWLOC_FOUND)
+    if (HWLOC_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${HWLOC_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${HWLOC_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    foreach(lib ${HWLOC_LIBRARIES})
+      if (EXISTS ${lib} OR ${lib} MATCHES "^-")
+	list(APPEND REQUIRED_LIBS "${lib}")
+      else()
+	list(APPEND REQUIRED_LIBS "-l${lib}")
+      endif()
+    endforeach()
+  endif()
+  # BLAS
+  if (BLAS_FOUND)
+    if (BLAS_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${BLAS_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${BLAS_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    list(APPEND REQUIRED_LIBS "${BLAS_LIBRARIES}")
+    if (BLAS_LINKER_FLAGS)
+      list(APPEND REQUIRED_LDFLAGS "${BLAS_LINKER_FLAGS}")
+    endif()
+  endif()
+  # SCOTCH
+  if (PASTIX_LOOK_FOR_SCOTCH AND SCOTCH_FOUND)
+    if (SCOTCH_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${SCOTCH_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${SCOTCH_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    list(APPEND REQUIRED_LIBS "${SCOTCH_LIBRARIES}")
+  endif()
+  # PTSCOTCH
+  if (PASTIX_LOOK_FOR_PTSCOTCH AND PTSCOTCH_FOUND)
+    if (PTSCOTCH_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${PTSCOTCH_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${PTSCOTCH_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    list(APPEND REQUIRED_LIBS "${PTSCOTCH_LIBRARIES}")
+  endif()
+  # METIS
+  if (PASTIX_LOOK_FOR_METIS AND METIS_FOUND)
+    if (METIS_INCLUDE_DIRS)
+      list(APPEND REQUIRED_INCDIRS "${METIS_INCLUDE_DIRS}")
+    endif()
+    foreach(libdir ${METIS_LIBRARY_DIRS})
+      if (libdir)
+	list(APPEND REQUIRED_LIBDIRS "${libdir}")
+      endif()
+    endforeach()
+    list(APPEND REQUIRED_LIBS "${METIS_LIBRARIES}")
+  endif()
+  # Fortran
+  if (CMAKE_C_COMPILER_ID MATCHES "GNU")
+    find_library(
+      FORTRAN_gfortran_LIBRARY
+      NAMES gfortran
+      HINTS ${_lib_env}
+      )
+    mark_as_advanced(FORTRAN_gfortran_LIBRARY)
+    if (FORTRAN_gfortran_LIBRARY)
+      list(APPEND REQUIRED_LIBS "${FORTRAN_gfortran_LIBRARY}")
+    endif()
+  elseif (CMAKE_C_COMPILER_ID MATCHES "Intel")
+    find_library(
+      FORTRAN_ifcore_LIBRARY
+      NAMES ifcore
+      HINTS ${_lib_env}
+      )
+    mark_as_advanced(FORTRAN_ifcore_LIBRARY)
+    if (FORTRAN_ifcore_LIBRARY)
+      list(APPEND REQUIRED_LIBS "${FORTRAN_ifcore_LIBRARY}")
+    endif()
+  endif()
+  # EXTRA LIBS such that pthread, m, rt
+  list(APPEND REQUIRED_LIBS ${PASTIX_EXTRA_LIBRARIES})
+
+  # set required libraries for link
+  set(CMAKE_REQUIRED_INCLUDES "${REQUIRED_INCDIRS}")
+  set(CMAKE_REQUIRED_LIBRARIES)
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LDFLAGS}")
+  foreach(lib_dir ${REQUIRED_LIBDIRS})
+    list(APPEND CMAKE_REQUIRED_LIBRARIES "-L${lib_dir}")
+  endforeach()
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LIBS}")
+  list(APPEND CMAKE_REQUIRED_FLAGS "${REQUIRED_FLAGS}")
+  string(REGEX REPLACE "^ -" "-" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+
+  # test link
+  unset(PASTIX_WORKS CACHE)
+  include(CheckFunctionExists)
+  check_function_exists(pastix PASTIX_WORKS)
+  mark_as_advanced(PASTIX_WORKS)
+
+  if(PASTIX_WORKS)
+    # save link with dependencies
+    set(PASTIX_LIBRARIES_DEP "${REQUIRED_LIBS}")
+    set(PASTIX_LIBRARY_DIRS_DEP "${REQUIRED_LIBDIRS}")
+    set(PASTIX_INCLUDE_DIRS_DEP "${REQUIRED_INCDIRS}")
+    set(PASTIX_LINKER_FLAGS "${REQUIRED_LDFLAGS}")
+    list(REMOVE_DUPLICATES PASTIX_LIBRARY_DIRS_DEP)
+    list(REMOVE_DUPLICATES PASTIX_INCLUDE_DIRS_DEP)
+    list(REMOVE_DUPLICATES PASTIX_LINKER_FLAGS)
+  else()
+    if(NOT PASTIX_FIND_QUIETLY)
+      message(STATUS "Looking for PASTIX : test of pastix() fails")
+      message(STATUS "CMAKE_REQUIRED_LIBRARIES: ${CMAKE_REQUIRED_LIBRARIES}")
+      message(STATUS "CMAKE_REQUIRED_INCLUDES: ${CMAKE_REQUIRED_INCLUDES}")
+      message(STATUS "Check in CMakeFiles/CMakeError.log to figure out why it fails")
+      message(STATUS "Maybe PASTIX is linked with specific libraries. "
+	"Have you tried with COMPONENTS (MPI/SEQ, STARPU, STARPU_CUDA, SCOTCH, PTSCOTCH, METIS)? "
+	"See the explanation in FindPASTIX.cmake.")
+    endif()
+  endif()
+  set(CMAKE_REQUIRED_INCLUDES)
+  set(CMAKE_REQUIRED_FLAGS)
+  set(CMAKE_REQUIRED_LIBRARIES)
+endif(PASTIX_LIBRARIES)
+
+if (PASTIX_LIBRARIES)
+  list(GET PASTIX_LIBRARIES 0 first_lib)
+  get_filename_component(first_lib_path "${first_lib}" PATH)
+  if (${first_lib_path} MATCHES "/lib(32|64)?$")
+    string(REGEX REPLACE "/lib(32|64)?$" "" not_cached_dir "${first_lib_path}")
+    set(PASTIX_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of PASTIX library" FORCE)
+  else()
+    set(PASTIX_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of PASTIX library" FORCE)
+  endif()
+endif()
+mark_as_advanced(PASTIX_DIR)
+mark_as_advanced(PASTIX_DIR_FOUND)
+
+# check that PASTIX has been found
+# ---------------------------------
 include(FindPackageHandleStandardArgs)
 find_package_handle_standard_args(PASTIX DEFAULT_MSG
-                                  PASTIX_INCLUDES PASTIX_LIBRARIES)
-
-mark_as_advanced(PASTIX_INCLUDES PASTIX_LIBRARIES)
+  PASTIX_LIBRARIES
+  PASTIX_WORKS)
diff --git a/cmake/FindScotch.cmake b/cmake/FindScotch.cmake
index 530340b16..89d295ac2 100644
--- a/cmake/FindScotch.cmake
+++ b/cmake/FindScotch.cmake
@@ -1,24 +1,369 @@
-# Pastix requires SCOTCH or METIS (partitioning and reordering tools)
+###
+#
+# @copyright (c) 2009-2014 The University of Tennessee and The University
+#                          of Tennessee Research Foundation.
+#                          All rights reserved.
+# @copyright (c) 2012-2014 Inria. All rights reserved.
+# @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
+#
+###
+#
+# - Find SCOTCH include dirs and libraries
+# Use this module by invoking find_package with the form:
+#  find_package(SCOTCH
+#               [REQUIRED]             # Fail with error if scotch is not found
+#               [COMPONENTS <comp1> <comp2> ...] # dependencies
+#              )
+#
+#  COMPONENTS can be some of the following:
+#   - ESMUMPS: to activate detection of Scotch with the esmumps interface
+#
+# This module finds headers and scotch library.
+# Results are reported in variables:
+#  SCOTCH_FOUND           - True if headers and requested libraries were found
+#  SCOTCH_INCLUDE_DIRS    - scotch include directories
+#  SCOTCH_LIBRARY_DIRS    - Link directories for scotch libraries
+#  SCOTCH_LIBRARIES       - scotch component libraries to be linked
+#  SCOTCH_INTSIZE         - Number of octets occupied by a SCOTCH_Num
+#
+# The user can give specific paths where to find the libraries adding cmake
+# options at configure (ex: cmake path/to/project -DSCOTCH=path/to/scotch):
+#  SCOTCH_DIR             - Where to find the base directory of scotch
+#  SCOTCH_INCDIR          - Where to find the header files
+#  SCOTCH_LIBDIR          - Where to find the library files
+# The module can also look for the following environment variables if paths
+# are not given as cmake variable: SCOTCH_DIR, SCOTCH_INCDIR, SCOTCH_LIBDIR
 
-if (SCOTCH_INCLUDES AND SCOTCH_LIBRARIES)
-  set(SCOTCH_FIND_QUIETLY TRUE)
-endif (SCOTCH_INCLUDES AND SCOTCH_LIBRARIES)
+#=============================================================================
+# Copyright 2012-2013 Inria
+# Copyright 2012-2013 Emmanuel Agullo
+# Copyright 2012-2013 Mathieu Faverge
+# Copyright 2012      Cedric Castagnede
+# Copyright 2013      Florent Pruvost
+#
+# Distributed under the OSI-approved BSD License (the "License");
+# see accompanying file MORSE-Copyright.txt for details.
+#
+# This software is distributed WITHOUT ANY WARRANTY; without even the
+# implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+# See the License for more information.
+#=============================================================================
+# (To distribute this file outside of Morse, substitute the full
+#  License text for the above reference.)
 
-find_path(SCOTCH_INCLUDES 
-  NAMES 
-  scotch.h 
-  PATHS 
-  $ENV{SCOTCHDIR} 
-  ${INCLUDE_INSTALL_DIR} 
-  PATH_SUFFIXES 
-  scotch
-)
+if (NOT SCOTCH_FOUND)
+  set(SCOTCH_DIR "" CACHE PATH "Installation directory of SCOTCH library")
+  if (NOT SCOTCH_FIND_QUIETLY)
+    message(STATUS "A cache variable, namely SCOTCH_DIR, has been set to specify the install directory of SCOTCH")
+  endif()
+endif()
+
+# Set the version to find
+set(SCOTCH_LOOK_FOR_ESMUMPS OFF)
+
+if( SCOTCH_FIND_COMPONENTS )
+  foreach( component ${SCOTCH_FIND_COMPONENTS} )
+    if (${component} STREQUAL "ESMUMPS")
+      # means we look for esmumps library
+      set(SCOTCH_LOOK_FOR_ESMUMPS ON)
+    endif()
+  endforeach()
+endif()
+
+# SCOTCH may depend on Threads, try to find it
+if (NOT THREADS_FOUND)
+  if (SCOTCH_FIND_REQUIRED)
+    find_package(Threads REQUIRED)
+  else()
+    find_package(Threads)
+  endif()
+endif()
+
+# Looking for include
+# -------------------
+
+# Add system include paths to search include
+# ------------------------------------------
+unset(_inc_env)
+set(ENV_SCOTCH_DIR "$ENV{SCOTCH_DIR}")
+set(ENV_SCOTCH_INCDIR "$ENV{SCOTCH_INCDIR}")
+if(ENV_SCOTCH_INCDIR)
+  list(APPEND _inc_env "${ENV_SCOTCH_INCDIR}")
+elseif(ENV_SCOTCH_DIR)
+  list(APPEND _inc_env "${ENV_SCOTCH_DIR}")
+  list(APPEND _inc_env "${ENV_SCOTCH_DIR}/include")
+  list(APPEND _inc_env "${ENV_SCOTCH_DIR}/include/scotch")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _inc_env "$ENV{INCLUDE}")
+  else()
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{C_INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{CPATH}")
+    list(APPEND _inc_env "${_path_env}")
+    string(REPLACE ":" ";" _path_env "$ENV{INCLUDE_PATH}")
+    list(APPEND _inc_env "${_path_env}")
+  endif()
+endif()
+list(APPEND _inc_env "${CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(APPEND _inc_env "${CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES}")
+list(REMOVE_DUPLICATES _inc_env)
 
 
-find_library(SCOTCH_LIBRARIES scotch PATHS $ENV{SCOTCHDIR} ${LIB_INSTALL_DIR})
+# Try to find the scotch header in the given paths
+# -------------------------------------------------
+# call cmake macro to find the header path
+if(SCOTCH_INCDIR)
+  set(SCOTCH_scotch.h_DIRS "SCOTCH_scotch.h_DIRS-NOTFOUND")
+  find_path(SCOTCH_scotch.h_DIRS
+    NAMES scotch.h
+    HINTS ${SCOTCH_INCDIR})
+else()
+  if(SCOTCH_DIR)
+    set(SCOTCH_scotch.h_DIRS "SCOTCH_scotch.h_DIRS-NOTFOUND")
+    find_path(SCOTCH_scotch.h_DIRS
+      NAMES scotch.h
+      HINTS ${SCOTCH_DIR}
+      PATH_SUFFIXES "include" "include/scotch")
+  else()
+    set(SCOTCH_scotch.h_DIRS "SCOTCH_scotch.h_DIRS-NOTFOUND")
+    find_path(SCOTCH_scotch.h_DIRS
+      NAMES scotch.h
+      HINTS ${_inc_env}
+      PATH_SUFFIXES "scotch")
+  endif()
+endif()
+mark_as_advanced(SCOTCH_scotch.h_DIRS)
 
+# If found, add path to cmake variable
+# ------------------------------------
+if (SCOTCH_scotch.h_DIRS)
+  set(SCOTCH_INCLUDE_DIRS "${SCOTCH_scotch.h_DIRS}")
+else ()
+  set(SCOTCH_INCLUDE_DIRS "SCOTCH_INCLUDE_DIRS-NOTFOUND")
+  if (NOT SCOTCH_FIND_QUIETLY)
+    message(STATUS "Looking for scotch -- scotch.h not found")
+  endif()
+endif()
+list(REMOVE_DUPLICATES SCOTCH_INCLUDE_DIRS)
+
+# Looking for lib
+# ---------------
+
+# Add system library paths to search lib
+# --------------------------------------
+unset(_lib_env)
+set(ENV_SCOTCH_LIBDIR "$ENV{SCOTCH_LIBDIR}")
+if(ENV_SCOTCH_LIBDIR)
+  list(APPEND _lib_env "${ENV_SCOTCH_LIBDIR}")
+elseif(ENV_SCOTCH_DIR)
+  list(APPEND _lib_env "${ENV_SCOTCH_DIR}")
+  list(APPEND _lib_env "${ENV_SCOTCH_DIR}/lib")
+else()
+  if(WIN32)
+    string(REPLACE ":" ";" _lib_env "$ENV{LIB}")
+  else()
+    if(APPLE)
+      string(REPLACE ":" ";" _lib_env "$ENV{DYLD_LIBRARY_PATH}")
+    else()
+      string(REPLACE ":" ";" _lib_env "$ENV{LD_LIBRARY_PATH}")
+    endif()
+    list(APPEND _lib_env "${CMAKE_PLATFORM_IMPLICIT_LINK_DIRECTORIES}")
+    list(APPEND _lib_env "${CMAKE_C_IMPLICIT_LINK_DIRECTORIES}")
+  endif()
+endif()
+list(REMOVE_DUPLICATES _lib_env)
+
+# Try to find the scotch lib in the given paths
+# ----------------------------------------------
+
+set(SCOTCH_libs_to_find "scotch;scotcherrexit")
+if (SCOTCH_LOOK_FOR_ESMUMPS)
+  list(INSERT SCOTCH_libs_to_find 0 "esmumps")
+endif()
+
+# call cmake macro to find the lib path
+if(SCOTCH_LIBDIR)
+  foreach(scotch_lib ${SCOTCH_libs_to_find})
+    set(SCOTCH_${scotch_lib}_LIBRARY "SCOTCH_${scotch_lib}_LIBRARY-NOTFOUND")
+    find_library(SCOTCH_${scotch_lib}_LIBRARY
+      NAMES ${scotch_lib}
+      HINTS ${SCOTCH_LIBDIR})
+  endforeach()
+else()
+  if(SCOTCH_DIR)
+    foreach(scotch_lib ${SCOTCH_libs_to_find})
+      set(SCOTCH_${scotch_lib}_LIBRARY "SCOTCH_${scotch_lib}_LIBRARY-NOTFOUND")
+      find_library(SCOTCH_${scotch_lib}_LIBRARY
+	NAMES ${scotch_lib}
+	HINTS ${SCOTCH_DIR}
+	PATH_SUFFIXES lib lib32 lib64)
+    endforeach()
+  else()
+    foreach(scotch_lib ${SCOTCH_libs_to_find})
+      set(SCOTCH_${scotch_lib}_LIBRARY "SCOTCH_${scotch_lib}_LIBRARY-NOTFOUND")
+      find_library(SCOTCH_${scotch_lib}_LIBRARY
+	NAMES ${scotch_lib}
+	HINTS ${_lib_env})
+    endforeach()
+  endif()
+endif()
+
+set(SCOTCH_LIBRARIES "")
+set(SCOTCH_LIBRARY_DIRS "")
+# If found, add path to cmake variable
+# ------------------------------------
+foreach(scotch_lib ${SCOTCH_libs_to_find})
+
+  if (SCOTCH_${scotch_lib}_LIBRARY)
+    get_filename_component(${scotch_lib}_lib_path "${SCOTCH_${scotch_lib}_LIBRARY}" PATH)
+    # set cmake variables
+    list(APPEND SCOTCH_LIBRARIES "${SCOTCH_${scotch_lib}_LIBRARY}")
+    list(APPEND SCOTCH_LIBRARY_DIRS "${${scotch_lib}_lib_path}")
+  else ()
+    list(APPEND SCOTCH_LIBRARIES "${SCOTCH_${scotch_lib}_LIBRARY}")
+    if (NOT SCOTCH_FIND_QUIETLY)
+      message(STATUS "Looking for scotch -- lib ${scotch_lib} not found")
+    endif()
+  endif ()
+
+  mark_as_advanced(SCOTCH_${scotch_lib}_LIBRARY)
+
+endforeach()
+list(REMOVE_DUPLICATES SCOTCH_LIBRARY_DIRS)
+
+# check a function to validate the find
+if(SCOTCH_LIBRARIES)
+
+  set(REQUIRED_INCDIRS)
+  set(REQUIRED_LIBDIRS)
+  set(REQUIRED_LIBS)
+
+  # SCOTCH
+  if (SCOTCH_INCLUDE_DIRS)
+    set(REQUIRED_INCDIRS  "${SCOTCH_INCLUDE_DIRS}")
+  endif()
+  if (SCOTCH_LIBRARY_DIRS)
+    set(REQUIRED_LIBDIRS "${SCOTCH_LIBRARY_DIRS}")
+  endif()
+  set(REQUIRED_LIBS "${SCOTCH_LIBRARIES}")
+  # THREADS
+  if(CMAKE_THREAD_LIBS_INIT)
+    list(APPEND REQUIRED_LIBS "${CMAKE_THREAD_LIBS_INIT}")
+  endif()
+  set(Z_LIBRARY "Z_LIBRARY-NOTFOUND")
+  find_library(Z_LIBRARY NAMES z)
+  mark_as_advanced(Z_LIBRARY)
+  if(Z_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lz")
+  endif()
+  set(M_LIBRARY "M_LIBRARY-NOTFOUND")
+  find_library(M_LIBRARY NAMES m)
+  mark_as_advanced(M_LIBRARY)
+  if(M_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lm")
+  endif()
+  set(RT_LIBRARY "RT_LIBRARY-NOTFOUND")
+  find_library(RT_LIBRARY NAMES rt)
+  mark_as_advanced(RT_LIBRARY)
+  if(RT_LIBRARY)
+    list(APPEND REQUIRED_LIBS "-lrt")
+  endif()
+
+  # set required libraries for link
+  set(CMAKE_REQUIRED_INCLUDES "${REQUIRED_INCDIRS}")
+  set(CMAKE_REQUIRED_LIBRARIES)
+  foreach(lib_dir ${REQUIRED_LIBDIRS})
+    list(APPEND CMAKE_REQUIRED_LIBRARIES "-L${lib_dir}")
+  endforeach()
+  list(APPEND CMAKE_REQUIRED_LIBRARIES "${REQUIRED_LIBS}")
+  string(REGEX REPLACE "^ -" "-" CMAKE_REQUIRED_LIBRARIES "${CMAKE_REQUIRED_LIBRARIES}")
+
+  # test link
+  unset(SCOTCH_WORKS CACHE)
+  include(CheckFunctionExists)
+  check_function_exists(SCOTCH_graphInit SCOTCH_WORKS)
+  mark_as_advanced(SCOTCH_WORKS)
+
+  if(SCOTCH_WORKS)
+    # save link with dependencies
+    set(SCOTCH_LIBRARIES "${REQUIRED_LIBS}")
+  else()
+    if(NOT SCOTCH_FIND_QUIETLY)
+      message(STATUS "Looking for SCOTCH : test of SCOTCH_graphInit with SCOTCH library fails")
+      message(STATUS "CMAKE_REQUIRED_LIBRARIES: ${CMAKE_REQUIRED_LIBRARIES}")
+      message(STATUS "CMAKE_REQUIRED_INCLUDES: ${CMAKE_REQUIRED_INCLUDES}")
+      message(STATUS "Check in CMakeFiles/CMakeError.log to figure out why it fails")
+    endif()
+  endif()
+  set(CMAKE_REQUIRED_INCLUDES)
+  set(CMAKE_REQUIRED_FLAGS)
+  set(CMAKE_REQUIRED_LIBRARIES)
+endif(SCOTCH_LIBRARIES)
+
+if (SCOTCH_LIBRARIES)
+  list(GET SCOTCH_LIBRARIES 0 first_lib)
+  get_filename_component(first_lib_path "${first_lib}" PATH)
+  if (${first_lib_path} MATCHES "/lib(32|64)?$")
+    string(REGEX REPLACE "/lib(32|64)?$" "" not_cached_dir "${first_lib_path}")
+    set(SCOTCH_DIR_FOUND "${not_cached_dir}" CACHE PATH "Installation directory of SCOTCH library" FORCE)
+  else()
+    set(SCOTCH_DIR_FOUND "${first_lib_path}" CACHE PATH "Installation directory of SCOTCH library" FORCE)
+  endif()
+endif()
+mark_as_advanced(SCOTCH_DIR)
+mark_as_advanced(SCOTCH_DIR_FOUND)
+
+# Check the size of SCOTCH_Num
+# ---------------------------------
+set(CMAKE_REQUIRED_INCLUDES ${SCOTCH_INCLUDE_DIRS})
+
+include(CheckCSourceRuns)
+#stdio.h and stdint.h should be included by scotch.h directly
+set(SCOTCH_C_TEST_SCOTCH_Num_4 "
+#include <stdio.h>
+#include <stdint.h>
+#include <scotch.h>
+int main(int argc, char **argv) {
+  if (sizeof(SCOTCH_Num) == 4)
+    return 0;
+  else
+    return 1;
+}
+")
+
+set(SCOTCH_C_TEST_SCOTCH_Num_8 "
+#include <stdio.h>
+#include <stdint.h>
+#include <scotch.h>
+int main(int argc, char **argv) {
+  if (sizeof(SCOTCH_Num) == 8)
+    return 0;
+  else
+    return 1;
+}
+")
+check_c_source_runs("${SCOTCH_C_TEST_SCOTCH_Num_4}" SCOTCH_Num_4)
+if(NOT SCOTCH_Num_4)
+  check_c_source_runs("${SCOTCH_C_TEST_SCOTCH_Num_8}" SCOTCH_Num_8)
+  if(NOT SCOTCH_Num_8)
+    set(SCOTCH_INTSIZE -1)
+  else()
+    set(SCOTCH_INTSIZE 8)
+  endif()
+else()
+  set(SCOTCH_INTSIZE 4)
+endif()
+set(CMAKE_REQUIRED_INCLUDES "")
+
+# check that SCOTCH has been found
+# ---------------------------------
 include(FindPackageHandleStandardArgs)
 find_package_handle_standard_args(SCOTCH DEFAULT_MSG
-                                  SCOTCH_INCLUDES SCOTCH_LIBRARIES)
-
-mark_as_advanced(SCOTCH_INCLUDES SCOTCH_LIBRARIES)
+  SCOTCH_LIBRARIES
+  SCOTCH_WORKS)
+#
+# TODO: Add possibility to check for specific functions in the library
+#
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index d337594f5..8da51ce57 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -27,7 +27,7 @@ endif()
 
 if(NOT EIGEN_Fortran_COMPILER_WORKS)
   # search for a default Lapack library to complete Eigen's one
-  find_package(LAPACK)
+  find_package(LAPACK QUIET)
 endif()
 
 # configure blas/lapack (use Eigen's ones)
@@ -80,23 +80,30 @@ else()
 endif()
 
 
-find_package(Pastix)
-find_package(Scotch)
-find_package(Metis 5.0 REQUIRED)
-if(PASTIX_FOUND)
+find_package(PASTIX QUIET COMPONENTS METIS SCOTCH)
+# check that the PASTIX found is a version without MPI
+find_path(PASTIX_pastix_nompi.h_INCLUDE_DIRS
+  NAMES pastix_nompi.h
+  HINTS ${PASTIX_INCLUDE_DIRS}
+)
+if (NOT PASTIX_pastix_nompi.h_INCLUDE_DIRS)
+  message(STATUS "A version of Pastix has been found but pastix_nompi.h does not exist in the include directory."
+                 " Because Eigen tests require a version without MPI, we disable the Pastix backend.")
+endif()
+if(PASTIX_FOUND AND PASTIX_pastix_nompi.h_INCLUDE_DIRS)
   add_definitions("-DEIGEN_PASTIX_SUPPORT")
-  include_directories(${PASTIX_INCLUDES})
+  include_directories(${PASTIX_INCLUDE_DIRS_DEP})
   if(SCOTCH_FOUND)
-    include_directories(${SCOTCH_INCLUDES})
+    include_directories(${SCOTCH_INCLUDE_DIRS})
     set(PASTIX_LIBRARIES ${PASTIX_LIBRARIES} ${SCOTCH_LIBRARIES})
   elseif(METIS_FOUND)
-    include_directories(${METIS_INCLUDES})
+    include_directories(${METIS_INCLUDE_DIRS})
     set(PASTIX_LIBRARIES ${PASTIX_LIBRARIES} ${METIS_LIBRARIES})
   else(SCOTCH_FOUND)
     ei_add_property(EIGEN_MISSING_BACKENDS  "PaStiX, ")
   endif(SCOTCH_FOUND)
-  set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES} ${ORDERING_LIBRARIES} ${EIGEN_BLAS_LIBRARIES})
-  set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES} ${EIGEN_BLAS_LIBRARIES})
+  set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES_DEP} ${ORDERING_LIBRARIES})
+  set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES_DEP})
   ei_add_property(EIGEN_TESTED_BACKENDS  "PaStiX, ")
 else()
   ei_add_property(EIGEN_MISSING_BACKENDS  "PaStiX, ")
@@ -104,7 +111,7 @@ endif()
 
 if(METIS_FOUND)
   add_definitions("-DEIGEN_METIS_SUPPORT")
-  include_directories(${METIS_INCLUDES})
+  include_directories(${METIS_INCLUDE_DIRS})
   ei_add_property(EIGEN_TESTED_BACKENDS "METIS, ")
 else()
   ei_add_property(EIGEN_MISSING_BACKENDS "METIS, ")

From f75dfdda7e6f133c4d45671efb6799dbb975b1ef Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 14 Apr 2017 10:34:30 +0200
Subject: [PATCH 094/680] Fix unwanted Real to Scalar to Real conversions in
 column-pivoting QR.

---
 Eigen/src/QR/ColPivHouseholderQR.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index d35395d04..1b828f5dc 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -505,8 +505,8 @@ void ColPivHouseholderQR<MatrixType>::computeInPlace()
     m_colNormsUpdated.coeffRef(k) = m_colNormsDirect.coeffRef(k);
   }
 
-  RealScalar threshold_helper =  numext::abs2<Scalar>(m_colNormsUpdated.maxCoeff() * NumTraits<Scalar>::epsilon()) / RealScalar(rows);
-  RealScalar norm_downdate_threshold = numext::sqrt(NumTraits<Scalar>::epsilon());
+  RealScalar threshold_helper =  numext::abs2<RealScalar>(m_colNormsUpdated.maxCoeff() * NumTraits<RealScalar>::epsilon()) / RealScalar(rows);
+  RealScalar norm_downdate_threshold = numext::sqrt(NumTraits<RealScalar>::epsilon());
 
   m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
   m_maxpivot = RealScalar(0);
@@ -556,8 +556,8 @@ void ColPivHouseholderQR<MatrixType>::computeInPlace()
         RealScalar temp = abs(m_qr.coeffRef(k, j)) / m_colNormsUpdated.coeffRef(j);
         temp = (RealScalar(1) + temp) * (RealScalar(1) - temp);
         temp = temp < 0 ? 0 : temp;
-        RealScalar temp2 = temp * numext::abs2<Scalar>(m_colNormsUpdated.coeffRef(j) /
-                                                       m_colNormsDirect.coeffRef(j));
+        RealScalar temp2 = temp * numext::abs2<RealScalar>(m_colNormsUpdated.coeffRef(j) /
+                                                           m_colNormsDirect.coeffRef(j));
         if (temp2 <= norm_downdate_threshold) {
           // The updated norm has become too inaccurate so re-compute the column
           // norm directly.

From d9084ac8e142697f0d767092a17ffc3a7a18a2e4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 14 Apr 2017 11:05:13 +0200
Subject: [PATCH 095/680] Improve mixing of complex and real in the vectorized
 path of apply_rotation_in_the_plane

---
 Eigen/src/Jacobi/Jacobi.h | 42 +++++++++++++++++++++++----------------
 1 file changed, 25 insertions(+), 17 deletions(-)

diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index d25af8e90..c30326e1d 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -302,8 +302,12 @@ template<typename VectorX, typename VectorY, typename OtherScalar>
 void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j)
 {
   typedef typename VectorX::Scalar Scalar;
-  enum { PacketSize = packet_traits<Scalar>::size };
+  enum {
+    PacketSize = packet_traits<Scalar>::size,
+    OtherPacketSize = packet_traits<OtherScalar>::size
+  };
   typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<OtherScalar>::type OtherPacket;
   eigen_assert(xpr_x.size() == xpr_y.size());
   Index size = xpr_x.size();
   Index incrx = xpr_x.derived().innerStride();
@@ -321,6 +325,7 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
 
   if(VectorX::SizeAtCompileTime == Dynamic &&
     (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
+    (PacketSize == OtherPacketSize) &&
     ((incrx==1 && incry==1) || PacketSize == 1))
   {
     // both vectors are sequentially stored in memory => vectorization
@@ -329,9 +334,10 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
     Index alignedStart = internal::first_default_aligned(y, size);
     Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
 
-    const Packet pc = pset1<Packet>(c);
-    const Packet ps = pset1<Packet>(s);
-    conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex,false> pcj;
+    const OtherPacket pc = pset1<OtherPacket>(c);
+    const OtherPacket ps = pset1<OtherPacket>(s);
+    conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
+    conj_helper<OtherPacket,Packet,false,false> pm;
 
     for(Index i=0; i<alignedStart; ++i)
     {
@@ -350,8 +356,8 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
       {
         Packet xi = pload<Packet>(px);
         Packet yi = pload<Packet>(py);
-        pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
+        pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
         px += PacketSize;
         py += PacketSize;
       }
@@ -365,10 +371,10 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
         Packet xi1  = ploadu<Packet>(px+PacketSize);
         Packet yi   = pload <Packet>(py);
         Packet yi1  = pload <Packet>(py+PacketSize);
-        pstoreu(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstoreu(px+PacketSize, padd(pmul(pc,xi1),pcj.pmul(ps,yi1)));
-        pstore (py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
-        pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pmul(ps,xi1)));
+        pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
+        pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+        pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
         px += Peeling*PacketSize;
         py += Peeling*PacketSize;
       }
@@ -376,8 +382,8 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
       {
         Packet xi = ploadu<Packet>(x+peelingEnd);
         Packet yi = pload <Packet>(y+peelingEnd);
-        pstoreu(x+peelingEnd, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
+        pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
       }
     }
 
@@ -393,19 +399,21 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
   /*** fixed-size vectorized path ***/
   else if(VectorX::SizeAtCompileTime != Dynamic &&
           (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
+          (PacketSize == OtherPacketSize) &&
           (EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment)>0)) // FIXME should be compared to the required alignment
   {
-    const Packet pc = pset1<Packet>(c);
-    const Packet ps = pset1<Packet>(s);
-    conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex,false> pcj;
+    const OtherPacket pc = pset1<OtherPacket>(c);
+    const OtherPacket ps = pset1<OtherPacket>(s);
+    conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
+    conj_helper<OtherPacket,Packet,false,false> pm;
     Scalar* EIGEN_RESTRICT px = x;
     Scalar* EIGEN_RESTRICT py = y;
     for(Index i=0; i<size; i+=PacketSize)
     {
       Packet xi = pload<Packet>(px);
       Packet yi = pload<Packet>(py);
-      pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
-      pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
+      pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+      pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
       px += PacketSize;
       py += PacketSize;
     }

From 949a2da38cbfebe358a25dc59b47abb67beb4126 Mon Sep 17 00:00:00 2001
From: RJ Ryan <rjryan@google.com>
Date: Fri, 14 Apr 2017 13:23:35 -0700
Subject: [PATCH 096/680] Use scalar_sum_op and scalar_quotient_op instead of
 operator+ and operator/ in MeanReducer.

Improves support for std::complex types when compiling for CUDA.

Expands on e2e9cdd16970914cf0a892fea5e7c4402b3ede41
 and 2bda1b0d93fb627d0c500ec48b20302d44c32cb7
.
---
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   |  8 +++--
 unsupported/test/cxx11_tensor_complex_cuda.cu | 36 +++++++++++++++++++
 2 files changed, 42 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 3b4f8eda1..5dcc3794c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -166,7 +166,8 @@ template <typename T> struct MeanReducer
     return pset1<Packet>(initialize());
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
-    return accum / scalarCount_;
+    internal::scalar_quotient_op<T> quotient_op;
+    return quotient_op(accum, T(scalarCount_));
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
@@ -175,7 +176,10 @@ template <typename T> struct MeanReducer
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
     internal::scalar_sum_op<T> sum_op;
-    return sum_op(saccum, predux(vaccum)) / (scalarCount_ + packetCount_ * unpacket_traits<Packet>::size);
+    internal::scalar_quotient_op<T> quotient_op;
+    return quotient_op(
+        sum_op(saccum, predux(vaccum)),
+        T(scalarCount_ + packetCount_ * unpacket_traits<Packet>::size));
   }
 
   protected:
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index d4e111f5d..87cf28920 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -107,6 +107,41 @@ static void test_cuda_sum_reductions() {
   gpu_device.deallocate(gpu_out_ptr);
 }
 
+static void test_cuda_mean_reductions() {
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  const int num_rows = internal::random<int>(1024, 5*1024);
+  const int num_cols = internal::random<int>(1024, 5*1024);
+
+  Tensor<std::complex<float>, 2> in(num_rows, num_cols);
+  in.setRandom();
+
+  Tensor<std::complex<float>, 0> full_redux;
+  full_redux = in.mean();
+
+  std::size_t in_bytes = in.size() * sizeof(std::complex<float>);
+  std::size_t out_bytes = full_redux.size() * sizeof(std::complex<float>);
+  std::complex<float>* gpu_in_ptr = static_cast<std::complex<float>*>(gpu_device.allocate(in_bytes));
+  std::complex<float>* gpu_out_ptr = static_cast<std::complex<float>*>(gpu_device.allocate(out_bytes));
+  gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
+
+  TensorMap<Tensor<std::complex<float>, 2> > in_gpu(gpu_in_ptr, num_rows, num_cols);
+  TensorMap<Tensor<std::complex<float>, 0> > out_gpu(gpu_out_ptr);
+
+  out_gpu.device(gpu_device) = in_gpu.mean();
+
+  Tensor<std::complex<float>, 0> full_redux_gpu;
+  gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
+  gpu_device.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  VERIFY_IS_APPROX(full_redux(), full_redux_gpu());
+
+  gpu_device.deallocate(gpu_in_ptr);
+  gpu_device.deallocate(gpu_out_ptr);
+}
 
 static void test_cuda_product_reductions() {
 
@@ -149,5 +184,6 @@ void test_cxx11_tensor_complex()
 {
   CALL_SUBTEST(test_cuda_nullary());
   CALL_SUBTEST(test_cuda_sum_reductions());
+  CALL_SUBTEST(test_cuda_mean_reductions());
   CALL_SUBTEST(test_cuda_product_reductions());
 }

From 891ac03483dd282569581076d4e0819608fc1155 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 25 Apr 2017 13:58:10 +0200
Subject: [PATCH 097/680] Fix dense * sparse-selfadjoint-view product.

---
 Eigen/src/SparseCore/SparseSelfAdjointView.h | 3 ++-
 test/sparse_product.cpp                      | 4 ++++
 2 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h
index 9e39be738..5ab64f1a8 100644
--- a/Eigen/src/SparseCore/SparseSelfAdjointView.h
+++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h
@@ -47,6 +47,7 @@ template<typename MatrixType, unsigned int _Mode> class SparseSelfAdjointView
     
     enum {
       Mode = _Mode,
+      TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0),
       RowsAtCompileTime = internal::traits<SparseSelfAdjointView>::RowsAtCompileTime,
       ColsAtCompileTime = internal::traits<SparseSelfAdjointView>::ColsAtCompileTime
     };
@@ -368,7 +369,7 @@ struct generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, Pr
     
     // transpose everything
     Transpose<Dest> dstT(dst);
-    internal::sparse_selfadjoint_time_dense_product<RhsView::Mode>(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha);
+    internal::sparse_selfadjoint_time_dense_product<RhsView::TransposeMode>(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha);
   }
 };
 
diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp
index c1edd26e3..197586741 100644
--- a/test/sparse_product.cpp
+++ b/test/sparse_product.cpp
@@ -297,6 +297,10 @@ template<typename SparseMatrixType> void sparse_product()
     VERIFY_IS_APPROX(x=mLo.template selfadjointView<Lower>()*b, refX=refS*b);
     VERIFY_IS_APPROX(x=mS.template selfadjointView<Upper|Lower>()*b, refX=refS*b);
 
+    VERIFY_IS_APPROX(x=b * mUp.template selfadjointView<Upper>(),       refX=b*refS);
+    VERIFY_IS_APPROX(x=b * mLo.template selfadjointView<Lower>(),       refX=b*refS);
+    VERIFY_IS_APPROX(x=b * mS.template selfadjointView<Upper|Lower>(),  refX=b*refS);
+
     VERIFY_IS_APPROX(x.noalias()+=mUp.template selfadjointView<Upper>()*b, refX+=refS*b);
     VERIFY_IS_APPROX(x.noalias()-=mLo.template selfadjointView<Lower>()*b, refX-=refS*b);
     VERIFY_IS_APPROX(x.noalias()+=mS.template selfadjointView<Upper|Lower>()*b, refX+=refS*b);

From 9bc0a35731d914a54a6fa7215d2ba0aeb53194a9 Mon Sep 17 00:00:00 2001
From: Abhijit Kundu <abhijit.kundu@gatech.edu>
Date: Thu, 27 Apr 2017 03:09:03 -0400
Subject: [PATCH 098/680] Fixed nested angle barckets >> issue when compiling
 with cuda 8

---
 Eigen/src/Core/arch/CUDA/Complex.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Complex.h b/Eigen/src/Core/arch/CUDA/Complex.h
index 9c2536509..ca0aaed32 100644
--- a/Eigen/src/Core/arch/CUDA/Complex.h
+++ b/Eigen/src/Core/arch/CUDA/Complex.h
@@ -55,7 +55,7 @@ template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T
 // Product
 template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
   enum {
-    Vectorizable = packet_traits<std::complex<T>>::HasMul
+    Vectorizable = packet_traits<std::complex<T> >::HasMul
   };
   typedef typename std::complex<T> result_type;
 
@@ -76,7 +76,7 @@ template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> >
 // Quotient
 template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
   enum {
-    Vectorizable = packet_traits<std::complex<T>>::HasDiv
+    Vectorizable = packet_traits<std::complex<T> >::HasDiv
   };
   typedef typename std::complex<T> result_type;
 

From 4343db84d8df4fbca74ff4321d0b84c1169f0628 Mon Sep 17 00:00:00 2001
From: Abhijit Kundu <abhijit.kundu@gatech.edu>
Date: Mon, 1 May 2017 10:36:27 -0400
Subject: [PATCH 099/680] updated warning number for nvcc relase 8 (V8.0.61)
 for the stupid warning message  'calling a __host__ function from a __host__
 __device__ function is not allowed'.

---
 Eigen/src/Core/util/DisableStupidWarnings.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index 4431f2fc4..b91d1d1af 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -72,6 +72,7 @@
   #pragma diag_suppress 2671
   #pragma diag_suppress 2735
   #pragma diag_suppress 2737
+  #pragma diag_suppress 2739
 #endif
 
 #endif // not EIGEN_WARNINGS_DISABLED

From 052426b824038bbee1c1c48c3df65dfccd79ae24 Mon Sep 17 00:00:00 2001
From: a-doumoulakis <anonymous@invalid.net>
Date: Fri, 5 May 2017 19:26:27 +0100
Subject: [PATCH 100/680] Add support for triSYCL

Eigen is now able to use triSYCL with EIGEN_SYCL_TRISYCL and TRISYCL_INCLUDE_DIR options

Fix contraction kernel with correct nd_item dimension
---
 CMakeLists.txt                                | 15 ++++-
 cmake/EigenTesting.cmake                      | 37 ++++++++----
 unsupported/Eigen/CXX11/Tensor                |  2 +-
 .../CXX11/src/Tensor/TensorContractionSycl.h  |  2 +-
 unsupported/test/CMakeLists.txt               | 59 +++++++++++--------
 5 files changed, 72 insertions(+), 43 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index fe4227cbb..54f3a7509 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -151,6 +151,10 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-Wenum-conversion")
   ei_add_cxx_compiler_flag("-Wc++11-extensions")
   ei_add_cxx_compiler_flag("-Wdouble-promotion")
+  ei_add_cxx_compiler_flag("-Wno-unused-parameter")
+  ei_add_cxx_compiler_flag("-Wno-ignored-attributes")
+  ei_add_cxx_compiler_flag("-Wno-ignored-qualifiers")
+  ei_add_cxx_compiler_flag("-Wno-sign-compare")
 #  ei_add_cxx_compiler_flag("-Wconversion")
   
   # -Wshadow is insanely too strict with gcc, hopefully it will become usable with gcc 6
@@ -437,10 +441,17 @@ endif()
 
 # add SYCL
 option(EIGEN_TEST_SYCL "Add Sycl support." OFF)
+option(EIGEN_SYCL_TRISYCL "Use the triSYCL Sycl implementation (ComputeCPP by default)." OFF)
 if(EIGEN_TEST_SYCL)
   set (CMAKE_MODULE_PATH "${CMAKE_ROOT}/Modules" "cmake/Modules/" "${CMAKE_MODULE_PATH}")
-  include(FindComputeCpp)
-endif()
+  if(EIGEN_SYCL_TRISYCL)
+    message(STATUS "Using triSYCL")
+    include(FindTriSYCL)
+  else(EIGEN_SYCL_TRISYCL)
+    message(STATUS "Using ComputeCPP SYCL")
+    include(FindComputeCpp)
+  endif(EIGEN_SYCL_TRISYCL)
+endif(EIGEN_TEST_SYCL)
 
 add_subdirectory(unsupported)
 
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index a92a2978b..848233342 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -111,7 +111,7 @@ endmacro(ei_add_test_internal)
 
 # SYCL
 macro(ei_add_test_internal_sycl testname testname_with_suffix)
-  include_directories( SYSTEM ${COMPUTECPP_PACKAGE_ROOT_DIR}/include)
+
   set(targetname ${testname_with_suffix})
 
   if(EIGEN_ADD_TEST_FILENAME_EXTENSION)
@@ -124,18 +124,25 @@ macro(ei_add_test_internal_sycl testname testname_with_suffix)
   set( bc_file ${CMAKE_CURRENT_BINARY_DIR}/${filename})
   set( host_file ${CMAKE_CURRENT_SOURCE_DIR}/${filename})
 
-  ADD_CUSTOM_COMMAND(
-    OUTPUT ${include_file}
-    COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${host_file}\\\"" > ${include_file}
-    COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${bc_file}.sycl\\\"" >> ${include_file}
-    DEPENDS ${filename} ${bc_file}.sycl
-    COMMENT "Building ComputeCpp integration header file ${include_file}"
-  )
-  # Add a custom target for the generated integration header
-  add_custom_target(${testname}_integration_header_sycl DEPENDS ${include_file})
+  if(NOT EIGEN_SYCL_TRISYCL)
+    include_directories( SYSTEM ${COMPUTECPP_PACKAGE_ROOT_DIR}/include)
+
+    ADD_CUSTOM_COMMAND(
+      OUTPUT ${include_file}
+      COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${host_file}\\\"" > ${include_file}
+      COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${bc_file}.sycl\\\"" >> ${include_file}
+      DEPENDS ${filename} ${bc_file}.sycl
+      COMMENT "Building ComputeCpp integration header file ${include_file}"
+      )
+
+    # Add a custom target for the generated integration header
+    add_custom_target(${testname}_integration_header_sycl DEPENDS ${include_file})
+    add_executable(${targetname} ${include_file})
+    add_dependencies(${targetname} ${testname}_integration_header_sycl)
+  else()
+    add_executable(${targetname} ${host_file})
+  endif()
 
-  add_executable(${targetname} ${include_file})
-  add_dependencies(${targetname} ${testname}_integration_header_sycl)
   add_sycl_to_target(${targetname} ${filename} ${CMAKE_CURRENT_BINARY_DIR})
 
   if (targetname MATCHES "^eigen2_")
@@ -467,7 +474,11 @@ macro(ei_testing_print_summary)
     endif()
 
     if(EIGEN_TEST_SYCL)
-      message(STATUS "SYCL:              ON")
+      if(EIGEN_SYCL_TRISYCL)
+        message(STATUS "SYCL:              ON (using triSYCL)")
+      else()
+        message(STATUS "SYCL:              ON (using computeCPP)")
+      endif()
     else()
       message(STATUS "SYCL:              OFF")
     endif()
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 39916092b..5d71a9c25 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -19,7 +19,7 @@
 #undef isnan
 #undef isinf
 #undef isfinite
-#include <SYCL/sycl.hpp>
+#include <CL/sycl.hpp>
 #include <iostream>
 #include <map>
 #include <memory>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
index 5b4c3c5bd..e6840bc87 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
@@ -195,7 +195,7 @@ typename HostExpr::Index LocalThreadSizeM, typename HostExpr::Index LocalThreadS
     m_j_strides(m_j_strides_),  m_right_nocontract_strides(m_right_nocontract_strides_),
     left_tuple_of_accessors(left_tuple_of_accessors_), right_tuple_of_accessors(right_tuple_of_accessors_), dev(dev_){}
 
-    void operator()(cl::sycl::nd_item<1> itemID) {
+    void operator()(cl::sycl::nd_item<2> itemID) {
       typedef typename Eigen::TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
       typedef typename Eigen::TensorSycl::internal::ConvertToDeviceExpression<LHSHostExpr>::Type LHSDevExpr;
       typedef typename Eigen::TensorSycl::internal::ConvertToDeviceExpression<RHSHostExpr>::Type RHSDevExpr;
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index cdf151f15..9a9124640 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -152,33 +152,40 @@ endif()
 
 if(EIGEN_TEST_CXX11)
   if(EIGEN_TEST_SYCL)
-    ei_add_test_sycl(cxx11_tensor_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_forced_eval_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_broadcast_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_device_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_reduction_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_morphing_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_shuffling_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_padding_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_builtins_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_contract_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_concatenation_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_reverse_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_convolution_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_striding_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_layout_swap_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_inflation_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_generator_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_image_patch_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_volume_patcP_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_argmax_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_custom_op_sycl "-std=c++11")
+    if(EIGEN_SYCL_TRISYCL)
+      set(CMAKE_CXX_STANDARD 14)
+      set(STD_CXX_FLAG "-std=c++1z")
+    else(EIGEN_SYCL_TRISYCL)
+      # It should be safe to always run these tests as there is some fallback code for
+      # older compiler that don't support cxx11.
+      set(CMAKE_CXX_STANDARD 11)
+      set(STD_CXX_FLAG "-std=c++11")
+    endif(EIGEN_SYCL_TRISYCL)
+
+    ei_add_test_sycl(cxx11_tensor_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_forced_eval_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_broadcast_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_device_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_reduction_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_morphing_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_shuffling_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_padding_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_builtins_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_contract_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_concatenation_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_reverse_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_convolution_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_striding_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_chipping_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_layout_swap_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_inflation_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_generator_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_patch_sycl ${STD_CXX_FLAG})
+    #ei_add_test_sycl(cxx11_tensor_image_patch_sycl ${STD_CXX_FLAG})
+    #ei_add_test_sycl(cxx11_tensor_volume_patch_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_argmax_sycl ${STD_CXX_FLAG})
+    ei_add_test_sycl(cxx11_tensor_custom_op_sycl ${STD_CXX_FLAG})
   endif(EIGEN_TEST_SYCL)
-  # It should be safe to always run these tests as there is some fallback code for
-  # older compiler that don't support cxx11.
-  set(CMAKE_CXX_STANDARD 11)
 
   ei_add_test(cxx11_eventcount "-pthread" "${CMAKE_THREAD_LIBS_INIT}")
   ei_add_test(cxx11_runqueue "-pthread" "${CMAKE_THREAD_LIBS_INIT}")

From a5226ce4f77f68a7f8590174ed3cdebb89cbe127 Mon Sep 17 00:00:00 2001
From: a-doumoulakis <anonymous@invalid.net>
Date: Wed, 17 May 2017 17:59:30 +0100
Subject: [PATCH 101/680] Add cmake file FindTriSYCL.cmake

---
 cmake/FindTriSYCL.cmake | 152 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 152 insertions(+)
 create mode 100644 cmake/FindTriSYCL.cmake

diff --git a/cmake/FindTriSYCL.cmake b/cmake/FindTriSYCL.cmake
new file mode 100644
index 000000000..cb2154192
--- /dev/null
+++ b/cmake/FindTriSYCL.cmake
@@ -0,0 +1,152 @@
+#.rst:
+# FindTriSYCL
+#---------------
+#
+# TODO : insert Copyright and licence
+
+#########################
+#  FindTriSYCL.cmake
+#########################
+#
+#  Tools for finding and building with TriSYCL.
+#
+#  User must define TRISYCL_INCLUDE_DIR pointing to the triSYCL
+#  include directory.
+#
+#  Latest version of this file can be found at:
+#    https://github.com/triSYCL/triSYCL
+
+# Requite CMake version 3.5 or higher
+cmake_minimum_required (VERSION 3.5)
+
+# Check that a supported host compiler can be found
+if(CMAKE_COMPILER_IS_GNUCXX)
+  # Require at least gcc 5.4
+  if (CMAKE_CXX_COMPILER_VERSION VERSION_LESS 5.4)
+    message(FATAL_ERROR
+      "host compiler - Not found! (gcc version must be at least 5.4)")
+  else()
+    message(STATUS "host compiler - gcc ${CMAKE_CXX_COMPILER_VERSION}")
+  endif()
+elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
+  # Require at least clang 3.9
+  if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.9)
+    message(FATAL_ERROR
+      "host compiler - Not found! (clang version must be at least 3.9)")
+  else()
+    message(STATUS "host compiler - clang ${CMAKE_CXX_COMPILER_VERSION}")
+  endif()
+else()
+  message(WARNING
+    "host compiler - Not found! (triSYCL supports GCC and Clang)")
+endif()
+
+#triSYCL options
+option(TRISYCL_OPENMP "triSYCL multi-threading with OpenMP" ON)
+option(TRISYCL_OPENCL "triSYCL OpenCL interoperability mode" OFF)
+option(TRISYCL_NO_ASYNC "triSYCL use synchronous kernel execution" OFF)
+option(TRISYCL_DEBUG "triSCYL use debug mode" OFF)
+option(TRISYCL_DEBUG_STRUCTORS "triSYCL trace of object lifetimes" OFF)
+option(TRISYCL_TRACE_KERNEL "triSYCL trace of kernel execution" OFF)
+
+mark_as_advanced(TRISYCL_OPENMP)
+mark_as_advanced(TRISYCL_OPENCL)
+mark_as_advanced(TRISYCL_NO_ASYNC)
+mark_as_advanced(TRISYCL_DEBUG)
+mark_as_advanced(TRISYCL_DEBUG_STRUCTORS)
+mark_as_advanced(TRISYCL_TRACE_KERNEL)
+
+#triSYCL definitions
+set(CL_SYCL_LANGUAGE_VERSION 220 CACHE VERSION
+  "Host language version to be used by trisYCL (default is: 220)")
+set(TRISYCL_CL_LANGUAGE_VERSION 220 CACHE VERSION
+  "Device language version to be used by trisYCL (default is: 220)")
+#set(TRISYCL_COMPILE_OPTIONS "-std=c++1z -Wall -Wextra")
+set(CMAKE_CXX_STANDARD 14)
+set(CXX_STANDARD_REQUIRED ON)
+
+
+# Find OpenCL package
+if(TRISYCL_OPENCL)
+  find_package(OpenCL REQUIRED)
+  if(UNIX)
+    set(BOOST_COMPUTE_INCPATH /usr/include/compute CACHE PATH
+      "Path to Boost.Compute headers (default is: /usr/include/compute)")
+  endif(UNIX)
+endif()
+
+# Find OpenMP package
+if(TRISYCL_OPENMP)
+  find_package(OpenMP REQUIRED)
+endif()
+
+# Find Boost
+find_package(Boost 1.58 REQUIRED COMPONENTS chrono log)
+
+# If debug or trace we need boost log
+if(TRISYCL_DEBUG OR TRISYCL_DEBUG_STRUCTORS OR TRISYCL_TRACE_KERNEL)
+  set(LOG_NEEDED ON)
+else()
+  set(LOG_NEEDED OFF)
+endif()
+
+find_package(Threads REQUIRED)
+
+# Find triSYCL directory
+if(NOT TRISYCL_INCLUDE_DIR)
+  message(FATAL_ERROR
+    "triSYCL include directory - Not found! (please set TRISYCL_INCLUDE_DIR")
+else()
+  message(STATUS "triSYCL include directory - Found ${TRISYCL_INCLUDE_DIR}")
+endif()
+
+#######################
+#  add_sycl_to_target
+#######################
+#
+#  Sets the proper flags and includes for the target compilation.
+#
+#  targetName : Name of the target to add a SYCL to.
+#  sourceFile : Source file to be compiled for SYCL.
+#  binaryDir : Intermediate directory to output the integration header.
+#
+function(add_sycl_to_target targetName sourceFile binaryDir)
+
+  # Add include directories to the "#include <>" paths
+  target_include_directories (${targetName} PUBLIC
+    ${TRISYCL_INCLUDE_DIR}
+    ${Boost_INCLUDE_DIRS}
+    $<$<BOOL:${TRISYCL_OPENCL}>:${OpenCL_INCLUDE_DIRS}>
+    $<$<BOOL:${TRISYCL_OPENCL}>:${BOOST_COMPUTE_INCPATH}>)
+
+
+  # Link dependencies
+  target_link_libraries(${targetName} PUBLIC
+    $<$<BOOL:${TRISYCL_OPENCL}>:${OpenCL_LIBRARIES}>
+    Threads::Threads
+    $<$<BOOL:${LOG_NEEDED}>:Boost::log>
+    Boost::chrono)
+
+
+  # Compile definitions
+  target_compile_definitions(${targetName} PUBLIC
+    $<$<BOOL:${TRISYCL_NO_ASYNC}>:TRISYCL_NO_ASYNC>
+    $<$<BOOL:${TRISYCL_OPENCL}>:TRISYCL_OPENCL>
+    $<$<BOOL:${TRISYCL_DEBUG}>:TRISYCL_DEBUG>
+    $<$<BOOL:${TRISYCL_DEBUG_STRUCTORS}>:TRISYCL_DEBUG_STRUCTORS>
+    $<$<BOOL:${TRISYCL_TRACE_KERNEL}>:TRISYCL_TRACE_KERNEL>
+    $<$<BOOL:${LOG_NEEDED}>:BOOST_LOG_DYN_LINK>)
+
+  # C++ and OpenMP requirements
+  target_compile_options(${targetName} PUBLIC
+    ${TRISYCL_COMPILE_OPTIONS}
+    $<$<BOOL:${TRISYCL_OPENMP}>:${OpenMP_CXX_FLAGS}>)
+
+  if(${TRISYCL_OPENMP} AND (NOT WIN32))
+    # Does not support generator expressions
+    set_target_properties(${targetName}
+      PROPERTIES
+      LINK_FLAGS ${OpenMP_CXX_FLAGS})
+  endif(${TRISYCL_OPENMP} AND (NOT WIN32))
+
+endfunction(add_sycl_to_target)

From 61d7f3664a06c64e39a0bbccb610592bcaee92cd Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 22 May 2017 14:58:28 +0100
Subject: [PATCH 102/680] Fixing Cmake Dependency for SYCL

---
 cmake/EigenTesting.cmake        | 18 +++++++++---------
 cmake/FindComputeCpp.cmake      |  8 +++++---
 unsupported/test/CMakeLists.txt |  2 +-
 3 files changed, 15 insertions(+), 13 deletions(-)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index a92a2978b..1fcdcc7bb 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -120,23 +120,23 @@ macro(ei_add_test_internal_sycl testname testname_with_suffix)
     set(filename ${testname}.cpp)
   endif()
 
-  set( include_file ${CMAKE_CURRENT_BINARY_DIR}/inc_${filename})
-  set( bc_file ${CMAKE_CURRENT_BINARY_DIR}/${filename})
-  set( host_file ${CMAKE_CURRENT_SOURCE_DIR}/${filename})
+  set( include_file "${CMAKE_CURRENT_BINARY_DIR}/inc_${filename}")
+  set( bc_file "${CMAKE_CURRENT_BINARY_DIR}/${filename}.sycl")
+  set( host_file "${CMAKE_CURRENT_SOURCE_DIR}/${filename}")
 
   ADD_CUSTOM_COMMAND(
     OUTPUT ${include_file}
     COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${host_file}\\\"" > ${include_file}
-    COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${bc_file}.sycl\\\"" >> ${include_file}
-    DEPENDS ${filename} ${bc_file}.sycl
+    COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${bc_file}\\\"" >> ${include_file}
+    DEPENDS ${host_file} ${bc_file}
     COMMENT "Building ComputeCpp integration header file ${include_file}"
   )
   # Add a custom target for the generated integration header
-  add_custom_target(${testname}_integration_header_sycl DEPENDS ${include_file})
+  add_custom_target("${testname}_integration_header_sycl" DEPENDS ${include_file})
 
   add_executable(${targetname} ${include_file})
-  add_dependencies(${targetname} ${testname}_integration_header_sycl)
-  add_sycl_to_target(${targetname} ${filename} ${CMAKE_CURRENT_BINARY_DIR})
+  add_dependencies(${targetname} "${testname}_integration_header_sycl")
+  add_sycl_to_target(${targetname} ${CMAKE_CURRENT_BINARY_DIR} ${filename})
 
   if (targetname MATCHES "^eigen2_")
     add_dependencies(eigen2_buildtests ${targetname})
@@ -720,4 +720,4 @@ macro(ei_test_get_compilerver_from_cxx_version_string)
   ei_test1_get_compilerver_from_cxx_version_string("i686-apple-darwin11-llvm-g++-4.2 (GCC) 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2335.15.00)" "llvm-g++" "4.2.1")
   ei_test1_get_compilerver_from_cxx_version_string("g++-mp-4.4 (GCC) 4.4.6" "g++" "4.4.6")
   ei_test1_get_compilerver_from_cxx_version_string("g++-mp-4.4 (GCC) 2011" "g++" "4.4")
-endmacro(ei_test_get_compilerver_from_cxx_version_string)
+endmacro(ei_test_get_compilerver_from_cxx_version_string)
\ No newline at end of file
diff --git a/cmake/FindComputeCpp.cmake b/cmake/FindComputeCpp.cmake
index 27e5c9b1f..90222ed0e 100644
--- a/cmake/FindComputeCpp.cmake
+++ b/cmake/FindComputeCpp.cmake
@@ -201,7 +201,8 @@ function(__build_spir targetName sourceFile binaryDir)
             ${device_compiler_includes}
             -o ${outputSyclFile}
             -c ${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile}
-    DEPENDS ${sourceFile}
+    DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile}
+    IMPLICIT_DEPENDS CXX "${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile}"
     WORKING_DIRECTORY ${binaryDir}
   COMMENT "Building ComputeCpp integration header file ${outputSyclFile}")
 
@@ -233,8 +234,9 @@ endfunction()
 #  sourceFile : Source file to be compiled for SYCL.
 #  binaryDir : Intermediate directory to output the integration header.
 #
-function(add_sycl_to_target targetName sourceFile binaryDir)
+function(add_sycl_to_target targetName binaryDir sourceFile)
 
+  set(sourceFiles ${sourceFiles} ${ARGN})
   # Add custom target to run compute++ and generate the integration header
   __build_spir(${targetName} ${sourceFile} ${binaryDir})
 
@@ -242,4 +244,4 @@ function(add_sycl_to_target targetName sourceFile binaryDir)
   target_link_libraries(${targetName} PUBLIC ${COMPUTECPP_RUNTIME_LIBRARY}
                         PUBLIC ${OpenCL_LIBRARIES})
 
-endfunction(add_sycl_to_target)
+endfunction(add_sycl_to_target)
\ No newline at end of file
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index cdf151f15..196a432ff 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -172,7 +172,7 @@ if(EIGEN_TEST_CXX11)
     ei_add_test_sycl(cxx11_tensor_generator_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_image_patch_sycl "-std=c++11")
-    ei_add_test_sycl(cxx11_tensor_volume_patcP_sycl "-std=c++11")
+    ei_add_test_sycl(cxx11_tensor_volume_patch_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_argmax_sycl "-std=c++11")
     ei_add_test_sycl(cxx11_tensor_custom_op_sycl "-std=c++11")
   endif(EIGEN_TEST_SYCL)

From 2d17128d6f885ab839f2ecde360268f5b523426c Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 22 May 2017 16:40:33 +0100
Subject: [PATCH 103/680] Fixing suported device list.

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 34 +++++++++----------
 1 file changed, 16 insertions(+), 18 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index c5142b7c9..06df1030b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -81,28 +81,26 @@ struct memsetCghFunctor{
   }
 };
 
-  //get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
+//get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
 EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device::get_devices()){
-  auto devices = cl::sycl::device::get_devices();
-  std::vector<cl::sycl::device>::iterator it =devices.begin();
-  while(it!=devices.end()) {
-    ///FIXME: Currently there is a bug in amd cpu OpenCL
-    auto name = (*it).template get_info<cl::sycl::info::device::name>();
-    std::transform(name.begin(), name.end(), name.begin(), ::tolower);
-    auto vendor = (*it).template get_info<cl::sycl::info::device::vendor>();
+std::vector<cl::sycl::device> supported_devices;
+auto plafrom_list =cl::sycl::platform::get_platforms();
+for(const auto& platform : plafrom_list){
+  auto device_list = platform.get_devices();
+  auto platform_name =platform.template get_info<cl::sycl::info::platform::name>();
+  std::transform(platform_name.begin(), platform_name.end(), platform_name.begin(), ::tolower);
+  for(const auto& device : device_list){
+    auto vendor = device.template get_info<cl::sycl::info::device::vendor>();
     std::transform(vendor.begin(), vendor.end(), vendor.begin(), ::tolower);
-
-    if((*it).is_cpu() && vendor.find("amd")!=std::string::npos && vendor.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
-      it = devices.erase(it);
-      //FIXME: currently there is a bug in intel gpu driver regarding memory allignment issue.
-    }else if((*it).is_gpu() && name.find("intel")!=std::string::npos){
-      it = devices.erase(it);
-    }
-    else{
-      ++it;
+    bool unsuported_condition = (device.is_cpu() && platform_name.find("amd")!=std::string::npos && vendor.find("apu") == std::string::npos) ||
+    (device.is_gpu() && platform_name.find("intel")!=std::string::npos);
+    if(!unsuported_condition){
+      std::cout << "Platform name "<< platform_name << std::endl;
+        supported_devices.push_back(device);
     }
   }
-  return devices;
+}
+return supported_devices;
 }
 
 class QueueInterface {

From 76c0fc1f955eda3d243db8960cb6fee9a5305112 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 22 May 2017 16:49:32 +0100
Subject: [PATCH 104/680] Fixing SYCL alignment issue required by TensorFlow.

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 23 +++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index c5142b7c9..627c0ab19 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -14,7 +14,23 @@
 
 #if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H)
 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
+template<size_t Align> struct CheckAlignStatically{
+  static const bool Val= (((Align&(Align-1))==0) && (Align >= sizeof(void *)));
+};
+template <bool IsAligned, size_t Align>
+struct Conditional_Allocate{
 
+EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){
+  return aligned_alloc(Align, elements);
+}
+};
+template <size_t Align>
+struct Conditional_Allocate<false, Align>{
+
+EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){
+  return malloc(elements);
+}
+};
 template <typename Scalar, size_t Align = EIGEN_MAX_ALIGN_BYTES, class Allocator = std::allocator<Scalar>>
 struct SyclAllocator {
   typedef Scalar value_type;
@@ -22,7 +38,10 @@ struct SyclAllocator {
   typedef typename std::allocator_traits<Allocator>::size_type size_type;
 
   SyclAllocator( ){};
-  Scalar* allocate(std::size_t elements) { return static_cast<Scalar*>(aligned_alloc(Align, elements)); }
+  Scalar* allocate(std::size_t elements) {
+    return static_cast<Scalar*>(Conditional_Allocate<CheckAlignStatically<Align>::Val, Align>::conditional_allocate(elements));
+  //  return static_cast<Scalar*>(aligned_alloc(Align, elements));
+  }
   void deallocate(Scalar * p, std::size_t size) { EIGEN_UNUSED_VARIABLE(size); free(p); }
 };
 
@@ -533,4 +552,4 @@ struct SyclKernelDevice:DefaultDevice{};
 
 }  // end namespace Eigen
 
-#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
\ No newline at end of file

From b42d775f1350838234f9f05bf64a4f7c12282218 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 23 May 2017 10:51:14 +0100
Subject: [PATCH 105/680] Temporarry branch for synch with upstream


From 8508db52ab832bfd52ac314c862b7beeb64dc182 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 6 Jun 2017 17:25:56 +0200
Subject: [PATCH 106/680] bug #1417: make LinSpace compatible with std::complex

---
 Eigen/src/Core/functors/NullaryFunctors.h | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h
index 6a30466fb..b03be0269 100644
--- a/Eigen/src/Core/functors/NullaryFunctors.h
+++ b/Eigen/src/Core/functors/NullaryFunctors.h
@@ -44,16 +44,16 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
 {
   linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
     m_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
-    m_interPacket(plset<Packet>(0)),
     m_flip(numext::abs(high)<numext::abs(low))
   {}
 
   template<typename IndexType>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
     if(m_flip)
-      return (i==0)? m_low : (m_high - (m_size1-i)*m_step);
+      return (i==0)? m_low : (m_high - RealScalar(m_size1-i)*m_step);
     else
-      return (i==m_size1)? m_high : (m_low + i*m_step);
+      return (i==m_size1)? m_high : (m_low + RealScalar(i)*m_step);
   }
 
   template<typename IndexType>
@@ -63,7 +63,7 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
     // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
     if(m_flip)
     {
-      Packet pi = padd(pset1<Packet>(Scalar(i-m_size1)),m_interPacket);
+      Packet pi = plset<Packet>(Scalar(i-m_size1));
       Packet res = padd(pset1<Packet>(m_high), pmul(pset1<Packet>(m_step), pi));
       if(i==0)
         res = pinsertfirst(res, m_low);
@@ -71,7 +71,7 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
     }
     else
     {
-      Packet pi = padd(pset1<Packet>(Scalar(i)),m_interPacket);
+      Packet pi = plset<Packet>(Scalar(i));
       Packet res = padd(pset1<Packet>(m_low), pmul(pset1<Packet>(m_step), pi));
       if(i==m_size1-unpacket_traits<Packet>::size+1)
         res = pinsertlast(res, m_high);
@@ -83,7 +83,6 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
   const Scalar m_high;
   const Index m_size1;
   const Scalar m_step;
-  const Packet m_interPacket;
   const bool m_flip;
 };
 

From e018142604b7f138d351dfc3c34890159cbe5fdc Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 6 Jun 2017 19:23:14 +0200
Subject: [PATCH 107/680] Make sure CholmodSupport works when included in
 multiple compilation units (issue was reported on stackoverflow.com)

---
 Eigen/src/CholmodSupport/CholmodSupport.h | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h
index 61faf43ba..dc199ece6 100644
--- a/Eigen/src/CholmodSupport/CholmodSupport.h
+++ b/Eigen/src/CholmodSupport/CholmodSupport.h
@@ -167,12 +167,12 @@ namespace internal {
 // template specializations for int and long that call the correct cholmod method
 
 #define EIGEN_CHOLMOD_SPECIALIZE0(ret, name) \
-    template<typename _StorageIndex> ret cm_ ## name       (cholmod_common &Common) { return cholmod_ ## name   (&Common); } \
-    template<>                       ret cm_ ## name<long> (cholmod_common &Common) { return cholmod_l_ ## name (&Common); }
+    template<typename _StorageIndex> inline ret cm_ ## name       (cholmod_common &Common) { return cholmod_ ## name   (&Common); } \
+    template<>                       inline ret cm_ ## name<long> (cholmod_common &Common) { return cholmod_l_ ## name (&Common); }
 
 #define EIGEN_CHOLMOD_SPECIALIZE1(ret, name, t1, a1) \
-    template<typename _StorageIndex> ret cm_ ## name       (t1& a1, cholmod_common &Common) { return cholmod_ ## name   (&a1, &Common); } \
-    template<>                       ret cm_ ## name<long> (t1& a1, cholmod_common &Common) { return cholmod_l_ ## name (&a1, &Common); }
+    template<typename _StorageIndex> inline ret cm_ ## name       (t1& a1, cholmod_common &Common) { return cholmod_ ## name   (&a1, &Common); } \
+    template<>                       inline ret cm_ ## name<long> (t1& a1, cholmod_common &Common) { return cholmod_l_ ## name (&a1, &Common); }
 
 EIGEN_CHOLMOD_SPECIALIZE0(int, start)
 EIGEN_CHOLMOD_SPECIALIZE0(int, finish)
@@ -183,16 +183,16 @@ EIGEN_CHOLMOD_SPECIALIZE1(int, free_sparse, cholmod_sparse*, A)
 
 EIGEN_CHOLMOD_SPECIALIZE1(cholmod_factor*, analyze, cholmod_sparse, A)
 
-template<typename _StorageIndex> cholmod_dense*  cm_solve         (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_solve     (sys, &L, &B, &Common); }
-template<>                       cholmod_dense*  cm_solve<long>   (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_l_solve   (sys, &L, &B, &Common); }
+template<typename _StorageIndex> inline cholmod_dense*  cm_solve         (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_solve     (sys, &L, &B, &Common); }
+template<>                       inline cholmod_dense*  cm_solve<long>   (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_l_solve   (sys, &L, &B, &Common); }
 
-template<typename _StorageIndex> cholmod_sparse* cm_spsolve       (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_spsolve   (sys, &L, &B, &Common); }
-template<>                       cholmod_sparse* cm_spsolve<long> (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_l_spsolve (sys, &L, &B, &Common); }
+template<typename _StorageIndex> inline cholmod_sparse* cm_spsolve       (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_spsolve   (sys, &L, &B, &Common); }
+template<>                       inline cholmod_sparse* cm_spsolve<long> (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_l_spsolve (sys, &L, &B, &Common); }
 
 template<typename _StorageIndex>
-int  cm_factorize_p       (cholmod_sparse*  A, double beta[2], _StorageIndex* fset, std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_factorize_p   (A, beta, fset, fsize, L, &Common); }
+inline int  cm_factorize_p       (cholmod_sparse*  A, double beta[2], _StorageIndex* fset, std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_factorize_p   (A, beta, fset, fsize, L, &Common); }
 template<>
-int  cm_factorize_p<long> (cholmod_sparse*  A, double beta[2], long* fset,          std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_l_factorize_p (A, beta, fset, fsize, L, &Common); }
+inline int  cm_factorize_p<long> (cholmod_sparse*  A, double beta[2], long* fset,          std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_l_factorize_p (A, beta, fset, fsize, L, &Common); }
 
 #undef EIGEN_CHOLMOD_SPECIALIZE0
 #undef EIGEN_CHOLMOD_SPECIALIZE1

From f2a553fb7b5dd2b8bf7b35af8fc9485f25b778b5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 7 Jun 2017 10:10:30 +0200
Subject: [PATCH 108/680] bug #1411: fix usage of alignment information in
 vectorization of quaternion product and conjugate.

---
 Eigen/src/Geometry/Quaternion.h        | 10 ++---
 Eigen/src/Geometry/arch/Geometry_SSE.h | 60 +++++++++++++++++---------
 2 files changed, 44 insertions(+), 26 deletions(-)

diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h
index f6ef1bcf6..3e5a9badb 100644
--- a/Eigen/src/Geometry/Quaternion.h
+++ b/Eigen/src/Geometry/Quaternion.h
@@ -423,7 +423,7 @@ typedef Map<Quaternion<double>, Aligned>  QuaternionMapAlignedd;
 // Generic Quaternion * Quaternion product
 // This product can be specialized for a given architecture via the Arch template argument.
 namespace internal {
-template<int Arch, class Derived1, class Derived2, typename Scalar, int _Options> struct quat_product
+template<int Arch, class Derived1, class Derived2, typename Scalar> struct quat_product
 {
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion<Scalar> run(const QuaternionBase<Derived1>& a, const QuaternionBase<Derived2>& b){
     return Quaternion<Scalar>
@@ -446,8 +446,7 @@ QuaternionBase<Derived>::operator* (const QuaternionBase<OtherDerived>& other) c
   EIGEN_STATIC_ASSERT((internal::is_same<typename Derived::Scalar, typename OtherDerived::Scalar>::value),
    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
   return internal::quat_product<Architecture::Target, Derived, OtherDerived,
-                         typename internal::traits<Derived>::Scalar,
-                         EIGEN_PLAIN_ENUM_MIN(internal::traits<Derived>::Alignment, internal::traits<OtherDerived>::Alignment)>::run(*this, other);
+                         typename internal::traits<Derived>::Scalar>::run(*this, other);
 }
 
 /** \sa operator*(Quaternion) */
@@ -672,7 +671,7 @@ EIGEN_DEVICE_FUNC inline Quaternion<typename internal::traits<Derived>::Scalar>
 
 // Generic conjugate of a Quaternion
 namespace internal {
-template<int Arch, class Derived, typename Scalar, int _Options> struct quat_conj
+template<int Arch, class Derived, typename Scalar> struct quat_conj
 {
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion<Scalar> run(const QuaternionBase<Derived>& q){
     return Quaternion<Scalar>(q.w(),-q.x(),-q.y(),-q.z());
@@ -691,8 +690,7 @@ EIGEN_DEVICE_FUNC inline Quaternion<typename internal::traits<Derived>::Scalar>
 QuaternionBase<Derived>::conjugate() const
 {
   return internal::quat_conj<Architecture::Target, Derived,
-                         typename internal::traits<Derived>::Scalar,
-                         internal::traits<Derived>::Alignment>::run(*this);
+                         typename internal::traits<Derived>::Scalar>::run(*this);
                          
 }
 
diff --git a/Eigen/src/Geometry/arch/Geometry_SSE.h b/Eigen/src/Geometry/arch/Geometry_SSE.h
index 1a86ff837..f68cab583 100644
--- a/Eigen/src/Geometry/arch/Geometry_SSE.h
+++ b/Eigen/src/Geometry/arch/Geometry_SSE.h
@@ -16,17 +16,23 @@ namespace Eigen {
 namespace internal {
 
 template<class Derived, class OtherDerived>
-struct quat_product<Architecture::SSE, Derived, OtherDerived, float, Aligned16>
+struct quat_product<Architecture::SSE, Derived, OtherDerived, float>
 {
+  enum {
+    AAlignment = traits<Derived>::Alignment,
+    BAlignment = traits<OtherDerived>::Alignment,
+    ResAlignment = traits<Quaternion<float> >::Alignment
+  };
   static inline Quaternion<float> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
   {
     Quaternion<float> res;
     const __m128 mask = _mm_setr_ps(0.f,0.f,0.f,-0.f);
-    __m128 a = _a.coeffs().template packet<Aligned16>(0);
-    __m128 b = _b.coeffs().template packet<Aligned16>(0);
+    __m128 a = _a.coeffs().template packet<AAlignment>(0);
+    __m128 b = _b.coeffs().template packet<BAlignment>(0);
     __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
     __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
-    pstore(&res.x(),
+    pstoret<float,Packet4f,ResAlignment>(
+              &res.x(),
               _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
                                     _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0),
                                                vec4f_swizzle1(b,1,2,0,0))),
@@ -36,14 +42,17 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, float, Aligned16>
   }
 };
 
-template<class Derived, int Alignment>
-struct quat_conj<Architecture::SSE, Derived, float, Alignment>
+template<class Derived>
+struct quat_conj<Architecture::SSE, Derived, float>
 {
+  enum {
+    ResAlignment = traits<Quaternion<float> >::Alignment
+  };
   static inline Quaternion<float> run(const QuaternionBase<Derived>& q)
   {
     Quaternion<float> res;
     const __m128 mask = _mm_setr_ps(-0.f,-0.f,-0.f,0.f);
-    pstore(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet<Alignment>(0)));
+    pstoret<float,Packet4f,ResAlignment>(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet<traits<Derived>::Alignment>(0)));
     return res;
   }
 };
@@ -52,6 +61,9 @@ struct quat_conj<Architecture::SSE, Derived, float, Alignment>
 template<typename VectorLhs,typename VectorRhs>
 struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
 {
+  enum {
+    ResAlignment = traits<typename plain_matrix_type<VectorLhs>::type>::Alignment
+  };
   static inline typename plain_matrix_type<VectorLhs>::type
   run(const VectorLhs& lhs, const VectorRhs& rhs)
   {
@@ -60,7 +72,7 @@ struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
     __m128 mul1=_mm_mul_ps(vec4f_swizzle1(a,1,2,0,3),vec4f_swizzle1(b,2,0,1,3));
     __m128 mul2=_mm_mul_ps(vec4f_swizzle1(a,2,0,1,3),vec4f_swizzle1(b,1,2,0,3));
     typename plain_matrix_type<VectorLhs>::type res;
-    pstore(&res.x(),_mm_sub_ps(mul1,mul2));
+    pstoret<float,Packet4f,ResAlignment>(&res.x(),_mm_sub_ps(mul1,mul2));
     return res;
   }
 };
@@ -68,9 +80,14 @@ struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
 
 
 
-template<class Derived, class OtherDerived, int Alignment>
-struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Alignment>
+template<class Derived, class OtherDerived>
+struct quat_product<Architecture::SSE, Derived, OtherDerived, double>
 {
+  enum {
+    BAlignment = traits<OtherDerived>::Alignment,
+    ResAlignment = traits<Quaternion<double> >::Alignment
+  };
+
   static inline Quaternion<double> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
   {
   const Packet2d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
@@ -78,8 +95,8 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Alignment>
   Quaternion<double> res;
 
   const double* a = _a.coeffs().data();
-  Packet2d b_xy = _b.coeffs().template packet<Alignment>(0);
-  Packet2d b_zw = _b.coeffs().template packet<Alignment>(2);
+  Packet2d b_xy = _b.coeffs().template packet<BAlignment>(0);
+  Packet2d b_zw = _b.coeffs().template packet<BAlignment>(2);
   Packet2d a_xx = pset1<Packet2d>(a[0]);
   Packet2d a_yy = pset1<Packet2d>(a[1]);
   Packet2d a_zz = pset1<Packet2d>(a[2]);
@@ -97,9 +114,9 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Alignment>
   t2 = psub(pmul(a_zz, b_xy), pmul(a_xx, b_zw));
 #ifdef EIGEN_VECTORIZE_SSE3
   EIGEN_UNUSED_VARIABLE(mask)
-  pstore(&res.x(), _mm_addsub_pd(t1, preverse(t2)));
+  pstoret<double,Packet2d,ResAlignment>(&res.x(), _mm_addsub_pd(t1, preverse(t2)));
 #else
-  pstore(&res.x(), padd(t1, pxor(mask,preverse(t2))));
+  pstoret<double,Packet2d,ResAlignment>(&res.x(), padd(t1, pxor(mask,preverse(t2))));
 #endif
   
   /*
@@ -111,25 +128,28 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Alignment>
   t2 = padd(pmul(a_zz, b_zw), pmul(a_xx, b_xy));
 #ifdef EIGEN_VECTORIZE_SSE3
   EIGEN_UNUSED_VARIABLE(mask)
-  pstore(&res.z(), preverse(_mm_addsub_pd(preverse(t1), t2)));
+  pstoret<double,Packet2d,ResAlignment>(&res.z(), preverse(_mm_addsub_pd(preverse(t1), t2)));
 #else
-  pstore(&res.z(), psub(t1, pxor(mask,preverse(t2))));
+  pstoret<double,Packet2d,ResAlignment>(&res.z(), psub(t1, pxor(mask,preverse(t2))));
 #endif
 
   return res;
 }
 };
 
-template<class Derived, int Alignment>
-struct quat_conj<Architecture::SSE, Derived, double, Alignment>
+template<class Derived>
+struct quat_conj<Architecture::SSE, Derived, double>
 {
+  enum {
+    ResAlignment = traits<Quaternion<double> >::Alignment
+  };
   static inline Quaternion<double> run(const QuaternionBase<Derived>& q)
   {
     Quaternion<double> res;
     const __m128d mask0 = _mm_setr_pd(-0.,-0.);
     const __m128d mask2 = _mm_setr_pd(-0.,0.);
-    pstore(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet<Alignment>(0)));
-    pstore(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet<Alignment>(2)));
+    pstoret<double,Packet2d,ResAlignment>(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet<traits<Derived>::Alignment>(0)));
+    pstoret<double,Packet2d,ResAlignment>(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet<traits<Derived>::Alignment>(2)));
     return res;
   }
 };

From 26e8f9171eb07e188af45d7583b03c5be40a5f48 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 7 Jun 2017 10:51:23 +0200
Subject: [PATCH 109/680] Fix compilation of matrix log with Map as input

---
 .../Eigen/src/MatrixFunctions/MatrixFunction.h        | 11 +++++------
 .../Eigen/src/MatrixFunctions/MatrixLogarithm.h       |  2 +-
 2 files changed, 6 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
index db2449d02..3f7d77710 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
@@ -398,8 +398,8 @@ struct matrix_function_compute
 template <typename MatrixType>
 struct matrix_function_compute<MatrixType, 0>
 {  
-  template <typename AtomicType, typename ResultType> 
-  static void run(const MatrixType& A, AtomicType& atomic, ResultType &result)
+  template <typename MatA, typename AtomicType, typename ResultType>
+  static void run(const MatA& A, AtomicType& atomic, ResultType &result)
   {
     typedef internal::traits<MatrixType> Traits;
     typedef typename Traits::Scalar Scalar;
@@ -422,11 +422,10 @@ struct matrix_function_compute<MatrixType, 0>
 template <typename MatrixType>
 struct matrix_function_compute<MatrixType, 1>
 {
-  template <typename AtomicType, typename ResultType> 
-  static void run(const MatrixType& A, AtomicType& atomic, ResultType &result)
+  template <typename MatA, typename AtomicType, typename ResultType>
+  static void run(const MatA& A, AtomicType& atomic, ResultType &result)
   {
     typedef internal::traits<MatrixType> Traits;
-    typedef typename MatrixType::Index Index;
     
     // compute Schur decomposition of A
     const ComplexSchur<MatrixType> schurOfA(A);  
@@ -514,7 +513,7 @@ template<typename Derived> class MatrixFunctionReturnValue
       typedef internal::MatrixFunctionAtomic<DynMatrixType> AtomicType;
       AtomicType atomic(m_f);
 
-      internal::matrix_function_compute<NestedEvalTypeClean>::run(m_A, atomic, result);
+      internal::matrix_function_compute<typename NestedEvalTypeClean::PlainObject>::run(m_A, atomic, result);
     }
 
     Index rows() const { return m_A.rows(); }
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
index 1acfbed9e..ff8f6e732 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
@@ -339,7 +339,7 @@ public:
     typedef internal::MatrixLogarithmAtomic<DynMatrixType> AtomicType;
     AtomicType atomic;
     
-    internal::matrix_function_compute<DerivedEvalTypeClean>::run(m_A, atomic, result);
+    internal::matrix_function_compute<typename DerivedEvalTypeClean::PlainObject>::run(m_A, atomic, result);
   }
 
   Index rows() const { return m_A.rows(); }

From 2971503fed85add086cb163d24ee5d402d631aab Mon Sep 17 00:00:00 2001
From: Mmanu Chaturvedi <mmanu.chaturvedi@kitware.com>
Date: Tue, 23 May 2017 17:12:36 -0400
Subject: [PATCH 110/680] Specializing numeric_limits For AutoDiffScalar

---
 unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h |  7 +++++++
 unsupported/test/autodiff_scalar.cpp            | 14 ++++++++++++++
 2 files changed, 21 insertions(+)

diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
index d2808860c..279fe5cd3 100755
--- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
+++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
@@ -683,4 +683,11 @@ template<typename DerType> struct NumTraits<AutoDiffScalar<DerType> >
 
 }
 
+namespace std {
+template <typename T>
+class numeric_limits<Eigen::AutoDiffScalar<T> >
+  : public numeric_limits<typename T::Scalar> {};
+
+}  // namespace std
+
 #endif // EIGEN_AUTODIFF_SCALAR_H
diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp
index 4df2f5c57..3a78d6b8d 100644
--- a/unsupported/test/autodiff_scalar.cpp
+++ b/unsupported/test/autodiff_scalar.cpp
@@ -72,6 +72,19 @@ template<typename Scalar> void check_hyperbolic_functions()
   VERIFY_IS_APPROX(res3.derivatives().x(), Scalar(0.339540557256150));
 }
 
+template <typename Scalar>
+void check_limits_specialization()
+{
+  typedef Eigen::Matrix<Scalar, 1, 1> Deriv;
+  typedef Eigen::AutoDiffScalar<Deriv> AD;
+
+  typedef std::numeric_limits<AD> A;
+  typedef std::numeric_limits<Scalar> B;
+
+  bool res = std::is_base_of<B, A>::value;
+  VERIFY_IS_EQUAL(res, true);
+}
+
 void test_autodiff_scalar()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -79,5 +92,6 @@ void test_autodiff_scalar()
     CALL_SUBTEST_2( check_atan2<double>() );
     CALL_SUBTEST_3( check_hyperbolic_functions<float>() );
     CALL_SUBTEST_4( check_hyperbolic_functions<double>() );
+    CALL_SUBTEST_5( check_limits_specialization<double>());
   }
 }

From 0cb3c7c7dd46d318921b5e140afc73fded87b068 Mon Sep 17 00:00:00 2001
From: Duncan McBain <duncan@codeplay.com>
Date: Wed, 24 May 2017 12:24:21 +0100
Subject: [PATCH 111/680] Update FindComputeCpp.cmake with new changes from SDK

---
 cmake/FindComputeCpp.cmake | 49 ++++++++++++++++++++++++++------------
 1 file changed, 34 insertions(+), 15 deletions(-)

diff --git a/cmake/FindComputeCpp.cmake b/cmake/FindComputeCpp.cmake
index 90222ed0e..e61dedc46 100644
--- a/cmake/FindComputeCpp.cmake
+++ b/cmake/FindComputeCpp.cmake
@@ -38,11 +38,6 @@ if(CMAKE_COMPILER_IS_GNUCXX)
     if (CMAKE_CXX_COMPILER_VERSION VERSION_LESS 4.8)
       message(FATAL_ERROR
         "host compiler - Not found! (gcc version must be at least 4.8)")
-    # Require the GCC dual ABI to be disabled for 5.1 or higher
-    elseif (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 5.1)
-      set(COMPUTECPP_DISABLE_GCC_DUAL_ABI "True")
-      message(STATUS
-        "host compiler - gcc ${CMAKE_CXX_COMPILER_VERSION} (note pre 5.1 gcc ABI enabled)")
     else()
       message(STATUS "host compiler - gcc ${CMAKE_CXX_COMPILER_VERSION}")
     endif()
@@ -64,6 +59,12 @@ option(COMPUTECPP_64_BIT_CODE "Compile device code in 64 bit mode"
         ${COMPUTECPP_64_BIT_DEFAULT})
 mark_as_advanced(COMPUTECPP_64_BIT_CODE)
 
+option(COMPUTECPP_DISABLE_GCC_DUAL_ABI "Compile with pre-5.1 ABI" OFF)
+mark_as_advanced(COMPUTECPP_DISABLE_GCC_DUAL_ABI)
+
+set(COMPUTECPP_USER_FLAGS "" CACHE STRING "User flags for compute++")
+mark_as_advanced(COMPUTECPP_USER_FLAGS)
+
 # Find OpenCL package
 find_package(OpenCL REQUIRED)
 
@@ -74,7 +75,6 @@ if(NOT COMPUTECPP_PACKAGE_ROOT_DIR)
 else()
   message(STATUS "ComputeCpp package - Found")
 endif()
-option(COMPUTECPP_PACKAGE_ROOT_DIR "Path to the ComputeCpp Package")
 
 # Obtain the path to compute++
 find_program(COMPUTECPP_DEVICE_COMPILER compute++ PATHS
@@ -138,8 +138,6 @@ else()
   message(STATUS "compute++ flags - ${COMPUTECPP_DEVICE_COMPILER_FLAGS}")
 endif()
 
-set(COMPUTECPP_DEVICE_COMPILER_FLAGS ${COMPUTECPP_DEVICE_COMPILER_FLAGS} -sycl-compress-name -Wall -no-serial-memop -DEIGEN_NO_ASSERTION_CHECKING=1)
-
 # Check if the platform is supported
 execute_process(COMMAND ${COMPUTECPP_INFO_TOOL} "--dump-is-supported"
   OUTPUT_VARIABLE COMPUTECPP_PLATFORM_IS_SUPPORTED
@@ -155,6 +153,13 @@ else()
   endif()
 endif()
 
+set(COMPUTECPP_USER_FLAGS
+  -sycl-compress-name
+  -Wall
+  -no-serial-memop
+  -DEIGEN_NO_ASSERTION_CHECKING=1
+  )
+
 ####################
 #   __build_sycl
 ####################
@@ -165,8 +170,11 @@ endif()
 #  targetName : Name of the target.
 #  sourceFile : Source file to be compiled.
 #  binaryDir : Intermediate directory to output the integration header.
+#  fileCounter : Counter included in name of custom target. Different counter
+#       values prevent duplicated names of custom target when source files with the same name,
+#       but located in different directories, are used for the same target.
 #
-function(__build_spir targetName sourceFile binaryDir)
+function(__build_spir targetName sourceFile binaryDir fileCounter)
 
   # Retrieve source file name.
   get_filename_component(sourceFileName ${sourceFile} NAME)
@@ -175,12 +183,16 @@ function(__build_spir targetName sourceFile binaryDir)
   set(outputSyclFile ${binaryDir}/${sourceFileName}.sycl)
 
   # Add any user-defined include to the device compiler
+  set(device_compiler_includes "")
   get_property(includeDirectories DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} PROPERTY
     INCLUDE_DIRECTORIES)
-  set(device_compiler_includes "")
   foreach(directory ${includeDirectories})
     set(device_compiler_includes "-I${directory}" ${device_compiler_includes})
   endforeach()
+  get_target_property(targetIncludeDirectories ${targetName} INCLUDE_DIRECTORIES)
+  foreach(directory ${targetIncludeDirectories})
+    set(device_compiler_includes "-I${directory}" ${device_compiler_includes})
+  endforeach()
   if (CMAKE_INCLUDE_PATH)
     foreach(directory ${CMAKE_INCLUDE_PATH})
       set(device_compiler_includes "-I${directory}"
@@ -188,6 +200,9 @@ function(__build_spir targetName sourceFile binaryDir)
     endforeach()
   endif()
 
+  set(COMPUTECPP_DEVICE_COMPILER_FLAGS
+    ${COMPUTECPP_DEVICE_COMPILER_FLAGS}
+    ${COMPUTECPP_USER_FLAGS})
   # Convert argument list format
   separate_arguments(COMPUTECPP_DEVICE_COMPILER_FLAGS)
 
@@ -204,7 +219,7 @@ function(__build_spir targetName sourceFile binaryDir)
     DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile}
     IMPLICIT_DEPENDS CXX "${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile}"
     WORKING_DIRECTORY ${binaryDir}
-  COMMENT "Building ComputeCpp integration header file ${outputSyclFile}")
+    COMMENT "Building ComputeCpp integration header file ${outputSyclFile}")
 
   # Add a custom target for the generated integration header
   add_custom_target(${targetName}_integration_header DEPENDS ${outputSyclFile})
@@ -231,17 +246,21 @@ endfunction()
 #  target and sets a dependancy on that new command.
 #
 #  targetName : Name of the target to add a SYCL to.
-#  sourceFile : Source file to be compiled for SYCL.
 #  binaryDir : Intermediate directory to output the integration header.
+#  sourceFiles : Source files to be compiled for SYCL.
 #
-function(add_sycl_to_target targetName binaryDir sourceFile)
+function(add_sycl_to_target targetName binaryDir sourceFiles)
 
   set(sourceFiles ${sourceFiles} ${ARGN})
+  set(fileCounter 0)
   # Add custom target to run compute++ and generate the integration header
-  __build_spir(${targetName} ${sourceFile} ${binaryDir})
+  foreach(sourceFile ${sourceFiles})
+    __build_spir(${targetName} ${sourceFile} ${binaryDir} ${fileCounter})
+    MATH(EXPR fileCounter "${fileCounter} + 1")
+  endforeach()
 
   # Link with the ComputeCpp runtime library
   target_link_libraries(${targetName} PUBLIC ${COMPUTECPP_RUNTIME_LIBRARY}
                         PUBLIC ${OpenCL_LIBRARIES})
 
-endfunction(add_sycl_to_target)
\ No newline at end of file
+endfunction(add_sycl_to_target)

From 9ef5c948ba0ae4e24941b1319aac0b7584f10795 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 24 May 2017 13:11:16 +0100
Subject: [PATCH 112/680] Fixing Cmake for gcc>=5.


From fb853a857a5d3e06687c6736084c20e984b7347d Mon Sep 17 00:00:00 2001
From: a-doumoulakis <anonymous@invalid.net>
Date: Wed, 24 May 2017 17:50:15 +0100
Subject: [PATCH 113/680] Restore misplaced comment

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 3105cc122..06df1030b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -81,7 +81,7 @@ struct memsetCghFunctor{
   }
 };
 
- //get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
+//get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
 EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device::get_devices()){
 std::vector<cl::sycl::device> supported_devices;
 auto plafrom_list =cl::sycl::platform::get_platforms();

From e3f964ed55a96d0c94814d07ab88d8805e0c2eec Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 25 May 2017 11:17:26 +0100
Subject: [PATCH 114/680] Applying Benoit's comment;removing dead code.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 627c0ab19..27c10a0cf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -40,7 +40,6 @@ struct SyclAllocator {
   SyclAllocator( ){};
   Scalar* allocate(std::size_t elements) {
     return static_cast<Scalar*>(Conditional_Allocate<CheckAlignStatically<Align>::Val, Align>::conditional_allocate(elements));
-  //  return static_cast<Scalar*>(aligned_alloc(Align, elements));
   }
   void deallocate(Scalar * p, std::size_t size) { EIGEN_UNUSED_VARIABLE(size); free(p); }
 };

From c3bd860de8f1d473a05072917f56b380ba59533b Mon Sep 17 00:00:00 2001
From: a-doumoulakis <anonymous@invalid.net>
Date: Thu, 25 May 2017 18:46:18 +0100
Subject: [PATCH 115/680] Modification upon request

- Remove warning suppression
---
 CMakeLists.txt | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 54f3a7509..2e1648fd1 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -151,10 +151,6 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-Wenum-conversion")
   ei_add_cxx_compiler_flag("-Wc++11-extensions")
   ei_add_cxx_compiler_flag("-Wdouble-promotion")
-  ei_add_cxx_compiler_flag("-Wno-unused-parameter")
-  ei_add_cxx_compiler_flag("-Wno-ignored-attributes")
-  ei_add_cxx_compiler_flag("-Wno-ignored-qualifiers")
-  ei_add_cxx_compiler_flag("-Wno-sign-compare")
 #  ei_add_cxx_compiler_flag("-Wconversion")
   
   # -Wshadow is insanely too strict with gcc, hopefully it will become usable with gcc 6

From 0370d3576e1df4f898b9030fbc269ec0488d3969 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Fri, 26 May 2017 16:01:48 +0100
Subject: [PATCH 116/680] Applying Ronnan's comments.

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 20 +++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 27c10a0cf..09c96317e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -14,22 +14,22 @@
 
 #if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H)
 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
-template<size_t Align> struct CheckAlignStatically{
+template<size_t Align> struct CheckAlignStatically {
   static const bool Val= (((Align&(Align-1))==0) && (Align >= sizeof(void *)));
 };
 template <bool IsAligned, size_t Align>
-struct Conditional_Allocate{
+struct Conditional_Allocate {
 
-EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){
-  return aligned_alloc(Align, elements);
-}
+  EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements) {
+    return aligned_alloc(Align, elements);
+  }
 };
 template <size_t Align>
-struct Conditional_Allocate<false, Align>{
+struct Conditional_Allocate<false, Align> {
 
-EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){
-  return malloc(elements);
-}
+  EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){
+    return malloc(elements);
+  }
 };
 template <typename Scalar, size_t Align = EIGEN_MAX_ALIGN_BYTES, class Allocator = std::allocator<Scalar>>
 struct SyclAllocator {
@@ -551,4 +551,4 @@ struct SyclKernelDevice:DefaultDevice{};
 
 }  // end namespace Eigen
 
-#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
\ No newline at end of file
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H

From 9341f258d4ee8a819c31cec8a9dc027a10669372 Mon Sep 17 00:00:00 2001
From: Hugh Perkins <hughperkins@gmail.com>
Date: Tue, 6 Jun 2017 15:51:06 +0100
Subject: [PATCH 117/680] Add labels to #ifdef, in TensorReductionCuda.h

---
 .../CXX11/src/Tensor/TensorReductionCuda.h    | 36 +++++++++----------
 1 file changed, 18 insertions(+), 18 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index edb0ab280..24a55a3d5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -62,9 +62,9 @@ __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer)
   else {
     assert(0 && "Wordsize not supported");
   }
-#else
+#else // __CUDA_ARCH__ >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif
+#endif // __CUDA_ARCH__ >= 300
 }
 
 // We extend atomicExch to support extra data types
@@ -98,15 +98,15 @@ __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer
     }
   }
 }
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
 template <>
 __device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
 #if __CUDA_ARCH__ >= 300
   atomicAdd(output, accum);
-#else
+#else // __CUDA_ARCH__ >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif
+#endif // __CUDA_ARCH__ >= 300
 }
 
 
@@ -179,9 +179,9 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
     // Let the last block reset the semaphore
     atomicInc(semaphore, gridDim.x + 1);
   }
-#else
+#else // __CUDA_ARCH__ >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif
+#endif // __CUDA_ARCH__ >= 300
 }
 
 
@@ -268,7 +268,7 @@ __global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2
   *output = tmp;
 }
 
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct FullReductionLauncher {
@@ -335,7 +335,7 @@ struct FullReductionLauncher<Self, Op, Eigen::half, true> {
     }
   }
 };
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
 
 template <typename Self, typename Op, bool Vectorizable>
@@ -348,11 +348,11 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else
+#else // EIGEN_HAS_CUDA_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
   template <typename OutputType>
   static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
@@ -433,9 +433,9 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
       }
     }
   }
-#else
+#else // __CUDA_ARCH__ >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif
+#endif // __CUDA_ARCH__ >= 300
 }
 
 #ifdef EIGEN_HAS_CUDA_FP16
@@ -533,7 +533,7 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
   }
 }
 
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct InnerReductionLauncher {
@@ -625,7 +625,7 @@ struct InnerReductionLauncher<Self, Op, Eigen::half, true> {
     return false;
   }
 };
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
 
 template <typename Self, typename Op>
@@ -638,11 +638,11 @@ struct InnerReducer<Self, Op, GpuDevice> {
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else
+#else // EIGEN_HAS_CUDA_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                  (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                   internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif
+#endif // EIGEN_HAS_CUDA_FP16
 
   template <typename OutputType>
   static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
@@ -740,7 +740,7 @@ struct OuterReducer<Self, Op, GpuDevice> {
   }
 };
 
-#endif
+#endif // defined(EIGEN_USE_GPU) && defined(__CUDACC__)
 
 
 } // end namespace internal

From 4bbc32046810f65bb0f77f6dbe538abad51de281 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 8 Jun 2017 12:55:25 +0200
Subject: [PATCH 118/680] bug #1435: fix aliasing issue in exressions like: A =
 C - B*A;

---
 Eigen/src/Core/ProductEvaluators.h | 6 ++++++
 test/product_notemporary.cpp       | 1 +
 2 files changed, 7 insertions(+)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 583b7f59e..c42725dbd 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -207,6 +207,12 @@ struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename
   static const bool value = true;
 };
 
+template<typename OtherXpr, typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_difference_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr,
+                                               const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > {
+  static const bool value = true;
+};
+
 template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2>
 struct assignment_from_xpr_op_product
 {
diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp
index 8bf71b4f2..30592b79e 100644
--- a/test/product_notemporary.cpp
+++ b/test/product_notemporary.cpp
@@ -51,6 +51,7 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m)
   VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * (m1 * m2.transpose()), 0);
 
   VERIFY_EVALUATION_COUNT( m3 = m3 + (m1 * m2.adjoint()), 1);
+  VERIFY_EVALUATION_COUNT( m3 = m3 - (m1 * m2.adjoint()), 1);
 
   VERIFY_EVALUATION_COUNT( m3 = m3 + (m1 * m2.adjoint()).transpose(), 1);
   VERIFY_EVALUATION_COUNT( m3.noalias() = m3 + m1 * m2.transpose(), 0);

From 682b2ef17ee360ac11cebbe7286dc4edd9accfa3 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 8 Jun 2017 15:06:27 +0200
Subject: [PATCH 119/680] bug #1423: fix LSCG\'s Jacobi preconditioner for
 row-major matrices.

---
 .../BasicPreconditioners.h                    | 27 ++++++++++++++-----
 test/lscg.cpp                                 |  8 ++++++
 2 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
index 358444aff..279c9173c 100644
--- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
+++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
@@ -152,13 +152,28 @@ class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
     {
       // Compute the inverse squared-norm of each column of mat
       m_invdiag.resize(mat.cols());
-      for(Index j=0; j<mat.outerSize(); ++j)
+      if(MatType::IsRowMajor)
       {
-        RealScalar sum = mat.innerVector(j).squaredNorm();
-        if(sum>0)
-          m_invdiag(j) = RealScalar(1)/sum;
-        else
-          m_invdiag(j) = RealScalar(1);
+        m_invdiag.setZero();
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          for(typename MatType::InnerIterator it(mat,j); it; ++it)
+            m_invdiag(it.index()) += it.value();
+        }
+        for(Index j=0; j<mat.cols(); ++j)
+          if(m_invdiag(j)>0)
+            m_invdiag(j) = RealScalar(1)/m_invdiag(j);
+      }
+      else
+      {
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          RealScalar sum = mat.innerVector(j).squaredNorm();
+          if(sum>0)
+            m_invdiag(j) = RealScalar(1)/sum;
+          else
+            m_invdiag(j) = RealScalar(1);
+        }
       }
       Base::m_isInitialized = true;
       return *this;
diff --git a/test/lscg.cpp b/test/lscg.cpp
index daa62a954..d49ee00c3 100644
--- a/test/lscg.cpp
+++ b/test/lscg.cpp
@@ -14,12 +14,20 @@ template<typename T> void test_lscg_T()
 {
   LeastSquaresConjugateGradient<SparseMatrix<T> > lscg_colmajor_diag;
   LeastSquaresConjugateGradient<SparseMatrix<T>, IdentityPreconditioner> lscg_colmajor_I;
+  LeastSquaresConjugateGradient<SparseMatrix<T,RowMajor> > lscg_rowmajor_diag;
+  LeastSquaresConjugateGradient<SparseMatrix<T,RowMajor>, IdentityPreconditioner> lscg_rowmajor_I;
 
   CALL_SUBTEST( check_sparse_square_solving(lscg_colmajor_diag)  );
   CALL_SUBTEST( check_sparse_square_solving(lscg_colmajor_I)     );
   
   CALL_SUBTEST( check_sparse_leastsquare_solving(lscg_colmajor_diag)  );
   CALL_SUBTEST( check_sparse_leastsquare_solving(lscg_colmajor_I)     );
+
+  CALL_SUBTEST( check_sparse_square_solving(lscg_rowmajor_diag)  );
+  CALL_SUBTEST( check_sparse_square_solving(lscg_rowmajor_I)     );
+
+  CALL_SUBTEST( check_sparse_leastsquare_solving(lscg_rowmajor_diag)  );
+  CALL_SUBTEST( check_sparse_leastsquare_solving(lscg_rowmajor_I)     );
 }
 
 void test_lscg()

From d58882277999508b1b97624771433422897e9aee Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 11:51:53 +0200
Subject: [PATCH 120/680] Add missing std::numeric_limits specialization for
 half, and complete NumTraits<half>

---
 Eigen/src/Core/arch/CUDA/Half.h | 50 +++++++++++++++++++++++++++++++++
 test/half_float.cpp             | 22 +++++++++++----
 2 files changed, 67 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 67518da9f..5a9bb52e6 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -493,9 +493,59 @@ template<> struct is_arithmetic<half> { enum { value = true }; };
 
 } // end namespace internal
 
+}  // end namespace Eigen
+
+namespace std {
+template<>
+struct numeric_limits<Eigen::half> {
+  static const bool is_specialized = true;
+  static const bool is_signed = true;
+  static const bool is_integer = false;
+  static const bool is_exact = false;
+  static const bool has_infinity = true;
+  static const bool has_quiet_NaN = true;
+  static const bool has_signaling_NaN = true;
+  static const float_denorm_style has_denorm = denorm_present;
+  static const bool has_denorm_loss = false;
+  std::float_round_style round_style = std::round_to_nearest;
+  static const bool is_iec559 = false;
+  static const bool is_bounded = false;
+  static const bool is_modulo = false;
+  static const int digits = 11;
+  static const int digits10 = 2;
+  //static const int max_digits10 = ;
+  static const int radix = 2;
+  static const int min_exponent = -13;
+  static const int min_exponent10 = -4;
+  static const int max_exponent = 16;
+  static const int max_exponent10 = 4;
+  static const bool traps = true;
+  static const bool tinyness_before = false;
+
+  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
+  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
+  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
+  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
+  static Eigen::half round_error() { return epsilon()/2; }
+  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
+  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
+};
+}
+
+namespace Eigen {
+
 template<> struct NumTraits<Eigen::half>
     : GenericNumTraits<Eigen::half>
 {
+  enum {
+    IsSigned = true,
+    IsInteger = false,
+    IsComplex = false,
+    RequireInitialization = false
+  };
+
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
     return half_impl::raw_uint16_to_half(0x0800);
   }
diff --git a/test/half_float.cpp b/test/half_float.cpp
index 6f3196852..7ec40f3ae 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -96,12 +96,24 @@ void test_conversion()
 
 void test_numtraits()
 {
-  std::cout << "epsilon  = " << NumTraits<half>::epsilon() << std::endl;
-  std::cout << "highest  = " << NumTraits<half>::highest() << std::endl;
-  std::cout << "lowest   = " << NumTraits<half>::lowest() << std::endl;
-  std::cout << "inifinty = " << NumTraits<half>::infinity() << std::endl;
-  std::cout << "nan      = " << NumTraits<half>::quiet_NaN() << std::endl;
+  std::cout << "epsilon       = " << NumTraits<half>::epsilon() << "  (0x" << std::hex << NumTraits<half>::epsilon().x << ")" << std::endl;
+  std::cout << "highest       = " << NumTraits<half>::highest() << "  (0x" << std::hex << NumTraits<half>::highest().x << ")" << std::endl;
+  std::cout << "lowest        = " << NumTraits<half>::lowest() << "  (0x" << std::hex << NumTraits<half>::lowest().x << ")" << std::endl;
+  std::cout << "min           = " << (std::numeric_limits<half>::min)() << "  (0x" << std::hex << half((std::numeric_limits<half>::min)()).x << ")" << std::endl;
+  std::cout << "denorm min    = " << (std::numeric_limits<half>::denorm_min)() << "  (0x" << std::hex << half((std::numeric_limits<half>::denorm_min)()).x << ")" << std::endl;
+  std::cout << "infinity      = " << NumTraits<half>::infinity() << "  (0x" << std::hex << NumTraits<half>::infinity().x << ")" << std::endl;
+  std::cout << "quiet nan     = " << NumTraits<half>::quiet_NaN() << "  (0x" << std::hex << NumTraits<half>::quiet_NaN().x << ")" << std::endl;
+  std::cout << "signaling nan = " << std::numeric_limits<half>::signaling_NaN() << "  (0x" << std::hex << std::numeric_limits<half>::signaling_NaN().x << ")" << std::endl;
 
+  VERIFY(NumTraits<half>::IsSigned);
+
+  VERIFY_IS_EQUAL( std::numeric_limits<half>::infinity().x, half(std::numeric_limits<float>::infinity()).x );
+  VERIFY_IS_EQUAL( std::numeric_limits<half>::quiet_NaN().x, half(std::numeric_limits<float>::quiet_NaN()).x );
+  VERIFY_IS_EQUAL( std::numeric_limits<half>::signaling_NaN().x, half(std::numeric_limits<float>::signaling_NaN()).x );
+  VERIFY( (std::numeric_limits<half>::min)() > half(0.f) );
+  VERIFY( (std::numeric_limits<half>::denorm_min)() > half(0.f) );
+  VERIFY( (std::numeric_limits<half>::min)()/2 > half(0.f) );
+  VERIFY_IS_EQUAL( (std::numeric_limits<half>::denorm_min)()/2, half(0.f) );
 }
 
 void test_arithmetic()

From 498aa95a8b2b505c919efff8857a6fd2a5bbc172 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 11:53:49 +0200
Subject: [PATCH 121/680] bug #1424: add numext::abs specialization for
 unsigned integer types.

---
 Eigen/src/Core/MathFunctions.h | 10 ++++++-
 test/CMakeLists.txt            |  1 +
 test/main.h                    | 11 +++++++
 test/numext.cpp                | 53 ++++++++++++++++++++++++++++++++++
 4 files changed, 74 insertions(+), 1 deletion(-)
 create mode 100644 test/numext.cpp

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 5ec6c395e..0be4a25da 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1232,11 +1232,19 @@ double log(const double &x) { return ::log(x); }
 
 template<typename T>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-typename NumTraits<T>::Real abs(const T &x) {
+typename internal::enable_if<NumTraits<T>::IsSigned || NumTraits<T>::IsComplex,typename NumTraits<T>::Real>::type
+abs(const T &x) {
   EIGEN_USING_STD_MATH(abs);
   return abs(x);
 }
 
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+typename internal::enable_if<!(NumTraits<T>::IsSigned || NumTraits<T>::IsComplex),typename NumTraits<T>::Real>::type
+abs(const T &x) {
+  return x;
+}
+
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
 EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 8da51ce57..e73ab92b4 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -148,6 +148,7 @@ add_custom_target(BuildOfficial)
 
 ei_add_test(rand)
 ei_add_test(meta)
+ei_add_test(numext)
 ei_add_test(sizeof)
 ei_add_test(dynalloc)
 ei_add_test(nomalloc)
diff --git a/test/main.h b/test/main.h
index 25d2dcf43..bd5325196 100644
--- a/test/main.h
+++ b/test/main.h
@@ -310,6 +310,17 @@ template<> inline float test_precision<std::complex<float> >() { return test_pre
 template<> inline double test_precision<std::complex<double> >() { return test_precision<double>(); }
 template<> inline long double test_precision<std::complex<long double> >() { return test_precision<long double>(); }
 
+inline bool test_isApprox(const short& a, const short& b)
+{ return internal::isApprox(a, b, test_precision<short>()); }
+inline bool test_isApprox(const unsigned short& a, const unsigned short& b)
+{ return internal::isApprox(a, b, test_precision<unsigned long>()); }
+inline bool test_isApprox(const unsigned int& a, const unsigned int& b)
+{ return internal::isApprox(a, b, test_precision<unsigned int>()); }
+inline bool test_isApprox(const long& a, const long& b)
+{ return internal::isApprox(a, b, test_precision<long>()); }
+inline bool test_isApprox(const unsigned long& a, const unsigned long& b)
+{ return internal::isApprox(a, b, test_precision<unsigned long>()); }
+
 inline bool test_isApprox(const int& a, const int& b)
 { return internal::isApprox(a, b, test_precision<int>()); }
 inline bool test_isMuchSmallerThan(const int& a, const int& b)
diff --git a/test/numext.cpp b/test/numext.cpp
new file mode 100644
index 000000000..3de33e2f9
--- /dev/null
+++ b/test/numext.cpp
@@ -0,0 +1,53 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+
+template<typename T>
+void check_abs() {
+  typedef typename NumTraits<T>::Real Real;
+
+  if(NumTraits<T>::IsSigned)
+    VERIFY_IS_EQUAL(numext::abs(-T(1)), T(1));
+  VERIFY_IS_EQUAL(numext::abs(T(0)), T(0));
+  VERIFY_IS_EQUAL(numext::abs(T(1)), T(1));
+
+  for(int k=0; k<g_repeat*100; ++k)
+  {
+    T x = internal::random<T>();
+    if(!internal::is_same<T,bool>::value)
+      x = x/Real(2);
+    if(NumTraits<T>::IsSigned)
+    {
+      VERIFY_IS_EQUAL(numext::abs(x), numext::abs(-x));
+      VERIFY( numext::abs(-x) >= Real(0));
+    }
+    VERIFY( numext::abs(x) >= Real(0));
+    VERIFY_IS_APPROX( numext::abs2(x), numext::abs2(numext::abs(x)) );
+  }
+}
+
+void test_numext() {
+  CALL_SUBTEST( check_abs<bool>() );
+  CALL_SUBTEST( check_abs<signed char>() );
+  CALL_SUBTEST( check_abs<unsigned char>() );
+  CALL_SUBTEST( check_abs<short>() );
+  CALL_SUBTEST( check_abs<unsigned short>() );
+  CALL_SUBTEST( check_abs<int>() );
+  CALL_SUBTEST( check_abs<unsigned int>() );
+  CALL_SUBTEST( check_abs<long>() );
+  CALL_SUBTEST( check_abs<unsigned long>() );
+  CALL_SUBTEST( check_abs<half>() );
+  CALL_SUBTEST( check_abs<float>() );
+  CALL_SUBTEST( check_abs<double>() );
+  CALL_SUBTEST( check_abs<long double>() );
+
+  CALL_SUBTEST( check_abs<std::complex<float> >() );
+  CALL_SUBTEST( check_abs<std::complex<double> >() );
+}

From a7be4cd1b1ea4d85165f003d793f1d46b199b7bd Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 11:57:53 +0200
Subject: [PATCH 122/680] Fix LeastSquareDiagonalPreconditioner for complexes
 (issue introduced in previous commit)

---
 Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
index 279c9173c..facdaf890 100644
--- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
+++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
@@ -158,18 +158,18 @@ class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
         for(Index j=0; j<mat.outerSize(); ++j)
         {
           for(typename MatType::InnerIterator it(mat,j); it; ++it)
-            m_invdiag(it.index()) += it.value();
+            m_invdiag(it.index()) += numext::abs2(it.value());
         }
         for(Index j=0; j<mat.cols(); ++j)
-          if(m_invdiag(j)>0)
-            m_invdiag(j) = RealScalar(1)/m_invdiag(j);
+          if(numext::real(m_invdiag(j))>RealScalar(0))
+            m_invdiag(j) = RealScalar(1)/numext::real(m_invdiag(j));
       }
       else
       {
         for(Index j=0; j<mat.outerSize(); ++j)
         {
           RealScalar sum = mat.innerVector(j).squaredNorm();
-          if(sum>0)
+          if(sum>RealScalar(0))
             m_invdiag(j) = RealScalar(1)/sum;
           else
             m_invdiag(j) = RealScalar(1);

From 8640093af1f66dd6c390420ee6919f7ba0c89e0c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 12:45:01 +0200
Subject: [PATCH 123/680] fix compilation in C++98

---
 unsupported/test/autodiff_scalar.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp
index 3a78d6b8d..9cf11280c 100644
--- a/unsupported/test/autodiff_scalar.cpp
+++ b/unsupported/test/autodiff_scalar.cpp
@@ -81,8 +81,9 @@ void check_limits_specialization()
   typedef std::numeric_limits<AD> A;
   typedef std::numeric_limits<Scalar> B;
 
-  bool res = std::is_base_of<B, A>::value;
-  VERIFY_IS_EQUAL(res, true);
+#if EIGEN_HAS_CXX11
+  VERIFY(bool(std::is_base_of<B, A>::value));
+#endif
 }
 
 void test_autodiff_scalar()

From 723a59ac26d28e42ad07221dc30b75ea8049ce1f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 12:54:40 +0200
Subject: [PATCH 124/680] add regression test for aliasing in product
 rewritting

---
 test/product.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/test/product.h b/test/product.h
index 3b6511270..0425a929e 100644
--- a/test/product.h
+++ b/test/product.h
@@ -216,6 +216,8 @@ template<typename MatrixType> void product(const MatrixType& m)
     // CwiseBinaryOp
     VERIFY_IS_APPROX(x = y + A*x, A*z);
     x = z;
+    VERIFY_IS_APPROX(x = y - A*x, A*(-z));
+    x = z;
     // CwiseUnaryOp
     VERIFY_IS_APPROX(x = Scalar(1.)*(A*x), A*z);
   }

From fb1ee04087f6d5cfe9009941c1a1eb25b3133a04 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 13:13:03 +0200
Subject: [PATCH 125/680] bug #1410: fix lvalue propagation of
 Array/Matrix-Wrapper with a const nested expression.

---
 Eigen/src/Core/ArrayWrapper.h |  6 ++++--
 test/array_for_matrix.cpp     | 22 +++++++++++++++++++++-
 2 files changed, 25 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h
index a04521a16..688aadd62 100644
--- a/Eigen/src/Core/ArrayWrapper.h
+++ b/Eigen/src/Core/ArrayWrapper.h
@@ -32,7 +32,8 @@ struct traits<ArrayWrapper<ExpressionType> >
   // Let's remove NestByRefBit
   enum {
     Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
-    Flags = Flags0 & ~NestByRefBit
+    LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0,
+    Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag
   };
 };
 }
@@ -129,7 +130,8 @@ struct traits<MatrixWrapper<ExpressionType> >
   // Let's remove NestByRefBit
   enum {
     Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
-    Flags = Flags0 & ~NestByRefBit
+    LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0,
+    Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag
   };
 };
 }
diff --git a/test/array_for_matrix.cpp b/test/array_for_matrix.cpp
index c1501947b..b8721391f 100644
--- a/test/array_for_matrix.cpp
+++ b/test/array_for_matrix.cpp
@@ -235,12 +235,31 @@ template<typename MatrixTraits> void resize(const MatrixTraits& t)
   VERIFY(a1.size()==cols);
 }
 
+template<int>
 void regression_bug_654()
 {
   ArrayXf a = RowVectorXf(3);
   VectorXf v = Array<float,1,Dynamic>(3);
 }
 
+// Check propagation of LvalueBit through Array/Matrix-Wrapper
+template<int>
+void regrrssion_bug_1410()
+{
+  const Matrix4i M;
+  const Array4i A;
+  ArrayWrapper<const Matrix4i> MA = M.array();
+  MA.row(0);
+  MatrixWrapper<const Array4i> AM = A.matrix();
+  AM.row(0);
+
+  VERIFY((internal::traits<ArrayWrapper<const Matrix4i> >::Flags&LvalueBit)==0);
+  VERIFY((internal::traits<MatrixWrapper<const Array4i> >::Flags&LvalueBit)==0);
+
+  VERIFY((internal::traits<ArrayWrapper<Matrix4i> >::Flags&LvalueBit)==LvalueBit);
+  VERIFY((internal::traits<MatrixWrapper<Array4i> >::Flags&LvalueBit)==LvalueBit);
+}
+
 void test_array_for_matrix()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -280,5 +299,6 @@ void test_array_for_matrix()
     CALL_SUBTEST_5( resize(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_6( resize(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
   }
-  CALL_SUBTEST_6( regression_bug_654() );
+  CALL_SUBTEST_6( regression_bug_654<0>() );
+  CALL_SUBTEST_6( regrrssion_bug_1410<0>() );
 }

From 1d59ca245811b5becc24bc89510b3d359f3a2dc1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 13:20:52 +0200
Subject: [PATCH 126/680] Fix compilation with gcc 4.3 and ARM NEON

---
 Eigen/src/Core/arch/NEON/PacketMath.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index aede4a6d5..836fbc0dd 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -116,7 +116,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t&    from)
 
 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)
 {
-  const float32_t f[] = {0, 1, 2, 3};
+  const float f[] = {0, 1, 2, 3};
   Packet4f countdown = vld1q_f32(f);
   return vaddq_f32(pset1<Packet4f>(a), countdown);
 }

From 26f552c18de48fa04f2a9a9a27540a7679328254 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 13:36:58 +0200
Subject: [PATCH 127/680] fix compilation of Half in C++98 (issue introduced in
 previous commit)

---
 Eigen/src/Core/arch/CUDA/Half.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 5a9bb52e6..e4e639fcd 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -507,7 +507,7 @@ struct numeric_limits<Eigen::half> {
   static const bool has_signaling_NaN = true;
   static const float_denorm_style has_denorm = denorm_present;
   static const bool has_denorm_loss = false;
-  std::float_round_style round_style = std::round_to_nearest;
+  static const std::float_round_style round_style = std::round_to_nearest;
   static const bool is_iec559 = false;
   static const bool is_bounded = false;
   static const bool is_modulo = false;
@@ -526,7 +526,7 @@ struct numeric_limits<Eigen::half> {
   static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
   static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
   static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
-  static Eigen::half round_error() { return epsilon()/2; }
+  static Eigen::half round_error() { return Eigen::half(0.5); }
   static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
   static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
   static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }

From 90168c003d80236e935589d56f9fbd8e9c7aaa75 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 14:01:44 +0200
Subject: [PATCH 128/680] bug #1414: doxygen, add EigenBase to CoreModule

---
 Eigen/src/Core/EigenBase.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h
index ccc122cfc..b195506a9 100644
--- a/Eigen/src/Core/EigenBase.h
+++ b/Eigen/src/Core/EigenBase.h
@@ -14,6 +14,7 @@
 namespace Eigen {
 
 /** \class EigenBase
+  * \ingroup Core_Module
   * 
   * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
   *

From ba5cab576a7615af12389ff159c6ed57b5312d5e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 14:38:04 +0200
Subject: [PATCH 129/680] bug #1405: enable StrictlyLower/StrictlyUpper
 triangularView as the destination of matrix*matrix products.

---
 .../Core/products/GeneralMatrixMatrixTriangular.h   | 13 +++++++++----
 .../products/GeneralMatrixMatrixTriangular_BLAS.h   |  2 +-
 test/product_mmtr.cpp                               | 11 +++++++++--
 3 files changed, 19 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index ad38bcf51..e436c50a4 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -269,10 +269,13 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
     enum {
       IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
       LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0,
-      RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0
+      RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0,
+      SkipDiag = (UpLo&(UnitDiag|ZeroDiag))!=0
     };
 
     Index size = mat.cols();
+    if(SkipDiag)
+      size--;
     Index depth = actualLhs.cols();
 
     typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,typename Lhs::Scalar,typename Rhs::Scalar,
@@ -283,10 +286,11 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
     internal::general_matrix_matrix_triangular_product<Index,
       typename Lhs::Scalar, LhsIsRowMajor ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
       typename Rhs::Scalar, RhsIsRowMajor ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
-      IsRowMajor ? RowMajor : ColMajor, UpLo>
+      IsRowMajor ? RowMajor : ColMajor, UpLo&(Lower|Upper)>
       ::run(size, depth,
-            &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &actualRhs.coeffRef(0,0), actualRhs.outerStride(),
-            mat.data(), mat.outerStride(), actualAlpha, blocking);
+            &actualLhs.coeffRef(SkipDiag&&(UpLo&Lower)==Lower ? 1 : 0,0), actualLhs.outerStride(),
+            &actualRhs.coeffRef(0,SkipDiag&&(UpLo&Upper)==Upper ? 1 : 0), actualRhs.outerStride(),
+            mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? 1 : mat.outerStride() ) : 0), mat.outerStride(), actualAlpha, blocking);
   }
 };
 
@@ -294,6 +298,7 @@ template<typename MatrixType, unsigned int UpLo>
 template<typename ProductType>
 EIGEN_DEVICE_FUNC TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
+  EIGEN_STATIC_ASSERT((UpLo&UnitDiag)==0, WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED);
   eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
 
   general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
index 5b7c15cca..41e18ff07 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
@@ -52,7 +52,7 @@ struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,Con
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \
                           const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \
   { \
-    if (lhs==rhs) { \
+    if ( lhs==rhs && ((UpLo&(Lower|Upper)==UpLo)) ) { \
       general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
       ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
     } else { \
diff --git a/test/product_mmtr.cpp b/test/product_mmtr.cpp
index f6e4bb1ae..d3e24b012 100644
--- a/test/product_mmtr.cpp
+++ b/test/product_mmtr.cpp
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010-2017 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -10,12 +10,19 @@
 #include "main.h"
 
 #define CHECK_MMTR(DEST, TRI, OP) {                   \
+    ref3 = DEST;                                      \
     ref2 = ref1 = DEST;                               \
     DEST.template triangularView<TRI>() OP;           \
     ref1 OP;                                          \
     ref2.template triangularView<TRI>()               \
       = ref1.template triangularView<TRI>();          \
     VERIFY_IS_APPROX(DEST,ref2);                      \
+    \
+    DEST = ref3;                                      \
+    ref3 = ref2;                                      \
+    ref3.diagonal() = DEST.diagonal();                \
+    DEST.template triangularView<TRI|ZeroDiag>() OP;  \
+    VERIFY_IS_APPROX(DEST,ref3);                      \
   }
 
 template<typename Scalar> void mmtr(int size)
@@ -27,7 +34,7 @@ template<typename Scalar> void mmtr(int size)
   
   MatrixColMaj matc = MatrixColMaj::Zero(size, size);
   MatrixRowMaj matr = MatrixRowMaj::Zero(size, size);
-  MatrixColMaj ref1(size, size), ref2(size, size);
+  MatrixColMaj ref1(size, size), ref2(size, size), ref3(size,size);
   
   MatrixColMaj soc(size,othersize); soc.setRandom();
   MatrixColMaj osc(othersize,size); osc.setRandom();

From 1bbcf19029ea3ac5493c3511807daa53fcde9e90 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 14:44:02 +0200
Subject: [PATCH 130/680] bug #1403: fix implicit scalar type conversion.

---
 Eigen/src/QR/ColPivHouseholderQR.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index 1b828f5dc..ae7c1775f 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -555,7 +555,7 @@ void ColPivHouseholderQR<MatrixType>::computeInPlace()
       if (m_colNormsUpdated.coeffRef(j) != 0) {
         RealScalar temp = abs(m_qr.coeffRef(k, j)) / m_colNormsUpdated.coeffRef(j);
         temp = (RealScalar(1) + temp) * (RealScalar(1) - temp);
-        temp = temp < 0 ? 0 : temp;
+        temp = temp < 0 ? RealScalar(0) : temp;
         RealScalar temp2 = temp * numext::abs2<RealScalar>(m_colNormsUpdated.coeffRef(j) /
                                                            m_colNormsDirect.coeffRef(j));
         if (temp2 <= norm_downdate_threshold) {

From 731c8c704df3c2a2f5872be712a12821e4acd869 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 15:45:49 +0200
Subject: [PATCH 131/680] bug #1403: more scalar conversions fixes in BDCSVD

---
 Eigen/src/QR/ColPivHouseholderQR.h     |  4 +-
 Eigen/src/SVD/BDCSVD.h                 | 81 +++++++++++++-------------
 Eigen/src/SVD/UpperBidiagonalization.h |  4 +-
 3 files changed, 46 insertions(+), 43 deletions(-)

diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index ae7c1775f..5270eaca2 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -552,10 +552,10 @@ void ColPivHouseholderQR<MatrixType>::computeInPlace()
       // http://www.netlib.org/lapack/lawnspdf/lawn176.pdf
       // and used in LAPACK routines xGEQPF and xGEQP3.
       // See lines 278-297 in http://www.netlib.org/lapack/explore-html/dc/df4/sgeqpf_8f_source.html
-      if (m_colNormsUpdated.coeffRef(j) != 0) {
+      if (m_colNormsUpdated.coeffRef(j) != RealScalar(0)) {
         RealScalar temp = abs(m_qr.coeffRef(k, j)) / m_colNormsUpdated.coeffRef(j);
         temp = (RealScalar(1) + temp) * (RealScalar(1) - temp);
-        temp = temp < 0 ? RealScalar(0) : temp;
+        temp = temp <  RealScalar(0) ? RealScalar(0) : temp;
         RealScalar temp2 = temp * numext::abs2<RealScalar>(m_colNormsUpdated.coeffRef(j) /
                                                            m_colNormsDirect.coeffRef(j));
         if (temp2 <= norm_downdate_threshold) {
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index 25fca6f4d..d7a4271cb 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -77,6 +77,7 @@ public:
   typedef _MatrixType MatrixType;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+  typedef typename NumTraits<RealScalar>::Literal Literal;
   enum {
     RowsAtCompileTime = MatrixType::RowsAtCompileTime, 
     ColsAtCompileTime = MatrixType::ColsAtCompileTime, 
@@ -259,7 +260,7 @@ BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsign
   
   //**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows
   RealScalar scale = matrix.cwiseAbs().maxCoeff();
-  if(scale==RealScalar(0)) scale = RealScalar(1);
+  if(scale==Literal(0)) scale = Literal(1);
   MatrixX copy;
   if (m_isTranspose) copy = matrix.adjoint()/scale;
   else               copy = matrix/scale;
@@ -351,13 +352,13 @@ void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, co
     Index k1=0, k2=0;
     for(Index j=0; j<n; ++j)
     {
-      if( (A.col(j).head(n1).array()!=0).any() )
+      if( (A.col(j).head(n1).array()!=Literal(0)).any() )
       {
         A1.col(k1) = A.col(j).head(n1);
         B1.row(k1) = B.row(j);
         ++k1;
       }
-      if( (A.col(j).tail(n2).array()!=0).any() )
+      if( (A.col(j).tail(n2).array()!=Literal(0)).any() )
       {
         A2.col(k2) = A.col(j).tail(n2);
         B2.row(k2) = B.row(j);
@@ -449,11 +450,11 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
     l = m_naiveU.row(1).segment(firstCol, k);
     f = m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1);
   }
-  if (m_compV) m_naiveV(firstRowW+k, firstColW) = 1;
+  if (m_compV) m_naiveV(firstRowW+k, firstColW) = Literal(1);
   if (r0<considerZero)
   {
-    c0 = 1;
-    s0 = 0;
+    c0 = Literal(1);
+    s0 = Literal(0);
   }
   else
   {
@@ -574,7 +575,7 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
   ArrayRef col0 = m_computed.col(firstCol).segment(firstCol, n);
   m_workspace.head(n) =  m_computed.block(firstCol, firstCol, n, n).diagonal();
   ArrayRef diag = m_workspace.head(n);
-  diag(0) = 0;
+  diag(0) = Literal(0);
 
   // Allocate space for singular values and vectors
   singVals.resize(n);
@@ -590,7 +591,7 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
   // but others are interleaved and we must ignore them at this stage.
   // To this end, let's compute a permutation skipping them:
   Index actual_n = n;
-  while(actual_n>1 && diag(actual_n-1)==0) --actual_n;
+  while(actual_n>1 && diag(actual_n-1)==Literal(0)) --actual_n;
   Index m = 0; // size of the deflated problem
   for(Index k=0;k<actual_n;++k)
     if(abs(col0(k))>considerZero)
@@ -691,7 +692,7 @@ template <typename MatrixType>
 typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar mu, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift)
 {
   Index m = perm.size();
-  RealScalar res = 1;
+  RealScalar res = Literal(1);
   for(Index i=0; i<m; ++i)
   {
     Index j = perm(i);
@@ -710,16 +711,16 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
 
   Index n = col0.size();
   Index actual_n = n;
-  while(actual_n>1 && col0(actual_n-1)==0) --actual_n;
+  while(actual_n>1 && col0(actual_n-1)==Literal(0)) --actual_n;
 
   for (Index k = 0; k < n; ++k)
   {
-    if (col0(k) == 0 || actual_n==1)
+    if (col0(k) == Literal(0) || actual_n==1)
     {
       // if col0(k) == 0, then entry is deflated, so singular value is on diagonal
       // if actual_n==1, then the deflated problem is already diagonalized
       singVals(k) = k==0 ? col0(0) : diag(k);
-      mus(k) = 0;
+      mus(k) = Literal(0);
       shifts(k) = k==0 ? col0(0) : diag(k);
       continue;
     } 
@@ -733,13 +734,13 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
     {
       // Skip deflated singular values
       Index l = k+1;
-      while(col0(l)==0) { ++l; eigen_internal_assert(l<actual_n); }
+      while(col0(l)==Literal(0)) { ++l; eigen_internal_assert(l<actual_n); }
       right = diag(l);
     }
 
     // first decide whether it's closer to the left end or the right end
-    RealScalar mid = left + (right-left) / 2;
-    RealScalar fMid = secularEq(mid, col0, diag, perm, diag, 0);
+    RealScalar mid = left + (right-left) / Literal(2);
+    RealScalar fMid = secularEq(mid, col0, diag, perm, diag, Literal(0));
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
     std::cout << right-left << "\n";
     std::cout << "fMid = " << fMid << " " << secularEq(mid-left, col0, diag, perm, diag-left, left) << " " << secularEq(mid-right, col0, diag, perm, diag-right, right)   << "\n";
@@ -755,7 +756,7 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
               << " "       << secularEq(0.8*(left+right), col0, diag, perm, diag, 0)
               << " "       << secularEq(0.9*(left+right), col0, diag, perm, diag, 0) << "\n";
 #endif
-    RealScalar shift = (k == actual_n-1 || fMid > 0) ? left : right;
+    RealScalar shift = (k == actual_n-1 || fMid > Literal(0)) ? left : right;
     
     // measure everything relative to shift
     Map<ArrayXr> diagShifted(m_workspace.data()+4*n, n);
@@ -785,13 +786,13 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
 
     // rational interpolation: fit a function of the form a / mu + b through the two previous
     // iterates and use its zero to compute the next iterate
-    bool useBisection = fPrev*fCur>0;
-    while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
+    bool useBisection = fPrev*fCur>Literal(0);
+    while (fCur!=Literal(0) && abs(muCur - muPrev) > Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
     {
       ++m_numIters;
 
       // Find a and b such that the function f(mu) = a / mu + b matches the current and previous samples.
-      RealScalar a = (fCur - fPrev) / (1/muCur - 1/muPrev);
+      RealScalar a = (fCur - fPrev) / (Literal(1)/muCur - Literal(1)/muPrev);
       RealScalar b = fCur - a / muCur;
       // And find mu such that f(mu)==0:
       RealScalar muZero = -a/b;
@@ -803,8 +804,8 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       fCur = fZero;
       
       
-      if (shift == left  && (muCur < 0 || muCur > right - left)) useBisection = true;
-      if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true;
+      if (shift == left  && (muCur < Literal(0) || muCur > right - left)) useBisection = true;
+      if (shift == right && (muCur < -(right - left) || muCur > Literal(0))) useBisection = true;
       if (abs(fCur)>abs(fPrev)) useBisection = true;
     }
 
@@ -841,13 +842,13 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
         std::cout << k << " : " <<  fLeft << " * " << fRight << " == " << fLeft * fRight << "  ;  " << left << " - " << right << " -> " <<  leftShifted << " " << rightShifted << "   shift=" << shift << "\n";
       }
 #endif
-      eigen_internal_assert(fLeft * fRight < 0);
+      eigen_internal_assert(fLeft * fRight < Literal(0));
       
-      while (rightShifted - leftShifted > 2 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(leftShifted), abs(rightShifted)))
+      while (rightShifted - leftShifted > Literal(2) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(leftShifted), abs(rightShifted)))
       {
-        RealScalar midShifted = (leftShifted + rightShifted) / 2;
+        RealScalar midShifted = (leftShifted + rightShifted) / Literal(2);
         fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift);
-        if (fLeft * fMid < 0)
+        if (fLeft * fMid < Literal(0))
         {
           rightShifted = midShifted;
         }
@@ -858,7 +859,7 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
         }
       }
 
-      muCur = (leftShifted + rightShifted) / 2;
+      muCur = (leftShifted + rightShifted) / Literal(2);
     }
       
     singVals[k] = shift + muCur;
@@ -892,8 +893,8 @@ void BDCSVD<MatrixType>::perturbCol0
   // The offset permits to skip deflated entries while computing zhat
   for (Index k = 0; k < n; ++k)
   {
-    if (col0(k) == 0) // deflated
-      zhat(k) = 0;
+    if (col0(k) == Literal(0)) // deflated
+      zhat(k) = Literal(0);
     else
     {
       // see equation (3.6)
@@ -918,7 +919,7 @@ void BDCSVD<MatrixType>::perturbCol0
       std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n";
 #endif
       RealScalar tmp = sqrt(prod);
-      zhat(k) = col0(k) > 0 ? tmp : -tmp;
+      zhat(k) = col0(k) > Literal(0) ? tmp : -tmp;
     }
   }
 }
@@ -934,7 +935,7 @@ void BDCSVD<MatrixType>::computeSingVecs
   
   for (Index k = 0; k < n; ++k)
   {
-    if (zhat(k) == 0)
+    if (zhat(k) == Literal(0))
     {
       U.col(k) = VectorType::Unit(n+1, k);
       if (m_compV) V.col(k) = VectorType::Unit(n, k);
@@ -947,7 +948,7 @@ void BDCSVD<MatrixType>::computeSingVecs
         Index i = perm(l);
         U(i,k) = zhat(i)/(((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
       }
-      U(n,k) = 0;      
+      U(n,k) = Literal(0);
       U.col(k).normalize();
     
       if (m_compV)
@@ -958,7 +959,7 @@ void BDCSVD<MatrixType>::computeSingVecs
           Index i = perm(l);
           V(i,k) = diag(i) * zhat(i) / (((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
         }
-        V(0,k) = -1;
+        V(0,k) = Literal(-1);
         V.col(k).normalize();
       }
     }
@@ -980,14 +981,14 @@ void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index
   RealScalar c = m_computed(start, start);
   RealScalar s = m_computed(start+i, start);
   RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
-  if (r == 0)
+  if (r == Literal(0))
   {
-    m_computed(start+i, start+i) = 0;
+    m_computed(start+i, start+i) = Literal(0);
     return;
   }
   m_computed(start,start) = r;  
-  m_computed(start+i, start) = 0;
-  m_computed(start+i, start+i) = 0;
+  m_computed(start+i, start) = Literal(0);
+  m_computed(start+i, start+i) = Literal(0);
   
   JacobiRotation<RealScalar> J(c/r,-s/r);
   if (m_compU)  m_naiveU.middleRows(firstCol, size+1).applyOnTheRight(firstCol, firstCol+i, J);
@@ -1020,7 +1021,7 @@ void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index fi
     << m_computed(firstColm + i+1, firstColm+i+1) << " "
     << m_computed(firstColm + i+2, firstColm+i+2) << "\n";
 #endif
-  if (r==0)
+  if (r==Literal(0))
   {
     m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j);
     return;
@@ -1029,7 +1030,7 @@ void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index fi
   s/=r;
   m_computed(firstColm + i, firstColm) = r;  
   m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i);
-  m_computed(firstColm + j, firstColm) = 0;
+  m_computed(firstColm + j, firstColm) = Literal(0);
 
   JacobiRotation<RealScalar> J(c,-s);
   if (m_compU)  m_naiveU.middleRows(firstColu, size+1).applyOnTheRight(firstColu + i, firstColu + j, J);
@@ -1053,7 +1054,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
   RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff();
   RealScalar epsilon_strict = numext::maxi<RealScalar>(considerZero,NumTraits<RealScalar>::epsilon() * maxDiag);
-  RealScalar epsilon_coarse = 8 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(col0.cwiseAbs().maxCoeff(), maxDiag);
+  RealScalar epsilon_coarse = Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(col0.cwiseAbs().maxCoeff(), maxDiag);
   
 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
   assert(m_naiveU.allFinite());
@@ -1081,7 +1082,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
       std::cout << "deflation 4.2, set z(" << i << ") to zero because " << abs(col0(i)) << " < " << epsilon_strict << "  (diag(" << i << ")=" << diag(i) << ")\n";
 #endif
-      col0(i) = 0;
+      col0(i) = Literal(0);
     }
 
   //condition 4.3
diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h
index 0b1460894..11ac847e1 100644
--- a/Eigen/src/SVD/UpperBidiagonalization.h
+++ b/Eigen/src/SVD/UpperBidiagonalization.h
@@ -159,6 +159,8 @@ void upperbidiagonalization_blocked_helper(MatrixType& A,
                                                       traits<MatrixType>::Flags & RowMajorBit> > Y)
 {
   typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef typename NumTraits<RealScalar>::Literal Literal;
   enum { StorageOrder = traits<MatrixType>::Flags & RowMajorBit };
   typedef InnerStride<int(StorageOrder) == int(ColMajor) ? 1 : Dynamic> ColInnerStride;
   typedef InnerStride<int(StorageOrder) == int(ColMajor) ? Dynamic : 1> RowInnerStride;
@@ -263,7 +265,7 @@ void upperbidiagonalization_blocked_helper(MatrixType& A,
     SubMatType A10( A.block(bs,0, brows-bs,bs) );
     SubMatType A01( A.block(0,bs, bs,bcols-bs) );
     Scalar tmp = A01(bs-1,0);
-    A01(bs-1,0) = 1;
+    A01(bs-1,0) = Literal(1);
     A11.noalias() -= A10 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint();
     A11.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A01;
     A01(bs-1,0) = tmp;

From c3e2afce0d2f5b4a3b9d44fc85e5103869299e47 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 16:25:18 +0200
Subject: [PATCH 132/680] Enable MSVC 2010 workaround from MSVC only

---
 Eigen/src/Core/util/Macros.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 14ec87da8..755646795 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -955,7 +955,7 @@ namespace Eigen {
                 const typename internal::plain_constant_type<EXPR,SCALAR>::type, const EXPR>
 
 // Workaround for MSVC 2010 (see ML thread "patch with compile for for MSVC 2010")
-#if EIGEN_COMP_MSVC_STRICT<=1600
+#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC_STRICT<=1600)
 #define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) typename internal::enable_if<true,X>::type
 #else
 #define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) X

From a4fd4233adbf4132f24bf989e3099022c5f3e207 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Jun 2017 23:02:02 +0200
Subject: [PATCH 133/680] Fix compilation with some compilers

---
 test/half_float.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/test/half_float.cpp b/test/half_float.cpp
index 7ec40f3ae..7b6208873 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -112,8 +112,8 @@ void test_numtraits()
   VERIFY_IS_EQUAL( std::numeric_limits<half>::signaling_NaN().x, half(std::numeric_limits<float>::signaling_NaN()).x );
   VERIFY( (std::numeric_limits<half>::min)() > half(0.f) );
   VERIFY( (std::numeric_limits<half>::denorm_min)() > half(0.f) );
-  VERIFY( (std::numeric_limits<half>::min)()/2 > half(0.f) );
-  VERIFY_IS_EQUAL( (std::numeric_limits<half>::denorm_min)()/2, half(0.f) );
+  VERIFY( (std::numeric_limits<half>::min)()/half(2) > half(0.f) );
+  VERIFY_IS_EQUAL( (std::numeric_limits<half>::denorm_min)()/half(2), half(0.f) );
 }
 
 void test_arithmetic()

From 50e09cca0f2312d89ced431b3a6cd7ff7c925d09 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 11 Jun 2017 15:30:36 +0200
Subject: [PATCH 134/680] fix tipo

---
 doc/AsciiQuickReference.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt
index 8409f8850..0ca54cef3 100644
--- a/doc/AsciiQuickReference.txt
+++ b/doc/AsciiQuickReference.txt
@@ -140,7 +140,7 @@ R.array().abs()                // abs(P)
 R.cwiseAbs2()                  // abs(P.^2)
 R.array().abs2()               // abs(P.^2)
 (R.array() < s).select(P,Q );  // (R < s ? P : Q)
-R = (Q.array()==0).select(P,A) // R(Q==0) = P(Q==0)
+R = (Q.array()==0).select(P,R) // R(Q==0) = P(Q==0)
 R = P.unaryExpr(ptr_fun(func)) // R = arrayfun(func, P)   // with: scalar func(const scalar &x);
 
 

From 6dcf966558ff9782f34c5df897c15eb37f1e95e3 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 12 Jun 2017 16:47:22 +0200
Subject: [PATCH 135/680] Avoid implicit scalar conversion with accuracy loss
 in pow(scalar,array)

---
 Eigen/src/Core/GlobalFunctions.h | 49 +++++++++++++++-----------------
 1 file changed, 23 insertions(+), 26 deletions(-)

diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h
index 12828a7c3..50406400b 100644
--- a/Eigen/src/Core/GlobalFunctions.h
+++ b/Eigen/src/Core/GlobalFunctions.h
@@ -103,17 +103,18 @@ namespace Eigen
   inline const CwiseBinaryOp<internal::scalar_pow_op<Derived::Scalar,ScalarExponent>,Derived,Constant<ScalarExponent> >
   pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent);
 #else
-  template<typename Derived,typename ScalarExponent>
-  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,ScalarExponent>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent),
-          const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,ScalarExponent,pow) >::type
-  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent) {
-    return x.derived().pow(exponent);
-  }
-
-  template<typename Derived>
-  inline const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename Derived::Scalar,pow)
-  pow(const Eigen::ArrayBase<Derived>& x, const typename Derived::Scalar& exponent) {
-    return x.derived().pow(exponent);
+  template <typename Derived,typename ScalarExponent>
+  EIGEN_DEVICE_FUNC inline
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(
+    const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename internal::promote_scalar_arg<typename Derived::Scalar
+                                                 EIGEN_COMMA ScalarExponent EIGEN_COMMA
+                                                 EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent)>::type,pow))
+  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent)
+  {
+    typedef typename internal::promote_scalar_arg<typename Derived::Scalar,ScalarExponent,
+                                                  EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent)>::type PromotedExponent;
+    return EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,PromotedExponent,pow)(x.derived(),
+           typename internal::plain_constant_type<Derived,PromotedExponent>::type(x.derived().rows(), x.derived().cols(), internal::scalar_constant_op<PromotedExponent>(exponent)));
   }
 #endif
 
@@ -156,21 +157,17 @@ namespace Eigen
   inline const CwiseBinaryOp<internal::scalar_pow_op<Scalar,Derived::Scalar>,Constant<Scalar>,Derived>
   pow(const Scalar& x,const Eigen::ArrayBase<Derived>& x);
 #else
-  template<typename Scalar, typename Derived>
-  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,Scalar>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar),
-          const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow) >::type
-  pow(const Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
-  {
-    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow)(
-            typename internal::plain_constant_type<Derived,Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
-  }
-
-  template<typename Derived>
-  inline const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)
-  pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
-  {
-    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)(
-      typename internal::plain_constant_type<Derived,typename Derived::Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
+  template <typename Scalar, typename Derived>
+  EIGEN_DEVICE_FUNC inline
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(
+    const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename internal::promote_scalar_arg<typename Derived::Scalar
+                                                 EIGEN_COMMA Scalar EIGEN_COMMA
+                                                 EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar)>::type,Derived,pow))
+  pow(const Scalar& x, const Eigen::ArrayBase<Derived>& exponents) {
+    typedef typename internal::promote_scalar_arg<typename Derived::Scalar,Scalar,
+                                                  EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar)>::type PromotedScalar;
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(PromotedScalar,Derived,pow)(
+           typename internal::plain_constant_type<Derived,PromotedScalar>::type(exponents.derived().rows(), exponents.derived().cols(), internal::scalar_constant_op<PromotedScalar>(x)), exponents.derived());
   }
 #endif
 

From c0e1d510fd7086a7e79987b7df0a789a4457c5dc Mon Sep 17 00:00:00 2001
From: Kyle Vedder <kyle.c.vedder@gmail.com>
Date: Wed, 4 Oct 2017 21:01:23 -0500
Subject: [PATCH 136/680] Add support for SuiteSparse's KLU routines

---
 cmake/FindKLU.cmake                           |  51 +++
 test/CMakeLists.txt                           |  16 +
 test/klu_support.cpp                          |  32 ++
 unsupported/Eigen/KLUSupport                  |  41 ++
 unsupported/Eigen/src/KLUSupport/KLUSupport.h | 364 ++++++++++++++++++
 5 files changed, 504 insertions(+)
 create mode 100644 cmake/FindKLU.cmake
 create mode 100644 test/klu_support.cpp
 create mode 100644 unsupported/Eigen/KLUSupport
 create mode 100644 unsupported/Eigen/src/KLUSupport/KLUSupport.h

diff --git a/cmake/FindKLU.cmake b/cmake/FindKLU.cmake
new file mode 100644
index 000000000..2783b63d2
--- /dev/null
+++ b/cmake/FindKLU.cmake
@@ -0,0 +1,51 @@
+# KLU lib usually requires linking to a blas library.
+# It is up to the user of this module to find a BLAS and link to it.
+
+if (KLU_INCLUDES AND KLU_LIBRARIES)
+  set(KLU_FIND_QUIETLY TRUE)
+endif (KLU_INCLUDES AND KLU_LIBRARIES)
+
+find_path(KLU_INCLUDES
+  NAMES
+  klu.h
+  PATHS
+  $ENV{KLUDIR}
+  ${INCLUDE_INSTALL_DIR}
+  PATH_SUFFIXES
+  suitesparse
+  ufsparse
+)
+
+if(KLU_LIBRARIES)
+
+  if(NOT KLU_LIBDIR)
+    get_filename_component(KLU_LIBDIR ${KLU_LIBRARIES} PATH)
+  endif(NOT KLU_LIBDIR)
+
+  find_library(COLAMD_LIBRARY colamd PATHS ${KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+  if(COLAMD_LIBRARY)
+    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${COLAMD_LIBRARY})
+  endif ()
+  
+  find_library(AMD_LIBRARY amd PATHS ${KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+  if(AMD_LIBRARY)
+    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${AMD_LIBRARY})
+  endif ()
+
+  find_library(SUITESPARSE_LIBRARY SuiteSparse PATHS ${KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+  if(SUITESPARSE_LIBRARY)
+    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${SUITESPARSE_LIBRARY})
+  endif ()
+
+  find_library(CHOLMOD_LIBRARY cholmod PATHS $ENV{KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+  if(CHOLMOD_LIBRARY)
+    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${CHOLMOD_LIBRARY})
+  endif()
+
+endif(KLU_LIBRARIES)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(KLU DEFAULT_MSG
+                                  KLU_INCLUDES KLU_LIBRARIES)
+
+mark_as_advanced(KLU_INCLUDES KLU_LIBRARIES AMD_LIBRARY COLAMD_LIBRARY CHOLMOD_LIBRARY SUITESPARSE_LIBRARY)
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index e73ab92b4..8bd086ce3 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -68,6 +68,17 @@ else()
   ei_add_property(EIGEN_MISSING_BACKENDS "UmfPack, ")
 endif()
 
+find_package(KLU)
+if(KLU_FOUND)
+  add_definitions("-DEIGEN_KLU_SUPPORT")
+  include_directories(${KLU_INCLUDES})
+  set(SPARSE_LIBS ${SPARSE_LIBS} ${KLU_LIBRARIES} ${EIGEN_BLAS_LIBRARIES})
+  set(KLU_ALL_LIBS ${KLU_LIBRARIES} ${EIGEN_BLAS_LIBRARIES})
+  ei_add_property(EIGEN_TESTED_BACKENDS "KLU, ")
+else()
+  ei_add_property(EIGEN_MISSING_BACKENDS "KLU, ")
+endif()
+
 find_package(SuperLU 4.0)
 if(SUPERLU_FOUND)
   add_definitions("-DEIGEN_SUPERLU_SUPPORT")
@@ -297,6 +308,11 @@ if(UMFPACK_FOUND)
   ei_add_test(umfpack_support "" "${UMFPACK_ALL_LIBS}")
 endif()
 
+if(KLU_FOUND OR SuiteSparse_FOUND)
+  message("ADDING KLU TEST")
+  ei_add_test(klu_support "" "${KLU_ALL_LIBS}")
+endif()
+
 if(SUPERLU_FOUND)
   ei_add_test(superlu_support "" "${SUPERLU_ALL_LIBS}")
 endif()
diff --git a/test/klu_support.cpp b/test/klu_support.cpp
new file mode 100644
index 000000000..8b1fdeb41
--- /dev/null
+++ b/test/klu_support.cpp
@@ -0,0 +1,32 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS
+#include "sparse_solver.h"
+
+#include <unsupported/Eigen/KLUSupport>
+
+template<typename T> void test_klu_support_T()
+{
+  KLU<SparseMatrix<T, ColMajor> > klu_colmajor;
+  KLU<SparseMatrix<T, RowMajor> > klu_rowmajor;
+  
+  check_sparse_square_solving(klu_colmajor);
+  check_sparse_square_solving(klu_rowmajor);
+  
+  //check_sparse_square_determinant(umfpack_colmajor);
+  //check_sparse_square_determinant(umfpack_rowmajor);
+}
+
+void test_klu_support()
+{
+  CALL_SUBTEST_1(test_klu_support_T<double>());
+  CALL_SUBTEST_2(test_klu_support_T<std::complex<double> >());
+}
+
diff --git a/unsupported/Eigen/KLUSupport b/unsupported/Eigen/KLUSupport
new file mode 100644
index 000000000..b23d90535
--- /dev/null
+++ b/unsupported/Eigen/KLUSupport
@@ -0,0 +1,41 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_KLUSUPPORT_MODULE_H
+#define EIGEN_KLUSUPPORT_MODULE_H
+
+#include <Eigen/SparseCore>
+
+#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+
+extern "C" {
+#include <btf.h>
+#include <klu.h>
+   }
+
+/** \ingroup Support_modules
+  * \defgroup KLUSupport_Module KLUSupport module
+  *
+  * This module provides an interface to the KLU library which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
+  * It provides the following factorization class:
+  * - class KLU: a sparse LU factorization, well-suited for circuit simulation.
+  *
+  * \code
+  * #include <Eigen/KLUSupport>
+  * \endcode
+  *
+  * In order to use this module, the klu and btf headers must be accessible from the include paths, and your binary must be linked to the klu library and its dependencies.
+  * The dependencies depend on how umfpack has been compiled.
+  * For a cmake based project, you can use our FindKLU.cmake module to help you in this task.
+  *
+  */
+
+#include "src/KLUSupport/KLUSupport.h"
+
+#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+
+#endif // EIGEN_KLUSUPPORT_MODULE_H
diff --git a/unsupported/Eigen/src/KLUSupport/KLUSupport.h b/unsupported/Eigen/src/KLUSupport/KLUSupport.h
new file mode 100644
index 000000000..d2781202e
--- /dev/null
+++ b/unsupported/Eigen/src/KLUSupport/KLUSupport.h
@@ -0,0 +1,364 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Kyle Macfarlan <kyle.macfarlan@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_KLUSUPPORT_H
+#define EIGEN_KLUSUPPORT_H
+
+namespace Eigen {
+
+/* TODO extract L, extract U, compute det, etc... */
+
+/** \ingroup KLUSupport_Module
+  * \brief A sparse LU factorization and solver based on KLU
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a LU factorization
+  * using the KLU library. The sparse matrix A must be squared and full rank.
+  * The vectors or matrices X and B can be either dense or sparse.
+  *
+  * \warning The input matrix A should be in a \b compressed and \b column-major form.
+  * Otherwise an expensive copy will be made. You can call the inexpensive makeCompressed() to get a compressed matrix.
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  */
+
+
+inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, double B [ ], klu_common *Common, double) {
+   return klu_solve(Symbolic, Numeric, ldim, nrhs, B, Common);
+}
+
+inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
+   return klu_z_solve(Symbolic, Numeric, ldim, nrhs, &numext::real_ref(B[0]), Common);
+}
+
+inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, double B[], klu_common *Common, double) {
+   return klu_tsolve(Symbolic, Numeric, ldim, nrhs, B, Common);
+}
+
+inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
+   return klu_z_tsolve(Symbolic, Numeric, ldim, nrhs, &numext::real_ref(B[0]), 0, Common);
+}
+
+inline klu_numeric* klu_factor(int Ap [ ], int Ai [ ], double Ax [ ], klu_symbolic *Symbolic, klu_common *Common, double) {
+   return klu_factor(Ap, Ai, Ax, Symbolic, Common);
+}
+
+inline klu_numeric* klu_factor(int Ap[], int Ai[], std::complex<double> Ax[], klu_symbolic *Symbolic, klu_common *Common, std::complex<double>) {
+   return klu_z_factor(Ap, Ai, &numext::real_ref(Ax[0]), Symbolic, Common);
+}
+
+
+template<typename _MatrixType>
+class KLU : public SparseSolverBase<KLU<_MatrixType> >
+{
+  protected:
+    typedef SparseSolverBase<KLU<_MatrixType> > Base;
+    using Base::m_isInitialized;
+  public:
+    using Base::_solve_impl;
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef Matrix<int, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
+    typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
+    typedef SparseMatrix<Scalar> LUMatrixType;
+    typedef SparseMatrix<Scalar,ColMajor,int> KLUMatrixType;
+    typedef Ref<const KLUMatrixType, StandardCompressedFormat> KLUMatrixRef;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+  public:
+
+    KLU()
+      : m_dummy(0,0), mp_matrix(m_dummy)
+    {
+      init();
+    }
+
+    template<typename InputMatrixType>
+    explicit KLU(const InputMatrixType& matrix)
+      : mp_matrix(matrix)
+    {
+      init();
+      compute(matrix);
+    }
+
+    ~KLU()
+    {
+      if(m_symbolic) klu_free_symbolic(&m_symbolic,&m_common);
+      if(m_numeric)  klu_free_numeric(&m_numeric,&m_common);
+    }
+
+    inline Index rows() const { return mp_matrix.rows(); }
+    inline Index cols() const { return mp_matrix.cols(); }
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+    inline const LUMatrixType& matrixL() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_l;
+    }
+
+    inline const LUMatrixType& matrixU() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_u;
+    }
+
+    inline const IntColVectorType& permutationP() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_p;
+    }
+
+    inline const IntRowVectorType& permutationQ() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_q;
+    }
+
+    /** Computes the sparse Cholesky decomposition of \a matrix
+     *  Note that the matrix should be column-major, and in compressed format for best performance.
+     *  \sa SparseMatrix::makeCompressed().
+     */
+    template<typename InputMatrixType>
+    void compute(const InputMatrixType& matrix)
+    {
+      if(m_symbolic) klu_free_symbolic(&m_symbolic, &m_common);
+      if(m_numeric)  klu_free_numeric(&m_numeric, &m_common);
+      grab(matrix.derived());
+      analyzePattern_impl();
+      factorize_impl();
+    }
+
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      *
+      * \sa factorize(), compute()
+      */
+    template<typename InputMatrixType>
+    void analyzePattern(const InputMatrixType& matrix)
+    {
+      if(m_symbolic) klu_free_symbolic(&m_symbolic, &m_common);
+      if(m_numeric)  klu_free_numeric(&m_numeric, &m_common);
+
+      grab(matrix.derived());
+
+      analyzePattern_impl();
+    }
+
+
+    /** Provides access to the control settings array used by KLU.
+      *
+      * See KLU documentation for details.
+      */
+    inline const klu_common& kluCommon() const
+    {
+      return m_common;
+    }
+
+    /** Provides access to the control settings array used by UmfPack.
+      *
+      * If this array contains NaN's, the default values are used.
+      *
+      * See KLU documentation for details.
+      */
+    inline klu_common& kluCommon()
+    {
+      return m_common;
+    }
+
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the pattern anylysis has been performed.
+      *
+      * \sa analyzePattern(), compute()
+      */
+    template<typename InputMatrixType>
+    void factorize(const InputMatrixType& matrix)
+    {
+      eigen_assert(m_analysisIsOk && "KLU: you must first call analyzePattern()");
+      if(m_numeric)
+        klu_free_numeric(&m_numeric,&m_common);
+
+      grab(matrix.derived());
+
+      factorize_impl();
+    }
+
+    /** \internal */
+    template<typename BDerived,typename XDerived>
+    bool _solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const;
+
+    Scalar determinant() const;
+
+    void extractData() const;
+
+  protected:
+
+    void init()
+    {
+      m_info                  = InvalidInput;
+      m_isInitialized         = false;
+      m_numeric               = 0;
+      m_symbolic              = 0;
+      m_extractedDataAreDirty = true;
+
+      klu_defaults(&m_common);
+    }
+
+    void analyzePattern_impl()
+    {
+      m_info = InvalidInput;
+      m_analysisIsOk = false;
+      m_factorizationIsOk = false;
+      m_symbolic = klu_analyze(internal::convert_index<int>(mp_matrix.rows()),
+                                     const_cast<StorageIndex*>(mp_matrix.outerIndexPtr()), const_cast<StorageIndex*>(mp_matrix.innerIndexPtr()),
+                                     &m_common);
+      if (m_symbolic) {
+         m_isInitialized = true;
+         m_info = Success;
+         m_analysisIsOk = true;
+         m_extractedDataAreDirty = true;
+      }
+    }
+
+    void factorize_impl()
+    {
+
+      m_numeric = klu_factor(const_cast<StorageIndex*>(mp_matrix.outerIndexPtr()), const_cast<StorageIndex*>(mp_matrix.innerIndexPtr()), const_cast<Scalar*>(mp_matrix.valuePtr()),
+                                    m_symbolic, &m_common, Scalar());
+                                         
+
+      m_info = m_numeric ? Success : NumericalIssue;
+      m_factorizationIsOk = m_numeric ? 1 : 0;
+      m_extractedDataAreDirty = true;
+    }
+
+    template<typename MatrixDerived>
+    void grab(const EigenBase<MatrixDerived> &A)
+    {
+      mp_matrix.~KLUMatrixRef();
+      ::new (&mp_matrix) KLUMatrixRef(A.derived());
+    }
+
+    void grab(const KLUMatrixRef &A)
+    {
+      if(&(A.derived()) != &mp_matrix)
+      {
+        mp_matrix.~KLUMatrixRef();
+        ::new (&mp_matrix) KLUMatrixRef(A);
+      }
+    }
+
+    // cached data to reduce reallocation, etc.
+    mutable LUMatrixType m_l;
+
+    mutable LUMatrixType m_u;
+    mutable IntColVectorType m_p;
+    mutable IntRowVectorType m_q;
+
+    KLUMatrixType m_dummy;
+    KLUMatrixRef mp_matrix;
+
+    klu_numeric* m_numeric;
+    klu_symbolic* m_symbolic;
+    klu_common m_common;
+    mutable ComputationInfo m_info;
+    int m_factorizationIsOk;
+    int m_analysisIsOk;
+    mutable bool m_extractedDataAreDirty;
+
+  private:
+    KLU(const KLU& ) { }
+};
+
+
+template<typename MatrixType>
+void KLU<MatrixType>::extractData() const
+{
+  if (m_extractedDataAreDirty)
+  {
+     eigen_assert(false && "KLU: extractData Not Yet Implemented");
+
+//     // get size of the data
+//     int lnz, unz, rows, cols, nz_udiag;
+//     umfpack_get_lunz(&lnz, &unz, &rows, &cols, &nz_udiag, m_numeric, Scalar());
+// 
+//     // allocate data
+//     m_l.resize(rows,(std::min)(rows,cols));
+//     m_l.resizeNonZeros(lnz);
+// 
+//     m_u.resize((std::min)(rows,cols),cols);
+//     m_u.resizeNonZeros(unz);
+// 
+//     m_p.resize(rows);
+//     m_q.resize(cols);
+// 
+//     // extract
+//     umfpack_get_numeric(m_l.outerIndexPtr(), m_l.innerIndexPtr(), m_l.valuePtr(),
+//                         m_u.outerIndexPtr(), m_u.innerIndexPtr(), m_u.valuePtr(),
+//                         m_p.data(), m_q.data(), 0, 0, 0, m_numeric);
+// 
+//     m_extractedDataAreDirty = false;
+  }
+}
+
+template<typename MatrixType>
+typename KLU<MatrixType>::Scalar KLU<MatrixType>::determinant() const
+{
+  eigen_assert(false && "KLU: extractData Not Yet Implemented");
+  return Scalar();
+}
+
+template<typename MatrixType>
+template<typename BDerived,typename XDerived>
+bool KLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const
+{
+  Index rhsCols = b.cols();
+  eigen_assert((BDerived::Flags&RowMajorBit)==0 && "KLU backend does not support non col-major rhs yet");
+  eigen_assert((XDerived::Flags&RowMajorBit)==0 && "KLU backend does not support non col-major result yet");
+  eigen_assert(b.derived().data() != x.derived().data() && " KLU does not support inplace solve");
+  eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()");
+
+  x = b;
+  int info = 0;
+  if (true/*(MatrixType::Flags&RowMajorBit) == 0*/)
+  {
+      info = klu_solve(m_symbolic, m_numeric, b.rows(), rhsCols, x.const_cast_derived().data(), const_cast<klu_common*>(&m_common), Scalar());
+  }
+  else
+  {
+     info = klu_tsolve(m_symbolic, m_numeric, b.rows(), rhsCols, x.const_cast_derived().data(), const_cast<klu_common*>(&m_common), Scalar());
+  }
+
+  m_info = info!=0 ? Success : NumericalIssue;
+  return true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_KLUSUPPORT_H

From a020d9b134ed804574845c0949c3b988ed56bcb5 Mon Sep 17 00:00:00 2001
From: Justin Carpentier <jcarpent@laas.fr>
Date: Tue, 17 Oct 2017 21:51:27 +0200
Subject: [PATCH 137/680] Use col method for column-major matrix

---
 Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
index facdaf890..f66c846ef 100644
--- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
+++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
@@ -168,7 +168,7 @@ class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
       {
         for(Index j=0; j<mat.outerSize(); ++j)
         {
-          RealScalar sum = mat.innerVector(j).squaredNorm();
+          RealScalar sum = mat.col(j).squaredNorm();
           if(sum>RealScalar(0))
             m_invdiag(j) = RealScalar(1)/sum;
           else

From 5fa79f96b8df069a5babb4e022fa43529062686f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 9 Nov 2017 10:58:38 +0100
Subject: [PATCH 138/680] Patch from Konstantin Arturov to enable MKL's direct
 call by default

---
 Eigen/src/Core/util/MKL_support.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/MKL_support.h b/Eigen/src/Core/util/MKL_support.h
index b7d6ecc76..014ba1e3e 100755
--- a/Eigen/src/Core/util/MKL_support.h
+++ b/Eigen/src/Core/util/MKL_support.h
@@ -55,7 +55,11 @@
 
 
 #if defined EIGEN_USE_MKL
-#   include <mkl.h> 
+#   ifndef MKL_DIRECT_CALL
+#       define MKL_DIRECT_CALL
+#       define MKL_DIRECT_CALL_JUST_SET
+#   endif
+#   include <mkl.h>
 /*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/
 #   ifndef INTEL_MKL_VERSION
 #       undef EIGEN_USE_MKL /* INTEL_MKL_VERSION is not even defined on older versions */
@@ -69,6 +73,9 @@
 #       undef   EIGEN_USE_MKL_VML
 #       undef   EIGEN_USE_LAPACKE_STRICT
 #       undef   EIGEN_USE_LAPACKE
+#       ifdef   MKL_DIRECT_CALL_JUST_SET
+#           undef MKL_DIRECT_CALL
+#       endif
 #   endif
 #endif
 

From f86bb89d39e6611806c3b31719509e661120f3a1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 9 Nov 2017 11:07:45 +0100
Subject: [PATCH 139/680] Add EIGEN_MKL_NO_DIRECT_CALL option

---
 Eigen/src/Core/util/MKL_support.h | 2 +-
 doc/UsingIntelMKL.dox             | 6 +++++-
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/MKL_support.h b/Eigen/src/Core/util/MKL_support.h
index 014ba1e3e..17963fad4 100755
--- a/Eigen/src/Core/util/MKL_support.h
+++ b/Eigen/src/Core/util/MKL_support.h
@@ -55,7 +55,7 @@
 
 
 #if defined EIGEN_USE_MKL
-#   ifndef MKL_DIRECT_CALL
+#   if (!defined MKL_DIRECT_CALL) && (!defined EIGEN_MKL_NO_DIRECT_CALL)
 #       define MKL_DIRECT_CALL
 #       define MKL_DIRECT_CALL_JUST_SET
 #   endif
diff --git a/doc/UsingIntelMKL.dox b/doc/UsingIntelMKL.dox
index 6de14afad..fc35c3cf0 100644
--- a/doc/UsingIntelMKL.dox
+++ b/doc/UsingIntelMKL.dox
@@ -63,7 +63,11 @@ In addition you can choose which parts will be substituted by defining one or mu
 <tr><td>\c EIGEN_USE_MKL_ALL </td><td>Defines \c EIGEN_USE_BLAS, \c EIGEN_USE_LAPACKE, and \c EIGEN_USE_MKL_VML </td></tr>
 </table>
 
-The options can be combined with \b MKL_DIRECT_CALL to enable MKL direct call feature. This may help to increase performance of some MKL BLAS (?GEMM, ?GEMV, ?TRSM, ?AXPY and ?DOT) and LAPACK (LU, Cholesky and QR) routines for very small matrices. To make it work properly, the macro \c EIGEN_USE_MKL must also be defined in the case none of the other \c EIGEN_USE_MKL_* macros has been defined.
+The \c EIGEN_USE_BLAS and \c EIGEN_USE_LAPACKE* macros can be combined with \c EIGEN_USE_MKL to explicitly tell Eigen that the underlying BLAS/Lapack implementation is Intel MKL.
+The main effect is to enable MKL direct call feature (\c MKL_DIRECT_CALL).
+This may help to increase performance of some MKL BLAS (?GEMM, ?GEMV, ?TRSM, ?AXPY and ?DOT) and LAPACK (LU, Cholesky and QR) routines for very small matrices.
+MKL direct call can be disabled by defining \c EIGEN_MKL_NO_DIRECT_CALL.
+
 
 Note that the BLAS and LAPACKE backends can be enabled for any F77 compatible BLAS and LAPACK libraries. See this \link TopicUsingBlasLapack page \endlink for the details.
 

From 0a1cc7394226c7439b586f5bac3e94cf287622f1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 10:25:41 +0100
Subject: [PATCH 140/680] bug #1484: restore deleted line for 128 bits long
 doubles, and improve dispatching logic.

---
 .../src/MatrixFunctions/MatrixExponential.h   | 41 +++++++++++--------
 1 file changed, 25 insertions(+), 16 deletions(-)

diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
index bb6d9e1fe..85ab3d97c 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
@@ -326,6 +326,7 @@ struct matrix_exp_computeUV<MatrixType, long double>
     } else if (l1norm < 1.125358383453143065081397882891878e+000L) {
       matrix_exp_pade13(arg, U, V);
     } else {
+      const long double maxnorm = 2.884233277829519311757165057717815L;
       frexp(l1norm / maxnorm, &squarings);
       if (squarings < 0) squarings = 0;
       MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<long double>(squarings));
@@ -342,6 +343,27 @@ struct matrix_exp_computeUV<MatrixType, long double>
   }
 };
 
+template<typename T> struct is_exp_known_type : false_type {};
+template<> struct is_exp_known_type<float> : true_type {};
+template<> struct is_exp_known_type<double> : true_type {};
+#if LDBL_MANT_DIG <= 112
+template<> struct is_exp_known_type<long double> : true_type {};
+#endif
+
+template <typename ArgType, typename ResultType>
+void matrix_exp_compute(const ArgType& arg, ResultType &result, true_type) // natively supported scalar type
+{
+  typedef typename ArgType::PlainObject MatrixType;
+  MatrixType U, V;
+  int squarings;
+  matrix_exp_computeUV<MatrixType>::run(arg, U, V, squarings); // Pade approximant is (U+V) / (-U+V)
+  MatrixType numer = U + V;
+  MatrixType denom = -U + V;
+  result = denom.partialPivLu().solve(numer);
+  for (int i=0; i<squarings; i++)
+    result *= result;   // undo scaling by repeated squaring
+}
+
 
 /* Computes the matrix exponential
  *
@@ -349,26 +371,13 @@ struct matrix_exp_computeUV<MatrixType, long double>
  * \param result variable in which result will be stored
  */
 template <typename ArgType, typename ResultType>
-void matrix_exp_compute(const ArgType& arg, ResultType &result)
+void matrix_exp_compute(const ArgType& arg, ResultType &result, false_type) // default
 {
   typedef typename ArgType::PlainObject MatrixType;
-#if LDBL_MANT_DIG > 112 // rarely happens
   typedef typename traits<MatrixType>::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef typename std::complex<RealScalar> ComplexScalar;
-  if (sizeof(RealScalar) > 14) {
-    result = arg.matrixFunction(internal::stem_function_exp<ComplexScalar>);
-    return;
-  }
-#endif
-  MatrixType U, V;
-  int squarings; 
-  matrix_exp_computeUV<MatrixType>::run(arg, U, V, squarings); // Pade approximant is (U+V) / (-U+V)
-  MatrixType numer = U + V;
-  MatrixType denom = -U + V;
-  result = denom.partialPivLu().solve(numer);
-  for (int i=0; i<squarings; i++)
-    result *= result;   // undo scaling by repeated squaring
+  result = arg.matrixFunction(internal::stem_function_exp<ComplexScalar>);
 }
 
 } // end namespace Eigen::internal
@@ -402,7 +411,7 @@ template<typename Derived> struct MatrixExponentialReturnValue
     inline void evalTo(ResultType& result) const
     {
       const typename internal::nested_eval<Derived, 10>::type tmp(m_src);
-      internal::matrix_exp_compute(tmp, result);
+      internal::matrix_exp_compute(tmp, result, internal::is_exp_known_type<typename Derived::Scalar>());
     }
 
     Index rows() const { return m_src.rows(); }

From 1b2dcf9a4742644e86a3da10d8b027c841ae9861 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 10:26:09 +0100
Subject: [PATCH 141/680] Check that Schur decomposition succeed.

---
 unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
index 3f7d77710..ef50c46a9 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
@@ -428,7 +428,8 @@ struct matrix_function_compute<MatrixType, 1>
     typedef internal::traits<MatrixType> Traits;
     
     // compute Schur decomposition of A
-    const ComplexSchur<MatrixType> schurOfA(A);  
+    const ComplexSchur<MatrixType> schurOfA(A);
+    eigen_assert(schurOfA.info()==Success);
     MatrixType T = schurOfA.matrixT();
     MatrixType U = schurOfA.matrixU();
 

From 6365f937d61ea0d1d7d7d21984f328a5965f2e92 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 13:58:52 +0100
Subject: [PATCH 142/680] KLU depends on BTF but not on libSuiteSparse nor
 Cholmod

---
 cmake/FindKLU.cmake | 15 ++++++---------
 1 file changed, 6 insertions(+), 9 deletions(-)

diff --git a/cmake/FindKLU.cmake b/cmake/FindKLU.cmake
index 2783b63d2..4a8f8e0b0 100644
--- a/cmake/FindKLU.cmake
+++ b/cmake/FindKLU.cmake
@@ -16,6 +16,8 @@ find_path(KLU_INCLUDES
   ufsparse
 )
 
+find_library(KLU_LIBRARIES klu PATHS $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+
 if(KLU_LIBRARIES)
 
   if(NOT KLU_LIBDIR)
@@ -32,14 +34,9 @@ if(KLU_LIBRARIES)
     set(KLU_LIBRARIES ${KLU_LIBRARIES} ${AMD_LIBRARY})
   endif ()
 
-  find_library(SUITESPARSE_LIBRARY SuiteSparse PATHS ${KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
-  if(SUITESPARSE_LIBRARY)
-    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${SUITESPARSE_LIBRARY})
-  endif ()
-
-  find_library(CHOLMOD_LIBRARY cholmod PATHS $ENV{KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
-  if(CHOLMOD_LIBRARY)
-    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${CHOLMOD_LIBRARY})
+  find_library(BTF_LIBRARY btf PATHS $ENV{KLU_LIBDIR} $ENV{KLUDIR} ${LIB_INSTALL_DIR})
+  if(BTF_LIBRARY)
+    set(KLU_LIBRARIES ${KLU_LIBRARIES} ${BTF_LIBRARY})
   endif()
 
 endif(KLU_LIBRARIES)
@@ -48,4 +45,4 @@ include(FindPackageHandleStandardArgs)
 find_package_handle_standard_args(KLU DEFAULT_MSG
                                   KLU_INCLUDES KLU_LIBRARIES)
 
-mark_as_advanced(KLU_INCLUDES KLU_LIBRARIES AMD_LIBRARY COLAMD_LIBRARY CHOLMOD_LIBRARY SUITESPARSE_LIBRARY)
+mark_as_advanced(KLU_INCLUDES KLU_LIBRARIES AMD_LIBRARY COLAMD_LIBRARY BTF_LIBRARY)

From b82cd93c01f60917c180d81d6321777f92e78464 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 14:09:01 +0100
Subject: [PATCH 143/680] KLU: truely disable unimplemented code, add proper
 static assertions in solve

---
 unsupported/Eigen/src/KLUSupport/KLUSupport.h | 84 +++++++++----------
 1 file changed, 39 insertions(+), 45 deletions(-)

diff --git a/unsupported/Eigen/src/KLUSupport/KLUSupport.h b/unsupported/Eigen/src/KLUSupport/KLUSupport.h
index d2781202e..a9e8633d9 100644
--- a/unsupported/Eigen/src/KLUSupport/KLUSupport.h
+++ b/unsupported/Eigen/src/KLUSupport/KLUSupport.h
@@ -27,24 +27,24 @@ namespace Eigen {
   *
   * \implsparsesolverconcept
   *
-  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  * \sa \ref TutorialSparseSolverConcept, class UmfPackLU, class SparseLU
   */
 
 
-inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, double B [ ], klu_common *Common, double) {
-   return klu_solve(Symbolic, Numeric, ldim, nrhs, B, Common);
+inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, Index ldim, Index nrhs, double B [ ], klu_common *Common, double) {
+   return klu_solve(Symbolic, Numeric, internal::convert_index<int>(ldim), internal::convert_index<int>(nrhs), B, Common);
 }
 
-inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
-   return klu_z_solve(Symbolic, Numeric, ldim, nrhs, &numext::real_ref(B[0]), Common);
+inline int klu_solve(klu_symbolic *Symbolic, klu_numeric *Numeric, Index ldim, Index nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
+   return klu_z_solve(Symbolic, Numeric, internal::convert_index<int>(ldim), internal::convert_index<int>(nrhs), &numext::real_ref(B[0]), Common);
 }
 
-inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, double B[], klu_common *Common, double) {
-   return klu_tsolve(Symbolic, Numeric, ldim, nrhs, B, Common);
+inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, Index ldim, Index nrhs, double B[], klu_common *Common, double) {
+   return klu_tsolve(Symbolic, Numeric, internal::convert_index<int>(ldim), internal::convert_index<int>(nrhs), B, Common);
 }
 
-inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, int ldim, int nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
-   return klu_z_tsolve(Symbolic, Numeric, ldim, nrhs, &numext::real_ref(B[0]), 0, Common);
+inline int klu_tsolve(klu_symbolic *Symbolic, klu_numeric *Numeric, Index ldim, Index nrhs, std::complex<double>B[], klu_common *Common, std::complex<double>) {
+   return klu_z_tsolve(Symbolic, Numeric, internal::convert_index<int>(ldim), internal::convert_index<int>(nrhs), &numext::real_ref(B[0]), 0, Common);
 }
 
 inline klu_numeric* klu_factor(int Ap [ ], int Ai [ ], double Ax [ ], klu_symbolic *Symbolic, klu_common *Common, double) {
@@ -114,7 +114,7 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
       eigen_assert(m_isInitialized && "Decomposition is not initialized.");
       return m_info;
     }
-
+#if 0 // not implemented yet
     inline const LUMatrixType& matrixL() const
     {
       if (m_extractedDataAreDirty) extractData();
@@ -138,7 +138,7 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
       if (m_extractedDataAreDirty) extractData();
       return m_q;
     }
-
+#endif
     /** Computes the sparse Cholesky decomposition of \a matrix
      *  Note that the matrix should be column-major, and in compressed format for best performance.
      *  \sa SparseMatrix::makeCompressed().
@@ -213,9 +213,11 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
     template<typename BDerived,typename XDerived>
     bool _solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const;
 
+#if 0 // not implemented yet
     Scalar determinant() const;
 
     void extractData() const;
+#endif
 
   protected:
 
@@ -275,11 +277,12 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
     }
 
     // cached data to reduce reallocation, etc.
+#if 0 // not implemented yet
     mutable LUMatrixType m_l;
-
     mutable LUMatrixType m_u;
     mutable IntColVectorType m_p;
     mutable IntRowVectorType m_q;
+#endif
 
     KLUMatrixType m_dummy;
     KLUMatrixRef mp_matrix;
@@ -296,7 +299,7 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
     KLU(const KLU& ) { }
 };
 
-
+#if 0 // not implemented yet
 template<typename MatrixType>
 void KLU<MatrixType>::extractData() const
 {
@@ -304,26 +307,26 @@ void KLU<MatrixType>::extractData() const
   {
      eigen_assert(false && "KLU: extractData Not Yet Implemented");
 
-//     // get size of the data
-//     int lnz, unz, rows, cols, nz_udiag;
-//     umfpack_get_lunz(&lnz, &unz, &rows, &cols, &nz_udiag, m_numeric, Scalar());
-// 
-//     // allocate data
-//     m_l.resize(rows,(std::min)(rows,cols));
-//     m_l.resizeNonZeros(lnz);
-// 
-//     m_u.resize((std::min)(rows,cols),cols);
-//     m_u.resizeNonZeros(unz);
-// 
-//     m_p.resize(rows);
-//     m_q.resize(cols);
-// 
-//     // extract
-//     umfpack_get_numeric(m_l.outerIndexPtr(), m_l.innerIndexPtr(), m_l.valuePtr(),
-//                         m_u.outerIndexPtr(), m_u.innerIndexPtr(), m_u.valuePtr(),
-//                         m_p.data(), m_q.data(), 0, 0, 0, m_numeric);
-// 
-//     m_extractedDataAreDirty = false;
+    // get size of the data
+    int lnz, unz, rows, cols, nz_udiag;
+    umfpack_get_lunz(&lnz, &unz, &rows, &cols, &nz_udiag, m_numeric, Scalar());
+
+    // allocate data
+    m_l.resize(rows,(std::min)(rows,cols));
+    m_l.resizeNonZeros(lnz);
+
+    m_u.resize((std::min)(rows,cols),cols);
+    m_u.resizeNonZeros(unz);
+
+    m_p.resize(rows);
+    m_q.resize(cols);
+
+    // extract
+    umfpack_get_numeric(m_l.outerIndexPtr(), m_l.innerIndexPtr(), m_l.valuePtr(),
+                        m_u.outerIndexPtr(), m_u.innerIndexPtr(), m_u.valuePtr(),
+                        m_p.data(), m_q.data(), 0, 0, 0, m_numeric);
+
+    m_extractedDataAreDirty = false;
   }
 }
 
@@ -333,27 +336,18 @@ typename KLU<MatrixType>::Scalar KLU<MatrixType>::determinant() const
   eigen_assert(false && "KLU: extractData Not Yet Implemented");
   return Scalar();
 }
+#endif
 
 template<typename MatrixType>
 template<typename BDerived,typename XDerived>
 bool KLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const
 {
   Index rhsCols = b.cols();
-  eigen_assert((BDerived::Flags&RowMajorBit)==0 && "KLU backend does not support non col-major rhs yet");
-  eigen_assert((XDerived::Flags&RowMajorBit)==0 && "KLU backend does not support non col-major result yet");
-  eigen_assert(b.derived().data() != x.derived().data() && " KLU does not support inplace solve");
+  EIGEN_STATIC_ASSERT((XDerived::Flags&RowMajorBit)==0, THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
   eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()");
 
   x = b;
-  int info = 0;
-  if (true/*(MatrixType::Flags&RowMajorBit) == 0*/)
-  {
-      info = klu_solve(m_symbolic, m_numeric, b.rows(), rhsCols, x.const_cast_derived().data(), const_cast<klu_common*>(&m_common), Scalar());
-  }
-  else
-  {
-     info = klu_tsolve(m_symbolic, m_numeric, b.rows(), rhsCols, x.const_cast_derived().data(), const_cast<klu_common*>(&m_common), Scalar());
-  }
+  int info = klu_solve(m_symbolic, m_numeric, b.rows(), rhsCols, x.const_cast_derived().data(), const_cast<klu_common*>(&m_common), Scalar());
 
   m_info = info!=0 ? Success : NumericalIssue;
   return true;

From 00bc67c374605c75d2cee4a0831885ac25e99b54 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 14:11:22 +0100
Subject: [PATCH 144/680] Move KLU support to official

---
 {unsupported/Eigen => Eigen}/KLUSupport                  | 0
 {unsupported/Eigen => Eigen}/src/KLUSupport/KLUSupport.h | 0
 2 files changed, 0 insertions(+), 0 deletions(-)
 rename {unsupported/Eigen => Eigen}/KLUSupport (100%)
 rename {unsupported/Eigen => Eigen}/src/KLUSupport/KLUSupport.h (100%)

diff --git a/unsupported/Eigen/KLUSupport b/Eigen/KLUSupport
similarity index 100%
rename from unsupported/Eigen/KLUSupport
rename to Eigen/KLUSupport
diff --git a/unsupported/Eigen/src/KLUSupport/KLUSupport.h b/Eigen/src/KLUSupport/KLUSupport.h
similarity index 100%
rename from unsupported/Eigen/src/KLUSupport/KLUSupport.h
rename to Eigen/src/KLUSupport/KLUSupport.h

From 103c0aa6ad83d641625d33cb9e1edade45dc6910 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 10 Nov 2017 14:13:29 +0100
Subject: [PATCH 145/680] Add KLU in the list of third-party sparse solvers

---
 doc/SparseLinearSystems.dox | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/doc/SparseLinearSystems.dox b/doc/SparseLinearSystems.dox
index fc33b93e7..38754e4af 100644
--- a/doc/SparseLinearSystems.dox
+++ b/doc/SparseLinearSystems.dox
@@ -70,6 +70,9 @@ They are summarized in the following tables:
 <tr><td>UmfPackLU</td><td>\link UmfPackSupport_Module UmfPackSupport \endlink</td><td>Direct LU factorization</td><td>Square</td><td>Fill-in reducing, Leverage fast dense algebra</td>
     <td>Requires the <a href="http://www.suitesparse.com">SuiteSparse</a> package, \b GPL </td>
     <td></td></tr>
+<tr><td>KLU</td><td>\link KLUSupport_Module KLUSupport \endlink</td><td>Direct LU factorization</td><td>Square</td><td>Fill-in reducing, suitted for circuit simulation</td>
+    <td>Requires the <a href="http://www.suitesparse.com">SuiteSparse</a> package, \b GPL </td>
+    <td></td></tr>
 <tr><td>SuperLU</td><td>\link SuperLUSupport_Module SuperLUSupport \endlink</td><td>Direct LU factorization</td><td>Square</td><td>Fill-in reducing, Leverage fast dense algebra</td>
     <td>Requires the <a href="http://crd-legacy.lbl.gov/~xiaoye/SuperLU/">SuperLU</a> library, (BSD-like)</td>
     <td></td></tr>

From 7cc503f9f57f6d3524115b222b62e584a04ff3ef Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 15 Nov 2017 21:33:37 +0100
Subject: [PATCH 146/680] bug #1485: fix linking issue of non template
 functions

---
 Eigen/src/PaStiXSupport/PaStiXSupport.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h
index d2ebfd7bb..160d8a523 100644
--- a/Eigen/src/PaStiXSupport/PaStiXSupport.h
+++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h
@@ -64,28 +64,28 @@ namespace internal
     typedef typename _MatrixType::StorageIndex StorageIndex;
   };
   
-  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, float *vals, int *perm, int * invp, float *x, int nbrhs, int *iparm, double *dparm)
+  inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, float *vals, int *perm, int * invp, float *x, int nbrhs, int *iparm, double *dparm)
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}
     s_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); 
   }
   
-  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, double *vals, int *perm, int * invp, double *x, int nbrhs, int *iparm, double *dparm)
+  inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, double *vals, int *perm, int * invp, double *x, int nbrhs, int *iparm, double *dparm)
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}
     d_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); 
   }
   
-  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<float> *vals, int *perm, int * invp, std::complex<float> *x, int nbrhs, int *iparm, double *dparm)
+  inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<float> *vals, int *perm, int * invp, std::complex<float> *x, int nbrhs, int *iparm, double *dparm)
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}
     c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<PASTIX_COMPLEX*>(vals), perm, invp, reinterpret_cast<PASTIX_COMPLEX*>(x), nbrhs, iparm, dparm); 
   }
   
-  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<double> *vals, int *perm, int * invp, std::complex<double> *x, int nbrhs, int *iparm, double *dparm)
+  inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<double> *vals, int *perm, int * invp, std::complex<double> *x, int nbrhs, int *iparm, double *dparm)
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}

From 7f42a93349c7ed175fda5a684f603dfdcf6724e8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 15 Nov 2017 20:45:09 +0000
Subject: [PATCH 147/680] Merged in
 alainvaucher/eigen/find-module-imported-target (pull request PR-324)

In the CMake find module, define the Eigen imported target as when installing with CMake

* In the CMake find module, define the Eigen imported target

* Add quotes to the imported location, in case there are spaces in the path.

Approved-by: Alain Vaucher <acvaucher@gmail.com>
---
 cmake/FindEigen3.cmake | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/cmake/FindEigen3.cmake b/cmake/FindEigen3.cmake
index 9e9697860..657440ba5 100644
--- a/cmake/FindEigen3.cmake
+++ b/cmake/FindEigen3.cmake
@@ -10,6 +10,10 @@
 #  EIGEN3_INCLUDE_DIR - the eigen include directory
 #  EIGEN3_VERSION - eigen version
 #
+# and the following imported target:
+#
+#  Eigen3::Eigen - The header-only Eigen library
+#
 # This module reads hints about search locations from 
 # the following enviroment variables:
 #
@@ -95,3 +99,8 @@ else (EIGEN3_INCLUDE_DIR)
 
 endif(EIGEN3_INCLUDE_DIR)
 
+if(EIGEN3_FOUND AND NOT TARGET Eigen3::Eigen)
+  add_library(Eigen3::Eigen INTERFACE IMPORTED)
+  set_target_properties(Eigen3::Eigen PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES "${EIGEN3_INCLUDE_DIR}")
+endif()

From 3f7fb5a6d63b0f0a278fa4d889f86d8001a5f45e Mon Sep 17 00:00:00 2001
From: Androbin42 <robin.richtsfeld@gmail.com>
Date: Mon, 12 Jun 2017 17:07:56 +0000
Subject: [PATCH 148/680] Make eigen_monitor_perf.sh more robust

---
 scripts/eigen_monitor_perf.sh | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/scripts/eigen_monitor_perf.sh b/scripts/eigen_monitor_perf.sh
index 39f8e7ecd..8f3425daf 100755
--- a/scripts/eigen_monitor_perf.sh
+++ b/scripts/eigen_monitor_perf.sh
@@ -12,9 +12,9 @@ export CXX_FLAGS="-mfma -w"
 ####
 
 BENCH_PATH=$EIGEN_SOURCE_PATH/bench/perf_monitoring/$PREFIX
-PREVPATH=`pwd`
-cd $EIGEN_SOURCE_PATH/bench/perf_monitoring && ./runall.sh "Haswell 2.6GHz, FMA, Apple's clang" $*
-cd $PREVPATH
+PREVPATH=$(pwd)
+cd $EIGEN_SOURCE_PATH/bench/perf_monitoring && ./runall.sh "Haswell 2.6GHz, FMA, Apple's clang" "$@"
+cd $PREVPATH || exit 1
 
 ALLFILES="$BENCH_PATH/*.png $BENCH_PATH/*.html $BENCH_PATH/index.html $BENCH_PATH/s1.js $BENCH_PATH/s2.js"
 

From 95ecb2b5d670f80133ffc3fb34215ab7e33288ac Mon Sep 17 00:00:00 2001
From: Androbin42 <robin.richtsfeld@gmail.com>
Date: Mon, 12 Jun 2017 17:11:06 +0000
Subject: [PATCH 149/680] Make buildtests.in more robust

---
 scripts/buildtests.in | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/buildtests.in b/scripts/buildtests.in
index 526d5b74b..ab9c18fb1 100755
--- a/scripts/buildtests.in
+++ b/scripts/buildtests.in
@@ -10,7 +10,7 @@ then
 fi
 
 TESTSLIST="@EIGEN_TESTS_LIST@"
-targets_to_make=`echo "$TESTSLIST" | egrep "$1" | xargs echo`
+targets_to_make=$(echo "$TESTSLIST" | grep -E "$1" | xargs echo)
 
 if [ -n "${EIGEN_MAKE_ARGS:+x}" ]
 then

From d9d7bd6d6215039f25d443c53a8f95cd760352e8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 12 Jun 2017 22:25:02 +0200
Subject: [PATCH 150/680] Fix 1x1 case in Solve expression with
 EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION==RowMajor

---
 Eigen/src/Core/Solve.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h
index 960a58597..a8daea511 100644
--- a/Eigen/src/Core/Solve.h
+++ b/Eigen/src/Core/Solve.h
@@ -34,12 +34,12 @@ template<typename Decomposition, typename RhsType,typename StorageKind> struct s
 template<typename Decomposition, typename RhsType>
 struct solve_traits<Decomposition,RhsType,Dense>
 {
-  typedef Matrix<typename RhsType::Scalar,
+  typedef typename make_proper_matrix_type<typename RhsType::Scalar,
                  Decomposition::ColsAtCompileTime,
                  RhsType::ColsAtCompileTime,
                  RhsType::PlainObject::Options,
                  Decomposition::MaxColsAtCompileTime,
-                 RhsType::MaxColsAtCompileTime> PlainObject;  
+                 RhsType::MaxColsAtCompileTime>::type PlainObject;
 };
 
 template<typename Decomposition, typename RhsType>

From e43d8fe9d778929051e740ef9a9143b6a2f26482 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 12 Jun 2017 22:26:26 +0200
Subject: [PATCH 151/680] Fix compilation of streaming nested Array, i.e., cout
 << Array<Array<>>

---
 Eigen/src/Core/NumTraits.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h
index aebc0c259..92a9ae1ea 100644
--- a/Eigen/src/Core/NumTraits.h
+++ b/Eigen/src/Core/NumTraits.h
@@ -215,6 +215,8 @@ struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
   static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); }
   EIGEN_DEVICE_FUNC
   static inline RealScalar dummy_precision() { return NumTraits<RealScalar>::dummy_precision(); }
+
+  static inline int digits10() { return NumTraits<Scalar>::digits10(); }
 };
 
 template<> struct NumTraits<std::string>

From 9fbdf020597cd198e3686ca786172aec6f009db6 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 12 Jun 2017 22:30:32 +0200
Subject: [PATCH 152/680] Enable Array(EigenBase<>) ctor for compatible scalar
 types only. This prevents nested arrays to look as being convertible from/to
 simple arrays.

---
 Eigen/src/Core/Array.h | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h
index 0d34269fd..e10020d4f 100644
--- a/Eigen/src/Core/Array.h
+++ b/Eigen/src/Core/Array.h
@@ -231,10 +231,16 @@ class Array
             : Base(other)
     { }
 
+  private:
+    struct PrivateType {};
+  public:
+
     /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other)
+    EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other,
+                              typename internal::enable_if<internal::is_convertible<typename OtherDerived::Scalar,Scalar>::value,
+                                                           PrivateType>::type = PrivateType())
       : Base(other.derived())
     { }
 

From 449936828c3d7f97b90a0cc3d25a1ec7eaec79ff Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 13 Jun 2017 12:54:57 -0700
Subject: [PATCH 153/680] Added missing __device__ qualifier

---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index b9a125b42..f4ae3c3c5 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -54,7 +54,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half*
   return __halves2half2(from[0], from[1]);
 }
 
-template<> EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
+template<> __device__ EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
   return __halves2half2(from[0], from[0]);
 }
 

From b240080e6443e4fb203ac1cba0ad5bd4fdad56b3 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 15 Jun 2017 10:16:30 +0200
Subject: [PATCH 154/680] bug #1436: fix compilation of Jacobi rotations with
 ARM NEON, some specializations of internal::conj_helper were missing.

---
 Eigen/Core                               |  1 +
 Eigen/src/Core/arch/AVX/Complex.h        | 36 ++----------------------
 Eigen/src/Core/arch/AltiVec/Complex.h    | 35 ++---------------------
 Eigen/src/Core/arch/Default/ConjHelper.h | 29 +++++++++++++++++++
 Eigen/src/Core/arch/NEON/Complex.h       |  4 +++
 Eigen/src/Core/arch/SSE/Complex.h        | 36 ++----------------------
 Eigen/src/Core/arch/ZVector/Complex.h    |  3 ++
 7 files changed, 43 insertions(+), 101 deletions(-)
 create mode 100644 Eigen/src/Core/arch/Default/ConjHelper.h

diff --git a/Eigen/Core b/Eigen/Core
index d18835613..a16942e19 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -371,6 +371,7 @@ using std::ptrdiff_t;
 #include "src/Core/MathFunctions.h"
 #include "src/Core/GenericPacketMath.h"
 #include "src/Core/MathFunctionsImpl.h"
+#include "src/Core/arch/Default/ConjHelper.h"
 
 #if defined EIGEN_VECTORIZE_AVX512
   #include "src/Core/arch/SSE/PacketMath.h"
diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h
index 99439c8aa..7fa61969d 100644
--- a/Eigen/src/Core/arch/AVX/Complex.h
+++ b/Eigen/src/Core/arch/AVX/Complex.h
@@ -204,23 +204,7 @@ template<> struct conj_helper<Packet4cf, Packet4cf, true,true>
   }
 };
 
-template<> struct conj_helper<Packet8f, Packet4cf, false,false>
-{
-  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const
-  { return Packet4cf(Eigen::internal::pmul(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet4cf, Packet8f, false,false>
-{
-  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const
-  { return Packet4cf(Eigen::internal::pmul(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet4cf,Packet8f)
 
 template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
 {
@@ -400,23 +384,7 @@ template<> struct conj_helper<Packet2cd, Packet2cd, true,true>
   }
 };
 
-template<> struct conj_helper<Packet4d, Packet2cd, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const
-  { return Packet2cd(Eigen::internal::pmul(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet2cd, Packet4d, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const
-  { return Packet2cd(Eigen::internal::pmul(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cd,Packet4d)
 
 template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
 {
diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h
index 67db2f8ee..3e665730c 100644
--- a/Eigen/src/Core/arch/AltiVec/Complex.h
+++ b/Eigen/src/Core/arch/AltiVec/Complex.h
@@ -224,23 +224,7 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
   }
 };
 
-template<> struct conj_helper<Packet4f, Packet2cf, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
-  { return Packet2cf(internal::pmul<Packet4f>(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet2cf, Packet4f, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
-  { return Packet2cf(internal::pmul<Packet4f>(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
 
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
@@ -416,23 +400,8 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
     return pconj(internal::pmul(a, b));
   }
 };
-template<> struct conj_helper<Packet2d, Packet1cd, false,false>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
-  { return padd(c, pmul(x,y)); }
 
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
-  { return Packet1cd(internal::pmul<Packet2d>(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet1cd, Packet2d, false,false>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
-  { return Packet1cd(internal::pmul<Packet2d>(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
 
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
diff --git a/Eigen/src/Core/arch/Default/ConjHelper.h b/Eigen/src/Core/arch/Default/ConjHelper.h
new file mode 100644
index 000000000..4cfe34e05
--- /dev/null
+++ b/Eigen/src/Core/arch/Default/ConjHelper.h
@@ -0,0 +1,29 @@
+
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARCH_CONJ_HELPER_H
+#define EIGEN_ARCH_CONJ_HELPER_H
+
+#define EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(PACKET_CPLX, PACKET_REAL)                                                          \
+  template<> struct conj_helper<PACKET_REAL, PACKET_CPLX, false,false> {                                          \
+    EIGEN_STRONG_INLINE PACKET_CPLX pmadd(const PACKET_REAL& x, const PACKET_CPLX& y, const PACKET_CPLX& c) const \
+    { return padd(c, pmul(x,y)); }                                                                                \
+    EIGEN_STRONG_INLINE PACKET_CPLX pmul(const PACKET_REAL& x, const PACKET_CPLX& y) const                        \
+    { return PACKET_CPLX(Eigen::internal::pmul<PACKET_REAL>(x, y.v)); }                                           \
+  };                                                                                                              \
+                                                                                                                  \
+  template<> struct conj_helper<PACKET_CPLX, PACKET_REAL, false,false> {                                          \
+    EIGEN_STRONG_INLINE PACKET_CPLX pmadd(const PACKET_CPLX& x, const PACKET_REAL& y, const PACKET_CPLX& c) const \
+    { return padd(c, pmul(x,y)); }                                                                                \
+    EIGEN_STRONG_INLINE PACKET_CPLX pmul(const PACKET_CPLX& x, const PACKET_REAL& y) const                        \
+    { return PACKET_CPLX(Eigen::internal::pmul<PACKET_REAL>(x.v, y)); }                                           \
+  };
+
+#endif // EIGEN_ARCH_CONJ_HELPER_H
diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h
index 57e9b431f..ef50ba303 100644
--- a/Eigen/src/Core/arch/NEON/Complex.h
+++ b/Eigen/src/Core/arch/NEON/Complex.h
@@ -265,6 +265,8 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
   }
 };
 
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
+
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
   // TODO optimize it for NEON
@@ -456,6 +458,8 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
   }
 };
 
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
+
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
   // TODO optimize it for NEON
diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h
index 5607fe0ab..23e717f28 100644
--- a/Eigen/src/Core/arch/SSE/Complex.h
+++ b/Eigen/src/Core/arch/SSE/Complex.h
@@ -229,23 +229,7 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
   }
 };
 
-template<> struct conj_helper<Packet4f, Packet2cf, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
-  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet2cf, Packet4f, false,false>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
-  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
 
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
@@ -430,23 +414,7 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
   }
 };
 
-template<> struct conj_helper<Packet2d, Packet1cd, false,false>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
-  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x, y.v)); }
-};
-
-template<> struct conj_helper<Packet1cd, Packet2d, false,false>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
-  { return padd(c, pmul(x,y)); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
-  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x.v, y)); }
-};
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
 
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index d39d2d105..1bfb73397 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -336,6 +336,9 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
   }
 };
 
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
+
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
   // TODO optimize it for AltiVec

From 157040d44f5ecea286fecb37cf376cfb42e1ff65 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 20 Jun 2017 16:58:03 +0200
Subject: [PATCH 155/680] Make sure CMAKE_Fortran_COMPILER is set before
 checking for Fortran functions

---
 cmake/FindBLAS.cmake | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cmake/FindBLAS.cmake b/cmake/FindBLAS.cmake
index 9f74b07fe..e3395bc10 100644
--- a/cmake/FindBLAS.cmake
+++ b/cmake/FindBLAS.cmake
@@ -152,7 +152,7 @@ set(_blas_ORIG_CMAKE_FIND_LIBRARY_SUFFIXES ${CMAKE_FIND_LIBRARY_SUFFIXES})
 
 # Check the language being used
 get_property( _LANGUAGES_ GLOBAL PROPERTY ENABLED_LANGUAGES )
-if( _LANGUAGES_ MATCHES Fortran )
+if( _LANGUAGES_ MATCHES Fortran AND CMAKE_Fortran_COMPILER)
   set( _CHECK_FORTRAN TRUE )
 elseif( (_LANGUAGES_ MATCHES C) OR (_LANGUAGES_ MATCHES CXX) )
   set( _CHECK_FORTRAN FALSE )

From b651ce0ffa1e5a7f0c6414689d0e1ebb4bd1621c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 26 Jun 2017 09:58:28 +0200
Subject: [PATCH 156/680] Fix a gcc7 warning: Wint-in-bool-context

---
 Eigen/src/Core/AssignEvaluator.h   |   4 +-
 Eigen/src/Core/util/StaticAssert.h | 116 ++++++++++++++---------------
 2 files changed, 60 insertions(+), 60 deletions(-)

diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h
index b0ec7b7ca..dbe435d86 100644
--- a/Eigen/src/Core/AssignEvaluator.h
+++ b/Eigen/src/Core/AssignEvaluator.h
@@ -39,7 +39,7 @@ public:
   enum {
     DstAlignment = DstEvaluator::Alignment,
     SrcAlignment = SrcEvaluator::Alignment,
-    DstHasDirectAccess = DstFlags & DirectAccessBit,
+    DstHasDirectAccess = (DstFlags & DirectAccessBit) == DirectAccessBit,
     JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment)
   };
 
@@ -83,7 +83,7 @@ private:
                        && int(OuterStride)!=Dynamic && int(OuterStride)%int(InnerPacketSize)==0
                        && (EIGEN_UNALIGNED_VECTORIZE  || int(JointAlignment)>=int(InnerRequiredAlignment)),
     MayLinearize = bool(StorageOrdersAgree) && (int(DstFlags) & int(SrcFlags) & LinearAccessBit),
-    MayLinearVectorize = bool(MightVectorize) && MayLinearize && DstHasDirectAccess
+    MayLinearVectorize = bool(MightVectorize) && bool(MayLinearize) && bool(DstHasDirectAccess)
                        && (EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)) || MaxSizeAtCompileTime == Dynamic),
       /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
          so it's only good for large enough sizes. */
diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h
index 983361a45..cb1678900 100644
--- a/Eigen/src/Core/util/StaticAssert.h
+++ b/Eigen/src/Core/util/StaticAssert.h
@@ -44,64 +44,64 @@
     struct static_assertion<true>
     {
       enum {
-        YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX,
-        YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES,
-        YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES,
-        THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE,
-        THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE,
-        THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE,
-        OUT_OF_RANGE_ACCESS,
-        YOU_MADE_A_PROGRAMMING_MISTAKE,
-        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT,
-        EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE,
-        YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR,
-        YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR,
-        UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC,
-        THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES,
-        FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED,
-        NUMERIC_TYPE_MUST_BE_REAL,
-        COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED,
-        WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED,
-        THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE,
-        INVALID_MATRIX_PRODUCT,
-        INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS,
-        INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION,
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY,
-        THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES,
-        THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES,
-        INVALID_MATRIX_TEMPLATE_PARAMETERS,
-        INVALID_MATRIXBASE_TEMPLATE_PARAMETERS,
-        BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER,
-        THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX,
-        THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE,
-        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES,
-        YOU_ALREADY_SPECIFIED_THIS_STRIDE,
-        INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION,
-        THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD,
-        PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1,
-        THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS,
-        YOU_CANNOT_MIX_ARRAYS_AND_MATRICES,
-        YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION,
-        THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY,
-        YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT,
-        THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS,
-        THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS,
-        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL,
-        THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES,
-        YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED,
-        YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED,
-        THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE,
-        THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH,
-        OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG,
-        IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY,
-        STORAGE_LAYOUT_DOES_NOT_MATCH,
-        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE,
-        THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS,
-        MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY,
-        THIS_TYPE_IS_NOT_SUPPORTED,
-        STORAGE_KIND_MUST_MATCH,
-        STORAGE_INDEX_MUST_MATCH,
-        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY
+        YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX=1,
+        YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES=1,
+        YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES=1,
+        THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE=1,
+        THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE=1,
+        THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE=1,
+        OUT_OF_RANGE_ACCESS=1,
+        YOU_MADE_A_PROGRAMMING_MISTAKE=1,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT=1,
+        EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE=1,
+        YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR=1,
+        YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR=1,
+        UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC=1,
+        THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES=1,
+        FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED=1,
+        NUMERIC_TYPE_MUST_BE_REAL=1,
+        COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED=1,
+        WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED=1,
+        THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE=1,
+        INVALID_MATRIX_PRODUCT=1,
+        INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS=1,
+        INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION=1,
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY=1,
+        THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES=1,
+        THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES=1,
+        INVALID_MATRIX_TEMPLATE_PARAMETERS=1,
+        INVALID_MATRIXBASE_TEMPLATE_PARAMETERS=1,
+        BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER=1,
+        THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX=1,
+        THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE=1,
+        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES=1,
+        YOU_ALREADY_SPECIFIED_THIS_STRIDE=1,
+        INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION=1,
+        THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD=1,
+        PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1=1,
+        THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS=1,
+        YOU_CANNOT_MIX_ARRAYS_AND_MATRICES=1,
+        YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION=1,
+        THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY=1,
+        YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT=1,
+        THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS=1,
+        THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS=1,
+        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL=1,
+        THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES=1,
+        YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED=1,
+        YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED=1,
+        THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE=1,
+        THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH=1,
+        OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG=1,
+        IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY=1,
+        STORAGE_LAYOUT_DOES_NOT_MATCH=1,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE=1,
+        THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS=1,
+        MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY=1,
+        THIS_TYPE_IS_NOT_SUPPORTED=1,
+        STORAGE_KIND_MUST_MATCH=1,
+        STORAGE_INDEX_MUST_MATCH=1,
+        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY=1
       };
     };
 

From 561f77707525644a10cbd971ad14bf47aa0e93e0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 27 Jun 2017 12:05:17 +0200
Subject: [PATCH 157/680] Fix a gcc7 warning about bool * bool in abs2 default
 implementation.

---
 Eigen/src/Core/MathFunctions.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 0be4a25da..f5635c076 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1059,6 +1059,9 @@ inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x)
   return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x);
 }
 
+EIGEN_DEVICE_FUNC
+inline bool abs2(bool x) { return x; }
+
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x)

From c010b173608550d515f0aafd95cf8488af546e1a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 27 Jun 2017 14:29:57 +0200
Subject: [PATCH 158/680] Fix warning

---
 test/integer_types.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/test/integer_types.cpp b/test/integer_types.cpp
index a21f73a81..25126315a 100644
--- a/test/integer_types.cpp
+++ b/test/integer_types.cpp
@@ -162,8 +162,8 @@ void test_integer_types()
   VERIFY_IS_EQUAL(internal::scalar_div_cost<int>::value, 8);
   VERIFY_IS_EQUAL(internal::scalar_div_cost<unsigned int>::value, 8);
   if(sizeof(long)>sizeof(int)) {
-    VERIFY(internal::scalar_div_cost<long>::value > internal::scalar_div_cost<int>::value);
-    VERIFY(internal::scalar_div_cost<unsigned long>::value > internal::scalar_div_cost<int>::value);
+    VERIFY(int(internal::scalar_div_cost<long>::value) > int(internal::scalar_div_cost<int>::value));
+    VERIFY(int(internal::scalar_div_cost<unsigned long>::value) > int(internal::scalar_div_cost<int>::value));
   }
 #endif
 }

From 53725c10b80dabd2a536f66e854c50f892496946 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 28 Jun 2017 17:55:23 +0000
Subject: [PATCH 159/680] Merged in mehdi_goli/opencl/DataDependancy (pull
 request PR-10)

DataDependancy

* Wrapping data type to the pointer class for sycl in non-terminal nodes; not having that breaks Tensorflow Conv2d code.

* Applying Ronnan's Comments.

* Applying benoit's comments
---
 .../Eigen/CXX11/src/Tensor/TensorAssign.h     |  3 ++-
 .../CXX11/src/Tensor/TensorBroadcasting.h     |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   |  3 ++-
 .../CXX11/src/Tensor/TensorConcatenation.h    |  4 +++-
 .../CXX11/src/Tensor/TensorContraction.h      |  4 +++-
 .../Eigen/CXX11/src/Tensor/TensorConversion.h |  3 ++-
 .../CXX11/src/Tensor/TensorConvolution.h      |  4 +++-
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   |  7 ++++--
 .../Eigen/CXX11/src/Tensor/TensorEvalTo.h     |  1 +
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  | 10 ++++----
 .../Eigen/CXX11/src/Tensor/TensorExpr.h       | 18 +++++++++++---
 .../Eigen/CXX11/src/Tensor/TensorFFT.h        |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |  4 +++-
 .../src/Tensor/TensorForwardDeclarations.h    | 16 +++++++++++++
 .../Eigen/CXX11/src/Tensor/TensorGenerator.h  |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorInflation.h  |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorLayoutSwap.h |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |  9 ++++---
 .../Eigen/CXX11/src/Tensor/TensorPadding.h    |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorReverse.h    |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorStriding.h   |  3 ++-
 .../Eigen/CXX11/src/Tensor/TensorSycl.h       | 24 +++++++++++++++----
 .../Eigen/CXX11/src/Tensor/TensorTraits.h     |  4 ++++
 .../CXX11/src/Tensor/TensorVolumePatch.h      |  3 ++-
 30 files changed, 118 insertions(+), 40 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
index 166be200c..027305586 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
@@ -34,6 +34,7 @@ struct traits<TensorAssignOp<LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const std::size_t NumDimensions = internal::traits<LhsXprType>::NumDimensions;
   static const int Layout = internal::traits<LhsXprType>::Layout;
+  typedef typename traits<LhsXprType>::PointerType PointerType;
 
   enum {
     Flags = 0
@@ -168,7 +169,7 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<RightArgType, Device>& right_impl() const { return m_rightImpl; }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_leftImpl.data(); }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_leftImpl.data(); }
 
  private:
   TensorEvaluator<LeftArgType, Device> m_leftImpl;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 23a74460e..b6c93aff9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -31,6 +31,7 @@ struct traits<TensorBroadcastingOp<Broadcast, XprType> > : public traits<XprType
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Broadcast, typename XprType>
@@ -372,7 +373,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index c46a778b5..21ffa2872 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -32,6 +32,7 @@ struct traits<TensorChippingOp<DimId, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions - 1;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<DenseIndex DimId, typename XprType>
@@ -264,7 +265,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
            TensorOpCost(0, 0, cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
     CoeffReturnType* result = const_cast<CoeffReturnType*>(m_impl.data());
     if (((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumDims) ||
          (static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) &&
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
index 2c7ba961c..363df876c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@@ -37,6 +37,8 @@ struct traits<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
   enum { Flags = 0 };
+  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 };
 
 template<typename Axis, typename LhsXprType, typename RhsXprType>
@@ -275,7 +277,7 @@ struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgTy
            TensorOpCost(0, 0, compute_cost);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<LeftArgType, Device>& left_impl() const { return m_leftImpl; }
   /// required by sycl in order to extract the accessor
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index bf4a476d9..d5ee0d036 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -104,6 +104,8 @@ struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >
   // From NumDims below.
   static const int NumDimensions = traits<RhsXprType>::NumDimensions + traits<RhsXprType>::NumDimensions - 2 * array_size<Dimensions>::value;
   static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+  typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 
   enum {
     Flags = 0
@@ -609,7 +611,7 @@ struct TensorContractionEvaluatorBase
     return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return m_result; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
 
 protected:
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void EnableXSMMIfPossible(const array<IndexPair<Index>, ContractDims>& eval_op_indices) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
index b29968b63..f2743ffb1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
@@ -32,6 +32,7 @@ struct traits<TensorConversionOp<TargetType, XprType> >
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
   enum { Flags = 0 };
+   typedef typename ::Eigen::internal::TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
 };
 
 template<typename TargetType, typename XprType>
@@ -244,7 +245,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
     }
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the sycl accessor
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 378f5cccb..543b4814b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -231,6 +231,8 @@ struct traits<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<InputXprType>::NumDimensions;
   static const int Layout = traits<InputXprType>::Layout;
+  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename InputXprType::Scalar, Scalar>::val,
+  typename traits<InputXprType>::PointerType, typename traits<KernelXprType>::PointerType>::type PointerType;
 
   enum {
     Flags = 0
@@ -465,7 +467,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
                                        PacketSize));
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
  private:
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index 2e6021b1e..da88bcb3b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -300,7 +300,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   /// used by sycl in order to build the sycl buffer
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const{return m_device;}
   /// used by sycl in order to build the sycl buffer
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const { return m_buf; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const { return m_buf; }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void preloadKernel() {
     // Don't make a local copy of the kernel unless we have to (i.e. it's an
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 7e4c129bb..3fd997278 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -30,6 +30,7 @@ struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
 };
 
 template<typename CustomUnaryFunc, typename XprType>
@@ -138,7 +139,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
@@ -184,6 +185,8 @@ struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 };
 
 template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
@@ -297,7 +300,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_result; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
index 82dd1e640..d0c027890 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -32,6 +32,7 @@ struct traits<TensorEvalToOp<XprType, MakePointer_> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename MakePointer_<Scalar>::Type PointerType;
 
   enum {
     Flags = 0
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index 8516b37b3..fcf330b10 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -283,7 +283,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
                         internal::unpacket_traits<PacketReturnType>::size);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC  typename Eigen::internal::traits<XprType>::PointerType  data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<ArgType, Device>& impl() const { return m_argImpl; }
@@ -353,7 +353,7 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
         TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<ArgType, Device> & impl() const { return m_argImpl; }
@@ -433,7 +433,7 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
            TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<LeftArgType, Device>& left_impl() const { return m_leftImpl; }
   /// required by sycl in order to extract the accessor
@@ -533,7 +533,7 @@ struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type,
            TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<Arg1Type, Device> & arg1Impl() const { return m_arg1Impl; }
@@ -625,7 +625,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>
         .cwiseMax(m_elseImpl.costPerCoeff(vectorized));
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<IfArgType, Device> & cond_impl() const { return m_condImpl; }
   /// required by sycl in order to extract the accessor
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
index 85dfc7a69..744cf7cc4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
@@ -38,7 +38,7 @@ struct traits<TensorCwiseNullaryOp<NullaryOp, XprType> >
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-
+  typedef typename XprTraits::PointerType PointerType;
   enum {
     Flags = 0
   };
@@ -89,6 +89,7 @@ struct traits<TensorCwiseUnaryOp<UnaryOp, XprType> >
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename ::Eigen::internal::TypeConversion<Scalar, typename XprTraits::PointerType>::type PointerType;
 };
 
 template<typename UnaryOp, typename XprType>
@@ -161,7 +162,11 @@ struct traits<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-
+  typedef typename ::Eigen::internal::TypeConversion<Scalar,
+                                                       typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                                       typename traits<LhsXprType>::PointerType,
+                                                       typename traits<RhsXprType>::PointerType>::type
+                                                       >::type PointerType;
   enum {
     Flags = 0
   };
@@ -238,7 +243,11 @@ struct traits<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprT
   typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-
+  typedef typename ::Eigen::internal::TypeConversion<Scalar,
+                                                       typename conditional<::Eigen::internal::Pointer_type_promotion<typename Arg2XprType::Scalar, Scalar>::val,
+                                                       typename traits<Arg2XprType>::PointerType,
+                                                       typename traits<Arg3XprType>::PointerType>::type
+                                                       >::type PointerType;
   enum {
     Flags = 0
   };
@@ -314,6 +323,9 @@ struct traits<TensorSelectOp<IfXprType, ThenXprType, ElseXprType> >
   typedef typename ElseXprType::Nested ElseNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename ThenXprType::Scalar, Scalar>::val,
+                                                     typename traits<ThenXprType>::PointerType,
+                                                     typename traits<ElseXprType>::PointerType>::type PointerType;
 };
 
 template<typename IfXprType, typename ThenXprType, typename ElseXprType>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
index f060191ab..10e0a8a6b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
@@ -71,6 +71,7 @@ struct traits<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir> > : public traits
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
 };
 
 template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
@@ -234,7 +235,7 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
         if (line_len > 1) {
           const RealScalar pi_over_len(EIGEN_PI / line_len);
           const ComplexScalar pos_j_base = ComplexScalar(
-	       std::cos(pi_over_len), std::sin(pi_over_len));
+              std::cos(pi_over_len), std::sin(pi_over_len));
           pos_j_base_powered[1] = pos_j_base;
           if (line_len > 2) {
             const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index abe85c860..c015ce196 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -38,6 +38,7 @@ struct traits<TensorForcedEvalOp<XprType> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 
   enum {
     Flags = 0
@@ -143,7 +144,8 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const { return m_buffer; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename Eigen::internal::traits<XprType>::PointerType  data() const { return m_buffer; }
 
   /// required by sycl in order to extract the sycl accessor
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 2e638992a..7a34f9419 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -22,6 +22,22 @@ template<typename T> struct MakePointer {
   typedef T* Type;
   typedef T& RefType;
 };
+
+namespace internal{
+template<typename A, typename B> struct Pointer_type_promotion {
+  static const bool val=false;
+};
+template<typename A> struct Pointer_type_promotion<A, A> {
+  static const bool val = true;
+};
+template<typename A, typename B> struct TypeConversion;
+#ifndef __SYCL_DEVICE_ONLY__
+template<typename A, typename B> struct TypeConversion{
+  typedef A* type;
+};
+#endif
+}
+
 #if defined(EIGEN_USE_SYCL)
 namespace TensorSycl {
 namespace internal{
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index 11ae21be9..fa269b8c6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -31,6 +31,7 @@ struct traits<TensorGeneratorOp<Generator, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Generator, typename XprType>
@@ -156,7 +157,7 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
                                   TensorOpCost::MulCost<Scalar>());
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType  data() const { return NULL; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_argImpl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index e1d5541bc..c56e3648a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -38,6 +38,7 @@ struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions + 1;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
@@ -423,7 +424,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
     return packetWithPossibleZero(index);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
index af6ecf5f4..6147fbdf1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
@@ -31,6 +31,7 @@ struct traits<TensorInflationOp<Strides, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Strides, typename XprType>
@@ -213,7 +214,7 @@ struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
                         compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
index cd0109ef4..4e384f9b9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
@@ -46,6 +46,7 @@ struct traits<TensorLayoutSwapOp<XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = (traits<XprType>::Layout == ColMajor) ? RowMajor : ColMajor;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename XprType>
@@ -159,7 +160,7 @@ struct TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
     return m_impl.costPerCoeff(vectorized);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return m_impl.data(); }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_impl.data(); }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 6ddd2ca18..329655817 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -31,6 +31,7 @@ struct traits<TensorReshapingOp<NewDimensions, XprType> > : public traits<XprTyp
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = array_size<NewDimensions>::value;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename NewDimensions, typename XprType>
@@ -146,7 +147,7 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
     return m_impl.costPerCoeff(vectorized);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return const_cast<Scalar*>(m_impl.data()); }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return const_cast<Scalar*>(m_impl.data()); }
 
   EIGEN_DEVICE_FUNC const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
@@ -214,6 +215,7 @@ struct traits<TensorSlicingOp<StartIndices, Sizes, XprType> > : public traits<Xp
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = array_size<StartIndices>::value;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename StartIndices, typename Sizes, typename XprType>
@@ -468,7 +470,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   }
 
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
     Scalar* result = m_impl.data();
     if (result) {
       Index offset = 0;
@@ -633,6 +635,7 @@ struct traits<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprTyp
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = array_size<StartIndices>::value;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
@@ -823,7 +826,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
     return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, NumDims);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
     return NULL;
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
index a8e255246..5956e513d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
@@ -31,6 +31,7 @@ struct traits<TensorPaddingOp<PaddingDimensions, XprType> > : public traits<XprT
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename PaddingDimensions, typename XprType>
@@ -198,7 +199,7 @@ struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device
     return cost;
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// used by sycl
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PaddingDimensions& padding() const { return m_padding; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 6e8e7885b..9e0a20abf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -31,6 +31,7 @@ struct traits<TensorPatchOp<PatchDim, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions + 1;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename PatchDim, typename XprType>
@@ -256,7 +257,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 7356334e1..69079805d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -44,6 +44,7 @@ namespace internal {
   typedef typename XprType::Nested Nested;
   static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 
   template <class T> struct MakePointer {
     // Intermediate typedef to workaround MSVC issue.
@@ -677,7 +678,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     }
   }
 
-  EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const { return m_result; }
+  EIGEN_DEVICE_FUNC typename MakePointer_<CoeffReturnType>::Type data() const { return m_result; }
 
 #if defined(EIGEN_USE_SYCL)
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
index e430b0826..14a50a029 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
@@ -31,6 +31,7 @@ struct traits<TensorReverseOp<ReverseDimensions,
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename ReverseDimensions, typename XprType>
@@ -222,7 +223,7 @@ struct TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>, Device
            TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<ArgType, Device> & impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 8501466ce..2a85ed840 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -24,6 +24,7 @@ struct traits<TensorScanOp<Op, XprType> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Op, typename XprType>
@@ -175,7 +176,7 @@ struct TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> {
     return internal::ploadt<PacketReturnType, LoadMode>(m_output + index);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const
   {
     return m_output;
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index edc9dd3f3..0697fd1ce 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -31,6 +31,7 @@ struct traits<TensorShufflingOp<Shuffle, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Shuffle, typename XprType>
@@ -185,7 +186,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   // required by sycl
   EIGEN_STRONG_INLINE const Shuffle& shufflePermutation() const {return m_shuffle;}
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
index 2237140e7..a7eea99b6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
@@ -31,6 +31,7 @@ struct traits<TensorStridingOp<Strides, XprType> > : public traits<XprType>
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<typename Strides, typename XprType>
@@ -222,7 +223,7 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
         TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
index 3d6270614..7b8bd2df7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
@@ -34,12 +34,26 @@ struct MakeLocalPointer {
 namespace Eigen {
   template<typename StrideDims, typename XprType> class TensorTupleReducerDeviceOp;
   template<typename StrideDims, typename ArgType> struct TensorEvaluator<const TensorTupleReducerDeviceOp<StrideDims, ArgType>, SyclKernelDevice>;
+namespace internal {
+
+#ifdef __SYCL_DEVICE_ONLY__
+template<typename A, typename B> struct TypeConversion {
+  template<typename T>
+  static typename MakeGlobalPointer<A>::Type get_address_space_pointer(typename MakeGlobalPointer<T>::Type p);
+  template<typename T>
+  static typename MakeLocalPointer<A>::Type get_address_space_pointer(typename MakeLocalPointer<T>::Type p);
+
+  template<typename T>
+  static A* get_address_space_pointer(T* p);
+  typedef decltype(get_address_space_pointer(B())) type;
+};
+
+#endif
+}
 namespace TensorSycl {
 namespace internal {
 
   template<typename CoeffReturnType, typename OP, typename OutputAccessor, typename InputAccessor, typename LocalAccessor> struct GenericKernelReducer;
-
-
 /// This struct is used for special expression nodes with no operations (for example assign and selectOP).
   struct NoOP;
 
@@ -51,10 +65,10 @@ template<typename T> struct GetType<false, T>{
 };
 
 template <bool Conds,  size_t X , size_t Y > struct ValueCondition {
-  static const size_t Res =X;
+  static constexpr size_t Res =X;
 };
-template<size_t X, size_t Y> struct ValueCondition<false, X , Y> {
-  static const size_t Res =Y;
+template<size_t X, size_t Y> struct ValueCondition<false, X, Y> {
+  static constexpr size_t Res =Y;
 };
 
 }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
index a1e944e59..006b37921 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
@@ -61,6 +61,7 @@ struct traits<Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
     typedef T& RefType;
 
   };
+  typedef typename MakePointer<Scalar>::Type PointerType;
 };
 
 
@@ -81,6 +82,7 @@ struct traits<TensorFixedSize<Scalar_, Dimensions, Options_, IndexType_> >
     typedef T& RefType;
 
   };
+  typedef typename MakePointer<Scalar>::Type PointerType;
 };
 
 
@@ -105,6 +107,7 @@ struct traits<TensorMap<PlainObjectType, Options_, MakePointer_> >
     typedef typename MakePointerT::RefType RefType;
 
   };
+  typedef typename MakePointer<Scalar>::Type PointerType;
 };
 
 template<typename PlainObjectType>
@@ -121,6 +124,7 @@ struct traits<TensorRef<PlainObjectType> >
     Options = BaseTraits::Options,
     Flags = BaseTraits::Flags
   };
+  typedef typename BaseTraits::PointerType PointerType;
 };
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index b0f970ae6..a47db0b39 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -33,6 +33,7 @@ struct traits<TensorVolumePatchOp<Planes, Rows, Cols, XprType> > : public traits
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions + 1;
   static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
 };
 
 template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
@@ -509,7 +510,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
     return TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 

From c92faf9d841deeecd7e87d9544fab949f7f59ed2 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 6 Jul 2017 05:05:57 +0000
Subject: [PATCH 160/680] Merged in mehdi_goli/upstr_benoit/HiperbolicOP (pull
 request PR-13)

Adding hyperbolic operations for sycl.

* Adding hyperbolic operations.

* Adding the hyperbolic operations for CPU as well.
---
 Eigen/src/Core/MathFunctions.h | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 0be4a25da..966a126c9 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1375,9 +1375,19 @@ T acos(const T &x) {
   return acos(x);
 }
 
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T acosh(const T &x) {
+  EIGEN_USING_STD_MATH(acosh);
+  return acosh(x);
+}
+
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   acos(float x) { return cl::sycl::acos(x); }
 EIGEN_ALWAYS_INLINE double  acos(double x) { return cl::sycl::acos(x); }
+EIGEN_ALWAYS_INLINE float   acosh(float x) { return cl::sycl::acosh(x); }
+EIGEN_ALWAYS_INLINE double  acosh(double x) { return cl::sycl::acosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #ifdef __CUDACC__
@@ -1395,9 +1405,18 @@ T asin(const T &x) {
   return asin(x);
 }
 
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T asinh(const T &x) {
+  EIGEN_USING_STD_MATH(asinh);
+  return asinh(x);
+}
+
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   asin(float x) { return cl::sycl::asin(x); }
 EIGEN_ALWAYS_INLINE double  asin(double x) { return cl::sycl::asin(x); }
+EIGEN_ALWAYS_INLINE float   asinh(float x) { return cl::sycl::asinh(x); }
+EIGEN_ALWAYS_INLINE double  asinh(double x) { return cl::sycl::asinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #ifdef __CUDACC__
@@ -1415,9 +1434,18 @@ T atan(const T &x) {
   return atan(x);
 }
 
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T atanh(const T &x) {
+  EIGEN_USING_STD_MATH(atanh);
+  return atanh(x);
+}
+
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   atan(float x) { return cl::sycl::atan(x); }
 EIGEN_ALWAYS_INLINE double  atan(double x) { return cl::sycl::atan(x); }
+EIGEN_ALWAYS_INLINE float   atanh(float x) { return cl::sycl::atanh(x); }
+EIGEN_ALWAYS_INLINE double  atanh(double x) { return cl::sycl::atanh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #ifdef __CUDACC__

From 62b4634ebe7cd7d391e91be812e5c18418db705a Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 6 Jul 2017 05:08:13 +0000
Subject: [PATCH 161/680] Merged in
 mehdi_goli/upstr_benoit/TensorSYCLImageVolumePatchFixed (pull request PR-14)

Applying Benoit's comment for Fixing ImageVolumePatch.

* Applying Benoit's comment for Fixing ImageVolumePatch. Fixing conflict on cmake file.

* Fixing dealocation of the memory in ImagePatch test for SYCL.

* Fixing the automerge issue.
---
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h | 60 +++++++++++++------
 .../src/Tensor/TensorSyclExprConstructor.h    |  8 +--
 .../CXX11/src/Tensor/TensorVolumePatch.h      | 49 +++++++++++----
 .../test/cxx11_tensor_image_patch_sycl.cpp    |  4 +-
 .../test/cxx11_tensor_volume_patch_sycl.cpp   |  4 +-
 5 files changed, 86 insertions(+), 39 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index c56e3648a..3c6a2e091 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -27,6 +27,7 @@ namespace Eigen {
   * patch_cols, and 1 for all the additional dimensions.
   */
 namespace internal {
+
 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
 struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType>
 {
@@ -71,12 +72,12 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                            PaddingType padding_type, Scalar padding_value)
-      : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+                                                           : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+                                                           m_padding_type(padding_type), m_padding_value(padding_value) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex row_strides, DenseIndex col_strides,
@@ -85,13 +86,31 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
                                                            Scalar padding_value)
-      : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
-        m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+                                                           : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+                                                           m_padding_left(padding_left), m_padding_right(padding_right),
+                                                           m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+
+#ifdef EIGEN_USE_SYCL // this is work around for sycl as Eigen could not use c++11 deligate constructor feature
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                         DenseIndex row_strides, DenseIndex col_strides,
+                                                         DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                         DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                         bool padding_explicit, DenseIndex padding_top, DenseIndex padding_bottom,
+                                                         DenseIndex padding_left, DenseIndex padding_right, PaddingType padding_type,
+                                                         Scalar padding_value)
+                                                         : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                         m_row_strides(row_strides), m_col_strides(col_strides),
+                                                         m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                         m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                         m_padding_explicit(padding_explicit), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+                                                         m_padding_left(padding_left), m_padding_right(padding_right),
+                                                         m_padding_type(padding_type), m_padding_value(padding_value) {}
+
+#endif
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_rows() const { return m_patch_rows; }
@@ -172,7 +191,11 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
     RawAccess = false
   };
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+  #ifdef __SYCL_DEVICE_ONLY__
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType op, const Device& device)
+  #else
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
+  #endif
       : m_impl(op.expression(), device)
 #ifdef EIGEN_USE_SYCL
       , m_op(op)
@@ -429,6 +452,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
 #ifdef EIGEN_USE_SYCL
+  // Required by SYCL in order to construct the expression tree on the device
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
 #endif
 
@@ -511,10 +535,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   Scalar m_paddingValue;
 
   TensorEvaluator<ArgType, Device> m_impl;
-
-#ifdef EIGEN_USE_SYCL
-  const XprType& m_op;
-#endif
+  #ifdef EIGEN_USE_SYCL
+  // Required for SYCL in order to construct the expression tree on the device
+   XprType m_op;
+   #endif
 };
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
index 24cc23f45..cbae4ea1d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -448,8 +448,8 @@ struct  ExprConstructor<CVQual TensorImagePatchOp<Rows, Cols, OrigXprType>, CVQu
   template <typename FuncDetector>\
   ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
   : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.m_patch_rows, funcD.m_patch_cols, funcD.m_row_strides, funcD.m_col_strides,\
-    funcD.m_in_row_strides, funcD.m_in_col_strides, funcD.m_row_inflate_strides, funcD.m_col_inflate_strides, \
-    funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, funcD.m_padding_value, funcD.m_padding_type, funcD.m_padding_explicit){}\
+    funcD.m_in_row_strides, funcD.m_in_col_strides, funcD.m_row_inflate_strides, funcD.m_col_inflate_strides, funcD.m_padding_explicit, \
+    funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, funcD.m_padding_type, funcD.m_padding_value){}\
 };
 
 SYCLTENSORIMAGEPATCHOPEXPR(const)
@@ -468,8 +468,8 @@ struct  ExprConstructor<CVQual TensorVolumePatchOp<Planes, Rows, Cols, OrigXprTy
   ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
   : xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.m_patch_planes, funcD.m_patch_rows, funcD.m_patch_cols, funcD.m_plane_strides, funcD.m_row_strides, funcD.m_col_strides,\
     funcD.m_in_plane_strides, funcD.m_in_row_strides, funcD.m_in_col_strides,funcD.m_plane_inflate_strides, funcD.m_row_inflate_strides, funcD.m_col_inflate_strides, \
-    funcD.m_padding_top_z, funcD.m_padding_bottom_z, funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, funcD.m_padding_value,\
-    funcD.m_padding_type, funcD.m_padding_explicit){\
+    funcD.m_padding_explicit, funcD.m_padding_top_z, funcD.m_padding_bottom_z, funcD.m_padding_top, funcD.m_padding_bottom, funcD.m_padding_left, funcD.m_padding_right, \
+    funcD.m_padding_type, funcD.m_padding_value ){\
     }\
 };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index a47db0b39..51c099591 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -22,6 +22,7 @@ namespace Eigen {
   * dimensions.
   */
 namespace internal {
+
 template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
 struct traits<TensorVolumePatchOp<Planes, Rows, Cols, XprType> > : public traits<XprType>
 {
@@ -34,6 +35,7 @@ struct traits<TensorVolumePatchOp<Planes, Rows, Cols, XprType> > : public traits
   static const int NumDimensions = XprTraits::NumDimensions + 1;
   static const int Layout = XprTraits::Layout;
   typedef typename XprTraits::PointerType PointerType;
+
 };
 
 template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
@@ -66,12 +68,12 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                             DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
                                                             DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
                                                             PaddingType padding_type, Scalar padding_value)
-      : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(false), m_padding_top_z(0), m_padding_bottom_z(0), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
-        m_padding_type(padding_type), m_padding_value(padding_value) {}
+                                                            : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                            m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+                                                            m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                            m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                            m_padding_explicit(false), m_padding_top_z(0), m_padding_bottom_z(0), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+                                                            m_padding_type(padding_type), m_padding_value(padding_value) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
                                                            DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
@@ -81,13 +83,31 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
                                                            DenseIndex padding_top, DenseIndex padding_bottom,
                                                            DenseIndex padding_left, DenseIndex padding_right,
                                                            Scalar padding_value)
-      : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
-        m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
-        m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
-        m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
-        m_padding_explicit(true), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
-        m_padding_left(padding_left), m_padding_right(padding_right),
-        m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+                                                           : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(true), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+                                                           m_padding_left(padding_left), m_padding_right(padding_right),
+                                                           m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+
+#ifdef EIGEN_USE_SYCL // this is work around for sycl as Eigen could not use c++11 deligate constructor feature
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                         DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
+                                                         DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                         DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                         bool padding_explicit, DenseIndex padding_top_z, DenseIndex padding_bottom_z,
+                                                         DenseIndex padding_top, DenseIndex padding_bottom, DenseIndex padding_left,
+                                                         DenseIndex padding_right, PaddingType padding_type, Scalar padding_value)
+                                                         : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                         m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+                                                         m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                         m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides),
+                                                         m_col_inflate_strides(col_inflate_strides), m_padding_explicit(padding_explicit), m_padding_top_z(padding_top_z),
+                                                         m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom), m_padding_left(padding_left),
+                                                         m_padding_right(padding_right), m_padding_type(padding_type), m_padding_value(padding_value) {}
+
+#endif
 
     EIGEN_DEVICE_FUNC
     DenseIndex patch_planes() const { return m_patch_planes; }
@@ -515,6 +535,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
 #ifdef EIGEN_USE_SYCL
+  // Required by SYCL in order to construct the expression on the device
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
 #endif
 
@@ -614,8 +635,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   TensorEvaluator<ArgType, Device> m_impl;
 
 #ifdef EIGEN_USE_SYCL
+// Required by SYCL in order to construct the expression on the device
   XprType m_op;
 #endif
+
 };
 
 
diff --git a/unsupported/test/cxx11_tensor_image_patch_sycl.cpp b/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
index e5ca4e388..eea18ec70 100644
--- a/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_image_patchOP_sycl
+#define EIGEN_TEST_FUNC cxx11_tensor_image_patch_sycl
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -1084,7 +1084,7 @@ test_patch_padding_same_sycl<DataType, int64_t>(sycl_device);
 test_patch_no_extra_dim_sycl<DataType, int64_t>(sycl_device);
 test_imagenet_patches_sycl<DataType, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_image_patchOP_sycl()
+void test_cxx11_tensor_image_patch_sycl()
 {
 for (const auto& device :Eigen::get_sycl_supported_devices()) {
   CALL_SUBTEST(sycl_tensor_image_patch_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp b/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
index ddc9e0d46..039715abc 100644
--- a/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_volume_patchOP_sycl
+#define EIGEN_TEST_FUNC cxx11_tensor_volume_patch_sycl
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -214,7 +214,7 @@ std::cout << "Running on " << s.template get_info<cl::sycl::info::device::name>(
 test_single_voxel_patch_sycl<DataType, int64_t>(sycl_device);
 test_entire_volume_patch_sycl<DataType, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_volume_patchOP_sycl()
+void test_cxx11_tensor_volume_patch_sycl()
 {
 for (const auto& device :Eigen::get_sycl_supported_devices()) {
   CALL_SUBTEST(sycl_tensor_volume_patch_test_per_device<float>(device));

From dc524ac7166a603de0dac9e0cfd5aa055dd117ef Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 6 Jul 2017 21:11:15 -0700
Subject: [PATCH 162/680] Fixed compilation warning

---
 .../Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h        | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 9dcc9dab7..1264a0270 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -24,9 +24,9 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
 
   NonBlockingThreadPoolTempl(int num_threads, bool allow_spinning,
                              Environment env = Environment())
-      : num_threads_(num_threads),
+      : env_(env),
+        num_threads_(num_threads),
         allow_spinning_(allow_spinning),
-        env_(env),
         threads_(num_threads),
         queues_(num_threads),
         coprimes_(num_threads),

From 6795512e5942b5fd1829f776fde6611a7405b5bf Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 6 Jul 2017 21:12:45 -0700
Subject: [PATCH 163/680] Improved the randomness of the tensor random
 generator

---
 .../Eigen/CXX11/src/Tensor/TensorRandom.h     | 68 ++++++++-----------
 1 file changed, 30 insertions(+), 38 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
index 1655a813e..f108c349f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
@@ -55,11 +55,11 @@ EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
 #endif
 }
 
-static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE unsigned PCG_XSH_RS_generator(uint64_t* state) {
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE unsigned PCG_XSH_RS_generator(uint64_t* state, uint64_t stream) {
   // TODO: Unify with the implementation in the non blocking thread pool.
   uint64_t current = *state;
   // Update the internal state
-  *state = current * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
+  *state = current * 6364136223846793005ULL + (stream << 1 | 1);
   // Generate the random output (using the PCG-XSH-RS scheme)
   return static_cast<unsigned>((current ^ (current >> 22)) >> (22 + (current >> 61)));
 }
@@ -73,17 +73,17 @@ static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE uint64_t PCG_XSH_RS_state(uint64_t
 
 
 template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-T RandomToTypeUniform(uint64_t* state) {
-  unsigned rnd = PCG_XSH_RS_generator(state);
+T RandomToTypeUniform(uint64_t* state, uint64_t stream) {
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
   return static_cast<T>(rnd);
 }
 
 
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state) {
+Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state, uint64_t stream) {
   Eigen::half result;
   // Generate 10 random bits for the mantissa
-  unsigned rnd = PCG_XSH_RS_generator(state);
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
   result.x = static_cast<uint16_t>(rnd & 0x3ffu);
   // Set the exponent
   result.x |= (static_cast<uint16_t>(15) << 10);
@@ -93,14 +93,14 @@ Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state) {
 
 
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-float RandomToTypeUniform<float>(uint64_t* state) {
+float RandomToTypeUniform<float>(uint64_t* state, uint64_t stream) {
   typedef union {
     uint32_t raw;
     float fp;
   } internal;
   internal result;
   // Generate 23 random bits for the mantissa mantissa
-  const unsigned rnd = PCG_XSH_RS_generator(state);
+  const unsigned rnd = PCG_XSH_RS_generator(state, stream);
   result.raw = rnd & 0x7fffffu;
   // Set the exponent
   result.raw |= (static_cast<uint32_t>(127) << 23);
@@ -109,7 +109,7 @@ float RandomToTypeUniform<float>(uint64_t* state) {
 }
 
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-double RandomToTypeUniform<double>(uint64_t* state) {
+double RandomToTypeUniform<double>(uint64_t* state, uint64_t stream) {
   typedef union {
     uint64_t raw;
     double dp;
@@ -118,9 +118,9 @@ double RandomToTypeUniform<double>(uint64_t* state) {
   result.raw = 0;
   // Generate 52 random bits for the mantissa
   // First generate the upper 20 bits
-  unsigned rnd1 = PCG_XSH_RS_generator(state) & 0xfffffu;
+  unsigned rnd1 = PCG_XSH_RS_generator(state, stream) & 0xfffffu;
   // The generate the lower 32 bits
-  unsigned rnd2 = PCG_XSH_RS_generator(state);
+  unsigned rnd2 = PCG_XSH_RS_generator(state, stream);
   result.raw = (static_cast<uint64_t>(rnd1) << 32) | rnd2;
   // Set the exponent
   result.raw |= (static_cast<uint64_t>(1023) << 52);
@@ -129,14 +129,14 @@ double RandomToTypeUniform<double>(uint64_t* state) {
 }
 
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-std::complex<float> RandomToTypeUniform<std::complex<float> >(uint64_t* state) {
-  return std::complex<float>(RandomToTypeUniform<float>(state),
-                             RandomToTypeUniform<float>(state));
+std::complex<float> RandomToTypeUniform<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeUniform<float>(state, stream),
+                             RandomToTypeUniform<float>(state, stream));
 }
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-std::complex<double> RandomToTypeUniform<std::complex<double> >(uint64_t* state) {
-  return std::complex<double>(RandomToTypeUniform<double>(state),
-                              RandomToTypeUniform<double>(state));
+std::complex<double> RandomToTypeUniform<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeUniform<double>(state, stream),
+                              RandomToTypeUniform<double>(state, stream));
 }
 
 template <typename T> class UniformRandomGenerator {
@@ -155,9 +155,7 @@ template <typename T> class UniformRandomGenerator {
 
   template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   T operator()(Index i) const {
-    uint64_t local_state = m_state + i;
-    T result = RandomToTypeUniform<T>(&local_state);
-    m_state = local_state;
+    T result = RandomToTypeUniform<T>(&m_state, i);
     return result;
   }
 
@@ -165,11 +163,9 @@ template <typename T> class UniformRandomGenerator {
   Packet packetOp(Index i) const {
     const int packetSize = internal::unpacket_traits<Packet>::size;
     EIGEN_ALIGN_MAX T values[packetSize];
-    uint64_t local_state = m_state + i;
     for (int j = 0; j < packetSize; ++j) {
-      values[j] = RandomToTypeUniform<T>(&local_state);
+      values[j] = RandomToTypeUniform<T>(&m_state, i);
     }
-    m_state = local_state;
     return internal::pload<Packet>(values);
   }
 
@@ -190,14 +186,14 @@ struct functor_traits<UniformRandomGenerator<Scalar> > {
 
 
 template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-T RandomToTypeNormal(uint64_t* state) {
+T RandomToTypeNormal(uint64_t* state, uint64_t stream) {
   // Use the ratio of uniform method to generate numbers following a normal
   // distribution. See for example Numerical Recipes chapter 7.3.9 for the
   // details.
   T u, v, q;
   do {
-    u = RandomToTypeUniform<T>(state);
-    v = T(1.7156) * (RandomToTypeUniform<T>(state) - T(0.5));
+    u = RandomToTypeUniform<T>(state, stream);
+    v = T(1.7156) * (RandomToTypeUniform<T>(state, stream) - T(0.5));
     const T x = u - T(0.449871);
     const T y = numext::abs(v) + T(0.386595);
     q = x*x + y * (T(0.196)*y - T(0.25472)*x);
@@ -208,14 +204,14 @@ T RandomToTypeNormal(uint64_t* state) {
 }
 
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-std::complex<float> RandomToTypeNormal<std::complex<float> >(uint64_t* state) {
-  return std::complex<float>(RandomToTypeNormal<float>(state),
-                             RandomToTypeNormal<float>(state));
+std::complex<float> RandomToTypeNormal<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeNormal<float>(state, stream),
+                             RandomToTypeNormal<float>(state, stream));
 }
 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-std::complex<double> RandomToTypeNormal<std::complex<double> >(uint64_t* state) {
-  return std::complex<double>(RandomToTypeNormal<double>(state),
-                              RandomToTypeNormal<double>(state));
+std::complex<double> RandomToTypeNormal<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeNormal<double>(state, stream),
+                              RandomToTypeNormal<double>(state, stream));
 }
 
 
@@ -234,9 +230,7 @@ template <typename T> class NormalRandomGenerator {
 
  template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   T operator()(Index i) const {
-    uint64_t local_state = m_state + i;
-    T result = RandomToTypeNormal<T>(&local_state);
-    m_state = local_state;
+    T result = RandomToTypeNormal<T>(&m_state, i);
     return result;
   }
 
@@ -244,11 +238,9 @@ template <typename T> class NormalRandomGenerator {
   Packet packetOp(Index i) const {
     const int packetSize = internal::unpacket_traits<Packet>::size;
     EIGEN_ALIGN_MAX T values[packetSize];
-    uint64_t local_state = m_state + i;
     for (int j = 0; j < packetSize; ++j) {
-      values[j] = RandomToTypeNormal<T>(&local_state);
+      values[j] = RandomToTypeNormal<T>(&m_state, i);
     }
-    m_state = local_state;
     return internal::pload<Packet>(values);
   }
 

From 9daed6795224ef93719db66b71098bb7ac1a30ec Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 7 Jul 2017 04:18:03 +0000
Subject: [PATCH 164/680] Merged in tntnatbry/eigen (pull request PR-319)

Tensor Trace op
---
 unsupported/Eigen/CXX11/Tensor                |   1 +
 unsupported/Eigen/CXX11/src/Tensor/README.md  |  52 ++++
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  12 +
 .../src/Tensor/TensorForwardDeclarations.h    |   1 +
 .../Eigen/CXX11/src/Tensor/TensorTrace.h      | 288 ++++++++++++++++++
 unsupported/test/CMakeLists.txt               |   1 +
 unsupported/test/cxx11_tensor_trace.cpp       | 171 +++++++++++
 7 files changed, 526 insertions(+)
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
 create mode 100644 unsupported/test/cxx11_tensor_trace.cpp

diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 5d71a9c25..d243fe035 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -141,6 +141,7 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorGenerator.h"
 #include "src/Tensor/TensorAssign.h"
 #include "src/Tensor/TensorScan.h"
+#include "src/Tensor/TensorTrace.h"
 
 #include "src/Tensor/TensorSycl.h"
 #include "src/Tensor/TensorExecutor.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index 8d989f13a..4423f81f7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -1168,6 +1168,58 @@ Reduce a tensor using a user-defined reduction operator.  See ```SumReducer```
 in TensorFunctors.h for information on how to implement a reduction operator.
 
 
+## Trace
+
+A *Trace* operation returns a tensor with fewer dimensions than the original
+tensor. It returns a tensor whose elements are the sum of the elements of the
+original tensor along the main diagonal for a list of specified dimensions, the
+"trace dimensions". Similar to the ```Reduction Dimensions```, the trace dimensions
+are passed as an input parameter to the operation, are of type ```<TensorType>::Dimensions```
+, and have the same requirements when passed as an input parameter. In addition,
+the trace dimensions must have the same size.
+
+Example: Trace along 2 dimensions.
+
+    // Create a tensor of 3 dimensions
+    Eigen::Tensor<int, 3> a(2, 2, 3);
+    a.setValues({{{1, 2, 3}, {4, 5, 6}}, {{7, 8, 9}, {10, 11, 12}}});
+    // Specify the dimensions along which the trace will be computed.
+    // In this example, the trace can only be computed along the dimensions
+    // with indices 0 and 1
+    Eigen::array<int, 2> dims({0, 1});
+    // The output tensor contains all but the trace dimensions.
+    Tensor<int, 1> a_trace = a.trace(dims);
+    cout << "a_trace:" << endl;
+    cout << a_trace << endl;
+    =>
+    a_trace:
+    11
+    13
+    15
+
+
+### <Operation> trace(const Dimensions& new_dims)
+### <Operation> trace()
+
+As a special case, if no parameter is passed to the operation, trace is computed
+along *all* dimensions of the input tensor.
+
+Example: Trace along all dimensions.
+
+    // Create a tensor of 3 dimensions, with all dimensions having the same size.
+    Eigen::Tensor<int, 3> a(3, 3, 3);
+    a.setValues({{{1, 2, 3}, {4, 5, 6}, {7, 8, 9}},
+                {{10, 11, 12}, {13, 14, 15}, {16, 17, 18}},
+                {{19, 20, 21}, {22, 23, 24}, {25, 26, 27}}});
+    // Result is a zero dimension tensor
+    Tensor<int, 0> a_trace = a.trace();
+    cout<<"a_trace:"<<endl;
+    cout<<a_trace<<endl;
+    =>
+    a_trace:
+    42
+
+
 ## Scan Operations
 
 A *Scan* operation returns a tensor with the same dimensions as the original
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 5b1235826..0d6331e9c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -671,6 +671,18 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return TensorReductionOp<Reducer, const Dims, const Derived>(derived(), dims, reducer);
     }
 
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTraceOp<const Dims, const Derived>
+    trace(const Dims& dims) const {
+      return TensorTraceOp<const Dims, const Derived>(derived(), dims);
+    }
+
+    const TensorTraceOp<const DimensionList<Index, NumDimensions>, const Derived>
+    trace() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorTraceOp<const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims);
+    }
+
     template <typename Broadcast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorBroadcastingOp<const Broadcast, const Derived>
     broadcast(const Broadcast& broadcast) const {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 2e638992a..2950e6963 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -70,6 +70,7 @@ template<typename Strides, typename XprType> class TensorInflationOp;
 template<typename Generator, typename XprType> class TensorGeneratorOp;
 template<typename LeftXprType, typename RightXprType> class TensorAssignOp;
 template<typename Op, typename XprType> class TensorScanOp;
+template<typename Dims, typename XprType> class TensorTraceOp;
 
 template<typename CustomUnaryFunc, typename XprType> class TensorCustomUnaryOp;
 template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType> class TensorCustomBinaryOp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
new file mode 100644
index 000000000..4b165399f
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@@ -0,0 +1,288 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gagan Goel <gagan.nith@gmail.com>
+// Copyright (C) 2017 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
+
+namespace Eigen {
+
+/** \class TensorTrace
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor Trace class.
+  *
+  *
+  */
+
+namespace internal {
+template<typename Dims, typename XprType>
+struct traits<TensorTraceOp<Dims, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
+  static const int Layout = XprTraits::Layout;
+};
+
+template<typename Dims, typename XprType>
+struct eval<TensorTraceOp<Dims, XprType>, Eigen::Dense>
+{
+  typedef const TensorTraceOp<Dims, XprType>& type;
+};
+
+template<typename Dims, typename XprType>
+struct nested<TensorTraceOp<Dims, XprType>, 1, typename eval<TensorTraceOp<Dims, XprType> >::type>
+{
+  typedef TensorTraceOp<Dims, XprType> type;
+};
+
+} // end namespace internal
+
+
+template<typename Dims, typename XprType>
+class TensorTraceOp : public TensorBase<TensorTraceOp<Dims, XprType> >
+{
+  public:
+    typedef typename Eigen::internal::traits<TensorTraceOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorTraceOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorTraceOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorTraceOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTraceOp(const XprType& expr, const Dims& dims)
+      : m_xpr(expr), m_dims(dims) {
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Dims& dims() const { return m_dims; }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename XprType::Nested>::type& expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Dims m_dims;
+};
+
+
+// Eval as rvalue
+template<typename Dims, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
+{
+  typedef TensorTraceOp<Dims, ArgType> XprType;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumReducedDims = internal::array_size<Dims>::value;
+  static const int NumOutputDims = NumInputDims - NumReducedDims;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumOutputDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : m_impl(op.expression(), device), m_device(device)
+  {
+
+    EIGEN_STATIC_ASSERT((NumOutputDims >= 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((NumReducedDims >= 2) || ((NumReducedDims == 0) && (NumInputDims == 0)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    for (int i = 0; i < NumInputDims; ++i) {
+      m_reduced[i] = false;
+    }
+    
+    const Dims& op_dims = op.dims();
+    for (int i = 0; i < NumReducedDims; ++i) {
+      eigen_assert(op_dims[i] >= 0);
+      eigen_assert(op_dims[i] < NumInputDims);
+      m_reduced[op_dims[i]] = true;
+    }
+
+    // All the dimensions should be distinct to compute the trace
+    int num_distinct_reduce_dims = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (m_reduced[i]) {
+        ++num_distinct_reduce_dims;
+      }
+    }
+
+    eigen_assert(num_distinct_reduce_dims == NumReducedDims);
+
+    // Compute the dimensions of the result. 
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+
+    int output_index = 0;
+    int reduced_index = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (m_reduced[i]) {
+        m_reducedDims[reduced_index] = input_dims[i];
+        if (reduced_index > 0) {
+          // All the trace dimensions must have the same size
+          eigen_assert(m_reducedDims[0] == m_reducedDims[reduced_index]);
+        }
+        ++reduced_index;
+      }
+      else {
+        m_dimensions[output_index] = input_dims[i];
+        ++output_index;
+      }
+    }
+
+    if (NumReducedDims != 0) {
+      m_traceDim = m_reducedDims[0];
+    }
+
+    // Compute the output strides
+    if (NumOutputDims > 0) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_outputStrides[0] = 1;
+        for (int i = 1; i < NumOutputDims; ++i) {
+          m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+        }
+      }
+      else {
+        m_outputStrides.back() = 1;
+        for (int i = NumOutputDims - 2; i >= 0; --i) {
+          m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+        }
+      }
+    }
+
+    // Compute the input strides
+    if (NumInputDims > 0) {
+      array<Index, NumInputDims> input_strides;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        input_strides[0] = 1;
+        for (int i = 1; i < NumInputDims; ++i) {
+          input_strides[i] = input_strides[i - 1] * input_dims[i - 1];
+        }
+      }
+      else {
+        input_strides.back() = 1;
+        for (int i = NumInputDims - 2; i >= 0; --i) {
+          input_strides[i] = input_strides[i + 1] * input_dims[i + 1];
+        }
+      }
+
+      output_index = 0;
+      reduced_index = 0;
+      for (int i = 0; i < NumInputDims; ++i) {
+        if(m_reduced[i]) {
+          m_reducedStrides[reduced_index] = input_strides[i];
+          ++reduced_index;
+        }
+        else {
+          m_preservedStrides[output_index] = input_strides[i];
+          ++output_index;
+        }
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_dimensions;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Initialize the result
+    CoeffReturnType result = internal::cast<int, CoeffReturnType>(0);
+    Index index_stride = 0;
+    for (int i = 0; i < NumReducedDims; ++i) {
+      index_stride += m_reducedStrides[i];
+    }
+
+    // If trace is requested along all dimensions, starting index would be 0
+    Index cur_index = 0;
+    if (NumOutputDims != 0)
+      cur_index = firstInput(index);
+    for (Index i = 0; i < m_traceDim; ++i) {
+        result += m_impl.coeff(cur_index);
+        cur_index += index_stride;
+    }
+      
+    return result;
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const {
+
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    for (int i = 0; i < PacketSize; ++i) {
+        values[i] = coeff(index + i);
+    }
+    PacketReturnType result = internal::ploadt<PacketReturnType, LoadMode>(values);
+    return result;
+  }
+
+ protected:
+  // Given the output index, finds the first index in the input tensor used to compute the trace
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
+    Index startInput = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumOutputDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      startInput += index * m_preservedStrides[0];
+    }
+    else {
+      for (int i = 0; i < NumOutputDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      startInput += index * m_preservedStrides[NumOutputDims - 1];
+    }
+    return startInput;
+  }
+
+  Dimensions m_dimensions;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Device& m_device;
+  array<bool, NumInputDims> m_reduced;
+  array<Index, NumReducedDims> m_reducedDims;
+  // Initialize the size of the trace dimension
+  Index m_traceDim = 1;
+  array<Index, NumOutputDims> m_outputStrides;
+  array<Index, NumReducedDims> m_reducedStrides;
+  array<Index, NumOutputDims> m_preservedStrides;
+};
+
+
+} // End namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index e639e7056..22647cadd 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -227,6 +227,7 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_fft)
   ei_add_test(cxx11_tensor_ifft)
   ei_add_test(cxx11_tensor_scan)
+  ei_add_test(cxx11_tensor_trace)
 
 endif()
 
diff --git a/unsupported/test/cxx11_tensor_trace.cpp b/unsupported/test/cxx11_tensor_trace.cpp
new file mode 100644
index 000000000..340d1211c
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_trace.cpp
@@ -0,0 +1,171 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gagan Goel <gagan.nith@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+using Eigen::array;
+
+template <int DataLayout>
+static void test_0D_trace() {
+  Tensor<float, 0, DataLayout> tensor;
+  tensor.setRandom();
+  array<ptrdiff_t, 0> dims;
+  Tensor<float, 0, DataLayout> result = tensor.trace(dims);
+  VERIFY_IS_EQUAL(result(), tensor());
+}
+
+
+template <int DataLayout>
+static void test_all_dimensions_trace() {
+  Tensor<float, 3, DataLayout> tensor1(5, 5, 5);
+  tensor1.setRandom();
+  Tensor<float, 0, DataLayout> result1 = tensor1.trace();
+  VERIFY_IS_EQUAL(result1.rank(), 0);
+  float sum = 0.0f;
+  for (int i = 0; i < 5; ++i) {
+    sum += tensor1(i, i, i);
+  }
+  VERIFY_IS_EQUAL(result1(), sum);
+
+  Tensor<float, 5, DataLayout> tensor2(7, 7, 7, 7, 7);
+  array<ptrdiff_t, 5> dims({{2, 1, 0, 3, 4}});
+  Tensor<float, 0, DataLayout> result2 = tensor2.trace(dims);
+  VERIFY_IS_EQUAL(result2.rank(), 0);
+  sum = 0.0f;
+  for (int i = 0; i < 7; ++i) {
+    sum += tensor2(i, i, i, i, i);
+  }
+  VERIFY_IS_EQUAL(result2(), sum);
+}
+
+
+template <int DataLayout>
+static void test_simple_trace() {
+  Tensor<float, 3, DataLayout> tensor1(3, 5, 3);
+  tensor1.setRandom();
+  array<ptrdiff_t, 2> dims1({{0, 2}});
+  Tensor<float, 1, DataLayout> result1 = tensor1.trace(dims1);
+  VERIFY_IS_EQUAL(result1.rank(), 1);
+  VERIFY_IS_EQUAL(result1.dimension(0), 5);
+  float sum = 0.0f;
+  for (int i = 0; i < 5; ++i) {
+    sum = 0.0f;
+    for (int j = 0; j < 3; ++j) {
+      sum += tensor1(j, i, j);
+    }
+    VERIFY_IS_EQUAL(result1(i), sum);
+  }
+
+  Tensor<float, 4, DataLayout> tensor2(5, 5, 7, 7);
+  tensor2.setRandom();
+  array<ptrdiff_t, 2> dims2({{2, 3}});
+  Tensor<float, 2, DataLayout> result2 = tensor2.trace(dims2);
+  VERIFY_IS_EQUAL(result2.rank(), 2);
+  VERIFY_IS_EQUAL(result2.dimension(0), 5);
+  VERIFY_IS_EQUAL(result2.dimension(1), 5);
+  for (int i = 0; i < 5; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      sum = 0.0f;
+      for (int k = 0; k < 7; ++k) {
+        sum += tensor2(i, j, k, k);
+      }
+      VERIFY_IS_EQUAL(result2(i, j), sum);
+    }
+  }
+
+  array<ptrdiff_t, 2> dims3({{1, 0}});
+  Tensor<float, 2, DataLayout> result3 = tensor2.trace(dims3);
+  VERIFY_IS_EQUAL(result3.rank(), 2);
+  VERIFY_IS_EQUAL(result3.dimension(0), 7);
+  VERIFY_IS_EQUAL(result3.dimension(1), 7);
+  for (int i = 0; i < 7; ++i) {
+    for (int j = 0; j < 7; ++j) {
+      sum = 0.0f;
+      for (int k = 0; k < 5; ++k) {
+        sum += tensor2(k, k, i, j);
+      }
+      VERIFY_IS_EQUAL(result3(i, j), sum);
+    }
+  }
+
+  Tensor<float, 5, DataLayout> tensor3(3, 7, 3, 7, 3);
+  tensor3.setRandom();
+  array<ptrdiff_t, 3> dims4({{0, 2, 4}});
+  Tensor<float, 2, DataLayout> result4 = tensor3.trace(dims4);
+  VERIFY_IS_EQUAL(result4.rank(), 2);
+  VERIFY_IS_EQUAL(result4.dimension(0), 7);
+  VERIFY_IS_EQUAL(result4.dimension(1), 7);
+  for (int i = 0; i < 7; ++i) {
+    for (int j = 0; j < 7; ++j) {
+      sum = 0.0f;
+      for (int k = 0; k < 3; ++k) {
+        sum += tensor3(k, i, k, j, k);
+      }
+      VERIFY_IS_EQUAL(result4(i, j), sum);
+    }
+  }
+
+  Tensor<float, 5, DataLayout> tensor4(3, 7, 4, 7, 5);
+  tensor4.setRandom();
+  array<ptrdiff_t, 2> dims5({{1, 3}});
+  Tensor<float, 3, DataLayout> result5 = tensor4.trace(dims5);
+  VERIFY_IS_EQUAL(result5.rank(), 3);
+  VERIFY_IS_EQUAL(result5.dimension(0), 3);
+  VERIFY_IS_EQUAL(result5.dimension(1), 4);
+  VERIFY_IS_EQUAL(result5.dimension(2), 5);
+  for (int i = 0; i < 3; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        sum = 0.0f;
+        for (int l = 0; l < 7; ++l) {
+          sum += tensor4(i, l, j, l, k);
+        }
+        VERIFY_IS_EQUAL(result5(i, j, k), sum);
+      }
+    }
+  }
+}
+
+
+template<int DataLayout>
+static void test_trace_in_expr() {
+  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 3);
+  tensor.setRandom();
+  array<ptrdiff_t, 2> dims({{1, 3}});
+  Tensor<float, 2, DataLayout> result(2, 5);
+  result = result.constant(1.0f) - tensor.trace(dims);
+  VERIFY_IS_EQUAL(result.rank(), 2);
+  VERIFY_IS_EQUAL(result.dimension(0), 2);
+  VERIFY_IS_EQUAL(result.dimension(1), 5);
+  float sum = 0.0f;
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      sum = 0.0f;
+      for (int k = 0; k < 3; ++k) {
+        sum += tensor(i, k, j, k);
+      }
+      VERIFY_IS_EQUAL(result(i, j), 1.0f - sum);
+    }
+  }
+}
+
+
+void test_cxx11_tensor_trace() {
+  CALL_SUBTEST(test_0D_trace<ColMajor>());
+  CALL_SUBTEST(test_0D_trace<RowMajor>());
+  CALL_SUBTEST(test_all_dimensions_trace<ColMajor>());
+  CALL_SUBTEST(test_all_dimensions_trace<RowMajor>());
+  CALL_SUBTEST(test_simple_trace<ColMajor>());
+  CALL_SUBTEST(test_simple_trace<RowMajor>());
+  CALL_SUBTEST(test_trace_in_expr<ColMajor>());
+  CALL_SUBTEST(test_trace_in_expr<RowMajor>());
+}

From 5ac27d5b519d63abeadd2537e2c607271f47536f Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Sat, 8 Jul 2017 21:58:44 -0700
Subject: [PATCH 165/680] Avoid relying on cxx11 features when possible.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
index 4b165399f..2b1968de1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@@ -101,7 +101,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-    : m_impl(op.expression(), device), m_device(device)
+    : m_impl(op.expression(), device), m_traceDim(1), m_device(device)
   {
 
     EIGEN_STATIC_ASSERT((NumOutputDims >= 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
@@ -276,7 +276,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
   array<bool, NumInputDims> m_reduced;
   array<Index, NumReducedDims> m_reducedDims;
   // Initialize the size of the trace dimension
-  Index m_traceDim = 1;
+  Index m_traceDim;
   array<Index, NumOutputDims> m_outputStrides;
   array<Index, NumReducedDims> m_reducedStrides;
   array<Index, NumOutputDims> m_preservedStrides;

From 575cda76b33a0169d427885e800cd7e5777bb34c Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Sun, 9 Jul 2017 11:39:01 -0700
Subject: [PATCH 166/680] Fixed syntax errors generated by xcode

---
 .../CXX11/src/Tensor/TensorConcatenation.h    |  4 +--
 .../CXX11/src/Tensor/TensorContraction.h      |  2 +-
 .../CXX11/src/Tensor/TensorConvolution.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   |  4 +--
 .../Eigen/CXX11/src/Tensor/TensorExpr.h       | 28 +++++++++----------
 5 files changed, 20 insertions(+), 20 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
index 363df876c..a7c1380b8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@@ -37,8 +37,8 @@ struct traits<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
   enum { Flags = 0 };
-  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
-                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                               typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 };
 
 template<typename Axis, typename LhsXprType, typename RhsXprType>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index d5ee0d036..e72ddb4a9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -104,7 +104,7 @@ struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >
   // From NumDims below.
   static const int NumDimensions = traits<RhsXprType>::NumDimensions + traits<RhsXprType>::NumDimensions - 2 * array_size<Dimensions>::value;
   static const int Layout = traits<LhsXprType>::Layout;
-  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
   typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 
   enum {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 543b4814b..6fa51fd64 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -231,7 +231,7 @@ struct traits<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<InputXprType>::NumDimensions;
   static const int Layout = traits<InputXprType>::Layout;
-  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename InputXprType::Scalar, Scalar>::val,
+  typedef typename conditional<Pointer_type_promotion<typename InputXprType::Scalar, Scalar>::val,
   typename traits<InputXprType>::PointerType, typename traits<KernelXprType>::PointerType>::type PointerType;
 
   enum {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 3fd997278..0e4db46de 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -185,7 +185,7 @@ struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
-  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
                                 typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 };
 
@@ -300,7 +300,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC typename internal::traits<XprType>::PointerType data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
index 744cf7cc4..4b6540c07 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h
@@ -89,7 +89,7 @@ struct traits<TensorCwiseUnaryOp<UnaryOp, XprType> >
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-  typedef typename ::Eigen::internal::TypeConversion<Scalar, typename XprTraits::PointerType>::type PointerType;
+  typedef typename TypeConversion<Scalar, typename XprTraits::PointerType>::type PointerType;
 };
 
 template<typename UnaryOp, typename XprType>
@@ -162,11 +162,11 @@ struct traits<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-  typedef typename ::Eigen::internal::TypeConversion<Scalar,
-                                                       typename conditional<::Eigen::internal::Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
-                                                       typename traits<LhsXprType>::PointerType,
-                                                       typename traits<RhsXprType>::PointerType>::type
-                                                       >::type PointerType;
+  typedef typename TypeConversion<Scalar,
+                                  typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                  typename traits<LhsXprType>::PointerType,
+                                  typename traits<RhsXprType>::PointerType>::type
+                                  >::type PointerType;
   enum {
     Flags = 0
   };
@@ -243,11 +243,11 @@ struct traits<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprT
   typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-  typedef typename ::Eigen::internal::TypeConversion<Scalar,
-                                                       typename conditional<::Eigen::internal::Pointer_type_promotion<typename Arg2XprType::Scalar, Scalar>::val,
-                                                       typename traits<Arg2XprType>::PointerType,
-                                                       typename traits<Arg3XprType>::PointerType>::type
-                                                       >::type PointerType;
+  typedef typename TypeConversion<Scalar,
+                                  typename conditional<Pointer_type_promotion<typename Arg2XprType::Scalar, Scalar>::val,
+                                  typename traits<Arg2XprType>::PointerType,
+                                  typename traits<Arg3XprType>::PointerType>::type
+                                  >::type PointerType;
   enum {
     Flags = 0
   };
@@ -323,9 +323,9 @@ struct traits<TensorSelectOp<IfXprType, ThenXprType, ElseXprType> >
   typedef typename ElseXprType::Nested ElseNested;
   static const int NumDimensions = XprTraits::NumDimensions;
   static const int Layout = XprTraits::Layout;
-  typedef typename conditional<::Eigen::internal::Pointer_type_promotion<typename ThenXprType::Scalar, Scalar>::val,
-                                                     typename traits<ThenXprType>::PointerType,
-                                                     typename traits<ElseXprType>::PointerType>::type PointerType;
+  typedef typename conditional<Pointer_type_promotion<typename ThenXprType::Scalar, Scalar>::val,
+                               typename traits<ThenXprType>::PointerType,
+                               typename traits<ElseXprType>::PointerType>::type PointerType;
 };
 
 template<typename IfXprType, typename ThenXprType, typename ElseXprType>

From f0b154a4b09914a9f11f5801220785f525217b9e Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 10 Jul 2017 09:54:09 -0700
Subject: [PATCH 167/680] Code cleanup

---
 unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
index f2743ffb1..182bef918 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
@@ -32,7 +32,7 @@ struct traits<TensorConversionOp<TargetType, XprType> >
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
   enum { Flags = 0 };
-   typedef typename ::Eigen::internal::TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
+   typedef typename TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
 };
 
 template<typename TargetType, typename XprType>

From 84d7be103a2d4234f4aab91a434da85fc39b09d6 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Sat, 22 Jul 2017 03:19:34 +0000
Subject: [PATCH 168/680] Fixing Argmax that was breaking upstream TensorFlow.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
index 8f76b8254..442639868 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
@@ -40,7 +40,7 @@ struct traits<TensorTupleReducerDeviceOp<StrideDims, XprType> > : public traits<
   typedef traits<XprType> XprTraits;
   typedef typename XprTraits::StorageKind StorageKind;
   typedef typename XprTraits::Index Index;
-  typedef typename XprType::Scalar Scalar;
+  typedef Index Scalar;
   typedef typename XprType::Nested Nested;
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = XprTraits::NumDimensions;
@@ -58,7 +58,8 @@ class TensorTupleReducerDeviceOp : public TensorBase<TensorTupleReducerDeviceOp<
   typedef typename Eigen::internal::nested<TensorTupleReducerDeviceOp>::type Nested;
   typedef typename Eigen::internal::traits<TensorTupleReducerDeviceOp>::StorageKind StorageKind;
   typedef typename Eigen::internal::traits<TensorTupleReducerDeviceOp>::Index Index;
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::CoeffReturnType TupleType;
+  typedef Index CoeffReturnType;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerDeviceOp(XprType expr,
                                                               const Index return_dim,
@@ -99,9 +100,9 @@ struct TensorEvaluator<const TensorTupleReducerDeviceOp<StrideDims, ArgType>, Sy
 {
   typedef TensorTupleReducerDeviceOp<StrideDims, ArgType> XprType;
   typedef typename XprType::Index Index;
-  typedef typename XprType::Index Scalar;
-  typedef Index CoeffReturnType;
-  typedef typename XprType::CoeffReturnType TupleType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::TupleType TupleType;
   typedef typename TensorEvaluator<ArgType, SyclKernelDevice>::Dimensions Dimensions;
 
   enum {

From e1e71ca4e49a504f638ae9fe449174425f565196 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Sun, 6 Aug 2017 19:53:18 -0400
Subject: [PATCH 169/680] initial support for z14

---
 Eigen/src/Core/arch/ZVector/PacketMath.h | 809 ++++++++++++++---------
 1 file changed, 514 insertions(+), 295 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index 57b01fc63..2e5f5774d 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -41,9 +41,14 @@ typedef __vector double              Packet2d;
 typedef __vector unsigned long long  Packet2ul;
 typedef __vector long long           Packet2l;
 
+// Z14 has builtin support for float vectors
+#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH < 14)
 typedef struct {
 	Packet2d  v4f[2];
 } Packet4f;
+#else
+typedef __vector float               Packet4f;
+#endif
 
 typedef union {
   int32_t   i[4];
@@ -51,11 +56,15 @@ typedef union {
   int64_t   l[2];
   uint64_t ul[2];
   double    d[2];
+  float     f[4];
   Packet4i  v4i;
   Packet4ui v4ui;
   Packet2l  v2l;
   Packet2ul v2ul;
   Packet2d  v2d;
+#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH >= 14)
+  Packet4f  v4f;
+#endif
 } Packet;
 
 // We don't want to write the same code all the time, but we need to reuse the constants
@@ -258,32 +267,6 @@ inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
   return s;
 }
 
-/* Helper function to simulate a vec_splat_packet4f
- */
-template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
-{
-  Packet4f splat;
-  switch (element) {
-  case 0:
-    splat.v4f[0] = vec_splat(from.v4f[0], 0);
-    splat.v4f[1] = splat.v4f[0];
-    break;
-  case 1:
-    splat.v4f[0] = vec_splat(from.v4f[0], 1);
-    splat.v4f[1] = splat.v4f[0];
-    break;
-  case 2:
-    splat.v4f[0] = vec_splat(from.v4f[1], 0);
-    splat.v4f[1] = splat.v4f[0];
-    break;
-  case 3:
-    splat.v4f[0] = vec_splat(from.v4f[1], 1);
-    splat.v4f[1] = splat.v4f[0];
-    break;
-  }
-  return splat;
-}
-
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
@@ -300,31 +283,6 @@ struct palign_impl<Offset,Packet4i>
   }
 };
 
-/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
- */
-template<int Offset>
-struct palign_impl<Offset,Packet4f>
-{
-  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
-  {
-    switch (Offset % 4) {
-    case 1:
-      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
-      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
-      break;
-    case 2:
-      first.v4f[0] = first.v4f[1];
-      first.v4f[1] = second.v4f[0];
-      break;
-    case 3:
-      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
-      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
-      break;
-    }
-  }
-};
-
-
 template<int Offset>
 struct palign_impl<Offset,Packet2d>
 {
@@ -344,16 +302,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
   return vfrom->v4i;
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
-{
-  // FIXME: No intrinsic yet
-  EIGEN_DEBUG_ALIGNED_LOAD
-  Packet4f vfrom;
-  vfrom.v4f[0] = vec_ld2f(&from[0]);
-  vfrom.v4f[1] = vec_ld2f(&from[2]);
-  return vfrom;
-}
-
 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
 {
   // FIXME: No intrinsic yet
@@ -372,15 +320,6 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& f
   vto->v4i = from;
 }
 
-template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
-{
-  // FIXME: No intrinsic yet
-  EIGEN_DEBUG_ALIGNED_STORE
-  vec_st2f(from.v4f[0], &to[0]);
-  vec_st2f(from.v4f[1], &to[2]);
-}
-
-
 template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
 {
   // FIXME: No intrinsic yet
@@ -397,13 +336,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
   return vec_splats(from);
 }
-template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
-{
-  Packet4f to;
-  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
-  to.v4f[1] = to.v4f[0];
-  return to;
-}
 
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet4i>(const int *a,
@@ -416,17 +348,6 @@ pbroadcast4<Packet4i>(const int *a,
   a3 = vec_splat(a3, 3);
 }
 
-template<> EIGEN_STRONG_INLINE void
-pbroadcast4<Packet4f>(const float *a,
-                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
-{
-  a3 = pload<Packet4f>(a);
-  a0 = vec_splat_packet4f<0>(a3);
-  a1 = vec_splat_packet4f<1>(a3);
-  a2 = vec_splat_packet4f<2>(a3);
-  a3 = vec_splat_packet4f<3>(a3);
-}
-
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet2d>(const double *a,
                       Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
@@ -449,16 +370,6 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f
  return pload<Packet4i>(ai);
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
-{
-  float EIGEN_ALIGN16 ai[4];
-  ai[0] = from[0*stride];
-  ai[1] = from[1*stride];
-  ai[2] = from[2*stride];
-  ai[3] = from[3*stride];
- return pload<Packet4f>(ai);
-}
-
 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -477,16 +388,6 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const
   to[3*stride] = ai[3];
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
-{
-  float EIGEN_ALIGN16 ai[4];
-  pstore<float>((float *)ai, from);
-  to[0*stride] = ai[0];
-  to[1*stride] = ai[1];
-  to[2*stride] = ai[2];
-  to[3*stride] = ai[3];
-}
-
 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -496,160 +397,52 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to,
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); }
-template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f c;
-  c.v4f[0] = a.v4f[0] + b.v4f[0];
-  c.v4f[1] = a.v4f[1] + b.v4f[1];
-  return c;
-}
 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); }
-template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f c;
-  c.v4f[0] = a.v4f[0] - b.v4f[0];
-  c.v4f[1] = a.v4f[1] - b.v4f[1];
-  return c;
-}
 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); }
-template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f c;
-  c.v4f[0] = a.v4f[0] * b.v4f[0];
-  c.v4f[1] = a.v4f[1] * b.v4f[1];
-  return c;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); }
-template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f c;
-  c.v4f[0] = a.v4f[0] / b.v4f[0];
-  c.v4f[1] = a.v4f[1] / b.v4f[1];
-  return c;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); }
-template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
-{
-  Packet4f c;
-  c.v4f[0] = -a.v4f[0];
-  c.v4f[1] = -a.v4f[1];
-  return c;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
-template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); }
-template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
-{
-  Packet4f res;
-  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
-  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
-  return res;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
 
 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)    { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); }
-template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
-template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
-template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
-template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
-template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
-template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); }
 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
-template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
-{
-  Packet4f res;
-  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
-  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
-  return res;
-}
 
-template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
-{
-  Packet4f res;
-  res.v4f[0] = vec_round(a.v4f[0]);
-  res.v4f[1] = vec_round(a.v4f[1]);
-  return res;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
-template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
-{
-  Packet4f res;
-  res.v4f[0] = vec_ceil(a.v4f[0]);
-  res.v4f[1] = vec_ceil(a.v4f[1]);
-  return res;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
-template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
-{
-  Packet4f res;
-  res.v4f[0] = vec_floor(a.v4f[0]);
-  res.v4f[1] = vec_floor(a.v4f[1]);
-  return res;
-}
 template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*       from) { return pload<Packet4i>(from); }
-template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double*    from) { return pload<Packet2d>(from); }
 
 
@@ -659,14 +452,6 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
-{
-  Packet4f p = pload<Packet4f>(from);
-  p.v4f[1] = vec_splat(p.v4f[0], 1);
-  p.v4f[0] = vec_splat(p.v4f[0], 0);
-  return p;
-}
-
 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 {
   Packet2d p = pload<Packet2d>(from);
@@ -674,15 +459,12 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 }
 
 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*        to, const Packet4i& from) { pstore<int>(to, from); }
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { pstore<double>(to, from); }
 
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int    EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
-template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
@@ -695,23 +477,8 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
   return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
-{
-  Packet4f rev;
-  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
-  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
-  return rev;
-}
-
 template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); }
 template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); }
-template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
-{
-  Packet4f res;
-  res.v4f[0] = pabs(a.v4f[0]);
-  res.v4f[1] = pabs(a.v4f[1]);
-  return res;
-}
 
 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
@@ -730,13 +497,6 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
   sum = padd<Packet2d>(a, b);
   return pfirst(sum);
 }
-template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
-{
-  Packet2d sum;
-  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
-  double first = predux<Packet2d>(sum);
-  return static_cast<float>(first);
-}
 
 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
 {
@@ -777,21 +537,6 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
   return sum;
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
-{
-  PacketBlock<Packet4f,4> transpose;
-  transpose.packet[0] = vecs[0];
-  transpose.packet[1] = vecs[1];
-  transpose.packet[2] = vecs[2];
-  transpose.packet[3] = vecs[3];
-  ptranspose(transpose);
-
-  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
-  sum = padd(sum, transpose.packet[2]);
-  sum = padd(sum, transpose.packet[3]);
-  return sum;
-}
-
 // Other reduction functions:
 // mul
 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
@@ -806,12 +551,6 @@ template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
   return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
-template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
-{
-  // Return predux_mul<Packet2d> of the subvectors product
-  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
-}
-
 // min
 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
 {
@@ -826,14 +565,6 @@ template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
   return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
-template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
-{
-  Packet2d b, res;
-  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
-  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
-  return static_cast<float>(pfirst(res));
-}
-
 // max
 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 {
@@ -849,14 +580,6 @@ template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
   return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
-template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
-{
-  Packet2d b, res;
-  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
-  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
-  return static_cast<float>(pfirst(res));
-}
-
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<Packet4i,4>& kernel) {
   Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
@@ -877,6 +600,321 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) {
   kernel.packet[1] = t1;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+/* z13 has no vector float support so we emulate that with double
+   z14 has proper vector float support.
+*/
+#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH < 14)
+/* Helper function to simulate a vec_splat_packet4f
+ */
+template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
+{
+  Packet4f splat;
+  switch (element) {
+  case 0:
+    splat.v4f[0] = vec_splat(from.v4f[0], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 1:
+    splat.v4f[0] = vec_splat(from.v4f[0], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 2:
+    splat.v4f[0] = vec_splat(from.v4f[1], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 3:
+    splat.v4f[0] = vec_splat(from.v4f[1], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  }
+  return splat;
+}
+
+/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
+ */
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
+      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
+      break;
+    case 2:
+      first.v4f[0] = first.v4f[1];
+      first.v4f[1] = second.v4f[0];
+      break;
+    case 3:
+      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
+      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
+      break;
+    }
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet4f vfrom;
+  vfrom.v4f[0] = vec_ld2f(&from[0]);
+  vfrom.v4f[1] = vec_ld2f(&from[2]);
+  return vfrom;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_st2f(from.v4f[0], &to[0]);
+  vec_st2f(from.v4f[1], &to[2]);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
+{
+  Packet4f to;
+  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
+  to.v4f[1] = to.v4f[0];
+  return to;
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat_packet4f<0>(a3);
+  a1 = vec_splat_packet4f<1>(a3);
+  a2 = vec_splat_packet4f<2>(a3);
+  a3 = vec_splat_packet4f<3>(a3);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4f>(ai);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  pstore<float>((float *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] + b.v4f[0];
+  c.v4f[1] = a.v4f[1] + b.v4f[1];
+  return c;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] - b.v4f[0];
+  c.v4f[1] = a.v4f[1] - b.v4f[1];
+  return c;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] * b.v4f[0];
+  c.v4f[1] = a.v4f[1] * b.v4f[1];
+  return c;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] / b.v4f[0];
+  c.v4f[1] = a.v4f[1] / b.v4f[1];
+  return c;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  Packet4f c;
+  c.v4f[0] = -a.v4f[0];
+  c.v4f[1] = -a.v4f[1];
+  return c;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
+{
+  Packet4f res;
+  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
+  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_round(a.v4f[0]);
+  res.v4f[1] = vec_round(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_ceil(a.v4f[0]);
+  res.v4f[1] = vec_ceil(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_floor(a.v4f[0]);
+  res.v4f[1] = vec_floor(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  p.v4f[1] = vec_splat(p.v4f[0], 1);
+  p.v4f[0] = vec_splat(p.v4f[0], 0);
+  return p;
+}
+
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  Packet4f rev;
+  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
+  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
+  return rev;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = pabs(a.v4f[0]);
+  res.v4f[1] = pabs(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet2d sum;
+  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
+  double first = predux<Packet2d>(sum);
+  return static_cast<float>(first);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  PacketBlock<Packet4f,4> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  transpose.packet[2] = vecs[2];
+  transpose.packet[3] = vecs[3];
+  ptranspose(transpose);
+
+  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
+  sum = padd(sum, transpose.packet[2]);
+  sum = padd(sum, transpose.packet[3]);
+  return sum;
+}
+
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  // Return predux_mul<Packet2d> of the subvectors product
+  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
 /* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
  */
 EIGEN_DEVICE_FUNC inline void
@@ -915,12 +953,6 @@ ptranspose(PacketBlock<Packet4f,4>& kernel) {
   kernel.packet[3].v4f[1] = t3.packet[1];
 }
 
-template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
-  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
-  Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
-  return vec_sel(elsePacket, thenPacket, mask);
-}
-
 template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
   Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] };
   Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] };
@@ -931,13 +963,200 @@ template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, cons
   result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
   return result;
 }
+#else
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+  }
+};
 
-template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
-  Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
-  Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet *vfrom;
+  vfrom = (Packet *) from;
+  return vfrom->v4f;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  Packet *vto;
+  vto = (Packet *) to;
+  vto->v4f = from;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from)
+{
+  return vec_splats(from);
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 af[4];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  af[2] = from[2*stride];
+  af[3] = from[3*stride];
+ return pload<Packet4f>(af);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 af[4];
+  pstore<float>((float*)af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+  to[2*stride] = af[2];
+  to[3*stride] = af[3];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a + b); }
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a - b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a * b); }
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a / b); }
+template<> EIGEN_STRONG_INLINE Packet4f pnegate<Packet4f>(const Packet4f& a) { return (-a); }
+template<> EIGEN_STRONG_INLINE Packet4f pconj<Packet4f>  (const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd<Packet4f>  (const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>  (const float& a)  { return padd<Packet4i>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>    (const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f> (const Packet4f& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>  (const Packet4f& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f> (const Packet4f& a) { return vec_floor(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>   (const Packet4f& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f> (const float* from) { return pload<Packet4f>(from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); }
+template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
+}
+
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4f b, sum;
+  b   = vec_sld(a, a, 8);
+  sum = padd<Packet4f>(a, b);
+  b   = vec_sld(sum, sum, 4);
+  sum = padd<Packet4f>(sum, b);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  Packet4f v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = padd<Packet4f>(sum[0], sum[1]);
+  // Lines 2+3
+  sum[1] = padd<Packet4f>(sum[2], sum[3]);
+  // Add the results
+  sum[0] = padd<Packet4f>(sum[0], sum[1]);
+
+  return sum[0];
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  return pfirst(pmul(a, reinterpret_cast<Packet4f>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8))));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b   = pmin<Packet4f>(a, vec_sld(a, a, 8));
+  res = pmin<Packet4f>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b = pmax<Packet4f>(a, vec_sld(a, a, 8));
+  res = pmax<Packet4f>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  Packet4f t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  Packet4f t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  Packet4f t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  Packet4f t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
   return vec_sel(elsePacket, thenPacket, mask);
 }
 
+#endif
+
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
+
 } // end namespace internal
 
 } // end namespace Eigen

From b223918ea99dcff9f6a3f8d017e7bd79ff4a7db7 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 14:12:47 +0200
Subject: [PATCH 170/680] Doc: warn about constness in LLT::solveInPlace

---
 Eigen/src/Cholesky/LLT.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 152837bfd..814174d47 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -489,6 +489,9 @@ void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
   *
   * This version avoids a copy when the right hand side matrix b is not needed anymore.
   *
+  * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
+  * This function will const_cast it, so constness isn't honored here.
+  *
   * \sa LLT::solve(), MatrixBase::llt()
   */
 template<typename MatrixType, int _UpLo>

From fc39d5954b72ca2307921beb8a784cd78c2a8d10 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 12:17:37 +0000
Subject: [PATCH 171/680] Merged in dtrebbien/eigen/patch-1 (pull request
 PR-312)

Work around a compilation error seen with nvcc V8.0.61
---
 Eigen/src/Core/products/TriangularMatrixMatrix.h | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 6ec5a8a0b..539b6c0c6 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -137,7 +137,13 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
+    // To work around an "error: member reference base type 'Matrix<...>
+    // (Eigen::internal::constructor_without_unaligned_array_assert (*)())' is
+    // not a structure or union" compilation error in nvcc (tested V8.0.61),
+    // create a dummy internal::constructor_without_unaligned_array_assert
+    // object to pass to the Matrix constructor.
+    internal::constructor_without_unaligned_array_assert a;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer(a);
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();
@@ -284,7 +290,8 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
+    internal::constructor_without_unaligned_array_assert a;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer(a);
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();

From bc91a2df8b9f1c5fa47bfeb9b03c2036890570b5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 15:10:42 +0200
Subject: [PATCH 172/680] bug #1461: fix compilation of Map<const
 Quaternion>::x()

---
 Eigen/src/Geometry/Quaternion.h | 29 +++++++++++++++++------------
 test/geo_quaternion.cpp         | 13 +++++++++++++
 2 files changed, 30 insertions(+), 12 deletions(-)

diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h
index 3e5a9badb..c3fd8c3e0 100644
--- a/Eigen/src/Geometry/Quaternion.h
+++ b/Eigen/src/Geometry/Quaternion.h
@@ -43,6 +43,11 @@ class QuaternionBase : public RotationBase<Derived, 3>
   typedef typename internal::traits<Derived>::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef typename internal::traits<Derived>::Coefficients Coefficients;
+  typedef typename Coefficients::CoeffReturnType CoeffReturnType;
+  typedef typename internal::conditional<bool(internal::traits<Derived>::Flags&LvalueBit),
+                                        Scalar&, CoeffReturnType>::type NonConstCoeffReturnType;
+
+
   enum {
     Flags = Eigen::internal::traits<Derived>::Flags
   };
@@ -58,22 +63,22 @@ class QuaternionBase : public RotationBase<Derived, 3>
 
 
   /** \returns the \c x coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar x() const { return this->derived().coeffs().coeff(0); }
+  EIGEN_DEVICE_FUNC inline CoeffReturnType x() const { return this->derived().coeffs().coeff(0); }
   /** \returns the \c y coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar y() const { return this->derived().coeffs().coeff(1); }
+  EIGEN_DEVICE_FUNC inline CoeffReturnType y() const { return this->derived().coeffs().coeff(1); }
   /** \returns the \c z coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar z() const { return this->derived().coeffs().coeff(2); }
+  EIGEN_DEVICE_FUNC inline CoeffReturnType z() const { return this->derived().coeffs().coeff(2); }
   /** \returns the \c w coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar w() const { return this->derived().coeffs().coeff(3); }
+  EIGEN_DEVICE_FUNC inline CoeffReturnType w() const { return this->derived().coeffs().coeff(3); }
 
-  /** \returns a reference to the \c x coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar& x() { return this->derived().coeffs().coeffRef(0); }
-  /** \returns a reference to the \c y coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar& y() { return this->derived().coeffs().coeffRef(1); }
-  /** \returns a reference to the \c z coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar& z() { return this->derived().coeffs().coeffRef(2); }
-  /** \returns a reference to the \c w coefficient */
-  EIGEN_DEVICE_FUNC inline Scalar& w() { return this->derived().coeffs().coeffRef(3); }
+  /** \returns a reference to the \c x coefficient (if Derived is a non-const lvalue) */
+  EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType x() { return this->derived().coeffs().x(); }
+  /** \returns a reference to the \c y coefficient (if Derived is a non-const lvalue) */
+  EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType y() { return this->derived().coeffs().y(); }
+  /** \returns a reference to the \c z coefficient (if Derived is a non-const lvalue) */
+  EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType z() { return this->derived().coeffs().z(); }
+  /** \returns a reference to the \c w coefficient (if Derived is a non-const lvalue) */
+  EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType w() { return this->derived().coeffs().w(); }
 
   /** \returns a read-only vector expression of the imaginary part (x,y,z) */
   EIGEN_DEVICE_FUNC inline const VectorBlock<const Coefficients,3> vec() const { return coeffs().template head<3>(); }
diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp
index 96889e722..8ee8fdb27 100644
--- a/test/geo_quaternion.cpp
+++ b/test/geo_quaternion.cpp
@@ -231,6 +231,19 @@ template<typename Scalar> void mapQuaternion(void){
   VERIFY_IS_APPROX(mq3*mq2, q3*q2);
   VERIFY_IS_APPROX(mcq1*mq2, q1*q2);
   VERIFY_IS_APPROX(mcq3*mq2, q3*q2);
+
+  // Bug 1461, compilation issue with Map<const Quat>::w(), and other reference/constness checks:
+  VERIFY_IS_APPROX(mcq3.coeffs().x() + mcq3.coeffs().y() + mcq3.coeffs().z() + mcq3.coeffs().w(), mcq3.coeffs().sum());
+  VERIFY_IS_APPROX(mcq3.x() + mcq3.y() + mcq3.z() + mcq3.w(), mcq3.coeffs().sum());
+  mq3.w() = 1;
+  const Quaternionx& cq3(q3);
+  VERIFY( &cq3.x() == &q3.x() );
+  const MQuaternionUA& cmq3(mq3);
+  VERIFY( &cmq3.x() == &mq3.x() );
+  // FIXME the following should be ok. The problem is that currently the LValueBit flag
+  // is used to determine wether we can return a coeff by reference or not, which is not enough for Map<const ...>.
+  //const MCQuaternionUA& cmcq3(mcq3);
+  //VERIFY( &cmcq3.x() == &mcq3.x() );
 }
 
 template<typename Scalar> void quaternionAlignment(void){

From bc4dae9aeb84cc3d3114ee496d55654cc7256584 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 15:59:08 +0200
Subject: [PATCH 173/680] bug #1449: fix redux_3 unit test

---
 test/redux.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/test/redux.cpp b/test/redux.cpp
index 989e1057b..2bade3735 100644
--- a/test/redux.cpp
+++ b/test/redux.cpp
@@ -9,6 +9,8 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #define TEST_ENABLE_TEMPORARY_TRACKING
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+// ^^ see bug 1449
 
 #include "main.h"
 

From 9deee79922c38415125e4d6c2cd34cd05bda7889 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 16:48:07 +0200
Subject: [PATCH 174/680] bug #1457: add setUnit() methods for consistency.

---
 Eigen/src/Core/CwiseNullaryOp.h | 36 +++++++++++++++++++++++++++++++++
 Eigen/src/Core/MatrixBase.h     |  2 ++
 doc/QuickReference.dox          | 10 ++++++++-
 test/nullary.cpp                | 18 +++++++++++++++++
 4 files changed, 65 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h
index 144608ec2..b1923da0f 100644
--- a/Eigen/src/Core/CwiseNullaryOp.h
+++ b/Eigen/src/Core/CwiseNullaryOp.h
@@ -861,6 +861,42 @@ template<typename Derived>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
 { return Derived::Unit(3); }
 
+/** \brief Set the coefficients of \c *this to the i-th unit (basis) vector
+  *
+  * \param i index of the unique coefficient to be set to 1
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Unit(Index,Index)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setUnit(Index i)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+  eigen_assert(i<size());
+  derived().setZero();
+  derived().coeffRef(i) = Scalar(1);
+  return derived();
+}
+
+/** \brief Resizes to the given \a newSize, and writes the i-th unit (basis) vector into *this.
+  *
+  * \param newSize the new size of the vector
+  * \param i index of the unique coefficient to be set to 1
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Unit(Index,Index)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setUnit(Index newSize, Index i)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+  eigen_assert(i<newSize);
+  derived().resize(newSize);
+  return setUnit(i);
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_CWISE_NULLARY_OP_H
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 5786f76e0..11435903b 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -268,6 +268,8 @@ template<typename Derived> class MatrixBase
     Derived& setIdentity();
     EIGEN_DEVICE_FUNC
     Derived& setIdentity(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC Derived& setUnit(Index i);
+    EIGEN_DEVICE_FUNC Derived& setUnit(Index newSize, Index i);
 
     bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
     bool isDiagonal(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
diff --git a/doc/QuickReference.dox b/doc/QuickReference.dox
index 44f5410db..59d7d05e4 100644
--- a/doc/QuickReference.dox
+++ b/doc/QuickReference.dox
@@ -261,6 +261,8 @@ x.setIdentity();
 Vector3f::UnitX() // 1 0 0
 Vector3f::UnitY() // 0 1 0
 Vector3f::UnitZ() // 0 0 1
+Vector4f::Unit(i)
+x.setUnit(i);
 \endcode
   </td>
   <td>
@@ -278,6 +280,7 @@ N/A
 
 
 VectorXf::Unit(size,i)
+x.setUnit(size,i);
 VectorXf::Unit(4,1) == Vector4f(0,1,0,0)
                     == Vector4f::UnitY()
 \endcode
@@ -285,7 +288,12 @@ VectorXf::Unit(4,1) == Vector4f(0,1,0,0)
 </tr>
 </table>
 
-
+Note that it is allowed to call any of the \c set* functions to a dynamic-sized vector or matrix without passing new sizes.
+For instance:
+\code
+MatrixXi M(3,3);
+M.setIdentity();
+\endcode
 
 \subsection QuickRef_Map Mapping external arrays
 
diff --git a/test/nullary.cpp b/test/nullary.cpp
index acd55506e..22ec92352 100644
--- a/test/nullary.cpp
+++ b/test/nullary.cpp
@@ -191,6 +191,24 @@ void testVectorType(const VectorType& base)
       }
     }
   }
+
+  // test setUnit()
+  if(m.size()>0)
+  {
+    for(Index k=0; k<10; ++k)
+    {
+      Index i = internal::random<Index>(0,m.size()-1);
+      m.setUnit(i);
+      VERIFY_IS_APPROX( m, VectorType::Unit(m.size(), i) );
+    }
+    if(VectorType::SizeAtCompileTime==Dynamic)
+    {
+      Index i = internal::random<Index>(0,2*m.size()-1);
+      m.setUnit(2*m.size(),i);
+      VERIFY_IS_APPROX( m, VectorType::Unit(m.size(),i) );
+    }
+  }
+
 }
 
 template<typename MatrixType>

From 600e52fc7f574504fa832d67c9d94c991e504bdc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 17:06:57 +0200
Subject: [PATCH 175/680] Add missing scalar conversion

---
 Eigen/src/Geometry/AngleAxis.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Geometry/AngleAxis.h b/Eigen/src/Geometry/AngleAxis.h
index 0af3c1b08..83ee1be46 100644
--- a/Eigen/src/Geometry/AngleAxis.h
+++ b/Eigen/src/Geometry/AngleAxis.h
@@ -178,7 +178,7 @@ EIGEN_DEVICE_FUNC AngleAxis<Scalar>& AngleAxis<Scalar>::operator=(const Quaterni
   if (n != Scalar(0))
   {
     m_angle = Scalar(2)*atan2(n, abs(q.w()));
-    if(q.w() < 0)
+    if(q.w() < Scalar(0))
       n = -n;
     m_axis  = q.vec() / n;
   }

From 39864ebe1eb7c8028769cf5d8750faaabce22446 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 Aug 2017 17:18:43 +0200
Subject: [PATCH 176/680] bug #336: improve doc for PlainObjectBase::Map

---
 Eigen/src/Core/PlainObjectBase.h   | 4 ++++
 doc/snippets/Matrix_Map_stride.cpp | 7 +++++++
 2 files changed, 11 insertions(+)
 create mode 100644 doc/snippets/Matrix_Map_stride.cpp

diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index 77f4f6066..1dc7e223a 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -577,6 +577,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned
       * \a data pointers.
       *
+      * Here is an example using strides:
+      * \include Matrix_Map_stride.cpp
+      * Output: \verbinclude Matrix_Map_stride.out
+      *
       * \see class Map
       */
     //@{
diff --git a/doc/snippets/Matrix_Map_stride.cpp b/doc/snippets/Matrix_Map_stride.cpp
new file mode 100644
index 000000000..ae42a127a
--- /dev/null
+++ b/doc/snippets/Matrix_Map_stride.cpp
@@ -0,0 +1,7 @@
+Matrix4i A;
+A << 1,  2,  3,  4,
+     5,  6,  7,  8,
+     9, 10, 11, 12,
+    13, 14, 15, 16;
+
+std::cout << Matrix2i::Map(&A(1,1),Stride<8,2>()) << std::endl;

From 12249849b5ef7ec0c64f74440690fb00708b8da6 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 24 Aug 2017 10:43:21 +0200
Subject: [PATCH 177/680] Make the threshold from gemm to coeff-based-product
 configurable, and add some explanations.

---
 Eigen/src/Core/GeneralProduct.h               | 10 ++++++++++
 Eigen/src/Core/products/GeneralMatrixMatrix.h | 12 +++++++++---
 2 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index dec24848d..483277fe6 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -18,6 +18,16 @@ enum {
   Small = 3
 };
 
+// Define the threshold value to fallback from the generic matrix-matrix product
+// implementation (heavy) to the lightweight coeff-based product one.
+// See generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
+// in products/GeneralMatrixMatrix.h for more details.
+// TODO This threshold should also be used in the compile-time selector below.
+#ifndef EIGEN_GEMM_TO_COEFFBASED_THRESHOLD
+// This default value has been obtained on a Haswell architecture.
+#define EIGEN_GEMM_TO_COEFFBASED_THRESHOLD 20
+#endif
+
 namespace internal {
 
 template<int Rows, int Cols, int Depth> struct product_type_selector;
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index 6440e1d09..ed4d3182b 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -427,7 +427,13 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=404 for a discussion and helper program
+    // to determine the following heuristic.
+    // EIGEN_GEMM_TO_COEFFBASED_THRESHOLD is typically defined to 20 in GeneralProduct.h,
+    // unless it has been specialized by the user or for a given architecture.
+    // Note that the condition rhs.rows()>0 was required because lazy produc is (was?) not happy with empty inputs.
+    // I'm not sure it is still required.
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::evalTo(dst, lhs, rhs);
     else
     {
@@ -439,7 +445,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::addTo(dst, lhs, rhs);
     else
       scaleAndAddTo(dst,lhs, rhs, Scalar(1));
@@ -448,7 +454,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::subTo(dst, lhs, rhs);
     else
       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));

From 21633e585b61564159d9cfbfbbad9006b8a09d64 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 24 Aug 2017 11:06:47 +0200
Subject: [PATCH 178/680] bug #1462: remove all occurences of the deprecated
 __CUDACC_VER__ macro by introducing EIGEN_CUDACC_VER

---
 Eigen/Core                                           | 10 +++++++++-
 Eigen/src/Core/arch/CUDA/Half.h                      | 12 ++++++------
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h            |  2 +-
 Eigen/src/Core/util/Macros.h                         |  7 ++++---
 test/cuda_basic.cu                                   |  2 +-
 unsupported/test/cxx11_tensor_argmax_cuda.cu         |  2 +-
 unsupported/test/cxx11_tensor_cast_float16_cuda.cu   |  2 +-
 unsupported/test/cxx11_tensor_complex_cuda.cu        |  2 +-
 .../test/cxx11_tensor_complex_cwise_ops_cuda.cu      |  2 +-
 unsupported/test/cxx11_tensor_contract_cuda.cu       |  2 +-
 unsupported/test/cxx11_tensor_cuda.cu                |  2 +-
 unsupported/test/cxx11_tensor_device.cu              |  2 +-
 unsupported/test/cxx11_tensor_of_float16_cuda.cu     |  2 +-
 unsupported/test/cxx11_tensor_random_cuda.cu         |  2 +-
 unsupported/test/cxx11_tensor_reduction_cuda.cu      |  2 +-
 unsupported/test/cxx11_tensor_scan_cuda.cu           |  2 +-
 16 files changed, 32 insertions(+), 23 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index f613ee9aa..f6fe4b0ec 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -22,6 +22,14 @@
   #define EIGEN_CUDA_ARCH __CUDA_ARCH__
 #endif
 
+#if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
+#define EIGEN_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
+#elif defined(__CUDACC_VER__)
+#define EIGEN_CUDACC_VER __CUDACC_VER__
+#else
+#define EIGEN_CUDACC_VER 0
+#endif
+
 // Handle NVCC/CUDA/SYCL
 #if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__)
   // Do not try asserts on CUDA and SYCL!
@@ -248,7 +256,7 @@
 #if defined EIGEN_CUDACC
   #define EIGEN_VECTORIZE_CUDA
   #include <vector_types.h>
-  #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+  #if EIGEN_CUDACC_VER >= 70500
     #define EIGEN_HAS_CUDA_FP16
   #endif
 #endif
diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index e99847e24..8cedd65ad 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -386,7 +386,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
   return result;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
   return half(hexp(a));
 #else
    return half(::expf(float(a)));
@@ -396,7 +396,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
   return half(numext::expm1(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return half(::hlog(a));
 #else
   return half(::logf(float(a)));
@@ -409,7 +409,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
   return half(::log10f(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
   return half(hsqrt(a));
 #else
     return half(::sqrtf(float(a)));
@@ -431,14 +431,14 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
   return half(::tanhf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
+#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
   return half(hfloor(a));
 #else
   return half(::floorf(float(a)));
 #endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
+#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
   return half(hceil(a));
 #else
   return half(::ceilf(float(a)));
@@ -576,7 +576,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
   return Eigen::half(::expf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
   return Eigen::half(::hlog(a));
 #else
   return Eigen::half(::logf(float(a)));
diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index 4a730a0e0..ba6a7f920 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -284,7 +284,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
 
 template<>  __device__ EIGEN_STRONG_INLINE
 half2 plog<half2>(const half2& a) {
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 322e8d899..b63ea2697 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -413,7 +413,7 @@
 // Does the compiler support variadic templates?
 #ifndef EIGEN_HAS_VARIADIC_TEMPLATES
 #if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \
-  && (!defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000) )
+  && (!defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (EIGEN_CUDACC_VER >= 80000) )
     // ^^ Disable the use of variadic templates when compiling with versions of nvcc older than 8.0 on ARM devices:
     //    this prevents nvcc from crashing when compiling Eigen on Tegra X1
 #define EIGEN_HAS_VARIADIC_TEMPLATES 1
@@ -429,7 +429,7 @@
 
 #if defined(EIGEN_CUDACC)
 // Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
-#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500))
+#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500))
   #define EIGEN_HAS_CONSTEXPR 1
 #endif
 #elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
@@ -837,7 +837,8 @@ namespace Eigen {
 // just an empty macro !
 #define EIGEN_EMPTY
 
-#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 ||  defined(__CUDACC_VER__)) // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324)
+#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 || EIGEN_CUDACC_VER>0)
+  // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324)
   #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
     using Base::operator =;
 #elif EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu
index cb2e4167a..0ff13477d 100644
--- a/test/cuda_basic.cu
+++ b/test/cuda_basic.cu
@@ -20,7 +20,7 @@
 
 #include <math_constants.h>
 #include <cuda.h>
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_cuda.cu
index 653443dc5..0dfd6cfe1 100644
--- a/unsupported/test/cxx11_tensor_argmax_cuda.cu
+++ b/unsupported/test/cxx11_tensor_argmax_cuda.cu
@@ -12,7 +12,7 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_cuda
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
index 88c233994..83a740e7a 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
@@ -13,7 +13,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index 87cf28920..cbff5a9b2 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -11,7 +11,7 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_complex
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
index 2baf5eaad..9133fce5a 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
@@ -11,7 +11,7 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_complex_cwise_ops
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_cuda.cu
index dd68430ce..0b2f3f0f4 100644
--- a/unsupported/test/cxx11_tensor_contract_cuda.cu
+++ b/unsupported/test/cxx11_tensor_contract_cuda.cu
@@ -14,7 +14,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index 0ba9d52e9..ad8c9662f 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -12,7 +12,7 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_cuda
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index fde20ddf2..ae21f492a 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -13,7 +13,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index 908a5e5a9..0ba7657b8 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -13,7 +13,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu
index b3be199e1..94d5f4e5a 100644
--- a/unsupported/test/cxx11_tensor_random_cuda.cu
+++ b/unsupported/test/cxx11_tensor_random_cuda.cu
@@ -13,7 +13,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_cuda.cu
index 6858b43a7..fd09d013b 100644
--- a/unsupported/test/cxx11_tensor_reduction_cuda.cu
+++ b/unsupported/test/cxx11_tensor_reduction_cuda.cu
@@ -12,7 +12,7 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_reduction_cuda
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if dEIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_cuda.cu
index 5f146f3c9..46571cfea 100644
--- a/unsupported/test/cxx11_tensor_scan_cuda.cu
+++ b/unsupported/test/cxx11_tensor_scan_cuda.cu
@@ -13,7 +13,7 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#if EIGEN_CUDACC_VER >= 70500
 #include <cuda_fp16.h>
 #endif
 #include "main.h"

From 304ef2957134be386e50592ad7120177c5f3a7c0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 24 Aug 2017 11:26:41 +0200
Subject: [PATCH 179/680] Handle min/max/inf/etc issue in cuda_fp16.h directly
 in test/main.h

---
 test/cuda_basic.cu                                  |  3 ---
 test/main.h                                         | 13 +++++++++++++
 unsupported/test/cxx11_tensor_argmax_cuda.cu        |  3 ---
 unsupported/test/cxx11_tensor_cast_float16_cuda.cu  |  3 ---
 unsupported/test/cxx11_tensor_complex_cuda.cu       |  3 ---
 .../test/cxx11_tensor_complex_cwise_ops_cuda.cu     |  3 ---
 unsupported/test/cxx11_tensor_contract_cuda.cu      |  3 ---
 unsupported/test/cxx11_tensor_cuda.cu               |  3 ---
 unsupported/test/cxx11_tensor_device.cu             |  3 ---
 unsupported/test/cxx11_tensor_of_float16_cuda.cu    |  3 ---
 unsupported/test/cxx11_tensor_random_cuda.cu        |  3 ---
 unsupported/test/cxx11_tensor_reduction_cuda.cu     |  3 ---
 unsupported/test/cxx11_tensor_scan_cuda.cu          |  3 ---
 13 files changed, 13 insertions(+), 36 deletions(-)

diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu
index 0ff13477d..ce66c2c78 100644
--- a/test/cuda_basic.cu
+++ b/test/cuda_basic.cu
@@ -20,9 +20,6 @@
 
 #include <math_constants.h>
 #include <cuda.h>
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include "cuda_common.h"
 
diff --git a/test/main.h b/test/main.h
index bd5325196..429c44f81 100644
--- a/test/main.h
+++ b/test/main.h
@@ -50,6 +50,19 @@
 #endif
 #endif
 
+// Same for cuda_fp16.h
+#if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
+#define EIGEN_TEST_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
+#elif defined(__CUDACC_VER__)
+#define EIGEN_TEST_CUDACC_VER __CUDACC_VER__
+#else
+#define EIGEN_TEST_CUDACC_VER 0
+#endif
+
+#if EIGEN_TEST_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 // To test that all calls from Eigen code to std::min() and std::max() are
 // protected by parenthesis against macro expansion, the min()/max() macros
 // are defined here and any not-parenthesized min/max call will cause a
diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_cuda.cu
index 0dfd6cfe1..3d73d491a 100644
--- a/unsupported/test/cxx11_tensor_argmax_cuda.cu
+++ b/unsupported/test/cxx11_tensor_argmax_cuda.cu
@@ -12,9 +12,6 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_cuda
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
index 83a740e7a..816e03220 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
@@ -13,9 +13,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index cbff5a9b2..a52350f85 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -11,9 +11,6 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_complex
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
index 9133fce5a..aac780905 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
@@ -11,9 +11,6 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_complex_cwise_ops
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_cuda.cu
index 0b2f3f0f4..e821ccf0c 100644
--- a/unsupported/test/cxx11_tensor_contract_cuda.cu
+++ b/unsupported/test/cxx11_tensor_contract_cuda.cu
@@ -14,9 +14,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index ad8c9662f..9584a539f 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -12,9 +12,6 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_cuda
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index ae21f492a..cbb43e210 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -13,9 +13,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index 0ba7657b8..b3aab0b9d 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -13,9 +13,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu
index 94d5f4e5a..fa1a46732 100644
--- a/unsupported/test/cxx11_tensor_random_cuda.cu
+++ b/unsupported/test/cxx11_tensor_random_cuda.cu
@@ -13,9 +13,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_cuda.cu
index fd09d013b..ec0669704 100644
--- a/unsupported/test/cxx11_tensor_reduction_cuda.cu
+++ b/unsupported/test/cxx11_tensor_reduction_cuda.cu
@@ -12,9 +12,6 @@
 #define EIGEN_TEST_FUNC cxx11_tensor_reduction_cuda
 #define EIGEN_USE_GPU
 
-#if dEIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_cuda.cu
index 46571cfea..de1c0ac95 100644
--- a/unsupported/test/cxx11_tensor_scan_cuda.cu
+++ b/unsupported/test/cxx11_tensor_scan_cuda.cu
@@ -13,9 +13,6 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
-#if EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 

From 6d991a9595ffe22a0924bb8140a37355aca4cccb Mon Sep 17 00:00:00 2001
From: Abhijit Kundu <abhijit.kundu@gatech.edu>
Date: Wed, 30 Aug 2017 13:26:30 -0400
Subject: [PATCH 180/680] bug #1464 : Fixes construction of EulerAngles from 3D
 vector expression.

---
 unsupported/Eigen/src/EulerAngles/EulerAngles.h | 2 +-
 unsupported/test/EulerAngles.cpp                | 3 +++
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/src/EulerAngles/EulerAngles.h b/unsupported/Eigen/src/EulerAngles/EulerAngles.h
index a5d034d71..e43cdb7fb 100644
--- a/unsupported/Eigen/src/EulerAngles/EulerAngles.h
+++ b/unsupported/Eigen/src/EulerAngles/EulerAngles.h
@@ -341,7 +341,7 @@ EIGEN_EULER_ANGLES_TYPEDEFS(double, d)
     
     // set from a vector of Euler angles
     template<class System, class Other>
-    struct eulerangles_assign_impl<System,Other,4,1>
+    struct eulerangles_assign_impl<System,Other,3,1>
     {
       typedef typename Other::Scalar Scalar;
       static void run(EulerAngles<Scalar, System>& e, const Other& vec)
diff --git a/unsupported/test/EulerAngles.cpp b/unsupported/test/EulerAngles.cpp
index 79ee72847..500fb2d17 100644
--- a/unsupported/test/EulerAngles.cpp
+++ b/unsupported/test/EulerAngles.cpp
@@ -278,6 +278,9 @@ void test_EulerAngles()
   EulerAnglesXYZd onesEd(1, 1, 1);
   EulerAnglesXYZf onesEf = onesEd.cast<float>();
   VERIFY_IS_APPROX(onesEd, onesEf.cast<double>());
+
+  // Simple Construction from Vector3 test
+  VERIFY_IS_APPROX(onesEd, EulerAnglesXYZd(Vector3d::Ones()));
   
   CALL_SUBTEST_1( eulerangles_manual<float>() );
   CALL_SUBTEST_2( eulerangles_manual<double>() );

From a4089991eb6bdb9e8ebfef93d81ca7b5e67ea77d Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 31 Aug 2017 02:49:39 +0000
Subject: [PATCH 181/680] Added support for CUDA 9.0.

---
 Eigen/Core                                    |  1 +
 Eigen/src/Core/arch/CUDA/Half.h               | 58 ++++++++++-------
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h     |  3 +-
 test/half_float.cpp                           | 38 +++++------
 .../CXX11/src/Tensor/TensorContractionCuda.h  |  9 +++
 .../CXX11/src/Tensor/TensorReductionCuda.h    | 63 +++++++++++++------
 unsupported/test/cxx11_tensor_argmax_cuda.cu  |  6 ++
 .../test/cxx11_tensor_cast_float16_cuda.cu    |  6 ++
 unsupported/test/cxx11_tensor_complex_cuda.cu |  6 ++
 .../cxx11_tensor_complex_cwise_ops_cuda.cu    |  6 ++
 .../test/cxx11_tensor_contract_cuda.cu        |  6 ++
 unsupported/test/cxx11_tensor_cuda.cu         |  6 ++
 unsupported/test/cxx11_tensor_device.cu       |  6 ++
 .../test/cxx11_tensor_of_float16_cuda.cu      |  6 ++
 unsupported/test/cxx11_tensor_random_cuda.cu  |  6 ++
 .../test/cxx11_tensor_reduction_cuda.cu       |  6 ++
 unsupported/test/cxx11_tensor_scan_cuda.cu    |  6 ++
 17 files changed, 175 insertions(+), 63 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index f6fe4b0ec..c66359b79 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -22,6 +22,7 @@
   #define EIGEN_CUDA_ARCH __CUDA_ARCH__
 #endif
 
+// Starting with CUDA 9 the composite __CUDACC_VER__ is not available.
 #if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
 #define EIGEN_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
 #elif defined(__CUDACC_VER__)
diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 8cedd65ad..1c557767a 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -50,38 +50,45 @@ struct half;
 namespace half_impl {
 
 #if !defined(EIGEN_HAS_CUDA_FP16)
-
-// Make our own __half definition that is similar to CUDA's.
-struct __half {
-  EIGEN_DEVICE_FUNC __half() : x(0) {}
-  explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
+// Make our own __half_raw definition that is similar to CUDA's.
+struct __half_raw {
+  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
   unsigned short x;
 };
-
+#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+// In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
+typedef __half __half_raw;
 #endif
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
 
-struct half_base : public __half {
+struct half_base : public __half_raw {
   EIGEN_DEVICE_FUNC half_base() {}
-  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
-  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
+#endif
 };
 
 } // namespace half_impl
 
 // Class definition.
 struct half : public half_impl::half_base {
-  #if !defined(EIGEN_HAS_CUDA_FP16)
-    typedef half_impl::__half __half;
+  #if !defined(EIGEN_HAS_CUDA_FP16) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
+    typedef half_impl::__half_raw __half_raw;
   #endif
 
   EIGEN_DEVICE_FUNC half() {}
 
-  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
   EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+#endif
 
   explicit EIGEN_DEVICE_FUNC half(bool b)
       : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
@@ -269,8 +276,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
 // these in hardware. If we need more performance on older/other CPUs, they are
 // also possible to vectorize directly.
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
-  __half h;
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x) {
+  __half_raw h;
   h.x = x;
   return h;
 }
@@ -280,12 +287,13 @@ union FP32 {
   float f;
 };
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
-  return __float2half(ff);
+  __half tmp_ff = __float2half(ff);
+  return *(__half_raw*)&tmp_ff;
 
 #elif defined(EIGEN_HAS_FP16_C)
-  __half h;
+  __half_raw h;
   h.x = _cvtss_sh(ff, 0);
   return h;
 
@@ -296,7 +304,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
   const FP32 f16max = { (127 + 16) << 23 };
   const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
   unsigned int sign_mask = 0x80000000u;
-  __half o;
+  __half_raw o;
   o.x = static_cast<unsigned short>(0x0u);
 
   unsigned int sign = f.u & sign_mask;
@@ -335,7 +343,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
 #endif
 }
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
   return __half2float(h);
 
@@ -612,11 +620,15 @@ struct hash<Eigen::half> {
 // Add the missing shfl_xor intrinsic
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+  #if EIGEN_CUDACC_VER < 90000
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
+  #else
+  return static_cast<Eigen::half>(__shfl_xor_sync(0xFFFFFFFF, static_cast<float>(var), laneMask, width));
+  #endif
 }
 #endif
 
-// ldg() has an overload for __half, but we also need one for Eigen::half.
+// ldg() has an overload for __half_raw, but we also need one for Eigen::half.
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
   return Eigen::half_impl::raw_uint16_to_half(
diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index ba6a7f920..ce48e4b31 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -100,7 +100,8 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2&
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
   half2 result;
-  result.x = a.x & 0x7FFF7FFF;
+  unsigned temp = *(reinterpret_cast<const unsigned*>(&(a)));
+  *(reinterpret_cast<unsigned*>(&(result))) = temp & 0x7FFF7FFF;
   return result;
 }
 
diff --git a/test/half_float.cpp b/test/half_float.cpp
index 7b6208873..7734f82cc 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -20,7 +20,7 @@ using Eigen::half;
 
 void test_conversion()
 {
-  using Eigen::half_impl::__half;
+  using Eigen::half_impl::__half_raw;
 
   // Conversion from float.
   VERIFY_IS_EQUAL(half(1.0f).x, 0x3c00);
@@ -37,9 +37,9 @@ void test_conversion()
   VERIFY_IS_EQUAL(half(1.19209e-07f).x, 0x0002);
 
   // Verify round-to-nearest-even behavior.
-  float val1 = float(half(__half(0x3c00)));
-  float val2 = float(half(__half(0x3c01)));
-  float val3 = float(half(__half(0x3c02)));
+  float val1 = float(half(__half_raw(0x3c00)));
+  float val2 = float(half(__half_raw(0x3c01)));
+  float val3 = float(half(__half_raw(0x3c02)));
   VERIFY_IS_EQUAL(half(0.5f * (val1 + val2)).x, 0x3c00);
   VERIFY_IS_EQUAL(half(0.5f * (val2 + val3)).x, 0x3c02);
 
@@ -55,21 +55,21 @@ void test_conversion()
   VERIFY_IS_EQUAL(half(true).x, 0x3c00);
 
   // Conversion to float.
-  VERIFY_IS_EQUAL(float(half(__half(0x0000))), 0.0f);
-  VERIFY_IS_EQUAL(float(half(__half(0x3c00))), 1.0f);
+  VERIFY_IS_EQUAL(float(half(__half_raw(0x0000))), 0.0f);
+  VERIFY_IS_EQUAL(float(half(__half_raw(0x3c00))), 1.0f);
 
   // Denormals.
-  VERIFY_IS_APPROX(float(half(__half(0x8001))), -5.96046e-08f);
-  VERIFY_IS_APPROX(float(half(__half(0x0001))), 5.96046e-08f);
-  VERIFY_IS_APPROX(float(half(__half(0x0002))), 1.19209e-07f);
+  VERIFY_IS_APPROX(float(half(__half_raw(0x8001))), -5.96046e-08f);
+  VERIFY_IS_APPROX(float(half(__half_raw(0x0001))), 5.96046e-08f);
+  VERIFY_IS_APPROX(float(half(__half_raw(0x0002))), 1.19209e-07f);
 
   // NaNs and infinities.
   VERIFY(!(numext::isinf)(float(half(65504.0f))));  // Largest finite number.
   VERIFY(!(numext::isnan)(float(half(0.0f))));
-  VERIFY((numext::isinf)(float(half(__half(0xfc00)))));
-  VERIFY((numext::isnan)(float(half(__half(0xfc01)))));
-  VERIFY((numext::isinf)(float(half(__half(0x7c00)))));
-  VERIFY((numext::isnan)(float(half(__half(0x7c01)))));
+  VERIFY((numext::isinf)(float(half(__half_raw(0xfc00)))));
+  VERIFY((numext::isnan)(float(half(__half_raw(0xfc01)))));
+  VERIFY((numext::isinf)(float(half(__half_raw(0x7c00)))));
+  VERIFY((numext::isnan)(float(half(__half_raw(0x7c01)))));
 
 #if !EIGEN_COMP_MSVC
   // Visual Studio errors out on divisions by 0
@@ -79,12 +79,12 @@ void test_conversion()
 #endif
 
   // Exactly same checks as above, just directly on the half representation.
-  VERIFY(!(numext::isinf)(half(__half(0x7bff))));
-  VERIFY(!(numext::isnan)(half(__half(0x0000))));
-  VERIFY((numext::isinf)(half(__half(0xfc00))));
-  VERIFY((numext::isnan)(half(__half(0xfc01))));
-  VERIFY((numext::isinf)(half(__half(0x7c00))));
-  VERIFY((numext::isnan)(half(__half(0x7c01))));
+  VERIFY(!(numext::isinf)(half(__half_raw(0x7bff))));
+  VERIFY(!(numext::isnan)(half(__half_raw(0x0000))));
+  VERIFY((numext::isinf)(half(__half_raw(0xfc00))));
+  VERIFY((numext::isnan)(half(__half_raw(0xfc01))));
+  VERIFY((numext::isinf)(half(__half_raw(0x7c00))));
+  VERIFY((numext::isnan)(half(__half_raw(0x7c01))));
 
 #if !EIGEN_COMP_MSVC
   // Visual Studio errors out on divisions by 0
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
index 428b18499..903bc51cc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
@@ -388,7 +388,11 @@ EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
   // the sum across all big k blocks of the product of little k block of index (x, y)
   // with block of index (y, z). To compute the final output, we need to reduce
   // the 8 threads over y by summation.
+#if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
 #define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
+#else
+#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor_sync(0xFFFFFFFF, res(i, j), mask)
+#endif
 
 #define reduceRow(i, mask)                      \
   shuffleInc(i, 0, mask);                       \
@@ -614,8 +618,13 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
       x1 = rhs_pf0.x;
       x2 = rhs_pf0.z;
     }
+    #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
     x1 = __shfl_xor(x1, 4);
     x2 = __shfl_xor(x2, 4);
+    #else
+    x1 = __shfl_xor_sync(0xFFFFFFFF, x1, 4);
+    x2 = __shfl_xor_sync(0xFFFFFFFF, x2, 4);
+    #endif
     if((threadIdx.x%8) < 4) {
       rhs_pf0.y = x1;
       rhs_pf0.w = x2;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index 974eb7deb..ebcbd6f41 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -62,9 +62,9 @@ __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer)
   else {
     assert(0 && "Wordsize not supported");
   }
-#else // __CUDA_ARCH__ >= 300
+#else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif // __CUDA_ARCH__ >= 300
+#endif // EIGEN_CUDA_ARCH >= 300
 }
 
 // We extend atomicExch to support extra data types
@@ -104,9 +104,9 @@ template <>
 __device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
 #if EIGEN_CUDA_ARCH >= 300
   atomicAdd(output, accum);
-#else // __CUDA_ARCH__ >= 300
+#else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif // __CUDA_ARCH__ >= 300
+#endif // EIGEN_CUDA_ARCH >= 300
 }
 
 
@@ -168,7 +168,11 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
 
 #pragma unroll
   for (int offset = warpSize/2; offset > 0; offset /= 2) {
+  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
     reducer.reduce(__shfl_down(accum, offset, warpSize), &accum);
+  #else
+    reducer.reduce(__shfl_down_sync(0xFFFFFFFF, accum, offset, warpSize), &accum);
+  #endif
   }
 
   if ((threadIdx.x & (warpSize - 1)) == 0) {
@@ -179,9 +183,9 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
     // Let the last block reset the semaphore
     atomicInc(semaphore, gridDim.x + 1);
   }
-#else // __CUDA_ARCH__ >= 300
+#else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif // __CUDA_ARCH__ >= 300
+#endif // EIGEN_CUDA_ARCH >= 300
 }
 
 
@@ -223,12 +227,14 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
   const Index first_index = blockIdx.x * BlockSize * NumPerThread + 2*threadIdx.x;
 
   // Initialize the output value if it wasn't initialized by the ReductionInitKernel
-  if (gridDim.x == 1 && first_index == 0) {
-    if (num_coeffs % 2 != 0) {
-      half last = input.m_impl.coeff(num_coeffs-1);
-      *scratch = __halves2half2(last, reducer.initialize());
-    } else {
-      *scratch = reducer.template initializePacket<half2>();
+  if (gridDim.x == 1) {
+    if (first_index == 0) {
+      if (num_coeffs % 2 != 0) {
+        half last = input.m_impl.coeff(num_coeffs-1);
+        *scratch = __halves2half2(last, reducer.initialize());
+      } else {
+        *scratch = reducer.template initializePacket<half2>();
+      }
     }
     __syncthreads();
   }
@@ -244,19 +250,25 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
 
 #pragma unroll
   for (int offset = warpSize/2; offset > 0; offset /= 2) {
+  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
     reducer.reducePacket(__shfl_down(accum, offset, warpSize), &accum);
+  #else
+    int temp = __shfl_down_sync(0xFFFFFFFF, *(int*)(&accum), (unsigned)offset, warpSize);
+    reducer.reducePacket(*(half2*)(&temp), &accum);
+  #endif
   }
 
   if ((threadIdx.x & (warpSize - 1)) == 0) {
     atomicReduce(scratch, accum, reducer);
   }
 
-  __syncthreads();
-
-  if (gridDim.x == 1 && first_index == 0) {
-    half tmp = __low2half(*scratch);
-    reducer.reduce(__high2half(*scratch), &tmp);
-    *output = tmp;
+  if (gridDim.x == 1) {
+    __syncthreads();
+    if (first_index == 0) {
+      half tmp = __low2half(*scratch);
+      reducer.reduce(__high2half(*scratch), &tmp);
+      *output = tmp;
+    }
   }
 }
 
@@ -425,7 +437,11 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
 
 #pragma unroll
       for (int offset = warpSize/2; offset > 0; offset /= 2) {
+      #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
         reducer.reduce(__shfl_down(reduced_val, offset), &reduced_val);
+      #else
+        reducer.reduce(__shfl_down_sync(0xFFFFFFFF, reduced_val, offset), &reduced_val);
+      #endif
       }
 
       if ((threadIdx.x & (warpSize - 1)) == 0) {
@@ -433,9 +449,9 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
       }
     }
   }
-#else // __CUDA_ARCH__ >= 300
+#else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
-#endif // __CUDA_ARCH__ >= 300
+#endif // EIGEN_CUDA_ARCH >= 300
 }
 
 #ifdef EIGEN_HAS_CUDA_FP16
@@ -515,8 +531,15 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
 
 #pragma unroll
       for (int offset = warpSize/2; offset > 0; offset /= 2) {
+      #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
         reducer.reducePacket(__shfl_down(reduced_val1, offset, warpSize), &reduced_val1);
         reducer.reducePacket(__shfl_down(reduced_val2, offset, warpSize), &reduced_val2);
+      #else
+        int temp1 = __shfl_down_sync(0xFFFFFFFF, *(int*)(&reduced_val1), (unsigned)offset, warpSize);
+        int temp2 = __shfl_down_sync(0xFFFFFFFF, *(int*)(&reduced_val2), (unsigned)offset, warpSize);
+        reducer.reducePacket(*(half2*)(&temp1), &reduced_val1);
+        reducer.reducePacket(*(half2*)(&temp2), &reduced_val2);
+      #endif
       }
 
       half val1 =  __low2half(reduced_val1);
diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_cuda.cu
index 3d73d491a..0e8b8125d 100644
--- a/unsupported/test/cxx11_tensor_argmax_cuda.cu
+++ b/unsupported/test/cxx11_tensor_argmax_cuda.cu
@@ -15,6 +15,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 template <int Layout>
diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
index 816e03220..dabf9e45f 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
@@ -16,6 +16,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 void test_cuda_conversion() {
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index a52350f85..d25e1bee1 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -14,6 +14,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 void test_cuda_nullary() {
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
index aac780905..4f0f621b4 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
@@ -14,6 +14,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 template<typename T>
diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_cuda.cu
index e821ccf0c..c68287e34 100644
--- a/unsupported/test/cxx11_tensor_contract_cuda.cu
+++ b/unsupported/test/cxx11_tensor_contract_cuda.cu
@@ -17,6 +17,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;
 
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index 9584a539f..d9059a2dc 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -15,6 +15,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 void test_cuda_nullary() {
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index cbb43e210..d5bfeeb39 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -16,6 +16,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 using Eigen::RowMajor;
 
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index b3aab0b9d..c9f3ae1ae 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -16,6 +16,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 
 template<typename>
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu
index fa1a46732..9d08605fc 100644
--- a/unsupported/test/cxx11_tensor_random_cuda.cu
+++ b/unsupported/test/cxx11_tensor_random_cuda.cu
@@ -16,6 +16,12 @@
 #include "main.h"
 #include <Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 
 void test_cuda_random_uniform()
 {
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_cuda.cu
index ec0669704..d6ce04f1c 100644
--- a/unsupported/test/cxx11_tensor_reduction_cuda.cu
+++ b/unsupported/test/cxx11_tensor_reduction_cuda.cu
@@ -15,6 +15,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 
 template<typename Type, int DataLayout>
 static void test_full_reductions() {
diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_cuda.cu
index de1c0ac95..e99724b91 100644
--- a/unsupported/test/cxx11_tensor_scan_cuda.cu
+++ b/unsupported/test/cxx11_tensor_scan_cuda.cu
@@ -16,6 +16,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+// The EIGEN_CUDACC_VER macro is provided by
+// unsupported/Eigen/CXX11/Tensor included above
+#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
+#include <cuda_fp16.h>
+#endif
+
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;
 

From a991c80365372f4cba2caddbd2d4c39352144793 Mon Sep 17 00:00:00 2001
From: Jonas Harsch <jonas.harsch@gmail.com>
Date: Fri, 1 Sep 2017 11:30:26 +0000
Subject: [PATCH 182/680] Added an example for a contraction to a scalar value,
 e.g. a double contraction of two second order tensors and how you can get the
 value of the result. I lost one day to get this doen so I think it will help
 some guys. I also added Eigen:: to the IndexPair and and array in the same
 example.

---
 unsupported/Eigen/CXX11/src/Tensor/README.md | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index 4423f81f7..30d553af7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -1016,13 +1016,20 @@ multidimensional case.
     a.setValues({{1, 2}, {4, 5}, {5, 6}});
 
     // Compute the traditional matrix product
-    array<IndexPair<int>, 1> product_dims = { IndexPair(1, 0) };
+    Eigen::array<Eigen::IndexPair<int>, 1> product_dims = { Eigen::IndexPair(1, 0) };
     Eigen::Tensor<int, 2> AB = a.contract(b, product_dims);
 
     // Compute the product of the transpose of the matrices
-    array<IndexPair<int>, 1> transpose_product_dims = { IndexPair(0, 1) };
+    Eigen::array<Eigen::IndexPair<int>, 1> transpose_product_dims = { Eigen::IndexPair(0, 1) };
     Eigen::Tensor<int, 2> AtBt = a.contract(b, transposed_product_dims);
-
+    
+    // Contraction to scalar value using a ouble contraction 
+    // First coordinate of both tensors are contracted as well as both second coordinates
+    Eigen::array<Eigen::IndexPair<int>, 2> double_contraction_product_dims = { Eigen::IndexPair<int>(0, 0), Eigen::IndexPair<int>(1, 1) };
+    Eigen::Tensor<int, 0> AdoubleontractedA = a.contract(a, double_contraction_product_dims);
+    
+    // Extracting the scalar value of the tensor contraction for further usage
+    int value = AdoublecontractedA(0);
 
 ## Reduction Operations
 

From a34fb212cd7ea96524d859a0c7f989f8657bc068 Mon Sep 17 00:00:00 2001
From: Jonas Harsch <jonas.harsch@gmail.com>
Date: Fri, 1 Sep 2017 12:01:39 +0000
Subject: [PATCH 183/680] Close branch
 JonasMu/added-an-example-for-a-contraction-to-a--1504265366851


From 80142362ac35ca77bfc5ccf7ba49c9f034b57abc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 2 Sep 2017 22:50:20 +0200
Subject: [PATCH 184/680] Fix mixing types in sparse matrix products.

---
 .../ConservativeSparseSparseProduct.h         | 67 ++++++++-------
 .../SparseSparseProductWithPruning.h          | 22 ++---
 test/sparse_product.cpp                       | 84 +++++++++++++++++++
 3 files changed, 132 insertions(+), 41 deletions(-)

diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
index 492eb0a29..9db119b67 100644
--- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
+++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
@@ -17,7 +17,9 @@ namespace internal {
 template<typename Lhs, typename Rhs, typename ResultType>
 static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false)
 {
-  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+  typedef typename remove_all<Lhs>::type::Scalar LhsScalar;
+  typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
+  typedef typename remove_all<ResultType>::type::Scalar ResScalar;
 
   // make sure to call innerSize/outerSize since we fake the storage order.
   Index rows = lhs.innerSize();
@@ -25,7 +27,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
   eigen_assert(lhs.outerSize() == rhs.innerSize());
   
   ei_declare_aligned_stack_constructed_variable(bool,   mask,     rows, 0);
-  ei_declare_aligned_stack_constructed_variable(Scalar, values,   rows, 0);
+  ei_declare_aligned_stack_constructed_variable(ResScalar, values,   rows, 0);
   ei_declare_aligned_stack_constructed_variable(Index,  indices,  rows, 0);
   
   std::memset(mask,0,sizeof(bool)*rows);
@@ -51,12 +53,12 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
     Index nnz = 0;
     for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
     {
-      Scalar y = rhsIt.value();
+      RhsScalar y = rhsIt.value();
       Index k = rhsIt.index();
       for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
       {
         Index i = lhsIt.index();
-        Scalar x = lhsIt.value();
+        LhsScalar x = lhsIt.value();
         if(!mask[i])
         {
           mask[i] = true;
@@ -166,11 +168,12 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,C
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-     typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
-     RowMajorMatrix rhsRow = rhs;
-     RowMajorMatrix resRow(lhs.rows(), rhs.cols());
-     internal::conservative_sparse_sparse_product_impl<RowMajorMatrix,Lhs,RowMajorMatrix>(rhsRow, lhs, resRow);
-     res = resRow;
+    typedef SparseMatrix<typename Rhs::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRhs;
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRes;
+    RowMajorRhs rhsRow = rhs;
+    RowMajorRes resRow(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<RowMajorRhs,Lhs,RowMajorRes>(rhsRow, lhs, resRow);
+    res = resRow;
   }
 };
 
@@ -179,10 +182,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,R
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
-    RowMajorMatrix lhsRow = lhs;
-    RowMajorMatrix resRow(lhs.rows(), rhs.cols());
-    internal::conservative_sparse_sparse_product_impl<Rhs,RowMajorMatrix,RowMajorMatrix>(rhs, lhsRow, resRow);
+    typedef SparseMatrix<typename Lhs::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorLhs;
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRes;
+    RowMajorLhs lhsRow = lhs;
+    RowMajorRes resRow(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Rhs,RowMajorLhs,RowMajorRes>(rhs, lhsRow, resRow);
     res = resRow;
   }
 };
@@ -219,10 +223,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,C
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
-    ColMajorMatrix lhsCol = lhs;
-    ColMajorMatrix resCol(lhs.rows(), rhs.cols());
-    internal::conservative_sparse_sparse_product_impl<ColMajorMatrix,Rhs,ColMajorMatrix>(lhsCol, rhs, resCol);
+    typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorLhs;
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRes;
+    ColMajorLhs lhsCol = lhs;
+    ColMajorRes resCol(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<ColMajorLhs,Rhs,ColMajorRes>(lhsCol, rhs, resCol);
     res = resCol;
   }
 };
@@ -232,10 +237,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,R
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
-    ColMajorMatrix rhsCol = rhs;
-    ColMajorMatrix resCol(lhs.rows(), rhs.cols());
-    internal::conservative_sparse_sparse_product_impl<Lhs,ColMajorMatrix,ColMajorMatrix>(lhs, rhsCol, resCol);
+    typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRhs;
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRes;
+    ColMajorRhs rhsCol = rhs;
+    ColMajorRes resCol(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Lhs,ColMajorRhs,ColMajorRes>(lhs, rhsCol, resCol);
     res = resCol;
   }
 };
@@ -263,7 +269,8 @@ namespace internal {
 template<typename Lhs, typename Rhs, typename ResultType>
 static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res)
 {
-  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+  typedef typename remove_all<Lhs>::type::Scalar LhsScalar;
+  typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
   Index cols = rhs.outerSize();
   eigen_assert(lhs.outerSize() == rhs.innerSize());
 
@@ -274,12 +281,12 @@ static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs,
   {
     for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
     {
-      Scalar y = rhsIt.value();
+      RhsScalar y = rhsIt.value();
       Index k = rhsIt.index();
       for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
       {
         Index i = lhsIt.index();
-        Scalar x = lhsIt.value();
+        LhsScalar x = lhsIt.value();
         res.coeffRef(i,j) += x * y;
       }
     }
@@ -310,9 +317,9 @@ struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,RowMajor,ColMa
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
-    ColMajorMatrix lhsCol(lhs);
-    internal::sparse_sparse_to_dense_product_impl<ColMajorMatrix,Rhs,ResultType>(lhsCol, rhs, res);
+    typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorLhs;
+    ColMajorLhs lhsCol(lhs);
+    internal::sparse_sparse_to_dense_product_impl<ColMajorLhs,Rhs,ResultType>(lhsCol, rhs, res);
   }
 };
 
@@ -321,9 +328,9 @@ struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,ColMajor,RowMa
 {
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
-    ColMajorMatrix rhsCol(rhs);
-    internal::sparse_sparse_to_dense_product_impl<Lhs,ColMajorMatrix,ResultType>(lhs, rhsCol, res);
+    typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRhs;
+    ColMajorRhs rhsCol(rhs);
+    internal::sparse_sparse_to_dense_product_impl<Lhs,ColMajorRhs,ResultType>(lhs, rhsCol, res);
   }
 };
 
diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h
index 21c419002..88820a48f 100644
--- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h
+++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h
@@ -21,7 +21,8 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
 {
   // return sparse_sparse_product_with_pruning_impl2(lhs,rhs,res);
 
-  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+  typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
+  typedef typename remove_all<ResultType>::type::Scalar ResScalar;
   typedef typename remove_all<Lhs>::type::StorageIndex StorageIndex;
 
   // make sure to call innerSize/outerSize since we fake the storage order.
@@ -31,7 +32,7 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
   eigen_assert(lhs.outerSize() == rhs.innerSize());
 
   // allocate a temporary buffer
-  AmbiVector<Scalar,StorageIndex> tempVector(rows);
+  AmbiVector<ResScalar,StorageIndex> tempVector(rows);
 
   // mimics a resizeByInnerOuter:
   if(ResultType::IsRowMajor)
@@ -63,14 +64,14 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
     {
       // FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index())
       tempVector.restart();
-      Scalar x = rhsIt.value();
+      RhsScalar x = rhsIt.value();
       for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt)
       {
         tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x;
       }
     }
     res.startVec(j);
-    for (typename AmbiVector<Scalar,StorageIndex>::Iterator it(tempVector,tolerance); it; ++it)
+    for (typename AmbiVector<ResScalar,StorageIndex>::Iterator it(tempVector,tolerance); it; ++it)
       res.insertBackByOuterInner(j,it.index()) = it.value();
   }
   res.finalize();
@@ -85,7 +86,6 @@ struct sparse_sparse_product_with_pruning_selector;
 template<typename Lhs, typename Rhs, typename ResultType>
 struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,ColMajor>
 {
-  typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
   typedef typename ResultType::RealScalar RealScalar;
 
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
@@ -129,8 +129,8 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,R
   typedef typename ResultType::RealScalar RealScalar;
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
+    typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
+    typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
     ColMajorMatrixLhs colLhs(lhs);
     ColMajorMatrixRhs colRhs(rhs);
     internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,ColMajorMatrixRhs,ResultType>(colLhs, colRhs, res, tolerance);
@@ -149,7 +149,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,R
   typedef typename ResultType::RealScalar RealScalar;
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixLhs;
+    typedef SparseMatrix<typename Lhs::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixLhs;
     RowMajorMatrixLhs rowLhs(lhs);
     sparse_sparse_product_with_pruning_selector<RowMajorMatrixLhs,Rhs,ResultType,RowMajor,RowMajor>(rowLhs,rhs,res,tolerance);
   }
@@ -161,7 +161,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,C
   typedef typename ResultType::RealScalar RealScalar;
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixRhs;
+    typedef SparseMatrix<typename Rhs::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixRhs;
     RowMajorMatrixRhs rowRhs(rhs);
     sparse_sparse_product_with_pruning_selector<Lhs,RowMajorMatrixRhs,ResultType,RowMajor,RowMajor,RowMajor>(lhs,rowRhs,res,tolerance);
   }
@@ -173,7 +173,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,R
   typedef typename ResultType::RealScalar RealScalar;
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
+    typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
     ColMajorMatrixRhs colRhs(rhs);
     internal::sparse_sparse_product_with_pruning_impl<Lhs,ColMajorMatrixRhs,ResultType>(lhs, colRhs, res, tolerance);
   }
@@ -185,7 +185,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,C
   typedef typename ResultType::RealScalar RealScalar;
   static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
   {
-    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
+    typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
     ColMajorMatrixLhs colLhs(lhs);
     internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,Rhs,ResultType>(colLhs, rhs, res, tolerance);
   }
diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp
index 197586741..f47170b72 100644
--- a/test/sparse_product.cpp
+++ b/test/sparse_product.cpp
@@ -371,6 +371,88 @@ void bug_942()
   VERIFY_IS_APPROX( ( d.asDiagonal()*cmA ).eval().coeff(0,0), res );
 }
 
+template<typename Real>
+void test_mixing_types()
+{
+  typedef std::complex<Real> Cplx;
+  typedef SparseMatrix<Real> SpMatReal;
+  typedef SparseMatrix<Cplx> SpMatCplx;
+  typedef SparseMatrix<Cplx,RowMajor> SpRowMatCplx;
+  typedef Matrix<Real,Dynamic,Dynamic> DenseMatReal;
+  typedef Matrix<Cplx,Dynamic,Dynamic> DenseMatCplx;
+
+  Index n = internal::random<Index>(1,100);
+  double density = (std::max)(8./(n*n), 0.2);
+
+  SpMatReal sR1(n,n);
+  SpMatCplx sC1(n,n), sC2(n,n), sC3(n,n);
+  SpRowMatCplx sCR(n,n);
+  DenseMatReal dR1(n,n);
+  DenseMatCplx dC1(n,n), dC2(n,n), dC3(n,n);
+
+  initSparse<Real>(density, dR1, sR1);
+  initSparse<Cplx>(density, dC1, sC1);
+  initSparse<Cplx>(density, dC2, sC2);
+
+  VERIFY_IS_APPROX( sC2 = (sR1 * sC1),                         dC3 = dR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( sC2 = (sC1 * sR1),                         dC3 = dC1 * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sC2 = (sR1.transpose() * sC1),             dC3 = dR1.template cast<Cplx>().transpose() * dC1 );
+  VERIFY_IS_APPROX( sC2 = (sC1.transpose() * sR1),             dC3 = dC1.transpose() * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sC2 = (sR1 * sC1.transpose()),             dC3 = dR1.template cast<Cplx>() * dC1.transpose() );
+  VERIFY_IS_APPROX( sC2 = (sC1 * sR1.transpose()),             dC3 = dC1 * dR1.template cast<Cplx>().transpose() );
+  VERIFY_IS_APPROX( sC2 = (sR1.transpose() * sC1.transpose()), dC3 = dR1.template cast<Cplx>().transpose() * dC1.transpose() );
+  VERIFY_IS_APPROX( sC2 = (sC1.transpose() * sR1.transpose()), dC3 = dC1.transpose() * dR1.template cast<Cplx>().transpose() );
+
+  VERIFY_IS_APPROX( sCR = (sR1 * sC1),                         dC3 = dR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( sCR = (sC1 * sR1),                         dC3 = dC1 * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sCR = (sR1.transpose() * sC1),             dC3 = dR1.template cast<Cplx>().transpose() * dC1 );
+  VERIFY_IS_APPROX( sCR = (sC1.transpose() * sR1),             dC3 = dC1.transpose() * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sCR = (sR1 * sC1.transpose()),             dC3 = dR1.template cast<Cplx>() * dC1.transpose() );
+  VERIFY_IS_APPROX( sCR = (sC1 * sR1.transpose()),             dC3 = dC1 * dR1.template cast<Cplx>().transpose() );
+  VERIFY_IS_APPROX( sCR = (sR1.transpose() * sC1.transpose()), dC3 = dR1.template cast<Cplx>().transpose() * dC1.transpose() );
+  VERIFY_IS_APPROX( sCR = (sC1.transpose() * sR1.transpose()), dC3 = dC1.transpose() * dR1.template cast<Cplx>().transpose() );
+
+
+  VERIFY_IS_APPROX( sC2 = (sR1 * sC1).pruned(),                         dC3 = dR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( sC2 = (sC1 * sR1).pruned(),                         dC3 = dC1 * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sC2 = (sR1.transpose() * sC1).pruned(),             dC3 = dR1.template cast<Cplx>().transpose() * dC1 );
+  VERIFY_IS_APPROX( sC2 = (sC1.transpose() * sR1).pruned(),             dC3 = dC1.transpose() * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sC2 = (sR1 * sC1.transpose()).pruned(),             dC3 = dR1.template cast<Cplx>() * dC1.transpose() );
+  VERIFY_IS_APPROX( sC2 = (sC1 * sR1.transpose()).pruned(),             dC3 = dC1 * dR1.template cast<Cplx>().transpose() );
+  VERIFY_IS_APPROX( sC2 = (sR1.transpose() * sC1.transpose()).pruned(), dC3 = dR1.template cast<Cplx>().transpose() * dC1.transpose() );
+  VERIFY_IS_APPROX( sC2 = (sC1.transpose() * sR1.transpose()).pruned(), dC3 = dC1.transpose() * dR1.template cast<Cplx>().transpose() );
+
+  VERIFY_IS_APPROX( sCR = (sR1 * sC1).pruned(),                         dC3 = dR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( sCR = (sC1 * sR1).pruned(),                         dC3 = dC1 * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sCR = (sR1.transpose() * sC1).pruned(),             dC3 = dR1.template cast<Cplx>().transpose() * dC1 );
+  VERIFY_IS_APPROX( sCR = (sC1.transpose() * sR1).pruned(),             dC3 = dC1.transpose() * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( sCR = (sR1 * sC1.transpose()).pruned(),             dC3 = dR1.template cast<Cplx>() * dC1.transpose() );
+  VERIFY_IS_APPROX( sCR = (sC1 * sR1.transpose()).pruned(),             dC3 = dC1 * dR1.template cast<Cplx>().transpose() );
+  VERIFY_IS_APPROX( sCR = (sR1.transpose() * sC1.transpose()).pruned(), dC3 = dR1.template cast<Cplx>().transpose() * dC1.transpose() );
+  VERIFY_IS_APPROX( sCR = (sC1.transpose() * sR1.transpose()).pruned(), dC3 = dC1.transpose() * dR1.template cast<Cplx>().transpose() );
+
+
+  VERIFY_IS_APPROX( dC2 = (sR1 * sC1),                         dC3 = dR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( dC2 = (sC1 * sR1),                         dC3 = dC1 * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( dC2 = (sR1.transpose() * sC1),             dC3 = dR1.template cast<Cplx>().transpose() * dC1 );
+  VERIFY_IS_APPROX( dC2 = (sC1.transpose() * sR1),             dC3 = dC1.transpose() * dR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( dC2 = (sR1 * sC1.transpose()),             dC3 = dR1.template cast<Cplx>() * dC1.transpose() );
+  VERIFY_IS_APPROX( dC2 = (sC1 * sR1.transpose()),             dC3 = dC1 * dR1.template cast<Cplx>().transpose() );
+  VERIFY_IS_APPROX( dC2 = (sR1.transpose() * sC1.transpose()), dC3 = dR1.template cast<Cplx>().transpose() * dC1.transpose() );
+  VERIFY_IS_APPROX( dC2 = (sC1.transpose() * sR1.transpose()), dC3 = dC1.transpose() * dR1.template cast<Cplx>().transpose() );
+
+
+  VERIFY_IS_APPROX( dC2 = dR1 * sC1, dC3 = dR1.template cast<Cplx>() * sC1 );
+  VERIFY_IS_APPROX( dC2 = sR1 * dC1, dC3 = sR1.template cast<Cplx>() * dC1 );
+  VERIFY_IS_APPROX( dC2 = dC1 * sR1, dC3 = dC1 * sR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( dC2 = sC1 * dR1, dC3 = sC1 * dR1.template cast<Cplx>() );
+
+  VERIFY_IS_APPROX( dC2 = dR1.row(0) * sC1, dC3 = dR1.template cast<Cplx>().row(0) * sC1 );
+  VERIFY_IS_APPROX( dC2 = sR1 * dC1.col(0), dC3 = sR1.template cast<Cplx>() * dC1.col(0) );
+  VERIFY_IS_APPROX( dC2 = dC1.row(0) * sR1, dC3 = dC1.row(0) * sR1.template cast<Cplx>() );
+  VERIFY_IS_APPROX( dC2 = sC1 * dR1.col(0), dC3 = sC1 * dR1.template cast<Cplx>().col(0) );
+}
+
 void test_sparse_product()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -381,5 +463,7 @@ void test_sparse_product()
     CALL_SUBTEST_2( (sparse_product<SparseMatrix<std::complex<double>, RowMajor > >()) );
     CALL_SUBTEST_3( (sparse_product<SparseMatrix<float,ColMajor,long int> >()) );
     CALL_SUBTEST_4( (sparse_product_regression_test<SparseMatrix<double,RowMajor>, Matrix<double, Dynamic, Dynamic, RowMajor> >()) );
+
+    CALL_SUBTEST_5( (test_mixing_types<float>()) );
   }
 }

From b35d1ce4a532a40e16927bb303a4c48b2bef1d7c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 6 Sep 2017 10:02:49 +0200
Subject: [PATCH 185/680] Implement true compile-time "if" for
 apply_rotation_in_the_plane. This fixes a compilation issue for vectorized
 real type with missing vectorization for complexes, e.g. AVX512.

---
 Eigen/src/Jacobi/Jacobi.h | 256 +++++++++++++++++++++-----------------
 1 file changed, 139 insertions(+), 117 deletions(-)

diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index 75595517a..af1228cb8 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -309,16 +309,144 @@ inline void MatrixBase<Derived>::applyOnTheRight(Index p, Index q, const JacobiR
 }
 
 namespace internal {
+
+template<typename Scalar, typename OtherScalar,
+         int SizeAtCompileTime, int MinAlignment, bool Vectorizable>
+struct apply_rotation_in_the_plane_selector
+{
+  static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
+  {
+    for(Index i=0; i<size; ++i)
+    {
+      Scalar xi = *x;
+      Scalar yi = *y;
+      *x =  c * xi + numext::conj(s) * yi;
+      *y = -s * xi + numext::conj(c) * yi;
+      x += incrx;
+      y += incry;
+    }
+  }
+};
+
+template<typename Scalar, typename OtherScalar,
+         int SizeAtCompileTime, int MinAlignment>
+struct apply_rotation_in_the_plane_selector<Scalar,OtherScalar,SizeAtCompileTime,MinAlignment,true /* vectorizable */>
+{
+  static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
+  {
+    enum {
+      PacketSize = packet_traits<Scalar>::size,
+      OtherPacketSize = packet_traits<OtherScalar>::size
+    };
+    typedef typename packet_traits<Scalar>::type Packet;
+    typedef typename packet_traits<OtherScalar>::type OtherPacket;
+
+    /*** dynamic-size vectorized paths ***/
+    if(SizeAtCompileTime == Dynamic && ((incrx==1 && incry==1) || PacketSize == 1))
+    {
+      // both vectors are sequentially stored in memory => vectorization
+      enum { Peeling = 2 };
+
+      Index alignedStart = internal::first_default_aligned(y, size);
+      Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
+
+      const OtherPacket pc = pset1<OtherPacket>(c);
+      const OtherPacket ps = pset1<OtherPacket>(s);
+      conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
+      conj_helper<OtherPacket,Packet,false,false> pm;
+
+      for(Index i=0; i<alignedStart; ++i)
+      {
+        Scalar xi = x[i];
+        Scalar yi = y[i];
+        x[i] =  c * xi + numext::conj(s) * yi;
+        y[i] = -s * xi + numext::conj(c) * yi;
+      }
+
+      Scalar* EIGEN_RESTRICT px = x + alignedStart;
+      Scalar* EIGEN_RESTRICT py = y + alignedStart;
+
+      if(internal::first_default_aligned(x, size)==alignedStart)
+      {
+        for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
+        {
+          Packet xi = pload<Packet>(px);
+          Packet yi = pload<Packet>(py);
+          pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+          pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+          px += PacketSize;
+          py += PacketSize;
+        }
+      }
+      else
+      {
+        Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
+        for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
+        {
+          Packet xi   = ploadu<Packet>(px);
+          Packet xi1  = ploadu<Packet>(px+PacketSize);
+          Packet yi   = pload <Packet>(py);
+          Packet yi1  = pload <Packet>(py+PacketSize);
+          pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+          pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
+          pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+          pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
+          px += Peeling*PacketSize;
+          py += Peeling*PacketSize;
+        }
+        if(alignedEnd!=peelingEnd)
+        {
+          Packet xi = ploadu<Packet>(x+peelingEnd);
+          Packet yi = pload <Packet>(y+peelingEnd);
+          pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+          pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+        }
+      }
+
+      for(Index i=alignedEnd; i<size; ++i)
+      {
+        Scalar xi = x[i];
+        Scalar yi = y[i];
+        x[i] =  c * xi + numext::conj(s) * yi;
+        y[i] = -s * xi + numext::conj(c) * yi;
+      }
+    }
+
+    /*** fixed-size vectorized path ***/
+    else if(SizeAtCompileTime != Dynamic && MinAlignment>0) // FIXME should be compared to the required alignment
+    {
+      const OtherPacket pc = pset1<OtherPacket>(c);
+      const OtherPacket ps = pset1<OtherPacket>(s);
+      conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
+      conj_helper<OtherPacket,Packet,false,false> pm;
+      Scalar* EIGEN_RESTRICT px = x;
+      Scalar* EIGEN_RESTRICT py = y;
+      for(Index i=0; i<size; i+=PacketSize)
+      {
+        Packet xi = pload<Packet>(px);
+        Packet yi = pload<Packet>(py);
+        pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+        px += PacketSize;
+        py += PacketSize;
+      }
+    }
+
+    /*** non-vectorized path ***/
+    else
+    {
+      apply_rotation_in_the_plane_selector<Scalar,OtherScalar,SizeAtCompileTime,MinAlignment,false>::run(x,incrx,y,incry,size,c,s);
+    }
+  }
+};
+
 template<typename VectorX, typename VectorY, typename OtherScalar>
 void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j)
 {
   typedef typename VectorX::Scalar Scalar;
-  enum {
-    PacketSize = packet_traits<Scalar>::size,
-    OtherPacketSize = packet_traits<OtherScalar>::size
-  };
-  typedef typename packet_traits<Scalar>::type Packet;
-  typedef typename packet_traits<OtherScalar>::type OtherPacket;
+  const bool Vectorizable =    (VectorX::Flags & VectorY::Flags & PacketAccessBit)
+                            && (int(packet_traits<Scalar>::size) == int(packet_traits<OtherScalar>::size));
+
   eigen_assert(xpr_x.size() == xpr_y.size());
   Index size = xpr_x.size();
   Index incrx = xpr_x.derived().innerStride();
@@ -332,117 +460,11 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
   if (c==OtherScalar(1) && s==OtherScalar(0))
     return;
 
-  /*** dynamic-size vectorized paths ***/
-
-  if(VectorX::SizeAtCompileTime == Dynamic &&
-    (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
-    (PacketSize == OtherPacketSize) &&
-    ((incrx==1 && incry==1) || PacketSize == 1))
-  {
-    // both vectors are sequentially stored in memory => vectorization
-    enum { Peeling = 2 };
-
-    Index alignedStart = internal::first_default_aligned(y, size);
-    Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
-
-    const OtherPacket pc = pset1<OtherPacket>(c);
-    const OtherPacket ps = pset1<OtherPacket>(s);
-    conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
-    conj_helper<OtherPacket,Packet,false,false> pm;
-
-    for(Index i=0; i<alignedStart; ++i)
-    {
-      Scalar xi = x[i];
-      Scalar yi = y[i];
-      x[i] =  c * xi + numext::conj(s) * yi;
-      y[i] = -s * xi + numext::conj(c) * yi;
-    }
-
-    Scalar* EIGEN_RESTRICT px = x + alignedStart;
-    Scalar* EIGEN_RESTRICT py = y + alignedStart;
-
-    if(internal::first_default_aligned(x, size)==alignedStart)
-    {
-      for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
-      {
-        Packet xi = pload<Packet>(px);
-        Packet yi = pload<Packet>(py);
-        pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
-        px += PacketSize;
-        py += PacketSize;
-      }
-    }
-    else
-    {
-      Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
-      for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
-      {
-        Packet xi   = ploadu<Packet>(px);
-        Packet xi1  = ploadu<Packet>(px+PacketSize);
-        Packet yi   = pload <Packet>(py);
-        Packet yi1  = pload <Packet>(py+PacketSize);
-        pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
-        pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
-        pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
-        px += Peeling*PacketSize;
-        py += Peeling*PacketSize;
-      }
-      if(alignedEnd!=peelingEnd)
-      {
-        Packet xi = ploadu<Packet>(x+peelingEnd);
-        Packet yi = pload <Packet>(y+peelingEnd);
-        pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
-        pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
-      }
-    }
-
-    for(Index i=alignedEnd; i<size; ++i)
-    {
-      Scalar xi = x[i];
-      Scalar yi = y[i];
-      x[i] =  c * xi + numext::conj(s) * yi;
-      y[i] = -s * xi + numext::conj(c) * yi;
-    }
-  }
-
-  /*** fixed-size vectorized path ***/
-  else if(VectorX::SizeAtCompileTime != Dynamic &&
-          (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
-          (PacketSize == OtherPacketSize) &&
-          (EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment)>0)) // FIXME should be compared to the required alignment
-  {
-    const OtherPacket pc = pset1<OtherPacket>(c);
-    const OtherPacket ps = pset1<OtherPacket>(s);
-    conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
-    conj_helper<OtherPacket,Packet,false,false> pm;
-    Scalar* EIGEN_RESTRICT px = x;
-    Scalar* EIGEN_RESTRICT py = y;
-    for(Index i=0; i<size; i+=PacketSize)
-    {
-      Packet xi = pload<Packet>(px);
-      Packet yi = pload<Packet>(py);
-      pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
-      pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
-      px += PacketSize;
-      py += PacketSize;
-    }
-  }
-
-  /*** non-vectorized path ***/
-  else
-  {
-    for(Index i=0; i<size; ++i)
-    {
-      Scalar xi = *x;
-      Scalar yi = *y;
-      *x =  c * xi + numext::conj(s) * yi;
-      *y = -s * xi + numext::conj(c) * yi;
-      x += incrx;
-      y += incry;
-    }
-  }
+  apply_rotation_in_the_plane_selector<
+    Scalar,OtherScalar,
+    VectorX::SizeAtCompileTime,
+    EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment),
+    Vectorizable>::run(x,incrx,y,incry,size,c,s);
 }
 
 } // end namespace internal

From 9c353dd1450c362175213c1018b15b24ac840826 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 6 Sep 2017 10:22:47 +0200
Subject: [PATCH 186/680] Add C++11 max_digits10 for half.

---
 Eigen/src/Core/arch/CUDA/Half.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 1c557767a..ee24e615a 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -520,8 +520,8 @@ struct numeric_limits<Eigen::half> {
   static const bool is_bounded = false;
   static const bool is_modulo = false;
   static const int digits = 11;
-  static const int digits10 = 2;
-  //static const int max_digits10 = ;
+  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
   static const int radix = 2;
   static const int min_exponent = -13;
   static const int min_exponent10 = -4;

From ea4e65bf413fd5e60cd13407879d7ccaae62f6fc Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 7 Sep 2017 09:13:52 +0000
Subject: [PATCH 187/680] Fixed compilation with cuda_clang.

---
 Eigen/src/Core/GeneralProduct.h | 1 +
 Eigen/src/Core/NoAlias.h        | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 483277fe6..694f7cbde 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -396,6 +396,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
   */
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 inline const Product<Derived, OtherDerived>
 MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 {
diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h
index 41fae5096..e94c8ee96 100644
--- a/Eigen/src/Core/NoAlias.h
+++ b/Eigen/src/Core/NoAlias.h
@@ -99,7 +99,7 @@ class NoAlias
   * \sa class NoAlias
   */
 template<typename Derived>
-NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
+NoAlias<Derived,MatrixBase> EIGEN_DEVICE_FUNC MatrixBase<Derived>::noalias()
 {
   return NoAlias<Derived, Eigen::MatrixBase >(derived());
 }

From 6d42309f1394396c984212abad73469e6a50c3a3 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Sep 2017 14:34:30 +0200
Subject: [PATCH 188/680] Fix compilation of Vector::operator()(enum) by
 treating enums as Index

---
 Eigen/src/Core/util/Macros.h           | 10 ++++++++++
 Eigen/src/Core/util/XprHelper.h        | 10 ++++++++++
 Eigen/src/plugins/IndexedViewMethods.h | 10 ++++------
 test/indexed_view.cpp                  |  5 +++++
 4 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index b63ea2697..caedb0b3a 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -410,6 +410,16 @@
 #endif
 #endif
 
+// Does the compiler support type_trais?
+#ifndef EIGEN_HAS_TYPE_TRAITS
+#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || __has_feature(is_enum) || EIGEN_COMP_MSVC >= 1700)
+#define EIGEN_HAS_TYPE_TRAITS 1
+#define EIGEN_INCLUDE_TYPE_TRAITS
+#else
+#define EIGEN_HAS_TYPE_TRAITS 0
+#endif
+#endif
+
 // Does the compiler support variadic templates?
 #ifndef EIGEN_HAS_VARIADIC_TEMPLATES
 #if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \
diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 4b337f29f..c88f1dbb9 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -34,6 +34,16 @@ inline IndexDest convert_index(const IndexSrc& idx) {
   return IndexDest(idx);
 }
 
+// true if T can be considered as an integral index (i.e., and integral type or enum)
+template<typename T> struct is_valid_index_type
+  : std::integral_constant<bool,
+#if EIGEN_HAS_TYPE_TRAITS
+   internal::is_integral<T>::value || std::is_enum<T>::value
+#else
+  // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
+  internal::is_convertible<T,Index>::value
+#endif
+> {};
 
 // promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
 //    expression * scalar
diff --git a/Eigen/src/plugins/IndexedViewMethods.h b/Eigen/src/plugins/IndexedViewMethods.h
index 22c1666c5..215bd0530 100644
--- a/Eigen/src/plugins/IndexedViewMethods.h
+++ b/Eigen/src/plugins/IndexedViewMethods.h
@@ -55,9 +55,7 @@ ivcSize(const Indices& indices) const {
 
 template<typename RowIndices, typename ColIndices>
 struct valid_indexed_view_overload {
-  // Here we use is_convertible to Index instead of is_integral in order to treat enums as Index.
-  // In c++11 we could use is_integral<T> && is_enum<T> if is_convertible appears to be too permissive.
-  enum { value = !(internal::is_convertible<RowIndices,Index>::value && internal::is_convertible<ColIndices,Index>::value) };
+  enum { value = !(internal::is_valid_index_type<RowIndices>::value && internal::is_valid_index_type<ColIndices>::value) };
 };
 
 public:
@@ -146,7 +144,7 @@ operator()(const RowIndicesT (&rowIndices)[RowIndicesN], const ColIndicesT (&col
 
 template<typename Indices>
 typename internal::enable_if<
-  IsRowMajor && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_integral<Indices>::value)),
+  IsRowMajor && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_valid_index_type<Indices>::value)),
   IndexedView<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,IvcIndex,typename IvcType<Indices>::type> >::type
 operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 {
@@ -157,7 +155,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 
 template<typename Indices>
 typename internal::enable_if<
-  (!IsRowMajor) && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_integral<Indices>::value)),
+  (!IsRowMajor) && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_valid_index_type<Indices>::value)),
   IndexedView<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,typename IvcType<Indices>::type,IvcIndex> >::type
 operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 {
@@ -168,7 +166,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 
 template<typename Indices>
 typename internal::enable_if<
-  (internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_integral<Indices>::value) && (!Symbolic::is_symbolic<Indices>::value),
+  (internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_valid_index_type<Indices>::value) && (!Symbolic::is_symbolic<Indices>::value),
   VectorBlock<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,internal::array_size<Indices>::value> >::type
 operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 {
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 7245cf378..8b3082cea 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -366,6 +366,11 @@ void check_indexed_view()
     VERIFY( is_same_eq( cA.middleRows<3>(1), cA.middleRows(1,fix<3>)) );
   }
 
+  // Check compilation of enums as index type:
+  enum { X=0, Y=1 };
+  a(X) = 1;
+  A(X,Y) = 1;
+
 }
 
 void test_indexed_view()

From 94e2213b3823ec1141077a7a7cc4f0b0af11876b Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 8 Sep 2017 15:49:55 -0700
Subject: [PATCH 189/680] Avoid undefined behavior in
 Eigen::TensorCostModel::numThreads.

If the cost is large enough then the thread count can be larger than the maximum
representable int, so just casting it to an int is undefined behavior.

Contributed by phurst@google.com.
---
 unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
index 83c449cf1..ac6717977 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
@@ -174,8 +174,9 @@ class TensorCostModel {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int numThreads(
       double output_size, const TensorOpCost& cost_per_coeff, int max_threads) {
     double cost = totalCost(output_size, cost_per_coeff);
-    int threads = (cost - kStartupCycles) / kPerThreadCycles + 0.9;
-    return numext::mini(max_threads, numext::maxi(1, threads));
+    // Make sure we don't invoke undefined behavior when we convert to an int.
+    threads = numext::mini<double>(threads, GenericNumTraits<int>::highest());
+    return numext::mini(max_threads, numext::maxi<int>(1, threads));
   }
 
   // taskSize assesses parallel task size.

From 1b7294f6fc5661ef196acf95f0eafa3242939be1 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 8 Sep 2017 16:35:58 -0700
Subject: [PATCH 190/680] Fix cut-and-paste error.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
index ac6717977..b148dae39 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
@@ -174,6 +174,7 @@ class TensorCostModel {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int numThreads(
       double output_size, const TensorOpCost& cost_per_coeff, int max_threads) {
     double cost = totalCost(output_size, cost_per_coeff);
+    double threads = (cost - kStartupCycles) / kPerThreadCycles + 0.9;
     // Make sure we don't invoke undefined behavior when we convert to an int.
     threads = numext::mini<double>(threads, GenericNumTraits<int>::highest());
     return numext::mini(max_threads, numext::maxi<int>(1, threads));

From 0c9ad2f52526bd7f07ba41161e68bee7231c05c7 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 14 Sep 2017 19:23:38 +0200
Subject: [PATCH 191/680] std::integral_constant is not C++03 compatible

---
 Eigen/src/Core/util/XprHelper.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index c88f1dbb9..10328be0d 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -36,14 +36,16 @@ inline IndexDest convert_index(const IndexSrc& idx) {
 
 // true if T can be considered as an integral index (i.e., and integral type or enum)
 template<typename T> struct is_valid_index_type
-  : std::integral_constant<bool,
+{
+  enum { value =
 #if EIGEN_HAS_TYPE_TRAITS
    internal::is_integral<T>::value || std::is_enum<T>::value
 #else
   // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
   internal::is_convertible<T,Index>::value
 #endif
-> {};
+  };
+};
 
 // promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
 //    expression * scalar

From 23f8b00bc884fe94e94ea273538e5546a4160e4f Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 14 Sep 2017 19:26:03 +0200
Subject: [PATCH 192/680] clang provides __has_feature(is_enum) (but not
 <type_traits>) in C++03 mode

---
 Eigen/src/Core/util/Macros.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index caedb0b3a..f5b071a9c 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -412,7 +412,7 @@
 
 // Does the compiler support type_trais?
 #ifndef EIGEN_HAS_TYPE_TRAITS
-#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || __has_feature(is_enum) || EIGEN_COMP_MSVC >= 1700)
+#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
 #define EIGEN_HAS_TYPE_TRAITS 1
 #define EIGEN_INCLUDE_TYPE_TRAITS
 #else

From 2062ac995864d51caad71a83bfc267a6aa17f8c3 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Mon, 18 Sep 2017 18:17:39 +0100
Subject: [PATCH 193/680] Changes required for new ComputeCpp CE version.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 2b5749f55..6158acbd9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -46,8 +46,8 @@ struct SyclAllocator {
 
 namespace Eigen {
 
-  #define ConvertToActualTypeSycl(Scalar, buf_acc) reinterpret_cast<typename cl::sycl::global_ptr<Scalar>::pointer_t>((&(*buf_acc.get_pointer())))
-  #define ConvertToActualSyclOffset(Scalar, offset) offset/sizeof(Scalar)
+#define ConvertToActualTypeSycl(Scalar, buf_acc) static_cast<Scalar*>(static_cast<void*>(((buf_acc.get_pointer().get()))))
+#define ConvertToActualSyclOffset(Scalar, offset) offset/sizeof(Scalar)
 
 
   template <typename Scalar, typename read_accessor, typename write_accessor> class MemCopyFunctor {

From 7c9b07dc5ccb2456ecbcdeef6870e7f4a59aba55 Mon Sep 17 00:00:00 2001
From: LaFeuille <pmoulon@gmail.com>
Date: Wed, 20 Sep 2017 06:38:39 +0000
Subject: [PATCH 194/680] Typo fix alignmeent ->alignment

---
 Eigen/src/Core/util/Macros.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index f5b071a9c..3a56b0e36 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -696,10 +696,10 @@ namespace Eigen {
 #if defined(EIGEN_DONT_VECTORIZE)
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
 #elif defined(EIGEN_VECTORIZE_AVX512)
-  // 64 bytes static alignmeent is preferred only if really required
+  // 64 bytes static alignment is preferred only if really required
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
 #elif defined(__AVX__)
-  // 32 bytes static alignmeent is preferred only if really required
+  // 32 bytes static alignment is preferred only if really required
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
 #else
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16

From 7ad07fc6f2e1dd74554ba576367883c4236c6b98 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 20 Sep 2017 10:22:00 +0200
Subject: [PATCH 195/680] Update documentation for aligned_allocator

---
 Eigen/src/Core/util/Memory.h | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 7d9053496..f48d115f4 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -696,7 +696,15 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 /** \class aligned_allocator
 * \ingroup Core_Module
 *
-* \brief STL compatible allocator to use with with 16 byte aligned types
+* \brief STL compatible allocator to use with types requiring a non standrad alignment.
+*
+* The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd.
+* By default, it will thus provide at least 16 bytes alignment and more in following cases:
+*  - 32 bytes alignment if AVX is enabled.
+*  - 64 bytes alignment if AVX512 is enabled.
+*
+* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
+* \link TopicPreprocessorDirectivesPerformance there \endlink.
 *
 * Example:
 * \code

From f92567fecc02f7653c5974ed9b162e49a813dc0c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 20 Sep 2017 10:22:23 +0200
Subject: [PATCH 196/680] Add link to a useful example.

---
 Eigen/src/plugins/IndexedViewMethods.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/src/plugins/IndexedViewMethods.h b/Eigen/src/plugins/IndexedViewMethods.h
index 215bd0530..a7ec63adf 100644
--- a/Eigen/src/plugins/IndexedViewMethods.h
+++ b/Eigen/src/plugins/IndexedViewMethods.h
@@ -248,6 +248,8 @@ operator()(const IndicesT (&indices)[IndicesN]) EIGEN_INDEXED_VIEW_METHOD_CONST
   *
   * For 1D vectors and arrays, you better use the operator()(const Indices&) overload, which behave the same way but taking a single parameter.
   *
+  * See also this <a href="https://stackoverflow.com/questions/46110917/eigen-replicate-items-along-one-dimension-without-useless-allocations">question</a> and its answer for an example of how to duplicate coefficients.
+  *
   * \sa operator()(const Indices&), class Block, class IndexedView, DenseBase::block(Index,Index,Index,Index)
   */
 template<typename RowIndices, typename ColIndices>

From 8579195169ba046b980b01769edb581b281e0b8a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 22 Sep 2017 09:23:24 +0200
Subject: [PATCH 197/680] bug #1468 (1/2) : add missing std:: to memcpy

---
 Eigen/QtAlignedMalloc             | 2 +-
 Eigen/src/Core/util/Memory.h      | 2 +-
 Eigen/src/SparseCore/AmbiVector.h | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/QtAlignedMalloc b/Eigen/QtAlignedMalloc
index c6571f129..4f07df02a 100644
--- a/Eigen/QtAlignedMalloc
+++ b/Eigen/QtAlignedMalloc
@@ -27,7 +27,7 @@ void qFree(void *ptr)
 void *qRealloc(void *ptr, std::size_t size)
 {
   void* newPtr = Eigen::internal::aligned_malloc(size);
-  memcpy(newPtr, ptr, size);
+  std::memcpy(newPtr, ptr, size);
   Eigen::internal::aligned_free(ptr);
   return newPtr;
 }
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index f48d115f4..c455f92a1 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -493,7 +493,7 @@ template<typename T> struct smart_copy_helper<T,true> {
     IntPtr size = IntPtr(end)-IntPtr(start);
     if(size==0) return;
     eigen_internal_assert(start!=0 && end!=0 && target!=0);
-    memcpy(target, start, size);
+    std::memcpy(target, start, size);
   }
 };
 
diff --git a/Eigen/src/SparseCore/AmbiVector.h b/Eigen/src/SparseCore/AmbiVector.h
index 8a5cc91f2..e0295f2af 100644
--- a/Eigen/src/SparseCore/AmbiVector.h
+++ b/Eigen/src/SparseCore/AmbiVector.h
@@ -94,7 +94,7 @@ class AmbiVector
       Index allocSize = m_allocatedElements * sizeof(ListEl);
       allocSize = (allocSize + sizeof(Scalar) - 1)/sizeof(Scalar);
       Scalar* newBuffer = new Scalar[allocSize];
-      memcpy(newBuffer,  m_buffer,  copyElements * sizeof(ListEl));
+      std::memcpy(newBuffer,  m_buffer,  copyElements * sizeof(ListEl));
       delete[] m_buffer;
       m_buffer = newBuffer;
     }

From 0e85a677e36956f13a3d85047d088d773b39b69b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 26 Sep 2017 10:53:33 +0200
Subject: [PATCH 198/680] bug #1472: fix warning

---
 Eigen/src/Core/Map.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index 7ca6a9280..c437f1a92 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -114,7 +114,7 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
-           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
+           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? Index(internal::traits<Map>::OuterStrideAtCompileTime)
            : IsVectorAtCompileTime ? (this->size() * innerStride())
            : int(Flags)&RowMajorBit ? (this->cols() * innerStride())
            : (this->rows() * innerStride());

From bc30305d2989443901c8ebe27045c9e996d32a40 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Mon, 9 Oct 2017 16:55:10 -0400
Subject: [PATCH 199/680] complete z14 port

---
 Eigen/src/Core/arch/ZVector/Complex.h       | 423 ++++++++++++++------
 Eigen/src/Core/arch/ZVector/MathFunctions.h |  15 +-
 Eigen/src/Core/arch/ZVector/PacketMath.h    |  42 +-
 3 files changed, 329 insertions(+), 151 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index 1bfb73397..3c72968e6 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -15,6 +15,10 @@ namespace Eigen {
 
 namespace internal {
 
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+#endif
+
 static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
 static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
 
@@ -29,10 +33,14 @@ struct Packet2cf
 {
   EIGEN_STRONG_INLINE Packet2cf() {}
   EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12)
   union {
     Packet4f v;
     Packet1cd cd[2];
   };
+#else
+  Packet4f v;
+#endif
 };
 
 template<> struct packet_traits<std::complex<float> >  : default_packet_traits
@@ -89,63 +97,27 @@ template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type
 /* Forward declaration */
 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel);
 
-template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+/* complex<double> first */
 template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
-template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
 
 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
 
-template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
-{
-  Packet2cf res;
-  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
-  res.cd[1] = res.cd[0];
-  return res;
-}
-template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
-{
-  std::complex<float> EIGEN_ALIGN16 af[2];
-  af[0] = from[0*stride];
-  af[1] = from[1*stride];
-  return pload<Packet2cf>(af);
-}
 template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED)
 {
   return pload<Packet1cd>(from);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
-{
-  std::complex<float> EIGEN_ALIGN16 af[2];
-  pstore<std::complex<float> >((std::complex<float> *) af, from);
-  to[0*stride] = af[0];
-  to[1*stride] = af[1];
-}
 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED)
 {
   pstore<std::complex<double> >(to, from);
 }
-
-template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
-template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
-template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
-{
-  Packet2cf res;
-  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
-  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
-  return res;
-}
-
 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
   Packet2d a_re, a_im, v1, v2;
@@ -163,27 +135,12 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
 
   return Packet1cd(v1 + v2);
 }
-template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
-{
-  Packet2cf res;
-  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
-  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
-  return res;
-}
-
-template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
-template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
-
+template<> EIGEN_STRONG_INLINE Packet1cd pand    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por     <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from) {  return pset1<Packet1cd>(*from); }
-template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
@@ -193,61 +150,20 @@ template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Pac
 
   return res;
 }
-template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
-{
-  std::complex<float> EIGEN_ALIGN16 res[2];
-  pstore<std::complex<float> >(res, a);
-
-  return res[0];
-}
 
 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
-template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
-{
-  Packet2cf res;
-  res.cd[0] = a.cd[1];
-  res.cd[1] = a.cd[0];
-  return res;
-}
-
 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
-template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
-{
-  std::complex<float> res;
-  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
-  vec_st2f(b.v, (float*)&res);
-  return res;
-}
-
 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
 {
   return vecs[0];
 }
-template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
-{
-  PacketBlock<Packet2cf,2> transpose;
-  transpose.packet[0] = vecs[0];
-  transpose.packet[1] = vecs[1];
-  ptranspose(transpose);
-
-  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
-} 
-
 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
-template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
-{
-  std::complex<float> res;
-  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
-  vec_st2f(b.v, (float*)&res);
-  return res;
-}
-
 template<int Offset>
 struct palign_impl<Offset,Packet1cd>
 {
@@ -258,18 +174,6 @@ struct palign_impl<Offset,Packet1cd>
   }
 };
 
-template<int Offset>
-struct palign_impl<Offset,Packet2cf>
-{
-  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
-  {
-    if (Offset == 1) {
-      first.cd[0] = first.cd[1];
-      first.cd[1] = second.cd[0];
-    }
-  }
-};
-
 template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
 {
   EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
@@ -303,6 +207,156 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
   }
 };
 
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_);
+  return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+/* complex<float> follows */
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<float> >(res, a);
+
+  return res[0];
+}
+
+
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12)
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
+  res.cd[1] = res.cd[0];
+  return res;
+}
+#else
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  if((std::ptrdiff_t(&from) % 16) == 0)
+    res.v = pload<Packet4f>((const float *)&from);
+  else
+    res.v = ploadu<Packet4f>((const float *)&from);
+  res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI);
+  return res;
+}
+#endif
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+
+
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12)
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
+  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
+  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.cd[0] = a.cd[1];
+  res.cd[1] = a.cd[0];
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  PacketBlock<Packet2cf,2> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  ptranspose(transpose);
+
+  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
+} 
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset == 1) {
+      first.cd[0] = first.cd[1];
+      first.cd[1] = second.cd[0];
+    }
+  }
+};
+
 template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
 {
   EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
@@ -337,15 +391,6 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
 };
 
 EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
-EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
-
-template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
-{
-  // TODO optimize it for AltiVec
-  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
-  Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_);
-  return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
-}
 
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
@@ -356,11 +401,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
   return res;
 }
 
-EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
-{
-  return Packet1cd(preverse(Packet2d(x.v)));
-}
-
 EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
 {
   Packet2cf res;
@@ -369,13 +409,6 @@ EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
   return res;
 }
 
-EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
-{
-  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
-  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
-  kernel.packet[0].v = tmp;
-}
-
 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
 {
   Packet1cd tmp = kernel.packet[0].cd[1];
@@ -389,6 +422,136 @@ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, con
   result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v);
   return result;
 }
+#else
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet4f v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
+  // Get the imaginary parts of a
+  v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
+  // multiply a_re * b 
+  v1 = vec_madd(v1, b.v, p4f_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(v2, b.v, p4f_ZERO);
+  v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)));
+  // permute back to a proper order
+  v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
+  
+  return Packet2cf(padd<Packet4f>(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet4f rev_a;
+  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2);
+  return Packet2cf(rev_a);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  b = vec_sld(a.v, a.v, 8);
+  b = padd<Packet4f>(a.v, b);
+  return pfirst<Packet2cf>(Packet2cf(b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  Packet4f b1, b2;
+  b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
+  b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
+  b2 = vec_sld(b2, b2, 8);
+  b2 = padd<Packet4f>(b1, b2);
+
+  return Packet2cf(b2);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  Packet2cf prod;
+  b = vec_sld(a.v, a.v, 8);
+  prod = pmul<Packet2cf>(a, Packet2cf(b));
+
+  return pfirst<Packet2cf>(prod);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset==1)
+    {
+      first.v = vec_sld(first.v, second.v, 8);
+    }
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b);
+  Packet4f s = pmul<Packet4f>(b.v, b.v);
+  return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
+{
+  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  result.v = reinterpret_cast<Packet4f>(pblend<Packet2d>(ifPacket, reinterpret_cast<Packet2d>(thenPacket.v), reinterpret_cast<Packet2d>(elsePacket.v)));
+  return result;
+}
+#endif
 
 } // end namespace internal
 
diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h
index 5c7aa7256..234d82be7 100644
--- a/Eigen/src/Core/arch/ZVector/MathFunctions.h
+++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -96,37 +96,48 @@ template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f pexp<Packet4f>(const Packet4f& x)
 {
   Packet4f res;
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+  res = pexp<Packet4f>(x);
+#else
   res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
   res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
+#endif
   return res;
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d psqrt<Packet2d>(const Packet2d& x)
 {
-  return  __builtin_s390_vfsqdb(x);
+  return vec_sqrt(x);
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f psqrt<Packet4f>(const Packet4f& x)
 {
   Packet4f res;
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+  res = vec_sqrt(x);
+#else
   res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
   res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
+#endif
   return res;
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d prsqrt<Packet2d>(const Packet2d& x) {
-  // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
   return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f prsqrt<Packet4f>(const Packet4f& x) {
   Packet4f res;
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+  res = pset1<Packet4f>(1.0) / psqrt<Packet4f>(x);
+#else
   res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
   res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
+#endif
   return res;
 }
 
diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index 2e5f5774d..a3294d318 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -17,7 +17,7 @@ namespace Eigen {
 namespace internal {
 
 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
-#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16
 #endif
 
 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
@@ -29,7 +29,7 @@ namespace internal {
 #endif
 
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  16
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
 #endif
 
 typedef __vector int                 Packet4i;
@@ -42,12 +42,12 @@ typedef __vector unsigned long long  Packet2ul;
 typedef __vector long long           Packet2l;
 
 // Z14 has builtin support for float vectors
-#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH < 14)
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+typedef __vector float               Packet4f;
+#else
 typedef struct {
 	Packet2d  v4f[2];
 } Packet4f;
-#else
-typedef __vector float               Packet4f;
 #endif
 
 typedef union {
@@ -62,7 +62,7 @@ typedef union {
   Packet2l  v2l;
   Packet2ul v2ul;
   Packet2d  v2d;
-#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH >= 14)
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
   Packet4f  v4f;
 #endif
 } Packet;
@@ -89,7 +89,7 @@ typedef union {
   Packet2l p2l_##NAME = pset1<Packet2l>(X)
 
 // These constants are endian-agnostic
-//static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
 
 static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
@@ -99,6 +99,15 @@ static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
 static Packet2d p2d_ONE = { 1.0, 1.0 }; 
 static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
 
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
+  Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
+
+static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
+static Packet4f p4f_MZERO = { 0x80000000, 0x80000000, 0x80000000, 0x80000000};
+#endif
+
 static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
 static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
@@ -129,9 +138,9 @@ static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0
 static Packet16uc p16uc_TRANSPOSE64_HI = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
 static Packet16uc p16uc_TRANSPOSE64_LO = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
 
-//static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
 
-//static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
 
 
 #if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
@@ -616,7 +625,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, cons
 /* z13 has no vector float support so we emulate that with double
    z14 has proper vector float support.
 */
-#if ~defined(__ARCH__) || (defined(__ARCH__) && __ARCH < 14)
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12)
 /* Helper function to simulate a vec_splat_packet4f
  */
 template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
@@ -1041,7 +1050,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pnegate<Packet4f>(const Packet4f& a) { return (-a); }
 template<> EIGEN_STRONG_INLINE Packet4f pconj<Packet4f>  (const Packet4f& a) { return a; }
 template<> EIGEN_STRONG_INLINE Packet4f pmadd<Packet4f>  (const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
-template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>  (const float& a)  { return padd<Packet4i>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>   (const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
@@ -1052,9 +1060,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f> (const Packet4f& a) { r
 template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>  (const Packet4f& a) { return vec_ceil(a); }
 template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f> (const Packet4f& a) { return vec_floor(a); }
 template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>   (const Packet4f& a) { return vec_abs(a); }
-template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f> (const float* from) { return pload<Packet4f>(from); }
-
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); }
 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
@@ -1068,7 +1073,7 @@ template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
   return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
 }
 
-template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 {
   Packet4f b, sum;
   b   = vec_sld(a, a, 8);
@@ -1152,10 +1157,9 @@ template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, cons
 #endif
 
 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
-template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
-template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
-template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f> (const float* from) { return pload<Packet4f>(from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>  (const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
 
 } // end namespace internal
 

From 374f750ad4708408a1255a98964719fd598b0659 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Mon, 9 Oct 2017 16:56:30 -0400
Subject: [PATCH 200/680] eliminate 'enumeral and non-enumeral type in
 conditional expression' warning

---
 Eigen/src/Core/Map.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index 7ca6a9280..169123824 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -114,7 +114,7 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
-           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
+           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? int(internal::traits<Map>::OuterStrideAtCompileTime)
            : IsVectorAtCompileTime ? (this->size() * innerStride())
            : int(Flags)&RowMajorBit ? (this->cols() * innerStride())
            : (this->rows() * innerStride());

From c2a224648919acb27bc208dcc24797345e3a1353 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Tue, 10 Oct 2017 13:38:32 -0400
Subject: [PATCH 201/680] fix predux_mul for z14/float

---
 Eigen/src/Core/arch/ZVector/PacketMath.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index a3294d318..21612ba91 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -1115,7 +1115,9 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 // mul
 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
 {
-  return pfirst(pmul(a, reinterpret_cast<Packet4f>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8))));
+  Packet4f prod;
+  prod = pmul(a, vec_sld(a, a, 8));
+  return pfirst(pmul(prod, vec_sld(prod, prod, 4)));
 }
 
 // min

From 3dcae2a27ffca563ef0a5b78d9d9d27cb05216a8 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Wed, 11 Oct 2017 09:40:45 -0400
Subject: [PATCH 202/680] initial pexp() for 32-bit floats, commented out due
 to vec_cts()

---
 Eigen/src/Core/arch/ZVector/PacketMath.h | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index 21612ba91..a72aea851 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -103,6 +103,12 @@ static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
 #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
   Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
 
+#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+  Packet4f p4f_##NAME = pset1<Packet4f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
+
 static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
 static Packet4f p4f_MZERO = { 0x80000000, 0x80000000, 0x80000000, 0x80000000};
@@ -187,7 +193,11 @@ template<> struct packet_traits<float> : default_packet_traits
     HasSin  = 0,
     HasCos  = 0,
     HasLog  = 0,
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+    HasExp  = 0,
+#else
     HasExp  = 1,
+#endif
     HasSqrt = 1,
     HasRsqrt = 1,
     HasRound = 1,

From df173f562062843a73454f2eb2479ae1e26dbcdf Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Wed, 11 Oct 2017 09:40:49 -0400
Subject: [PATCH 203/680] initial pexp() for 32-bit floats, commented out due
 to vec_cts()

---
 Eigen/src/Core/arch/ZVector/MathFunctions.h | 96 +++++++++++++++++++--
 1 file changed, 90 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h
index 234d82be7..ff33a975f 100644
--- a/Eigen/src/Core/arch/ZVector/MathFunctions.h
+++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -20,6 +20,50 @@ namespace Eigen {
 
 namespace internal {
 
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+static _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+/* the smallest non denormalized float number */
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000); // -1.f/0.f
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan,     0xffffffff);
+  
+/* natural logarithm computed for 4 simultaneous float
+  return NaN for x <= 0
+*/
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+#endif
+
 static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
 static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
 static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
@@ -93,16 +137,56 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
-Packet4f pexp<Packet4f>(const Packet4f& x)
+Packet4f pexp<Packet4f>(const Packet4f& _x)
 {
-  Packet4f res;
 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
-  res = pexp<Packet4f>(x);
+/*
+  Packet4f x = _x;
+
+  Packet4f tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
+
+  // express exp(x) as exp(g + n*log(2))
+  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
+
+  fx = pfloor(fx);
+
+  tmp = pmul(fx, p4f_cephes_exp_C1);
+  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  z = pmul(x,x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // build 2^n
+  emm0 = vec_cts(fx, 0);
+  emm0 = emm0 + p4i_0x7f;
+  emm0 = emm0 << reinterpret_cast<Packet4i>(p4i_23);
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet4ui isnumber_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x),
+                 isnumber_mask);*/
+  return _x;
 #else
-  res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
-  res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
-#endif
+  Packet4f res;
+  res.v4f[0] = pexp<Packet2d>(_x.v4f[0]);
+  res.v4f[1] = pexp<Packet2d>(_x.v4f[1]);
   return res;
+#endif
 }
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED

From d0b7b9d0d321905776326ce99c5c3ff3d48f4ce7 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Wed, 11 Oct 2017 10:17:22 -0400
Subject: [PATCH 204/680] some Packet2cf pmul fixes

---
 Eigen/src/Core/arch/ZVector/Complex.h | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index 3c72968e6..f9e3a480a 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -426,21 +426,22 @@ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, con
 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); }
 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  Packet4f v1, v2;
+  Packet4f a_re, a_im, prod, prod_im;
 
   // Permute and multiply the real parts of a and b
-  v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
+  a_re = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
   // Get the imaginary parts of a
-  v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
-  // multiply a_re * b 
-  v1 = vec_madd(v1, b.v, p4f_ZERO);
+  a_im = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
   // multiply a_im * b and get the conjugate result
-  v2 = vec_madd(v2, b.v, p4f_ZERO);
-  v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)));
+  prod_im = a_im * b.v;
+  prod_im = pxor<Packet4f>(prod_im, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR));
   // permute back to a proper order
-  v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
-  
-  return Packet2cf(padd<Packet4f>(v1, v2));
+  prod_im = vec_perm(prod_im, prod_im, p16uc_COMPLEX32_REV);
+
+  // multiply a_re * b, add prod_im
+  prod = pmadd<Packet4f>(a_re, b.v, prod_im);
+ 
+  return Packet2cf(prod);
 }
 
 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)

From 380d41fd76f2e8aeb4f8d4b9138515a05c9d70e3 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Wed, 11 Oct 2017 10:40:12 -0400
Subject: [PATCH 205/680] added some extra debugging

---
 Eigen/src/Core/arch/ZVector/Complex.h    | 10 ++++++++++
 Eigen/src/Core/arch/ZVector/PacketMath.h | 11 +++++++++++
 2 files changed, 21 insertions(+)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index f9e3a480a..cd4316687 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -428,18 +428,28 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
 {
   Packet4f a_re, a_im, prod, prod_im;
 
+  std::cout << "a = " << a.v << std::endl;
+  std::cout << ", b = " << b.v << std::endl;
   // Permute and multiply the real parts of a and b
   a_re = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
+  
+  std::cout << "a_re = " << a_re << std::endl;
   // Get the imaginary parts of a
   a_im = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
+  std::cout << "a_im = " << a_im << std::endl;
+
   // multiply a_im * b and get the conjugate result
   prod_im = a_im * b.v;
+  std::cout << "prod_im = " << prod_im << std::endl;
   prod_im = pxor<Packet4f>(prod_im, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR));
+  std::cout << "prod_im = " << prod_im << std::endl;
   // permute back to a proper order
   prod_im = vec_perm(prod_im, prod_im, p16uc_COMPLEX32_REV);
+  std::cout << "prod_im = " << prod_im << std::endl;
 
   // multiply a_re * b, add prod_im
   prod = pmadd<Packet4f>(a_re, b.v, prod_im);
+  std::cout << "prod = " << prod << std::endl;
  
   return Packet2cf(prod);
 }
diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index a72aea851..0b37f4992 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -286,6 +286,17 @@ inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
   return s;
 }
 
+#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
+inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
+{
+  Packet vt;
+  vt.v4f = v;
+  s << vt.f[0] << ", " << vt.f[1] << ", " << vt.f[2] << ", " << vt.f[3];
+  return s;
+}
+#endif
+
+
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {

From c4ad358565a1099bc0bdad797c3ca8bd297ea770 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Wed, 11 Oct 2017 11:05:29 -0400
Subject: [PATCH 206/680] explicitly set conjugate mask

---
 Eigen/src/Core/arch/ZVector/Complex.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index cd4316687..7504acff9 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -16,7 +16,7 @@ namespace Eigen {
 namespace internal {
 
 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
-static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+static Packet4ui  p4ui_CONJ_XOR = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; //vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);
 #endif
 
 static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };

From 98e52cc770ac26fbd29aaa7583443009d7937084 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Thu, 12 Oct 2017 15:22:10 -0400
Subject: [PATCH 207/680] rollback 374f750ad4708408a1255a98964719fd598b0659

---
 Eigen/src/Core/Map.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index 169123824..7ca6a9280 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -114,7 +114,7 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
-           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? int(internal::traits<Map>::OuterStrideAtCompileTime)
+           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
            : IsVectorAtCompileTime ? (this->size() * innerStride())
            : int(Flags)&RowMajorBit ? (this->cols() * innerStride())
            : (this->rows() * innerStride());

From 6c3475f110a017b3872a94a4ef13d0c8b543faca Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Thu, 12 Oct 2017 15:34:55 -0400
Subject: [PATCH 208/680] remove debugging

---
 Eigen/src/Core/arch/ZVector/Complex.h | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index 7504acff9..95aba428f 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -428,28 +428,20 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
 {
   Packet4f a_re, a_im, prod, prod_im;
 
-  std::cout << "a = " << a.v << std::endl;
-  std::cout << ", b = " << b.v << std::endl;
   // Permute and multiply the real parts of a and b
   a_re = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
   
-  std::cout << "a_re = " << a_re << std::endl;
   // Get the imaginary parts of a
   a_im = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
-  std::cout << "a_im = " << a_im << std::endl;
 
   // multiply a_im * b and get the conjugate result
   prod_im = a_im * b.v;
-  std::cout << "prod_im = " << prod_im << std::endl;
   prod_im = pxor<Packet4f>(prod_im, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR));
-  std::cout << "prod_im = " << prod_im << std::endl;
   // permute back to a proper order
   prod_im = vec_perm(prod_im, prod_im, p16uc_COMPLEX32_REV);
-  std::cout << "prod_im = " << prod_im << std::endl;
 
   // multiply a_re * b, add prod_im
   prod = pmadd<Packet4f>(a_re, b.v, prod_im);
-  std::cout << "prod = " << prod << std::endl;
  
   return Packet2cf(prod);
 }

From 0e6e027e91c134104d3595e5bfcd4a7d2a44a6dd Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <markos@freevec.org>
Date: Thu, 12 Oct 2017 15:38:34 -0400
Subject: [PATCH 209/680] check both z13 and z14 arches

---
 CMakeLists.txt | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2e1648fd1..62bf823c8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -271,12 +271,18 @@ if(NOT MSVC)
     message(STATUS "Enabling NEON in tests/examples")
   endif()
 
-  option(EIGEN_TEST_ZVECTOR "Enable/Disable S390X(zEC13) ZVECTOR in tests/examples" OFF)
-  if(EIGEN_TEST_ZVECTOR)
+  option(EIGEN_TEST_Z13 "Enable/Disable S390X(zEC13) ZVECTOR in tests/examples" OFF)
+  if(EIGEN_TEST_Z13)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=z13 -mzvector")
     message(STATUS "Enabling S390X(zEC13) ZVECTOR in tests/examples")
   endif()
 
+  option(EIGEN_TEST_Z14 "Enable/Disable S390X(zEC14) ZVECTOR in tests/examples" OFF)
+  if(EIGEN_TEST_Z14)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=z14 -mzvector")
+    message(STATUS "Enabling S390X(zEC13) ZVECTOR in tests/examples")
+  endif()
+
   check_cxx_compiler_flag("-fopenmp" COMPILER_SUPPORT_OPENMP)
   if(COMPILER_SUPPORT_OPENMP)
     option(EIGEN_TEST_OPENMP "Enable/Disable OpenMP in tests/examples" OFF)

From f349507e02b7623d439d90006289d0d74a75898d Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 13 Oct 2017 15:58:12 -0700
Subject: [PATCH 210/680] Specialize ThreadPoolDevice::enqueueNotification for
 the case with no args. As an example this reduces binary size of an
 TensorFlow demo app for Android by about 2.5%.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index 16180ca69..ec6802e85 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -154,7 +154,11 @@ struct ThreadPoolDevice {
 
   template <class Function, class... Args>
   EIGEN_STRONG_INLINE void enqueueNoNotification(Function&& f, Args&&... args) const {
-    pool_->Schedule(std::bind(f, args...));
+    if (sizeof...(args) > 0) {
+      pool_->Schedule(std::bind(f, args...));
+    } else {
+      pool_->Schedule(f);
+    }
   }
 
   // Returns a logical thread index between 0 and pool_->NumThreads() - 1 if

From 4245475d22f35572199a05a8119e8caf9f3e6c36 Mon Sep 17 00:00:00 2001
From: Henry Schreiner <henryschreineriii@utexas.edu>
Date: Fri, 20 Oct 2017 03:20:13 +0000
Subject: [PATCH 211/680] Fixing missing inlines on device functions for newer
 CUDA cards

---
 Eigen/src/Core/arch/CUDA/Half.h | 32 ++++++++++++++++----------------
 1 file changed, 16 insertions(+), 16 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index ee24e615a..666b23c6e 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -154,55 +154,55 @@ namespace half_impl {
 // versions to get the ALU speed increased), but you do save the
 // conversion steps back and forth.
 
-__device__ half operator + (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
   return __hadd(a, b);
 }
-__device__ half operator * (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
   return __hmul(a, b);
 }
-__device__ half operator - (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
   return __hsub(a, b);
 }
-__device__ half operator / (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
   float num = __half2float(a);
   float denom = __half2float(b);
   return __float2half(num / denom);
 }
-__device__ half operator - (const half& a) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
   return __hneg(a);
 }
-__device__ half& operator += (half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
   a = a + b;
   return a;
 }
-__device__ half& operator *= (half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
   a = a * b;
   return a;
 }
-__device__ half& operator -= (half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
   a = a - b;
   return a;
 }
-__device__ half& operator /= (half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
   a = a / b;
   return a;
 }
-__device__ bool operator == (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
   return __heq(a, b);
 }
-__device__ bool operator != (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
   return __hne(a, b);
 }
-__device__ bool operator < (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
   return __hlt(a, b);
 }
-__device__ bool operator <= (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
   return __hle(a, b);
 }
-__device__ bool operator > (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
   return __hgt(a, b);
 }
-__device__ bool operator >= (const half& a, const half& b) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
   return __hge(a, b);
 }
 
@@ -619,7 +619,7 @@ struct hash<Eigen::half> {
 
 // Add the missing shfl_xor intrinsic
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
-__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
   #if EIGEN_CUDACC_VER < 90000
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
   #else

From 9bb26eb8f1438d7856ae6c37cd5ba2c4414605fb Mon Sep 17 00:00:00 2001
From: Henry Schreiner <henryschreineriii@utexas.edu>
Date: Sat, 21 Oct 2017 00:50:38 +0000
Subject: [PATCH 212/680] Restore `__device__`

---
 Eigen/src/Core/arch/CUDA/Half.h | 32 ++++++++++++++++----------------
 1 file changed, 16 insertions(+), 16 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 666b23c6e..bfda39df5 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -154,55 +154,55 @@ namespace half_impl {
 // versions to get the ALU speed increased), but you do save the
 // conversion steps back and forth.
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
   return __hadd(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
   return __hmul(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
   return __hsub(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
   float num = __half2float(a);
   float denom = __half2float(b);
   return __float2half(num / denom);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
+EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
   return __hneg(a);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) {
   a = a + b;
   return a;
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half& operator *= (half& a, const half& b) {
   a = a * b;
   return a;
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half& operator -= (half& a, const half& b) {
   a = a - b;
   return a;
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) {
   a = a / b;
   return a;
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) {
   return __heq(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) {
   return __hne(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) {
   return __hlt(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) {
   return __hle(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) {
   return __hgt(a, b);
 }
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
+EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
   return __hge(a, b);
 }
 
@@ -619,7 +619,7 @@ struct hash<Eigen::half> {
 
 // Add the missing shfl_xor intrinsic
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
   #if EIGEN_CUDACC_VER < 90000
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
   #else

From a6d875bac8df04844781cc5a50c5063c8d920478 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Sun, 22 Oct 2017 08:12:45 -0700
Subject: [PATCH 213/680] Removed unecesasry #include

---
 unsupported/test/cxx11_tensor_argmax_cuda.cu            | 6 ------
 unsupported/test/cxx11_tensor_cast_float16_cuda.cu      | 6 ------
 unsupported/test/cxx11_tensor_complex_cuda.cu           | 6 ------
 unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu | 6 ------
 unsupported/test/cxx11_tensor_contract_cuda.cu          | 5 -----
 unsupported/test/cxx11_tensor_cuda.cu                   | 6 ------
 unsupported/test/cxx11_tensor_device.cu                 | 5 -----
 unsupported/test/cxx11_tensor_of_float16_cuda.cu        | 5 -----
 unsupported/test/cxx11_tensor_random_cuda.cu            | 6 ------
 unsupported/test/cxx11_tensor_reduction_cuda.cu         | 6 ------
 unsupported/test/cxx11_tensor_scan_cuda.cu              | 5 -----
 11 files changed, 62 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_cuda.cu
index 0e8b8125d..3d73d491a 100644
--- a/unsupported/test/cxx11_tensor_argmax_cuda.cu
+++ b/unsupported/test/cxx11_tensor_argmax_cuda.cu
@@ -15,12 +15,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 using Eigen::Tensor;
 
 template <int Layout>
diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
index dabf9e45f..816e03220 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu
@@ -16,12 +16,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 using Eigen::Tensor;
 
 void test_cuda_conversion() {
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index d25e1bee1..a52350f85 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -14,12 +14,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 using Eigen::Tensor;
 
 void test_cuda_nullary() {
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
index 4f0f621b4..aac780905 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
@@ -14,12 +14,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 using Eigen::Tensor;
 
 template<typename T>
diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_cuda.cu
index c68287e34..3621e2aa6 100644
--- a/unsupported/test/cxx11_tensor_contract_cuda.cu
+++ b/unsupported/test/cxx11_tensor_contract_cuda.cu
@@ -17,11 +17,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index d9059a2dc..9584a539f 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -15,12 +15,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 using Eigen::Tensor;
 
 void test_cuda_nullary() {
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index d5bfeeb39..7c14bc187 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -16,11 +16,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 
 using Eigen::Tensor;
 using Eigen::RowMajor;
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index c9f3ae1ae..167b75d25 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -16,11 +16,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 
 using Eigen::Tensor;
 
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu
index 9d08605fc..fa1a46732 100644
--- a/unsupported/test/cxx11_tensor_random_cuda.cu
+++ b/unsupported/test/cxx11_tensor_random_cuda.cu
@@ -16,12 +16,6 @@
 #include "main.h"
 #include <Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 
 void test_cuda_random_uniform()
 {
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_cuda.cu
index d6ce04f1c..ec0669704 100644
--- a/unsupported/test/cxx11_tensor_reduction_cuda.cu
+++ b/unsupported/test/cxx11_tensor_reduction_cuda.cu
@@ -15,12 +15,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
-
 
 template<typename Type, int DataLayout>
 static void test_full_reductions() {
diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_cuda.cu
index e99724b91..1d4edef11 100644
--- a/unsupported/test/cxx11_tensor_scan_cuda.cu
+++ b/unsupported/test/cxx11_tensor_scan_cuda.cu
@@ -16,11 +16,6 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-// The EIGEN_CUDACC_VER macro is provided by
-// unsupported/Eigen/CXX11/Tensor included above
-#if defined EIGEN_CUDACC_VER && EIGEN_CUDACC_VER >= 70500
-#include <cuda_fp16.h>
-#endif
 
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;

From 11ddac57e5fb7ed8b377b58ca955689bb637afe2 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Mon, 23 Oct 2017 13:22:22 +0000
Subject: [PATCH 214/680] Merged in guillaume_michel/eigen (pull request
 PR-334)

- Add support for NEON plog PacketMath function
---
 Eigen/src/Core/arch/NEON/MathFunctions.h | 92 ++++++++++++++++++++++++
 Eigen/src/Core/arch/NEON/PacketMath.h    |  2 +-
 2 files changed, 93 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/NEON/MathFunctions.h b/Eigen/src/Core/arch/NEON/MathFunctions.h
index 6bb05bb92..c48c61023 100644
--- a/Eigen/src/Core/arch/NEON/MathFunctions.h
+++ b/Eigen/src/Core/arch/NEON/MathFunctions.h
@@ -84,6 +84,98 @@ Packet4f pexp<Packet4f>(const Packet4f& _x)
   return y;
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f plog<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+
+  _EIGEN_DECLARE_CONST_Packet4i(inv_mant_mask, ~0x7f800000);
+
+  /* natural logarithm computed for 4 simultaneous float
+    return NaN for x <= 0
+  */
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+  x = vmaxq_f32(x, vdupq_n_f32(0)); /* force flush to zero on denormal values */
+  Packet4ui invalid_mask = vcleq_f32(x, vdupq_n_f32(0));
+
+  Packet4i ux = vreinterpretq_s32_f32(x);
+
+  Packet4i emm0 = vshrq_n_s32(ux, 23);
+
+  /* keep only the fractional part */
+  ux = vandq_s32(ux, p4i_inv_mant_mask);
+  ux = vorrq_s32(ux, vreinterpretq_s32_f32(p4f_half));
+  x = vreinterpretq_f32_s32(ux);
+
+  emm0 = vsubq_s32(emm0, p4i_0x7f);
+  Packet4f e = vcvtq_f32_s32(emm0);
+
+  e = vaddq_f32(e, p4f_1);
+
+  /* part2:
+     if( x < SQRTHF ) {
+       e -= 1;
+       x = x + x - 1.0;
+     } else { x = x - 1.0; }
+  */
+  Packet4ui mask = vcltq_f32(x, p4f_cephes_SQRTHF);
+  Packet4f tmp = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(x), mask));
+  x = vsubq_f32(x, p4f_1);
+  e = vsubq_f32(e, vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(p4f_1), mask)));
+  x = vaddq_f32(x, tmp);
+
+  Packet4f z = vmulq_f32(x,x);
+
+  Packet4f y = p4f_cephes_log_p0;
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p1);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p2);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p3);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p4);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p5);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p6);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p7);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_log_p8);
+  y = vmulq_f32(y, x);
+
+  y = vmulq_f32(y, z);
+
+  tmp = vmulq_f32(e, p4f_cephes_log_q1);
+  y = vaddq_f32(y, tmp);
+
+
+  tmp = vmulq_f32(z, p4f_half);
+  y = vsubq_f32(y, tmp);
+
+  tmp = vmulq_f32(e, p4f_cephes_log_q2);
+  x = vaddq_f32(x, y);
+  x = vaddq_f32(x, tmp);
+  x = vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(x), invalid_mask)); // negative arg will be NAN
+  return x;
+}
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 836fbc0dd..6283b6cfa 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -81,7 +81,7 @@ template<> struct packet_traits<float>  : default_packet_traits
     // FIXME check the Has*
     HasSin  = 0,
     HasCos  = 0,
-    HasLog  = 0,
+    HasLog  = 1,
     HasExp  = 1,
     HasSqrt = 0
   };

From e8468ea91b45e6b09e1a58626a78fd723da9b64f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 8 Nov 2017 10:24:28 +0100
Subject: [PATCH 215/680] Fix overflow issues in BDCSVD

---
 Eigen/src/SVD/BDCSVD.h | 33 ++++++++++++++++++++++++---------
 1 file changed, 24 insertions(+), 9 deletions(-)

diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index b8c41c560..effa643fd 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -11,7 +11,7 @@
 // Copyright (C) 2013 Jean Ceccato <jean.ceccato@ensimag.fr>
 // Copyright (C) 2013 Pierre Zoppitelli <pierre.zoppitelli@ensimag.fr>
 // Copyright (C) 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
-// Copyright (C) 2014-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2014-2017 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -696,7 +696,9 @@ typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar
   for(Index i=0; i<m; ++i)
   {
     Index j = perm(i);
-    res += numext::abs2(col0(j)) / ((diagShifted(j) - mu) * (diag(j) + shift + mu));
+    // The following expression could be rewritten to involve only a single division,
+    // but this would make the expression more sensitive to overflow.
+    res += (col0(j) / (diagShifted(j) - mu)) * (col0(j) / (diag(j) + shift + mu));
   }
   return res;
 
@@ -708,9 +710,12 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
 {
   using std::abs;
   using std::swap;
+  using std::sqrt;
 
   Index n = col0.size();
   Index actual_n = n;
+  // Note that here actual_n is computed based on col0(i)==0 instead of diag(i)==0 as above
+  // because 1) we have diag(i)==0 => col0(i)==0 and 2) if col0(i)==0, then diag(i) is already a singular value.
   while(actual_n>1 && col0(actual_n-1)==Literal(0)) --actual_n;
 
   for (Index k = 0; k < n; ++k)
@@ -732,7 +737,9 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       right = (diag(actual_n-1) + col0.matrix().norm());
     else
     {
-      // Skip deflated singular values
+      // Skip deflated singular values,
+      // recall that at this stage we assume that z[j]!=0 and all entries for which z[j]==0 have been put aside.
+      // This should be equivalent to using perm[]
       Index l = k+1;
       while(col0(l)==Literal(0)) { ++l; eigen_internal_assert(l<actual_n); }
       right = diag(l);
@@ -818,15 +825,23 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       RealScalar leftShifted, rightShifted;
       if (shift == left)
       {
-        leftShifted = (std::numeric_limits<RealScalar>::min)();
+        // to avoid overflow, we must have mu > max(real_min, |z(k)|/sqrt(real_max)),
+        // the factor 2 is to be more conservative
+        leftShifted = numext::maxi( (std::numeric_limits<RealScalar>::min)(), Literal(2) * abs(col0(k)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
+
+        // check that we did it right:
+        eigen_internal_assert( (std::isfinite)( (col0(k)/leftShifted)*(col0(k)/(diag(k)+shift+leftShifted)) ) );
         // I don't understand why the case k==0 would be special there:
-        // if (k == 0) rightShifted = right - left; else 
-        rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.6)); // theoretically we can take 0.5, but let's be safe
+        // if (k == 0) rightShifted = right - left; else
+        rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.51)); // theoretically we can take 0.5, but let's be safe
       }
       else
       {
-        leftShifted = -(right - left) * RealScalar(0.6);
-        rightShifted = -(std::numeric_limits<RealScalar>::min)();
+        leftShifted = -(right - left) * RealScalar(0.51);
+        if(k+1<n)
+          rightShifted = -numext::maxi( (std::numeric_limits<RealScalar>::min)(), abs(col0(k+1)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
+        else
+          rightShifted = -(std::numeric_limits<RealScalar>::min)();
       }
       
       RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift);
@@ -980,7 +995,7 @@ void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index
   Index start = firstCol + shift;
   RealScalar c = m_computed(start, start);
   RealScalar s = m_computed(start+i, start);
-  RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
+  RealScalar r = numext::hypot(c,s);
   if (r == Literal(0))
   {
     m_computed(start+i, start+i) = Literal(0);

From e9d2888e74f9afaab16e2b5ec5264d44613ff366 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 8 Nov 2017 10:26:03 +0100
Subject: [PATCH 216/680] Improve debugging tests and output in BDCSVD

---
 Eigen/src/SVD/BDCSVD.h | 131 ++++++++++++++++++++++++++++++++---------
 1 file changed, 102 insertions(+), 29 deletions(-)

diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index effa643fd..b440901d6 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -22,6 +22,11 @@
 // #define EIGEN_BDCSVD_DEBUG_VERBOSE
 // #define EIGEN_BDCSVD_SANITY_CHECKS
 
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+#undef eigen_internal_assert
+#define eigen_internal_assert(X) assert(X);
+#endif
+
 namespace Eigen {
 
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
@@ -591,7 +596,7 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
   // but others are interleaved and we must ignore them at this stage.
   // To this end, let's compute a permutation skipping them:
   Index actual_n = n;
-  while(actual_n>1 && diag(actual_n-1)==Literal(0)) --actual_n;
+  while(actual_n>1 && diag(actual_n-1)==Literal(0)) {--actual_n; eigen_internal_assert(col0(actual_n)==Literal(0)); }
   Index m = 0; // size of the deflated problem
   for(Index k=0;k<actual_n;++k)
     if(abs(col0(k))>considerZero)
@@ -618,13 +623,11 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
   std::cout << "  shift:    " << shifts.transpose() << "\n";
   
   {
-    Index actual_n = n;
-    while(actual_n>1 && abs(col0(actual_n-1))<considerZero) --actual_n;
     std::cout << "\n\n    mus:    " << mus.head(actual_n).transpose() << "\n\n";
     std::cout << "    check1 (expect0) : " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n";
+    assert((((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n) >= 0).all());
     std::cout << "    check2 (>0)      : " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n";
-    std::cout << "    check3 (>0)      : " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n";
-    std::cout << "    check4 (>0)      : " << ((singVals.segment(1,actual_n-1)-singVals.head(actual_n-1))).transpose() << "\n\n\n";
+    assert((((singVals.array()-diag) / singVals.array()).head(actual_n) >= 0).all());
   }
 #endif
   
@@ -652,13 +655,13 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
 #endif
   
 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
-  assert(U.allFinite());
-  assert(V.allFinite());
-  assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n);
-  assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n);
   assert(m_naiveU.allFinite());
   assert(m_naiveV.allFinite());
   assert(m_computed.allFinite());
+  assert(U.allFinite());
+  assert(V.allFinite());
+//   assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 100*NumTraits<RealScalar>::epsilon() * n);
+//   assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 100*NumTraits<RealScalar>::epsilon() * n);
 #endif
   
   // Because of deflation, the singular values might not be completely sorted.
@@ -673,6 +676,15 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
       if(m_compV) V.col(i).swap(V.col(i+1));
     }
   }
+
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  {
+    bool singular_values_sorted = (((singVals.segment(1,actual_n-1)-singVals.head(actual_n-1))).array() >= 0).all();
+    if(!singular_values_sorted)
+      std::cout << "Singular values are not sorted: " << singVals.segment(1,actual_n).transpose() << "\n";
+    assert(singular_values_sorted);
+  }
+#endif
   
   // Reverse order so that singular values in increased order
   // Because of deflation, the zeros singular-values are already at the end
@@ -749,19 +761,22 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
     RealScalar mid = left + (right-left) / Literal(2);
     RealScalar fMid = secularEq(mid, col0, diag, perm, diag, Literal(0));
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
-    std::cout << right-left << "\n";
-    std::cout << "fMid = " << fMid << " " << secularEq(mid-left, col0, diag, perm, diag-left, left) << " " << secularEq(mid-right, col0, diag, perm, diag-right, right)   << "\n";
-    std::cout << "     = " << secularEq(0.1*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.2*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.3*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.4*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.49*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.5*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.51*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.6*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.7*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.8*(left+right), col0, diag, perm, diag, 0)
-              << " "       << secularEq(0.9*(left+right), col0, diag, perm, diag, 0) << "\n";
+    std::cout << "right-left = " << right-left << "\n";
+//     std::cout << "fMid = " << fMid << " " << secularEq(mid-left, col0, diag, perm, ArrayXr(diag-left), left)
+//                            << " " << secularEq(mid-right, col0, diag, perm, ArrayXr(diag-right), right)   << "\n";
+    std::cout << "     = " << secularEq(left+RealScalar(0.000001)*(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.1)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.2)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.3)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.4)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.49)    *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.5)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.51)    *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.6)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.7)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.8)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.9)     *(right-left), col0, diag, perm, diag, 0)
+              << " "       << secularEq(left+RealScalar(0.999999)*(right-left), col0, diag, perm, diag, 0) << "\n";
 #endif
     RealScalar shift = (k == actual_n-1 || fMid > Literal(0)) ? left : right;
     
@@ -804,13 +819,16 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       // And find mu such that f(mu)==0:
       RealScalar muZero = -a/b;
       RealScalar fZero = secularEq(muZero, col0, diag, perm, diagShifted, shift);
+
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+      assert((std::isfinite)(fZero));
+#endif
       
       muPrev = muCur;
       fPrev = fCur;
       muCur = muZero;
       fCur = fZero;
       
-      
       if (shift == left  && (muCur < Literal(0) || muCur > right - left)) useBisection = true;
       if (shift == right && (muCur < -(right - left) || muCur > Literal(0))) useBisection = true;
       if (abs(fCur)>abs(fPrev)) useBisection = true;
@@ -843,18 +861,30 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
         else
           rightShifted = -(std::numeric_limits<RealScalar>::min)();
       }
-      
+
       RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift);
 
-#if defined EIGEN_INTERNAL_DEBUGGING || defined EIGEN_BDCSVD_DEBUG_VERBOSE
+#if defined EIGEN_INTERNAL_DEBUGGING || defined EIGEN_BDCSVD_SANITY_CHECKS
       RealScalar fRight = secularEq(rightShifted, col0, diag, perm, diagShifted, shift);
 #endif
 
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+      if(!(std::isfinite)(fLeft))
+        std::cout << "f(" << leftShifted << ") =" << fLeft << " ; " << left << " " << shift << " " << right << "\n";
+      assert((std::isfinite)(fLeft));
+
+      if(!(std::isfinite)(fRight))
+        std::cout << "f(" << rightShifted << ") =" << fRight << " ; " << left << " " << shift << " " << right << "\n";
+//       assert((std::isfinite)(fRight));
+#endif
+
 #ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
       if(!(fLeft * fRight<0))
       {
-        std::cout << "fLeft: " << leftShifted << " - " << diagShifted.head(10).transpose()  << "\n ; " << bool(left==shift) << " " << (left-shift) << "\n";
-        std::cout << k << " : " <<  fLeft << " * " << fRight << " == " << fLeft * fRight << "  ;  " << left << " - " << right << " -> " <<  leftShifted << " " << rightShifted << "   shift=" << shift << "\n";
+        std::cout << "f(leftShifted) using  leftShifted=" << leftShifted << " ;  diagShifted(1:10):" << diagShifted.head(10).transpose()  << "\n ; "
+                  << "left==shift=" << bool(left==shift) << " ; left-shift = " << (left-shift) << "\n";
+        std::cout << "k=" << k << ", " <<  fLeft << " * " << fRight << " == " << fLeft * fRight << "  ;  "
+                  << "[" << left << " .. " << right << "] -> [" << leftShifted << " " << rightShifted << "], shift=" << shift << " ,  f(right)=" <<  secularEq(0, col0, diag, perm, diagShifted, shift) << " == " << secularEq(right, col0, diag, perm, diag, 0) << "\n";
       }
 #endif
       eigen_internal_assert(fLeft * fRight < Literal(0));
@@ -863,6 +893,8 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       {
         RealScalar midShifted = (leftShifted + rightShifted) / Literal(2);
         fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift);
+        eigen_internal_assert((numext::isfinite)(fMid));
+
         if (fLeft * fMid < Literal(0))
         {
           rightShifted = midShifted;
@@ -881,6 +913,15 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
     shifts[k] = shift;
     mus[k] = muCur;
 
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+    if(k+1<n)
+      std::cout << "found " << singVals[k] << " == " << shift << " + " << muCur << " from " << diag(k) << " .. "  << diag(k+1) << "\n";
+#endif
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+    assert(k==0 || singVals[k]>=singVals[k-1]);
+    assert(singVals[k]>=diag(k));
+#endif
+
     // perturb singular value slightly if it equals diagonal entry to avoid division by zero later
     // (deflation is supposed to avoid this from happening)
     // - this does no seem to be necessary anymore -
@@ -915,14 +956,42 @@ void BDCSVD<MatrixType>::perturbCol0
       // see equation (3.6)
       RealScalar dk = diag(k);
       RealScalar prod = (singVals(last) + dk) * (mus(last) + (shifts(last) - dk));
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+      if(prod<0) {
+        std::cout << "k = " << k << " ;  z(k)=" << col0(k) << ", diag(k)=" << dk << "\n";
+        std::cout << "prod = " << "(" << singVals(last) << " + " << dk << ") * (" << mus(last) << " + (" << shifts(last) << " - " << dk << "))" << "\n";
+        std::cout << "     = " << singVals(last) + dk << " * " << mus(last) + (shifts(last) - dk) <<  "\n";
+      }
+      assert(prod>=0);
+#endif
 
       for(Index l = 0; l<m; ++l)
       {
         Index i = perm(l);
         if(i!=k)
         {
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+          if(i>=k && (l==0 || l-1>=m))
+          {
+            std::cout << "Error in perturbCol0\n";
+            std::cout << "  " << k << "/" << n << " "  << l << "/" << m << " " << i << "/" << n << " ; " << col0(k) << " " << diag(k) << " "  <<  "\n";
+            std::cout << "  " <<diag(i) << "\n";
+            Index j = (i<k /*|| l==0*/) ? i : perm(l-1);
+            std::cout << "  " << "j=" << j << "\n";
+          }
+#endif
           Index j = i<k ? i : perm(l-1);
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+          if(!(dk!=Literal(0) || diag(i)!=Literal(0)))
+          {
+            std::cout << "k=" << k << ", i=" << i << ", l=" << l << ", perm.size()=" << perm.size() << "\n";
+          }
+          assert(dk!=Literal(0) || diag(i)!=Literal(0));
+#endif
           prod *= ((singVals(j)+dk) / ((diag(i)+dk))) * ((mus(j)+(shifts(j)-dk)) / ((diag(i)-dk)));
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+          assert(prod>=0);
+#endif
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
           if(i!=k && std::abs(((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) - 1) > 0.9 )
             std::cout << "     " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk))
@@ -934,6 +1003,9 @@ void BDCSVD<MatrixType>::perturbCol0
       std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n";
 #endif
       RealScalar tmp = sqrt(prod);
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+      assert((std::isfinite)(tmp));
+#endif
       zhat(k) = col0(k) > Literal(0) ? tmp : -tmp;
     }
   }
@@ -1043,7 +1115,7 @@ void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index fi
   }
   c/=r;
   s/=r;
-  m_computed(firstColm + i, firstColm) = r;  
+  m_computed(firstColm + i, firstColm) = r;
   m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i);
   m_computed(firstColm + j, firstColm) = Literal(0);
 
@@ -1117,6 +1189,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
 #endif
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
   std::cout << "to be sorted: " << diag.transpose() << "\n\n";
+  std::cout << "            : " << col0.transpose() << "\n\n";
 #endif
   {
     // Check for total deflation
@@ -1207,7 +1280,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
        if( (diag(i) - diag(i-1)) < NumTraits<RealScalar>::epsilon()*maxDiag )
       {
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
-        std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i) - diag(i-1)) << " < " << NumTraits<RealScalar>::epsilon()*diag(i) << "\n";
+        std::cout << "deflation 4.4 with i = " << i << " because " << diag(i) << " - " << diag(i-1) << " == " << (diag(i) - diag(i-1)) << " < " << NumTraits<RealScalar>::epsilon()*/*diag(i)*/maxDiag << "\n";
 #endif
         eigen_internal_assert(abs(diag(i) - diag(i-1))<epsilon_coarse && " diagonal entries are not properly sorted");
         deflation44(firstCol, firstCol + shift, firstRowW, firstColW, i-1, i, length);

From 4c03b3511e055817a4c1aa2858bd5735cb5bb4e7 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 8 Nov 2017 23:28:01 +0100
Subject: [PATCH 217/680] Fix issue with boost::multiprec in previous commit

---
 Eigen/src/SVD/BDCSVD.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index b440901d6..0abd4c1bb 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -845,10 +845,10 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       {
         // to avoid overflow, we must have mu > max(real_min, |z(k)|/sqrt(real_max)),
         // the factor 2 is to be more conservative
-        leftShifted = numext::maxi( (std::numeric_limits<RealScalar>::min)(), Literal(2) * abs(col0(k)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
+        leftShifted = numext::maxi<RealScalar>( (std::numeric_limits<RealScalar>::min)(), Literal(2) * abs(col0(k)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
 
         // check that we did it right:
-        eigen_internal_assert( (std::isfinite)( (col0(k)/leftShifted)*(col0(k)/(diag(k)+shift+leftShifted)) ) );
+        eigen_internal_assert( (numext::isfinite)( (col0(k)/leftShifted)*(col0(k)/(diag(k)+shift+leftShifted)) ) );
         // I don't understand why the case k==0 would be special there:
         // if (k == 0) rightShifted = right - left; else
         rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.51)); // theoretically we can take 0.5, but let's be safe
@@ -857,7 +857,7 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& d
       {
         leftShifted = -(right - left) * RealScalar(0.51);
         if(k+1<n)
-          rightShifted = -numext::maxi( (std::numeric_limits<RealScalar>::min)(), abs(col0(k+1)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
+          rightShifted = -numext::maxi<RealScalar>( (std::numeric_limits<RealScalar>::min)(), abs(col0(k+1)) / sqrt((std::numeric_limits<RealScalar>::max)()) );
         else
           rightShifted = -(std::numeric_limits<RealScalar>::min)();
       }

From 912e9965ef64e97ed496e04b5c2a3376a21bcfd1 Mon Sep 17 00:00:00 2001
From: zhouzhaoping <zhouzhaoping11@gmail.com>
Date: Thu, 9 Nov 2017 08:49:01 +0000
Subject: [PATCH 218/680] a small mistake QuickReference.dox edited online with
 Bitbucket

---
 doc/QuickReference.dox | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/QuickReference.dox b/doc/QuickReference.dox
index 59d7d05e4..18c90a2a9 100644
--- a/doc/QuickReference.dox
+++ b/doc/QuickReference.dox
@@ -68,7 +68,7 @@ Array<float,4,1>                <=>   Array4f
 
 Conversion between the matrix and array worlds:
 \code
-Array44f a1, a1;
+Array44f a1, a2;
 Matrix4f m1, m2;
 m1 = a1 * a2;                     // coeffwise product, implicit conversion from array to matrix.
 a1 = m1 * m2;                     // matrix product, implicit conversion from matrix to array.

From 42a83346683861e7a58e8c9aa407eb001d56befa Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Yan=20Facai=20=28=E9=A2=9C=E5=8F=91=E6=89=8D=29?=
 <facai.yan@gmail.com>
Date: Thu, 4 Jan 2018 16:01:01 +0800
Subject: [PATCH 219/680] ENH: exp supports complex type for cuda

---
 Eigen/src/Core/MathFunctions.h | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 5ba5293a0..1b864a405 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1289,6 +1289,22 @@ float exp(const float &x) { return ::expf(x); }
 
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 double exp(const double &x) { return ::exp(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+std::complex<float> exp(const std::complex<float>& x) {
+  auto com = ::expf(x.real());
+  auto res_real = com * ::cosf(x.imag());
+  auto res_imag = com * ::sinf(x.imag());
+  return std::complex<float>(res_real, res_imag);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+std::complex<double> exp(const std::complex<double>& x) {
+  auto com = ::exp(x.real());
+  auto res_real = com * ::cos(x.imag());
+  auto res_imag = com * ::sin(x.imag());
+  return std::complex<double>(res_real, res_imag);
+}
 #endif
 
 template<typename Scalar>

From 0c57be407d790c6013e9bad7c8f0123f4689d172 Mon Sep 17 00:00:00 2001
From: Daniel Trebbien <dtrebbien@gmail.com>
Date: Sun, 18 Feb 2018 15:35:45 -0800
Subject: [PATCH 220/680] Move up the specialization of std::numeric_limits

This fixes a compilation error seen when building TensorFlow on macOS:
https://github.com/tensorflow/tensorflow/issues/17067
---
 Eigen/src/Core/arch/CUDA/Half.h | 97 ++++++++++++++++++---------------
 1 file changed, 54 insertions(+), 43 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index bfda39df5..2bb3d0d9b 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -145,6 +145,60 @@ struct half : public half_impl::half_base {
   }
 };
 
+} // end namespace Eigen
+
+namespace std {
+template<>
+struct numeric_limits<Eigen::half> {
+  static const bool is_specialized = true;
+  static const bool is_signed = true;
+  static const bool is_integer = false;
+  static const bool is_exact = false;
+  static const bool has_infinity = true;
+  static const bool has_quiet_NaN = true;
+  static const bool has_signaling_NaN = true;
+  static const float_denorm_style has_denorm = denorm_present;
+  static const bool has_denorm_loss = false;
+  static const std::float_round_style round_style = std::round_to_nearest;
+  static const bool is_iec559 = false;
+  static const bool is_bounded = false;
+  static const bool is_modulo = false;
+  static const int digits = 11;
+  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int radix = 2;
+  static const int min_exponent = -13;
+  static const int min_exponent10 = -4;
+  static const int max_exponent = 16;
+  static const int max_exponent10 = 4;
+  static const bool traps = true;
+  static const bool tinyness_before = false;
+
+  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
+  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
+  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
+  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
+  static Eigen::half round_error() { return Eigen::half(0.5); }
+  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
+  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
+};
+
+// If std::numeric_limits<T> is specialized, should also specialize
+// std::numeric_limits<const T>, std::numeric_limits<volatile T>, and
+// std::numeric_limits<const volatile T>
+// https://stackoverflow.com/a/16519653/
+template<>
+struct numeric_limits<const Eigen::half> : numeric_limits<Eigen::half> {};
+template<>
+struct numeric_limits<volatile Eigen::half> : numeric_limits<Eigen::half> {};
+template<>
+struct numeric_limits<const volatile Eigen::half> : numeric_limits<Eigen::half> {};
+} // end namespace std
+
+namespace Eigen {
+
 namespace half_impl {
 
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
@@ -501,49 +555,6 @@ template<> struct is_arithmetic<half> { enum { value = true }; };
 
 } // end namespace internal
 
-}  // end namespace Eigen
-
-namespace std {
-template<>
-struct numeric_limits<Eigen::half> {
-  static const bool is_specialized = true;
-  static const bool is_signed = true;
-  static const bool is_integer = false;
-  static const bool is_exact = false;
-  static const bool has_infinity = true;
-  static const bool has_quiet_NaN = true;
-  static const bool has_signaling_NaN = true;
-  static const float_denorm_style has_denorm = denorm_present;
-  static const bool has_denorm_loss = false;
-  static const std::float_round_style round_style = std::round_to_nearest;
-  static const bool is_iec559 = false;
-  static const bool is_bounded = false;
-  static const bool is_modulo = false;
-  static const int digits = 11;
-  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
-  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
-  static const int radix = 2;
-  static const int min_exponent = -13;
-  static const int min_exponent10 = -4;
-  static const int max_exponent = 16;
-  static const int max_exponent10 = 4;
-  static const bool traps = true;
-  static const bool tinyness_before = false;
-
-  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
-  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
-  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
-  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
-  static Eigen::half round_error() { return Eigen::half(0.5); }
-  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
-  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
-  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
-  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
-};
-}
-
-namespace Eigen {
-
 template<> struct NumTraits<Eigen::half>
     : GenericNumTraits<Eigen::half>
 {

From d820ab9edc0b38af4cdb3d545714a0c9083e5a78 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Mar 2018 09:33:43 +0100
Subject: [PATCH 221/680] Add static assertion on selfadjoint-view's UpLo
 parameter.

---
 Eigen/src/Core/SelfAdjointView.h   | 4 +++-
 Eigen/src/Core/util/StaticAssert.h | 3 ++-
 test/selfadjoint.cpp               | 4 ++++
 3 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h
index 7e71fe3c0..be219726b 100644
--- a/Eigen/src/Core/SelfAdjointView.h
+++ b/Eigen/src/Core/SelfAdjointView.h
@@ -71,7 +71,9 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
 
     EIGEN_DEVICE_FUNC
     explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
-    {}
+    {
+      EIGEN_STATIC_ASSERT(UpLo==Lower || UpLo==Upper,SELFADJOINTVIEW_ACCEPTS_UPPER_AND_LOWER_MODE_ONLY);
+    }
 
     EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows(); }
diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h
index cb1678900..987d5fce4 100644
--- a/Eigen/src/Core/util/StaticAssert.h
+++ b/Eigen/src/Core/util/StaticAssert.h
@@ -101,7 +101,8 @@
         THIS_TYPE_IS_NOT_SUPPORTED=1,
         STORAGE_KIND_MUST_MATCH=1,
         STORAGE_INDEX_MUST_MATCH=1,
-        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY=1
+        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY=1,
+        SELFADJOINTVIEW_ACCEPTS_UPPER_AND_LOWER_MODE_ONLY=1
       };
     };
 
diff --git a/test/selfadjoint.cpp b/test/selfadjoint.cpp
index 92401e506..5ac219352 100644
--- a/test/selfadjoint.cpp
+++ b/test/selfadjoint.cpp
@@ -7,6 +7,7 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
+#define EIGEN_NO_STATIC_ASSERT
 #include "main.h"
 
 // This file tests the basic selfadjointView API,
@@ -45,6 +46,9 @@ template<typename MatrixType> void selfadjoint(const MatrixType& m)
   m4 = m2;
   m4 -= m1.template selfadjointView<Lower>();
   VERIFY_IS_APPROX(m4, m2-m3);
+
+  VERIFY_RAISES_ASSERT(m2.template selfadjointView<StrictlyUpper>());
+  VERIFY_RAISES_ASSERT(m2.template selfadjointView<UnitLower>());
 }
 
 void bug_159()

From f6be7289d703ea608db01bfae1fb41cb93465363 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Mar 2018 10:00:51 +0100
Subject: [PATCH 222/680] Implement better static assertion checking to make
 sure that the first assertion is a static one and not a runtime one.

---
 Eigen/src/Core/util/StaticAssert.h |  3 ++-
 test/main.h                        | 37 ++++++++++++++++++++++++++++--
 test/selfadjoint.cpp               |  6 ++---
 3 files changed, 40 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h
index 987d5fce4..500e47792 100644
--- a/Eigen/src/Core/util/StaticAssert.h
+++ b/Eigen/src/Core/util/StaticAssert.h
@@ -24,6 +24,7 @@
  *
  */
 
+#ifndef EIGEN_STATIC_ASSERT
 #ifndef EIGEN_NO_STATIC_ASSERT
 
   #if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600))
@@ -132,7 +133,7 @@
   #define EIGEN_STATIC_ASSERT(CONDITION,MSG) eigen_assert((CONDITION) && #MSG);
 
 #endif // EIGEN_NO_STATIC_ASSERT
-
+#endif // EIGEN_STATIC_ASSERT
 
 // static assertion failing if the type \a TYPE is not a vector type
 #define EIGEN_STATIC_ASSERT_VECTOR_ONLY(TYPE) \
diff --git a/test/main.h b/test/main.h
index 429c44f81..6079cbd06 100644
--- a/test/main.h
+++ b/test/main.h
@@ -175,6 +175,12 @@ namespace Eigen
       eigen_assert_exception(void) {}
       ~eigen_assert_exception() { Eigen::no_more_assert = false; }
     };
+
+    struct eigen_static_assert_exception
+    {
+      eigen_static_assert_exception(void) {}
+      ~eigen_static_assert_exception() { Eigen::no_more_assert = false; }
+    };
   }
   // If EIGEN_DEBUG_ASSERTS is defined and if no assertion is triggered while
   // one should have been, then the list of excecuted assertions is printed out.
@@ -238,6 +244,7 @@ namespace Eigen
         else                                  \
           EIGEN_THROW_X(Eigen::eigen_assert_exception()); \
       }
+
     #ifdef EIGEN_EXCEPTIONS
       #define VERIFY_RAISES_ASSERT(a) {                           \
         Eigen::no_more_assert = false;                            \
@@ -249,13 +256,39 @@ namespace Eigen
         catch (Eigen::eigen_assert_exception&) { VERIFY(true); }  \
         Eigen::report_on_cerr_on_assert_failure = true;           \
       }
-    #endif //EIGEN_EXCEPTIONS
+    #endif // EIGEN_EXCEPTIONS
   #endif // EIGEN_DEBUG_ASSERTS
 
+  #if defined(TEST_CHECK_STATIC_ASSERTIONS) && defined(EIGEN_EXCEPTIONS)
+    #define EIGEN_STATIC_ASSERT(a,MSG) \
+      if( (!Eigen::internal::copy_bool(a)) && (!no_more_assert) )\
+      {                                       \
+        Eigen::no_more_assert = true;         \
+        if(report_on_cerr_on_assert_failure)  \
+          eigen_plain_assert(a && #MSG);      \
+        else                                  \
+          EIGEN_THROW_X(Eigen::eigen_static_assert_exception()); \
+      }
+    #define VERIFY_RAISES_STATIC_ASSERT(a) {                    \
+      Eigen::no_more_assert = false;                            \
+      Eigen::report_on_cerr_on_assert_failure = false;          \
+      try {                                                     \
+        a;                                                      \
+        VERIFY(Eigen::should_raise_an_assert && # a);           \
+      }                                                         \
+      catch (Eigen::eigen_static_assert_exception&) { VERIFY(true); }  \
+      Eigen::report_on_cerr_on_assert_failure = true;           \
+    }
+  #endif // TEST_CHECK_STATIC_ASSERTIONS
+
 #ifndef VERIFY_RAISES_ASSERT
   #define VERIFY_RAISES_ASSERT(a) \
     std::cout << "Can't VERIFY_RAISES_ASSERT( " #a " ) with exceptions disabled\n";
 #endif
+#ifndef VERIFY_RAISES_STATIC_ASSERT
+  #define VERIFY_RAISES_STATIC_ASSERT(a) \
+    std::cout << "Can't VERIFY_RAISES_STATIC_ASSERT( " #a " ) with exceptions disabled\n";
+#endif
     
   #if !defined(__CUDACC__)
   #define EIGEN_USE_CUSTOM_ASSERT
@@ -264,10 +297,10 @@ namespace Eigen
 #else // EIGEN_NO_ASSERTION_CHECKING
 
   #define VERIFY_RAISES_ASSERT(a) {}
+  #define VERIFY_RAISES_STATIC_ASSERT(a) {}
 
 #endif // EIGEN_NO_ASSERTION_CHECKING
 
-
 #define EIGEN_INTERNAL_DEBUGGING
 #include <Eigen/QR> // required for createRandomPIMatrixOfRank
 
diff --git a/test/selfadjoint.cpp b/test/selfadjoint.cpp
index 5ac219352..aaa4888bd 100644
--- a/test/selfadjoint.cpp
+++ b/test/selfadjoint.cpp
@@ -7,7 +7,7 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#define EIGEN_NO_STATIC_ASSERT
+#define TEST_CHECK_STATIC_ASSERTIONS
 #include "main.h"
 
 // This file tests the basic selfadjointView API,
@@ -47,8 +47,8 @@ template<typename MatrixType> void selfadjoint(const MatrixType& m)
   m4 -= m1.template selfadjointView<Lower>();
   VERIFY_IS_APPROX(m4, m2-m3);
 
-  VERIFY_RAISES_ASSERT(m2.template selfadjointView<StrictlyUpper>());
-  VERIFY_RAISES_ASSERT(m2.template selfadjointView<UnitLower>());
+  VERIFY_RAISES_STATIC_ASSERT(m2.template selfadjointView<StrictlyUpper>());
+  VERIFY_RAISES_STATIC_ASSERT(m2.template selfadjointView<UnitLower>());
 }
 
 void bug_159()

From f7d17689a5e4609418c55ec47d7e5924e2785460 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Mar 2018 10:11:13 +0100
Subject: [PATCH 223/680] Add static assertion for fixed sizes Ref<>

---
 Eigen/src/Core/Ref.h |  2 ++
 test/ref.cpp         | 14 +++++++++++++-
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h
index abb1e5121..ac9502bc4 100644
--- a/Eigen/src/Core/Ref.h
+++ b/Eigen/src/Core/Ref.h
@@ -95,6 +95,8 @@ protected:
   template<typename Expression>
   EIGEN_DEVICE_FUNC void construct(Expression& expr)
   {
+    EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(PlainObjectType,Expression);
+
     if(PlainObjectType::RowsAtCompileTime==1)
     {
       eigen_assert(expr.rows()==1 || expr.cols()==1);
diff --git a/test/ref.cpp b/test/ref.cpp
index 769db0414..9dd2c0449 100644
--- a/test/ref.cpp
+++ b/test/ref.cpp
@@ -13,7 +13,7 @@
 #endif
 
 #define TEST_ENABLE_TEMPORARY_TRACKING
-
+#define TEST_CHECK_STATIC_ASSERTIONS
 #include "main.h"
 
 // test Ref.h
@@ -255,6 +255,17 @@ void test_ref_overloads()
   test_ref_ambiguous(A, B);
 }
 
+void test_ref_fixed_size_assert()
+{
+  Vector4f v4;
+  VectorXf vx(10);
+  VERIFY_RAISES_STATIC_ASSERT( Ref<Vector3f> y = v4; (void)y; );
+  VERIFY_RAISES_STATIC_ASSERT( Ref<Vector3f> y = vx.head<4>(); (void)y; );
+  VERIFY_RAISES_STATIC_ASSERT( Ref<const Vector3f> y = v4; (void)y; );
+  VERIFY_RAISES_STATIC_ASSERT( Ref<const Vector3f> y = vx.head<4>(); (void)y; );
+  VERIFY_RAISES_STATIC_ASSERT( Ref<const Vector3f> y = 2*v4; (void)y; );
+}
+
 void test_ref()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -277,4 +288,5 @@ void test_ref()
   }
   
   CALL_SUBTEST_7( test_ref_overloads() );
+  CALL_SUBTEST_7( test_ref_fixed_size_assert() );
 }

From e900b010c899b04b1710c5e99486b52494a40ea1 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 19 Mar 2018 09:04:54 -0700
Subject: [PATCH 224/680] Improve robustness of igamma and igammac to bad
 inputs.

Check for nan inputs and propagate them immediately. Limit the number of internal iterations to 2000 (same number as used by scipy.special.gammainc). This prevents an infinite loop when the function is called with nan or very large arguments.

Original change by mfirgunov@google.com
---
 .../src/SpecialFunctions/SpecialFunctionsImpl.h      | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 5d1b8fcc2..7deca3e12 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -535,6 +535,10 @@ struct igammac_impl {
       return nan;
     }
 
+    if (numext::isnan(a) || numext::isnan(x)) { // propagate nans
+      return nan;
+    }
+
     if ((x < one) || (x < a)) {
       /* The checks above ensure that we meet the preconditions for
        * igamma_impl::Impl(), so call it, rather than igamma_impl::Run().
@@ -592,7 +596,7 @@ struct igammac_impl {
     qkm1 = z * x;
     ans = pkm1 / qkm1;
 
-    while (true) {
+    for (int i = 0; i < 2000; i++) {
       c += one;
       y += one;
       z += two;
@@ -724,6 +728,10 @@ struct igamma_impl {
       return nan;
     }
 
+    if (numext::isnan(a) || numext::isnan(x)) { // propagate nans
+      return nan;
+    }
+
     if ((x > one) && (x > a)) {
       /* The checks above ensure that we meet the preconditions for
        * igammac_impl::Impl(), so call it, rather than igammac_impl::Run().
@@ -770,7 +778,7 @@ struct igamma_impl {
     c = one;
     ans = one;
 
-    while (true) {
+    for (int i = 0; i < 2000; i++) {
       r += one;
       c *= x/r;
       ans += c;

From 624df5094597ef4427ba8877dcf00804493160fe Mon Sep 17 00:00:00 2001
From: Basil Fierz <basil.fierz@hotmail.com>
Date: Thu, 26 Oct 2017 22:44:28 +0200
Subject: [PATCH 225/680] Adds missing EIGEN_STRONG_INLINE to support MSVC
 properly inlining small vector calculations

When working with MSVC often small vector operations are not properly inlined. This behaviour is observed even on the most recent compiler versions.
---
 Eigen/src/Core/Dot.h               | 17 ++++++++++-------
 Eigen/src/Core/Product.h           | 10 +++++-----
 Eigen/src/Core/ProductEvaluators.h | 20 ++++++++++----------
 Eigen/src/Core/Redux.h             |  2 +-
 4 files changed, 26 insertions(+), 23 deletions(-)

diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h
index bb8e3fecc..11da432b2 100644
--- a/Eigen/src/Core/Dot.h
+++ b/Eigen/src/Core/Dot.h
@@ -31,7 +31,8 @@ struct dot_nocheck
   typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
   typedef typename conj_prod::result_type ResScalar;
   EIGEN_DEVICE_FUNC
-  static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
+  EIGEN_STRONG_INLINE
+  static ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
     return a.template binaryExpr<conj_prod>(b).sum();
   }
@@ -43,7 +44,8 @@ struct dot_nocheck<T, U, true>
   typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
   typedef typename conj_prod::result_type ResScalar;
   EIGEN_DEVICE_FUNC
-  static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
+  EIGEN_STRONG_INLINE
+  static ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
     return a.transpose().template binaryExpr<conj_prod>(b).sum();
   }
@@ -65,6 +67,7 @@ struct dot_nocheck<T, U, true>
 template<typename Derived>
 template<typename OtherDerived>
 EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE
 typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
 MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
 {
@@ -102,7 +105,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename NumTraits<typename internal::trai
   * \sa lpNorm(), dot(), squaredNorm()
   */
 template<typename Derived>
-EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
 {
   return numext::sqrt(squaredNorm());
 }
@@ -117,7 +120,7 @@ EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::
   * \sa norm(), normalize()
   */
 template<typename Derived>
-EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::normalized() const
 {
   typedef typename internal::nested_eval<Derived,2>::type _Nested;
@@ -139,7 +142,7 @@ MatrixBase<Derived>::normalized() const
   * \sa norm(), normalized()
   */
 template<typename Derived>
-EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::normalize()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void MatrixBase<Derived>::normalize()
 {
   RealScalar z = squaredNorm();
   // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
@@ -160,7 +163,7 @@ EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::normalize()
   * \sa stableNorm(), stableNormalize(), normalized()
   */
 template<typename Derived>
-EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::stableNormalized() const
 {
   typedef typename internal::nested_eval<Derived,3>::type _Nested;
@@ -185,7 +188,7 @@ MatrixBase<Derived>::stableNormalized() const
   * \sa stableNorm(), stableNormalized(), normalize()
   */
 template<typename Derived>
-EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::stableNormalize()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void MatrixBase<Derived>::stableNormalize()
 {
   RealScalar w = cwiseAbs().maxCoeff();
   RealScalar z = (derived()/w).squaredNorm();
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index ae0c94b38..676c48027 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -97,8 +97,8 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option,
         && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
     }
 
-    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
-    EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); }
 
     EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; }
     EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; }
@@ -127,7 +127,7 @@ public:
   using Base::derived;
   typedef typename Base::Scalar Scalar;
   
-  operator const Scalar() const
+  EIGEN_STRONG_INLINE operator const Scalar() const
   {
     return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
   }
@@ -162,7 +162,7 @@ class ProductImpl<Lhs,Rhs,Option,Dense>
     
   public:
   
-    EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar coeff(Index row, Index col) const
     {
       EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
       eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
@@ -170,7 +170,7 @@ class ProductImpl<Lhs,Rhs,Option,Dense>
       return internal::evaluator<Derived>(derived()).coeff(row,col);
     }
 
-    EIGEN_DEVICE_FUNC Scalar coeff(Index i) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar coeff(Index i) const
     {
       EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
       eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 86966abdb..fb637191d 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -32,7 +32,7 @@ struct evaluator<Product<Lhs, Rhs, Options> >
   typedef Product<Lhs, Rhs, Options> XprType;
   typedef product_evaluator<XprType> Base;
   
-  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr) : Base(xpr) {}
 };
  
 // Catch "scalar * ( A * B )" and transform it to "(A*scalar) * B"
@@ -55,7 +55,7 @@ struct evaluator<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
                                const Product<Lhs, Rhs, DefaultProduct> > XprType;
   typedef evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> > Base;
 
-  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr)
     : Base(xpr.lhs().functor().m_other * xpr.rhs().lhs() * xpr.rhs().rhs())
   {}
 };
@@ -68,7 +68,7 @@ struct evaluator<Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> >
   typedef Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> XprType;
   typedef evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> > Base;
   
-  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr)
     : Base(Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex>(
         Product<Lhs, Rhs, LazyProduct>(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()),
         xpr.index() ))
@@ -246,19 +246,19 @@ template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
 {
   template<typename Dst>
-  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
   }
   
   template<typename Dst>
-  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum();
   }
   
   template<typename Dst>
-  static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); }
 };
 
@@ -312,25 +312,25 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
   };
   
   template<typename Dst>
-  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>());
   }
diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index 2b5b73bf7..0a99acb21 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -407,7 +407,7 @@ protected:
   */
 template<typename Derived>
 template<typename Func>
-EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
   eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");

From 672bdc126b0923e6228a024ce62d1f18b05840ea Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 16 Nov 2017 17:55:24 +0100
Subject: [PATCH 226/680] bug #1479: fix failure detection in LDLT

---
 Eigen/src/Cholesky/LDLT.h |  2 ++
 test/cholesky.cpp         | 17 +++++++++++++++++
 2 files changed, 19 insertions(+)

diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h
index 968427b3a..13a8f6d14 100644
--- a/Eigen/src/Cholesky/LDLT.h
+++ b/Eigen/src/Cholesky/LDLT.h
@@ -375,6 +375,8 @@ template<> struct ldlt_inplace<Lower>
 
       if((rs>0) && pivot_is_valid)
         A21 /= realAkk;
+      else if(rs>0)
+        ret = ret && (A21.array()==Scalar(0)).all();
 
       if(found_zero_pivot && pivot_is_valid) ret = false; // factorization failed
       else if(!pivot_is_valid) found_zero_pivot = true;
diff --git a/test/cholesky.cpp b/test/cholesky.cpp
index 8ad5ac639..b4b6bda7d 100644
--- a/test/cholesky.cpp
+++ b/test/cholesky.cpp
@@ -373,6 +373,7 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m)
     VERIFY(ldlt.info()==Success);
     VERIFY(!ldlt.isNegative());
     VERIFY(!ldlt.isPositive());
+    VERIFY_IS_APPROX(mat,ldlt.reconstructedMatrix());
   }
   {
     mat << 1, 2, 2, 1;
@@ -380,6 +381,7 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m)
     VERIFY(ldlt.info()==Success);
     VERIFY(!ldlt.isNegative());
     VERIFY(!ldlt.isPositive());
+    VERIFY_IS_APPROX(mat,ldlt.reconstructedMatrix());
   }
   {
     mat << 0, 0, 0, 0;
@@ -387,6 +389,7 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m)
     VERIFY(ldlt.info()==Success);
     VERIFY(ldlt.isNegative());
     VERIFY(ldlt.isPositive());
+    VERIFY_IS_APPROX(mat,ldlt.reconstructedMatrix());
   }
   {
     mat << 0, 0, 0, 1;
@@ -394,6 +397,7 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m)
     VERIFY(ldlt.info()==Success);
     VERIFY(!ldlt.isNegative());
     VERIFY(ldlt.isPositive());
+    VERIFY_IS_APPROX(mat,ldlt.reconstructedMatrix());
   }
   {
     mat << -1, 0, 0, 0;
@@ -401,6 +405,7 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m)
     VERIFY(ldlt.info()==Success);
     VERIFY(ldlt.isNegative());
     VERIFY(!ldlt.isPositive());
+    VERIFY_IS_APPROX(mat,ldlt.reconstructedMatrix());
   }
 }
 
@@ -452,6 +457,18 @@ void cholesky_faillure_cases()
     VERIFY(ldlt.info()==NumericalIssue);
     VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix());
   }
+
+  // bug 1479
+  {
+    mat.resize(4,4);
+    mat <<  1, 2, 0, 1,
+            2, 4, 0, 2,
+            0, 0, 0, 1,
+            1, 2, 1, 1;
+    ldlt.compute(mat);
+    VERIFY(ldlt.info()==NumericalIssue);
+    VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix());
+  }
 }
 
 template<typename MatrixType> void cholesky_verify_assert()

From 599a88da273ecaf00fec7cf9263a574ed44fef4f Mon Sep 17 00:00:00 2001
From: Zvi Rackover <zvirack@gmail.com>
Date: Thu, 16 Nov 2017 19:53:38 +0000
Subject: [PATCH 227/680] Disable gcc-specific workaround for Clang to allow
 build with AVX512

There is currently a workaround for an issue in gcc that requires invoking gcc with the -fabi-version flag. This workaround is not needed for Clang and moreover is not supported.
---
 CMakeLists.txt | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 62bf823c8..8a068739b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -232,7 +232,10 @@ if(NOT MSVC)
 
   option(EIGEN_TEST_AVX512 "Enable/Disable AVX512 in tests/examples" OFF)
   if(EIGEN_TEST_AVX512)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -fabi-version=6 -DEIGEN_ENABLE_AVX512")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -DEIGEN_ENABLE_AVX512")
+    if (NOT "${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fabi-version=6")
+    endif()
     message(STATUS "Enabling AVX512 in tests/examples")
   endif()
 

From 3dc6ff73cae64da97244e493905de7be40d1bcde Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 17 Nov 2017 22:54:39 +0100
Subject: [PATCH 228/680] Handle PGI compiler

---
 cmake/EigenTesting.cmake | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 8a995f118..4a34ddef5 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -549,6 +549,8 @@ macro(ei_get_compilerver VAR)
       else()
         set(${VAR} "na")
       endif()
+    elseif(${CMAKE_CXX_COMPILER_ID} MATCHES "PGI")
+      set(${VAR} "${CMAKE_CXX_COMPILER_ID}-${CMAKE_CXX_COMPILER_VERSION}")
     else()
     # on all other system we rely on ${CMAKE_CXX_COMPILER}
     # supporting a "--version" or "/version" flag

From dd6de618c3fda4275aff3a57c590f82b6e628ac1 Mon Sep 17 00:00:00 2001
From: nluehr <nluehr@nvidia.com>
Date: Tue, 21 Nov 2017 10:47:00 -0800
Subject: [PATCH 229/680] Fix incorrect integer cast in predux<half2>().

Bug corrupts results on Maxwell and earlier GPU architectures.
---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index ce48e4b31..e3e45b03a 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -231,7 +231,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
 #else
   float a1 = __low2float(a);
   float a2 = __high2float(a);
-  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 + a2)));
+  return Eigen::half(__float2half_rn(a1 + a2));
 #endif
 }
 
@@ -265,7 +265,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const ha
 #else
   float a1 = __low2float(a);
   float a2 = __high2float(a);
-  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 * a2)));
+  return Eigen::half(__float2half_rn(a1 * a2));
 #endif
 }
 

From 3587e481fb58710272d4aa71d082bf0669f03698 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 27 Nov 2017 21:53:02 +0100
Subject: [PATCH 230/680] silent MSVC warning

---
 Eigen/src/Core/AssignEvaluator.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h
index dbe435d86..ebf5590de 100644
--- a/Eigen/src/Core/AssignEvaluator.h
+++ b/Eigen/src/Core/AssignEvaluator.h
@@ -97,7 +97,7 @@ private:
 
 public:
   enum {
-    Traversal = int(MayLinearVectorize) && (LinearPacketSize>InnerPacketSize) ? int(LinearVectorizedTraversal)
+    Traversal = (int(MayLinearVectorize) && (LinearPacketSize>InnerPacketSize)) ? int(LinearVectorizedTraversal)
               : int(MayInnerVectorize)   ? int(InnerVectorizedTraversal)
               : int(MayLinearVectorize)  ? int(LinearVectorizedTraversal)
               : int(MaySliceVectorize)   ? int(SliceVectorizedTraversal)

From d0b028e1731e7670c5ace4f79c8d48137dd55483 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 27 Nov 2017 22:11:57 +0100
Subject: [PATCH 231/680] clarify Pastix requirements

---
 Eigen/PaStiXSupport | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/PaStiXSupport b/Eigen/PaStiXSupport
index de3a63b4d..234619acc 100644
--- a/Eigen/PaStiXSupport
+++ b/Eigen/PaStiXSupport
@@ -36,6 +36,7 @@ extern "C" {
   * \endcode
   *
   * In order to use this module, the PaSTiX headers must be accessible from the include paths, and your binary must be linked to the PaSTiX library and its dependencies.
+  * This wrapper resuires PaStiX version 5.x compiled without MPI support.
   * The dependencies depend on how PaSTiX has been compiled.
   * For a cmake based project, you can use our FindPaSTiX.cmake module to help you in this task.
   *

From aefd5fd5c4331f0265abb692d4742b558f13f01d Mon Sep 17 00:00:00 2001
From: nluehr <nluehr@nvidia.com>
Date: Tue, 28 Nov 2017 10:15:46 -0800
Subject: [PATCH 232/680] Replace __float2half_rn with __float2half

The latter provides a consistent definition for CUDA 8.0 and 9.0.
---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index e3e45b03a..8a6f209c4 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -231,7 +231,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
 #else
   float a1 = __low2float(a);
   float a2 = __high2float(a);
-  return Eigen::half(__float2half_rn(a1 + a2));
+  return Eigen::half(__float2half(a1 + a2));
 #endif
 }
 
@@ -265,7 +265,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const ha
 #else
   float a1 = __low2float(a);
   float a2 = __high2float(a);
-  return Eigen::half(__float2half_rn(a1 * a2));
+  return Eigen::half(__float2half(a1 * a2));
 #endif
 }
 

From 3122477c8660f4e66e9cf4bf24e4fdfd6d56378c Mon Sep 17 00:00:00 2001
From: Yangzihao Wang <yangzihao@google.com>
Date: Tue, 12 Dec 2017 11:15:24 -0800
Subject: [PATCH 233/680] Update the padding computation for PADDING_SAME to be
 consistent with TensorFlow.

---
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |  4 ++
 unsupported/test/cxx11_tensor_image_patch.cpp | 52 +++++++++++++++++++
 2 files changed, 56 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 3c6a2e091..91d4ead28 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -265,6 +265,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
           // Calculate the padding
           m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2;
           m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2;
+          // The padding size calculation for PADDING_SAME has been updated to
+          // be consistent with how TensorFlow extracts its paddings.
+          m_rowPaddingTop = numext::maxi<Index>(0, m_rowPaddingTop);
+          m_colPaddingLeft = numext::maxi<Index>(0, m_colPaddingLeft);
           break;
         default:
           eigen_assert(false && "unexpected padding");
diff --git a/unsupported/test/cxx11_tensor_image_patch.cpp b/unsupported/test/cxx11_tensor_image_patch.cpp
index 475c59651..105d32fb4 100644
--- a/unsupported/test/cxx11_tensor_image_patch.cpp
+++ b/unsupported/test/cxx11_tensor_image_patch.cpp
@@ -405,6 +405,57 @@ void test_patch_padding_same()
   }
 }
 
+// Verifies that SAME padding, when computed as negative values, will be clipped
+// to zero.
+void test_patch_padding_same_negative_padding_clip_to_zero() {
+  int input_depth = 1;
+  int input_rows = 15;
+  int input_cols = 1;
+  int input_batches = 1;
+  int ksize = 1;  // Corresponds to the Rows and Cols for
+                  // tensor.extract_image_patches<>.
+  int row_stride = 5;
+  int col_stride = 1;
+  // ColMajor
+  Tensor<float, 4> tensor(input_depth, input_rows, input_cols, input_batches);
+  // Initializes tensor with incrementing numbers.
+  for (int i = 0; i < tensor.size(); ++i) {
+    tensor.data()[i] = i + 1;
+  }
+  Tensor<float, 5> result = tensor.extract_image_patches(
+      ksize, ksize, row_stride, col_stride, 1, 1, PADDING_SAME);
+  // row padding will be computed as -2 originally and then be clipped to 0.
+  VERIFY_IS_EQUAL(result.coeff(0), 1.0f);
+  VERIFY_IS_EQUAL(result.coeff(1), 6.0f);
+  VERIFY_IS_EQUAL(result.coeff(2), 11.0f);
+
+  VERIFY_IS_EQUAL(result.dimension(0), input_depth);    // depth
+  VERIFY_IS_EQUAL(result.dimension(1), ksize);          // kernel rows
+  VERIFY_IS_EQUAL(result.dimension(2), ksize);          // kernel cols
+  VERIFY_IS_EQUAL(result.dimension(3), 3);              // number of patches
+  VERIFY_IS_EQUAL(result.dimension(4), input_batches);  // number of batches
+
+  // RowMajor
+  Tensor<float, 4, RowMajor> tensor_row_major = tensor.swap_layout();
+  VERIFY_IS_EQUAL(tensor.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor.dimension(3), tensor_row_major.dimension(0));
+
+  Tensor<float, 5, RowMajor> result_row_major =
+      tensor_row_major.extract_image_patches(ksize, ksize, row_stride,
+                                             col_stride, 1, 1, PADDING_SAME);
+  VERIFY_IS_EQUAL(result_row_major.coeff(0), 1.0f);
+  VERIFY_IS_EQUAL(result_row_major.coeff(1), 6.0f);
+  VERIFY_IS_EQUAL(result_row_major.coeff(2), 11.0f);
+
+  VERIFY_IS_EQUAL(result.dimension(0), result_row_major.dimension(4));
+  VERIFY_IS_EQUAL(result.dimension(1), result_row_major.dimension(3));
+  VERIFY_IS_EQUAL(result.dimension(2), result_row_major.dimension(2));
+  VERIFY_IS_EQUAL(result.dimension(3), result_row_major.dimension(1));
+  VERIFY_IS_EQUAL(result.dimension(4), result_row_major.dimension(0));
+}
+
 void test_patch_no_extra_dim()
 {
   Tensor<float, 3> tensor(2,3,5);
@@ -754,4 +805,5 @@ void test_cxx11_tensor_image_patch()
   CALL_SUBTEST_4(test_patch_padding_valid_same_value());
   CALL_SUBTEST_5(test_patch_padding_same());
   CALL_SUBTEST_6(test_imagenet_patches());
+  CALL_SUBTEST_7(test_patch_padding_same_negative_padding_clip_to_zero());
 }

From b2cacd189e2d3c2d841207029c399b80eecadd01 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 10:01:02 +0100
Subject: [PATCH 234/680] fix header inclusion

---
 test/klu_support.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/klu_support.cpp b/test/klu_support.cpp
index 8b1fdeb41..138dcc301 100644
--- a/test/klu_support.cpp
+++ b/test/klu_support.cpp
@@ -10,7 +10,7 @@
 #define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS
 #include "sparse_solver.h"
 
-#include <unsupported/Eigen/KLUSupport>
+#include <Eigen/KLUSupport>
 
 template<typename T> void test_klu_support_T()
 {

From 76c7dae600efc26a6fda212be518b804f101244f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 14:22:14 +0100
Subject: [PATCH 235/680] ignore all *build* sub directories

---
 .hgignore | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.hgignore b/.hgignore
index dcd9f4431..ebbf746bf 100644
--- a/.hgignore
+++ b/.hgignore
@@ -13,7 +13,7 @@ core
 core.*
 *.bak
 *~
-build*
+*build*
 *.moc.*
 *.moc
 ui_*

From 9c3aed9d48d7dbc0f88d2fb92ca232dcbf0d402e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 14:24:33 +0100
Subject: [PATCH 236/680] Fix packet and alignment propagation logic of
 Block<Xpr> expressions. In particular, (A+B).col(j) lost vectorisation.

---
 Eigen/src/Core/CoreEvaluators.h | 6 ++++--
 test/vectorization_logic.cpp    | 6 ++++++
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 15b361b38..dfdb63e98 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1034,7 +1034,7 @@ struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
     OuterStrideAtCompileTime = HasSameStorageOrderAsArgType
                              ? int(outer_stride_at_compile_time<ArgType>::ret)
                              : int(inner_stride_at_compile_time<ArgType>::ret),
-    MaskPacketAccessBit = (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0,
+    MaskPacketAccessBit = (InnerStrideAtCompileTime == 1 || HasSameStorageOrderAsArgType) ? PacketAccessBit : 0,
     
     FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator<ArgType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,    
     FlagsRowMajorBit = XprType::Flags&RowMajorBit,
@@ -1044,7 +1044,9 @@ struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
     Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit,
     
     PacketAlignment = unpacket_traits<PacketScalar>::alignment,
-    Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0,
+    Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic)
+                             && (OuterStrideAtCompileTime!=0)
+                             && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0,
     Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ArgType>::Alignment, Alignment0)
   };
   typedef block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> block_evaluator_type;
diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp
index 83c1439ad..37e7495f5 100644
--- a/test/vectorization_logic.cpp
+++ b/test/vectorization_logic.cpp
@@ -207,6 +207,12 @@ struct vectorization_logic
     VERIFY(test_redux(Vector1(),
       LinearVectorizedTraversal,CompleteUnrolling));
 
+    VERIFY(test_redux(Vector1().array()*Vector1().array(),
+      LinearVectorizedTraversal,CompleteUnrolling));
+
+    VERIFY(test_redux((Vector1().array()*Vector1().array()).col(0),
+      LinearVectorizedTraversal,CompleteUnrolling));
+
     VERIFY(test_redux(Matrix<Scalar,PacketSize,3>(),
       LinearVectorizedTraversal,CompleteUnrolling));
 

From 546ab97d7679c75fadbf9e003c7c75faa9a179e5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 14:47:38 +0100
Subject: [PATCH 237/680] Add possibility to overwrite EIGEN_STRONG_INLINE.

---
 Eigen/src/Core/util/Macros.h   | 2 ++
 doc/PreprocessorDirectives.dox | 4 ++++
 2 files changed, 6 insertions(+)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 3a56b0e36..e351b7ad9 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -514,11 +514,13 @@
 // EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC,
 // but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline
 // but GCC is still doing fine with just inline.
+#ifndef EIGEN_STRONG_INLINE
 #if EIGEN_COMP_MSVC || EIGEN_COMP_ICC
 #define EIGEN_STRONG_INLINE __forceinline
 #else
 #define EIGEN_STRONG_INLINE inline
 #endif
+#endif
 
 // EIGEN_ALWAYS_INLINE is the stronget, it has the effect of making the function inline and adding every possible
 // attribute to maximize inlining. This should only be used when really necessary: in particular,
diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox
index 0919d4190..b6d08c700 100644
--- a/doc/PreprocessorDirectives.dox
+++ b/doc/PreprocessorDirectives.dox
@@ -122,6 +122,10 @@ run time. However, these assertions do cost time and can thus be turned off.
    this threshold raises a compile time assertion. Use 0 to set no limit. Default is 128 KB.
  - \b \c EIGEN_NO_CUDA - disables CUDA support when defined. Might be useful in .cu files for which Eigen is used on the host only,
    and never called from device code.
+ - \b \c EIGEN_STRONG_INLINE - This macro is used to qualify critical functions and methods that we expect the compiler to inline.
+   By default it is defined to \c __forceinline for MSVC and ICC, and to \c inline for other compilers. A tipical usage is to
+   define it to \c inline for MSVC users wanting faster compilation times, at the risk of performance degradations in some rare
+   cases for which MSVC inliner fails to do a good job.
 
 
  - \c EIGEN_DONT_ALIGN - Deprecated, it is a synonym for \c EIGEN_MAX_ALIGN_BYTES=0. It disables alignment completely. %Eigen will not try to align its objects and does not expect that any objects passed to it are aligned. This will turn off vectorization if \b EIGEN_UNALIGNED_VECTORIZE=1. Not defined by default.

From 26a2c6fc1676d75f4360a993461d8e2778d454f2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 15:11:04 +0100
Subject: [PATCH 238/680] fix unit test

---
 test/sparse_basic.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp
index 384985028..f84b6e3f5 100644
--- a/test/sparse_basic.cpp
+++ b/test/sparse_basic.cpp
@@ -228,8 +228,8 @@ template<typename SparseMatrixType> void sparse_basic(const SparseMatrixType& re
       VERIFY_RAISES_ASSERT( m1 -= m1.innerVector(0) );
       VERIFY_RAISES_ASSERT( refM1 -= m1.innerVector(0) );
       VERIFY_RAISES_ASSERT( refM1 += m1.innerVector(0) );
-      m1 = m4; refM1 = refM4;
     }
+    m1 = m4; refM1 = refM4;
 
     // test aliasing
     VERIFY_IS_APPROX((m1 = -m1), (refM1 = -refM1));

From 31e0bda2e3b44b908e63f3b19ade2f4af12c7e10 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 14 Dec 2017 15:48:27 +0100
Subject: [PATCH 239/680] Fix cmake warning

---
 bench/spbench/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/bench/spbench/CMakeLists.txt b/bench/spbench/CMakeLists.txt
index 932735698..029ba6d6b 100644
--- a/bench/spbench/CMakeLists.txt
+++ b/bench/spbench/CMakeLists.txt
@@ -60,7 +60,7 @@ if(PASTIX_FOUND AND PASTIX_pastix_nompi.h_INCLUDE_DIRS AND BLAS_FOUND)
   endif(SCOTCH_FOUND)
   set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES_DEP} ${ORDERING_LIBRARIES})
   set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES_DEP})
-endif(PASTIX_FOUND AND BLAS_FOUND)
+endif()
 
 if(METIS_FOUND)
   include_directories(${METIS_INCLUDE_DIRS})

From 73214c4bd06108ee3a26204acbbda205a40131bc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 15 Dec 2017 14:10:59 +0100
Subject: [PATCH 240/680] Workaround nvcc 9.0 issue. See PR 351.
 https://bitbucket.org/eigen/eigen/pull-requests/351

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 0d6331e9c..1d459a3a0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -836,7 +836,8 @@ class TensorBase<Derived, ReadOnlyAccessors>
   protected:
     template <typename Scalar, int NumIndices, int Options, typename IndexType> friend class Tensor;
     template <typename Scalar, typename Dimensions, int Option, typename IndexTypes> friend class TensorFixedSize;
-    template <typename OtherDerived, int AccessLevel> friend class TensorBase;
+    // the Eigen:: prefix is required to workaround a compilation issue with nvcc 9.0
+    template <typename OtherDerived, int AccessLevel> friend class Eigen::TensorBase;
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast<const Derived*>(this); }
 };
@@ -852,7 +853,8 @@ class TensorBase : public TensorBase<Derived, ReadOnlyAccessors> {
 
     template <typename Scalar, int NumIndices, int Options, typename IndexType> friend class Tensor;
     template <typename Scalar, typename Dimensions, int Option, typename IndexTypes> friend class TensorFixedSize;
-    template <typename OtherDerived, int OtherAccessLevel> friend class TensorBase;
+    // the Eigen:: prefix is required to workaround a compilation issue with nvcc 9.0
+    template <typename OtherDerived, int OtherAccessLevel> friend class Eigen::TensorBase;
 
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& setZero() {

From 06bf1047f99afec61062e188ec4e68efbf203d86 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 15 Dec 2017 15:15:37 +0100
Subject: [PATCH 241/680] Fix compilation of stableNorm with some expressions
 as input

---
 Eigen/src/Core/StableNorm.h |  2 +-
 test/stable_norm.cpp        | 10 ++++++++++
 2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h
index 4ea598ba8..e0e7a0c8f 100644
--- a/Eigen/src/Core/StableNorm.h
+++ b/Eigen/src/Core/StableNorm.h
@@ -165,7 +165,7 @@ MatrixBase<Derived>::stableNorm() const
   
   typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
   typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
-  DerivedCopy copy(derived());
+  const DerivedCopy copy(derived());
   
   enum {
     CanAlign = (   (int(DerivedCopyClean::Flags)&DirectAccessBit)
diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp
index c3eb5ff31..3c02474b8 100644
--- a/test/stable_norm.cpp
+++ b/test/stable_norm.cpp
@@ -65,6 +65,8 @@ template<typename MatrixType> void stable_norm(const MatrixType& m)
     factor = internal::random<Scalar>();
   Scalar small = factor * ((std::numeric_limits<RealScalar>::min)() * RealScalar(1e4));
 
+  Scalar one(1);
+
   MatrixType  vzero = MatrixType::Zero(rows, cols),
               vrand = MatrixType::Random(rows, cols),
               vbig(rows, cols),
@@ -78,6 +80,14 @@ template<typename MatrixType> void stable_norm(const MatrixType& m)
   VERIFY_IS_APPROX(vrand.blueNorm(),        vrand.norm());
   VERIFY_IS_APPROX(vrand.hypotNorm(),       vrand.norm());
 
+  // test with expressions as input
+  VERIFY_IS_APPROX((one*vrand).stableNorm(),      vrand.norm());
+  VERIFY_IS_APPROX((one*vrand).blueNorm(),        vrand.norm());
+  VERIFY_IS_APPROX((one*vrand).hypotNorm(),       vrand.norm());
+  VERIFY_IS_APPROX((one*vrand+one*vrand-one*vrand).stableNorm(),      vrand.norm());
+  VERIFY_IS_APPROX((one*vrand+one*vrand-one*vrand).blueNorm(),        vrand.norm());
+  VERIFY_IS_APPROX((one*vrand+one*vrand-one*vrand).hypotNorm(),       vrand.norm());
+
   RealScalar size = static_cast<RealScalar>(m.size());
 
   // test numext::isfinite

From f9bdcea022e24bac4a66a937c37de92f7f22b9da Mon Sep 17 00:00:00 2001
From: nluehr <nluehr@nvidia.com>
Date: Mon, 18 Dec 2017 16:51:15 -0800
Subject: [PATCH 242/680] For cuda 9.1 replace math_functions.hpp with
 cuda_runtime.h

---
 Eigen/Core | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index c66359b79..5a6dec8cc 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -54,9 +54,9 @@
     #endif
 
     #define EIGEN_DEVICE_FUNC __host__ __device__
-    // We need math_functions.hpp to ensure that that EIGEN_USING_STD_MATH macro
+    // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
     // works properly on the device side
-    #include <math_functions.hpp>
+    #include <cuda_runtime.h>
   #else
     #define EIGEN_DEVICE_FUNC
   #endif

From 59985cfd26416fb6b196af868c187e90d237c352 Mon Sep 17 00:00:00 2001
From: RJ Ryan <rryan@mixxx.org>
Date: Sun, 31 Dec 2017 10:44:56 -0500
Subject: [PATCH 243/680] Disable use of recurrence for computing twiddle
 factors. Fixes FFT precision issues for large FFTs.
 https://github.com/tensorflow/tensorflow/issues/10749#issuecomment-354557689

---
 .../Eigen/CXX11/src/Tensor/TensorFFT.h        | 40 ++++++++++++-------
 unsupported/test/cxx11_tensor_fft.cpp         | 28 +++++++++++++
 2 files changed, 54 insertions(+), 14 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
index 10e0a8a6b..f81da318c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
@@ -231,20 +231,32 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
         //   t_n = exp(sqrt(-1) * pi * n^2 / line_len)
         // for n = 0, 1,..., line_len-1.
         // For n > 2 we use the recurrence t_n = t_{n-1}^2 / t_{n-2} * t_1^2
-        pos_j_base_powered[0] = ComplexScalar(1, 0);
-        if (line_len > 1) {
-          const RealScalar pi_over_len(EIGEN_PI / line_len);
-          const ComplexScalar pos_j_base = ComplexScalar(
-              std::cos(pi_over_len), std::sin(pi_over_len));
-          pos_j_base_powered[1] = pos_j_base;
-          if (line_len > 2) {
-            const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
-            for (int j = 2; j < line_len + 1; ++j) {
-              pos_j_base_powered[j] = pos_j_base_powered[j - 1] *
-                                      pos_j_base_powered[j - 1] /
-                                      pos_j_base_powered[j - 2] * pos_j_base_sq;
-            }
-          }
+
+        // The recurrence is correct in exact arithmetic, but causes
+        // numerical issues for large transforms, especially in
+        // single-precision floating point.
+        //
+        // pos_j_base_powered[0] = ComplexScalar(1, 0);
+        // if (line_len > 1) {
+        //   const ComplexScalar pos_j_base = ComplexScalar(
+        //       numext::cos(M_PI / line_len), numext::sin(M_PI / line_len));
+        //   pos_j_base_powered[1] = pos_j_base;
+        //   if (line_len > 2) {
+        //     const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
+        //     for (int i = 2; i < line_len + 1; ++i) {
+        //       pos_j_base_powered[i] = pos_j_base_powered[i - 1] *
+        //           pos_j_base_powered[i - 1] /
+        //           pos_j_base_powered[i - 2] *
+        //           pos_j_base_sq;
+        //     }
+        //   }
+        // }
+        // TODO(rmlarsen): Find a way to use Eigen's vectorized sin
+        // and cosine functions here.
+        for (int j = 0; j < line_len + 1; ++j) {
+          double arg = ((EIGEN_PI * j) * j) / line_len;
+          std::complex<double> tmp(numext::cos(arg), numext::sin(arg));
+          pos_j_base_powered[j] = static_cast<ComplexScalar>(tmp);
         }
       }
 
diff --git a/unsupported/test/cxx11_tensor_fft.cpp b/unsupported/test/cxx11_tensor_fft.cpp
index 2f14ebc62..a55369477 100644
--- a/unsupported/test/cxx11_tensor_fft.cpp
+++ b/unsupported/test/cxx11_tensor_fft.cpp
@@ -224,6 +224,32 @@ static void test_fft_real_input_energy() {
   }
 }
 
+template <typename RealScalar>
+static void test_fft_non_power_of_2_round_trip(int exponent) {
+  int n = (1 << exponent) + 1;
+
+  Eigen::DSizes<long, 1> dimensions;
+  dimensions[0] = n;
+  const DSizes<long, 1> arr = dimensions;
+  Tensor<RealScalar, 1, ColMajor, long> input;
+
+  input.resize(arr);
+  input.setRandom();
+
+  array<int, 1> fft;
+  fft[0] = 0;
+
+  Tensor<std::complex<RealScalar>, 1, ColMajor> forward =
+      input.template fft<BothParts, FFT_FORWARD>(fft);
+
+  Tensor<RealScalar, 1, ColMajor, long> output =
+      forward.template fft<RealPart, FFT_REVERSE>(fft);
+
+  for (int i = 0; i < n; ++i) {
+    VERIFY_IS_APPROX(input[i], output[i]);
+  }
+}
+
 void test_cxx11_tensor_fft() {
     test_fft_complex_input_golden();
     test_fft_real_input_golden();
@@ -270,4 +296,6 @@ void test_cxx11_tensor_fft() {
     test_fft_real_input_energy<RowMajor, double, true,  Eigen::BothParts, FFT_FORWARD, 4>();
     test_fft_real_input_energy<RowMajor, float,  false,  Eigen::BothParts, FFT_FORWARD, 4>();
     test_fft_real_input_energy<RowMajor, double, false,  Eigen::BothParts, FFT_FORWARD, 4>();
+
+    test_fft_non_power_of_2_round_trip<float>(7);
 }

From 73629f8b680fb4107b490ea99ce7dfc3b5332775 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 9 Jan 2018 08:59:27 +0100
Subject: [PATCH 244/680] Fix gcc7 warning

---
 Eigen/src/Core/CoreEvaluators.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index dfdb63e98..d1454117a 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -136,7 +136,9 @@ template<typename Scalar,int OuterStride> class plainobjectbase_evaluator_data {
 public:
   EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr)
   {
-    EIGEN_ONLY_USED_FOR_DEBUG(outerStride);
+#ifndef EIGEN_INTERNAL_DEBUGGING
+    EIGEN_UNUSED_VARIABLE(outerStride);
+#endif
     eigen_internal_assert(outerStride==OuterStride);
   }
   EIGEN_DEVICE_FUNC Index outerStride() const { return OuterStride; }

From f558ad2955ef4899f208883f46c410273e21451d Mon Sep 17 00:00:00 2001
From: Eugene Chereshnev <eugene.chereshnev@intel.com>
Date: Wed, 3 Jan 2018 12:55:52 -0800
Subject: [PATCH 245/680] Fix incorrect ldvt in LAPACKE call from JacobiSVD

---
 Eigen/src/SVD/JacobiSVD_LAPACKE.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/SVD/JacobiSVD_LAPACKE.h b/Eigen/src/SVD/JacobiSVD_LAPACKE.h
index 50272154f..ff0516f61 100644
--- a/Eigen/src/SVD/JacobiSVD_LAPACKE.h
+++ b/Eigen/src/SVD/JacobiSVD_LAPACKE.h
@@ -61,9 +61,10 @@ JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPiv
     u    = (LAPACKE_TYPE*)m_matrixU.data(); \
   } else { ldu=1; u=&dummy; }\
   MatrixType localV; \
-  ldvt = (m_computeFullV) ? internal::convert_index<lapack_int>(m_cols) : (m_computeThinV) ? internal::convert_index<lapack_int>(m_diagSize) : 1; \
+  lapack_int vt_rows = (m_computeFullV) ? internal::convert_index<lapack_int>(m_cols) : (m_computeThinV) ? internal::convert_index<lapack_int>(m_diagSize) : 1; \
   if (computeV()) { \
-    localV.resize(ldvt, m_cols); \
+    localV.resize(vt_rows, m_cols); \
+    ldvt  = internal::convert_index<lapack_int>(localV.outerStride()); \
     vt   = (LAPACKE_TYPE*)localV.data(); \
   } else { ldvt=1; vt=&dummy; }\
   Matrix<LAPACKE_RTYPE, Dynamic, Dynamic> superb; superb.resize(m_diagSize, 1); \

From 5b3c367926bf0e55cb2ec5359605349909f9f7e4 Mon Sep 17 00:00:00 2001
From: "Lee.Deokjae" <lee.deokjae@gmail.com>
Date: Sat, 6 Jan 2018 14:36:19 +0900
Subject: [PATCH 246/680] Fix typos in the contraction example of tensor README

---
 unsupported/Eigen/CXX11/src/Tensor/README.md | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index 30d553af7..49cc33c5d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -1013,23 +1013,23 @@ multidimensional case.
     Eigen::Tensor<int, 2> a(2, 3);
     a.setValues({{1, 2, 3}, {6, 5, 4}});
     Eigen::Tensor<int, 2> b(3, 2);
-    a.setValues({{1, 2}, {4, 5}, {5, 6}});
+    b.setValues({{1, 2}, {4, 5}, {5, 6}});
 
     // Compute the traditional matrix product
-    Eigen::array<Eigen::IndexPair<int>, 1> product_dims = { Eigen::IndexPair(1, 0) };
+    Eigen::array<Eigen::IndexPair<int>, 1> product_dims = { Eigen::IndexPair<int>(1, 0) };
     Eigen::Tensor<int, 2> AB = a.contract(b, product_dims);
 
     // Compute the product of the transpose of the matrices
-    Eigen::array<Eigen::IndexPair<int>, 1> transpose_product_dims = { Eigen::IndexPair(0, 1) };
+    Eigen::array<Eigen::IndexPair<int>, 1> transposed_product_dims = { Eigen::IndexPair<int>(0, 1) };
     Eigen::Tensor<int, 2> AtBt = a.contract(b, transposed_product_dims);
-    
-    // Contraction to scalar value using a ouble contraction 
-    // First coordinate of both tensors are contracted as well as both second coordinates
+
+    // Contraction to scalar value using a double contraction.
+    // First coordinate of both tensors are contracted as well as both second coordinates, i.e., this computes the sum of the squares of the elements.
     Eigen::array<Eigen::IndexPair<int>, 2> double_contraction_product_dims = { Eigen::IndexPair<int>(0, 0), Eigen::IndexPair<int>(1, 1) };
-    Eigen::Tensor<int, 0> AdoubleontractedA = a.contract(a, double_contraction_product_dims);
-    
+    Eigen::Tensor<int, 0> AdoubleContractedA = a.contract(a, double_contraction_product_dims);
+
     // Extracting the scalar value of the tensor contraction for further usage
-    int value = AdoublecontractedA(0);
+    int value = AdoubleContractedA(0);
 
 ## Reduction Operations
 

From 09a16ba42fa1acc7bb0ace489ee51b3eb958ffa0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 17 Jan 2018 23:13:16 +0100
Subject: [PATCH 247/680] bug #1412: fix compilation with nvcc+MSVC

---
 Eigen/src/Geometry/Scaling.h | 24 +++++++++++++++++++++---
 Eigen/src/SVD/BDCSVD.h       | 12 ++++++------
 Eigen/src/SVD/JacobiSVD.h    |  2 +-
 3 files changed, 28 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Geometry/Scaling.h b/Eigen/src/Geometry/Scaling.h
index f58ca03d9..c1899a029 100755
--- a/Eigen/src/Geometry/Scaling.h
+++ b/Eigen/src/Geometry/Scaling.h
@@ -29,6 +29,22 @@ namespace Eigen {
   *
   * \sa Scaling(), class DiagonalMatrix, MatrixBase::asDiagonal(), class Translation, class Transform
   */
+
+namespace internal
+{
+  // This helper helps nvcc+MSVC to properly parse this file.
+  // See bug 1412.
+  template <typename Scalar, int Dim, int Mode>
+  struct uniformscaling_times_affine_returntype
+  {
+    enum
+    {
+      NewMode = int(Mode) == int(Isometry) ? Affine : Mode
+    };
+    typedef Transform <Scalar, Dim, NewMode> type;
+  };
+}
+
 template<typename _Scalar>
 class UniformScaling
 {
@@ -60,9 +76,11 @@ public:
 
   /** Concatenates a uniform scaling and an affine transformation */
   template<int Dim, int Mode, int Options>
-  inline Transform<Scalar,Dim,(int(Mode)==int(Isometry)?Affine:Mode)> operator* (const Transform<Scalar,Dim, Mode, Options>& t) const
+  inline typename
+	internal::uniformscaling_times_affine_returntype <Scalar,Dim,Mode>::type
+	operator* (const Transform<Scalar, Dim, Mode, Options>& t) const
   {
-    Transform<Scalar,Dim,(int(Mode)==int(Isometry)?Affine:Mode)> res = t;
+    typename internal::uniformscaling_times_affine_returntype <Scalar,Dim,Mode> res = t;
     res.prescale(factor());
     return res;
   }
@@ -70,7 +88,7 @@ public:
   /** Concatenates a uniform scaling and a linear transformation matrix */
   // TODO returns an expression
   template<typename Derived>
-  inline typename internal::plain_matrix_type<Derived>::type operator* (const MatrixBase<Derived>& other) const
+  inline typename Eigen::internal::plain_matrix_type<Derived>::type operator* (const MatrixBase<Derived>& other) const
   { return other * m_factor; }
 
   template<typename Derived,int Dim>
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index 0abd4c1bb..06865a331 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -217,7 +217,7 @@ public:
 
 // Method to allocate and initialize matrix and attributes
 template<typename MatrixType>
-void BDCSVD<MatrixType>::allocate(Index rows, Index cols, unsigned int computationOptions)
+void BDCSVD<MatrixType>::allocate(Eigen::Index rows, Eigen::Index cols, unsigned int computationOptions)
 {
   m_isTranspose = (cols > rows);
 
@@ -393,7 +393,7 @@ void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, co
 //@param shift : Each time one takes the left submatrix, one must add 1 to the shift. Why? Because! We actually want the last column of the U submatrix 
 // to become the first column (*coeff) and to shift all the other columns to the right. There are more details on the reference paper.
 template<typename MatrixType>
-void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift)
+void BDCSVD<MatrixType>::divide (Eigen::Index firstCol, Eigen::Index lastCol, Eigen::Index firstRowW, Eigen::Index firstColW, Eigen::Index shift)
 {
   // requires rows = cols + 1;
   using std::pow;
@@ -573,7 +573,7 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
 // handling of round-off errors, be consistent in ordering
 // For instance, to solve the secular equation using FMM, see http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf
 template <typename MatrixType>
-void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V)
+void BDCSVD<MatrixType>::computeSVDofM(Eigen::Index firstCol, Eigen::Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V)
 {
   const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
   using std::abs;
@@ -1059,7 +1059,7 @@ void BDCSVD<MatrixType>::computeSingVecs
 // i >= 1, di almost null and zi non null.
 // We use a rotation to zero out zi applied to the left of M
 template <typename MatrixType>
-void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index size)
+void BDCSVD<MatrixType>::deflation43(Eigen::Index firstCol, Eigen::Index shift, Eigen::Index i, Eigen::Index size)
 {
   using std::abs;
   using std::sqrt;
@@ -1088,7 +1088,7 @@ void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index
 // We apply two rotations to have zj = 0;
 // TODO deflation44 is still broken and not properly tested
 template <typename MatrixType>
-void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size)
+void BDCSVD<MatrixType>::deflation44(Eigen::Index firstColu , Eigen::Index firstColm, Eigen::Index firstRowW, Eigen::Index firstColW, Eigen::Index i, Eigen::Index j, Eigen::Index size)
 {
   using std::abs;
   using std::sqrt;
@@ -1128,7 +1128,7 @@ void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index fi
 
 // acts on block from (firstCol+shift, firstCol+shift) to (lastCol+shift, lastCol+shift) [inclusive]
 template <typename MatrixType>
-void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift)
+void BDCSVD<MatrixType>::deflation(Eigen::Index firstCol, Eigen::Index lastCol, Eigen::Index k, Eigen::Index firstRowW, Eigen::Index firstColW, Eigen::Index shift)
 {
   using std::sqrt;
   using std::abs;
diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h
index 43488b1e0..1c7c80376 100644
--- a/Eigen/src/SVD/JacobiSVD.h
+++ b/Eigen/src/SVD/JacobiSVD.h
@@ -610,7 +610,7 @@ template<typename _MatrixType, int QRPreconditioner> class JacobiSVD
 };
 
 template<typename MatrixType, int QRPreconditioner>
-void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Index rows, Index cols, unsigned int computationOptions)
+void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Eigen::Index rows, Eigen::Index cols, unsigned int computationOptions)
 {
   eigen_assert(rows >= 0 && cols >= 0);
 

From 5539587b1f5b5922b2419b0a4468cf2f393def51 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 10:29:12 +0200
Subject: [PATCH 248/680] Some warning fixes

---
 Eigen/src/Householder/HouseholderSequence.h     | 2 +-
 Eigen/src/SparseLU/SparseLU.h                   | 3 ---
 unsupported/Eigen/src/IterativeSolvers/DGMRES.h | 1 -
 3 files changed, 1 insertion(+), 5 deletions(-)

diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h
index af19701a4..e62befcb6 100644
--- a/Eigen/src/Householder/HouseholderSequence.h
+++ b/Eigen/src/Householder/HouseholderSequence.h
@@ -356,7 +356,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       if(m_length>=BlockSize && dst.cols()>1)
       {
         // Make sure we have at least 2 useful blocks, otherwise it is point-less:
-        Index blockSize = m_length<2*BlockSize ? (m_length+1)/2 : BlockSize;
+        Index blockSize = m_length<Index(2*BlockSize) ? (m_length+1)/2 : Index(BlockSize);
         for(Index i = 0; i < m_length; i+=blockSize)
         {
           Index end = m_reverse ? (std::min)(m_length,i+blockSize) : m_length-i;
diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h
index 5d2419353..91b6369ac 100644
--- a/Eigen/src/SparseLU/SparseLU.h
+++ b/Eigen/src/SparseLU/SparseLU.h
@@ -499,11 +499,8 @@ void SparseLU<MatrixType, OrderingType>::factorize(const MatrixType& matrix)
   eigen_assert(m_analysisIsOk && "analyzePattern() should be called first"); 
   eigen_assert((matrix.rows() == matrix.cols()) && "Only for squared matrices");
   
-  typedef typename IndexVector::Scalar StorageIndex; 
-  
   m_isInitialized = true;
   
-  
   // Apply the column permutation computed in analyzepattern()
   //   m_mat = matrix * m_perm_c.inverse(); 
   m_mat = matrix;
diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index be039e07f..85a4f696c 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -392,7 +392,6 @@ inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_Matr
 template< typename _MatrixType, typename _Preconditioner>
 inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_MatrixType, _Preconditioner>::schurValues(const RealSchur<DenseMatrix>& schurofH) const
 {
-  typedef typename MatrixType::Index Index;
   const DenseMatrix& T = schurofH.matrixT();
   Index it = T.rows();
   ComplexVector eig(it);

From 01fd4096d395e7b816459f571bf2328c8435cc37 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 10 Jul 2018 13:16:38 -0700
Subject: [PATCH 249/680] Fuse computations into the Tensor contractions using
 output kernel

---
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  10 +-
 .../CXX11/src/Tensor/TensorContraction.h      | 113 ++++++++++++++----
 .../src/Tensor/TensorContractionThreadPool.h  |  51 ++++++--
 .../src/Tensor/TensorForwardDeclarations.h    |   4 +-
 unsupported/test/cxx11_tensor_contraction.cpp |  51 ++++++++
 unsupported/test/cxx11_tensor_thread_pool.cpp |  56 +++++++++
 6 files changed, 248 insertions(+), 37 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index bdc1a17a7..97f90f638 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -517,9 +517,15 @@ class TensorBase<Derived, ReadOnlyAccessors>
     typedef Eigen::IndexPair<Index> DimensionPair;
 
     template<typename OtherDerived, typename Dimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const TensorContractionOp<const Dimensions, const Derived, const OtherDerived>
+    const TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const NoOpOutputKernel>
     contract(const OtherDerived& other, const Dimensions& dims) const {
-      return TensorContractionOp<const Dimensions, const Derived, const OtherDerived>(derived(), other.derived(), dims);
+      return TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const NoOpOutputKernel>(derived(), other.derived(), dims);
+    }
+
+    template<typename OtherDerived, typename Dimensions, typename OutputKernel> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const OutputKernel>
+    contract(const OtherDerived& other, const Dimensions& dims, const OutputKernel& output_kernel) const {
+      return TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const OutputKernel>(derived(), other.derived(), dims, output_kernel);
     }
 
     // Convolutions.
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 979fcf4d9..85126a127 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -85,8 +85,8 @@ template<typename LhsScalar, typename RhsScalar, typename Scalar>
 #endif
 
 
-template<typename Dimensions, typename LhsXprType, typename RhsXprType>
-struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> >
 {
   // Type promotion to handle the case where the types of the lhs and the rhs are different.
   typedef typename gebp_traits<typename remove_const<typename LhsXprType::Scalar>::type,
@@ -112,23 +112,24 @@ struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >
   };
 };
 
-template<typename Dimensions, typename LhsXprType, typename RhsXprType>
-struct eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType>, Eigen::Dense>
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>, Eigen::Dense>
 {
-  typedef const TensorContractionOp<Dimensions, LhsXprType, RhsXprType>& type;
+  typedef const TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>& type;
 };
 
-template<typename Dimensions, typename LhsXprType, typename RhsXprType>
-struct nested<TensorContractionOp<Dimensions, LhsXprType, RhsXprType>, 1, typename eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >::type>
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct nested<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>, 1, typename eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> >::type>
 {
-  typedef TensorContractionOp<Dimensions, LhsXprType, RhsXprType> type;
+  typedef TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> type;
 };
 
-template<typename Indices_, typename LeftArgType_, typename RightArgType_, typename Device_>
-struct traits<TensorEvaluator<const TensorContractionOp<Indices_, LeftArgType_, RightArgType_>, Device_> > {
+template<typename Indices_, typename LeftArgType_, typename RightArgType_, typename OutputKernelType_, typename Device_>
+struct traits<TensorEvaluator<const TensorContractionOp<Indices_, LeftArgType_, RightArgType_, OutputKernelType_>, Device_> > {
   typedef Indices_ Indices;
   typedef LeftArgType_ LeftArgType;
   typedef RightArgType_ RightArgType;
+  typedef OutputKernelType_ OutputKernelType;
   typedef Device_ Device;
 
   // From NumDims below.
@@ -137,8 +138,52 @@ struct traits<TensorEvaluator<const TensorContractionOp<Indices_, LeftArgType_,
 
 }  // end namespace internal
 
-template<typename Indices, typename LhsXprType, typename RhsXprType>
-class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXprType, RhsXprType>, ReadOnlyAccessors>
+// Tensor contraction params that should enable to get from output matrix
+// 2-dimensional coordinates to the output tensor dimensions.
+struct TensorContractionParams {
+  // TensorContraction evaluator assumes that both tensors are in ColMajor
+  // layout, if tensors are in RowMajor evaluator swap lhs with rhs.
+  bool swapped_arguments;
+};
+
+// Output kernel allows to fuse operations into the tensor contraction.
+//
+// Examples:
+//   1. Elementwise Relu transformation following Conv2D.
+//   2. AddBias to the Conv2D output channels dimension.
+//
+// See expected implementation in NoOpOutputKernel.
+struct OutputKernel {
+  template <typename Index, typename Scalar>
+  using OutputMapper = internal::blas_data_mapper<Scalar, Index, ColMajor>;
+};
+
+// Output kernel that does absolutely nothing.
+struct NoOpOutputKernel {
+  /**
+   * Tensor contraction evaluator calls this kernel after finishing each block
+   * of output matrix. Output blocks belong to the 2-dimensional output tensor.
+   *
+   * TensorContractionParams contains contraction dimensions information
+   * required to map output 2-d space into the expected output tensor space
+   * (potentially higher dimensional).
+   *
+   * \param[in] output_mapper Access to output tensor memory
+   * \param[in] params   Tensor contraction parameters
+   * \param[in] i        Index of a first row available through output_mapper
+   * \param[in] j        Index of a first column available through output_mapper
+   * \param[in] num_rows Number of available rows
+   * \param[in] num_cols Number of available columns
+   */
+  template <typename Index, typename Scalar>
+  EIGEN_ALWAYS_INLINE void operator()(
+      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
+      const TensorContractionParams& params, Index i, Index j, Index num_rows,
+      Index num_cols) const {}
+};
+
+template<typename Indices, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXprType, RhsXprType, OutputKernelType>, ReadOnlyAccessors>
 {
   public:
   typedef typename Eigen::internal::traits<TensorContractionOp>::Scalar Scalar;
@@ -149,8 +194,10 @@ class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXp
   typedef typename Eigen::internal::traits<TensorContractionOp>::Index Index;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionOp(
-      const LhsXprType& lhs, const RhsXprType& rhs, const Indices& dims)
-      : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_indices(dims) {}
+      const LhsXprType& lhs, const RhsXprType& rhs, const Indices& dims,
+      const OutputKernelType& output_kernel = OutputKernelType())
+      : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_indices(dims),
+        m_output_kernel(output_kernel) {}
 
   EIGEN_DEVICE_FUNC
   const Indices& indices() const { return m_indices; }
@@ -164,10 +211,14 @@ class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXp
   const typename internal::remove_all<typename RhsXprType::Nested>::type&
   rhsExpression() const { return m_rhs_xpr; }
 
+  EIGEN_DEVICE_FUNC
+  const OutputKernelType& outputKernel() const { return m_output_kernel; }
+
   protected:
     typename LhsXprType::Nested m_lhs_xpr;
     typename RhsXprType::Nested m_rhs_xpr;
     const Indices m_indices;
+    const OutputKernelType m_output_kernel;
 };
 
 
@@ -177,9 +228,10 @@ struct TensorContractionEvaluatorBase
   typedef typename internal::traits<Derived>::Indices Indices;
   typedef typename internal::traits<Derived>::LeftArgType LeftArgType;
   typedef typename internal::traits<Derived>::RightArgType RightArgType;
+  typedef typename internal::traits<Derived>::OutputKernelType OutputKernelType;
   typedef typename internal::traits<Derived>::Device Device;
 
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
@@ -221,6 +273,7 @@ struct TensorContractionEvaluatorBase
                           op.lhsExpression(), op.rhsExpression()), device),
     m_rightImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
                           op.rhsExpression(), op.lhsExpression()), device),
+        m_output_kernel(op.outputKernel()),
         m_device(device),
         m_result(NULL) {
     EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
@@ -391,6 +444,13 @@ struct TensorContractionEvaluatorBase
         numext::swap(m_dimensions[i], m_dimensions[j]);
       }
     }
+
+    // A set of parameters that will allow output kernel to get from output
+    // tensor dimensions (i, j) into the original tensor dimensions.
+    // TODO(ezhulenev): Add parameters required to infer output tensor index for
+    // more complex contractions than 2x2 on internal dimension.
+    m_tensor_contraction_params = {
+        /**swapped_arguments=*/static_cast<int>(Layout) == RowMajor};
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -585,7 +645,15 @@ struct TensorContractionEvaluatorBase
 
           // call gebp (matrix kernel)
           // The parameters here are copied from Eigen's GEMM implementation
-          gebp(output.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, Scalar(1), -1, -1, 0, 0);
+          const auto output_mapper = output.getSubMapper(i2, j2);
+          gebp(output_mapper, blockA, blockB, actual_mc, actual_kc, actual_nc,
+               Scalar(1), -1, -1, 0, 0);
+
+          // We are done with this [i2, j2] output block.
+          if (k2 + kc >= k) {
+            m_output_kernel(output_mapper, m_tensor_contraction_params, i2, j2,
+                            actual_mc, actual_nc);
+          }
         }
       }
     }
@@ -848,23 +916,26 @@ protected:
   Index m_j_size;
   Index m_k_size;
 
+  TensorContractionParams m_tensor_contraction_params;
+
   TensorEvaluator<EvalLeftArgType, Device> m_leftImpl;
   TensorEvaluator<EvalRightArgType, Device> m_rightImpl;
   const Device& m_device;
+  OutputKernelType m_output_kernel;
   Scalar* m_result;
   bool m_can_use_xsmm;
 };
 
 
 // evaluator for default device
-template<typename Indices, typename LeftArgType, typename RightArgType, typename Device>
-struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> :
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType, typename Device>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> :
     public TensorContractionEvaluatorBase<
-      TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> > {
-  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
+      TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> > {
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
   typedef TensorContractionEvaluatorBase<Self> Base;
 
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 3c007b183..d7536bd6a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -56,16 +56,16 @@ struct packRhsAndKernelArg {
 }  // end namespace internal
 #endif  // EIGEN_USE_SIMPLE_THREAD_POOL
 
-template<typename Indices, typename LeftArgType, typename RightArgType>
-struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, ThreadPoolDevice> :
-    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, ThreadPoolDevice> > {
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> > {
 
   typedef ThreadPoolDevice Device;
 
-  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
   typedef TensorContractionEvaluatorBase<Self> Base;
 
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
@@ -308,7 +308,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
                   this->m_k_strides);
 
     Context<LhsPacker, RhsPacker, GebpKernel, LhsMapper, RhsMapper,
-            OutputMapper>(this->m_device, num_threads, lhs, rhs, buffer, m, n,
+            OutputMapper>(this, num_threads, lhs, rhs, buffer, m, n,
                           k, bm, bn, bk, nm, nn, nk, gm, gn, nm0, nn0,
                           shard_by_col, parallel_pack)
         .run();
@@ -319,16 +319,18 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
             typename LhsMapper, typename RhsMapper, typename OutputMapper>
   class Context {
    public:
-    Context(const Device& device, int num_threads, LhsMapper& lhs,
+    Context(const Self* self, int num_threads, LhsMapper& lhs,
             RhsMapper& rhs, Scalar* buffer, Index tm, Index tn, Index tk, Index bm,
             Index bn, Index bk, Index nm, Index nn, Index nk, Index gm,
             Index gn, Index nm0, Index nn0, bool shard_by_col,
             bool parallel_pack)
-        : device_(device),
+        : device_(self->m_device),
           lhs_(lhs),
           rhs_(rhs),
           buffer_(buffer),
           output_(buffer, tm),
+          output_kernel_(self->m_output_kernel),
+          tensor_contraction_params_(self->m_tensor_contraction_params),
           num_threads_(num_threads),
           shard_by_col_(shard_by_col),
           parallel_pack_(parallel_pack),
@@ -420,6 +422,8 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     RhsMapper& rhs_;
     Scalar* const buffer_;
     OutputMapper output_;
+    OutputKernelType output_kernel_;
+    TensorContractionParams tensor_contraction_params_;
     const int num_threads_;
     const bool shard_by_col_;
     const bool parallel_pack_;
@@ -536,19 +540,32 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
       const Index mend = m * gm_ + gm(m);
       if (shard_by_col_) {
         for (Index n1 = n * gn_; n1 < nend; n1++) {
-          for (Index m1 = m * gm_; m1 < mend; m1++)
-            GebpKernel()(output_.getSubMapper(m1 * bm_, n1 * bn_),
-                         packed_lhs_[k % (P - 1)][m1],
+          for (Index m1 = m * gm_; m1 < mend; m1++) {
+            const auto output_mapper = output_.getSubMapper(m1 * bm_, n1 * bn_);
+            GebpKernel()(output_mapper, packed_lhs_[k % (P - 1)][m1],
                          packed_rhs_[k % (P - 1)][n1], bm(m1), bk(k), bn(n1),
                          Scalar(1), -1, -1, 0, 0);
+
+            // We are done with the last task for the [m1, n1] block.
+            if (k + 1 == nk_) {
+              output_kernel_(output_mapper, tensor_contraction_params_,
+                             m1 * bm_, n1 * bn_, bm(m1), bn(n1));
+            }
+          }
         }
       } else {
         for (Index m1 = m * gm_; m1 < mend; m1++)
           for (Index n1 = n * gn_; n1 < nend; n1++) {
-            GebpKernel()(output_.getSubMapper(m1 * bm_, n1 * bn_),
-                         packed_lhs_[k % (P - 1)][m1],
+            const auto output_mapper = output_.getSubMapper(m1 * bm_, n1 * bn_);
+            GebpKernel()(output_mapper, packed_lhs_[k % (P - 1)][m1],
                          packed_rhs_[k % (P - 1)][n1], bm(m1), bk(k), bn(n1),
                          Scalar(1), -1, -1, 0, 0);
+
+            // We are done with the last task for the [m1, n1] block.
+            if (k + 1 == nk_) {
+              output_kernel_(output_mapper, tensor_contraction_params_,
+                             m1 * bm_, n1 * bn_, bm(m1), bn(n1));
+            }
           }
       }
       signal_kernel(m, n, k + 1, false);
@@ -747,6 +764,10 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   }
 
 #else  // EIGEN_USE_SIMPLE_THREAD_POOL
+  // TODO(ezhulenev): SimpleThreadPool will be removed in the future, and seems
+  // like it's not worth adding output kernel support here.
+  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
+                "SimpleThreadPool does not support contraction output kernels.");
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
   void evalProduct(Scalar* buffer) const {
@@ -1065,6 +1086,10 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   }
 
 #if defined(EIGEN_VECTORIZE_AVX) && defined(EIGEN_USE_LIBXSMM)
+  // TODO(ezhulenev): Add support for output kernels and LIBXSMM.
+  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
+                "XSMM does not support contraction output kernels.");
+
   template<int Alignment>
   class ContextXsmm {
    public:
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 6c237bac3..19e456e19 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -65,7 +65,7 @@ template<typename Op, typename Dims, typename XprType, template <class> class Ma
 template<typename XprType> class TensorIndexTupleOp;
 template<typename ReduceOp, typename Dims, typename XprType> class TensorTupleReducerOp;
 template<typename Axis, typename LeftXprType, typename RightXprType> class TensorConcatenationOp;
-template<typename Dimensions, typename LeftXprType, typename RightXprType> class TensorContractionOp;
+template<typename Dimensions, typename LeftXprType, typename RightXprType, typename OutputKernelType> class TensorContractionOp;
 template<typename TargetType, typename XprType> class TensorConversionOp;
 template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
 template<typename FFT, typename XprType, int FFTDataType, int FFTDirection> class TensorFFTOp;
@@ -97,6 +97,8 @@ template<typename XprType> class TensorForcedEvalOp;
 template<typename ExpressionType, typename DeviceType> class TensorDevice;
 template<typename Derived, typename Device> struct TensorEvaluator;
 
+class NoOpOutputKernel;
+
 struct DefaultDevice;
 struct ThreadPoolDevice;
 struct GpuDevice;
diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp
index ace97057f..918c96277 100644
--- a/unsupported/test/cxx11_tensor_contraction.cpp
+++ b/unsupported/test/cxx11_tensor_contraction.cpp
@@ -510,6 +510,55 @@ static void test_const_inputs()
   VERIFY_IS_APPROX(mat3(1,1), mat1(1,0)*mat2(0,1) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(2,1));
 }
 
+// Apply Sqrt to all output elements.
+struct SqrtOutputKernel {
+  template <typename Index, typename Scalar>
+  EIGEN_ALWAYS_INLINE void operator()(
+      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
+      const TensorContractionParams&, Index, Index, Index num_rows,
+      Index num_cols) const {
+    for (int i = 0; i < num_rows; ++i) {
+      for (int j = 0; j < num_cols; ++j) {
+        output_mapper(i, j) = std::sqrt(output_mapper(i, j));
+      }
+    }
+  }
+};
+
+template <int DataLayout>
+static void test_large_contraction_with_output_kernel() {
+  Tensor<float, 4, DataLayout> t_left(30, 50, 8, 31);
+  Tensor<float, 5, DataLayout> t_right(8, 31, 7, 20, 10);
+  Tensor<float, 5, DataLayout> t_result(30, 50, 7, 20, 10);
+
+  t_left.setRandom();
+  t_right.setRandom();
+  // Put trash in mat4 to verify contraction clears output memory.
+  t_result.setRandom();
+
+  // Add a little offset so that the results won't be close to zero.
+  t_left += t_left.constant(1.0f);
+  t_right += t_right.constant(1.0f);
+
+  typedef Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>> MapXf;
+  MapXf m_left(t_left.data(), 1500, 248);
+  MapXf m_right(t_right.data(), 248, 1400);
+  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(1500, 1400);
+
+  // this contraction should be equivalent to a single matrix multiplication
+  Eigen::array<DimPair, 2> dims({{DimPair(2, 0), DimPair(3, 1)}});
+
+  // compute results by separate methods
+  t_result = t_left.contract(t_right, dims, SqrtOutputKernel());
+
+  m_result = m_left * m_right;
+
+  for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
+    VERIFY(&t_result.data()[i] != &m_result.data()[i]);
+    VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
+  }
+}
+
 void test_cxx11_tensor_contraction()
 {
   CALL_SUBTEST(test_evals<ColMajor>());
@@ -542,4 +591,6 @@ void test_cxx11_tensor_contraction()
   CALL_SUBTEST(test_tensor_product<RowMajor>());
   CALL_SUBTEST(test_const_inputs<ColMajor>());
   CALL_SUBTEST(test_const_inputs<RowMajor>());
+  CALL_SUBTEST(test_large_contraction_with_output_kernel<ColMajor>());
+  CALL_SUBTEST(test_large_contraction_with_output_kernel<RowMajor>());
 }
diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index 2ef665f30..ea9d8afdc 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -232,6 +232,60 @@ void test_multithread_contraction_agrees_with_singlethread() {
   }
 }
 
+// Apply Sqrt to all output elements.
+struct SqrtOutputKernel {
+  template <typename Index, typename Scalar>
+  EIGEN_ALWAYS_INLINE void operator()(
+      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
+      const TensorContractionParams&, Index, Index, Index num_rows,
+      Index num_cols) const {
+    for (int i = 0; i < num_rows; ++i) {
+      for (int j = 0; j < num_cols; ++j) {
+        output_mapper(i, j) = std::sqrt(output_mapper(i, j));
+      }
+    }
+  }
+};
+
+template <int DataLayout>
+static void test_multithread_contraction_with_output_kernel() {
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+
+  const int num_threads = internal::random<int>(2, 11);
+  ThreadPool threads(num_threads);
+  Eigen::ThreadPoolDevice device(&threads, num_threads);
+
+  Tensor<float, 4, DataLayout> t_left(30, 50, 8, 31);
+  Tensor<float, 5, DataLayout> t_right(8, 31, 7, 20, 10);
+  Tensor<float, 5, DataLayout> t_result(30, 50, 7, 20, 10);
+
+  t_left.setRandom();
+  t_right.setRandom();
+  // Put trash in mat4 to verify contraction clears output memory.
+  t_result.setRandom();
+
+  // Add a little offset so that the results won't be close to zero.
+  t_left += t_left.constant(1.0f);
+  t_right += t_right.constant(1.0f);
+
+  typedef Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>> MapXf;
+  MapXf m_left(t_left.data(), 1500, 248);
+  MapXf m_right(t_right.data(), 248, 1400);
+  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(1500, 1400);
+
+  // this contraction should be equivalent to a single matrix multiplication
+  Eigen::array<DimPair, 2> dims({{DimPair(2, 0), DimPair(3, 1)}});
+
+  // compute results by separate methods
+  t_result.device(device) = t_left.contract(t_right, dims, SqrtOutputKernel());
+
+  m_result = m_left * m_right;
+
+  for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
+    VERIFY(&t_result.data()[i] != &m_result.data()[i]);
+    VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
+  }
+}
 
 template<int DataLayout>
 void test_full_contraction() {
@@ -355,6 +409,8 @@ void test_cxx11_tensor_thread_pool()
 
   CALL_SUBTEST_3(test_multithread_contraction_agrees_with_singlethread<ColMajor>());
   CALL_SUBTEST_3(test_multithread_contraction_agrees_with_singlethread<RowMajor>());
+  CALL_SUBTEST_3(test_multithread_contraction_with_output_kernel<ColMajor>());
+  CALL_SUBTEST_3(test_multithread_contraction_with_output_kernel<RowMajor>());
 
   // Exercise various cases that have been problematic in the past.
   CALL_SUBTEST_4(test_contraction_corner_cases<ColMajor>());

From b324ed55d969b28ff84343b0840137a6b56300f1 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 12 Jul 2018 14:52:23 -0700
Subject: [PATCH 250/680] Call OutputKernel in evalGemv


From e204ecdaafa6c5642a4286a1ffb19e9964e32201 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Mon, 16 Jul 2018 15:06:57 -0700
Subject: [PATCH 251/680] Remove SimpleThreadPool and always use
 {NonBlocking}ThreadPool

---
 unsupported/Eigen/CXX11/ThreadPool            |  13 -
 .../src/Tensor/TensorContractionThreadPool.h  | 324 ------------------
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |   8 -
 .../src/ThreadPool/NonBlockingThreadPool.h    |  16 +-
 .../CXX11/src/ThreadPool/SimpleThreadPool.h   | 162 ---------
 5 files changed, 8 insertions(+), 515 deletions(-)
 delete mode 100644 unsupported/Eigen/CXX11/src/ThreadPool/SimpleThreadPool.h

diff --git a/unsupported/Eigen/CXX11/ThreadPool b/unsupported/Eigen/CXX11/ThreadPool
index c34614194..cbb3bbf2c 100644
--- a/unsupported/Eigen/CXX11/ThreadPool
+++ b/unsupported/Eigen/CXX11/ThreadPool
@@ -55,21 +55,8 @@
 #include "src/ThreadPool/RunQueue.h"
 #include "src/ThreadPool/ThreadPoolInterface.h"
 #include "src/ThreadPool/ThreadEnvironment.h"
-#include "src/ThreadPool/SimpleThreadPool.h"
 #include "src/ThreadPool/NonBlockingThreadPool.h"
 
-
-// Use the more efficient NonBlockingThreadPool by default.
-namespace Eigen {
-#ifndef EIGEN_USE_SIMPLE_THREAD_POOL
-template <typename Env> using ThreadPoolTempl = NonBlockingThreadPoolTempl<Env>;
-typedef NonBlockingThreadPool ThreadPool;
-#else
-template <typename Env> using ThreadPoolTempl = SimpleThreadPoolTempl<Env>;
-typedef SimpleThreadPool ThreadPool;
-#endif
-}  // namespace Eigen
-
 #endif
 
 #include <Eigen/src/Core/util/ReenableStupidWarnings.h>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index d7536bd6a..c774d15e2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -15,47 +15,6 @@
 
 namespace Eigen {
 
-#ifdef EIGEN_USE_SIMPLE_THREAD_POOL
-namespace internal {
-
-template<typename LhsScalar, typename LhsMapper, typename Index>
-struct packLhsArg {
-  LhsScalar* blockA;
-  const LhsMapper& lhs;
-  const Index m_start;
-  const Index k_start;
-  const Index mc;
-  const Index kc;
-};
-
-template<typename LhsScalar, typename RhsScalar, typename RhsMapper, typename OutputMapper, typename Index>
-struct packRhsAndKernelArg {
-  const MaxSizeVector<LhsScalar*>* blockAs;
-  RhsScalar* blockB;
-  const RhsMapper& rhs;
-  OutputMapper& output;
-  const Index m;
-  const Index k;
-  const Index n;
-  const Index mc;
-  const Index kc;
-  const Index nc;
-  const Index num_threads;
-  const Index num_blockAs;
-  const Index max_m;
-  const Index k_block_idx;
-  const Index m_block_idx;
-  const Index n_block_idx;
-  const Index m_blocks;
-  const Index n_blocks;
-  MaxSizeVector<Notification*>* kernel_notifications;
-  const MaxSizeVector<Notification*>* lhs_notifications;
-  const bool need_to_pack;
-};
-
-}  // end namespace internal
-#endif  // EIGEN_USE_SIMPLE_THREAD_POOL
-
 template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
 struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> :
     public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> > {
@@ -112,7 +71,6 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   TensorEvaluator(const XprType& op, const Device& device) :
       Base(op, device) {}
 
-#ifndef EIGEN_USE_SIMPLE_THREAD_POOL
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
             bool rhs_inner_dim_reordered, int Alignment>
   void evalProduct(Scalar* buffer) const {
@@ -763,288 +721,6 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     return 0;
   }
 
-#else  // EIGEN_USE_SIMPLE_THREAD_POOL
-  // TODO(ezhulenev): SimpleThreadPool will be removed in the future, and seems
-  // like it's not worth adding output kernel support here.
-  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
-                "SimpleThreadPool does not support contraction output kernels.");
-
-  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  void evalProduct(Scalar* buffer) const {
-    if (this->m_j_size == 1) {
-      this->template evalGemv<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
-      return;
-    }
-
-    evalGemm<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
-  }
-
-  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  void evalGemm(Scalar* buffer) const {
-    // columns in left side, rows in right side
-    const Index k = this->m_k_size;
-
-    // rows in left side
-    const Index m = this->m_i_size;
-
-    // columns in right side
-    const Index n = this->m_j_size;
-
-    // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar)
-    this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
-
-
-    const int lhs_packet_size = internal::unpacket_traits<typename LeftEvaluator::PacketReturnType>::size;
-    const int rhs_packet_size = internal::unpacket_traits<typename RightEvaluator::PacketReturnType>::size;
-
-    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
-                                                   LeftEvaluator, left_nocontract_t,
-                                                   contract_t, lhs_packet_size,
-                                                   lhs_inner_dim_contiguous,
-                                                   false, Unaligned> LhsMapper;
-
-    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
-                                                   RightEvaluator, right_nocontract_t,
-                                                   contract_t, rhs_packet_size,
-                                                   rhs_inner_dim_contiguous,
-                                                   rhs_inner_dim_reordered, Unaligned> RhsMapper;
-
-    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
-
-    // TODO: packing could be faster sometimes if we supported row major tensor mappers
-    typedef internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, Traits::mr,
-                                    Traits::LhsProgress, ColMajor> LhsPacker;
-    typedef internal::gemm_pack_rhs<RhsScalar, Index, typename RhsMapper::SubMapper, Traits::nr, ColMajor> RhsPacker;
-
-    // TODO: replace false, false with conjugate values?
-    typedef internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper,
-                                  Traits::mr, Traits::nr, false, false> GebpKernel;
-
-    typedef internal::packLhsArg<LhsScalar, LhsMapper, Index> packLArg;
-    typedef internal::packRhsAndKernelArg<LhsScalar, RhsScalar, RhsMapper, OutputMapper, Index> packRKArg;
-
-    // initialize data mappers
-    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
-                  this->m_left_contracting_strides, this->m_k_strides);
-
-    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
-                  this->m_right_contracting_strides, this->m_k_strides);
-
-    OutputMapper output(buffer, m);
-
-    // compute block sizes (which depend on number of threads)
-    const Index num_threads = this->m_device.numThreads();
-    internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index, internal::ShardByCol> blocking(k, m, n, num_threads);
-    Index mc = blocking.mc();
-    Index nc = blocking.nc();
-    Index kc = blocking.kc();
-    eigen_assert(mc <= m);
-    eigen_assert(nc <= n);
-    eigen_assert(kc <= k);
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
-    const Index k_blocks = CEIL_DIV(k, kc);
-    const Index n_blocks = CEIL_DIV(n, nc);
-    const Index m_blocks = CEIL_DIV(m, mc);
-    const Index sizeA = mc * kc;
-    const Index sizeB = kc * nc;
-
-    /*    cout << "m: " << m << " n: " << n << " k: " << k << endl;
-    cout << "mc: " << mc << " nc: " << nc << " kc: " << kc << endl;
-    cout << "m_blocks: " << m_blocks << " n_blocks: " << n_blocks << " k_blocks: " << k_blocks << endl;
-    cout << "num threads: " << num_threads << endl;
-    */
-
-    // note: m_device.allocate should return 16 byte aligned pointers, but if blockA and blockB
-    //       aren't 16 byte aligned segfaults will happen due to SIMD instructions
-    // note: You can get away with allocating just a single blockA and offsets and meet the
-    //       the alignment requirements with the assumption that
-    //       (Traits::mr * sizeof(ResScalar)) % 16 == 0
-    const Index numBlockAs = numext::mini(num_threads, m_blocks);
-    MaxSizeVector<LhsScalar *> blockAs(num_threads);
-    for (int i = 0; i < num_threads; i++) {
-      blockAs.push_back(static_cast<LhsScalar *>(this->m_device.allocate(sizeA * sizeof(LhsScalar))));
-    }
-
-    // To circumvent alignment issues, I'm just going to separately allocate the memory for each thread
-    // TODO: is this too much memory to allocate? This simplifies coding a lot, but is wasteful.
-    //       Other options: (1) reuse memory when a thread finishes. con: tricky
-    //                      (2) allocate block B memory in each thread. con: overhead
-    MaxSizeVector<RhsScalar *> blockBs(n_blocks);
-    for (int i = 0; i < n_blocks; i++) {
-      blockBs.push_back(static_cast<RhsScalar *>(this->m_device.allocate(sizeB * sizeof(RhsScalar))));
-    }
-
-    // lhs_notifications starts with all null Notifications
-    MaxSizeVector<Notification*> lhs_notifications(num_threads, nullptr);
-
-    // this should really be numBlockAs * n_blocks;
-    const Index num_kernel_notifications = num_threads * n_blocks;
-    MaxSizeVector<Notification*> kernel_notifications(num_kernel_notifications,
-                                                    nullptr);
-
-    for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) {
-      const Index k_start = k_block_idx * kc;
-      // make sure we don't overshoot right edge of left matrix
-      const Index actual_kc = numext::mini(k_start + kc, k) - k_start;
-
-      for (Index m_block_idx = 0; m_block_idx < m_blocks; m_block_idx += numBlockAs) {
-        const Index num_blocks = numext::mini(m_blocks-m_block_idx, numBlockAs);
-
-        for (Index mt_block_idx = m_block_idx; mt_block_idx < m_block_idx+num_blocks; mt_block_idx++) {
-          const Index m_start = mt_block_idx * mc;
-          const Index actual_mc = numext::mini(m_start + mc, m) - m_start;
-          eigen_assert(actual_mc > 0);
-
-          Index blockAId = (k_block_idx * m_blocks + mt_block_idx) % num_threads;
-
-          for (int i = 0; i < n_blocks; ++i) {
-            Index notification_id = (blockAId * n_blocks + i);
-            // Wait for any current kernels using this slot to complete
-            // before using it.
-            if (kernel_notifications[notification_id]) {
-              wait_until_ready(kernel_notifications[notification_id]);
-              delete kernel_notifications[notification_id];
-            }
-            kernel_notifications[notification_id] = new Notification();
-          }
-          const packLArg arg = {
-            blockAs[blockAId], // blockA
-            lhs,        // lhs
-            m_start,    // m
-            k_start,    // k
-            actual_mc,  // mc
-            actual_kc,  // kc
-          };
-
-          // Delete any existing notification since we may be
-          // replacing it.  The algorithm should ensure that there are
-          // no existing waiters on this notification.
-          delete lhs_notifications[blockAId];
-          lhs_notifications[blockAId] =
-          this->m_device.enqueue(&Self::packLhs<packLArg, LhsPacker>, arg);
-        }
-
-        // now start kernels.
-        const Index m_base_start = m_block_idx * mc;
-        const bool need_to_pack = m_block_idx == 0;
-
-        for (Index n_block_idx = 0; n_block_idx < n_blocks; n_block_idx++) {
-          const Index n_start = n_block_idx * nc;
-          const Index actual_nc = numext::mini(n_start + nc, n) - n_start;
-
-          // first make sure the previous kernels are all done before overwriting rhs. Also wait if
-          // we're going to start new k. In both cases need_to_pack is true.
-          if (need_to_pack) {
-            for (Index i = num_blocks; i < num_threads; ++i) {
-              Index blockAId = (k_block_idx * m_blocks + i + m_block_idx) % num_threads;
-              Index future_id = (blockAId * n_blocks + n_block_idx);
-              wait_until_ready(kernel_notifications[future_id]);
-            }
-          }
-
-          packRKArg arg = {
-            &blockAs, // blockA
-            blockBs[n_block_idx], // blockB
-            rhs,          // rhs
-            output,       // output
-            m_base_start, // m
-            k_start,      // k
-            n_start,      // n
-            mc,           // mc
-            actual_kc,    // kc
-            actual_nc,    // nc
-            num_threads,
-            numBlockAs,
-            m,
-            k_block_idx,
-            m_block_idx,
-            n_block_idx, // n_block_idx
-            m_blocks, // m_blocks
-            n_blocks, // n_blocks
-            &kernel_notifications, // kernel notifications
-            &lhs_notifications,    // lhs notifications
-            need_to_pack, // need_to_pack
-          };
-
-          // We asynchronously kick off this function, which ends up
-          // notifying the appropriate kernel_notifications objects,
-          // which this thread waits on before exiting.
-          this->m_device.enqueueNoNotification(&Self::packRhsAndKernel<packRKArg, RhsPacker, GebpKernel>, arg);
-        }
-      }
-    }
-
-    // Make sure all the kernels are done.
-    for (size_t i = 0; i < kernel_notifications.size(); ++i) {
-      wait_until_ready(kernel_notifications[i]);
-      delete kernel_notifications[i];
-    }
-
-    // No need to wait for lhs notifications since they should have
-    // already been waited on.  Just clean them up.
-    for (size_t i = 0; i < lhs_notifications.size(); ++i) {
-      delete lhs_notifications[i];
-    }
-
-    // deallocate all of the memory for both A and B's
-    for (size_t i = 0; i < blockAs.size(); i++) {
-      this->m_device.deallocate(blockAs[i]);
-    }
-    for (size_t i = 0; i < blockBs.size(); i++) {
-      this->m_device.deallocate(blockBs[i]);
-    }
-
-#undef CEIL_DIV
-  }
-
-  /*
-   * Packs a LHS block of size (mt, kc) starting at lhs(m, k). Before packing
-   * the LHS block, check that all of the kernels that worked on the same
-   * mt_block_idx in the previous m_block are done.
-   */
-  template <typename packLArg, typename LhsPacker>
-  static void packLhs(const packLArg arg) {
-    // perform actual packing
-    LhsPacker pack_lhs;
-    pack_lhs(arg.blockA, arg.lhs.getSubMapper(arg.m_start, arg.k_start), arg.kc, arg.mc);
-  }
-
-  /*
-   * Packs a RHS block of size (kc, nc) starting at (k, n) after checking that
-   * all kernels in the previous block are done.
-   * Then for each LHS future, we wait on the future and then call GEBP
-   * on the area packed by the future (which starts at
-   * blockA + future_idx * mt * kc) on the LHS and with the full packed
-   * RHS block.
-   * The output of this GEBP is written to output(m + i * mt, n).
-   */
-  template <typename packRKArg, typename RhsPacker, typename GebpKernel>
-  static void packRhsAndKernel(packRKArg arg) {
-    if (arg.need_to_pack) {
-      RhsPacker pack_rhs;
-      pack_rhs(arg.blockB, arg.rhs.getSubMapper(arg.k, arg.n), arg.kc, arg.nc);
-    }
-
-    GebpKernel gebp;
-    for (Index mt_block_idx = 0; mt_block_idx < arg.num_blockAs; mt_block_idx++) {
-      const Index m_base_start = arg.m + arg.mc*mt_block_idx;
-      if (m_base_start < arg.max_m) {
-        Index blockAId = (arg.k_block_idx * arg.m_blocks + mt_block_idx + arg.m_block_idx) % arg.num_threads;
-        wait_until_ready((*arg.lhs_notifications)[blockAId]);
-        const Index actual_mc = numext::mini(m_base_start + arg.mc, arg.max_m) - m_base_start;
-        gebp(arg.output.getSubMapper(m_base_start, arg.n),
-             (*arg.blockAs)[blockAId], arg.blockB,
-             actual_mc, arg.kc, arg.nc, Scalar(1), -1, -1, 0, 0);
-
-        // Notify that the kernel is done.
-        const Index set_idx = blockAId * arg.n_blocks + arg.n_block_idx;
-        (*arg.kernel_notifications)[set_idx]->Notify();
-      }
-    }
-  }
-#endif  // EIGEN_USE_SIMPLE_THREAD_POOL
-
   TensorOpCost contractionCost(Index m, Index n, Index bm, Index bn, Index bk,
                                bool shard_by_col, bool prepacked) const {
     const int packed_size = std::min<int>(PacketType<LhsScalar, Device>::size,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 1181c2753..53640c6aa 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -150,13 +150,6 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> {
     if (needs_assign)
     {
       const Index size = array_prod(evaluator.dimensions());
-#if !defined(EIGEN_USE_SIMPLE_THREAD_POOL)
-      device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
-                         EvalRange<Evaluator, Index, Vectorizable>::alignBlockSize,
-                         [&evaluator](Index first, Index last) {
-                           EvalRange<Evaluator, Index, Vectorizable>::run(&evaluator, first, last);
-                         });
-#else
       size_t num_threads = device.numThreads();
       if (num_threads > 1) {
         num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
@@ -182,7 +175,6 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> {
         }
         barrier.Wait();
       }
-#endif  // defined(!EIGEN_USE_SIMPLE_THREAD_POOL)
     }
     evaluator.cleanup();
   }
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 1264a0270..ecd49f382 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -14,15 +14,15 @@
 namespace Eigen {
 
 template <typename Environment>
-class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
+class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
  public:
   typedef typename Environment::Task Task;
   typedef RunQueue<Task, 1024> Queue;
 
-  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
-      : NonBlockingThreadPoolTempl(num_threads, true, env) {}
+  ThreadPoolTempl(int num_threads, Environment env = Environment())
+      : ThreadPoolTempl(num_threads, true, env) {}
 
-  NonBlockingThreadPoolTempl(int num_threads, bool allow_spinning,
+  ThreadPoolTempl(int num_threads, bool allow_spinning,
                              Environment env = Environment())
       : env_(env),
         num_threads_(num_threads),
@@ -66,7 +66,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
     }
   }
 
-  ~NonBlockingThreadPoolTempl() {
+  ~ThreadPoolTempl() {
     done_ = true;
 
     // Now if all threads block without work, they will start exiting.
@@ -136,7 +136,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
 
   int CurrentThreadId() const final {
     const PerThread* pt =
-        const_cast<NonBlockingThreadPoolTempl*>(this)->GetPerThread();
+        const_cast<ThreadPoolTempl*>(this)->GetPerThread();
     if (pt->pool == this) {
       return pt->thread_id;
     } else {
@@ -149,7 +149,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
 
   struct PerThread {
     constexpr PerThread() : pool(NULL), rand(0), thread_id(-1) { }
-    NonBlockingThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
+    ThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
     uint64_t rand;  // Random generator state.
     int thread_id;  // Worker thread index in pool.
   };
@@ -337,7 +337,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
   }
 };
 
-typedef NonBlockingThreadPoolTempl<StlThreadEnvironment> NonBlockingThreadPool;
+typedef ThreadPoolTempl<StlThreadEnvironment> ThreadPool;
 
 }  // namespace Eigen
 
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/SimpleThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/SimpleThreadPool.h
deleted file mode 100644
index 335728665..000000000
--- a/unsupported/Eigen/CXX11/src/ThreadPool/SimpleThreadPool.h
+++ /dev/null
@@ -1,162 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_CXX11_THREADPOOL_SIMPLE_THREAD_POOL_H
-#define EIGEN_CXX11_THREADPOOL_SIMPLE_THREAD_POOL_H
-
-namespace Eigen {
-
-// The implementation of the ThreadPool type ensures that the Schedule method
-// runs the functions it is provided in FIFO order when the scheduling is done
-// by a single thread.
-// Environment provides a way to create threads and also allows to intercept
-// task submission and execution.
-template <typename Environment>
-class SimpleThreadPoolTempl : public ThreadPoolInterface {
- public:
-  // Construct a pool that contains "num_threads" threads.
-  explicit SimpleThreadPoolTempl(int num_threads, Environment env = Environment())
-      : env_(env), threads_(num_threads), waiters_(num_threads) {
-    for (int i = 0; i < num_threads; i++) {
-      threads_.push_back(env.CreateThread([this, i]() { WorkerLoop(i); }));
-    }
-  }
-
-  // Wait until all scheduled work has finished and then destroy the
-  // set of threads.
-  ~SimpleThreadPoolTempl() {
-    {
-      // Wait for all work to get done.
-      std::unique_lock<std::mutex> l(mu_);
-      while (!pending_.empty()) {
-        empty_.wait(l);
-      }
-      exiting_ = true;
-
-      // Wakeup all waiters.
-      for (auto w : waiters_) {
-        w->ready = true;
-        w->task.f = nullptr;
-        w->cv.notify_one();
-      }
-    }
-
-    // Wait for threads to finish.
-    for (auto t : threads_) {
-      delete t;
-    }
-  }
-
-  // Schedule fn() for execution in the pool of threads. The functions are
-  // executed in the order in which they are scheduled.
-  void Schedule(std::function<void()> fn) final {
-    Task t = env_.CreateTask(std::move(fn));
-    std::unique_lock<std::mutex> l(mu_);
-    if (waiters_.empty()) {
-      pending_.push_back(std::move(t));
-    } else {
-      Waiter* w = waiters_.back();
-      waiters_.pop_back();
-      w->ready = true;
-      w->task = std::move(t);
-      w->cv.notify_one();
-    }
-  }
-
-  void Cancel() {
-#ifdef EIGEN_THREAD_ENV_SUPPORTS_CANCELLATION
-    for (size_t i = 0; i < threads_.size(); i++) {
-      threads_[i]->OnCancel();
-    }
-#endif
-  }
-
-  int NumThreads() const final {
-    return static_cast<int>(threads_.size());
-  }
-
-  int CurrentThreadId() const final {
-    const PerThread* pt = this->GetPerThread();
-    if (pt->pool == this) {
-      return pt->thread_id;
-    } else {
-      return -1;
-    }
-  }
-
- protected:
-  void WorkerLoop(int thread_id) {
-    std::unique_lock<std::mutex> l(mu_);
-    PerThread* pt = GetPerThread();
-    pt->pool = this;
-    pt->thread_id = thread_id;
-    Waiter w;
-    Task t;
-    while (!exiting_) {
-      if (pending_.empty()) {
-        // Wait for work to be assigned to me
-        w.ready = false;
-        waiters_.push_back(&w);
-        while (!w.ready) {
-          w.cv.wait(l);
-        }
-        t = w.task;
-        w.task.f = nullptr;
-      } else {
-        // Pick up pending work
-        t = std::move(pending_.front());
-        pending_.pop_front();
-        if (pending_.empty()) {
-          empty_.notify_all();
-        }
-      }
-      if (t.f) {
-        mu_.unlock();
-        env_.ExecuteTask(t);
-        t.f = nullptr;
-        mu_.lock();
-      }
-    }
-  }
-
- private:
-  typedef typename Environment::Task Task;
-  typedef typename Environment::EnvThread Thread;
-
-  struct Waiter {
-    std::condition_variable cv;
-    Task task;
-    bool ready;
-  };
-
-  struct PerThread {
-    constexpr PerThread() : pool(NULL), thread_id(-1) { }
-    SimpleThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
-    int thread_id;                // Worker thread index in pool.
-  };
-
-  Environment env_;
-  std::mutex mu_;
-  MaxSizeVector<Thread*> threads_;  // All threads
-  MaxSizeVector<Waiter*> waiters_;  // Stack of waiting threads.
-  std::deque<Task> pending_;        // Queue of pending work
-  std::condition_variable empty_;   // Signaled on pending_.empty()
-  bool exiting_ = false;
-
-  PerThread* GetPerThread() const {
-    EIGEN_THREAD_LOCAL PerThread per_thread;
-    return &per_thread;
-  }
-};
-
-typedef SimpleThreadPoolTempl<StlThreadEnvironment> SimpleThreadPool;
-
-}  // namespace Eigen
-
-#endif  // EIGEN_CXX11_THREADPOOL_SIMPLE_THREAD_POOL_H

From 43206ac4de09e348aaf0c8e7134f59872eaf55f4 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 12 Jul 2018 14:52:23 -0700
Subject: [PATCH 252/680] Call OutputKernel in evalGemv

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 85126a127..0fbffa34c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -549,6 +549,11 @@ struct TensorContractionEvaluatorBase
     internal::general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,false,RhsScalar,RhsMapper,false>::run(
         rows, cols, lhs, rhs,
         buffer, resIncr, alpha);
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+    m_output_kernel(OutputMapper(buffer, rows), m_tensor_contraction_params,
+                    static_cast<Index>(0), static_cast<Index>(0), rows,
+                    static_cast<Index>(1));
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>

From 2b2cd85694c7d340eb8b3d3ba944d33360666381 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 11:11:33 +0200
Subject: [PATCH 253/680] bug #1573: add noexcept move constructor and move
 assignment operator to Quaternion

---
 Eigen/src/Geometry/Quaternion.h | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h
index c3fd8c3e0..6fad6564a 100644
--- a/Eigen/src/Geometry/Quaternion.h
+++ b/Eigen/src/Geometry/Quaternion.h
@@ -105,6 +105,22 @@ class QuaternionBase : public RotationBase<Derived, 3>
   EIGEN_DEVICE_FUNC Derived& operator=(const AngleAxisType& aa);
   template<class OtherDerived> EIGEN_DEVICE_FUNC Derived& operator=(const MatrixBase<OtherDerived>& m);
 
+#if EIGEN_HAS_RVALUE_REFERENCES
+  // Because we have a user-defined copy assignment operator, we don't get an implicit move constructor or assignment operator.
+  /** Default move constructor */
+  EIGEN_DEVICE_FUNC inline QuaternionBase(QuaternionBase&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value) = default;
+
+  /** Default move assignment operator */
+  EIGEN_DEVICE_FUNC QuaternionBase& operator=(QuaternionBase&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) = default;
+
+  // And now because we declared a constructor, we don't get a default constructor or copy constructor. Say we want them.
+  /** Default constructor */
+  EIGEN_DEVICE_FUNC QuaternionBase() = default;
+
+  /** Copy constructor */
+  EIGEN_DEVICE_FUNC QuaternionBase(const QuaternionBase& other) = default;
+#endif
+
   /** \returns a quaternion representing an identity rotation
     * \sa MatrixBase::Identity()
     */
@@ -276,6 +292,19 @@ public:
   EIGEN_DEVICE_FUNC explicit inline Quaternion(const Quaternion<OtherScalar, OtherOptions>& other)
   { m_coeffs = other.coeffs().template cast<Scalar>(); }
 
+#if EIGEN_HAS_RVALUE_REFERENCES
+  // We define a copy constructor, which means we don't get an implicit move constructor or assignment operator.
+  /** Default move constructor */
+  EIGEN_DEVICE_FUNC inline Quaternion(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value) = default;
+
+  /** Default move assignment operator */
+  EIGEN_DEVICE_FUNC Quaternion& operator=(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) = default;
+
+  // And now because we declared a constructor, we don't get an implicit copy constructor. Say we want one.
+  /** Default copy constructor */
+  EIGEN_DEVICE_FUNC Quaternion(const Quaternion& other) = default;
+#endif
+
   EIGEN_DEVICE_FUNC static Quaternion UnitRandom();
 
   template<typename Derived1, typename Derived2>

From 37f4bdd97df454f2c85c6047d190f5db21c3a09b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 13:20:49 +0200
Subject: [PATCH 254/680] Fix VERIFY_EVALUATION_COUNT(EXPR,N) with a complex
 expression as N

---
 test/main.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/test/main.h b/test/main.h
index 95bbc9eb0..51399f7ad 100644
--- a/test/main.h
+++ b/test/main.h
@@ -122,8 +122,8 @@ inline void on_temporary_creation(long int size) {
 #define VERIFY_EVALUATION_COUNT(XPR,N) {\
     nb_temporaries = 0; \
     XPR; \
-    if(nb_temporaries!=N) { std::cerr << "nb_temporaries == " << nb_temporaries << "\n"; }\
-    VERIFY( (#XPR) && nb_temporaries==N ); \
+    if(nb_temporaries!=(N)) { std::cerr << "nb_temporaries == " << nb_temporaries << "\n"; }\
+    VERIFY( (#XPR) && nb_temporaries==(N) ); \
   }
   
 #endif

From 82f0ce27261df3b21037d93d4595655b3df754a6 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 14:46:15 +0200
Subject: [PATCH 255/680] Get rid of EIGEN_TEST_FUNC, unit tests must now be
 declared with EIGEN_DECLARE_TEST(mytest) { /* code */ }. This provide several
 advantages: - more flexibility in designing unit tests - unit tests can be
 glued to speed up compilation - unit tests are compiled with same predefined
 macros, which is a requirement for zapcc

---
 cmake/EigenTesting.cmake                      |  4 --
 test/adjoint.cpp                              |  2 +-
 test/array.cpp                                |  2 +-
 test/array_for_matrix.cpp                     |  2 +-
 test/array_of_string.cpp                      |  2 +-
 test/array_replicate.cpp                      |  2 +-
 test/array_reverse.cpp                        |  2 +-
 test/bandmatrix.cpp                           |  2 +-
 test/basicstuff.cpp                           |  2 +-
 test/bdcsvd.cpp                               |  2 +-
 test/bicgstab.cpp                             |  2 +-
 test/block.cpp                                |  2 +-
 test/boostmultiprec.cpp                       |  2 +-
 test/cholesky.cpp                             |  2 +-
 test/cholmod_support.cpp                      |  2 +-
 test/commainitializer.cpp                     |  2 +-
 test/conjugate_gradient.cpp                   |  2 +-
 test/conservative_resize.cpp                  |  2 +-
 test/constructor.cpp                          |  2 +-
 test/corners.cpp                              |  2 +-
 test/ctorleak.cpp                             |  2 +-
 test/denseLM.cpp                              |  2 +-
 test/dense_storage.cpp                        |  2 +-
 test/determinant.cpp                          |  2 +-
 test/diagonal.cpp                             |  2 +-
 test/diagonalmatrices.cpp                     |  2 +-
 test/dontalign.cpp                            |  2 +-
 test/dynalloc.cpp                             |  2 +-
 test/eigen2support.cpp                        |  2 +-
 test/eigensolver_complex.cpp                  |  2 +-
 test/eigensolver_generalized_real.cpp         |  2 +-
 test/eigensolver_generic.cpp                  |  2 +-
 test/eigensolver_selfadjoint.cpp              |  2 +-
 test/evaluators.cpp                           |  2 +-
 test/exceptions.cpp                           |  2 +-
 test/fastmath.cpp                             |  2 +-
 test/first_aligned.cpp                        |  2 +-
 test/geo_alignedbox.cpp                       |  2 +-
 test/geo_eulerangles.cpp                      |  2 +-
 test/geo_homogeneous.cpp                      |  2 +-
 test/geo_hyperplane.cpp                       |  2 +-
 test/geo_orthomethods.cpp                     |  2 +-
 test/geo_parametrizedline.cpp                 |  2 +-
 test/geo_quaternion.cpp                       |  2 +-
 test/geo_transformations.cpp                  |  2 +-
 test/gpu_basic.cu                             |  3 +-
 test/half_float.cpp                           |  2 +-
 test/hessenberg.cpp                           |  2 +-
 test/householder.cpp                          |  2 +-
 test/incomplete_cholesky.cpp                  |  2 +-
 test/indexed_view.cpp                         |  2 +-
 test/inplace_decomposition.cpp                |  2 +-
 test/integer_types.cpp                        |  2 +-
 test/inverse.cpp                              |  2 +-
 test/is_same_dense.cpp                        |  2 +-
 test/jacobi.cpp                               |  2 +-
 test/jacobisvd.cpp                            |  2 +-
 test/klu_support.cpp                          |  2 +-
 test/linearstructure.cpp                      |  2 +-
 test/lscg.cpp                                 |  2 +-
 test/lu.cpp                                   |  2 +-
 test/main.h                                   | 69 +++++++++++++------
 test/mapped_matrix.cpp                        |  2 +-
 test/mapstaticmethods.cpp                     |  2 +-
 test/mapstride.cpp                            |  2 +-
 test/meta.cpp                                 |  2 +-
 test/metis_support.cpp                        |  2 +-
 test/miscmatrices.cpp                         |  2 +-
 test/mixingtypes.cpp                          |  2 +-
 test/nesting_ops.cpp                          |  2 +-
 test/nomalloc.cpp                             |  2 +-
 test/nullary.cpp                              |  2 +-
 test/num_dimensions.cpp                       |  2 +-
 test/numext.cpp                               |  2 +-
 test/packetmath.cpp                           |  2 +-
 test/pardiso_support.cpp                      |  2 +-
 test/pastix_support.cpp                       |  2 +-
 test/permutationmatrices.cpp                  |  2 +-
 test/prec_inverse_4x4.cpp                     |  2 +-
 test/product_extra.cpp                        |  2 +-
 test/product_large.cpp                        |  2 +-
 test/product_mmtr.cpp                         |  2 +-
 test/product_notemporary.cpp                  |  2 +-
 test/product_selfadjoint.cpp                  |  2 +-
 test/product_small.cpp                        |  2 +-
 test/product_symm.cpp                         |  2 +-
 test/product_syrk.cpp                         |  2 +-
 test/product_trmm.cpp                         |  2 +-
 test/product_trmv.cpp                         |  2 +-
 test/product_trsolve.cpp                      |  2 +-
 test/qr.cpp                                   |  2 +-
 test/qr_colpivoting.cpp                       |  2 +-
 test/qr_fullpivoting.cpp                      |  2 +-
 test/qtvector.cpp                             |  2 +-
 test/rand.cpp                                 |  2 +-
 test/real_qz.cpp                              |  2 +-
 test/redux.cpp                                |  2 +-
 test/ref.cpp                                  |  2 +-
 test/resize.cpp                               |  2 +-
 test/rvalue_types.cpp                         |  2 +-
 test/schur_complex.cpp                        |  2 +-
 test/schur_real.cpp                           |  2 +-
 test/selfadjoint.cpp                          |  2 +-
 test/simplicial_cholesky.cpp                  |  2 +-
 test/sizeof.cpp                               |  2 +-
 test/sizeoverflow.cpp                         |  2 +-
 test/smallvectors.cpp                         |  2 +-
 test/sparseLM.cpp                             |  2 +-
 test/sparse_basic.cpp                         |  2 +-
 test/sparse_block.cpp                         |  2 +-
 test/sparse_permutations.cpp                  |  2 +-
 test/sparse_product.cpp                       |  2 +-
 test/sparse_ref.cpp                           |  2 +-
 test/sparse_solvers.cpp                       |  2 +-
 test/sparse_vector.cpp                        |  2 +-
 test/sparselu.cpp                             |  2 +-
 test/sparseqr.cpp                             |  2 +-
 test/special_numbers.cpp                      |  2 +-
 test/spqr_support.cpp                         |  2 +-
 test/stable_norm.cpp                          |  2 +-
 test/stddeque.cpp                             |  2 +-
 test/stddeque_overload.cpp                    |  2 +-
 test/stdlist.cpp                              |  2 +-
 test/stdlist_overload.cpp                     |  2 +-
 test/stdvector.cpp                            |  2 +-
 test/stdvector_overload.cpp                   |  2 +-
 test/superlu_support.cpp                      |  2 +-
 test/swap.cpp                                 |  2 +-
 test/symbolic_index.cpp                       |  2 +-
 test/triangular.cpp                           |  2 +-
 test/umeyama.cpp                              |  2 +-
 test/umfpack_support.cpp                      |  2 +-
 test/unalignedassert.cpp                      |  2 +-
 test/unalignedcount.cpp                       |  2 +-
 test/upperbidiagonalization.cpp               |  2 +-
 test/vectorization_logic.cpp                  |  2 +-
 test/vectorwiseop.cpp                         |  2 +-
 test/visitor.cpp                              |  2 +-
 test/zerosized.cpp                            |  2 +-
 .../LevenbergMarquardt/LevenbergMarquardt.h   |  2 +-
 unsupported/test/BVH.cpp                      |  2 +-
 unsupported/test/EulerAngles.cpp              |  2 +-
 unsupported/test/FFTW.cpp                     |  2 +-
 unsupported/test/NonLinearOptimization.cpp    |  2 +-
 unsupported/test/NumericalDiff.cpp            |  2 +-
 unsupported/test/alignedvector3.cpp           |  2 +-
 unsupported/test/autodiff.cpp                 |  2 +-
 unsupported/test/autodiff_scalar.cpp          |  2 +-
 unsupported/test/cxx11_eventcount.cpp         |  2 +-
 unsupported/test/cxx11_meta.cpp               |  2 +-
 .../test/cxx11_non_blocking_thread_pool.cpp   |  2 +-
 unsupported/test/cxx11_runqueue.cpp           |  2 +-
 unsupported/test/cxx11_tensor_argmax.cpp      |  2 +-
 unsupported/test/cxx11_tensor_argmax_gpu.cu   |  2 +-
 unsupported/test/cxx11_tensor_argmax_sycl.cpp |  4 +-
 unsupported/test/cxx11_tensor_assign.cpp      |  2 +-
 .../test/cxx11_tensor_broadcast_sycl.cpp      |  4 +-
 .../test/cxx11_tensor_broadcasting.cpp        |  2 +-
 .../test/cxx11_tensor_builtins_sycl.cpp       |  4 +-
 .../test/cxx11_tensor_cast_float16_gpu.cu     |  2 +-
 unsupported/test/cxx11_tensor_casts.cpp       |  2 +-
 unsupported/test/cxx11_tensor_chipping.cpp    |  2 +-
 .../test/cxx11_tensor_chipping_sycl.cpp       |  4 +-
 unsupported/test/cxx11_tensor_comparisons.cpp |  2 +-
 .../cxx11_tensor_complex_cwise_ops_gpu.cu     |  2 +-
 unsupported/test/cxx11_tensor_complex_gpu.cu  |  2 +-
 .../test/cxx11_tensor_concatenation.cpp       |  2 +-
 .../test/cxx11_tensor_concatenation_sycl.cpp  |  4 +-
 unsupported/test/cxx11_tensor_const.cpp       |  2 +-
 unsupported/test/cxx11_tensor_contract_gpu.cu |  2 +-
 .../test/cxx11_tensor_contract_sycl.cpp       |  4 +-
 unsupported/test/cxx11_tensor_contraction.cpp |  2 +-
 unsupported/test/cxx11_tensor_convolution.cpp |  2 +-
 .../test/cxx11_tensor_convolution_sycl.cpp    |  4 +-
 .../test/cxx11_tensor_custom_index.cpp        |  2 +-
 unsupported/test/cxx11_tensor_custom_op.cpp   |  2 +-
 .../test/cxx11_tensor_custom_op_sycl.cpp      |  4 +-
 unsupported/test/cxx11_tensor_device.cu       |  2 +-
 unsupported/test/cxx11_tensor_device_sycl.cpp |  4 +-
 unsupported/test/cxx11_tensor_dimension.cpp   |  2 +-
 unsupported/test/cxx11_tensor_empty.cpp       |  2 +-
 unsupported/test/cxx11_tensor_expr.cpp        |  2 +-
 unsupported/test/cxx11_tensor_fft.cpp         |  2 +-
 unsupported/test/cxx11_tensor_fixed_size.cpp  |  2 +-
 unsupported/test/cxx11_tensor_forced_eval.cpp |  2 +-
 .../test/cxx11_tensor_forced_eval_sycl.cpp    |  4 +-
 unsupported/test/cxx11_tensor_generator.cpp   |  2 +-
 .../test/cxx11_tensor_generator_sycl.cpp      |  4 +-
 unsupported/test/cxx11_tensor_gpu.cu          |  2 +-
 unsupported/test/cxx11_tensor_ifft.cpp        |  2 +-
 unsupported/test/cxx11_tensor_image_patch.cpp |  2 +-
 .../test/cxx11_tensor_image_patch_sycl.cpp    |  4 +-
 unsupported/test/cxx11_tensor_index_list.cpp  |  2 +-
 unsupported/test/cxx11_tensor_inflation.cpp   |  2 +-
 .../test/cxx11_tensor_inflation_sycl.cpp      |  4 +-
 unsupported/test/cxx11_tensor_intdiv.cpp      |  2 +-
 unsupported/test/cxx11_tensor_io.cpp          |  2 +-
 unsupported/test/cxx11_tensor_layout_swap.cpp |  2 +-
 .../test/cxx11_tensor_layout_swap_sycl.cpp    |  4 +-
 unsupported/test/cxx11_tensor_lvalue.cpp      |  2 +-
 unsupported/test/cxx11_tensor_map.cpp         |  2 +-
 unsupported/test/cxx11_tensor_math.cpp        |  2 +-
 .../test/cxx11_tensor_mixed_indices.cpp       |  2 +-
 unsupported/test/cxx11_tensor_morphing.cpp    |  2 +-
 .../test/cxx11_tensor_morphing_sycl.cpp       |  4 +-
 .../test/cxx11_tensor_notification.cpp        |  2 +-
 unsupported/test/cxx11_tensor_of_complex.cpp  |  2 +-
 .../test/cxx11_tensor_of_const_values.cpp     |  2 +-
 .../test/cxx11_tensor_of_float16_gpu.cu       |  2 +-
 unsupported/test/cxx11_tensor_of_strings.cpp  |  2 +-
 unsupported/test/cxx11_tensor_padding.cpp     |  2 +-
 .../test/cxx11_tensor_padding_sycl.cpp        |  4 +-
 unsupported/test/cxx11_tensor_patch.cpp       |  2 +-
 unsupported/test/cxx11_tensor_patch_sycl.cpp  |  4 +-
 unsupported/test/cxx11_tensor_random.cpp      |  2 +-
 unsupported/test/cxx11_tensor_random_gpu.cu   |  2 +-
 unsupported/test/cxx11_tensor_reduction.cpp   |  2 +-
 .../test/cxx11_tensor_reduction_gpu.cu        |  2 +-
 .../test/cxx11_tensor_reduction_sycl.cpp      |  4 +-
 unsupported/test/cxx11_tensor_ref.cpp         |  2 +-
 unsupported/test/cxx11_tensor_reverse.cpp     |  2 +-
 .../test/cxx11_tensor_reverse_sycl.cpp        |  4 +-
 unsupported/test/cxx11_tensor_roundings.cpp   |  2 +-
 unsupported/test/cxx11_tensor_scan.cpp        |  2 +-
 unsupported/test/cxx11_tensor_scan_gpu.cu     |  2 +-
 unsupported/test/cxx11_tensor_shuffling.cpp   |  2 +-
 .../test/cxx11_tensor_shuffling_sycl.cpp      |  4 +-
 unsupported/test/cxx11_tensor_simple.cpp      |  2 +-
 unsupported/test/cxx11_tensor_striding.cpp    |  2 +-
 .../test/cxx11_tensor_striding_sycl.cpp       |  4 +-
 unsupported/test/cxx11_tensor_sugar.cpp       |  2 +-
 unsupported/test/cxx11_tensor_sycl.cpp        |  4 +-
 unsupported/test/cxx11_tensor_symmetry.cpp    |  2 +-
 unsupported/test/cxx11_tensor_thread_pool.cpp |  2 +-
 unsupported/test/cxx11_tensor_trace.cpp       |  2 +-
 unsupported/test/cxx11_tensor_uint128.cpp     |  2 +-
 .../test/cxx11_tensor_volume_patch.cpp        |  2 +-
 .../test/cxx11_tensor_volume_patch_sycl.cpp   |  4 +-
 unsupported/test/dgmres.cpp                   |  2 +-
 unsupported/test/forward_adolc.cpp            |  2 +-
 unsupported/test/gmres.cpp                    |  2 +-
 unsupported/test/kronecker_product.cpp        |  4 +-
 unsupported/test/levenberg_marquardt.cpp      |  2 +-
 unsupported/test/matrix_exponential.cpp       |  2 +-
 unsupported/test/matrix_function.cpp          |  2 +-
 unsupported/test/matrix_power.cpp             |  2 +-
 unsupported/test/matrix_square_root.cpp       |  2 +-
 unsupported/test/minres.cpp                   |  2 +-
 unsupported/test/mpreal_support.cpp           |  2 +-
 unsupported/test/openglsupport.cpp            |  2 +-
 unsupported/test/polynomialsolver.cpp         |  2 +-
 unsupported/test/polynomialutils.cpp          |  2 +-
 unsupported/test/sparse_extra.cpp             |  2 +-
 unsupported/test/special_functions.cpp        |  2 +-
 unsupported/test/splines.cpp                  |  2 +-
 255 files changed, 325 insertions(+), 303 deletions(-)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index d60873554..35deed509 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -68,8 +68,6 @@ macro(ei_add_test_internal testname testname_with_suffix)
 
   ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_MAX_SIZE=${EIGEN_TEST_MAX_SIZE}")
 
-  ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_FUNC=${testname}")
-
   if(MSVC)
     ei_add_target_property(${targetname} COMPILE_FLAGS "/bigobj")
   endif()
@@ -170,8 +168,6 @@ macro(ei_add_test_internal_sycl testname testname_with_suffix)
 
   ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_MAX_SIZE=${EIGEN_TEST_MAX_SIZE}")
 
-  ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_FUNC=${testname}")
-
   if(MSVC AND NOT EIGEN_SPLIT_LARGE_TESTS)
     ei_add_target_property(${targetname} COMPILE_FLAGS "/bigobj")
   endif()
diff --git a/test/adjoint.cpp b/test/adjoint.cpp
index 37032d220..ca44466b2 100644
--- a/test/adjoint.cpp
+++ b/test/adjoint.cpp
@@ -145,7 +145,7 @@ template<typename MatrixType> void adjoint(const MatrixType& m)
   VERIFY_IS_APPROX(rv1.template cast<Scalar>().dot(v1), rv1.dot(v1));
 }
 
-void test_adjoint()
+EIGEN_DECLARE_TEST(adjoint)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( adjoint(Matrix<float, 1, 1>()) );
diff --git a/test/array.cpp b/test/array.cpp
index 7e228ed54..c01653668 100644
--- a/test/array.cpp
+++ b/test/array.cpp
@@ -449,7 +449,7 @@ template<typename ArrayType> void min_max(const ArrayType& m)
 
 }
 
-void test_array()
+EIGEN_DECLARE_TEST(array)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( array(Array<float, 1, 1>()) );
diff --git a/test/array_for_matrix.cpp b/test/array_for_matrix.cpp
index 3dc5dc487..6b03abb10 100644
--- a/test/array_for_matrix.cpp
+++ b/test/array_for_matrix.cpp
@@ -256,7 +256,7 @@ void regrrssion_bug_1410()
   VERIFY((internal::traits<MatrixWrapper<Array4i> >::Flags&LvalueBit)==LvalueBit);
 }
 
-void test_array_for_matrix()
+EIGEN_DECLARE_TEST(array_for_matrix)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( array_for_matrix(Matrix<float, 1, 1>()) );
diff --git a/test/array_of_string.cpp b/test/array_of_string.cpp
index e23b7c59e..23e51529b 100644
--- a/test/array_of_string.cpp
+++ b/test/array_of_string.cpp
@@ -9,7 +9,7 @@
 
 #include "main.h"
 
-void test_array_of_string()
+EIGEN_DECLARE_TEST(array_of_string)
 {
   typedef Array<std::string,1,Dynamic> ArrayXs;
   ArrayXs a1(3), a2(3), a3(3), a3ref(3);
diff --git a/test/array_replicate.cpp b/test/array_replicate.cpp
index 0dad5bace..057c3c77b 100644
--- a/test/array_replicate.cpp
+++ b/test/array_replicate.cpp
@@ -68,7 +68,7 @@ template<typename MatrixType> void replicate(const MatrixType& m)
   VERIFY_IS_APPROX(vx1, v1.colwise().replicate(f2));
 }
 
-void test_array_replicate()
+EIGEN_DECLARE_TEST(array_replicate)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( replicate(Matrix<float, 1, 1>()) );
diff --git a/test/array_reverse.cpp b/test/array_reverse.cpp
index 9d5b9a66d..317b9a6e2 100644
--- a/test/array_reverse.cpp
+++ b/test/array_reverse.cpp
@@ -123,7 +123,7 @@ template<typename MatrixType> void reverse(const MatrixType& m)
   VERIFY_IS_APPROX(x, m1(r, cols - 1 - c));
 }
 
-void test_array_reverse()
+EIGEN_DECLARE_TEST(array_reverse)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( reverse(Matrix<float, 1, 1>()) );
diff --git a/test/bandmatrix.cpp b/test/bandmatrix.cpp
index f8c38f7c3..66a1b0db4 100644
--- a/test/bandmatrix.cpp
+++ b/test/bandmatrix.cpp
@@ -59,7 +59,7 @@ template<typename MatrixType> void bandmatrix(const MatrixType& _m)
 
 using Eigen::internal::BandMatrix;
 
-void test_bandmatrix()
+EIGEN_DECLARE_TEST(bandmatrix)
 {
   for(int i = 0; i < 10*g_repeat ; i++) {
     Index rows = internal::random<Index>(1,10);
diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp
index 0fbeb42eb..5ad4090cd 100644
--- a/test/basicstuff.cpp
+++ b/test/basicstuff.cpp
@@ -268,7 +268,7 @@ void fixedSizeMatrixConstruction()
   }
 }
 
-void test_basicstuff()
+EIGEN_DECLARE_TEST(basicstuff)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( basicStuff(Matrix<float, 1, 1>()) );
diff --git a/test/bdcsvd.cpp b/test/bdcsvd.cpp
index 109218766..3065ff015 100644
--- a/test/bdcsvd.cpp
+++ b/test/bdcsvd.cpp
@@ -62,7 +62,7 @@ void compare_bdc_jacobi(const MatrixType& a = MatrixType(), unsigned int computa
   if(computationOptions & ComputeThinV) VERIFY_IS_APPROX(bdc_svd.matrixV(), jacobi_svd.matrixV());
 }
 
-void test_bdcsvd()
+EIGEN_DECLARE_TEST(bdcsvd)
 {
   CALL_SUBTEST_3(( svd_verify_assert<BDCSVD<Matrix3f>  >(Matrix3f()) ));
   CALL_SUBTEST_4(( svd_verify_assert<BDCSVD<Matrix4d>  >(Matrix4d()) ));
diff --git a/test/bicgstab.cpp b/test/bicgstab.cpp
index 4cc0dd31c..89d6a45ef 100644
--- a/test/bicgstab.cpp
+++ b/test/bicgstab.cpp
@@ -26,7 +26,7 @@ template<typename T, typename I> void test_bicgstab_T()
   //CALL_SUBTEST( check_sparse_square_solving(bicgstab_colmajor_ssor)     );
 }
 
-void test_bicgstab()
+EIGEN_DECLARE_TEST(bicgstab)
 {
   CALL_SUBTEST_1((test_bicgstab_T<double,int>()) );
   CALL_SUBTEST_2((test_bicgstab_T<std::complex<double>, int>()));
diff --git a/test/block.cpp b/test/block.cpp
index a45dd501d..ca13539a9 100644
--- a/test/block.cpp
+++ b/test/block.cpp
@@ -276,7 +276,7 @@ void data_and_stride(const MatrixType& m)
   compare_using_data_and_stride(m1.col(c1).transpose());
 }
 
-void test_block()
+EIGEN_DECLARE_TEST(block)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( block(Matrix<float, 1, 1>()) );
diff --git a/test/boostmultiprec.cpp b/test/boostmultiprec.cpp
index bafbd3b96..1588eb86d 100644
--- a/test/boostmultiprec.cpp
+++ b/test/boostmultiprec.cpp
@@ -145,7 +145,7 @@ namespace Eigen {
 
 }
 
-void test_boostmultiprec()
+EIGEN_DECLARE_TEST(boostmultiprec)
 {
   typedef Matrix<Real,Dynamic,Dynamic> Mat;
   typedef Matrix<std::complex<Real>,Dynamic,Dynamic> MatC;
diff --git a/test/cholesky.cpp b/test/cholesky.cpp
index 62c391de4..b871351e0 100644
--- a/test/cholesky.cpp
+++ b/test/cholesky.cpp
@@ -488,7 +488,7 @@ template<typename MatrixType> void cholesky_verify_assert()
   VERIFY_RAISES_ASSERT(ldlt.solveInPlace(&tmp))
 }
 
-void test_cholesky()
+EIGEN_DECLARE_TEST(cholesky)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/cholmod_support.cpp b/test/cholmod_support.cpp
index 931207334..89b9cf41e 100644
--- a/test/cholmod_support.cpp
+++ b/test/cholmod_support.cpp
@@ -55,7 +55,7 @@ template<typename T, int flags, typename IdxType> void test_cholmod_T()
     test_cholmod_ST<SparseMatrix<T, flags, IdxType> >();
 }
 
-void test_cholmod_support()
+EIGEN_DECLARE_TEST(cholmod_support)
 {
   CALL_SUBTEST_11( (test_cholmod_T<double              , ColMajor, int >()) );
   CALL_SUBTEST_12( (test_cholmod_T<double              , ColMajor, long>()) );
diff --git a/test/commainitializer.cpp b/test/commainitializer.cpp
index 9844adbd2..3cb94da62 100644
--- a/test/commainitializer.cpp
+++ b/test/commainitializer.cpp
@@ -65,7 +65,7 @@ struct test_block_recursion<-1>
   static void run() { }
 };
 
-void test_commainitializer()
+EIGEN_DECLARE_TEST(commainitializer)
 {
   Matrix3d m3;
   Matrix4d m4;
diff --git a/test/conjugate_gradient.cpp b/test/conjugate_gradient.cpp
index 9622fd86d..47a4ca707 100644
--- a/test/conjugate_gradient.cpp
+++ b/test/conjugate_gradient.cpp
@@ -26,7 +26,7 @@ template<typename T, typename I> void test_conjugate_gradient_T()
   CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_upper_I)     );
 }
 
-void test_conjugate_gradient()
+EIGEN_DECLARE_TEST(conjugate_gradient)
 {
   CALL_SUBTEST_1(( test_conjugate_gradient_T<double,int>() ));
   CALL_SUBTEST_2(( test_conjugate_gradient_T<std::complex<double>, int>() ));
diff --git a/test/conservative_resize.cpp b/test/conservative_resize.cpp
index 21a1db4ac..d3d0a5a31 100644
--- a/test/conservative_resize.cpp
+++ b/test/conservative_resize.cpp
@@ -109,7 +109,7 @@ void run_vector_tests()
   }
 }
 
-void test_conservative_resize()
+EIGEN_DECLARE_TEST(conservative_resize)
 {
   for(int i=0; i<g_repeat; ++i)
   {
diff --git a/test/constructor.cpp b/test/constructor.cpp
index eec9e2192..1dd3bc3c0 100644
--- a/test/constructor.cpp
+++ b/test/constructor.cpp
@@ -37,7 +37,7 @@ template<typename MatrixType> void ctor_init1(const MatrixType& m)
 }
 
 
-void test_constructor()
+EIGEN_DECLARE_TEST(constructor)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( ctor_init1(Matrix<float, 1, 1>()) );
diff --git a/test/corners.cpp b/test/corners.cpp
index 32edadb25..73342a8dd 100644
--- a/test/corners.cpp
+++ b/test/corners.cpp
@@ -101,7 +101,7 @@ template<typename MatrixType, int CRows, int CCols, int SRows, int SCols> void c
   VERIFY_IS_EQUAL((const_matrix.template rightCols<c>()), (const_matrix.template block<rows,c>(0,cols-c)));
 }
 
-void test_corners()
+EIGEN_DECLARE_TEST(corners)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( corners(Matrix<float, 1, 1>()) );
diff --git a/test/ctorleak.cpp b/test/ctorleak.cpp
index c158f5e4e..7202e90dd 100644
--- a/test/ctorleak.cpp
+++ b/test/ctorleak.cpp
@@ -33,7 +33,7 @@ Index Foo::object_limit = 0;
 #undef EIGEN_TEST_MAX_SIZE
 #define EIGEN_TEST_MAX_SIZE 3
 
-void test_ctorleak()
+EIGEN_DECLARE_TEST(ctorleak)
 {
   typedef Matrix<Foo, Dynamic, Dynamic> MatrixX;
   typedef Matrix<Foo, Dynamic, 1> VectorX;
diff --git a/test/denseLM.cpp b/test/denseLM.cpp
index 0aa736ea3..afb8004b1 100644
--- a/test/denseLM.cpp
+++ b/test/denseLM.cpp
@@ -182,7 +182,7 @@ void test_denseLM_T()
   
 }
 
-void test_denseLM()
+EIGEN_DECLARE_TEST(denseLM)
 {
   CALL_SUBTEST_2(test_denseLM_T<double>());
   
diff --git a/test/dense_storage.cpp b/test/dense_storage.cpp
index e63712b1a..1150ec52b 100644
--- a/test/dense_storage.cpp
+++ b/test/dense_storage.cpp
@@ -52,7 +52,7 @@ void dense_storage_assignment()
     VERIFY_IS_EQUAL(raw_reference[i], raw_copied_reference[i]);
 }
 
-void test_dense_storage()
+EIGEN_DECLARE_TEST(dense_storage)
 {
   dense_storage_copy<int,Dynamic,Dynamic>();  
   dense_storage_copy<int,Dynamic,3>();
diff --git a/test/determinant.cpp b/test/determinant.cpp
index b8c9babb3..7dd33c373 100644
--- a/test/determinant.cpp
+++ b/test/determinant.cpp
@@ -50,7 +50,7 @@ template<typename MatrixType> void determinant(const MatrixType& m)
   VERIFY_IS_APPROX(m2.block(0,0,0,0).determinant(), Scalar(1));
 }
 
-void test_determinant()
+EIGEN_DECLARE_TEST(determinant)
 {
   for(int i = 0; i < g_repeat; i++) {
     int s = 0;
diff --git a/test/diagonal.cpp b/test/diagonal.cpp
index 8ed9b4682..4e8c4b3c9 100644
--- a/test/diagonal.cpp
+++ b/test/diagonal.cpp
@@ -88,7 +88,7 @@ template<typename MatrixType> void diagonal_assert(const MatrixType& m) {
   VERIFY_RAISES_ASSERT( m1.diagonal(-(rows+1)) );
 }
 
-void test_diagonal()
+EIGEN_DECLARE_TEST(diagonal)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( diagonal(Matrix<float, 1, 1>()) );
diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp
index c55733df8..ba58ca8d1 100644
--- a/test/diagonalmatrices.cpp
+++ b/test/diagonalmatrices.cpp
@@ -144,7 +144,7 @@ void bug987()
   VERIFY_IS_APPROX(( res1 = points.topLeftCorner<2,2>()*diag.asDiagonal()) , res2 = tmp2*diag.asDiagonal() );
 }
 
-void test_diagonalmatrices()
+EIGEN_DECLARE_TEST(diagonalmatrices)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( diagonalmatrices(Matrix<float, 1, 1>()) );
diff --git a/test/dontalign.cpp b/test/dontalign.cpp
index ac00112ed..2e4102b86 100644
--- a/test/dontalign.cpp
+++ b/test/dontalign.cpp
@@ -44,7 +44,7 @@ void dontalign(const MatrixType& m)
   internal::aligned_delete(array, rows);
 }
 
-void test_dontalign()
+EIGEN_DECLARE_TEST(dontalign)
 {
 #if defined EIGEN_TEST_PART_1 || defined EIGEN_TEST_PART_5
   dontalign(Matrix3d());
diff --git a/test/dynalloc.cpp b/test/dynalloc.cpp
index f1cc70bee..ceecd76e3 100644
--- a/test/dynalloc.cpp
+++ b/test/dynalloc.cpp
@@ -119,7 +119,7 @@ template<typename T> void check_custom_new_delete()
 #endif
 }
 
-void test_dynalloc()
+EIGEN_DECLARE_TEST(dynalloc)
 {
   // low level dynamic memory allocation
   CALL_SUBTEST(check_handmade_aligned_malloc());
diff --git a/test/eigen2support.cpp b/test/eigen2support.cpp
index ac6931a0e..49d7328e9 100644
--- a/test/eigen2support.cpp
+++ b/test/eigen2support.cpp
@@ -52,7 +52,7 @@ template<typename MatrixType> void eigen2support(const MatrixType& m)
   m1.minor(0,0);
 }
 
-void test_eigen2support()
+EIGEN_DECLARE_TEST(eigen2support)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( eigen2support(Matrix<double,1,1>()) );
diff --git a/test/eigensolver_complex.cpp b/test/eigensolver_complex.cpp
index 8f279dc00..c5373f420 100644
--- a/test/eigensolver_complex.cpp
+++ b/test/eigensolver_complex.cpp
@@ -152,7 +152,7 @@ template<typename MatrixType> void eigensolver_verify_assert(const MatrixType& m
   VERIFY_RAISES_ASSERT(eig.eigenvectors());
 }
 
-void test_eigensolver_complex()
+EIGEN_DECLARE_TEST(eigensolver_complex)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/eigensolver_generalized_real.cpp b/test/eigensolver_generalized_real.cpp
index 9dd44c89d..95ed431db 100644
--- a/test/eigensolver_generalized_real.cpp
+++ b/test/eigensolver_generalized_real.cpp
@@ -85,7 +85,7 @@ template<typename MatrixType> void generalized_eigensolver_real(const MatrixType
   }
 }
 
-void test_eigensolver_generalized_real()
+EIGEN_DECLARE_TEST(eigensolver_generalized_real)
 {
   for(int i = 0; i < g_repeat; i++) {
     int s = 0;
diff --git a/test/eigensolver_generic.cpp b/test/eigensolver_generic.cpp
index 07bf65e03..b41186819 100644
--- a/test/eigensolver_generic.cpp
+++ b/test/eigensolver_generic.cpp
@@ -100,7 +100,7 @@ template<typename MatrixType> void eigensolver_verify_assert(const MatrixType& m
   VERIFY_RAISES_ASSERT(eig.pseudoEigenvectors());
 }
 
-void test_eigensolver_generic()
+EIGEN_DECLARE_TEST(eigensolver_generic)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp
index 0e39b5364..65b80c3fb 100644
--- a/test/eigensolver_selfadjoint.cpp
+++ b/test/eigensolver_selfadjoint.cpp
@@ -230,7 +230,7 @@ void bug_1204()
   SelfAdjointEigenSolver<Eigen::SparseMatrix<double> > eig(A);
 }
 
-void test_eigensolver_selfadjoint()
+EIGEN_DECLARE_TEST(eigensolver_selfadjoint)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/evaluators.cpp b/test/evaluators.cpp
index aed5a05a7..f4fdaf053 100644
--- a/test/evaluators.cpp
+++ b/test/evaluators.cpp
@@ -101,7 +101,7 @@ using namespace std;
 #define VERIFY_IS_APPROX_EVALUATOR(DEST,EXPR) VERIFY_IS_APPROX(copy_using_evaluator(DEST,(EXPR)), (EXPR).eval());
 #define VERIFY_IS_APPROX_EVALUATOR2(DEST,EXPR,REF) VERIFY_IS_APPROX(copy_using_evaluator(DEST,(EXPR)), (REF).eval());
 
-void test_evaluators()
+EIGEN_DECLARE_TEST(evaluators)
 {
   // Testing Matrix evaluator and Transpose
   Vector2d v = Vector2d::Random();
diff --git a/test/exceptions.cpp b/test/exceptions.cpp
index b83fb82ba..182082e39 100644
--- a/test/exceptions.cpp
+++ b/test/exceptions.cpp
@@ -107,7 +107,7 @@ void memoryleak()
   VERIFY(ScalarWithExceptions::instances==0 && "global memory leak detected in " && EIGEN_MAKESTRING(OP)); \
 }
 
-void test_exceptions()
+EIGEN_DECLARE_TEST(exceptions)
 {
   CALL_SUBTEST( memoryleak() );
 }
diff --git a/test/fastmath.cpp b/test/fastmath.cpp
index cc5db0746..c30f0a846 100644
--- a/test/fastmath.cpp
+++ b/test/fastmath.cpp
@@ -88,7 +88,7 @@ void check_inf_nan(bool dryrun) {
   }
 }
 
-void test_fastmath() {
+EIGEN_DECLARE_TEST(fastmath) {
   std::cout << "*** float *** \n\n"; check_inf_nan<float>(true);
   std::cout << "*** double ***\n\n"; check_inf_nan<double>(true);
   std::cout << "*** long double *** \n\n"; check_inf_nan<long double>(true);
diff --git a/test/first_aligned.cpp b/test/first_aligned.cpp
index ae2d4bc42..ed9945077 100644
--- a/test/first_aligned.cpp
+++ b/test/first_aligned.cpp
@@ -26,7 +26,7 @@ void test_none_aligned_helper(Scalar *array, int size)
 
 struct some_non_vectorizable_type { float x; };
 
-void test_first_aligned()
+EIGEN_DECLARE_TEST(first_aligned)
 {
   EIGEN_ALIGN16 float array_float[100];
   test_first_aligned_helper(array_float, 50);
diff --git a/test/geo_alignedbox.cpp b/test/geo_alignedbox.cpp
index 87bf1140f..c6c051ce4 100644
--- a/test/geo_alignedbox.cpp
+++ b/test/geo_alignedbox.cpp
@@ -169,7 +169,7 @@ void specificTest2()
 }
 
 
-void test_geo_alignedbox()
+EIGEN_DECLARE_TEST(geo_alignedbox)
 {
   for(int i = 0; i < g_repeat; i++)
   {
diff --git a/test/geo_eulerangles.cpp b/test/geo_eulerangles.cpp
index 932ebe773..693c627a9 100644
--- a/test/geo_eulerangles.cpp
+++ b/test/geo_eulerangles.cpp
@@ -103,7 +103,7 @@ template<typename Scalar> void eulerangles()
   check_all_var(ea);
 }
 
-void test_geo_eulerangles()
+EIGEN_DECLARE_TEST(geo_eulerangles)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( eulerangles<float>() );
diff --git a/test/geo_homogeneous.cpp b/test/geo_homogeneous.cpp
index 2187c7bf9..9aebe6226 100644
--- a/test/geo_homogeneous.cpp
+++ b/test/geo_homogeneous.cpp
@@ -115,7 +115,7 @@ template<typename Scalar,int Size> void homogeneous(void)
   VERIFY_IS_APPROX( (t2.template triangularView<Lower>() * v0.homogeneous()).eval(), (t2.template triangularView<Lower>()*hv0) );
 }
 
-void test_geo_homogeneous()
+EIGEN_DECLARE_TEST(geo_homogeneous)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(( homogeneous<float,1>() ));
diff --git a/test/geo_hyperplane.cpp b/test/geo_hyperplane.cpp
index b3a48c585..a26709301 100644
--- a/test/geo_hyperplane.cpp
+++ b/test/geo_hyperplane.cpp
@@ -180,7 +180,7 @@ template<typename Scalar> void hyperplane_alignment()
 }
 
 
-void test_geo_hyperplane()
+EIGEN_DECLARE_TEST(geo_hyperplane)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( hyperplane(Hyperplane<float,2>()) );
diff --git a/test/geo_orthomethods.cpp b/test/geo_orthomethods.cpp
index e178df257..b7b660740 100644
--- a/test/geo_orthomethods.cpp
+++ b/test/geo_orthomethods.cpp
@@ -115,7 +115,7 @@ template<typename Scalar, int Size> void orthomethods(int size=Size)
   VERIFY_IS_APPROX(mcrossN3.row(i), matN3.row(i).cross(vec3));
 }
 
-void test_geo_orthomethods()
+EIGEN_DECLARE_TEST(geo_orthomethods)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( orthomethods_3<float>() );
diff --git a/test/geo_parametrizedline.cpp b/test/geo_parametrizedline.cpp
index 14cf4fd48..7135c8fa5 100644
--- a/test/geo_parametrizedline.cpp
+++ b/test/geo_parametrizedline.cpp
@@ -117,7 +117,7 @@ template<typename Scalar> void parametrizedline_alignment()
   #endif
 }
 
-void test_geo_parametrizedline()
+EIGEN_DECLARE_TEST(geo_parametrizedline)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( parametrizedline(ParametrizedLine<float,2>()) );
diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp
index 5854d39c5..3541b8b25 100644
--- a/test/geo_quaternion.cpp
+++ b/test/geo_quaternion.cpp
@@ -285,7 +285,7 @@ template<typename PlainObjectType> void check_const_correctness(const PlainObjec
   VERIFY( !(Map<ConstPlainObjectType, Aligned>::Flags & LvalueBit) );
 }
 
-void test_geo_quaternion()
+EIGEN_DECLARE_TEST(geo_quaternion)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(( quaternion<float,AutoAlign>() ));
diff --git a/test/geo_transformations.cpp b/test/geo_transformations.cpp
index 278e527c2..bf920696b 100755
--- a/test/geo_transformations.cpp
+++ b/test/geo_transformations.cpp
@@ -612,7 +612,7 @@ template<typename Scalar, int Dim, int Options> void transform_products()
   VERIFY_IS_APPROX((ac*p).matrix(), a_m*p_m);
 }
 
-void test_geo_transformations()
+EIGEN_DECLARE_TEST(geo_transformations)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(( transformations<double,Affine,AutoAlign>() ));
diff --git a/test/gpu_basic.cu b/test/gpu_basic.cu
index 285b87910..e8069f185 100644
--- a/test/gpu_basic.cu
+++ b/test/gpu_basic.cu
@@ -15,7 +15,6 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC gpu_basic
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 
 #include "main.h"
@@ -162,7 +161,7 @@ struct matrix_inverse {
   }
 };
 
-void test_gpu_basic()
+EIGEN_DECLARE_TEST(gpu_basic)
 {
   ei_test_init_gpu();
   
diff --git a/test/half_float.cpp b/test/half_float.cpp
index 5a881680a..2a7f9b497 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -272,7 +272,7 @@ void test_product()
   VERIFY_IS_APPROX(Ch.noalias()+=Ah*Bh, (Cf.noalias()+=Af*Bf).cast<half>());
 }
 
-void test_half_float()
+EIGEN_DECLARE_TEST(half_float)
 {
   CALL_SUBTEST(test_numtraits());
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/hessenberg.cpp b/test/hessenberg.cpp
index 96bc19e2e..0e1b0098d 100644
--- a/test/hessenberg.cpp
+++ b/test/hessenberg.cpp
@@ -49,7 +49,7 @@ template<typename Scalar,int Size> void hessenberg(int size = Size)
   // TODO: Add tests for packedMatrix() and householderCoefficients()
 }
 
-void test_hessenberg()
+EIGEN_DECLARE_TEST(hessenberg)
 {
   CALL_SUBTEST_1(( hessenberg<std::complex<double>,1>() ));
   CALL_SUBTEST_2(( hessenberg<std::complex<double>,2>() ));
diff --git a/test/householder.cpp b/test/householder.cpp
index ccda46811..cad8138a2 100644
--- a/test/householder.cpp
+++ b/test/householder.cpp
@@ -133,7 +133,7 @@ template<typename MatrixType> void householder(const MatrixType& m)
   VERIFY_IS_APPROX(m3 * m5, m1); // test evaluating rhseq to a dense matrix, then applying
 }
 
-void test_householder()
+EIGEN_DECLARE_TEST(householder)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( householder(Matrix<double,2,2>()) );
diff --git a/test/incomplete_cholesky.cpp b/test/incomplete_cholesky.cpp
index 59ffe9259..52235c2c2 100644
--- a/test/incomplete_cholesky.cpp
+++ b/test/incomplete_cholesky.cpp
@@ -29,7 +29,7 @@ template<typename T, typename I> void test_incomplete_cholesky_T()
   CALL_SUBTEST( check_sparse_spd_solving(cg_illt_uplo_amd) );
 }
 
-void test_incomplete_cholesky()
+EIGEN_DECLARE_TEST(incomplete_cholesky)
 {
   CALL_SUBTEST_1(( test_incomplete_cholesky_T<double,int>() ));
   CALL_SUBTEST_2(( test_incomplete_cholesky_T<std::complex<double>, int>() ));
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 551dc55b0..4cc0954f7 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -388,7 +388,7 @@ void check_indexed_view()
 
 }
 
-void test_indexed_view()
+EIGEN_DECLARE_TEST(indexed_view)
 {
 //   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( check_indexed_view() );
diff --git a/test/inplace_decomposition.cpp b/test/inplace_decomposition.cpp
index 92d0d91b6..e3aa9957d 100644
--- a/test/inplace_decomposition.cpp
+++ b/test/inplace_decomposition.cpp
@@ -79,7 +79,7 @@ template<typename DecType,typename MatrixType> void inplace(bool square = false,
 }
 
 
-void test_inplace_decomposition()
+EIGEN_DECLARE_TEST(inplace_decomposition)
 {
   EIGEN_UNUSED typedef Matrix<double,4,3> Matrix43d;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/integer_types.cpp b/test/integer_types.cpp
index 36295598f..88573c568 100644
--- a/test/integer_types.cpp
+++ b/test/integer_types.cpp
@@ -131,7 +131,7 @@ template<typename MatrixType> void integer_type_tests(const MatrixType& m)
   VERIFY_IS_APPROX((m1 * m2.transpose()) * m1, m1 * (m2.transpose() * m1));
 }
 
-void test_integer_types()
+EIGEN_DECLARE_TEST(integer_types)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( integer_type_tests(Matrix<unsigned int, 1, 1>()) );
diff --git a/test/inverse.cpp b/test/inverse.cpp
index be607cc8b..b989c7905 100644
--- a/test/inverse.cpp
+++ b/test/inverse.cpp
@@ -92,7 +92,7 @@ template<typename MatrixType> void inverse(const MatrixType& m)
   }
 }
 
-void test_inverse()
+EIGEN_DECLARE_TEST(inverse)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/is_same_dense.cpp b/test/is_same_dense.cpp
index c4e2fbc92..23dd806eb 100644
--- a/test/is_same_dense.cpp
+++ b/test/is_same_dense.cpp
@@ -11,7 +11,7 @@
 
 using internal::is_same_dense;
 
-void test_is_same_dense()
+EIGEN_DECLARE_TEST(is_same_dense)
 {
   typedef Matrix<double,Dynamic,Dynamic,ColMajor> ColMatrixXd;
   typedef Matrix<std::complex<double>,Dynamic,Dynamic,ColMajor> ColMatrixXcd;
diff --git a/test/jacobi.cpp b/test/jacobi.cpp
index 319e4767a..5604797f5 100644
--- a/test/jacobi.cpp
+++ b/test/jacobi.cpp
@@ -57,7 +57,7 @@ void jacobi(const MatrixType& m = MatrixType())
   }
 }
 
-void test_jacobi()
+EIGEN_DECLARE_TEST(jacobi)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(( jacobi<Matrix3f, float>() ));
diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp
index 72007bff5..f9a59e0e7 100644
--- a/test/jacobisvd.cpp
+++ b/test/jacobisvd.cpp
@@ -69,7 +69,7 @@ void jacobisvd_method()
   VERIFY_IS_APPROX(m.jacobiSvd(ComputeFullU|ComputeFullV).solve(m), m);
 }
 
-void test_jacobisvd()
+EIGEN_DECLARE_TEST(jacobisvd)
 {
   CALL_SUBTEST_3(( jacobisvd_verify_assert(Matrix3f()) ));
   CALL_SUBTEST_4(( jacobisvd_verify_assert(Matrix4d()) ));
diff --git a/test/klu_support.cpp b/test/klu_support.cpp
index 138dcc301..f806ad50e 100644
--- a/test/klu_support.cpp
+++ b/test/klu_support.cpp
@@ -24,7 +24,7 @@ template<typename T> void test_klu_support_T()
   //check_sparse_square_determinant(umfpack_rowmajor);
 }
 
-void test_klu_support()
+EIGEN_DECLARE_TEST(klu_support)
 {
   CALL_SUBTEST_1(test_klu_support_T<double>());
   CALL_SUBTEST_2(test_klu_support_T<std::complex<double> >());
diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp
index b6559b2a0..4137b0251 100644
--- a/test/linearstructure.cpp
+++ b/test/linearstructure.cpp
@@ -110,7 +110,7 @@ template<typename MatrixType> void real_complex(DenseIndex rows = MatrixType::Ro
   VERIFY(g_called && "matrix<complex> - real not properly optimized");
 }
 
-void test_linearstructure()
+EIGEN_DECLARE_TEST(linearstructure)
 {
   g_called = true;
   VERIFY(g_called); // avoid `unneeded-internal-declaration` warning.
diff --git a/test/lscg.cpp b/test/lscg.cpp
index d49ee00c3..feb2347a8 100644
--- a/test/lscg.cpp
+++ b/test/lscg.cpp
@@ -30,7 +30,7 @@ template<typename T> void test_lscg_T()
   CALL_SUBTEST( check_sparse_leastsquare_solving(lscg_rowmajor_I)     );
 }
 
-void test_lscg()
+EIGEN_DECLARE_TEST(lscg)
 {
   CALL_SUBTEST_1(test_lscg_T<double>());
   CALL_SUBTEST_2(test_lscg_T<std::complex<double> >());
diff --git a/test/lu.cpp b/test/lu.cpp
index cad703dd0..144496e91 100644
--- a/test/lu.cpp
+++ b/test/lu.cpp
@@ -242,7 +242,7 @@ template<typename MatrixType> void lu_verify_assert()
   VERIFY_RAISES_ASSERT(plu.inverse())
 }
 
-void test_lu()
+EIGEN_DECLARE_TEST(lu)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( lu_non_invertible<Matrix3f>() );
diff --git a/test/main.h b/test/main.h
index 51399f7ad..5d64bc736 100644
--- a/test/main.h
+++ b/test/main.h
@@ -128,7 +128,6 @@ inline void on_temporary_creation(long int size) {
   
 #endif
 
-// the following file is automatically generated by cmake
 #include "split_test_helper.h"
 
 #ifdef NDEBUG
@@ -145,10 +144,6 @@ inline void on_temporary_creation(long int size) {
 #define EIGEN_MAKING_DOCS
 #endif
 
-#ifndef EIGEN_TEST_FUNC
-#error EIGEN_TEST_FUNC must be defined
-#endif
-
 #define DEFAULT_REPEAT 10
 
 namespace Eigen
@@ -157,17 +152,45 @@ namespace Eigen
   // level == 0 <=> abort if test fail
   // level >= 1 <=> warning message to std::cerr if test fail
   static int g_test_level = 0;
-  static int g_repeat;
-  static unsigned int g_seed;
-  static bool g_has_set_repeat, g_has_set_seed;
+  static int g_repeat = 1;
+  static unsigned int g_seed = 0;
+  static bool g_has_set_repeat = false, g_has_set_seed = false;
+
+  class EigenTest
+  {
+  public:
+    EigenTest() : m_func(0) {}
+    EigenTest(const char* a_name, void (*func)(void))
+      : m_name(a_name), m_func(func)
+    {
+      ms_registered_tests.push_back(this);
+    }
+    const std::string& name() const { return m_name; }
+    void operator()() const { m_func(); }
+
+    static const std::vector<EigenTest*>& all() { return ms_registered_tests; }
+  protected:
+    std::string m_name;
+    void (*m_func)(void);
+    static std::vector<EigenTest*> ms_registered_tests;
+  };
+
+  std::vector<EigenTest*> EigenTest::ms_registered_tests;
+
+  // Declare and register a test, e.g.:
+  //    EIGEN_DECLARE_TEST(mytest) { ... }
+  // will create a function:
+  //    void test_mytest() { ... }
+  // that will be automatically called.
+  #define EIGEN_DECLARE_TEST(X) \
+    void EIGEN_CAT(test_,X) (); \
+    static EigenTest EIGEN_CAT(test_handler_,X) (EIGEN_MAKESTRING(X), & EIGEN_CAT(test_,X)); \
+    void EIGEN_CAT(test_,X) ()
 }
 
 #define TRACK std::cerr << __FILE__ << " " << __LINE__ << std::endl
 // #define TRACK while()
 
-#define EI_PP_MAKE_STRING2(S) #S
-#define EI_PP_MAKE_STRING(S) EI_PP_MAKE_STRING2(S)
-
 #define EIGEN_DEFAULT_IO_FORMAT IOFormat(4, 0, "  ", "\n", "", "", "", "")
 
 #if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
@@ -225,7 +248,7 @@ namespace Eigen
       }                                     \
       else if (Eigen::internal::push_assert)       \
       {                                     \
-        eigen_assert_list.push_back(std::string(EI_PP_MAKE_STRING(__FILE__) " (" EI_PP_MAKE_STRING(__LINE__) ") : " #a) ); \
+        eigen_assert_list.push_back(std::string(EIGEN_MAKESTRING(__FILE__) " (" EIGEN_MAKESTRING(__LINE__) ") : " #a) ); \
       }
 
     #ifdef EIGEN_EXCEPTIONS
@@ -338,10 +361,10 @@ inline void verify_impl(bool condition, const char *testname, const char *file,
   }
 }
 
-#define VERIFY(a) ::verify_impl(a, g_test_stack.back().c_str(), __FILE__, __LINE__, EI_PP_MAKE_STRING(a))
+#define VERIFY(a) ::verify_impl(a, g_test_stack.back().c_str(), __FILE__, __LINE__, EIGEN_MAKESTRING(a))
 
-#define VERIFY_GE(a, b) ::verify_impl(a >= b, g_test_stack.back().c_str(), __FILE__, __LINE__, EI_PP_MAKE_STRING(a >= b))
-#define VERIFY_LE(a, b) ::verify_impl(a <= b, g_test_stack.back().c_str(), __FILE__, __LINE__, EI_PP_MAKE_STRING(a <= b))
+#define VERIFY_GE(a, b) ::verify_impl(a >= b, g_test_stack.back().c_str(), __FILE__, __LINE__, EIGEN_MAKESTRING(a >= b))
+#define VERIFY_LE(a, b) ::verify_impl(a <= b, g_test_stack.back().c_str(), __FILE__, __LINE__, EIGEN_MAKESTRING(a <= b))
 
 
 #define VERIFY_IS_EQUAL(a, b) VERIFY(test_is_equal(a, b, true))
@@ -358,7 +381,7 @@ inline void verify_impl(bool condition, const char *testname, const char *file,
 #define STATIC_CHECK(COND) EIGEN_STATIC_ASSERT( (COND) , EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT )
 
 #define CALL_SUBTEST(FUNC) do { \
-    g_test_stack.push_back(EI_PP_MAKE_STRING(FUNC)); \
+    g_test_stack.push_back(EIGEN_MAKESTRING(FUNC)); \
     FUNC; \
     g_test_stack.pop_back(); \
   } while (0)
@@ -709,9 +732,6 @@ template<> std::string type_name<std::complex<double> >()       { return "comple
 template<> std::string type_name<std::complex<long double> >()  { return "complex<long double>"; }
 template<> std::string type_name<std::complex<int> >()          { return "complex<int>"; }
 
-// forward declaration of the main test function
-void EIGEN_CAT(test_,EIGEN_TEST_FUNC)();
-
 using namespace Eigen;
 
 inline void set_repeat_from_string(const char *str)
@@ -798,9 +818,16 @@ int main(int argc, char *argv[])
     srand(g_seed);
     std::cout << "Repeating each test " << g_repeat << " times" << std::endl;
 
-    Eigen::g_test_stack.push_back(std::string(EI_PP_MAKE_STRING(EIGEN_TEST_FUNC)));
+    VERIFY(EigenTest::all().size()>0);
+
+    for(std::size_t i=0; i<EigenTest::all().size(); ++i)
+    {
+      const EigenTest& current_test = *EigenTest::all()[i];
+      Eigen::g_test_stack.push_back(current_test.name());
+      current_test();
+      Eigen::g_test_stack.pop_back();
+    }
 
-    EIGEN_CAT(test_,EIGEN_TEST_FUNC)();
     return 0;
 }
 
diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp
index a530fab41..e1106ba42 100644
--- a/test/mapped_matrix.cpp
+++ b/test/mapped_matrix.cpp
@@ -177,7 +177,7 @@ void map_not_aligned_on_scalar()
   internal::aligned_delete(array1, (size+1)*(size+1)+1);
 }
 
-void test_mapped_matrix()
+EIGEN_DECLARE_TEST(mapped_matrix)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( map_class_vector(Matrix<float, 1, 1>()) );
diff --git a/test/mapstaticmethods.cpp b/test/mapstaticmethods.cpp
index 8156ca939..f32c0beec 100644
--- a/test/mapstaticmethods.cpp
+++ b/test/mapstaticmethods.cpp
@@ -143,7 +143,7 @@ void mapstaticmethods(const PlainObjectType& m)
   VERIFY(true); // just to avoid 'unused function' warning
 }
 
-void test_mapstaticmethods()
+EIGEN_DECLARE_TEST(mapstaticmethods)
 {
   ptr = internal::aligned_new<float>(1000);
   for(int i = 0; i < 1000; i++) ptr[i] = float(i);
diff --git a/test/mapstride.cpp b/test/mapstride.cpp
index d785148cf..09196600b 100644
--- a/test/mapstride.cpp
+++ b/test/mapstride.cpp
@@ -197,7 +197,7 @@ void bug1453()
   VERIFY_IS_APPROX(RowMatrix32i::Map(data, InnerStride<>(2)), RowMatrixXi::Map(data, 3, 2, Stride<4,2>()));
 }
 
-void test_mapstride()
+EIGEN_DECLARE_TEST(mapstride)
 {
   for(int i = 0; i < g_repeat; i++) {
     int maxn = 30;
diff --git a/test/meta.cpp b/test/meta.cpp
index 34c5520a6..a6a67b85c 100644
--- a/test/meta.cpp
+++ b/test/meta.cpp
@@ -27,7 +27,7 @@ struct MyImpl : public MyInterface {
   void func() {}
 };
 
-void test_meta()
+EIGEN_DECLARE_TEST(meta)
 {
   VERIFY((internal::conditional<(3<4),internal::true_type, internal::false_type>::type::value));
   VERIFY(( internal::is_same<float,float>::value));
diff --git a/test/metis_support.cpp b/test/metis_support.cpp
index d87c56a13..b490dacde 100644
--- a/test/metis_support.cpp
+++ b/test/metis_support.cpp
@@ -19,7 +19,7 @@ template<typename T> void test_metis_T()
   check_sparse_square_solving(sparselu_metis); 
 }
 
-void test_metis_support()
+EIGEN_DECLARE_TEST(metis_support)
 {
   CALL_SUBTEST_1(test_metis_T<double>());
 }
diff --git a/test/miscmatrices.cpp b/test/miscmatrices.cpp
index f17291c40..e71712f33 100644
--- a/test/miscmatrices.cpp
+++ b/test/miscmatrices.cpp
@@ -34,7 +34,7 @@ template<typename MatrixType> void miscMatrices(const MatrixType& m)
   VERIFY_IS_APPROX(square, MatrixType::Identity(rows, rows));
 }
 
-void test_miscmatrices()
+EIGEN_DECLARE_TEST(miscmatrices)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( miscMatrices(Matrix<float, 1, 1>()) );
diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp
index b796082cd..38f062f1e 100644
--- a/test/mixingtypes.cpp
+++ b/test/mixingtypes.cpp
@@ -286,7 +286,7 @@ template<int SizeAtCompileType> void mixingtypes(int size = SizeAtCompileType)
   VERIFY_IS_APPROX( rcd.noalias() -= mcd + md*md,           - ((md*md).eval().template cast<CD>()) );
 }
 
-void test_mixingtypes()
+EIGEN_DECLARE_TEST(mixingtypes)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(mixingtypes<3>());
diff --git a/test/nesting_ops.cpp b/test/nesting_ops.cpp
index a419b0e44..4b5fc21f2 100644
--- a/test/nesting_ops.cpp
+++ b/test/nesting_ops.cpp
@@ -91,7 +91,7 @@ template <typename MatrixType> void run_nesting_ops_2(const MatrixType& _m)
 }
 
 
-void test_nesting_ops()
+EIGEN_DECLARE_TEST(nesting_ops)
 {
   CALL_SUBTEST_1(run_nesting_ops_1(MatrixXf::Random(25,25)));
   CALL_SUBTEST_2(run_nesting_ops_1(MatrixXcd::Random(25,25)));
diff --git a/test/nomalloc.cpp b/test/nomalloc.cpp
index d1ca79d68..cb4c073e9 100644
--- a/test/nomalloc.cpp
+++ b/test/nomalloc.cpp
@@ -202,7 +202,7 @@ template<typename MatrixType> void test_reference(const MatrixType& m) {
 
 }
 
-void test_nomalloc()
+EIGEN_DECLARE_TEST(nomalloc)
 {
   // create some dynamic objects
   Eigen::MatrixXd M1 = MatrixXd::Random(3,3);
diff --git a/test/nullary.cpp b/test/nullary.cpp
index 22ec92352..9e6e6eaa0 100644
--- a/test/nullary.cpp
+++ b/test/nullary.cpp
@@ -239,7 +239,7 @@ void testMatrixType(const MatrixType& m)
   VERIFY_IS_APPROX( A(i,j), s1 );
 }
 
-void test_nullary()
+EIGEN_DECLARE_TEST(nullary)
 {
   CALL_SUBTEST_1( testMatrixType(Matrix2d()) );
   CALL_SUBTEST_2( testMatrixType(MatrixXcf(internal::random<int>(1,300),internal::random<int>(1,300))) );
diff --git a/test/num_dimensions.cpp b/test/num_dimensions.cpp
index f5209283d..7ad7ef697 100644
--- a/test/num_dimensions.cpp
+++ b/test/num_dimensions.cpp
@@ -60,7 +60,7 @@ using TMatrix = Matrix<Scalar,Rows,Cols>;
 
 #endif
 
-void test_num_dimensions()
+EIGEN_DECLARE_TEST(num_dimensions)
 {
   int n = 10;
   ArrayXXd A(n,n);
diff --git a/test/numext.cpp b/test/numext.cpp
index 3de33e2f9..6307f5979 100644
--- a/test/numext.cpp
+++ b/test/numext.cpp
@@ -33,7 +33,7 @@ void check_abs() {
   }
 }
 
-void test_numext() {
+EIGEN_DECLARE_TEST(numext) {
   CALL_SUBTEST( check_abs<bool>() );
   CALL_SUBTEST( check_abs<signed char>() );
   CALL_SUBTEST( check_abs<unsigned char>() );
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 56e017383..58a1c60bf 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -627,7 +627,7 @@ template<typename Scalar> void packetmath_scatter_gather()
   }
 }
 
-void test_packetmath()
+EIGEN_DECLARE_TEST(packetmath)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( packetmath<float>() );
diff --git a/test/pardiso_support.cpp b/test/pardiso_support.cpp
index 67efad6d8..9c16ded5b 100644
--- a/test/pardiso_support.cpp
+++ b/test/pardiso_support.cpp
@@ -20,7 +20,7 @@ template<typename T> void test_pardiso_T()
   check_sparse_square_solving(pardiso_lu);
 }
 
-void test_pardiso_support()
+EIGEN_DECLARE_TEST(pardiso_support)
 {
   CALL_SUBTEST_1(test_pardiso_T<float>());
   CALL_SUBTEST_2(test_pardiso_T<double>());
diff --git a/test/pastix_support.cpp b/test/pastix_support.cpp
index b62f85739..9b64417c1 100644
--- a/test/pastix_support.cpp
+++ b/test/pastix_support.cpp
@@ -45,7 +45,7 @@ template<typename T> void test_pastix_T_LU()
   check_sparse_square_solving(pastix_lu);
 }
 
-void test_pastix_support()
+EIGEN_DECLARE_TEST(pastix_support)
 {
   CALL_SUBTEST_1(test_pastix_T<float>());
   CALL_SUBTEST_2(test_pastix_T<double>());
diff --git a/test/permutationmatrices.cpp b/test/permutationmatrices.cpp
index e885f0e04..71f09be0f 100644
--- a/test/permutationmatrices.cpp
+++ b/test/permutationmatrices.cpp
@@ -152,7 +152,7 @@ void bug890()
   VERIFY_IS_APPROX(v1, (P.inverse() * rhs).eval());
 }
 
-void test_permutationmatrices()
+EIGEN_DECLARE_TEST(permutationmatrices)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( permutationmatrices(Matrix<float, 1, 1>()) );
diff --git a/test/prec_inverse_4x4.cpp b/test/prec_inverse_4x4.cpp
index eb6ad18c9..072466467 100644
--- a/test/prec_inverse_4x4.cpp
+++ b/test/prec_inverse_4x4.cpp
@@ -68,7 +68,7 @@ template<typename MatrixType> void inverse_general_4x4(int repeat)
   }
 }
 
-void test_prec_inverse_4x4()
+EIGEN_DECLARE_TEST(prec_inverse_4x4)
 {
   CALL_SUBTEST_1((inverse_permutation_4x4<Matrix4f>()));
   CALL_SUBTEST_1(( inverse_general_4x4<Matrix4f>(200000 * g_repeat) ));
diff --git a/test/product_extra.cpp b/test/product_extra.cpp
index de2709d8b..bd31df84d 100644
--- a/test/product_extra.cpp
+++ b/test/product_extra.cpp
@@ -352,7 +352,7 @@ void bug_1308()
   VERIFY_IS_APPROX(r44.noalias() += Vector4d::Ones() * m44.col(0).transpose(), ones44);
 }
 
-void test_product_extra()
+EIGEN_DECLARE_TEST(product_extra)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( product_extra(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
diff --git a/test/product_large.cpp b/test/product_large.cpp
index 14a4f739d..1b6fec738 100644
--- a/test/product_large.cpp
+++ b/test/product_large.cpp
@@ -30,7 +30,7 @@ void test_aliasing()
   x = z;
 }
 
-void test_product_large()
+EIGEN_DECLARE_TEST(product_large)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( product(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
diff --git a/test/product_mmtr.cpp b/test/product_mmtr.cpp
index d3e24b012..bb19e6e52 100644
--- a/test/product_mmtr.cpp
+++ b/test/product_mmtr.cpp
@@ -84,7 +84,7 @@ template<typename Scalar> void mmtr(int size)
   VERIFY_IS_APPROX(matc, ref2);
 }
 
-void test_product_mmtr()
+EIGEN_DECLARE_TEST(product_mmtr)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp
index b90165b2f..dffb07608 100644
--- a/test/product_notemporary.cpp
+++ b/test/product_notemporary.cpp
@@ -150,7 +150,7 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m)
   VERIFY_EVALUATION_COUNT( rvres.noalias() = rv1 * (m1 * m2.adjoint()), 1 );
 }
 
-void test_product_notemporary()
+EIGEN_DECLARE_TEST(product_notemporary)
 {
   int s;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/product_selfadjoint.cpp b/test/product_selfadjoint.cpp
index 88d68391b..bdccd0491 100644
--- a/test/product_selfadjoint.cpp
+++ b/test/product_selfadjoint.cpp
@@ -59,7 +59,7 @@ template<typename MatrixType> void product_selfadjoint(const MatrixType& m)
   }
 }
 
-void test_product_selfadjoint()
+EIGEN_DECLARE_TEST(product_selfadjoint)
 {
   int s = 0;
   for(int i = 0; i < g_repeat ; i++) {
diff --git a/test/product_small.cpp b/test/product_small.cpp
index fdfdd9f6c..16138631a 100644
--- a/test/product_small.cpp
+++ b/test/product_small.cpp
@@ -228,7 +228,7 @@ void bug_1311()
   VERIFY_IS_APPROX(res, A*b);
 }
 
-void test_product_small()
+EIGEN_DECLARE_TEST(product_small)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( product(Matrix<float, 3, 2>()) );
diff --git a/test/product_symm.cpp b/test/product_symm.cpp
index 7d1042a4f..7d786467d 100644
--- a/test/product_symm.cpp
+++ b/test/product_symm.cpp
@@ -94,7 +94,7 @@ template<typename Scalar, int Size, int OtherSize> void symm(int size = Size, in
 
 }
 
-void test_product_symm()
+EIGEN_DECLARE_TEST(product_symm)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/product_syrk.cpp b/test/product_syrk.cpp
index 3ebbe14ca..23406fe4b 100644
--- a/test/product_syrk.cpp
+++ b/test/product_syrk.cpp
@@ -117,7 +117,7 @@ template<typename MatrixType> void syrk(const MatrixType& m)
                    ((s1 * m1.row(c).adjoint() * m1.row(c).adjoint().adjoint()).eval().template triangularView<Upper>().toDenseMatrix()));
 }
 
-void test_product_syrk()
+EIGEN_DECLARE_TEST(product_syrk)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/product_trmm.cpp b/test/product_trmm.cpp
index e08d9f39f..c7594e512 100644
--- a/test/product_trmm.cpp
+++ b/test/product_trmm.cpp
@@ -115,7 +115,7 @@ void trmm(int rows=get_random_size<Scalar>(), int cols=get_random_size<Scalar>()
   CALL_ALL_ORDERS(EIGEN_CAT(3,NB),SCALAR,StrictlyLower)
   
 
-void test_product_trmm()
+EIGEN_DECLARE_TEST(product_trmm)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/product_trmv.cpp b/test/product_trmv.cpp
index 65d66e57b..5eb1b5ac0 100644
--- a/test/product_trmv.cpp
+++ b/test/product_trmv.cpp
@@ -70,7 +70,7 @@ template<typename MatrixType> void trmv(const MatrixType& m)
   // TODO check with sub-matrices
 }
 
-void test_product_trmv()
+EIGEN_DECLARE_TEST(product_trmv)
 {
   int s = 0;
   for(int i = 0; i < g_repeat ; i++) {
diff --git a/test/product_trsolve.cpp b/test/product_trsolve.cpp
index 4b97fa9d6..0c22cccf6 100644
--- a/test/product_trsolve.cpp
+++ b/test/product_trsolve.cpp
@@ -73,7 +73,7 @@ template<typename Scalar,int Size, int Cols> void trsolve(int size=Size,int cols
   VERIFY_TRSM(cmLhs.template triangularView<Lower>(), rmRhs.col(c));
 }
 
-void test_product_trsolve()
+EIGEN_DECLARE_TEST(product_trsolve)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/qr.cpp b/test/qr.cpp
index 02e3ed74c..4799aa9ef 100644
--- a/test/qr.cpp
+++ b/test/qr.cpp
@@ -100,7 +100,7 @@ template<typename MatrixType> void qr_verify_assert()
   VERIFY_RAISES_ASSERT(qr.logAbsDeterminant())
 }
 
-void test_qr()
+EIGEN_DECLARE_TEST(qr)
 {
   for(int i = 0; i < g_repeat; i++) {
    CALL_SUBTEST_1( qr(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE),internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
diff --git a/test/qr_colpivoting.cpp b/test/qr_colpivoting.cpp
index 96c0badb7..d224a9436 100644
--- a/test/qr_colpivoting.cpp
+++ b/test/qr_colpivoting.cpp
@@ -296,7 +296,7 @@ template<typename MatrixType> void qr_verify_assert()
   VERIFY_RAISES_ASSERT(qr.logAbsDeterminant())
 }
 
-void test_qr_colpivoting()
+EIGEN_DECLARE_TEST(qr_colpivoting)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( qr<MatrixXf>() );
diff --git a/test/qr_fullpivoting.cpp b/test/qr_fullpivoting.cpp
index c270edf1a..150b4256c 100644
--- a/test/qr_fullpivoting.cpp
+++ b/test/qr_fullpivoting.cpp
@@ -125,7 +125,7 @@ template<typename MatrixType> void qr_verify_assert()
   VERIFY_RAISES_ASSERT(qr.logAbsDeterminant())
 }
 
-void test_qr_fullpivoting()
+EIGEN_DECLARE_TEST(qr_fullpivoting)
 {
   for(int i = 0; i < 1; i++) {
     CALL_SUBTEST_5( qr<Matrix3f>() );
diff --git a/test/qtvector.cpp b/test/qtvector.cpp
index 22df0d515..4ec79b1e6 100644
--- a/test/qtvector.cpp
+++ b/test/qtvector.cpp
@@ -125,7 +125,7 @@ void check_qtvector_quaternion(const QuaternionType&)
   }
 }
 
-void test_qtvector()
+EIGEN_DECLARE_TEST(qtvector)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST(check_qtvector_matrix(Vector2f()));
diff --git a/test/rand.cpp b/test/rand.cpp
index 51cf01773..1b5c058ab 100644
--- a/test/rand.cpp
+++ b/test/rand.cpp
@@ -54,7 +54,7 @@ template<typename Scalar> void check_histogram(Scalar x, Scalar y, int bins)
   VERIFY( (((hist.cast<double>()/double(f))-1.0).abs()<0.02).all() );
 }
 
-void test_rand()
+EIGEN_DECLARE_TEST(rand)
 {
   long long_ref = NumTraits<long>::highest()/10;
   signed char char_offset = (std::min)(g_repeat,64);
diff --git a/test/real_qz.cpp b/test/real_qz.cpp
index 3c1492e4b..1cf7aba2d 100644
--- a/test/real_qz.cpp
+++ b/test/real_qz.cpp
@@ -75,7 +75,7 @@ template<typename MatrixType> void real_qz(const MatrixType& m)
   VERIFY_IS_APPROX(qz.matrixZ()*qz.matrixZ().adjoint(), MatrixType::Identity(dim,dim));
 }
 
-void test_real_qz()
+EIGEN_DECLARE_TEST(real_qz)
 {
   int s = 0;
   for(int i = 0; i < g_repeat; i++) {
diff --git a/test/redux.cpp b/test/redux.cpp
index 213f080aa..9e3ed4546 100644
--- a/test/redux.cpp
+++ b/test/redux.cpp
@@ -146,7 +146,7 @@ template<typename VectorType> void vectorRedux(const VectorType& w)
   VERIFY_RAISES_ASSERT(v.head(0).maxCoeff());
 }
 
-void test_redux()
+EIGEN_DECLARE_TEST(redux)
 {
   // the max size cannot be too large, otherwise reduxion operations obviously generate large errors.
   int maxsize = (std::min)(100,EIGEN_TEST_MAX_SIZE);
diff --git a/test/ref.cpp b/test/ref.cpp
index 704495aff..a94ea9677 100644
--- a/test/ref.cpp
+++ b/test/ref.cpp
@@ -264,7 +264,7 @@ void test_ref_fixed_size_assert()
   VERIFY_RAISES_STATIC_ASSERT( Ref<const Vector3f> y = 2*v4; (void)y; );
 }
 
-void test_ref()
+EIGEN_DECLARE_TEST(ref)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( ref_vector(Matrix<float, 1, 1>()) );
diff --git a/test/resize.cpp b/test/resize.cpp
index 4adaafe56..646a75b8f 100644
--- a/test/resize.cpp
+++ b/test/resize.cpp
@@ -33,7 +33,7 @@ void resizeLikeTest12() { resizeLikeTest<1,2>(); }
 void resizeLikeTest1020() { resizeLikeTest<10,20>(); }
 void resizeLikeTest31() { resizeLikeTest<3,1>(); }
 
-void test_resize()
+EIGEN_DECLARE_TEST(resize)
 {
   CALL_SUBTEST(resizeLikeTest12() );
   CALL_SUBTEST(resizeLikeTest1020() );
diff --git a/test/rvalue_types.cpp b/test/rvalue_types.cpp
index 8887f1b1b..5f52fb3bc 100644
--- a/test/rvalue_types.cpp
+++ b/test/rvalue_types.cpp
@@ -43,7 +43,7 @@ template <typename MatrixType>
 void rvalue_copyassign(const MatrixType&) {}
 #endif
 
-void test_rvalue_types()
+EIGEN_DECLARE_TEST(rvalue_types)
 {
   CALL_SUBTEST_1(rvalue_copyassign( MatrixXf::Random(50,50).eval() ));
   CALL_SUBTEST_1(rvalue_copyassign( ArrayXXf::Random(50,50).eval() ));
diff --git a/test/schur_complex.cpp b/test/schur_complex.cpp
index deb78e44e..03e17e81d 100644
--- a/test/schur_complex.cpp
+++ b/test/schur_complex.cpp
@@ -79,7 +79,7 @@ template<typename MatrixType> void schur(int size = MatrixType::ColsAtCompileTim
   }
 }
 
-void test_schur_complex()
+EIGEN_DECLARE_TEST(schur_complex)
 {
   CALL_SUBTEST_1(( schur<Matrix4cd>() ));
   CALL_SUBTEST_2(( schur<MatrixXcf>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/4)) ));
diff --git a/test/schur_real.cpp b/test/schur_real.cpp
index e5229e6e8..945461027 100644
--- a/test/schur_real.cpp
+++ b/test/schur_real.cpp
@@ -98,7 +98,7 @@ template<typename MatrixType> void schur(int size = MatrixType::ColsAtCompileTim
   }
 }
 
-void test_schur_real()
+EIGEN_DECLARE_TEST(schur_real)
 {
   CALL_SUBTEST_1(( schur<Matrix4f>() ));
   CALL_SUBTEST_2(( schur<MatrixXd>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/4)) ));
diff --git a/test/selfadjoint.cpp b/test/selfadjoint.cpp
index bb11cc351..9ca9cef9e 100644
--- a/test/selfadjoint.cpp
+++ b/test/selfadjoint.cpp
@@ -56,7 +56,7 @@ void bug_159()
   EIGEN_UNUSED_VARIABLE(m)
 }
 
-void test_selfadjoint()
+EIGEN_DECLARE_TEST(selfadjoint)
 {
   for(int i = 0; i < g_repeat ; i++)
   {
diff --git a/test/simplicial_cholesky.cpp b/test/simplicial_cholesky.cpp
index 5bb1e0d6c..314b903e2 100644
--- a/test/simplicial_cholesky.cpp
+++ b/test/simplicial_cholesky.cpp
@@ -39,7 +39,7 @@ template<typename T, typename I> void test_simplicial_cholesky_T()
   check_sparse_spd_solving(ldlt_colmajor_upper_nat, (std::min)(300,EIGEN_TEST_MAX_SIZE), 1000);
 }
 
-void test_simplicial_cholesky()
+EIGEN_DECLARE_TEST(simplicial_cholesky)
 {
   CALL_SUBTEST_1(( test_simplicial_cholesky_T<double,int>() ));
   CALL_SUBTEST_2(( test_simplicial_cholesky_T<std::complex<double>, int>() ));
diff --git a/test/sizeof.cpp b/test/sizeof.cpp
index e552bae7a..af34e97dd 100644
--- a/test/sizeof.cpp
+++ b/test/sizeof.cpp
@@ -18,7 +18,7 @@ template<typename MatrixType> void verifySizeOf(const MatrixType&)
     VERIFY_IS_EQUAL(sizeof(MatrixType),sizeof(Scalar*) + 2 * sizeof(Index));
 }
 
-void test_sizeof()
+EIGEN_DECLARE_TEST(sizeof)
 {
   CALL_SUBTEST(verifySizeOf(Matrix<float, 1, 1>()) );
   CALL_SUBTEST(verifySizeOf(Array<float, 2, 1>()) );
diff --git a/test/sizeoverflow.cpp b/test/sizeoverflow.cpp
index 240d22294..421351233 100644
--- a/test/sizeoverflow.cpp
+++ b/test/sizeoverflow.cpp
@@ -34,7 +34,7 @@ void triggerVectorBadAlloc(Index size)
   VERIFY_THROWS_BADALLOC( VectorType v; v.conservativeResize(size) );
 }
 
-void test_sizeoverflow()
+EIGEN_DECLARE_TEST(sizeoverflow)
 {
   // there are 2 levels of overflow checking. first in PlainObjectBase.h we check for overflow in rows*cols computations.
   // this is tested in tests of the form times_itself_gives_0 * times_itself_gives_0
diff --git a/test/smallvectors.cpp b/test/smallvectors.cpp
index 781511397..f9803acbb 100644
--- a/test/smallvectors.cpp
+++ b/test/smallvectors.cpp
@@ -57,7 +57,7 @@ template<typename Scalar> void smallVectors()
   }
 }
 
-void test_smallvectors()
+EIGEN_DECLARE_TEST(smallvectors)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST(smallVectors<int>() );
diff --git a/test/sparseLM.cpp b/test/sparseLM.cpp
index 8e148f9bc..a48fcb685 100644
--- a/test/sparseLM.cpp
+++ b/test/sparseLM.cpp
@@ -168,7 +168,7 @@ void test_sparseLM_T()
       return ;
 
 }
-void test_sparseLM()
+EIGEN_DECLARE_TEST(sparseLM)
 {
   CALL_SUBTEST_1(test_sparseLM_T<double>());
   
diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp
index f84b6e3f5..bfd61f69f 100644
--- a/test/sparse_basic.cpp
+++ b/test/sparse_basic.cpp
@@ -641,7 +641,7 @@ void big_sparse_triplet(Index rows, Index cols, double density) {
 }
 
 
-void test_sparse_basic()
+EIGEN_DECLARE_TEST(sparse_basic)
 {
   for(int i = 0; i < g_repeat; i++) {
     int r = Eigen::internal::random<int>(1,200), c = Eigen::internal::random<int>(1,200);
diff --git a/test/sparse_block.cpp b/test/sparse_block.cpp
index 2a0b3b617..e7303eab0 100644
--- a/test/sparse_block.cpp
+++ b/test/sparse_block.cpp
@@ -288,7 +288,7 @@ template<typename SparseMatrixType> void sparse_block(const SparseMatrixType& re
   }
 }
 
-void test_sparse_block()
+EIGEN_DECLARE_TEST(sparse_block)
 {
   for(int i = 0; i < g_repeat; i++) {
     int r = Eigen::internal::random<int>(1,200), c = Eigen::internal::random<int>(1,200);
diff --git a/test/sparse_permutations.cpp b/test/sparse_permutations.cpp
index b82cceff8..e93493c39 100644
--- a/test/sparse_permutations.cpp
+++ b/test/sparse_permutations.cpp
@@ -220,7 +220,7 @@ template<typename Scalar> void sparse_permutations_all(int size)
   CALL_SUBTEST(( sparse_permutations<RowMajor>(SparseMatrix<Scalar, RowMajor>(size,size)) ));
 }
 
-void test_sparse_permutations()
+EIGEN_DECLARE_TEST(sparse_permutations)
 {
   for(int i = 0; i < g_repeat; i++) {
     int s = Eigen::internal::random<int>(1,50);
diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp
index f47170b72..d6b7638b1 100644
--- a/test/sparse_product.cpp
+++ b/test/sparse_product.cpp
@@ -453,7 +453,7 @@ void test_mixing_types()
   VERIFY_IS_APPROX( dC2 = sC1 * dR1.col(0), dC3 = sC1 * dR1.template cast<Cplx>().col(0) );
 }
 
-void test_sparse_product()
+EIGEN_DECLARE_TEST(sparse_product)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( (sparse_product<SparseMatrix<double,ColMajor> >()) );
diff --git a/test/sparse_ref.cpp b/test/sparse_ref.cpp
index 5e9607234..12b6f8a9d 100644
--- a/test/sparse_ref.cpp
+++ b/test/sparse_ref.cpp
@@ -126,7 +126,7 @@ void call_ref()
   VERIFY_EVALUATION_COUNT( call_ref_5(A.row(2), A.row(2).transpose()),  1);
 }
 
-void test_sparse_ref()
+EIGEN_DECLARE_TEST(sparse_ref)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( check_const_correctness(SparseMatrix<float>()) );
diff --git a/test/sparse_solvers.cpp b/test/sparse_solvers.cpp
index 3a8873d43..aaf3d39c9 100644
--- a/test/sparse_solvers.cpp
+++ b/test/sparse_solvers.cpp
@@ -101,7 +101,7 @@ template<typename Scalar> void sparse_solvers(int rows, int cols)
   }
 }
 
-void test_sparse_solvers()
+EIGEN_DECLARE_TEST(sparse_solvers)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(sparse_solvers<double>(8, 8) );
diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp
index b3e1dda25..35129278b 100644
--- a/test/sparse_vector.cpp
+++ b/test/sparse_vector.cpp
@@ -145,7 +145,7 @@ template<typename Scalar,typename StorageIndex> void sparse_vector(int rows, int
 
 }
 
-void test_sparse_vector()
+EIGEN_DECLARE_TEST(sparse_vector)
 {
   for(int i = 0; i < g_repeat; i++) {
     int r = Eigen::internal::random<int>(1,500), c = Eigen::internal::random<int>(1,500);
diff --git a/test/sparselu.cpp b/test/sparselu.cpp
index bd000baf1..84cc6ebe5 100644
--- a/test/sparselu.cpp
+++ b/test/sparselu.cpp
@@ -36,7 +36,7 @@ template<typename T> void test_sparselu_T()
   check_sparse_square_determinant(sparselu_amd);
 }
 
-void test_sparselu()
+EIGEN_DECLARE_TEST(sparselu)
 {
   CALL_SUBTEST_1(test_sparselu_T<float>()); 
   CALL_SUBTEST_2(test_sparselu_T<double>());
diff --git a/test/sparseqr.cpp b/test/sparseqr.cpp
index f0e721fce..3ffe62314 100644
--- a/test/sparseqr.cpp
+++ b/test/sparseqr.cpp
@@ -117,7 +117,7 @@ template<typename Scalar> void test_sparseqr_scalar()
   dQ = solver.matrixQ();
   VERIFY_IS_APPROX(Q, dQ);
 }
-void test_sparseqr()
+EIGEN_DECLARE_TEST(sparseqr)
 {
   for(int i=0; i<g_repeat; ++i)
   {
diff --git a/test/special_numbers.cpp b/test/special_numbers.cpp
index 2f1b704be..1e1a63631 100644
--- a/test/special_numbers.cpp
+++ b/test/special_numbers.cpp
@@ -49,7 +49,7 @@ template<typename Scalar> void special_numbers()
   VERIFY(!mboth.array().allFinite());
 }
 
-void test_special_numbers()
+EIGEN_DECLARE_TEST(special_numbers)
 {
   for(int i = 0; i < 10*g_repeat; i++) {
     CALL_SUBTEST_1( special_numbers<float>() );
diff --git a/test/spqr_support.cpp b/test/spqr_support.cpp
index 81e63b6a5..79c2c12fc 100644
--- a/test/spqr_support.cpp
+++ b/test/spqr_support.cpp
@@ -57,7 +57,7 @@ template<typename Scalar> void test_spqr_scalar()
   refX = dA.colPivHouseholderQr().solve(b);
   VERIFY(x.isApprox(refX,test_precision<Scalar>()));
 }
-void test_spqr_support()
+EIGEN_DECLARE_TEST(spqr_support)
 {
   CALL_SUBTEST_1(test_spqr_scalar<double>());
   CALL_SUBTEST_2(test_spqr_scalar<std::complex<double> >());
diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp
index 255cae9b3..ee5f91674 100644
--- a/test/stable_norm.cpp
+++ b/test/stable_norm.cpp
@@ -220,7 +220,7 @@ void test_hypot()
   VERIFY((numext::isnan)(numext::hypot(a,nan)));
 }
 
-void test_stable_norm()
+EIGEN_DECLARE_TEST(stable_norm)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_3( test_hypot<double>() );
diff --git a/test/stddeque.cpp b/test/stddeque.cpp
index b511c4e61..0a6aa2572 100644
--- a/test/stddeque.cpp
+++ b/test/stddeque.cpp
@@ -100,7 +100,7 @@ void check_stddeque_quaternion(const QuaternionType&)
   VERIFY_IS_APPROX(v.back(), x);
 }
 
-void test_stddeque()
+EIGEN_DECLARE_TEST(stddeque)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stddeque_matrix(Vector2f()));
diff --git a/test/stddeque_overload.cpp b/test/stddeque_overload.cpp
index 0541bc1bd..eebe93d81 100644
--- a/test/stddeque_overload.cpp
+++ b/test/stddeque_overload.cpp
@@ -128,7 +128,7 @@ void check_stddeque_quaternion(const QuaternionType&)
   }
 }
 
-void test_stddeque_overload()
+EIGEN_DECLARE_TEST(stddeque_overload)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stddeque_matrix(Vector2f()));
diff --git a/test/stdlist.cpp b/test/stdlist.cpp
index 23cbe9039..c9e5b7286 100644
--- a/test/stdlist.cpp
+++ b/test/stdlist.cpp
@@ -100,7 +100,7 @@ void check_stdlist_quaternion(const QuaternionType&)
   VERIFY_IS_APPROX(v.back(), x);
 }
 
-void test_stdlist()
+EIGEN_DECLARE_TEST(stdlist)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stdlist_matrix(Vector2f()));
diff --git a/test/stdlist_overload.cpp b/test/stdlist_overload.cpp
index 386124ad4..93b6fc9ed 100644
--- a/test/stdlist_overload.cpp
+++ b/test/stdlist_overload.cpp
@@ -162,7 +162,7 @@ void check_stdlist_quaternion(const QuaternionType&)
   }
 }
 
-void test_stdlist_overload()
+EIGEN_DECLARE_TEST(stdlist_overload)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stdlist_matrix(Vector2f()));
diff --git a/test/stdvector.cpp b/test/stdvector.cpp
index cf3646870..e2b7bd061 100644
--- a/test/stdvector.cpp
+++ b/test/stdvector.cpp
@@ -117,7 +117,7 @@ void check_stdvector_quaternion(const QuaternionType&)
   }
 }
 
-void test_stdvector()
+EIGEN_DECLARE_TEST(stdvector)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stdvector_matrix(Vector2f()));
diff --git a/test/stdvector_overload.cpp b/test/stdvector_overload.cpp
index 0c86ce8a4..5c042a64c 100644
--- a/test/stdvector_overload.cpp
+++ b/test/stdvector_overload.cpp
@@ -131,7 +131,7 @@ void check_stdvector_quaternion(const QuaternionType&)
   }
 }
 
-void test_stdvector_overload()
+EIGEN_DECLARE_TEST(stdvector_overload)
 {
   // some non vectorizable fixed sizes
   CALL_SUBTEST_1(check_stdvector_matrix(Vector2f()));
diff --git a/test/superlu_support.cpp b/test/superlu_support.cpp
index 98a7bc5c8..55450c868 100644
--- a/test/superlu_support.cpp
+++ b/test/superlu_support.cpp
@@ -12,7 +12,7 @@
 
 #include <Eigen/SuperLUSupport>
 
-void test_superlu_support()
+EIGEN_DECLARE_TEST(superlu_support)
 {
   SuperLU<SparseMatrix<double> > superlu_double_colmajor;
   SuperLU<SparseMatrix<std::complex<double> > > superlu_cplxdouble_colmajor;
diff --git a/test/swap.cpp b/test/swap.cpp
index a294da719..5b259d3ec 100644
--- a/test/swap.cpp
+++ b/test/swap.cpp
@@ -83,7 +83,7 @@ template<typename MatrixType> void swap(const MatrixType& m)
   }
 }
 
-void test_swap()
+EIGEN_DECLARE_TEST(swap)
 {
   int s = internal::random<int>(1,EIGEN_TEST_MAX_SIZE);
   CALL_SUBTEST_1( swap(Matrix3f()) ); // fixed size, no vectorization 
diff --git a/test/symbolic_index.cpp b/test/symbolic_index.cpp
index 4c079857b..cf89b9fbe 100644
--- a/test/symbolic_index.cpp
+++ b/test/symbolic_index.cpp
@@ -113,7 +113,7 @@ void check_symbolic_index()
 #endif
 }
 
-void test_symbolic_index()
+EIGEN_DECLARE_TEST(symbolic_index)
 {
   CALL_SUBTEST_1( check_symbolic_index() );
   CALL_SUBTEST_2( check_symbolic_index() );
diff --git a/test/triangular.cpp b/test/triangular.cpp
index d50342ccc..73b6fc46b 100644
--- a/test/triangular.cpp
+++ b/test/triangular.cpp
@@ -220,7 +220,7 @@ void bug_159()
   EIGEN_UNUSED_VARIABLE(m)
 }
 
-void test_triangular()
+EIGEN_DECLARE_TEST(triangular)
 {
   int maxsize = (std::min)(EIGEN_TEST_MAX_SIZE,20);
   for(int i = 0; i < g_repeat ; i++)
diff --git a/test/umeyama.cpp b/test/umeyama.cpp
index 2e8092434..1590a0a81 100644
--- a/test/umeyama.cpp
+++ b/test/umeyama.cpp
@@ -155,7 +155,7 @@ void run_fixed_size_test(int num_elements)
   VERIFY(error < Scalar(16)*std::numeric_limits<Scalar>::epsilon());
 }
 
-void test_umeyama()
+EIGEN_DECLARE_TEST(umeyama)
 {
   for (int i=0; i<g_repeat; ++i)
   {
diff --git a/test/umfpack_support.cpp b/test/umfpack_support.cpp
index 0378801f4..d8f2a6f80 100644
--- a/test/umfpack_support.cpp
+++ b/test/umfpack_support.cpp
@@ -24,7 +24,7 @@ template<typename T1, typename T2> void test_umfpack_support_T()
   check_sparse_square_determinant(umfpack_rowmajor);
 }
 
-void test_umfpack_support()
+EIGEN_DECLARE_TEST(umfpack_support)
 {
   CALL_SUBTEST_1((test_umfpack_support_T<double, int>()));
   CALL_SUBTEST_2((test_umfpack_support_T<std::complex<double>, int>()));
diff --git a/test/unalignedassert.cpp b/test/unalignedassert.cpp
index 731a08977..120cc42bb 100644
--- a/test/unalignedassert.cpp
+++ b/test/unalignedassert.cpp
@@ -174,7 +174,7 @@ void unalignedassert()
 #endif
 }
 
-void test_unalignedassert()
+EIGEN_DECLARE_TEST(unalignedassert)
 {
   CALL_SUBTEST(unalignedassert());
 }
diff --git a/test/unalignedcount.cpp b/test/unalignedcount.cpp
index d6ffeafdf..069fc1bb9 100644
--- a/test/unalignedcount.cpp
+++ b/test/unalignedcount.cpp
@@ -28,7 +28,7 @@ static int nb_storeu;
 
 #include "main.h"
 
-void test_unalignedcount()
+EIGEN_DECLARE_TEST(unalignedcount)
 {
   #if defined(EIGEN_VECTORIZE_AVX)
   VectorXf a(40), b(40);
diff --git a/test/upperbidiagonalization.cpp b/test/upperbidiagonalization.cpp
index 27a747875..945c99959 100644
--- a/test/upperbidiagonalization.cpp
+++ b/test/upperbidiagonalization.cpp
@@ -29,7 +29,7 @@ template<typename MatrixType> void upperbidiag(const MatrixType& m)
   VERIFY_IS_APPROX(a.adjoint(),d);
 }
 
-void test_upperbidiagonalization()
+EIGEN_DECLARE_TEST(upperbidiagonalization)
 {
   for(int i = 0; i < g_repeat; i++) {
    CALL_SUBTEST_1( upperbidiag(MatrixXf(3,3)) );
diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp
index 37e7495f5..01b55e192 100644
--- a/test/vectorization_logic.cpp
+++ b/test/vectorization_logic.cpp
@@ -386,7 +386,7 @@ template<typename Scalar> struct vectorization_logic_half<Scalar,false>
   static void run() {}
 };
 
-void test_vectorization_logic()
+EIGEN_DECLARE_TEST(vectorization_logic)
 {
 
 #ifdef EIGEN_VECTORIZE
diff --git a/test/vectorwiseop.cpp b/test/vectorwiseop.cpp
index a099d17c8..2d7ddbed1 100644
--- a/test/vectorwiseop.cpp
+++ b/test/vectorwiseop.cpp
@@ -237,7 +237,7 @@ template<typename MatrixType> void vectorwiseop_matrix(const MatrixType& m)
   VERIFY_EVALUATION_COUNT( m2 = (m1.rowwise() - m1.colwise().sum()/RealScalar(m1.rows())), (MatrixType::RowsAtCompileTime!=1 ? 1 : 0) );
 }
 
-void test_vectorwiseop()
+EIGEN_DECLARE_TEST(vectorwiseop)
 {
   CALL_SUBTEST_1( vectorwiseop_array(Array22cd()) );
   CALL_SUBTEST_2( vectorwiseop_array(Array<double, 3, 2>()) );
diff --git a/test/visitor.cpp b/test/visitor.cpp
index 7f4efab97..de938fc95 100644
--- a/test/visitor.cpp
+++ b/test/visitor.cpp
@@ -113,7 +113,7 @@ template<typename VectorType> void vectorVisitor(const VectorType& w)
   VERIFY(eigen_maxidx == (std::min)(idx0,idx2));
 }
 
-void test_visitor()
+EIGEN_DECLARE_TEST(visitor)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( matrixVisitor(Matrix<float, 1, 1>()) );
diff --git a/test/zerosized.cpp b/test/zerosized.cpp
index 477ff0070..edd1f6925 100644
--- a/test/zerosized.cpp
+++ b/test/zerosized.cpp
@@ -81,7 +81,7 @@ template<typename VectorType> void zeroSizedVector()
   }
 }
 
-void test_zerosized()
+EIGEN_DECLARE_TEST(zerosized)
 {
   zeroSizedMatrix<Matrix2d>();
   zeroSizedMatrix<Matrix3i>();
diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
index 5f64501be..62561da1d 100644
--- a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
+++ b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
@@ -117,7 +117,7 @@ class LevenbergMarquardt : internal::no_assignment_operator
     typedef typename JacobianType::RealScalar RealScalar; 
     typedef typename QRSolver::StorageIndex PermIndex;
     typedef Matrix<Scalar,Dynamic,1> FVectorType;
-    typedef PermutationMatrix<Dynamic,Dynamic> PermutationType;
+    typedef PermutationMatrix<Dynamic,Dynamic,int> PermutationType;
   public:
     LevenbergMarquardt(FunctorType& functor) 
     : m_functor(functor),m_nfev(0),m_njev(0),m_fnorm(0.0),m_gnorm(0),
diff --git a/unsupported/test/BVH.cpp b/unsupported/test/BVH.cpp
index ff5b3299d..d8c39d556 100644
--- a/unsupported/test/BVH.cpp
+++ b/unsupported/test/BVH.cpp
@@ -192,7 +192,7 @@ struct TreeTest
 };
 
 
-void test_BVH()
+EIGEN_DECLARE_TEST(BVH)
 {
   for(int i = 0; i < g_repeat; i++) {
 #ifdef EIGEN_TEST_PART_1
diff --git a/unsupported/test/EulerAngles.cpp b/unsupported/test/EulerAngles.cpp
index 500fb2d17..572fc08a3 100644
--- a/unsupported/test/EulerAngles.cpp
+++ b/unsupported/test/EulerAngles.cpp
@@ -272,7 +272,7 @@ template<typename Scalar> void eulerangles_rand()
   check_all_var(ea);
 }
 
-void test_EulerAngles()
+EIGEN_DECLARE_TEST(EulerAngles)
 {
   // Simple cast test
   EulerAnglesXYZd onesEd(1, 1, 1);
diff --git a/unsupported/test/FFTW.cpp b/unsupported/test/FFTW.cpp
index 8b7528fb7..cfe559ebd 100644
--- a/unsupported/test/FFTW.cpp
+++ b/unsupported/test/FFTW.cpp
@@ -225,7 +225,7 @@ void test_return_by_value(int len)
     VERIFY( (in1-in).norm() < test_precision<float>() );
 }
 
-void test_FFTW()
+EIGEN_DECLARE_TEST(FFTW)
 {
   CALL_SUBTEST( test_return_by_value(32) );
   //CALL_SUBTEST( ( test_complex2d<float,4,8> () ) ); CALL_SUBTEST( ( test_complex2d<double,4,8> () ) );
diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp
index 1d682dd83..cc95cea29 100644
--- a/unsupported/test/NonLinearOptimization.cpp
+++ b/unsupported/test/NonLinearOptimization.cpp
@@ -1818,7 +1818,7 @@ void testNistEckerle4(void)
   VERIFY_IS_APPROX(x[2], 4.5154121844E+02);
 }
 
-void test_NonLinearOptimization()
+EIGEN_DECLARE_TEST(NonLinearOptimization)
 {
     // Tests using the examples provided by (c)minpack
     CALL_SUBTEST/*_1*/(testChkder());
diff --git a/unsupported/test/NumericalDiff.cpp b/unsupported/test/NumericalDiff.cpp
index 27d888056..35f2f6d7c 100644
--- a/unsupported/test/NumericalDiff.cpp
+++ b/unsupported/test/NumericalDiff.cpp
@@ -107,7 +107,7 @@ void test_central()
     VERIFY_IS_APPROX(jac, actual_jac);
 }
 
-void test_NumericalDiff()
+EIGEN_DECLARE_TEST(NumericalDiff)
 {
     CALL_SUBTEST(test_forward());
     CALL_SUBTEST(test_central());
diff --git a/unsupported/test/alignedvector3.cpp b/unsupported/test/alignedvector3.cpp
index 252cb1d3f..fcc89daab 100644
--- a/unsupported/test/alignedvector3.cpp
+++ b/unsupported/test/alignedvector3.cpp
@@ -76,7 +76,7 @@ void alignedvector3()
   VERIFY(ss1.str()==ss2.str());
 }
 
-void test_alignedvector3()
+EIGEN_DECLARE_TEST(alignedvector3)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST( alignedvector3<float>() );
diff --git a/unsupported/test/autodiff.cpp b/unsupported/test/autodiff.cpp
index 1c5e0dc66..495cee03e 100644
--- a/unsupported/test/autodiff.cpp
+++ b/unsupported/test/autodiff.cpp
@@ -354,7 +354,7 @@ double bug_1264() {
 
 #endif
 
-void test_autodiff()
+EIGEN_DECLARE_TEST(autodiff)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( test_autodiff_scalar<1>() );
diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp
index 1d4418203..e81a7788b 100644
--- a/unsupported/test/autodiff_scalar.cpp
+++ b/unsupported/test/autodiff_scalar.cpp
@@ -89,7 +89,7 @@ void check_limits_specialization()
 #endif
 }
 
-void test_autodiff_scalar()
+EIGEN_DECLARE_TEST(autodiff_scalar)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( check_atan2<float>() );
diff --git a/unsupported/test/cxx11_eventcount.cpp b/unsupported/test/cxx11_eventcount.cpp
index 3b598bf42..2f1418684 100644
--- a/unsupported/test/cxx11_eventcount.cpp
+++ b/unsupported/test/cxx11_eventcount.cpp
@@ -135,7 +135,7 @@ static void test_stress_eventcount()
   }
 }
 
-void test_cxx11_eventcount()
+EIGEN_DECLARE_TEST(cxx11_eventcount)
 {
   CALL_SUBTEST(test_basic_eventcount());
   CALL_SUBTEST(test_stress_eventcount());
diff --git a/unsupported/test/cxx11_meta.cpp b/unsupported/test/cxx11_meta.cpp
index 8911c59d8..510e11032 100644
--- a/unsupported/test/cxx11_meta.cpp
+++ b/unsupported/test/cxx11_meta.cpp
@@ -340,7 +340,7 @@ static void test_array_misc()
   VERIFY_IS_EQUAL((instantiate_by_c_array<dummy_inst, int, 5>(data).c), 5);
 }
 
-void test_cxx11_meta()
+EIGEN_DECLARE_TEST(cxx11_meta)
 {
   CALL_SUBTEST(test_gen_numeric_list());
   CALL_SUBTEST(test_concat());
diff --git a/unsupported/test/cxx11_non_blocking_thread_pool.cpp b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
index 48cd2d4e4..5a8080ea3 100644
--- a/unsupported/test/cxx11_non_blocking_thread_pool.cpp
+++ b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
@@ -116,7 +116,7 @@ static void test_cancel()
   tp.Cancel();
 }
 
-void test_cxx11_non_blocking_thread_pool()
+EIGEN_DECLARE_TEST(cxx11_non_blocking_thread_pool)
 {
   CALL_SUBTEST(test_create_destroy_empty_pool());
   CALL_SUBTEST(test_parallelism(true));
diff --git a/unsupported/test/cxx11_runqueue.cpp b/unsupported/test/cxx11_runqueue.cpp
index 91f690114..8fc5a3074 100644
--- a/unsupported/test/cxx11_runqueue.cpp
+++ b/unsupported/test/cxx11_runqueue.cpp
@@ -227,7 +227,7 @@ void test_stress_runqueue()
   VERIFY(total.load() == 0);
 }
 
-void test_cxx11_runqueue()
+EIGEN_DECLARE_TEST(cxx11_runqueue)
 {
   CALL_SUBTEST_1(test_basic_runqueue());
   CALL_SUBTEST_2(test_empty_runqueue());
diff --git a/unsupported/test/cxx11_tensor_argmax.cpp b/unsupported/test/cxx11_tensor_argmax.cpp
index 037767270..4a0c8967b 100644
--- a/unsupported/test/cxx11_tensor_argmax.cpp
+++ b/unsupported/test/cxx11_tensor_argmax.cpp
@@ -273,7 +273,7 @@ static void test_argmin_dim()
   }
 }
 
-void test_cxx11_tensor_argmax()
+EIGEN_DECLARE_TEST(cxx11_tensor_argmax)
 {
   CALL_SUBTEST(test_simple_index_tuples<RowMajor>());
   CALL_SUBTEST(test_simple_index_tuples<ColMajor>());
diff --git a/unsupported/test/cxx11_tensor_argmax_gpu.cu b/unsupported/test/cxx11_tensor_argmax_gpu.cu
index 541a27865..f6c3a9908 100644
--- a/unsupported/test/cxx11_tensor_argmax_gpu.cu
+++ b/unsupported/test/cxx11_tensor_argmax_gpu.cu
@@ -9,7 +9,7 @@
 
 
 #define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_FUNC cxx11_tensor_gpu
+
 #define EIGEN_USE_GPU
 
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_argmax_sycl.cpp b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
index 521a7f82c..0bbb0f6dc 100644
--- a/unsupported/test/cxx11_tensor_argmax_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_argmax_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_argmax_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -237,7 +237,7 @@ template<typename DataType, typename Device_Selector> void sycl_argmax_test_per_
   test_sycl_argmin_dim<DataType, RowMajor, int64_t>(sycl_device);
 }
 
-void test_cxx11_tensor_argmax_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_argmax_sycl) {
  for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_argmax_test_per_device<double>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp
index 8fe85d83c..ce9d24369 100644
--- a/unsupported/test/cxx11_tensor_assign.cpp
+++ b/unsupported/test/cxx11_tensor_assign.cpp
@@ -358,7 +358,7 @@ static void test_std_initializers_tensor() {
 #endif  // EIGEN_HAS_VARIADIC_TEMPLATES
 }
 
-void test_cxx11_tensor_assign()
+EIGEN_DECLARE_TEST(cxx11_tensor_assign)
 {
   CALL_SUBTEST(test_1d());
   CALL_SUBTEST(test_2d());
diff --git a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
index 21fdfca22..20f84b8e0 100644
--- a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_broadcast_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -137,7 +137,7 @@ template<typename DataType> void sycl_broadcast_test_per_device(const cl::sycl::
   test_broadcast_sycl_fixed<DataType, ColMajor, int64_t>(sycl_device);
 }
 
-void test_cxx11_tensor_broadcast_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_broadcast_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_broadcast_test_per_device<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp
index f0ff03184..2f8ab6afd 100644
--- a/unsupported/test/cxx11_tensor_broadcasting.cpp
+++ b/unsupported/test/cxx11_tensor_broadcasting.cpp
@@ -292,7 +292,7 @@ static void test_simple_broadcasting_one_by_n_by_one_2d()
   }
 }
 
-void test_cxx11_tensor_broadcasting()
+EIGEN_DECLARE_TEST(cxx11_tensor_broadcasting)
 {
   CALL_SUBTEST(test_simple_broadcasting<ColMajor>());
   CALL_SUBTEST(test_simple_broadcasting<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_builtins_sycl.cpp b/unsupported/test/cxx11_tensor_builtins_sycl.cpp
index 400a31d09..db2975783 100644
--- a/unsupported/test/cxx11_tensor_builtins_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_builtins_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_builtins_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -257,7 +257,7 @@ static void test_builtin_binary_sycl(const Eigen::SyclDevice &sycl_device) {
   TEST_BINARY_BUILTINS_OPERATORS_THAT_TAKES_SCALAR(int, %, ColMajor)
 }
 
-void test_cxx11_tensor_builtins_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_builtins_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     QueueInterface queueInterface(device);
     Eigen::SyclDevice sycl_device(&queueInterface);
diff --git a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
index a2928b0b3..0a37a555c 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_gpu
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_casts.cpp b/unsupported/test/cxx11_tensor_casts.cpp
index 3c6d0d2ff..c4fe9a798 100644
--- a/unsupported/test/cxx11_tensor_casts.cpp
+++ b/unsupported/test/cxx11_tensor_casts.cpp
@@ -105,7 +105,7 @@ static void test_small_to_big_type_cast()
 }
 
 
-void test_cxx11_tensor_casts()
+EIGEN_DECLARE_TEST(cxx11_tensor_casts)
 {
    CALL_SUBTEST(test_simple_cast());
    CALL_SUBTEST(test_vectorized_cast());
diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp
index 89cf5c7b7..922274462 100644
--- a/unsupported/test/cxx11_tensor_chipping.cpp
+++ b/unsupported/test/cxx11_tensor_chipping.cpp
@@ -410,7 +410,7 @@ static void test_chip_raw_data_row_major()
   VERIFY_IS_EQUAL(chip4.data(), static_cast<float*>(0));
 }
 
-void test_cxx11_tensor_chipping()
+EIGEN_DECLARE_TEST(cxx11_tensor_chipping)
 {
   CALL_SUBTEST(test_simple_chip<ColMajor>());
   CALL_SUBTEST(test_simple_chip<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_chipping_sycl.cpp b/unsupported/test/cxx11_tensor_chipping_sycl.cpp
index 39e4f0a7f..a91efe00c 100644
--- a/unsupported/test/cxx11_tensor_chipping_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_chipping_sycl.cpp
@@ -15,7 +15,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_chipping_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -614,7 +614,7 @@ template<typename DataType, typename dev_Selector> void sycl_chipping_test_per_d
   test_chip_as_lvalue_sycl<DataType, RowMajor, int64_t>(sycl_device);
   test_chip_as_lvalue_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_chipping_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_chipping_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_chipping_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_comparisons.cpp b/unsupported/test/cxx11_tensor_comparisons.cpp
index b1ff8aecb..1a18e07cc 100644
--- a/unsupported/test/cxx11_tensor_comparisons.cpp
+++ b/unsupported/test/cxx11_tensor_comparisons.cpp
@@ -77,7 +77,7 @@ static void test_equality()
 }
 
 
-void test_cxx11_tensor_comparisons()
+EIGEN_DECLARE_TEST(cxx11_tensor_comparisons)
 {
   CALL_SUBTEST(test_orderings());
   CALL_SUBTEST(test_equality());
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
index 3b078d1d1..aa28457b1 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
@@ -8,7 +8,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_FUNC cxx11_tensor_complex_cwise_ops
+
 #define EIGEN_USE_GPU
 
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_complex_gpu.cu b/unsupported/test/cxx11_tensor_complex_gpu.cu
index 45b49d266..7cf06aa7a 100644
--- a/unsupported/test/cxx11_tensor_complex_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_gpu.cu
@@ -8,7 +8,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_FUNC cxx11_tensor_complex
+
 #define EIGEN_USE_GPU
 
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp
index 03ef12e63..9189a609b 100644
--- a/unsupported/test/cxx11_tensor_concatenation.cpp
+++ b/unsupported/test/cxx11_tensor_concatenation.cpp
@@ -123,7 +123,7 @@ static void test_concatenation_as_lvalue()
 }
 
 
-void test_cxx11_tensor_concatenation()
+EIGEN_DECLARE_TEST(cxx11_tensor_concatenation)
 {
    CALL_SUBTEST(test_dimension_failures<ColMajor>());
    CALL_SUBTEST(test_dimension_failures<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_concatenation_sycl.cpp b/unsupported/test/cxx11_tensor_concatenation_sycl.cpp
index e3023a368..765991b35 100644
--- a/unsupported/test/cxx11_tensor_concatenation_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_concatenation_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_concatenation_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -173,7 +173,7 @@ template <typename DataType, typename Dev_selector> void tensorConcat_perDevice(
   test_simple_concatenation<DataType, ColMajor, int64_t>(sycl_device);
   test_concatenation_as_lvalue<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_concatenation_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_concatenation_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(tensorConcat_perDevice<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_const.cpp b/unsupported/test/cxx11_tensor_const.cpp
index ad9c9da39..9d806ee3c 100644
--- a/unsupported/test/cxx11_tensor_const.cpp
+++ b/unsupported/test/cxx11_tensor_const.cpp
@@ -55,7 +55,7 @@ static void test_assign_of_const_tensor()
 }
 
 
-void test_cxx11_tensor_const()
+EIGEN_DECLARE_TEST(cxx11_tensor_const)
 {
   CALL_SUBTEST(test_simple_assign());
   CALL_SUBTEST(test_assign_of_const_tensor());
diff --git a/unsupported/test/cxx11_tensor_contract_gpu.cu b/unsupported/test/cxx11_tensor_contract_gpu.cu
index 061d0464e..cb1416478 100644
--- a/unsupported/test/cxx11_tensor_contract_gpu.cu
+++ b/unsupported/test/cxx11_tensor_contract_gpu.cu
@@ -10,7 +10,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_gpu
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_contract_sycl.cpp b/unsupported/test/cxx11_tensor_contract_sycl.cpp
index 5bace66c5..c8e86e69f 100644
--- a/unsupported/test/cxx11_tensor_contract_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_contract_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_contract_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -283,7 +283,7 @@ template <typename Dev_selector> void tensorContractionPerDevice(Dev_selector& s
 
 }
 
-void test_cxx11_tensor_contract_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_contract_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(tensorContractionPerDevice(device));
   }
diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp
index 918c96277..d4cfbd0da 100644
--- a/unsupported/test/cxx11_tensor_contraction.cpp
+++ b/unsupported/test/cxx11_tensor_contraction.cpp
@@ -559,7 +559,7 @@ static void test_large_contraction_with_output_kernel() {
   }
 }
 
-void test_cxx11_tensor_contraction()
+EIGEN_DECLARE_TEST(cxx11_tensor_contraction)
 {
   CALL_SUBTEST(test_evals<ColMajor>());
   CALL_SUBTEST(test_evals<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp
index e3d4675eb..01bc77bc1 100644
--- a/unsupported/test/cxx11_tensor_convolution.cpp
+++ b/unsupported/test/cxx11_tensor_convolution.cpp
@@ -136,7 +136,7 @@ static void test_strides() {
                                input(12)*kernel(2)));
 }
 
-void test_cxx11_tensor_convolution()
+EIGEN_DECLARE_TEST(cxx11_tensor_convolution)
 {
   CALL_SUBTEST(test_evals<ColMajor>());
   CALL_SUBTEST(test_evals<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_convolution_sycl.cpp b/unsupported/test/cxx11_tensor_convolution_sycl.cpp
index a4226a63a..3954c8a28 100644
--- a/unsupported/test/cxx11_tensor_convolution_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_convolution_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_convolution_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -462,7 +462,7 @@ template <typename Dev_selector> void tensorConvolutionPerDevice(Dev_selector& s
   test_strides<float, RowMajor, int64_t>(sycl_device);
 }
 
-void test_cxx11_tensor_convolution_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_convolution_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(tensorConvolutionPerDevice(device));
   }
diff --git a/unsupported/test/cxx11_tensor_custom_index.cpp b/unsupported/test/cxx11_tensor_custom_index.cpp
index 4528cc176..b5dbc97bd 100644
--- a/unsupported/test/cxx11_tensor_custom_index.cpp
+++ b/unsupported/test/cxx11_tensor_custom_index.cpp
@@ -88,7 +88,7 @@ static void test_sizes_as_index()
 }
 
 
-void test_cxx11_tensor_custom_index() {
+EIGEN_DECLARE_TEST(cxx11_tensor_custom_index) {
   test_map_as_index<ColMajor>();
   test_map_as_index<RowMajor>();
   test_matrix_as_index<ColMajor>();
diff --git a/unsupported/test/cxx11_tensor_custom_op.cpp b/unsupported/test/cxx11_tensor_custom_op.cpp
index 8baa477cc..875ea57d2 100644
--- a/unsupported/test/cxx11_tensor_custom_op.cpp
+++ b/unsupported/test/cxx11_tensor_custom_op.cpp
@@ -104,7 +104,7 @@ static void test_custom_binary_op()
 }
 
 
-void test_cxx11_tensor_custom_op()
+EIGEN_DECLARE_TEST(cxx11_tensor_custom_op)
 {
   CALL_SUBTEST(test_custom_unary_op());
   CALL_SUBTEST(test_custom_binary_op());
diff --git a/unsupported/test/cxx11_tensor_custom_op_sycl.cpp b/unsupported/test/cxx11_tensor_custom_op_sycl.cpp
index 9ff287fff..cc3b02448 100644
--- a/unsupported/test/cxx11_tensor_custom_op_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_custom_op_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_custom_op_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -158,7 +158,7 @@ template <typename DataType, typename Dev_selector> void custom_op_perDevice(Dev
   test_custom_binary_op_sycl<DataType, RowMajor, int64_t>(sycl_device);
 
 }
-void test_cxx11_tensor_custom_op_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_custom_op_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(custom_op_perDevice<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index 52215fc39..cd9ba3ecd 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_device
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_device_sycl.cpp b/unsupported/test/cxx11_tensor_device_sycl.cpp
index 3ecc68df0..5095cb078 100644
--- a/unsupported/test/cxx11_tensor_device_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_device_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_device_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -70,7 +70,7 @@ template<typename DataType> void sycl_device_test_per_device(const cl::sycl::dev
   //test_device_exceptions<DataType, ColMajor>(sycl_device);
 }
 
-void test_cxx11_tensor_device_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_device_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_device_test_per_device<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_dimension.cpp b/unsupported/test/cxx11_tensor_dimension.cpp
index 16f168ed4..10364d4b4 100644
--- a/unsupported/test/cxx11_tensor_dimension.cpp
+++ b/unsupported/test/cxx11_tensor_dimension.cpp
@@ -60,7 +60,7 @@ static void test_rank_zero()
   VERIFY_IS_EQUAL((int)dscalar.rank(), 0);
 }
 
-void test_cxx11_tensor_dimension()
+EIGEN_DECLARE_TEST(cxx11_tensor_dimension)
 {
   CALL_SUBTEST(test_dynamic_size());
   CALL_SUBTEST(test_fixed_size());
diff --git a/unsupported/test/cxx11_tensor_empty.cpp b/unsupported/test/cxx11_tensor_empty.cpp
index d7eea42d7..fd889c46c 100644
--- a/unsupported/test/cxx11_tensor_empty.cpp
+++ b/unsupported/test/cxx11_tensor_empty.cpp
@@ -33,7 +33,7 @@ static void test_empty_fixed_size_tensor()
 }
 
 
-void test_cxx11_tensor_empty()
+EIGEN_DECLARE_TEST(cxx11_tensor_empty)
 {
    CALL_SUBTEST(test_empty_tensor());
    CALL_SUBTEST(test_empty_fixed_size_tensor());
diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp
index d64301318..30924b6b6 100644
--- a/unsupported/test/cxx11_tensor_expr.cpp
+++ b/unsupported/test/cxx11_tensor_expr.cpp
@@ -366,7 +366,7 @@ static void test_minmax_nan_propagation()
   test_minmax_nan_propagation_templ<double>();
 }
 
-void test_cxx11_tensor_expr()
+EIGEN_DECLARE_TEST(cxx11_tensor_expr)
 {
   CALL_SUBTEST(test_1d());
   CALL_SUBTEST(test_2d());
diff --git a/unsupported/test/cxx11_tensor_fft.cpp b/unsupported/test/cxx11_tensor_fft.cpp
index ef5f1a312..4e4c9c4ec 100644
--- a/unsupported/test/cxx11_tensor_fft.cpp
+++ b/unsupported/test/cxx11_tensor_fft.cpp
@@ -253,7 +253,7 @@ static void test_fft_non_power_of_2_round_trip(int exponent) {
   }
 }
 
-void test_cxx11_tensor_fft() {
+EIGEN_DECLARE_TEST(cxx11_tensor_fft) {
     test_fft_complex_input_golden();
     test_fft_real_input_golden();
 
diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp
index e6274f8eb..456ce6bea 100644
--- a/unsupported/test/cxx11_tensor_fixed_size.cpp
+++ b/unsupported/test/cxx11_tensor_fixed_size.cpp
@@ -250,7 +250,7 @@ static void test_array()
   }
 }
 
-void test_cxx11_tensor_fixed_size()
+EIGEN_DECLARE_TEST(cxx11_tensor_fixed_size)
 {
   CALL_SUBTEST(test_0d());
   CALL_SUBTEST(test_1d());
diff --git a/unsupported/test/cxx11_tensor_forced_eval.cpp b/unsupported/test/cxx11_tensor_forced_eval.cpp
index 45d7345e9..f76e2ea97 100644
--- a/unsupported/test/cxx11_tensor_forced_eval.cpp
+++ b/unsupported/test/cxx11_tensor_forced_eval.cpp
@@ -72,7 +72,7 @@ static void test_const()
 }
 
 
-void test_cxx11_tensor_forced_eval()
+EIGEN_DECLARE_TEST(cxx11_tensor_forced_eval)
 {
   CALL_SUBTEST(test_simple());
   CALL_SUBTEST(test_const());
diff --git a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
index a21514d56..74d38a644 100644
--- a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_forced_eval_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -69,7 +69,7 @@ template <typename DataType, typename Dev_selector> void tensorForced_evalperDev
   test_forced_eval_sycl<DataType, RowMajor, int64_t>(sycl_device);
   test_forced_eval_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_forced_eval_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_forced_eval_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(tensorForced_evalperDevice<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_generator.cpp b/unsupported/test/cxx11_tensor_generator.cpp
index dcb928714..ee5e29b77 100644
--- a/unsupported/test/cxx11_tensor_generator.cpp
+++ b/unsupported/test/cxx11_tensor_generator.cpp
@@ -80,7 +80,7 @@ static void test_gaussian()
 }
 
 
-void test_cxx11_tensor_generator()
+EIGEN_DECLARE_TEST(cxx11_tensor_generator)
 {
   CALL_SUBTEST(test_1D<ColMajor>());
   CALL_SUBTEST(test_1D<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_generator_sycl.cpp b/unsupported/test/cxx11_tensor_generator_sycl.cpp
index f551c8d0c..fb6e3d9d0 100644
--- a/unsupported/test/cxx11_tensor_generator_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_generator_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_generator_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 static const float error_threshold =1e-8f;
@@ -139,7 +139,7 @@ template<typename DataType, typename dev_Selector> void sycl_generator_test_per_
   test_gaussian_sycl<DataType, RowMajor, int64_t>(sycl_device);
   test_gaussian_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_generator_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_generator_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_generator_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_gpu.cu b/unsupported/test/cxx11_tensor_gpu.cu
index 285441182..faaac73cf 100644
--- a/unsupported/test/cxx11_tensor_gpu.cu
+++ b/unsupported/test/cxx11_tensor_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_gpu
+
 #define EIGEN_USE_GPU
 
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_ifft.cpp b/unsupported/test/cxx11_tensor_ifft.cpp
index 5fd88fa6c..c20edd9ac 100644
--- a/unsupported/test/cxx11_tensor_ifft.cpp
+++ b/unsupported/test/cxx11_tensor_ifft.cpp
@@ -131,7 +131,7 @@ static void test_sub_fft_ifft_invariant(int dim0, int dim1, int dim2, int dim3)
   }
 }
 
-void test_cxx11_tensor_ifft() {
+EIGEN_DECLARE_TEST(cxx11_tensor_ifft) {
   CALL_SUBTEST(test_1D_fft_ifft_invariant<ColMajor>(4));
   CALL_SUBTEST(test_1D_fft_ifft_invariant<ColMajor>(16));
   CALL_SUBTEST(test_1D_fft_ifft_invariant<ColMajor>(32));
diff --git a/unsupported/test/cxx11_tensor_image_patch.cpp b/unsupported/test/cxx11_tensor_image_patch.cpp
index 105d32fb4..862f1f7f0 100644
--- a/unsupported/test/cxx11_tensor_image_patch.cpp
+++ b/unsupported/test/cxx11_tensor_image_patch.cpp
@@ -797,7 +797,7 @@ void test_imagenet_patches()
   }
 }
 
-void test_cxx11_tensor_image_patch()
+EIGEN_DECLARE_TEST(cxx11_tensor_image_patch)
 {
   CALL_SUBTEST_1(test_simple_patch());
   CALL_SUBTEST_2(test_patch_no_extra_dim());
diff --git a/unsupported/test/cxx11_tensor_image_patch_sycl.cpp b/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
index eea18ec70..c1828a0ec 100644
--- a/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_image_patch_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_image_patch_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -1084,7 +1084,7 @@ test_patch_padding_same_sycl<DataType, int64_t>(sycl_device);
 test_patch_no_extra_dim_sycl<DataType, int64_t>(sycl_device);
 test_imagenet_patches_sycl<DataType, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_image_patch_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_image_patch_sycl)
 {
 for (const auto& device :Eigen::get_sycl_supported_devices()) {
   CALL_SUBTEST(sycl_tensor_image_patch_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_index_list.cpp b/unsupported/test/cxx11_tensor_index_list.cpp
index 4cf5df666..e81fa5e40 100644
--- a/unsupported/test/cxx11_tensor_index_list.cpp
+++ b/unsupported/test/cxx11_tensor_index_list.cpp
@@ -373,7 +373,7 @@ static void test_dim_check()
 
 #endif
 
-void test_cxx11_tensor_index_list()
+EIGEN_DECLARE_TEST(cxx11_tensor_index_list)
 {
 #ifdef EIGEN_HAS_INDEX_LIST
   CALL_SUBTEST(test_static_index_list());
diff --git a/unsupported/test/cxx11_tensor_inflation.cpp b/unsupported/test/cxx11_tensor_inflation.cpp
index 4997935e9..75089e856 100644
--- a/unsupported/test/cxx11_tensor_inflation.cpp
+++ b/unsupported/test/cxx11_tensor_inflation.cpp
@@ -74,7 +74,7 @@ static void test_simple_inflation()
   }
 }
 
-void test_cxx11_tensor_inflation()
+EIGEN_DECLARE_TEST(cxx11_tensor_inflation)
 {
   CALL_SUBTEST(test_simple_inflation<ColMajor>());
   CALL_SUBTEST(test_simple_inflation<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_inflation_sycl.cpp b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
index cf3e29f4c..521ae0cc3 100644
--- a/unsupported/test/cxx11_tensor_inflation_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_inflation_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -128,7 +128,7 @@ template<typename DataType, typename dev_Selector> void sycl_inflation_test_per_
   test_simple_inflation_sycl<DataType, RowMajor, int64_t>(sycl_device);
   test_simple_inflation_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_inflation_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_inflation_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_inflation_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_intdiv.cpp b/unsupported/test/cxx11_tensor_intdiv.cpp
index 8e2b70b75..d18a05ec4 100644
--- a/unsupported/test/cxx11_tensor_intdiv.cpp
+++ b/unsupported/test/cxx11_tensor_intdiv.cpp
@@ -135,7 +135,7 @@ void test_specific() {
   VERIFY_IS_EQUAL(result, result_op);
 }
 
-void test_cxx11_tensor_intdiv()
+EIGEN_DECLARE_TEST(cxx11_tensor_intdiv)
 {
   CALL_SUBTEST_1(test_signed_32bit());
   CALL_SUBTEST_2(test_unsigned_32bit());
diff --git a/unsupported/test/cxx11_tensor_io.cpp b/unsupported/test/cxx11_tensor_io.cpp
index 489960529..2c638f9bf 100644
--- a/unsupported/test/cxx11_tensor_io.cpp
+++ b/unsupported/test/cxx11_tensor_io.cpp
@@ -119,7 +119,7 @@ static void test_output_const()
 }
 
 
-void test_cxx11_tensor_io()
+EIGEN_DECLARE_TEST(cxx11_tensor_io)
 {
   CALL_SUBTEST(test_output_0d<ColMajor>());
   CALL_SUBTEST(test_output_0d<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_layout_swap.cpp b/unsupported/test/cxx11_tensor_layout_swap.cpp
index ae297a9da..efb333360 100644
--- a/unsupported/test/cxx11_tensor_layout_swap.cpp
+++ b/unsupported/test/cxx11_tensor_layout_swap.cpp
@@ -54,7 +54,7 @@ static void test_swap_as_lvalue()
 }
 
 
-void test_cxx11_tensor_layout_swap()
+EIGEN_DECLARE_TEST(cxx11_tensor_layout_swap)
 {
   CALL_SUBTEST(test_simple_swap());
   CALL_SUBTEST(test_swap_as_lvalue());
diff --git a/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp b/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp
index 9e8db8b4b..9546b911c 100644
--- a/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_layout_swap_sycl.cpp
@@ -14,7 +14,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_layout_swap_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -118,7 +118,7 @@ template<typename DataType, typename dev_Selector> void sycl_tensor_layout_swap_
   test_simple_swap_sycl<DataType, int64_t>(sycl_device);
   test_swap_as_lvalue_sycl<DataType, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_layout_swap_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_layout_swap_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_tensor_layout_swap_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_lvalue.cpp b/unsupported/test/cxx11_tensor_lvalue.cpp
index 071f5b406..6ba9a212d 100644
--- a/unsupported/test/cxx11_tensor_lvalue.cpp
+++ b/unsupported/test/cxx11_tensor_lvalue.cpp
@@ -36,7 +36,7 @@ static void test_compound_assignment()
 }
 
 
-void test_cxx11_tensor_lvalue()
+EIGEN_DECLARE_TEST(cxx11_tensor_lvalue)
 {
   CALL_SUBTEST(test_compound_assignment());
 }
diff --git a/unsupported/test/cxx11_tensor_map.cpp b/unsupported/test/cxx11_tensor_map.cpp
index 3db0ee7c0..ce608aca7 100644
--- a/unsupported/test/cxx11_tensor_map.cpp
+++ b/unsupported/test/cxx11_tensor_map.cpp
@@ -265,7 +265,7 @@ static void test_casting()
   VERIFY_IS_EQUAL(sum1, 861);
 }
 
-void test_cxx11_tensor_map()
+EIGEN_DECLARE_TEST(cxx11_tensor_map)
 {
   CALL_SUBTEST(test_0d());
   CALL_SUBTEST(test_1d());
diff --git a/unsupported/test/cxx11_tensor_math.cpp b/unsupported/test/cxx11_tensor_math.cpp
index 61c742a16..82a1a26d8 100644
--- a/unsupported/test/cxx11_tensor_math.cpp
+++ b/unsupported/test/cxx11_tensor_math.cpp
@@ -39,7 +39,7 @@ static void test_sigmoid()
 }
 
 
-void test_cxx11_tensor_math()
+EIGEN_DECLARE_TEST(cxx11_tensor_math)
 {
   CALL_SUBTEST(test_tanh());
   CALL_SUBTEST(test_sigmoid());
diff --git a/unsupported/test/cxx11_tensor_mixed_indices.cpp b/unsupported/test/cxx11_tensor_mixed_indices.cpp
index 4fba6fdd1..ee2616fd7 100644
--- a/unsupported/test/cxx11_tensor_mixed_indices.cpp
+++ b/unsupported/test/cxx11_tensor_mixed_indices.cpp
@@ -47,7 +47,7 @@ static void test_simple()
 }
 
 
-void test_cxx11_tensor_mixed_indices()
+EIGEN_DECLARE_TEST(cxx11_tensor_mixed_indices)
 {
   CALL_SUBTEST(test_simple());
 }
diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp
index f7de43110..6365cd89a 100644
--- a/unsupported/test/cxx11_tensor_morphing.cpp
+++ b/unsupported/test/cxx11_tensor_morphing.cpp
@@ -459,7 +459,7 @@ static void test_composition()
 }
 
 
-void test_cxx11_tensor_morphing()
+EIGEN_DECLARE_TEST(cxx11_tensor_morphing)
 {
   CALL_SUBTEST_1(test_simple_reshape<void>());
   CALL_SUBTEST_1(test_reshape_in_expr<void>());
diff --git a/unsupported/test/cxx11_tensor_morphing_sycl.cpp b/unsupported/test/cxx11_tensor_morphing_sycl.cpp
index 9b521bc6b..93dabe3ec 100644
--- a/unsupported/test/cxx11_tensor_morphing_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_morphing_sycl.cpp
@@ -15,7 +15,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_morphing_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -240,7 +240,7 @@ template<typename DataType, typename dev_Selector> void sycl_morphing_test_per_d
   test_strided_slice_write_sycl<DataType, ColMajor, int64_t>(sycl_device);
   test_strided_slice_write_sycl<DataType, RowMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_morphing_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_morphing_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_morphing_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_notification.cpp b/unsupported/test/cxx11_tensor_notification.cpp
index 183ef02c1..f63fee013 100644
--- a/unsupported/test/cxx11_tensor_notification.cpp
+++ b/unsupported/test/cxx11_tensor_notification.cpp
@@ -65,7 +65,7 @@ static void test_notification_multiple()
   VERIFY_IS_EQUAL(counter, 4);
 }
 
-void test_cxx11_tensor_notification()
+EIGEN_DECLARE_TEST(cxx11_tensor_notification)
 {
   CALL_SUBTEST(test_notification_single());
   CALL_SUBTEST(test_notification_multiple());
diff --git a/unsupported/test/cxx11_tensor_of_complex.cpp b/unsupported/test/cxx11_tensor_of_complex.cpp
index e9d1b2d3c..99e18076a 100644
--- a/unsupported/test/cxx11_tensor_of_complex.cpp
+++ b/unsupported/test/cxx11_tensor_of_complex.cpp
@@ -94,7 +94,7 @@ static void test_contractions()
 }
 
 
-void test_cxx11_tensor_of_complex()
+EIGEN_DECLARE_TEST(cxx11_tensor_of_complex)
 {
   CALL_SUBTEST(test_additions());
   CALL_SUBTEST(test_abs());
diff --git a/unsupported/test/cxx11_tensor_of_const_values.cpp b/unsupported/test/cxx11_tensor_of_const_values.cpp
index f179a0c21..344d678ef 100644
--- a/unsupported/test/cxx11_tensor_of_const_values.cpp
+++ b/unsupported/test/cxx11_tensor_of_const_values.cpp
@@ -97,7 +97,7 @@ static void test_plus_equal()
 }
 
 
-void test_cxx11_tensor_of_const_values()
+EIGEN_DECLARE_TEST(cxx11_tensor_of_const_values)
 {
   CALL_SUBTEST(test_assign());
   CALL_SUBTEST(test_plus());
diff --git a/unsupported/test/cxx11_tensor_of_float16_gpu.cu b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
index 150fde8bf..1f1ec26c2 100644
--- a/unsupported/test/cxx11_tensor_of_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_gpu
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_of_strings.cpp b/unsupported/test/cxx11_tensor_of_strings.cpp
index 4ef9aed91..159656276 100644
--- a/unsupported/test/cxx11_tensor_of_strings.cpp
+++ b/unsupported/test/cxx11_tensor_of_strings.cpp
@@ -141,7 +141,7 @@ static void test_initialization()
 }
 
 
-void test_cxx11_tensor_of_strings()
+EIGEN_DECLARE_TEST(cxx11_tensor_of_strings)
 {
   // Beware: none of this is likely to ever work on a GPU.
   CALL_SUBTEST(test_assign());
diff --git a/unsupported/test/cxx11_tensor_padding.cpp b/unsupported/test/cxx11_tensor_padding.cpp
index ffa19896e..b8a329deb 100644
--- a/unsupported/test/cxx11_tensor_padding.cpp
+++ b/unsupported/test/cxx11_tensor_padding.cpp
@@ -84,7 +84,7 @@ static void test_padded_expr()
   }
 }
 
-void test_cxx11_tensor_padding()
+EIGEN_DECLARE_TEST(cxx11_tensor_padding)
 {
   CALL_SUBTEST(test_simple_padding<ColMajor>());
   CALL_SUBTEST(test_simple_padding<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_padding_sycl.cpp b/unsupported/test/cxx11_tensor_padding_sycl.cpp
index dc748b73e..727a9ffd7 100644
--- a/unsupported/test/cxx11_tensor_padding_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_padding_sycl.cpp
@@ -15,7 +15,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_padding_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -149,7 +149,7 @@ template<typename DataType, typename dev_Selector> void sycl_padding_test_per_de
   test_padded_expr<DataType, ColMajor, int64_t>(sycl_device);
 
 }
-void test_cxx11_tensor_padding_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_padding_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_padding_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_patch.cpp b/unsupported/test/cxx11_tensor_patch.cpp
index 434359730..498ab8ca7 100644
--- a/unsupported/test/cxx11_tensor_patch.cpp
+++ b/unsupported/test/cxx11_tensor_patch.cpp
@@ -164,7 +164,7 @@ static void test_simple_patch()
   }
 }
 
-void test_cxx11_tensor_patch()
+EIGEN_DECLARE_TEST(cxx11_tensor_patch)
 {
    CALL_SUBTEST(test_simple_patch<ColMajor>());
    CALL_SUBTEST(test_simple_patch<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_patch_sycl.cpp b/unsupported/test/cxx11_tensor_patch_sycl.cpp
index 88a29cb31..7f92bec78 100644
--- a/unsupported/test/cxx11_tensor_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_patch_sycl.cpp
@@ -14,7 +14,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_patch_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -241,7 +241,7 @@ template<typename DataType, typename dev_Selector> void sycl_tensor_patch_test_p
   test_simple_patch_sycl<DataType, RowMajor, int64_t>(sycl_device);
   test_simple_patch_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_patch_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_patch_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_tensor_patch_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_random.cpp b/unsupported/test/cxx11_tensor_random.cpp
index 0f3dc5787..4740d5811 100644
--- a/unsupported/test/cxx11_tensor_random.cpp
+++ b/unsupported/test/cxx11_tensor_random.cpp
@@ -70,7 +70,7 @@ static void test_custom()
   }
 }
 
-void test_cxx11_tensor_random()
+EIGEN_DECLARE_TEST(cxx11_tensor_random)
 {
   CALL_SUBTEST(test_default());
   CALL_SUBTEST(test_normal());
diff --git a/unsupported/test/cxx11_tensor_random_gpu.cu b/unsupported/test/cxx11_tensor_random_gpu.cu
index da5977f09..262182d30 100644
--- a/unsupported/test/cxx11_tensor_random_gpu.cu
+++ b/unsupported/test/cxx11_tensor_random_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_random_gpu
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp
index 1490ec3da..4c8a95c34 100644
--- a/unsupported/test/cxx11_tensor_reduction.cpp
+++ b/unsupported/test/cxx11_tensor_reduction.cpp
@@ -484,7 +484,7 @@ static void test_reduce_middle_dims() {
   }
 }
 
-void test_cxx11_tensor_reduction() {
+EIGEN_DECLARE_TEST(cxx11_tensor_reduction) {
   CALL_SUBTEST(test_trivial_reductions<ColMajor>());
   CALL_SUBTEST(test_trivial_reductions<RowMajor>());
   CALL_SUBTEST(test_simple_reductions<ColMajor>());
diff --git a/unsupported/test/cxx11_tensor_reduction_gpu.cu b/unsupported/test/cxx11_tensor_reduction_gpu.cu
index a36759303..7b8ac2309 100644
--- a/unsupported/test/cxx11_tensor_reduction_gpu.cu
+++ b/unsupported/test/cxx11_tensor_reduction_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_reduction_gpu
+
 #define EIGEN_USE_GPU
 
 #include "main.h"
diff --git a/unsupported/test/cxx11_tensor_reduction_sycl.cpp b/unsupported/test/cxx11_tensor_reduction_sycl.cpp
index 440d48bca..f526299c6 100644
--- a/unsupported/test/cxx11_tensor_reduction_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_reduction_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_reduction_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -174,7 +174,7 @@ template<typename DataType> void sycl_reduction_test_per_device(const cl::sycl::
   test_first_dim_reductions_max_sycl<DataType, ColMajor, int64_t>(sycl_device);
   test_last_dim_reductions_sum_sycl<DataType, ColMajor, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_reduction_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_reduction_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_reduction_test_per_device<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_ref.cpp b/unsupported/test/cxx11_tensor_ref.cpp
index c8f105e3d..7dbd0478c 100644
--- a/unsupported/test/cxx11_tensor_ref.cpp
+++ b/unsupported/test/cxx11_tensor_ref.cpp
@@ -235,7 +235,7 @@ static void test_nested_ops_with_ref()
 }
 
 
-void test_cxx11_tensor_ref()
+EIGEN_DECLARE_TEST(cxx11_tensor_ref)
 {
   CALL_SUBTEST(test_simple_lvalue_ref());
   CALL_SUBTEST(test_simple_rvalue_ref());
diff --git a/unsupported/test/cxx11_tensor_reverse.cpp b/unsupported/test/cxx11_tensor_reverse.cpp
index b35b8d29e..5e44ec007 100644
--- a/unsupported/test/cxx11_tensor_reverse.cpp
+++ b/unsupported/test/cxx11_tensor_reverse.cpp
@@ -179,7 +179,7 @@ static void test_expr_reverse(bool LValue)
 }
 
 
-void test_cxx11_tensor_reverse()
+EIGEN_DECLARE_TEST(cxx11_tensor_reverse)
 {
   CALL_SUBTEST(test_simple_reverse<ColMajor>());
   CALL_SUBTEST(test_simple_reverse<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_reverse_sycl.cpp b/unsupported/test/cxx11_tensor_reverse_sycl.cpp
index 2f5484484..77c2235d1 100644
--- a/unsupported/test/cxx11_tensor_reverse_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_reverse_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_reverse_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -214,7 +214,7 @@ template<typename DataType> void sycl_reverse_test_per_device(const cl::sycl::de
   test_expr_reverse<DataType, RowMajor, int64_t>(sycl_device, true);
   test_expr_reverse<DataType, ColMajor, int64_t>(sycl_device, true);
 }
-void test_cxx11_tensor_reverse_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_reverse_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_reverse_test_per_device<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_roundings.cpp b/unsupported/test/cxx11_tensor_roundings.cpp
index 2c26151ab..83b592384 100644
--- a/unsupported/test/cxx11_tensor_roundings.cpp
+++ b/unsupported/test/cxx11_tensor_roundings.cpp
@@ -54,7 +54,7 @@ static void test_float_ceiling()
   }
 }
 
-void test_cxx11_tensor_roundings()
+EIGEN_DECLARE_TEST(cxx11_tensor_roundings)
 {
    CALL_SUBTEST(test_float_rounding());
    CALL_SUBTEST(test_float_ceiling());
diff --git a/unsupported/test/cxx11_tensor_scan.cpp b/unsupported/test/cxx11_tensor_scan.cpp
index af59aa3ef..dccee9e84 100644
--- a/unsupported/test/cxx11_tensor_scan.cpp
+++ b/unsupported/test/cxx11_tensor_scan.cpp
@@ -98,7 +98,7 @@ static void test_tensor_maps() {
   }
 }
 
-void test_cxx11_tensor_scan() {
+EIGEN_DECLARE_TEST(cxx11_tensor_scan) {
   CALL_SUBTEST((test_1d_scan<ColMajor, float, true>()));
   CALL_SUBTEST((test_1d_scan<ColMajor, float, false>()));
   CALL_SUBTEST((test_1d_scan<RowMajor, float, true>()));
diff --git a/unsupported/test/cxx11_tensor_scan_gpu.cu b/unsupported/test/cxx11_tensor_scan_gpu.cu
index 51cd3a3cf..f3e773db5 100644
--- a/unsupported/test/cxx11_tensor_scan_gpu.cu
+++ b/unsupported/test/cxx11_tensor_scan_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_scan_gpu
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp
index d11444a14..ab19b6e6b 100644
--- a/unsupported/test/cxx11_tensor_shuffling.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling.cpp
@@ -215,7 +215,7 @@ static void test_shuffle_unshuffle()
 }
 
 
-void test_cxx11_tensor_shuffling()
+EIGEN_DECLARE_TEST(cxx11_tensor_shuffling)
 {
   CALL_SUBTEST(test_simple_shuffling<ColMajor>());
   CALL_SUBTEST(test_simple_shuffling<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_shuffling_sycl.cpp b/unsupported/test/cxx11_tensor_shuffling_sycl.cpp
index c88db7c72..0e8cc3bd2 100644
--- a/unsupported/test/cxx11_tensor_shuffling_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling_sycl.cpp
@@ -15,7 +15,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_shuffling_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -111,7 +111,7 @@ template<typename DataType, typename dev_Selector> void sycl_shuffling_test_per_
   test_simple_shuffling_sycl<DataType, ColMajor, int64_t>(sycl_device);
 
 }
-void test_cxx11_tensor_shuffling_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_shuffling_sycl)
 {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_shuffling_test_per_device<float>(device));
diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp
index 5a0d339ef..6d70f5435 100644
--- a/unsupported/test/cxx11_tensor_simple.cpp
+++ b/unsupported/test/cxx11_tensor_simple.cpp
@@ -316,7 +316,7 @@ static void test_resize()
   VERIFY_IS_EQUAL(epsilon.size(), 3*5*7);
 }
 
-void test_cxx11_tensor_simple()
+EIGEN_DECLARE_TEST(cxx11_tensor_simple)
 {
   CALL_SUBTEST(test_0d());
   CALL_SUBTEST(test_1d());
diff --git a/unsupported/test/cxx11_tensor_striding.cpp b/unsupported/test/cxx11_tensor_striding.cpp
index 935b908cc..aefdfa9b4 100644
--- a/unsupported/test/cxx11_tensor_striding.cpp
+++ b/unsupported/test/cxx11_tensor_striding.cpp
@@ -110,7 +110,7 @@ static void test_striding_as_lvalue()
 }
 
 
-void test_cxx11_tensor_striding()
+EIGEN_DECLARE_TEST(cxx11_tensor_striding)
 {
   CALL_SUBTEST(test_simple_striding<ColMajor>());
   CALL_SUBTEST(test_simple_striding<RowMajor>());
diff --git a/unsupported/test/cxx11_tensor_striding_sycl.cpp b/unsupported/test/cxx11_tensor_striding_sycl.cpp
index 603c3746f..d3b1fa77c 100644
--- a/unsupported/test/cxx11_tensor_striding_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_striding_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_striding_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -196,7 +196,7 @@ template <typename Dev_selector> void tensorStridingPerDevice(Dev_selector& s){
   test_striding_as_lvalue<float, RowMajor, int64_t>(sycl_device);
 }
 
-void test_cxx11_tensor_striding_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_striding_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(tensorStridingPerDevice(device));
   }
diff --git a/unsupported/test/cxx11_tensor_sugar.cpp b/unsupported/test/cxx11_tensor_sugar.cpp
index 2f56eb495..2ca5c47db 100644
--- a/unsupported/test/cxx11_tensor_sugar.cpp
+++ b/unsupported/test/cxx11_tensor_sugar.cpp
@@ -73,7 +73,7 @@ static void test_scalar_sugar_sub_div() {
   }
 }
 
-void test_cxx11_tensor_sugar()
+EIGEN_DECLARE_TEST(cxx11_tensor_sugar)
 {
   CALL_SUBTEST(test_comparison_sugar());
   CALL_SUBTEST(test_scalar_sugar_add_mul());
diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp
index 5cd0f4c71..9357bed02 100644
--- a/unsupported/test/cxx11_tensor_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_sycl.cpp
@@ -15,7 +15,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -269,7 +269,7 @@ template<typename DataType, typename dev_Selector> void sycl_computing_test_per_
   test_sycl_cast<DataType, int, ColMajor, int64_t>(sycl_device);
 }
 
-void test_cxx11_tensor_sycl() {
+EIGEN_DECLARE_TEST(cxx11_tensor_sycl) {
   for (const auto& device :Eigen::get_sycl_supported_devices()) {
     CALL_SUBTEST(sycl_computing_test_per_device<float>(device));
   }
diff --git a/unsupported/test/cxx11_tensor_symmetry.cpp b/unsupported/test/cxx11_tensor_symmetry.cpp
index d680e9b3b..fed269a9a 100644
--- a/unsupported/test/cxx11_tensor_symmetry.cpp
+++ b/unsupported/test/cxx11_tensor_symmetry.cpp
@@ -801,7 +801,7 @@ static void test_tensor_randacc()
   }
 }
 
-void test_cxx11_tensor_symmetry()
+EIGEN_DECLARE_TEST(cxx11_tensor_symmetry)
 {
   CALL_SUBTEST(test_symgroups_static());
   CALL_SUBTEST(test_symgroups_dynamic());
diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index ea9d8afdc..20a197f2b 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -399,7 +399,7 @@ void test_multithread_shuffle()
 }
 
 
-void test_cxx11_tensor_thread_pool()
+EIGEN_DECLARE_TEST(cxx11_tensor_thread_pool)
 {
   CALL_SUBTEST_1(test_multithread_elementwise());
   CALL_SUBTEST_1(test_multithread_compound_assignment());
diff --git a/unsupported/test/cxx11_tensor_trace.cpp b/unsupported/test/cxx11_tensor_trace.cpp
index 340d1211c..1579bc1eb 100644
--- a/unsupported/test/cxx11_tensor_trace.cpp
+++ b/unsupported/test/cxx11_tensor_trace.cpp
@@ -159,7 +159,7 @@ static void test_trace_in_expr() {
 }
 
 
-void test_cxx11_tensor_trace() {
+EIGEN_DECLARE_TEST(cxx11_tensor_trace) {
   CALL_SUBTEST(test_0D_trace<ColMajor>());
   CALL_SUBTEST(test_0D_trace<RowMajor>());
   CALL_SUBTEST(test_all_dimensions_trace<ColMajor>());
diff --git a/unsupported/test/cxx11_tensor_uint128.cpp b/unsupported/test/cxx11_tensor_uint128.cpp
index d2a1e8673..07691df98 100644
--- a/unsupported/test/cxx11_tensor_uint128.cpp
+++ b/unsupported/test/cxx11_tensor_uint128.cpp
@@ -144,7 +144,7 @@ void test_misc2() {
 #endif
 
 
-void test_cxx11_tensor_uint128()
+EIGEN_DECLARE_TEST(cxx11_tensor_uint128)
 {
 #ifdef EIGEN_NO_INT128
   // Skip the test on compilers that don't support 128bit integers natively
diff --git a/unsupported/test/cxx11_tensor_volume_patch.cpp b/unsupported/test/cxx11_tensor_volume_patch.cpp
index ca6840f3b..3aa363eb5 100644
--- a/unsupported/test/cxx11_tensor_volume_patch.cpp
+++ b/unsupported/test/cxx11_tensor_volume_patch.cpp
@@ -105,7 +105,7 @@ static void test_entire_volume_patch()
   }
 }
 
-void test_cxx11_tensor_volume_patch()
+EIGEN_DECLARE_TEST(cxx11_tensor_volume_patch)
 {
   CALL_SUBTEST(test_single_voxel_patch());
   CALL_SUBTEST(test_entire_volume_patch());
diff --git a/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp b/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
index 039715abc..ca7994fd9 100644
--- a/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_volume_patch_sycl.cpp
@@ -13,7 +13,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_volume_patch_sycl
+
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
 #define EIGEN_USE_SYCL
 
@@ -214,7 +214,7 @@ std::cout << "Running on " << s.template get_info<cl::sycl::info::device::name>(
 test_single_voxel_patch_sycl<DataType, int64_t>(sycl_device);
 test_entire_volume_patch_sycl<DataType, int64_t>(sycl_device);
 }
-void test_cxx11_tensor_volume_patch_sycl()
+EIGEN_DECLARE_TEST(cxx11_tensor_volume_patch_sycl)
 {
 for (const auto& device :Eigen::get_sycl_supported_devices()) {
   CALL_SUBTEST(sycl_tensor_volume_patch_test_per_device<float>(device));
diff --git a/unsupported/test/dgmres.cpp b/unsupported/test/dgmres.cpp
index 2b11807c8..04f5ad670 100644
--- a/unsupported/test/dgmres.cpp
+++ b/unsupported/test/dgmres.cpp
@@ -24,7 +24,7 @@ template<typename T> void test_dgmres_T()
   //CALL_SUBTEST( check_sparse_square_solving(dgmres_colmajor_ssor)     );
 }
 
-void test_dgmres()
+EIGEN_DECLARE_TEST(dgmres)
 {
   CALL_SUBTEST_1(test_dgmres_T<double>());
   CALL_SUBTEST_2(test_dgmres_T<std::complex<double> >());
diff --git a/unsupported/test/forward_adolc.cpp b/unsupported/test/forward_adolc.cpp
index 6d0ae738d..688594eff 100644
--- a/unsupported/test/forward_adolc.cpp
+++ b/unsupported/test/forward_adolc.cpp
@@ -119,7 +119,7 @@ template<typename Func> void adolc_forward_jacobian(const Func& f)
     VERIFY_IS_APPROX(j, jref);
 }
 
-void test_forward_adolc()
+EIGEN_DECLARE_TEST(forward_adolc)
 {
   adtl::setNumDir(NUMBER_DIRECTIONS);
 
diff --git a/unsupported/test/gmres.cpp b/unsupported/test/gmres.cpp
index f2969116b..8d2254b5b 100644
--- a/unsupported/test/gmres.cpp
+++ b/unsupported/test/gmres.cpp
@@ -24,7 +24,7 @@ template<typename T> void test_gmres_T()
   //CALL_SUBTEST( check_sparse_square_solving(gmres_colmajor_ssor)     );
 }
 
-void test_gmres()
+EIGEN_DECLARE_TEST(gmres)
 {
   CALL_SUBTEST_1(test_gmres_T<double>());
   CALL_SUBTEST_2(test_gmres_T<std::complex<double> >());
diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp
index e770049e5..4f143b6de 100644
--- a/unsupported/test/kronecker_product.cpp
+++ b/unsupported/test/kronecker_product.cpp
@@ -83,7 +83,7 @@ void check_sparse_kronecker_product(const MatrixType& ab)
 }
 
 
-void test_kronecker_product()
+EIGEN_DECLARE_TEST(kronecker_product)
 {
   // DM = dense matrix; SM = sparse matrix
 
@@ -240,7 +240,7 @@ void test_kronecker_product()
 #include "main.h"
 #include <Eigen/KroneckerProduct>
 
-void test_kronecker_product()
+EIGEN_DECLARE_TEST(kronecker_product)
 {
   MatrixXd a(2,2), b(3,3), c;
   a.setRandom();
diff --git a/unsupported/test/levenberg_marquardt.cpp b/unsupported/test/levenberg_marquardt.cpp
index 64f168c16..7f9a81cd3 100644
--- a/unsupported/test/levenberg_marquardt.cpp
+++ b/unsupported/test/levenberg_marquardt.cpp
@@ -1445,7 +1445,7 @@ void testNistEckerle4(void)
   VERIFY_IS_APPROX(x[2], 4.5154121844E+02);
 }
 
-void test_levenberg_marquardt()
+EIGEN_DECLARE_TEST(levenberg_marquardt)
 {
     // Tests using the examples provided by (c)minpack
     CALL_SUBTEST(testLmder1());
diff --git a/unsupported/test/matrix_exponential.cpp b/unsupported/test/matrix_exponential.cpp
index 50dec083d..b032cbf1d 100644
--- a/unsupported/test/matrix_exponential.cpp
+++ b/unsupported/test/matrix_exponential.cpp
@@ -119,7 +119,7 @@ void randomTest(const MatrixType& m, double tol)
   }
 }
 
-void test_matrix_exponential()
+EIGEN_DECLARE_TEST(matrix_exponential)
 {
   CALL_SUBTEST_2(test2dRotation<double>(1e-13));
   CALL_SUBTEST_1(test2dRotation<float>(2e-5));  // was 1e-5, relaxed for clang 2.8 / linux / x86-64
diff --git a/unsupported/test/matrix_function.cpp b/unsupported/test/matrix_function.cpp
index 7c9b68a3c..93fb71430 100644
--- a/unsupported/test/matrix_function.cpp
+++ b/unsupported/test/matrix_function.cpp
@@ -181,7 +181,7 @@ void testMatrixType(const MatrixType& m)
   }
 }
 
-void test_matrix_function()
+EIGEN_DECLARE_TEST(matrix_function)
 {
   CALL_SUBTEST_1(testMatrixType(Matrix<float,1,1>()));
   CALL_SUBTEST_2(testMatrixType(Matrix3cf()));
diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp
index 7ccfacfdf..fa52d256e 100644
--- a/unsupported/test/matrix_power.cpp
+++ b/unsupported/test/matrix_power.cpp
@@ -150,7 +150,7 @@ typedef Matrix<double,3,3,RowMajor>         Matrix3dRowMajor;
 typedef Matrix<long double,3,3>             Matrix3e;
 typedef Matrix<long double,Dynamic,Dynamic> MatrixXe;
  
-void test_matrix_power()
+EIGEN_DECLARE_TEST(matrix_power)
 {
   CALL_SUBTEST_2(test2dRotation<double>(1e-13));
   CALL_SUBTEST_1(test2dRotation<float>(2e-5));  // was 1e-5, relaxed for clang 2.8 / linux / x86-64
diff --git a/unsupported/test/matrix_square_root.cpp b/unsupported/test/matrix_square_root.cpp
index ea541e1ea..034f29217 100644
--- a/unsupported/test/matrix_square_root.cpp
+++ b/unsupported/test/matrix_square_root.cpp
@@ -18,7 +18,7 @@ void testMatrixSqrt(const MatrixType& m)
   VERIFY_IS_APPROX(sqrtA * sqrtA, A);
 }
 
-void test_matrix_square_root()
+EIGEN_DECLARE_TEST(matrix_square_root)
 {
   for (int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(testMatrixSqrt(Matrix3cf()));
diff --git a/unsupported/test/minres.cpp b/unsupported/test/minres.cpp
index 8b300b78a..2eb40fef6 100644
--- a/unsupported/test/minres.cpp
+++ b/unsupported/test/minres.cpp
@@ -36,7 +36,7 @@ template<typename T> void test_minres_T()
 
 }
 
-void test_minres()
+EIGEN_DECLARE_TEST(minres)
 {
   CALL_SUBTEST_1(test_minres_T<double>());
 //  CALL_SUBTEST_2(test_minres_T<std::compex<double> >());
diff --git a/unsupported/test/mpreal_support.cpp b/unsupported/test/mpreal_support.cpp
index 685e7ea45..4a25e993c 100644
--- a/unsupported/test/mpreal_support.cpp
+++ b/unsupported/test/mpreal_support.cpp
@@ -7,7 +7,7 @@
 using namespace mpfr;
 using namespace Eigen;
 
-void test_mpreal_support()
+EIGEN_DECLARE_TEST(mpreal_support)
 {
   // set precision to 256 bits (double has only 53 bits)
   mpreal::set_default_prec(256);
diff --git a/unsupported/test/openglsupport.cpp b/unsupported/test/openglsupport.cpp
index 706a816f7..460830086 100644
--- a/unsupported/test/openglsupport.cpp
+++ b/unsupported/test/openglsupport.cpp
@@ -106,7 +106,7 @@ GLint createShader(const char* vtx, const char* frg)
   return prg_id;
 }
 
-void test_openglsupport()
+EIGEN_DECLARE_TEST(openglsupport)
 {
   int argc = 0;
   glutInit(&argc, 0);
diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp
index 7ad4aa69d..65efea0cb 100644
--- a/unsupported/test/polynomialsolver.cpp
+++ b/unsupported/test/polynomialsolver.cpp
@@ -191,7 +191,7 @@ void polynomialsolver(int deg)
       realRoots );
 }
 
-void test_polynomialsolver()
+EIGEN_DECLARE_TEST(polynomialsolver)
 {
   for(int i = 0; i < g_repeat; i++)
   {
diff --git a/unsupported/test/polynomialutils.cpp b/unsupported/test/polynomialutils.cpp
index 5fc968402..8ff451996 100644
--- a/unsupported/test/polynomialutils.cpp
+++ b/unsupported/test/polynomialutils.cpp
@@ -101,7 +101,7 @@ template<typename _Scalar> void CauchyBounds_scalar()
           internal::random<int>(18,26) )) );
 }
 
-void test_polynomialutils()
+EIGEN_DECLARE_TEST(polynomialutils)
 {
   for(int i = 0; i < g_repeat; i++)
   {
diff --git a/unsupported/test/sparse_extra.cpp b/unsupported/test/sparse_extra.cpp
index 7cf4a77c3..4ac53a9a7 100644
--- a/unsupported/test/sparse_extra.cpp
+++ b/unsupported/test/sparse_extra.cpp
@@ -142,7 +142,7 @@ void check_marketio()
   VERIFY_IS_EQUAL(DenseMatrix(m1),DenseMatrix(m2));
 }
 
-void test_sparse_extra()
+EIGEN_DECLARE_TEST(sparse_extra)
 {
   for(int i = 0; i < g_repeat; i++) {
     int s = Eigen::internal::random<int>(1,50);
diff --git a/unsupported/test/special_functions.cpp b/unsupported/test/special_functions.cpp
index 802e16150..50dae6c93 100644
--- a/unsupported/test/special_functions.cpp
+++ b/unsupported/test/special_functions.cpp
@@ -472,7 +472,7 @@ template<typename ArrayType> void array_special_functions()
 #endif  // EIGEN_HAS_C99_MATH
 }
 
-void test_special_functions()
+EIGEN_DECLARE_TEST(special_functions)
 {
   CALL_SUBTEST_1(array_special_functions<ArrayXf>());
   CALL_SUBTEST_2(array_special_functions<ArrayXd>());
diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp
index 3be020434..88ec87b97 100644
--- a/unsupported/test/splines.cpp
+++ b/unsupported/test/splines.cpp
@@ -268,7 +268,7 @@ void check_global_interpolation_with_derivatives2d()
   }
 }
 
-void test_splines()
+EIGEN_DECLARE_TEST(splines)
 {
   for (int i = 0; i < g_repeat; ++i)
   {

From dff3a92d527fd38c28152ab9259af1904a01f248 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 15:52:58 +0200
Subject: [PATCH 256/680] Remove usage of #if EIGEN_TEST_PART_XX in unit tests
 that does not require them (splitting can thus be avoided for them)

---
 test/adjoint.cpp             | 56 +++++++++++-----------
 test/array_reverse.cpp       | 16 ++++---
 test/basicstuff.cpp          |  5 +-
 test/eigensolver_generic.cpp | 54 +++++++++++----------
 test/incomplete_cholesky.cpp | 20 ++++----
 test/integer_types.cpp       | 20 ++++----
 test/inverse.cpp             | 91 ++++++++++++++++++++----------------
 test/linearstructure.cpp     | 29 ++++++------
 test/nullary.cpp             | 76 ++++++++++++++++--------------
 test/product_large.cpp       | 34 ++++++++------
 test/product_small.cpp       | 58 ++++++++++++-----------
 test/sparse_basic.cpp        | 29 ++++++------
 12 files changed, 267 insertions(+), 221 deletions(-)

diff --git a/test/adjoint.cpp b/test/adjoint.cpp
index ca44466b2..4e1e4b5e8 100644
--- a/test/adjoint.cpp
+++ b/test/adjoint.cpp
@@ -145,6 +145,34 @@ template<typename MatrixType> void adjoint(const MatrixType& m)
   VERIFY_IS_APPROX(rv1.template cast<Scalar>().dot(v1), rv1.dot(v1));
 }
 
+template<int>
+void adjoint_extra()
+{
+  MatrixXcf a(10,10), b(10,10);
+  VERIFY_RAISES_ASSERT(a = a.transpose());
+  VERIFY_RAISES_ASSERT(a = a.transpose() + b);
+  VERIFY_RAISES_ASSERT(a = b + a.transpose());
+  VERIFY_RAISES_ASSERT(a = a.conjugate().transpose());
+  VERIFY_RAISES_ASSERT(a = a.adjoint());
+  VERIFY_RAISES_ASSERT(a = a.adjoint() + b);
+  VERIFY_RAISES_ASSERT(a = b + a.adjoint());
+
+  // no assertion should be triggered for these cases:
+  a.transpose() = a.transpose();
+  a.transpose() += a.transpose();
+  a.transpose() += a.transpose() + b;
+  a.transpose() = a.adjoint();
+  a.transpose() += a.adjoint();
+  a.transpose() += a.adjoint() + b;
+
+  // regression tests for check_for_aliasing
+  MatrixXd c(10,10);
+  c = 1.0 * MatrixXd::Ones(10,10) + c;
+  c = MatrixXd::Ones(10,10) * 1.0 + c;
+  c = c + MatrixXd::Ones(10,10) .cwiseProduct( MatrixXd::Zero(10,10) );
+  c = MatrixXd::Ones(10,10) * MatrixXd::Zero(10,10);
+}
+
 EIGEN_DECLARE_TEST(adjoint)
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -168,32 +196,6 @@ EIGEN_DECLARE_TEST(adjoint)
   // test a large static matrix only once
   CALL_SUBTEST_7( adjoint(Matrix<float, 100, 100>()) );
 
-#ifdef EIGEN_TEST_PART_13
-  {
-    MatrixXcf a(10,10), b(10,10);
-    VERIFY_RAISES_ASSERT(a = a.transpose());
-    VERIFY_RAISES_ASSERT(a = a.transpose() + b);
-    VERIFY_RAISES_ASSERT(a = b + a.transpose());
-    VERIFY_RAISES_ASSERT(a = a.conjugate().transpose());
-    VERIFY_RAISES_ASSERT(a = a.adjoint());
-    VERIFY_RAISES_ASSERT(a = a.adjoint() + b);
-    VERIFY_RAISES_ASSERT(a = b + a.adjoint());
-
-    // no assertion should be triggered for these cases:
-    a.transpose() = a.transpose();
-    a.transpose() += a.transpose();
-    a.transpose() += a.transpose() + b;
-    a.transpose() = a.adjoint();
-    a.transpose() += a.adjoint();
-    a.transpose() += a.adjoint() + b;
-
-    // regression tests for check_for_aliasing
-    MatrixXd c(10,10);
-    c = 1.0 * MatrixXd::Ones(10,10) + c;
-    c = MatrixXd::Ones(10,10) * 1.0 + c;
-    c = c + MatrixXd::Ones(10,10) .cwiseProduct( MatrixXd::Zero(10,10) );
-    c = MatrixXd::Ones(10,10) * MatrixXd::Zero(10,10);
-  }
-#endif
+  CALL_SUBTEST_13( adjoint_extra<0>() );
 }
 
diff --git a/test/array_reverse.cpp b/test/array_reverse.cpp
index 317b9a6e2..e23159def 100644
--- a/test/array_reverse.cpp
+++ b/test/array_reverse.cpp
@@ -123,6 +123,15 @@ template<typename MatrixType> void reverse(const MatrixType& m)
   VERIFY_IS_APPROX(x, m1(r, cols - 1 - c));
 }
 
+template<int>
+void array_reverse_extra()
+{
+  Vector4f x; x << 1, 2, 3, 4;
+  Vector4f y; y << 4, 3, 2, 1;
+  VERIFY(x.reverse()[1] == 3);
+  VERIFY(x.reverse() == y);
+}
+
 EIGEN_DECLARE_TEST(array_reverse)
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -136,10 +145,5 @@ EIGEN_DECLARE_TEST(array_reverse)
     CALL_SUBTEST_8( reverse(Matrix<float, 100, 100>()) );
     CALL_SUBTEST_9( reverse(Matrix<float,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
   }
-#ifdef EIGEN_TEST_PART_3
-  Vector4f x; x << 1, 2, 3, 4;
-  Vector4f y; y << 4, 3, 2, 1;
-  VERIFY(x.reverse()[1] == 3);
-  VERIFY(x.reverse() == y);
-#endif
+  CALL_SUBTEST_3( array_reverse_extra<0>() );
 }
diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp
index 5ad4090cd..85af603d8 100644
--- a/test/basicstuff.cpp
+++ b/test/basicstuff.cpp
@@ -194,7 +194,7 @@ template<typename MatrixType> void basicStuffComplex(const MatrixType& m)
   VERIFY(!static_cast<const MatrixType&>(cm).imag().isZero());
 }
 
-#ifdef EIGEN_TEST_PART_2
+template<int>
 void casting()
 {
   Matrix4f m = Matrix4f::Random(), m2;
@@ -203,7 +203,6 @@ void casting()
   m2 = m.cast<float>(); // check the specialization when NewType == Type
   VERIFY(m.isApprox(m2));
 }
-#endif
 
 template <typename Scalar>
 void fixedSizeMatrixConstruction()
@@ -290,5 +289,5 @@ EIGEN_DECLARE_TEST(basicstuff)
   CALL_SUBTEST_1(fixedSizeMatrixConstruction<long int>());
   CALL_SUBTEST_1(fixedSizeMatrixConstruction<std::ptrdiff_t>());
 
-  CALL_SUBTEST_2(casting());
+  CALL_SUBTEST_2(casting<0>());
 }
diff --git a/test/eigensolver_generic.cpp b/test/eigensolver_generic.cpp
index b41186819..e0e435151 100644
--- a/test/eigensolver_generic.cpp
+++ b/test/eigensolver_generic.cpp
@@ -100,6 +100,34 @@ template<typename MatrixType> void eigensolver_verify_assert(const MatrixType& m
   VERIFY_RAISES_ASSERT(eig.pseudoEigenvectors());
 }
 
+template<int>
+void eigensolver_generic_extra()
+{
+  {
+    // regression test for bug 793
+    MatrixXd a(3,3);
+    a << 0,  0,  1,
+        1,  1, 1,
+        1, 1e+200,  1;
+    Eigen::EigenSolver<MatrixXd> eig(a);
+    double scale = 1e-200; // scale to avoid overflow during the comparisons
+    VERIFY_IS_APPROX(a * eig.pseudoEigenvectors()*scale, eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix()*scale);
+    VERIFY_IS_APPROX(a * eig.eigenvectors()*scale, eig.eigenvectors() * eig.eigenvalues().asDiagonal()*scale);
+  }
+  {
+    // check a case where all eigenvalues are null.
+    MatrixXd a(2,2);
+    a << 1,  1,
+        -1, -1;
+    Eigen::EigenSolver<MatrixXd> eig(a);
+    VERIFY_IS_APPROX(eig.pseudoEigenvectors().squaredNorm(), 2.);
+    VERIFY_IS_APPROX((a * eig.pseudoEigenvectors()).norm()+1., 1.);
+    VERIFY_IS_APPROX((eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix()).norm()+1., 1.);
+    VERIFY_IS_APPROX((a * eig.eigenvectors()).norm()+1., 1.);
+    VERIFY_IS_APPROX((eig.eigenvectors() * eig.eigenvalues().asDiagonal()).norm()+1., 1.);
+  }
+}
+
 EIGEN_DECLARE_TEST(eigensolver_generic)
 {
   int s = 0;
@@ -135,31 +163,7 @@ EIGEN_DECLARE_TEST(eigensolver_generic)
   }
   );
   
-#ifdef EIGEN_TEST_PART_2
-  {
-    // regression test for bug 793
-    MatrixXd a(3,3);
-    a << 0,  0,  1,
-        1,  1, 1,
-        1, 1e+200,  1;
-    Eigen::EigenSolver<MatrixXd> eig(a);
-    double scale = 1e-200; // scale to avoid overflow during the comparisons
-    VERIFY_IS_APPROX(a * eig.pseudoEigenvectors()*scale, eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix()*scale);
-    VERIFY_IS_APPROX(a * eig.eigenvectors()*scale, eig.eigenvectors() * eig.eigenvalues().asDiagonal()*scale);
-  }
-  {
-    // check a case where all eigenvalues are null.
-    MatrixXd a(2,2);
-    a << 1,  1,
-        -1, -1;
-    Eigen::EigenSolver<MatrixXd> eig(a);
-    VERIFY_IS_APPROX(eig.pseudoEigenvectors().squaredNorm(), 2.);
-    VERIFY_IS_APPROX((a * eig.pseudoEigenvectors()).norm()+1., 1.);
-    VERIFY_IS_APPROX((eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix()).norm()+1., 1.);
-    VERIFY_IS_APPROX((a * eig.eigenvectors()).norm()+1., 1.);
-    VERIFY_IS_APPROX((eig.eigenvectors() * eig.eigenvalues().asDiagonal()).norm()+1., 1.);
-  }
-#endif
+  CALL_SUBTEST_2( eigensolver_generic_extra<0>() );
   
   TEST_SET_BUT_UNUSED_VARIABLE(s)
 }
diff --git a/test/incomplete_cholesky.cpp b/test/incomplete_cholesky.cpp
index 52235c2c2..68fe7d507 100644
--- a/test/incomplete_cholesky.cpp
+++ b/test/incomplete_cholesky.cpp
@@ -29,14 +29,10 @@ template<typename T, typename I> void test_incomplete_cholesky_T()
   CALL_SUBTEST( check_sparse_spd_solving(cg_illt_uplo_amd) );
 }
 
-EIGEN_DECLARE_TEST(incomplete_cholesky)
+template<int>
+void bug1150()
 {
-  CALL_SUBTEST_1(( test_incomplete_cholesky_T<double,int>() ));
-  CALL_SUBTEST_2(( test_incomplete_cholesky_T<std::complex<double>, int>() ));
-  CALL_SUBTEST_3(( test_incomplete_cholesky_T<double,long int>() ));
-
-#ifdef EIGEN_TEST_PART_1
-    // regression for bug 1150
+  // regression for bug 1150
   for(int N = 1; N<20; ++N)
   {
     Eigen::MatrixXd b( N, N );
@@ -61,5 +57,13 @@ EIGEN_DECLARE_TEST(incomplete_cholesky)
     VERIFY(solver.preconditioner().info() == Eigen::Success);
     VERIFY(solver.info() == Eigen::Success);
   }
-#endif
+}
+
+EIGEN_DECLARE_TEST(incomplete_cholesky)
+{
+  CALL_SUBTEST_1(( test_incomplete_cholesky_T<double,int>() ));
+  CALL_SUBTEST_2(( test_incomplete_cholesky_T<std::complex<double>, int>() ));
+  CALL_SUBTEST_3(( test_incomplete_cholesky_T<double,long int>() ));
+
+  CALL_SUBTEST_1(( bug1150<0>() ));
 }
diff --git a/test/integer_types.cpp b/test/integer_types.cpp
index 88573c568..e9e514f12 100644
--- a/test/integer_types.cpp
+++ b/test/integer_types.cpp
@@ -131,6 +131,17 @@ template<typename MatrixType> void integer_type_tests(const MatrixType& m)
   VERIFY_IS_APPROX((m1 * m2.transpose()) * m1, m1 * (m2.transpose() * m1));
 }
 
+template<int>
+void integer_types_extra()
+{
+  VERIFY_IS_EQUAL(internal::scalar_div_cost<int>::value, 8);
+  VERIFY_IS_EQUAL(internal::scalar_div_cost<unsigned int>::value, 8);
+  if(sizeof(long)>sizeof(int)) {
+    VERIFY(int(internal::scalar_div_cost<long>::value) > int(internal::scalar_div_cost<int>::value));
+    VERIFY(int(internal::scalar_div_cost<unsigned long>::value) > int(internal::scalar_div_cost<int>::value));
+  }
+}
+
 EIGEN_DECLARE_TEST(integer_types)
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -156,12 +167,5 @@ EIGEN_DECLARE_TEST(integer_types)
 
     CALL_SUBTEST_8( integer_type_tests(Matrix<unsigned long long, Dynamic, 5>(1, 5)) );
   }
-#ifdef EIGEN_TEST_PART_9
-  VERIFY_IS_EQUAL(internal::scalar_div_cost<int>::value, 8);
-  VERIFY_IS_EQUAL(internal::scalar_div_cost<unsigned int>::value, 8);
-  if(sizeof(long)>sizeof(int)) {
-    VERIFY(int(internal::scalar_div_cost<long>::value) > int(internal::scalar_div_cost<int>::value));
-    VERIFY(int(internal::scalar_div_cost<unsigned long>::value) > int(internal::scalar_div_cost<int>::value));
-  }
-#endif
+  CALL_SUBTEST_9( integer_types_extra<0>() );
 }
diff --git a/test/inverse.cpp b/test/inverse.cpp
index b989c7905..8754cb7e5 100644
--- a/test/inverse.cpp
+++ b/test/inverse.cpp
@@ -11,9 +11,59 @@
 #include "main.h"
 #include <Eigen/LU>
 
-template<typename MatrixType> void inverse(const MatrixType& m)
+template<typename MatrixType>
+void inverse_for_fixed_size(const MatrixType&, typename internal::enable_if<MatrixType::SizeAtCompileTime==Dynamic>::type* = 0)
+{
+}
+
+template<typename MatrixType>
+void inverse_for_fixed_size(const MatrixType& m1, typename internal::enable_if<MatrixType::SizeAtCompileTime!=Dynamic>::type* = 0)
 {
   using std::abs;
+
+  MatrixType m2, identity = MatrixType::Identity();
+
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Matrix<Scalar, MatrixType::ColsAtCompileTime, 1> VectorType;
+  
+  //computeInverseAndDetWithCheck tests
+  //First: an invertible matrix
+  bool invertible;
+  Scalar det;
+
+  m2.setZero();
+  m1.computeInverseAndDetWithCheck(m2, det, invertible);
+  VERIFY(invertible);
+  VERIFY_IS_APPROX(identity, m1*m2);
+  VERIFY_IS_APPROX(det, m1.determinant());
+
+  m2.setZero();
+  m1.computeInverseWithCheck(m2, invertible);
+  VERIFY(invertible);
+  VERIFY_IS_APPROX(identity, m1*m2);
+
+  //Second: a rank one matrix (not invertible, except for 1x1 matrices)
+  VectorType v3 = VectorType::Random();
+  MatrixType m3 = v3*v3.transpose(), m4;
+  m3.computeInverseAndDetWithCheck(m4, det, invertible);
+  VERIFY( m1.rows()==1 ? invertible : !invertible );
+  VERIFY_IS_MUCH_SMALLER_THAN(abs(det-m3.determinant()), RealScalar(1));
+  m3.computeInverseWithCheck(m4, invertible);
+  VERIFY( m1.rows()==1 ? invertible : !invertible );
+  
+  // check with submatrices
+  {
+    Matrix<Scalar, MatrixType::RowsAtCompileTime+1, MatrixType::RowsAtCompileTime+1, MatrixType::Options> m5;
+    m5.setRandom();
+    m5.topLeftCorner(m1.rows(),m1.rows()) = m1;
+    m2 = m5.template topLeftCorner<MatrixType::RowsAtCompileTime,MatrixType::ColsAtCompileTime>().inverse();
+    VERIFY_IS_APPROX( (m5.template topLeftCorner<MatrixType::RowsAtCompileTime,MatrixType::ColsAtCompileTime>()), m2.inverse() );
+  }
+}
+
+template<typename MatrixType> void inverse(const MatrixType& m)
+{
   /* this test covers the following files:
      Inverse.h
   */
@@ -39,44 +89,7 @@ template<typename MatrixType> void inverse(const MatrixType& m)
   // since for the general case we implement separately row-major and col-major, test that
   VERIFY_IS_APPROX(MatrixType(m1.transpose().inverse()), MatrixType(m1.inverse().transpose()));
 
-#if !defined(EIGEN_TEST_PART_5) && !defined(EIGEN_TEST_PART_6)
-  typedef typename NumTraits<Scalar>::Real RealScalar;
-  typedef Matrix<Scalar, MatrixType::ColsAtCompileTime, 1> VectorType;
-  
-  //computeInverseAndDetWithCheck tests
-  //First: an invertible matrix
-  bool invertible;
-  Scalar det;
-
-  m2.setZero();
-  m1.computeInverseAndDetWithCheck(m2, det, invertible);
-  VERIFY(invertible);
-  VERIFY_IS_APPROX(identity, m1*m2);
-  VERIFY_IS_APPROX(det, m1.determinant());
-
-  m2.setZero();
-  m1.computeInverseWithCheck(m2, invertible);
-  VERIFY(invertible);
-  VERIFY_IS_APPROX(identity, m1*m2);
-
-  //Second: a rank one matrix (not invertible, except for 1x1 matrices)
-  VectorType v3 = VectorType::Random(rows);
-  MatrixType m3 = v3*v3.transpose(), m4(rows,cols);
-  m3.computeInverseAndDetWithCheck(m4, det, invertible);
-  VERIFY( rows==1 ? invertible : !invertible );
-  VERIFY_IS_MUCH_SMALLER_THAN(abs(det-m3.determinant()), RealScalar(1));
-  m3.computeInverseWithCheck(m4, invertible);
-  VERIFY( rows==1 ? invertible : !invertible );
-  
-  // check with submatrices
-  {
-    Matrix<Scalar, MatrixType::RowsAtCompileTime+1, MatrixType::RowsAtCompileTime+1, MatrixType::Options> m5;
-    m5.setRandom();
-    m5.topLeftCorner(rows,rows) = m1;
-    m2 = m5.template topLeftCorner<MatrixType::RowsAtCompileTime,MatrixType::ColsAtCompileTime>().inverse();
-    VERIFY_IS_APPROX( (m5.template topLeftCorner<MatrixType::RowsAtCompileTime,MatrixType::ColsAtCompileTime>()), m2.inverse() );
-  }
-#endif
+  inverse_for_fixed_size(m1);
 
   // check in-place inversion
   if(MatrixType::RowsAtCompileTime>=2 && MatrixType::RowsAtCompileTime<=4)
diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp
index 4137b0251..46ee5162b 100644
--- a/test/linearstructure.cpp
+++ b/test/linearstructure.cpp
@@ -110,6 +110,19 @@ template<typename MatrixType> void real_complex(DenseIndex rows = MatrixType::Ro
   VERIFY(g_called && "matrix<complex> - real not properly optimized");
 }
 
+template<int>
+void linearstructure_overflow()
+{
+  // make sure that /=scalar and /scalar do not overflow
+  // rational: 1.0/4.94e-320 overflow, but m/4.94e-320 should not
+  Matrix4d m2, m3;
+  m3 = m2 =  Matrix4d::Random()*1e-20;
+  m2 = m2 / 4.9e-320;
+  VERIFY_IS_APPROX(m2.cwiseQuotient(m2), Matrix4d::Ones());
+  m3 /= 4.9e-320;
+  VERIFY_IS_APPROX(m3.cwiseQuotient(m3), Matrix4d::Ones());
+}
+
 EIGEN_DECLARE_TEST(linearstructure)
 {
   g_called = true;
@@ -130,19 +143,5 @@ EIGEN_DECLARE_TEST(linearstructure)
     CALL_SUBTEST_11( real_complex<MatrixXcf>(10,10) );
     CALL_SUBTEST_11( real_complex<ArrayXXcf>(10,10) );
   }
-  
-#ifdef EIGEN_TEST_PART_4
-  {
-    // make sure that /=scalar and /scalar do not overflow
-    // rational: 1.0/4.94e-320 overflow, but m/4.94e-320 should not
-    Matrix4d m2, m3;
-    m3 = m2 =  Matrix4d::Random()*1e-20;
-    m2 = m2 / 4.9e-320;
-    VERIFY_IS_APPROX(m2.cwiseQuotient(m2), Matrix4d::Ones());
-    m3 /= 4.9e-320;
-    VERIFY_IS_APPROX(m3.cwiseQuotient(m3), Matrix4d::Ones());
-    
-    
-  }
-#endif
+  CALL_SUBTEST_4( linearstructure_overflow<0>() );
 }
diff --git a/test/nullary.cpp b/test/nullary.cpp
index 9e6e6eaa0..12b9e122f 100644
--- a/test/nullary.cpp
+++ b/test/nullary.cpp
@@ -239,45 +239,28 @@ void testMatrixType(const MatrixType& m)
   VERIFY_IS_APPROX( A(i,j), s1 );
 }
 
-EIGEN_DECLARE_TEST(nullary)
+template<int>
+void bug79()
 {
-  CALL_SUBTEST_1( testMatrixType(Matrix2d()) );
-  CALL_SUBTEST_2( testMatrixType(MatrixXcf(internal::random<int>(1,300),internal::random<int>(1,300))) );
-  CALL_SUBTEST_3( testMatrixType(MatrixXf(internal::random<int>(1,300),internal::random<int>(1,300))) );
-  
-  for(int i = 0; i < g_repeat*10; i++) {
-    CALL_SUBTEST_4( testVectorType(VectorXd(internal::random<int>(1,30000))) );
-    CALL_SUBTEST_5( testVectorType(Vector4d()) );  // regression test for bug 232
-    CALL_SUBTEST_6( testVectorType(Vector3d()) );
-    CALL_SUBTEST_7( testVectorType(VectorXf(internal::random<int>(1,30000))) );
-    CALL_SUBTEST_8( testVectorType(Vector3f()) );
-    CALL_SUBTEST_8( testVectorType(Vector4f()) );
-    CALL_SUBTEST_8( testVectorType(Matrix<float,8,1>()) );
-    CALL_SUBTEST_8( testVectorType(Matrix<float,1,1>()) );
-
-    CALL_SUBTEST_9( testVectorType(VectorXi(internal::random<int>(1,10))) );
-    CALL_SUBTEST_9( testVectorType(VectorXi(internal::random<int>(9,300))) );
-    CALL_SUBTEST_9( testVectorType(Matrix<int,1,1>()) );
-  }
-
-#ifdef EIGEN_TEST_PART_6
   // Assignment of a RowVectorXd to a MatrixXd (regression test for bug #79).
   VERIFY( (MatrixXd(RowVectorXd::LinSpaced(3, 0, 1)) - RowVector3d(0, 0.5, 1)).norm() < std::numeric_limits<double>::epsilon() );
-#endif
+}
 
-#ifdef EIGEN_TEST_PART_9
+template<int>
+void nullary_overflow()
+{
   // Check possible overflow issue
-  {
-    int n = 60000;
-    ArrayXi a1(n), a2(n);
-    a1.setLinSpaced(n, 0, n-1);
-    for(int i=0; i<n; ++i)
-      a2(i) = i;
-    VERIFY_IS_APPROX(a1,a2);
-  }
-#endif
+  int n = 60000;
+  ArrayXi a1(n), a2(n);
+  a1.setLinSpaced(n, 0, n-1);
+  for(int i=0; i<n; ++i)
+    a2(i) = i;
+  VERIFY_IS_APPROX(a1,a2);
+}
 
-#ifdef EIGEN_TEST_PART_10
+template<int>
+void nullary_internal_logic()
+{
   // check some internal logic
   VERIFY((  internal::has_nullary_operator<internal::scalar_constant_op<double> >::value ));
   VERIFY(( !internal::has_unary_operator<internal::scalar_constant_op<double> >::value ));
@@ -318,5 +301,30 @@ EIGEN_DECLARE_TEST(nullary)
     VERIFY(( !internal::has_binary_operator<internal::linspaced_op<int,int> >::value ));
     VERIFY((  internal::functor_has_linear_access<internal::linspaced_op<int,int> >::ret ));
   }
-#endif
+}
+
+EIGEN_DECLARE_TEST(nullary)
+{
+  CALL_SUBTEST_1( testMatrixType(Matrix2d()) );
+  CALL_SUBTEST_2( testMatrixType(MatrixXcf(internal::random<int>(1,300),internal::random<int>(1,300))) );
+  CALL_SUBTEST_3( testMatrixType(MatrixXf(internal::random<int>(1,300),internal::random<int>(1,300))) );
+  
+  for(int i = 0; i < g_repeat*10; i++) {
+    CALL_SUBTEST_4( testVectorType(VectorXd(internal::random<int>(1,30000))) );
+    CALL_SUBTEST_5( testVectorType(Vector4d()) );  // regression test for bug 232
+    CALL_SUBTEST_6( testVectorType(Vector3d()) );
+    CALL_SUBTEST_7( testVectorType(VectorXf(internal::random<int>(1,30000))) );
+    CALL_SUBTEST_8( testVectorType(Vector3f()) );
+    CALL_SUBTEST_8( testVectorType(Vector4f()) );
+    CALL_SUBTEST_8( testVectorType(Matrix<float,8,1>()) );
+    CALL_SUBTEST_8( testVectorType(Matrix<float,1,1>()) );
+
+    CALL_SUBTEST_9( testVectorType(VectorXi(internal::random<int>(1,10))) );
+    CALL_SUBTEST_9( testVectorType(VectorXi(internal::random<int>(9,300))) );
+    CALL_SUBTEST_9( testVectorType(Matrix<int,1,1>()) );
+  }
+
+  CALL_SUBTEST_6( bug79<0>() );
+  CALL_SUBTEST_9( nullary_overflow<0>() );
+  CALL_SUBTEST_10( nullary_internal_logic<0>() );
 }
diff --git a/test/product_large.cpp b/test/product_large.cpp
index 1b6fec738..ae14b714c 100644
--- a/test/product_large.cpp
+++ b/test/product_large.cpp
@@ -30,21 +30,9 @@ void test_aliasing()
   x = z;
 }
 
-EIGEN_DECLARE_TEST(product_large)
+template<int>
+void product_large_regressions()
 {
-  for(int i = 0; i < g_repeat; i++) {
-    CALL_SUBTEST_1( product(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
-    CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
-    CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,10), internal::random<int>(1,10))) );
-
-    CALL_SUBTEST_3( product(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
-    CALL_SUBTEST_4( product(MatrixXcf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2), internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2))) );
-    CALL_SUBTEST_5( product(Matrix<float,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
-
-    CALL_SUBTEST_1( test_aliasing<float>() );
-  }
-
-#if defined EIGEN_TEST_PART_6
   {
     // test a specific issue in DiagonalProduct
     int N = 1000000;
@@ -97,7 +85,23 @@ EIGEN_DECLARE_TEST(product_large)
                 * (((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)) * ((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)));
     VERIFY_IS_APPROX(B,C);
   }
-#endif
+}
+
+EIGEN_DECLARE_TEST(product_large)
+{
+  for(int i = 0; i < g_repeat; i++) {
+    CALL_SUBTEST_1( product(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,10), internal::random<int>(1,10))) );
+
+    CALL_SUBTEST_3( product(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_4( product(MatrixXcf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2), internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2))) );
+    CALL_SUBTEST_5( product(Matrix<float,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+
+    CALL_SUBTEST_1( test_aliasing<float>() );
+  }
+
+  CALL_SUBTEST_6( product_large_regressions<0>() );
 
   // Regression test for bug 714:
 #if defined EIGEN_HAS_OPENMP
diff --git a/test/product_small.cpp b/test/product_small.cpp
index 16138631a..b8ce37d90 100644
--- a/test/product_small.cpp
+++ b/test/product_small.cpp
@@ -228,6 +228,36 @@ void bug_1311()
   VERIFY_IS_APPROX(res, A*b);
 }
 
+template<int>
+void product_small_regressions()
+{
+  {
+    // test compilation of (outer_product) * vector
+    Vector3f v = Vector3f::Random();
+    VERIFY_IS_APPROX( (v * v.transpose()) * v, (v * v.transpose()).eval() * v);
+  }
+  
+  {
+    // regression test for pull-request #93
+    Eigen::Matrix<double, 1, 1> A;  A.setRandom();
+    Eigen::Matrix<double, 18, 1> B; B.setRandom();
+    Eigen::Matrix<double, 1, 18> C; C.setRandom();
+    VERIFY_IS_APPROX(B * A.inverse(), B * A.inverse()[0]);
+    VERIFY_IS_APPROX(A.inverse() * C, A.inverse()[0] * C);
+  }
+
+  {
+    Eigen::Matrix<double, 10, 10> A, B, C;
+    A.setRandom();
+    C = A;
+    for(int k=0; k<79; ++k)
+      C = C * A;
+    B.noalias() = (((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)) * ((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)))
+                * (((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)) * ((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)));
+    VERIFY_IS_APPROX(B,C);
+  }
+}
+
 EIGEN_DECLARE_TEST(product_small)
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -263,31 +293,5 @@ EIGEN_DECLARE_TEST(product_small)
     CALL_SUBTEST_6( bug_1311<5>() );
   }
 
-#ifdef EIGEN_TEST_PART_6
-  {
-    // test compilation of (outer_product) * vector
-    Vector3f v = Vector3f::Random();
-    VERIFY_IS_APPROX( (v * v.transpose()) * v, (v * v.transpose()).eval() * v);
-  }
-  
-  {
-    // regression test for pull-request #93
-    Eigen::Matrix<double, 1, 1> A;  A.setRandom();
-    Eigen::Matrix<double, 18, 1> B; B.setRandom();
-    Eigen::Matrix<double, 1, 18> C; C.setRandom();
-    VERIFY_IS_APPROX(B * A.inverse(), B * A.inverse()[0]);
-    VERIFY_IS_APPROX(A.inverse() * C, A.inverse()[0] * C);
-  }
-
-  {
-    Eigen::Matrix<double, 10, 10> A, B, C;
-    A.setRandom();
-    C = A;
-    for(int k=0; k<79; ++k)
-      C = C * A;
-    B.noalias() = (((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)) * ((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)))
-                * (((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)) * ((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A))*((A*A)*(A*A)));
-    VERIFY_IS_APPROX(B,C);
-  }
-#endif
+  CALL_SUBTEST_6( product_small_regressions<0>() );
 }
diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp
index bfd61f69f..3691e8dad 100644
--- a/test/sparse_basic.cpp
+++ b/test/sparse_basic.cpp
@@ -640,6 +640,20 @@ void big_sparse_triplet(Index rows, Index cols, double density) {
   VERIFY_IS_APPROX(sum, m.sum());
 }
 
+template<int>
+void bug1105()
+{
+  // Regression test for bug 1105
+  int n = Eigen::internal::random<int>(200,600);
+  SparseMatrix<std::complex<double>,0, long> mat(n, n);
+  std::complex<double> val;
+
+  for(int i=0; i<n; ++i)
+  {
+    mat.coeffRef(i, i%(n/10)) = val;
+    VERIFY(mat.data().allocatedSize()<20*n);
+  }
+}
 
 EIGEN_DECLARE_TEST(sparse_basic)
 {
@@ -671,18 +685,5 @@ EIGEN_DECLARE_TEST(sparse_basic)
   CALL_SUBTEST_3((big_sparse_triplet<SparseMatrix<float, RowMajor, int> >(10000, 10000, 0.125)));
   CALL_SUBTEST_4((big_sparse_triplet<SparseMatrix<double, ColMajor, long int> >(10000, 10000, 0.125)));
 
-  // Regression test for bug 1105
-#ifdef EIGEN_TEST_PART_7
-  {
-    int n = Eigen::internal::random<int>(200,600);
-    SparseMatrix<std::complex<double>,0, long> mat(n, n);
-    std::complex<double> val;
-
-    for(int i=0; i<n; ++i)
-    {
-      mat.coeffRef(i, i%(n/10)) = val;
-      VERIFY(mat.data().allocatedSize()<20*n);
-    }
-  }
-#endif
+  CALL_SUBTEST_7( bug1105<0>() );
 }

From f124f0796533081cb7b061c4aa8667df6ed58863 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Tue, 17 Jul 2018 14:16:48 -0400
Subject: [PATCH 257/680] applying EIGEN_DECLARE_TEST to *gpu* tests

Also, a few minor fixes for GPU tests running in HIP mode.

1. Adding an include for hip/hip_runtime.h in the Macros.h file
   For HIP __host__ and __device__ are macros which are defined in hip headers.
   Their definitions need to be included before their use in the file.

2. Fixing the compile failure in TensorContractionGpu introduced by the commit to
   "Fuse computations into the Tensor contractions using output kernel"

3. Fixing a HIP/clang specific compile error by making the struct-member assignment explicit
---
 Eigen/src/Core/util/Macros.h                           |  6 ++++--
 Eigen/src/Core/util/Memory.h                           |  2 +-
 Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h         |  7 ++++---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h |  3 +--
 .../Eigen/CXX11/src/Tensor/TensorContractionGpu.h      | 10 +++++-----
 unsupported/test/cxx11_tensor_argmax_gpu.cu            |  2 +-
 unsupported/test/cxx11_tensor_cast_float16_gpu.cu      |  2 +-
 unsupported/test/cxx11_tensor_contract_gpu.cu          |  2 +-
 unsupported/test/cxx11_tensor_device.cu                |  2 +-
 unsupported/test/cxx11_tensor_gpu.cu                   |  2 +-
 unsupported/test/cxx11_tensor_of_float16_gpu.cu        |  2 +-
 unsupported/test/cxx11_tensor_random_gpu.cu            |  2 +-
 unsupported/test/cxx11_tensor_reduction_gpu.cu         |  2 +-
 unsupported/test/cxx11_tensor_scan_gpu.cu              |  2 +-
 14 files changed, 24 insertions(+), 22 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index fc4c0815c..adf25ee9b 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -395,8 +395,10 @@
   // Means the compiler is HIPCC (analogous to EIGEN_CUDACC, but for HIP)
   #define EIGEN_HIPCC __HIPCC__
 
-  // We need hip_common.h here because __HIP_DEVICE_COMPILE__ is defined in this header.
-  #include <hip/hip_common.h>
+  // We need to include hip_runtime.h here because it pulls in
+  // ++ hip_common.h which contains the define for  __HIP_DEVICE_COMPILE__ 
+  // ++ host_defines.h which contains the defines for the __host__ and __device__ macros
+  #include <hip/hip_runtime.h>
 
   #if defined(__HIP_DEVICE_COMPILE__)
     // analogous to EIGEN_CUDA_ARCH, but for HIP
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 956f64d1d..8dbd4d93b 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -580,7 +580,7 @@ template<typename T> struct smart_memmove_helper<T,false> {
 
 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
 // to the appropriate stack allocation function
-#if ! defined EIGEN_ALLOCA && ! defined EIGEN_CUDA_ARCH
+#if ! defined EIGEN_ALLOCA && ! defined EIGEN_GPU_COMPILE_PHASE
   #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
     #define EIGEN_ALLOCA alloca
   #elif EIGEN_COMP_MSVC
diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
index fbc1ee2f6..2d1a82c57 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -610,6 +610,7 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
   static inline bool extract_kernel(MatrixType& mat, Ref<VectorType> res, Ref<VectorType> representative)
   {
     EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
     Index i0;
     // Find non-zero column i0 (by construction, there must exist a non zero coefficient on the diagonal):
     mat.diagonal().cwiseAbs().maxCoeff(&i0);
@@ -620,8 +621,8 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
     VectorType c0, c1;
     n0 = (c0 = representative.cross(mat.col((i0+1)%3))).squaredNorm();
     n1 = (c1 = representative.cross(mat.col((i0+2)%3))).squaredNorm();
-    if(n0>n1) res = c0/std::sqrt(n0);
-    else      res = c1/std::sqrt(n1);
+    if(n0>n1) res = c0/sqrt(n0);
+    else      res = c1/sqrt(n1);
 
     return true;
   }
@@ -723,7 +724,7 @@ struct direct_selfadjoint_eigenvalues<SolverType,2,false>
   EIGEN_DEVICE_FUNC
   static inline void computeRoots(const MatrixType& m, VectorType& roots)
   {
-    using std::sqrt;
+    EIGEN_USING_STD_MATH(sqrt);
     const Scalar t0 = Scalar(0.5) * sqrt( numext::abs2(m(0,0)-m(1,1)) + Scalar(4)*numext::abs2(m(1,0)));
     const Scalar t1 = Scalar(0.5) * (m(0,0) + m(1,1));
     roots(0) = t1 - t0;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 0fbffa34c..c7c443a59 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -449,8 +449,7 @@ struct TensorContractionEvaluatorBase
     // tensor dimensions (i, j) into the original tensor dimensions.
     // TODO(ezhulenev): Add parameters required to infer output tensor index for
     // more complex contractions than 2x2 on internal dimension.
-    m_tensor_contraction_params = {
-        /**swapped_arguments=*/static_cast<int>(Layout) == RowMajor};
+    m_tensor_contraction_params.swapped_arguments = static_cast<int>(Layout) == RowMajor;
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index a4f92ee44..b9956cd43 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -1215,16 +1215,16 @@ EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
 }
 
 
-template<typename Indices, typename LeftArgType, typename RightArgType>
-struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> :
-    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> > {
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> > {
 
   typedef GpuDevice Device;
 
-  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
   typedef TensorContractionEvaluatorBase<Self> Base;
 
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
diff --git a/unsupported/test/cxx11_tensor_argmax_gpu.cu b/unsupported/test/cxx11_tensor_argmax_gpu.cu
index f6c3a9908..79f4066e9 100644
--- a/unsupported/test/cxx11_tensor_argmax_gpu.cu
+++ b/unsupported/test/cxx11_tensor_argmax_gpu.cu
@@ -242,7 +242,7 @@ void test_gpu_argmin_dim()
   }
 }
 
-void test_cxx11_tensor_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_argmax_gpu)
 {
   CALL_SUBTEST_1(test_gpu_simple_argmax<RowMajor>());
   CALL_SUBTEST_1(test_gpu_simple_argmax<ColMajor>());
diff --git a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
index 0a37a555c..97923d15f 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
@@ -72,7 +72,7 @@ void test_fallback_conversion() {
 }
 
 
-void test_cxx11_tensor_cast_float16_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_cast_float16_gpu)
 {
   CALL_SUBTEST(test_gpu_conversion());
   CALL_SUBTEST(test_fallback_conversion());
diff --git a/unsupported/test/cxx11_tensor_contract_gpu.cu b/unsupported/test/cxx11_tensor_contract_gpu.cu
index cb1416478..575bdc1f9 100644
--- a/unsupported/test/cxx11_tensor_contract_gpu.cu
+++ b/unsupported/test/cxx11_tensor_contract_gpu.cu
@@ -193,7 +193,7 @@ void test_gpu_contraction_sizes() {
   }
 }
 
-void test_cxx11_tensor_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_contract_gpu)
 {
   CALL_SUBTEST_1(test_gpu_contraction<ColMajor>(128, 128, 128));
   CALL_SUBTEST_1(test_gpu_contraction<RowMajor>(128, 128, 128));
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index cd9ba3ecd..c9f78d2d3 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -389,7 +389,7 @@ void test_gpu() {
 }
 
 
-void test_cxx11_tensor_device()
+EIGEN_DECLARE_TEST(cxx11_tensor_device)
 {
   CALL_SUBTEST_1(test_cpu());
   CALL_SUBTEST_2(test_gpu());
diff --git a/unsupported/test/cxx11_tensor_gpu.cu b/unsupported/test/cxx11_tensor_gpu.cu
index faaac73cf..14fc0bd04 100644
--- a/unsupported/test/cxx11_tensor_gpu.cu
+++ b/unsupported/test/cxx11_tensor_gpu.cu
@@ -1472,7 +1472,7 @@ void test_gpu_gamma_sample_der_alpha()
   gpuFree(d_out);
 }
 
-void test_cxx11_tensor_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_gpu)
 {
   CALL_SUBTEST_1(test_gpu_nullary());
   CALL_SUBTEST_1(test_gpu_elementwise_small());
diff --git a/unsupported/test/cxx11_tensor_of_float16_gpu.cu b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
index 1f1ec26c2..4d74e6138 100644
--- a/unsupported/test/cxx11_tensor_of_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
@@ -479,7 +479,7 @@ void test_gpu_forced_evals() {
 #endif
 
 
-void test_cxx11_tensor_of_float16_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_of_float16_gpu)
 {
   CALL_SUBTEST_1(test_gpu_numext<void>());
 
diff --git a/unsupported/test/cxx11_tensor_random_gpu.cu b/unsupported/test/cxx11_tensor_random_gpu.cu
index 262182d30..090986ebc 100644
--- a/unsupported/test/cxx11_tensor_random_gpu.cu
+++ b/unsupported/test/cxx11_tensor_random_gpu.cu
@@ -78,7 +78,7 @@ static void test_complex()
 }
 
 
-void test_cxx11_tensor_random_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_random_gpu)
 {
   CALL_SUBTEST(test_gpu_random_uniform());
   CALL_SUBTEST(test_gpu_random_normal());
diff --git a/unsupported/test/cxx11_tensor_reduction_gpu.cu b/unsupported/test/cxx11_tensor_reduction_gpu.cu
index 7b8ac2309..122ac946b 100644
--- a/unsupported/test/cxx11_tensor_reduction_gpu.cu
+++ b/unsupported/test/cxx11_tensor_reduction_gpu.cu
@@ -134,7 +134,7 @@ static void test_last_dim_reductions() {
 }
 
 
-void test_cxx11_tensor_reduction_gpu() {
+EIGEN_DECLARE_TEST(cxx11_tensor_reduction_gpu) {
   CALL_SUBTEST_1((test_full_reductions<float, ColMajor>()));
   CALL_SUBTEST_1((test_full_reductions<double, ColMajor>()));
   CALL_SUBTEST_2((test_full_reductions<float, RowMajor>()));
diff --git a/unsupported/test/cxx11_tensor_scan_gpu.cu b/unsupported/test/cxx11_tensor_scan_gpu.cu
index f3e773db5..770a144f1 100644
--- a/unsupported/test/cxx11_tensor_scan_gpu.cu
+++ b/unsupported/test/cxx11_tensor_scan_gpu.cu
@@ -71,7 +71,7 @@ void test_gpu_cumsum(int m_size, int k_size, int n_size)
 }
 
 
-void test_cxx11_tensor_scan_gpu()
+EIGEN_DECLARE_TEST(cxx11_tensor_scan_gpu)
 {
   CALL_SUBTEST_1(test_gpu_cumsum<ColMajor>(128, 128, 128));
   CALL_SUBTEST_2(test_gpu_cumsum<RowMajor>(128, 128, 128));

From c95aacab90e9d8bb9f9e082395b3b843a530fa41 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 17 Jul 2018 14:09:37 -0700
Subject: [PATCH 258/680] Fix TensorContractionOp evaluators for GPU and SYCL

---
 .../Eigen/CXX11/src/Tensor/TensorContractionGpu.h   | 10 +++++-----
 .../Eigen/CXX11/src/Tensor/TensorContractionSycl.h  | 13 ++++++++-----
 2 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index b9956cd43..6d3aa24c8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -505,9 +505,9 @@ template<typename Scalar, typename Index, typename LhsMapper,
 __global__ void
 #if defined(EIGEN_HIPCC)
 __launch_bounds__(512, 1)
-#else  
+#else
 __launch_bounds__(512)
-#endif  
+#endif
 EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
                        const OutputMapper output,
                        const Index m_size, const Index n_size, const Index k_size) {
@@ -698,7 +698,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
 
 #undef prefetch_lhs
 #undef add_vals
-  
+
   Index horiz_base = threadIdx.y*4+base_n;
   if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
     for (int i = 0; i < 4; i++) {
@@ -1137,7 +1137,7 @@ template<typename Index, typename LhsMapper,
 __global__ void
 #if defined(EIGEN_HIPCC)
 __launch_bounds__(256, 1)
-#else  
+#else
 __launch_bounds__(256)
 #endif
 EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
@@ -1184,7 +1184,7 @@ template<typename Index, typename LhsMapper,
 __global__ void
 #if defined(EIGEN_HIPCC)
 __launch_bounds__(256, 1)
-#else  
+#else
 __launch_bounds__(256)
 #endif
 EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
index e6840bc87..35f931c53 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h
@@ -23,15 +23,18 @@
 namespace Eigen {
 
 template <typename Index, typename LhsScalar, typename RhsScalar,bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered> struct LaunchSyclKernels;
-template<typename Indices, typename LeftArgType, typename RightArgType>
-struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, const Eigen::SyclDevice> :
-    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, const Eigen::SyclDevice> > {
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, const Eigen::SyclDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, const Eigen::SyclDevice> > {
+
+  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
+                "SYCL tensor contraction does not support output kernels.");
 
   typedef const Eigen::SyclDevice Device;
 
-  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
   typedef TensorContractionEvaluatorBase<Self> Base;
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;

From 3a2dc3869ed77cb6eca2beaad5c0af7d5134f15c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 02:26:43 -0700
Subject: [PATCH 259/680] Fix weird issue with MSVC 2013

---
 test/mapped_matrix.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp
index e1106ba42..0ea136ae6 100644
--- a/test/mapped_matrix.cpp
+++ b/test/mapped_matrix.cpp
@@ -62,8 +62,9 @@ template<typename MatrixType> void map_class_matrix(const MatrixType& m)
   for(int i = 0; i < size; i++) array2[i] = Scalar(1);
   // array3unaligned -> unaligned pointer to heap
   Scalar* array3 = new Scalar[size+1];
-  for(int i = 0; i < size+1; i++) array3[i] = Scalar(1);
-  Scalar* array3unaligned = internal::UIntPtr(array3)%EIGEN_MAX_ALIGN_BYTES == 0 ? array3+1 : array3;
+  Index sizep1 = size + 1; // <- without this temporary MSVC 2103 generates bad code
+  for(Index i = 0; i < sizep1; i++) array3[i] = Scalar(1);
+  Scalar* array3unaligned = (internal::UIntPtr(array3)%EIGEN_MAX_ALIGN_BYTES) == 0 ? array3+1 : array3;
   Scalar array4[256];
   if(size<=256)
     for(int i = 0; i < size; i++) array4[i] = Scalar(1);

From 000840cae0568da3c9148ddfe18b2af5b3e24067 Mon Sep 17 00:00:00 2001
From: Viktor Csomor <viktor.csomor@gmail.com>
Date: Wed, 7 Feb 2018 19:10:54 +0100
Subject: [PATCH 260/680] Added a move constructor and move assignment operator
 to Tensor and wrote some tests.

---
 unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 15 ++++
 unsupported/test/CMakeLists.txt             |  1 +
 unsupported/test/cxx11_tensor_move.cpp      | 81 +++++++++++++++++++++
 3 files changed, 97 insertions(+)
 create mode 100644 unsupported/test/cxx11_tensor_move.cpp

diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
index 4fd96448f..e3f6e37f0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
@@ -398,6 +398,21 @@ class Tensor : public TensorBase<Tensor<Scalar_, NumIndices_, Options_, IndexTyp
       internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
     }
 
+    #if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor(Self&& other)
+      : Tensor()
+    {
+      m_storage.swap(other.m_storage);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor& operator=(Self&& other)
+    {
+      m_storage.swap(other.m_storage);
+      return *this;
+    }
+    #endif
+
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Tensor& operator=(const Tensor& other)
     {
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 73befbca1..0fc864eca 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -225,6 +225,7 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_ifft)
   ei_add_test(cxx11_tensor_scan)
   ei_add_test(cxx11_tensor_trace)
+  ei_add_test(cxx11_tensor_move)
 
 endif()
 
diff --git a/unsupported/test/cxx11_tensor_move.cpp b/unsupported/test/cxx11_tensor_move.cpp
new file mode 100644
index 000000000..0ab2b7786
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_move.cpp
@@ -0,0 +1,81 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Viktor Csomor <viktor.csomor@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+#include <utility>
+
+using Eigen::Tensor;
+using Eigen::RowMajor;
+
+static void calc_indices(int i, int& x, int& y, int& z)
+{
+  x = i / 4;
+  y = (i % 4) / 2;
+  z = i % 2;
+}
+
+static void test_move()
+{
+  int x;
+  int y;
+  int z;
+
+  Tensor<int,3> tensor1(2, 2, 2);
+  Tensor<int,3,RowMajor> tensor2(2, 2, 2);
+
+  for (int i = 0; i < 8; i++)
+  {
+    calc_indices(i, x, y, z);
+    tensor1(x,y,z) = i;
+    tensor2(x,y,z) = 2 * i;
+  }
+
+  // Invokes the move constructor.
+  Tensor<int,3> moved_tensor1 = std::move(tensor1);
+  Tensor<int,3,RowMajor> moved_tensor2 = std::move(tensor2);
+
+  VERIFY_IS_EQUAL(tensor1.size(), 0);
+  VERIFY_IS_EQUAL(tensor2.size(), 0);
+
+  for (int i = 0; i < 8; i++)
+  {
+    calc_indices(i, x, y, z);
+    VERIFY_IS_EQUAL(moved_tensor1(x,y,z), i);
+    VERIFY_IS_EQUAL(moved_tensor2(x,y,z), 2 * i);
+  }
+
+  Tensor<int,3> moved_tensor3(2,2,2);
+  Tensor<int,3,RowMajor> moved_tensor4(2,2,2);
+
+  moved_tensor3.setZero();
+  moved_tensor4.setZero();
+
+  // Invokes the move assignment operator.
+  moved_tensor3 = std::move(moved_tensor1);
+  moved_tensor4 = std::move(moved_tensor2);
+
+  VERIFY_IS_EQUAL(moved_tensor1.size(), 8);
+  VERIFY_IS_EQUAL(moved_tensor2.size(), 8);
+
+  for (int i = 0; i < 8; i++)
+  {
+    calc_indices(i, x, y, z);
+    VERIFY_IS_EQUAL(moved_tensor1(x,y,z), 0);
+    VERIFY_IS_EQUAL(moved_tensor2(x,y,z), 0);
+    VERIFY_IS_EQUAL(moved_tensor3(x,y,z), i);
+    VERIFY_IS_EQUAL(moved_tensor4(x,y,z), 2 * i);
+  }
+}
+
+EIGEN_DECLARE_TEST(cxx11_tensor_move)
+{
+  CALL_SUBTEST(test_move());
+}

From f4a6863c75a14e72bead8b844000ee5239e95f70 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Feb 2018 16:43:49 +0100
Subject: [PATCH 261/680] Fix typo

---
 Eigen/src/Geometry/Scaling.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Geometry/Scaling.h b/Eigen/src/Geometry/Scaling.h
index c1899a029..8d9acf252 100755
--- a/Eigen/src/Geometry/Scaling.h
+++ b/Eigen/src/Geometry/Scaling.h
@@ -77,10 +77,10 @@ public:
   /** Concatenates a uniform scaling and an affine transformation */
   template<int Dim, int Mode, int Options>
   inline typename
-	internal::uniformscaling_times_affine_returntype <Scalar,Dim,Mode>::type
+	internal::uniformscaling_times_affine_returntype<Scalar,Dim,Mode>::type
 	operator* (const Transform<Scalar, Dim, Mode, Options>& t) const
   {
-    typename internal::uniformscaling_times_affine_returntype <Scalar,Dim,Mode> res = t;
+    typename internal::uniformscaling_times_affine_returntype<Scalar,Dim,Mode>::type res = t;
     res.prescale(factor());
     return res;
   }

From 12efc7d41b80259b996be5781bf596c249c90d3f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Feb 2018 16:45:49 +0100
Subject: [PATCH 262/680] Fix linear indexing in generic block evaluation.

---
 Eigen/src/Core/CoreEvaluators.h | 30 ++++++++++++++++++++++--------
 test/block.cpp                  | 11 +++++++++--
 2 files changed, 31 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index d1454117a..8419798c1 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1079,7 +1079,8 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block)
     : m_argImpl(block.nestedExpression()), 
       m_startRow(block.startRow()), 
-      m_startCol(block.startCol()) 
+      m_startCol(block.startCol()),
+      m_linear_offset(InnerPanel?(XprType::IsRowMajor ? block.startRow()*block.cols() : block.startCol()*block.rows()):0)
   { }
  
   typedef typename XprType::Scalar Scalar;
@@ -1098,7 +1099,10 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   CoeffReturnType coeff(Index index) const
   { 
-    return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+    if (InnerPanel)
+      return m_argImpl.coeff(m_linear_offset.value() + index); 
+    else
+      return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -1110,7 +1114,10 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   Scalar& coeffRef(Index index)
   { 
-    return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+    if (InnerPanel)
+      return m_argImpl.coeffRef(m_linear_offset.value() + index); 
+    else
+      return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
   }
  
   template<int LoadMode, typename PacketType>
@@ -1124,8 +1131,11 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_STRONG_INLINE
   PacketType packet(Index index) const 
   { 
-    return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
-                                       RowsAtCompileTime == 1 ? index : 0);
+    if (InnerPanel)
+      return m_argImpl.template packet<LoadMode,PacketType>(m_linear_offset.value() + index);
+    else
+      return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                         RowsAtCompileTime == 1 ? index : 0);
   }
   
   template<int StoreMode, typename PacketType>
@@ -1139,15 +1149,19 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketType& x) 
   {
-    return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
-                                             RowsAtCompileTime == 1 ? index : 0,
-                                             x);
+    if (InnerPanel)
+      return m_argImpl.template writePacket<StoreMode,PacketType>(m_linear_offset.value() + index, x);
+    else
+      return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                              RowsAtCompileTime == 1 ? index : 0,
+                                              x);
   }
  
 protected:
   evaluator<ArgType> m_argImpl;
   const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
   const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+  const variable_if_dynamic<Index, InnerPanel ? Dynamic : 0> m_linear_offset;
 };
 
 // TODO: This evaluator does not actually use the child evaluator; 
diff --git a/test/block.cpp b/test/block.cpp
index d61059874..8c4dd87be 100644
--- a/test/block.cpp
+++ b/test/block.cpp
@@ -44,7 +44,7 @@ template<typename MatrixType> void block(const MatrixType& m)
   typedef typename MatrixType::RealScalar RealScalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   typedef Matrix<Scalar, 1, MatrixType::ColsAtCompileTime> RowVectorType;
-  typedef Matrix<Scalar, Dynamic, Dynamic> DynamicMatrixType;
+  typedef Matrix<Scalar, Dynamic, Dynamic, MatrixType::IsRowMajor?RowMajor:ColMajor> DynamicMatrixType;
   typedef Matrix<Scalar, Dynamic, 1> DynamicVectorType;
   
   Index rows = m.rows();
@@ -142,7 +142,7 @@ template<typename MatrixType> void block(const MatrixType& m)
   VERIFY(numext::real(ones.col(c1).dot(ones.col(c2))) == RealScalar(rows));
   VERIFY(numext::real(ones.row(r1).dot(ones.row(r2))) == RealScalar(cols));
   
-  // chekc that linear acccessors works on blocks
+  // check that linear acccessors works on blocks
   m1 = m1_copy;
   if((MatrixType::Flags&RowMajorBit)==0)
     VERIFY_IS_EQUAL(m1.leftCols(c1).coeff(r1+c1*rows), m1(r1,c1));
@@ -166,6 +166,13 @@ template<typename MatrixType> void block(const MatrixType& m)
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).transpose().col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
   VERIFY_IS_APPROX( ((m1+m2).transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
 
+  VERIFY_IS_APPROX( (m1*1).topRows(r1),  m1.topRows(r1) );
+  VERIFY_IS_APPROX( (m1*1).leftCols(c1), m1.leftCols(c1) );
+  VERIFY_IS_APPROX( (m1*1).transpose().topRows(c1), m1.transpose().topRows(c1) );
+  VERIFY_IS_APPROX( (m1*1).transpose().leftCols(r1), m1.transpose().leftCols(r1) );
+  VERIFY_IS_APPROX( (m1*1).transpose().middleRows(c1,c2-c1+1), m1.transpose().middleRows(c1,c2-c1+1) );
+  VERIFY_IS_APPROX( (m1*1).transpose().middleCols(r1,r2-r1+1), m1.transpose().middleCols(r1,r2-r1+1) );
+
   // evaluation into plain matrices from expressions with direct access (stress MapBase)
   DynamicMatrixType dm;
   DynamicVectorType dv;

From 5deeb19e7bb19c67abeac0a6cfa26ad3d14e215b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 9 Feb 2018 16:52:35 +0100
Subject: [PATCH 263/680] bug #1517: fix triangular product with unit diagonal
 and nested scaling factor: (s*A).triangularView<UpperUnit>()*B

---
 .../Core/products/TriangularMatrixMatrix.h    | 24 ++++++++++++++--
 .../Core/products/TriangularMatrixVector.h    | 22 ++++++++++++---
 test/product_trmm.cpp                         | 28 +++++++++++++------
 3 files changed, 59 insertions(+), 15 deletions(-)

diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 539b6c0c6..f784507e7 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -400,7 +400,9 @@ struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
 {
   template<typename Dest> static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Scalar     Scalar;
+    typedef typename Lhs::Scalar  LhsScalar;
+    typedef typename Rhs::Scalar  RhsScalar;
+    typedef typename Dest::Scalar Scalar;
     
     typedef internal::blas_traits<Lhs> LhsBlasTraits;
     typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
@@ -412,8 +414,9 @@ struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
     typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
     typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
 
-    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
-                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(a_lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(a_rhs);
+    Scalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
 
     typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
               Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType;
@@ -438,6 +441,21 @@ struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
         &dst.coeffRef(0,0), dst.outerStride(),    // result info
         actualAlpha, blocking
       );
+
+    // Apply correction if the diagonal is unit and a scalar factor was nested:
+    if ((Mode&UnitDiag)==UnitDiag)
+    {
+      if (LhsIsTriangular && lhs_alpha!=LhsScalar(1))
+      {
+        Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+        dst.topRows(diagSize) -= ((lhs_alpha-LhsScalar(1))*a_rhs).topRows(diagSize);
+      }
+      else if ((!LhsIsTriangular) && rhs_alpha!=RhsScalar(1))
+      {
+        Index diagSize = (std::min)(rhs.rows(),rhs.cols());
+        dst.leftCols(diagSize) -= (rhs_alpha-RhsScalar(1))*a_lhs.leftCols(diagSize);
+      }
+    }
   }
 };
 
diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h
index 4b292e74d..76bfa159c 100644
--- a/Eigen/src/Core/products/TriangularMatrixVector.h
+++ b/Eigen/src/Core/products/TriangularMatrixVector.h
@@ -221,8 +221,9 @@ template<int Mode> struct trmv_selector<Mode,ColMajor>
     typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
     typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
-                                  * RhsBlasTraits::extractScalarFactor(rhs);
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs);
+    ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
@@ -274,6 +275,12 @@ template<int Mode> struct trmv_selector<Mode,ColMajor>
       else
         dest = MappedDest(actualDestPtr, dest.size());
     }
+
+    if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) )
+    {
+      Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+      dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize);
+    }
   }
 };
 
@@ -295,8 +302,9 @@ template<int Mode> struct trmv_selector<Mode,RowMajor>
     typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
     typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
-                                  * RhsBlasTraits::extractScalarFactor(rhs);
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs);
+    ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
 
     enum {
       DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
@@ -326,6 +334,12 @@ template<int Mode> struct trmv_selector<Mode,RowMajor>
             actualRhsPtr,1,
             dest.data(),dest.innerStride(),
             actualAlpha);
+
+    if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) )
+    {
+      Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+      dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize);
+    }
   }
 };
 
diff --git a/test/product_trmm.cpp b/test/product_trmm.cpp
index 12e554410..e08d9f39f 100644
--- a/test/product_trmm.cpp
+++ b/test/product_trmm.cpp
@@ -29,7 +29,7 @@ void trmm(int rows=get_random_size<Scalar>(),
   typedef Matrix<Scalar,Dynamic,OtherCols,OtherCols==1?ColMajor:ResOrder> ResXS;
   typedef Matrix<Scalar,OtherCols,Dynamic,OtherCols==1?RowMajor:ResOrder> ResSX;
 
-  TriMatrix  mat(rows,cols), tri(rows,cols), triTr(cols,rows);
+  TriMatrix  mat(rows,cols), tri(rows,cols), triTr(cols,rows), s1tri(rows,cols), s1triTr(cols,rows);
   
   OnTheRight  ge_right(cols,otherCols);
   OnTheLeft   ge_left(otherCols,rows);
@@ -42,6 +42,8 @@ void trmm(int rows=get_random_size<Scalar>(),
   mat.setRandom();
   tri = mat.template triangularView<Mode>();
   triTr = mat.transpose().template triangularView<Mode>();
+  s1tri = (s1*mat).template triangularView<Mode>();
+  s1triTr = (s1*mat).transpose().template triangularView<Mode>();
   ge_right.setRandom();
   ge_left.setRandom();
 
@@ -51,19 +53,29 @@ void trmm(int rows=get_random_size<Scalar>(),
   VERIFY_IS_APPROX( ge_xs.noalias() = mat.template triangularView<Mode>() * ge_right, tri * ge_right);
   VERIFY_IS_APPROX( ge_sx.noalias() = ge_left * mat.template triangularView<Mode>(), ge_left * tri);
 
-  VERIFY_IS_APPROX( ge_xs.noalias() = (s1*mat.adjoint()).template triangularView<Mode>() * (s2*ge_left.transpose()), s1*triTr.conjugate() * (s2*ge_left.transpose()));
-  VERIFY_IS_APPROX( ge_sx.noalias() = ge_right.transpose() * mat.adjoint().template triangularView<Mode>(), ge_right.transpose() * triTr.conjugate());
+  if((Mode&UnitDiag)==0)
+    VERIFY_IS_APPROX( ge_xs.noalias() = (s1*mat.adjoint()).template triangularView<Mode>() * (s2*ge_left.transpose()), s1*triTr.conjugate() * (s2*ge_left.transpose()));
   
-  VERIFY_IS_APPROX( ge_xs.noalias() = (s1*mat.adjoint()).template triangularView<Mode>() * (s2*ge_left.adjoint()), s1*triTr.conjugate() * (s2*ge_left.adjoint()));
-  VERIFY_IS_APPROX( ge_sx.noalias() = ge_right.adjoint() * mat.adjoint().template triangularView<Mode>(), ge_right.adjoint() * triTr.conjugate());
+  VERIFY_IS_APPROX( ge_xs.noalias() = (s1*mat.transpose()).template triangularView<Mode>() * (s2*ge_left.transpose()), s1triTr * (s2*ge_left.transpose()));
+  VERIFY_IS_APPROX( ge_sx.noalias() = (s2*ge_left) * (s1*mat).template triangularView<Mode>(), (s2*ge_left)*s1tri);
 
+  VERIFY_IS_APPROX( ge_sx.noalias() = ge_right.transpose() * mat.adjoint().template triangularView<Mode>(), ge_right.transpose() * triTr.conjugate());
+  VERIFY_IS_APPROX( ge_sx.noalias() = ge_right.adjoint() * mat.adjoint().template triangularView<Mode>(), ge_right.adjoint() * triTr.conjugate());
+  
   ge_xs_save = ge_xs;
-  VERIFY_IS_APPROX( (ge_xs_save + s1*triTr.conjugate() * (s2*ge_left.adjoint())).eval(), ge_xs.noalias() += (s1*mat.adjoint()).template triangularView<Mode>() * (s2*ge_left.adjoint()) );
+  if((Mode&UnitDiag)==0)
+    VERIFY_IS_APPROX( (ge_xs_save + s1*triTr.conjugate() * (s2*ge_left.adjoint())).eval(), ge_xs.noalias() += (s1*mat.adjoint()).template triangularView<Mode>() * (s2*ge_left.adjoint()) );
+  ge_xs_save = ge_xs;
+  VERIFY_IS_APPROX( (ge_xs_save + s1triTr * (s2*ge_left.adjoint())).eval(), ge_xs.noalias() += (s1*mat.transpose()).template triangularView<Mode>() * (s2*ge_left.adjoint()) );
   ge_sx.setRandom();
   ge_sx_save = ge_sx;
-  VERIFY_IS_APPROX( ge_sx_save - (ge_right.adjoint() * (-s1 * triTr).conjugate()).eval(), ge_sx.noalias() -= (ge_right.adjoint() * (-s1 * mat).adjoint().template triangularView<Mode>()).eval());
+  if((Mode&UnitDiag)==0)
+    VERIFY_IS_APPROX( ge_sx_save - (ge_right.adjoint() * (-s1 * triTr).conjugate()).eval(), ge_sx.noalias() -= (ge_right.adjoint() * (-s1 * mat).adjoint().template triangularView<Mode>()).eval());
   
-  VERIFY_IS_APPROX( ge_xs = (s1*mat).adjoint().template triangularView<Mode>() * ge_left.adjoint(), numext::conj(s1) * triTr.conjugate() * ge_left.adjoint());
+  if((Mode&UnitDiag)==0)
+    VERIFY_IS_APPROX( ge_xs = (s1*mat).adjoint().template triangularView<Mode>() * ge_left.adjoint(), numext::conj(s1) * triTr.conjugate() * ge_left.adjoint());
+  VERIFY_IS_APPROX( ge_xs = (s1*mat).transpose().template triangularView<Mode>() * ge_left.adjoint(), s1triTr * ge_left.adjoint());
+
   
   // TODO check with sub-matrix expressions ?
 }

From e3912f5e63b0c08a0f592ad425e926a5d61e1b8a Mon Sep 17 00:00:00 2001
From: "luz.paz" <luzpaz@users.noreply.github.com>
Date: Sun, 11 Mar 2018 10:01:44 -0400
Subject: [PATCH 264/680] MIsc. source and comment typos

Found using `codespell` and `grep` from downstream FreeCAD
---
 Eigen/Core                                             |  2 +-
 Eigen/src/Cholesky/LDLT.h                              |  2 +-
 Eigen/src/Cholesky/LLT.h                               |  2 +-
 Eigen/src/Core/AssignEvaluator.h                       |  4 ++--
 Eigen/src/Core/DenseBase.h                             |  2 +-
 Eigen/src/Core/DenseStorage.h                          |  2 +-
 Eigen/src/Core/MathFunctions.h                         |  2 +-
 Eigen/src/Core/NoAlias.h                               |  4 ++--
 Eigen/src/Core/PlainObjectBase.h                       |  2 +-
 Eigen/src/Core/Product.h                               |  2 +-
 Eigen/src/Core/Transpositions.h                        |  2 +-
 Eigen/src/Core/TriangularMatrix.h                      |  4 ++--
 Eigen/src/Core/arch/AltiVec/PacketMath.h               |  4 ++--
 Eigen/src/Core/arch/SSE/MathFunctions.h                |  2 +-
 Eigen/src/Core/products/GeneralBlockPanelKernel.h      |  4 ++--
 Eigen/src/Core/products/GeneralMatrixVector.h          |  2 +-
 Eigen/src/Core/products/Parallelizer.h                 |  2 +-
 Eigen/src/Core/products/SelfadjointMatrixVector.h      |  2 +-
 Eigen/src/Core/util/Macros.h                           |  4 ++--
 Eigen/src/Core/util/Memory.h                           |  2 +-
 Eigen/src/Core/util/Meta.h                             |  2 +-
 Eigen/src/Eigenvalues/ComplexEigenSolver.h             |  2 +-
 Eigen/src/Eigenvalues/ComplexSchur.h                   |  2 +-
 Eigen/src/Eigenvalues/EigenSolver.h                    |  2 +-
 .../Eigenvalues/GeneralizedSelfAdjointEigenSolver.h    |  2 +-
 Eigen/src/Eigenvalues/RealQZ.h                         |  2 +-
 Eigen/src/Eigenvalues/RealSchur.h                      |  2 +-
 Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h         |  2 +-
 Eigen/src/Geometry/Scaling.h                           |  2 +-
 Eigen/src/IterativeLinearSolvers/IncompleteLUT.h       |  4 ++--
 Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h |  2 +-
 Eigen/src/KLUSupport/KLUSupport.h                      |  2 +-
 Eigen/src/LU/FullPivLU.h                               |  2 +-
 Eigen/src/LU/PartialPivLU.h                            |  4 ++--
 Eigen/src/OrderingMethods/Eigen_Colamd.h               |  6 +++---
 Eigen/src/PaStiXSupport/PaStiXSupport.h                |  2 +-
 Eigen/src/PardisoSupport/PardisoSupport.h              |  2 +-
 Eigen/src/QR/ColPivHouseholderQR.h                     |  2 +-
 Eigen/src/QR/CompleteOrthogonalDecomposition.h         |  2 +-
 Eigen/src/SPQRSupport/SuiteSparseQRSupport.h           |  2 +-
 Eigen/src/SVD/BDCSVD.h                                 |  2 +-
 Eigen/src/SVD/UpperBidiagonalization.h                 |  4 ++--
 Eigen/src/SparseCholesky/SimplicialCholesky.h          |  2 +-
 Eigen/src/SparseCore/SparseMatrix.h                    |  6 +++---
 Eigen/src/SparseCore/SparseProduct.h                   |  2 +-
 Eigen/src/SparseCore/SparseVector.h                    |  2 +-
 Eigen/src/SparseLU/SparseLU.h                          |  2 +-
 Eigen/src/SparseLU/SparseLU_Memory.h                   |  2 +-
 Eigen/src/SparseLU/SparseLU_column_dfs.h               |  4 ++--
 Eigen/src/SparseLU/SparseLU_gemm_kernel.h              |  2 +-
 Eigen/src/SparseLU/SparseLU_panel_bmod.h               |  2 +-
 Eigen/src/SuperLUSupport/SuperLUSupport.h              |  2 +-
 Eigen/src/UmfPackSupport/UmfPackSupport.h              |  2 +-
 Eigen/src/plugins/IndexedViewMethods.h                 |  2 +-
 bench/analyze-blocking-sizes.cpp                       |  2 +-
 bench/btl/README                                       |  2 +-
 bench/btl/generic_bench/bench.hh                       |  2 +-
 bench/btl/generic_bench/utils/size_log.hh              |  2 +-
 bench/btl/generic_bench/utils/xy_file.hh               |  2 +-
 bench/btl/libs/ublas/ublas_interface.hh                |  2 +-
 bench/eig33.cpp                                        |  2 +-
 bench/spbench/spbenchsolver.cpp                        |  2 +-
 blas/f2c/ctbmv.c                                       |  2 +-
 blas/f2c/dtbmv.c                                       |  2 +-
 blas/f2c/stbmv.c                                       |  2 +-
 blas/f2c/ztbmv.c                                       |  2 +-
 blas/level1_impl.h                                     |  2 +-
 blas/testing/cblat1.f                                  |  2 +-
 blas/testing/dblat1.f                                  |  2 +-
 blas/testing/sblat1.f                                  |  2 +-
 blas/testing/zblat1.f                                  |  2 +-
 cmake/EigenConfigureTesting.cmake                      |  4 ++--
 cmake/EigenTesting.cmake                               | 10 +++++-----
 cmake/FindComputeCpp.cmake                             |  2 +-
 cmake/FindEigen3.cmake                                 |  2 +-
 debug/msvc/eigen_autoexp_part.dat                      |  2 +-
 doc/Doxyfile.in                                        |  2 +-
 doc/FunctionsTakingEigenTypes.dox                      |  4 ++--
 doc/PreprocessorDirectives.dox                         |  2 +-
 doc/QuickStartGuide.dox                                |  2 +-
 doc/SparseQuickReference.dox                           |  2 +-
 doc/TemplateKeyword.dox                                |  2 +-
 doc/TopicLazyEvaluation.dox                            |  2 +-
 doc/TopicLinearAlgebraDecompositions.dox               |  2 +-
 doc/TopicMultithreading.dox                            |  2 +-
 doc/TutorialMapClass.dox                               |  4 ++--
 doc/TutorialSparse.dox                                 |  4 ++--
 doc/UnalignedArrayAssert.dox                           |  4 ++--
 doc/UsingNVCC.dox                                      |  2 +-
 doc/eigendoxy.css                                      |  2 +-
 doc/special_examples/Tutorial_sparse_example.cpp       |  2 +-
 lapack/CMakeLists.txt                                  |  4 ++--
 test/CMakeLists.txt                                    |  2 +-
 test/bdcsvd.cpp                                        |  2 +-
 test/eigensolver_complex.cpp                           |  2 +-
 test/geo_quaternion.cpp                                |  2 +-
 test/main.h                                            |  2 +-
 test/packetmath.cpp                                    |  2 +-
 98 files changed, 122 insertions(+), 122 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index 5a6dec8cc..a9bbfe276 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -360,7 +360,7 @@ inline static const char *SimdInstructionSetsInUse(void) {
 
 namespace Eigen {
 
-// we use size_t frequently and we'll never remember to prepend it with std:: everytime just to
+// we use size_t frequently and we'll never remember to prepend it with std:: every time just to
 // ensure QNX/QCC support
 using std::size_t;
 // gcc 4.6.0 wants std:: for ptrdiff_t
diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h
index 13a8f6d14..5be58377b 100644
--- a/Eigen/src/Cholesky/LDLT.h
+++ b/Eigen/src/Cholesky/LDLT.h
@@ -247,7 +247,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the factorization failed because of a zero pivot.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 814174d47..22e4be75d 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -180,7 +180,7 @@ template<typename _MatrixType, int _UpLo> class LLT
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears not to be positive definite.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h
index ebf5590de..362d905d2 100644
--- a/Eigen/src/Core/AssignEvaluator.h
+++ b/Eigen/src/Core/AssignEvaluator.h
@@ -756,7 +756,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType
 // AssignmentKind must define a Kind typedef.
 template<typename DstShape, typename SrcShape> struct AssignmentKind;
 
-// Assignement kind defined in this file:
+// Assignment kind defined in this file:
 struct Dense2Dense {};
 struct EigenBase2EigenBase {};
 
@@ -899,7 +899,7 @@ struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Weak>
     src.evalTo(dst);
   }
 
-  // NOTE The following two functions are templated to avoid their instanciation if not needed
+  // NOTE The following two functions are templated to avoid their instantiation if not needed
   //      This is needed because some expressions supports evalTo only and/or have 'void' as scalar type.
   template<typename SrcScalarType>
   EIGEN_DEVICE_FUNC
diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index fd933eed4..53b427b17 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -395,7 +395,7 @@ template<typename Derived> class DenseBase
       * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
       * a const reference, in order to avoid a useless copy.
       * 
-      * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
+      * \warning Be careful with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
       */
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE EvalReturnType eval() const
diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h
index 7958feeb9..9e58fbf88 100644
--- a/Eigen/src/Core/DenseStorage.h
+++ b/Eigen/src/Core/DenseStorage.h
@@ -61,7 +61,7 @@ struct plain_array
 #if defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
   #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask)
 #elif EIGEN_GNUC_AT_LEAST(4,7) 
-  // GCC 4.7 is too aggressive in its optimizations and remove the alignement test based on the fact the array is declared to be aligned.
+  // GCC 4.7 is too aggressive in its optimizations and remove the alignment test based on the fact the array is declared to be aligned.
   // See this bug report: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53900
   // Hiding the origin of the array pointer behind a function argument seems to do the trick even if the function is inlined:
   template<typename PtrType>
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 1b864a405..e981129b2 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -749,7 +749,7 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
   return EIGEN_MATHFUNC_IMPL(random, Scalar)::run();
 }
 
-// Implementatin of is* functions
+// Implementation of is* functions
 
 // std::is* do not work with fast-math and gcc, std::is* are available on MSVC 2013 and newer, as well as in clang.
 #if (EIGEN_HAS_CXX11_MATH && !(EIGEN_COMP_GNUC_STRICT && __FINITE_MATH_ONLY__)) || (EIGEN_COMP_MSVC>=1800) || (EIGEN_COMP_CLANG)
diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h
index e94c8ee96..570283d90 100644
--- a/Eigen/src/Core/NoAlias.h
+++ b/Eigen/src/Core/NoAlias.h
@@ -75,10 +75,10 @@ class NoAlias
   *
   * More precisely, noalias() allows to bypass the EvalBeforeAssignBit flag.
   * Currently, even though several expressions may alias, only product
-  * expressions have this flag. Therefore, noalias() is only usefull when
+  * expressions have this flag. Therefore, noalias() is only useful when
   * the source expression contains a matrix product.
   *
-  * Here are some examples where noalias is usefull:
+  * Here are some examples where noalias is useful:
   * \code
   * D.noalias()  = A * B;
   * D.noalias() += A.transpose() * B;
diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index 1dc7e223a..6c0a42ec7 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -780,7 +780,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       resize(size);
     }
     
-    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted)
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitly converted)
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0)
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index 676c48027..3d67d9489 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -116,7 +116,7 @@ class dense_product_base
  : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
 {};
 
-/** Convertion to scalar for inner-products */
+/** Conversion to scalar for inner-products */
 template<typename Lhs, typename Rhs, int Option>
 class dense_product_base<Lhs, Rhs, Option, InnerProduct>
  : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h
index 19c17bb4a..8798deca5 100644
--- a/Eigen/src/Core/Transpositions.h
+++ b/Eigen/src/Core/Transpositions.h
@@ -84,7 +84,7 @@ class TranspositionsBase
     }
 
     // FIXME: do we want such methods ?
-    // might be usefull when the target matrix expression is complex, e.g.:
+    // might be useful when the target matrix expression is complex, e.g.:
     // object.matrix().block(..,..,..,..) = trans * object.matrix().block(..,..,..,..);
     /*
     template<typename MatrixType>
diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h
index ed80da36a..ab73fcf21 100644
--- a/Eigen/src/Core/TriangularMatrix.h
+++ b/Eigen/src/Core/TriangularMatrix.h
@@ -470,7 +470,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
       * \a Side==OnTheRight.
       *
-      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      * Note that the template parameter \c Side can be omitted, in which case \c Side==OnTheLeft
       *
       * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
       * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
@@ -496,7 +496,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
       * This function will const_cast it, so constness isn't honored here.
       *
-      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      * Note that the template parameter \c Side can be omitted, in which case \c Side==OnTheLeft
       *
       * See TriangularView:solve() for the details.
       */
diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index b3f1ea199..31bb896ca 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -434,7 +434,7 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
   return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
 }
 #else
-// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+// We also need to redefine little endian loading of Packet4i/Packet4f using VSX
 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
 {
   EIGEN_DEBUG_UNALIGNED_LOAD
@@ -500,7 +500,7 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& f
   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
 }
 #else
-// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+// We also need to redefine little endian loading of Packet4i/Packet4f using VSX
 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from)
 {
   EIGEN_DEBUG_ALIGNED_STORE
diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h
index 7b5f948e1..4af2c6cae 100644
--- a/Eigen/src/Core/arch/SSE/MathFunctions.h
+++ b/Eigen/src/Core/arch/SSE/MathFunctions.h
@@ -242,7 +242,7 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
   return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x);
 }
 
-/* evaluation of 4 sines at onces, using SSE2 intrinsics.
+/* evaluation of 4 sines at once, using SSE2 intrinsics.
 
    The code is the exact rewriting of the cephes sinf function.
    Precision is excellent as long as x < 8192 (I did not bother to
diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 45230bce5..9072d0ff3 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -1523,7 +1523,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           prefetch(&blA[0]);
           const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
 
-          // The following piece of code wont work for 512 bit registers
+          // The following piece of code won't work for 512 bit registers
           // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size
           // as nr (which is currently 4) for the return type.
           typedef typename unpacket_traits<SResPacket>::half SResPacketHalf;
@@ -1924,7 +1924,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Co
 //       const Scalar* b6 = &rhs[(j2+6)*rhsStride];
 //       const Scalar* b7 = &rhs[(j2+7)*rhsStride];
 //       Index k=0;
-//       if(PacketSize==8) // TODO enbale vectorized transposition for PacketSize==4
+//       if(PacketSize==8) // TODO enable vectorized transposition for PacketSize==4
 //       {
 //         for(; k<peeled_k; k+=PacketSize) {
 //           PacketBlock<Packet> kernel;
diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h
index 41d8242e1..b2a71bc6f 100644
--- a/Eigen/src/Core/products/GeneralMatrixVector.h
+++ b/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -201,7 +201,7 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,C
 }
 
 /* Optimized row-major matrix * vector product:
- * This algorithm processes 4 rows at onces that allows to both reduce
+ * This algorithm processes 4 rows at once that allows to both reduce
  * the number of load/stores of the result by a factor 4 and to reduce
  * the instruction dependency. Moreover, we know that all bands have the
  * same alignment pattern.
diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h
index c2f084c82..e4d13103b 100644
--- a/Eigen/src/Core/products/Parallelizer.h
+++ b/Eigen/src/Core/products/Parallelizer.h
@@ -117,7 +117,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth,
   // compute the number of threads we are going to use
   Index threads = std::min<Index>(nbThreads(), pb_max_threads);
 
-  // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
+  // if multi-threading is explicitly disabled, not useful, or if we already are in a parallel session,
   // then abort multi-threading
   // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
   if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h
index 3fd180e6c..67390f1d7 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixVector.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h
@@ -15,7 +15,7 @@ namespace Eigen {
 namespace internal {
 
 /* Optimized selfadjoint matrix * vector product:
- * This algorithm processes 2 columns at onces that allows to both reduce
+ * This algorithm processes 2 columns at once that allows to both reduce
  * the number of load/stores of the result by a factor 2 and to reduce
  * the instruction dependency.
  */
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index e351b7ad9..5872ade26 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -719,7 +719,7 @@ namespace Eigen {
 #error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
 #endif
 
-// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprectated
+// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprecated
 // They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
 #if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
   #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
@@ -778,7 +778,7 @@ namespace Eigen {
 #endif
 
 // At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
-// It takes into account both the user choice to explicitly enable/disable alignment (by settting EIGEN_MAX_STATIC_ALIGN_BYTES)
+// It takes into account both the user choice to explicitly enable/disable alignment (by setting EIGEN_MAX_STATIC_ALIGN_BYTES)
 // and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
 // Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
 
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index c455f92a1..006b0bfba 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -703,7 +703,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 *  - 32 bytes alignment if AVX is enabled.
 *  - 64 bytes alignment if AVX512 is enabled.
 *
-* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
+* This can be controlled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
 * \link TopicPreprocessorDirectivesPerformance there \endlink.
 *
 * Example:
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 0fa818008..998b8921a 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -272,7 +272,7 @@ template<> struct numeric_limits<unsigned long long>
 #endif
 
 /** \internal
-  * A base class do disable default copy ctor and copy assignement operator.
+  * A base class do disable default copy ctor and copy assignment operator.
   */
 class noncopyable
 {
diff --git a/Eigen/src/Eigenvalues/ComplexEigenSolver.h b/Eigen/src/Eigenvalues/ComplexEigenSolver.h
index dc5fae06a..081e918f1 100644
--- a/Eigen/src/Eigenvalues/ComplexEigenSolver.h
+++ b/Eigen/src/Eigenvalues/ComplexEigenSolver.h
@@ -214,7 +214,7 @@ template<typename _MatrixType> class ComplexEigenSolver
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      * \returns \c Success if computation was successful, \c NoConvergence otherwise.
       */
     ComputationInfo info() const
     {
diff --git a/Eigen/src/Eigenvalues/ComplexSchur.h b/Eigen/src/Eigenvalues/ComplexSchur.h
index 7f38919f7..b8b3490c6 100644
--- a/Eigen/src/Eigenvalues/ComplexSchur.h
+++ b/Eigen/src/Eigenvalues/ComplexSchur.h
@@ -212,7 +212,7 @@ template<typename _MatrixType> class ComplexSchur
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      * \returns \c Success if computation was successful, \c NoConvergence otherwise.
       */
     ComputationInfo info() const
     {
diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h
index f205b185d..997bebe7b 100644
--- a/Eigen/src/Eigenvalues/EigenSolver.h
+++ b/Eigen/src/Eigenvalues/EigenSolver.h
@@ -277,7 +277,7 @@ template<typename _MatrixType> class EigenSolver
     template<typename InputType>
     EigenSolver& compute(const EigenBase<InputType>& matrix, bool computeEigenvectors = true);
 
-    /** \returns NumericalIssue if the input contains INF or NaN values or overflow occured. Returns Success otherwise. */
+    /** \returns NumericalIssue if the input contains INF or NaN values or overflow occurred. Returns Success otherwise. */
     ComputationInfo info() const
     {
       eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
diff --git a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h
index 5f6bb8289..d0f9091be 100644
--- a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h
@@ -121,7 +121,7 @@ class GeneralizedSelfAdjointEigenSolver : public SelfAdjointEigenSolver<_MatrixT
       *
       * \returns    Reference to \c *this
       *
-      * Accoring to \p options, this function computes eigenvalues and (if requested)
+      * According to \p options, this function computes eigenvalues and (if requested)
       * the eigenvectors of one of the following three generalized eigenproblems:
       * - \c Ax_lBx: \f$ Ax = \lambda B x \f$
       * - \c ABx_lx: \f$ ABx = \lambda x \f$
diff --git a/Eigen/src/Eigenvalues/RealQZ.h b/Eigen/src/Eigenvalues/RealQZ.h
index b3a910dd9..e2b37f40e 100644
--- a/Eigen/src/Eigenvalues/RealQZ.h
+++ b/Eigen/src/Eigenvalues/RealQZ.h
@@ -161,7 +161,7 @@ namespace Eigen {
 
       /** \brief Reports whether previous computation was successful.
        *
-       * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+       * \returns \c Success if computation was successful, \c NoConvergence otherwise.
        */
       ComputationInfo info() const
       {
diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h
index f5c86041d..9e71f3040 100644
--- a/Eigen/src/Eigenvalues/RealSchur.h
+++ b/Eigen/src/Eigenvalues/RealSchur.h
@@ -190,7 +190,7 @@ template<typename _MatrixType> class RealSchur
     RealSchur& computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ,  bool computeU);
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      * \returns \c Success if computation was successful, \c NoConvergence otherwise.
       */
     ComputationInfo info() const
     {
diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
index 9ddd553f2..040f8d3bb 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -337,7 +337,7 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      * \returns \c Success if computation was successful, \c NoConvergence otherwise.
       */
     EIGEN_DEVICE_FUNC
     ComputationInfo info() const
diff --git a/Eigen/src/Geometry/Scaling.h b/Eigen/src/Geometry/Scaling.h
index 8d9acf252..df650fda6 100755
--- a/Eigen/src/Geometry/Scaling.h
+++ b/Eigen/src/Geometry/Scaling.h
@@ -128,7 +128,7 @@ public:
 /** Concatenates a linear transformation matrix and a uniform scaling
   * \relates UniformScaling
   */
-// NOTE this operator is defiend in MatrixBase and not as a friend function
+// NOTE this operator is defined in MatrixBase and not as a friend function
 // of UniformScaling to fix an internal crash of Intel's ICC
 template<typename Derived,typename Scalar>
 EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,Scalar,product)
diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h
index 338e6f10a..43bd8e8f6 100644
--- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h
+++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h
@@ -136,7 +136,7 @@ class IncompleteLUT : public SparseSolverBase<IncompleteLUT<_Scalar, _StorageInd
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears to be negative.
       */
     ComputationInfo info() const
@@ -230,7 +230,7 @@ void IncompleteLUT<Scalar,StorageIndex>::analyzePattern(const _MatrixType& amat)
   SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
   SparseMatrix<Scalar,ColMajor, StorageIndex> mat2 = amat.transpose();
   // FIXME for a matrix with nearly symmetric pattern, mat2+mat1 is the appropriate choice.
-  //       on the other hand for a really non-symmetric pattern, mat2*mat1 should be prefered...
+  //       on the other hand for a really non-symmetric pattern, mat2*mat1 should be preferred...
   SparseMatrix<Scalar,ColMajor, StorageIndex> AtA = mat2 + mat1;
   AMDOrdering<StorageIndex> ordering;
   ordering(AtA,m_P);
diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h
index 7c2326eb7..bfeee71cd 100644
--- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h
+++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h
@@ -275,7 +275,7 @@ public:
   const Preconditioner& preconditioner() const { return m_preconditioner; }
 
   /** \returns the max number of iterations.
-    * It is either the value setted by setMaxIterations or, by default,
+    * It is either the value set by setMaxIterations or, by default,
     * twice the number of columns of the matrix.
     */
   Index maxIterations() const
diff --git a/Eigen/src/KLUSupport/KLUSupport.h b/Eigen/src/KLUSupport/KLUSupport.h
index a9e8633d9..d2633a935 100644
--- a/Eigen/src/KLUSupport/KLUSupport.h
+++ b/Eigen/src/KLUSupport/KLUSupport.h
@@ -106,7 +106,7 @@ class KLU : public SparseSolverBase<KLU<_MatrixType> >
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears to be negative.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h
index ec61086d5..50d1bb41b 100644
--- a/Eigen/src/LU/FullPivLU.h
+++ b/Eigen/src/LU/FullPivLU.h
@@ -48,7 +48,7 @@ template<typename _MatrixType> struct traits<FullPivLU<_MatrixType> >
   * The data of the LU decomposition can be directly accessed through the methods matrixLU(),
   * permutationP(), permutationQ().
   *
-  * As an exemple, here is how the original matrix can be retrieved:
+  * As an example, here is how the original matrix can be retrieved:
   * \include class_FullPivLU.cpp
   * Output: \verbinclude class_FullPivLU.out
   *
diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h
index d43961887..bfcd2c95b 100644
--- a/Eigen/src/LU/PartialPivLU.h
+++ b/Eigen/src/LU/PartialPivLU.h
@@ -420,8 +420,8 @@ struct partial_lu_impl
     * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
     *
     * \note This very low level interface using pointers, etc. is to:
-    *   1 - reduce the number of instanciations to the strict minimum
-    *   2 - avoid infinite recursion of the instanciations with Block<Block<Block<...> > >
+    *   1 - reduce the number of instantiations to the strict minimum
+    *   2 - avoid infinite recursion of the instantiations with Block<Block<Block<...> > >
     */
   static Index blocked_lu(Index rows, Index cols, Scalar* lu_data, Index luStride, PivIndex* row_transpositions, PivIndex& nb_transpositions, Index maxBlockSize=256)
   {
diff --git a/Eigen/src/OrderingMethods/Eigen_Colamd.h b/Eigen/src/OrderingMethods/Eigen_Colamd.h
index da85b4d6e..67fcad3f7 100644
--- a/Eigen/src/OrderingMethods/Eigen_Colamd.h
+++ b/Eigen/src/OrderingMethods/Eigen_Colamd.h
@@ -1493,7 +1493,7 @@ static inline  void order_children
 	c = Col [c].shared1.parent ;
 
 	/* continue until we hit an ordered column.  There are */
-	/* guarranteed not to be anymore unordered columns */
+	/* guaranteed not to be anymore unordered columns */
 	/* above an ordered column */
       } while (Col [c].shared2.order == COLAMD_EMPTY) ;
 
@@ -1638,7 +1638,7 @@ static void detect_super_cols
 	  COLAMD_ASSERT (ROW_IS_ALIVE (*cp1))  ;
 	  COLAMD_ASSERT (ROW_IS_ALIVE (*cp2))  ;
 	  /* row indices will same order for both supercols, */
-	  /* no gather scatter nessasary */
+	  /* no gather scatter necessary */
 	  if (*cp1++ != *cp2++)
 	  {
 	    break ;
@@ -1688,7 +1688,7 @@ static void detect_super_cols
 
 /*
   Defragments and compacts columns and rows in the workspace A.  Used when
-  all avaliable memory has been used while performing row merging.  Returns
+  all available memory has been used while performing row merging.  Returns
   the index of the first free position in A, after garbage collection.  The
   time taken by this routine is linear is the size of the array A, which is
   itself linear in the number of nonzeros in the input matrix.
diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h
index 160d8a523..37426877a 100644
--- a/Eigen/src/PaStiXSupport/PaStiXSupport.h
+++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h
@@ -203,7 +203,7 @@ class PastixBase : public SparseSolverBase<Derived>
     
      /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the PaStiX reports a problem
       *          \c InvalidInput if the input matrix is invalid
       *
diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h
index 091c3970e..fb2ba04b4 100644
--- a/Eigen/src/PardisoSupport/PardisoSupport.h
+++ b/Eigen/src/PardisoSupport/PardisoSupport.h
@@ -140,7 +140,7 @@ class PardisoImpl : public SparseSolverBase<Derived>
   
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix appears to be negative.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index 5270eaca2..ed47b05e3 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -402,7 +402,7 @@ template<typename _MatrixType> class ColPivHouseholderQR
       */
     RealScalar maxPivot() const { return m_maxpivot; }
 
-    /** \brief Reports whether the QR factorization was succesful.
+    /** \brief Reports whether the QR factorization was successful.
       *
       * \note This function always returns \c Success. It is provided for compatibility
       * with other factorization routines.
diff --git a/Eigen/src/QR/CompleteOrthogonalDecomposition.h b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
index 13b61fcdb..880becb25 100644
--- a/Eigen/src/QR/CompleteOrthogonalDecomposition.h
+++ b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
@@ -353,7 +353,7 @@ class CompleteOrthogonalDecomposition {
   inline RealScalar maxPivot() const { return m_cpqr.maxPivot(); }
 
   /** \brief Reports whether the complete orthogonal decomposition was
-   * succesful.
+   * successful.
    *
    * \note This function always returns \c Success. It is provided for
    * compatibility
diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h
index 953d57c9d..1a5c5254e 100644
--- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h
+++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h
@@ -220,7 +220,7 @@ class SPQR : public SparseSolverBase<SPQR<_MatrixType> >
     
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the sparse QR can not be computed
       */
     ComputationInfo info() const
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index 06865a331..a24deb96a 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -62,7 +62,7 @@ struct traits<BDCSVD<_MatrixType> >
  * recommended and can several order of magnitude faster.
  *
  * \warning this algorithm is unlikely to provide accurate result when compiled with unsafe math optimizations.
- * For instance, this concerns Intel's compiler (ICC), which perfroms such optimization by default unless
+ * For instance, this concerns Intel's compiler (ICC), which performs such optimization by default unless
  * you compile with the \c -fp-model \c precise option. Likewise, the \c -ffast-math option of GCC or clang will
  * significantly degrade the accuracy.
  *
diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h
index 11ac847e1..0526ac931 100644
--- a/Eigen/src/SVD/UpperBidiagonalization.h
+++ b/Eigen/src/SVD/UpperBidiagonalization.h
@@ -202,7 +202,7 @@ void upperbidiagonalization_blocked_helper(MatrixType& A,
       {
         SubColumnType y_k( Y.col(k).tail(remainingCols) );
         
-        // let's use the begining of column k of Y as a temporary vector
+        // let's use the beginning of column k of Y as a temporary vector
         SubColumnType tmp( Y.col(k).head(k) );
         y_k.noalias()  = A.block(k,k+1, remainingRows,remainingCols).adjoint() * v_k; // bottleneck
         tmp.noalias()  = V_k1.adjoint()  * v_k;
@@ -231,7 +231,7 @@ void upperbidiagonalization_blocked_helper(MatrixType& A,
       {
         SubColumnType x_k ( X.col(k).tail(remainingRows-1) );
         
-        // let's use the begining of column k of X as a temporary vectors
+        // let's use the beginning of column k of X as a temporary vectors
         // note that tmp0 and tmp1 overlaps
         SubColumnType tmp0 ( X.col(k).head(k) ),
                       tmp1 ( X.col(k).head(k+1) );
diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h
index 2907f6529..b9ca94bc3 100644
--- a/Eigen/src/SparseCholesky/SimplicialCholesky.h
+++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h
@@ -101,7 +101,7 @@ class SimplicialCholeskyBase : public SparseSolverBase<Derived>
     
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears to be negative.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h
index 323c2323b..8f77194b6 100644
--- a/Eigen/src/SparseCore/SparseMatrix.h
+++ b/Eigen/src/SparseCore/SparseMatrix.h
@@ -21,7 +21,7 @@ namespace Eigen {
   * This class implements a more versatile variants of the common \em compressed row/column storage format.
   * Each colmun's (resp. row) non zeros are stored as a pair of value with associated row (resp. colmiun) index.
   * All the non zeros are stored in a single large buffer. Unlike the \em compressed format, there might be extra
-  * space inbetween the nonzeros of two successive colmuns (resp. rows) such that insertion of new non-zero
+  * space in between the nonzeros of two successive colmuns (resp. rows) such that insertion of new non-zero
   * can be done with limited memory reallocation and copies.
   *
   * A call to the function makeCompressed() turns the matrix into the standard \em compressed format
@@ -503,7 +503,7 @@ class SparseMatrix
       }
     }
     
-    /** Suppresses all nonzeros which are \b much \b smaller \b than \a reference under the tolerence \a epsilon */
+    /** Suppresses all nonzeros which are \b much \b smaller \b than \a reference under the tolerance \a epsilon */
     void prune(const Scalar& reference, const RealScalar& epsilon = NumTraits<RealScalar>::dummy_precision())
     {
       prune(default_prunning_func(reference,epsilon));
@@ -986,7 +986,7 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa
   *
   * \warning The list of triplets is read multiple times (at least twice). Therefore, it is not recommended to define
   * an abstract iterator over a complex data-structure that would be expensive to evaluate. The triplets should rather
-  * be explicitely stored into a std::vector for instance.
+  * be explicitly stored into a std::vector for instance.
   */
 template<typename Scalar, int _Options, typename _StorageIndex>
 template<typename InputIterators>
diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h
index 4cbf68781..c495a7398 100644
--- a/Eigen/src/SparseCore/SparseProduct.h
+++ b/Eigen/src/SparseCore/SparseProduct.h
@@ -17,7 +17,7 @@ namespace Eigen {
   * The automatic pruning of the small values can be achieved by calling the pruned() function
   * in which case a totally different product algorithm is employed:
   * \code
-  * C = (A*B).pruned();             // supress numerical zeros (exact)
+  * C = (A*B).pruned();             // suppress numerical zeros (exact)
   * C = (A*B).pruned(ref);
   * C = (A*B).pruned(ref,epsilon);
   * \endcode
diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h
index 19b0fbc9d..05779be68 100644
--- a/Eigen/src/SparseCore/SparseVector.h
+++ b/Eigen/src/SparseCore/SparseVector.h
@@ -281,7 +281,7 @@ class SparseVector
     }
 
     /** Swaps the values of \c *this and \a other.
-      * Overloaded for performance: this version performs a \em shallow swap by swaping pointers and attributes only.
+      * Overloaded for performance: this version performs a \em shallow swap by swapping pointers and attributes only.
       * \sa SparseMatrixBase::swap()
       */
     inline void swap(SparseVector& other)
diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h
index f883ab383..383a203b4 100644
--- a/Eigen/src/SparseLU/SparseLU.h
+++ b/Eigen/src/SparseLU/SparseLU.h
@@ -193,7 +193,7 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
     
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the LU factorization reports a problem, zero diagonal for instance
       *          \c InvalidInput if the input matrix is invalid
       *
diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h
index 4dc42e87b..349bfd585 100644
--- a/Eigen/src/SparseLU/SparseLU_Memory.h
+++ b/Eigen/src/SparseLU/SparseLU_Memory.h
@@ -51,7 +51,7 @@ inline Index LUTempSpace(Index&m, Index& w)
 
 
 /** 
-  * Expand the existing storage to accomodate more fill-ins
+  * Expand the existing storage to accommodate more fill-ins
   * \param vec Valid pointer to the vector to allocate or expand
   * \param[in,out] length  At input, contain the current length of the vector that is to be increased. At output, length of the newly allocated vector
   * \param[in] nbElts Current number of elements in the factors
diff --git a/Eigen/src/SparseLU/SparseLU_column_dfs.h b/Eigen/src/SparseLU/SparseLU_column_dfs.h
index c98b30e32..5a2c941b4 100644
--- a/Eigen/src/SparseLU/SparseLU_column_dfs.h
+++ b/Eigen/src/SparseLU/SparseLU_column_dfs.h
@@ -151,7 +151,7 @@ Index SparseLUImpl<Scalar,StorageIndex>::column_dfs(const Index m, const Index j
         StorageIndex ito = glu.xlsub(fsupc+1);
         glu.xlsub(jcolm1) = ito; 
         StorageIndex istop = ito + jptr - jm1ptr; 
-        xprune(jcolm1) = istop; // intialize xprune(jcol-1)
+        xprune(jcolm1) = istop; // initialize xprune(jcol-1)
         glu.xlsub(jcol) = istop; 
         
         for (StorageIndex ifrom = jm1ptr; ifrom < nextl; ++ifrom, ++ito)
@@ -166,7 +166,7 @@ Index SparseLUImpl<Scalar,StorageIndex>::column_dfs(const Index m, const Index j
   // Tidy up the pointers before exit
   glu.xsup(nsuper+1) = jcolp1; 
   glu.supno(jcolp1) = nsuper; 
-  xprune(jcol) = StorageIndex(nextl);  // Intialize upper bound for pruning
+  xprune(jcol) = StorageIndex(nextl);  // Initialize upper bound for pruning
   glu.xlsub(jcolp1) = StorageIndex(nextl); 
   
   return 0; 
diff --git a/Eigen/src/SparseLU/SparseLU_gemm_kernel.h b/Eigen/src/SparseLU/SparseLU_gemm_kernel.h
index 95ba7413f..e37c2fe0d 100644
--- a/Eigen/src/SparseLU/SparseLU_gemm_kernel.h
+++ b/Eigen/src/SparseLU/SparseLU_gemm_kernel.h
@@ -215,7 +215,7 @@ void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const
         if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize);  }\
                    pstore(C0+i+(I)*PacketSize, c0);
         
-        // agressive vectorization and peeling
+        // aggressive vectorization and peeling
         for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
         {
           EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
diff --git a/Eigen/src/SparseLU/SparseLU_panel_bmod.h b/Eigen/src/SparseLU/SparseLU_panel_bmod.h
index 822cf32c3..f052001c8 100644
--- a/Eigen/src/SparseLU/SparseLU_panel_bmod.h
+++ b/Eigen/src/SparseLU/SparseLU_panel_bmod.h
@@ -38,7 +38,7 @@ namespace internal {
  * \brief Performs numeric block updates (sup-panel) in topological order.
  * 
  * Before entering this routine, the original nonzeros in the panel
- * were already copied i nto the spa[m,w]
+ * were already copied into the spa[m,w]
  * 
  * \param m number of rows in the matrix
  * \param w Panel size
diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h
index 50a69f306..4bb95eb8b 100644
--- a/Eigen/src/SuperLUSupport/SuperLUSupport.h
+++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h
@@ -352,7 +352,7 @@ class SuperLUBase : public SparseSolverBase<Derived>
     
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears to be negative.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h
index 9568cc1d5..c636f17ac 100644
--- a/Eigen/src/UmfPackSupport/UmfPackSupport.h
+++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -201,7 +201,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
 
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the matrix.appears to be negative.
       */
     ComputationInfo info() const
diff --git a/Eigen/src/plugins/IndexedViewMethods.h b/Eigen/src/plugins/IndexedViewMethods.h
index a7ec63adf..9ad2d9aee 100644
--- a/Eigen/src/plugins/IndexedViewMethods.h
+++ b/Eigen/src/plugins/IndexedViewMethods.h
@@ -112,7 +112,7 @@ operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_IND
 
 #if EIGEN_HAS_STATIC_ARRAY_TEMPLATE
 
-// The folowing three overloads are needed to handle raw Index[N] arrays.
+// The following three overloads are needed to handle raw Index[N] arrays.
 
 template<typename RowIndicesT, std::size_t RowIndicesN, typename ColIndices>
 IndexedView<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,const RowIndicesT (&)[RowIndicesN],typename IvcColType<ColIndices>::type>
diff --git a/bench/analyze-blocking-sizes.cpp b/bench/analyze-blocking-sizes.cpp
index d563a1d2d..6bc4aca3d 100644
--- a/bench/analyze-blocking-sizes.cpp
+++ b/bench/analyze-blocking-sizes.cpp
@@ -825,7 +825,7 @@ int main(int argc, char* argv[])
   }
   for (int i = 1; i < argc; i++) {
     bool arg_handled = false;
-    // Step 1. Try to match action invokation names.
+    // Step 1. Try to match action invocation names.
     for (auto it = available_actions.begin(); it != available_actions.end(); ++it) {
       if (!strcmp(argv[i], (*it)->invokation_name())) {
         if (!action) {
diff --git a/bench/btl/README b/bench/btl/README
index f3f5fb36f..ebed88960 100644
--- a/bench/btl/README
+++ b/bench/btl/README
@@ -36,7 +36,7 @@ For instance:
 
 You can also select a given set of actions defining the environment variable BTL_CONFIG this way:
   BTL_CONFIG="-a action1{:action2}*" ctest -V
-An exemple:
+An example:
   BTL_CONFIG="-a axpy:vector_matrix:trisolve:ata" ctest -V -R eigen2
 
 Finally, if bench results already exist (the bench*.dat files) then they merges by keeping the best for each matrix size. If you want to overwrite the previous ones you can simply add the "--overwrite" option:
diff --git a/bench/btl/generic_bench/bench.hh b/bench/btl/generic_bench/bench.hh
index 7b7b951b5..0732940d5 100644
--- a/bench/btl/generic_bench/bench.hh
+++ b/bench/btl/generic_bench/bench.hh
@@ -159,7 +159,7 @@ BTL_DONT_INLINE void bench( int size_min, int size_max, int nb_point ){
 //    bench<Mixed_Perf_Analyzer,Action>(size_min,size_max,nb_point);
 
 
-  // Only for small problem size. Otherwize it will be too long
+  // Only for small problem size. Otherwise it will be too long
 //   bench<X86_Perf_Analyzer,Action>(size_min,size_max,nb_point);
 //   bench<STL_Perf_Analyzer,Action>(size_min,size_max,nb_point);
 
diff --git a/bench/btl/generic_bench/utils/size_log.hh b/bench/btl/generic_bench/utils/size_log.hh
index 13a3da7a8..68945e7cc 100644
--- a/bench/btl/generic_bench/utils/size_log.hh
+++ b/bench/btl/generic_bench/utils/size_log.hh
@@ -23,7 +23,7 @@
 #include "math.h"
 // The Vector class must satisfy the following part of STL vector concept :
 //            resize() method
-//            [] operator for seting element
+//            [] operator for setting element
 // the vector element are int compatible.
 template<class Vector>
 void size_log(const int nb_point, const int size_min, const int size_max, Vector & X)
diff --git a/bench/btl/generic_bench/utils/xy_file.hh b/bench/btl/generic_bench/utils/xy_file.hh
index 4571bed8f..0492faf09 100644
--- a/bench/btl/generic_bench/utils/xy_file.hh
+++ b/bench/btl/generic_bench/utils/xy_file.hh
@@ -55,7 +55,7 @@ bool read_xy_file(const std::string & filename, std::vector<int> & tab_sizes,
 
 // The Vector class must satisfy the following part of STL vector concept :
 //            resize() method
-//            [] operator for seting element
+//            [] operator for setting element
 // the vector element must have the << operator define
 
 using namespace std;
diff --git a/bench/btl/libs/ublas/ublas_interface.hh b/bench/btl/libs/ublas/ublas_interface.hh
index 95cad5195..f59b7cf2f 100644
--- a/bench/btl/libs/ublas/ublas_interface.hh
+++ b/bench/btl/libs/ublas/ublas_interface.hh
@@ -100,7 +100,7 @@ public :
     Y+=coef*X;
   }
 
-  // alias free assignements
+  // alias free assignments
 
   static inline void matrix_vector_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
     X.assign(prod(A,B));
diff --git a/bench/eig33.cpp b/bench/eig33.cpp
index 47947a9be..f003d8a53 100644
--- a/bench/eig33.cpp
+++ b/bench/eig33.cpp
@@ -101,7 +101,7 @@ void eigen33(const Matrix& mat, Matrix& evecs, Vector& evals)
   computeRoots(scaledMat,evals);
 
   // compute the eigen vectors
-  // **here we assume 3 differents eigenvalues**
+  // **here we assume 3 different eigenvalues**
 
   // "optimized version" which appears to be slower with gcc!
 //     Vector base;
diff --git a/bench/spbench/spbenchsolver.cpp b/bench/spbench/spbenchsolver.cpp
index 4acd0039c..2a7351124 100644
--- a/bench/spbench/spbenchsolver.cpp
+++ b/bench/spbench/spbenchsolver.cpp
@@ -54,7 +54,7 @@ int main(int argc, char ** args)
       statbuf.close();
     }
     else
-      std::cerr << "Unable to open the provided file for writting... \n";
+      std::cerr << "Unable to open the provided file for writing... \n";
   }       
   
   // Get the maximum number of iterations and the tolerance
diff --git a/blas/f2c/ctbmv.c b/blas/f2c/ctbmv.c
index 790fd581f..a6e0dae80 100644
--- a/blas/f2c/ctbmv.c
+++ b/blas/f2c/ctbmv.c
@@ -147,7 +147,7 @@
 /*           ( 1 + ( n - 1 )*abs( INCX ) ). */
 /*           Before entry, the incremented array X must contain the n */
 /*           element vector x. On exit, X is overwritten with the */
-/*           tranformed vector x. */
+/*           transformed vector x. */
 
 /*  INCX   - INTEGER. */
 /*           On entry, INCX specifies the increment for the elements of */
diff --git a/blas/f2c/dtbmv.c b/blas/f2c/dtbmv.c
index fdf73ebb5..aa67d19da 100644
--- a/blas/f2c/dtbmv.c
+++ b/blas/f2c/dtbmv.c
@@ -143,7 +143,7 @@
 /*           ( 1 + ( n - 1 )*abs( INCX ) ). */
 /*           Before entry, the incremented array X must contain the n */
 /*           element vector x. On exit, X is overwritten with the */
-/*           tranformed vector x. */
+/*           transformed vector x. */
 
 /*  INCX   - INTEGER. */
 /*           On entry, INCX specifies the increment for the elements of */
diff --git a/blas/f2c/stbmv.c b/blas/f2c/stbmv.c
index fcf9ce336..b5a68b545 100644
--- a/blas/f2c/stbmv.c
+++ b/blas/f2c/stbmv.c
@@ -143,7 +143,7 @@
 /*           ( 1 + ( n - 1 )*abs( INCX ) ). */
 /*           Before entry, the incremented array X must contain the n */
 /*           element vector x. On exit, X is overwritten with the */
-/*           tranformed vector x. */
+/*           transformed vector x. */
 
 /*  INCX   - INTEGER. */
 /*           On entry, INCX specifies the increment for the elements of */
diff --git a/blas/f2c/ztbmv.c b/blas/f2c/ztbmv.c
index 4cdcd7f88..3bf0beb01 100644
--- a/blas/f2c/ztbmv.c
+++ b/blas/f2c/ztbmv.c
@@ -147,7 +147,7 @@
 /*           ( 1 + ( n - 1 )*abs( INCX ) ). */
 /*           Before entry, the incremented array X must contain the n */
 /*           element vector x. On exit, X is overwritten with the */
-/*           tranformed vector x. */
+/*           transformed vector x. */
 
 /*  INCX   - INTEGER. */
 /*           On entry, INCX specifies the increment for the elements of */
diff --git a/blas/level1_impl.h b/blas/level1_impl.h
index f857bfa20..6e7f8c976 100644
--- a/blas/level1_impl.h
+++ b/blas/level1_impl.h
@@ -33,7 +33,7 @@ int EIGEN_BLAS_FUNC(copy)(int *n, RealScalar *px, int *incx, RealScalar *py, int
   Scalar* x = reinterpret_cast<Scalar*>(px);
   Scalar* y = reinterpret_cast<Scalar*>(py);
 
-  // be carefull, *incx==0 is allowed !!
+  // be careful, *incx==0 is allowed !!
   if(*incx==1 && *incy==1)
     make_vector(y,*n) = make_vector(x,*n);
   else
diff --git a/blas/testing/cblat1.f b/blas/testing/cblat1.f
index 8ca67fb19..73015f5a9 100644
--- a/blas/testing/cblat1.f
+++ b/blas/testing/cblat1.f
@@ -619,7 +619,7 @@
       SUBROUTINE STEST1(SCOMP1,STRUE1,SSIZE,SFAC)
 *     ************************* STEST1 *****************************
 *
-*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMODATE THE FORTRAN
+*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMMODATE THE FORTRAN
 *     REQUIREMENT THAT WHEN A DUMMY ARGUMENT IS AN ARRAY, THE
 *     ACTUAL ARGUMENT MUST ALSO BE AN ARRAY OR AN ARRAY ELEMENT.
 *
diff --git a/blas/testing/dblat1.f b/blas/testing/dblat1.f
index 30691f9bf..03d9f1345 100644
--- a/blas/testing/dblat1.f
+++ b/blas/testing/dblat1.f
@@ -990,7 +990,7 @@
       SUBROUTINE STEST1(SCOMP1,STRUE1,SSIZE,SFAC)
 *     ************************* STEST1 *****************************
 *
-*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMODATE THE FORTRAN
+*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMMODATE THE FORTRAN
 *     REQUIREMENT THAT WHEN A DUMMY ARGUMENT IS AN ARRAY, THE
 *     ACTUAL ARGUMENT MUST ALSO BE AN ARRAY OR AN ARRAY ELEMENT.
 *
diff --git a/blas/testing/sblat1.f b/blas/testing/sblat1.f
index 6657c2693..4d43d9b48 100644
--- a/blas/testing/sblat1.f
+++ b/blas/testing/sblat1.f
@@ -946,7 +946,7 @@
       SUBROUTINE STEST1(SCOMP1,STRUE1,SSIZE,SFAC)
 *     ************************* STEST1 *****************************
 *
-*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMODATE THE FORTRAN
+*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMMODATE THE FORTRAN
 *     REQUIREMENT THAT WHEN A DUMMY ARGUMENT IS AN ARRAY, THE
 *     ACTUAL ARGUMENT MUST ALSO BE AN ARRAY OR AN ARRAY ELEMENT.
 *
diff --git a/blas/testing/zblat1.f b/blas/testing/zblat1.f
index d30112c63..c00b67dc8 100644
--- a/blas/testing/zblat1.f
+++ b/blas/testing/zblat1.f
@@ -619,7 +619,7 @@
       SUBROUTINE STEST1(SCOMP1,STRUE1,SSIZE,SFAC)
 *     ************************* STEST1 *****************************
 *
-*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMODATE THE FORTRAN
+*     THIS IS AN INTERFACE SUBROUTINE TO ACCOMMODATE THE FORTRAN
 *     REQUIREMENT THAT WHEN A DUMMY ARGUMENT IS AN ARRAY, THE
 *     ACTUAL ARGUMENT MUST ALSO BE AN ARRAY OR AN ARRAY ELEMENT.
 *
diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake
index afc24b5e9..a2a4f54b9 100644
--- a/cmake/EigenConfigureTesting.cmake
+++ b/cmake/EigenConfigureTesting.cmake
@@ -20,7 +20,7 @@ include(CTest)
 set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command of unit tests")
 
 # Overwrite default DartConfiguration.tcl such that ctest can build our unit tests.
-# Recall that our unit tests are not in the "all" target, so we have to explicitely ask ctest to build our custom 'buildtests' target.
+# Recall that our unit tests are not in the "all" target, so we have to explicitly ask ctest to build our custom 'buildtests' target.
 # At this stage, we can also add custom flags to the build tool through the user defined EIGEN_TEST_BUILD_FLAGS variable.
 file(READ  "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" EIGEN_DART_CONFIG_FILE)
 # try to grab the default flags
@@ -39,7 +39,7 @@ ei_init_testing()
 
 # configure Eigen related testing options
 option(EIGEN_NO_ASSERTION_CHECKING "Disable checking of assertions using exceptions" OFF)
-option(EIGEN_DEBUG_ASSERTS "Enable advanced debuging of assertions" OFF)
+option(EIGEN_DEBUG_ASSERTS "Enable advanced debugging of assertions" OFF)
 
 if(CMAKE_COMPILER_IS_GNUCXX)
   option(EIGEN_COVERAGE_TESTING "Enable/disable gcov" OFF)
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 4a34ddef5..16d6d279f 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -247,7 +247,7 @@ endmacro(ei_add_test_internal_sycl)
 #
 # If EIGEN_SPLIT_LARGE_TESTS is ON, the test is split into multiple executables
 #   test_<testname>_<N>
-# where N runs from 1 to the greatest occurence found in the source file. Each of these
+# where N runs from 1 to the greatest occurrence found in the source file. Each of these
 # executables is built passing -DEIGEN_TEST_PART_N. This allows to split large tests
 # into smaller executables.
 #
@@ -269,8 +269,8 @@ macro(ei_add_test testname)
   file(READ "${filename}" test_source)
   set(parts 0)
   string(REGEX MATCHALL "CALL_SUBTEST_[0-9]+|EIGEN_TEST_PART_[0-9]+|EIGEN_SUFFIXES(;[0-9]+)+"
-         occurences "${test_source}")
-  string(REGEX REPLACE "CALL_SUBTEST_|EIGEN_TEST_PART_|EIGEN_SUFFIXES" "" suffixes "${occurences}")
+         occurrences "${test_source}")
+  string(REGEX REPLACE "CALL_SUBTEST_|EIGEN_TEST_PART_|EIGEN_SUFFIXES" "" suffixes "${occurrences}")
   list(REMOVE_DUPLICATES suffixes)
   if(EIGEN_SPLIT_LARGE_TESTS AND suffixes)
     add_custom_target(${testname})
@@ -303,8 +303,8 @@ macro(ei_add_test_sycl testname)
   file(READ "${filename}" test_source)
   set(parts 0)
   string(REGEX MATCHALL "CALL_SUBTEST_[0-9]+|EIGEN_TEST_PART_[0-9]+|EIGEN_SUFFIXES(;[0-9]+)+"
-         occurences "${test_source}")
-  string(REGEX REPLACE "CALL_SUBTEST_|EIGEN_TEST_PART_|EIGEN_SUFFIXES" "" suffixes "${occurences}")
+         occurrences "${test_source}")
+  string(REGEX REPLACE "CALL_SUBTEST_|EIGEN_TEST_PART_|EIGEN_SUFFIXES" "" suffixes "${occurrences}")
   list(REMOVE_DUPLICATES suffixes)
   if(EIGEN_SPLIT_LARGE_TESTS AND suffixes)
     add_custom_target(${testname})
diff --git a/cmake/FindComputeCpp.cmake b/cmake/FindComputeCpp.cmake
index e61dedc46..29f2a5007 100644
--- a/cmake/FindComputeCpp.cmake
+++ b/cmake/FindComputeCpp.cmake
@@ -243,7 +243,7 @@ endfunction()
 #######################
 #
 #  Adds a SYCL compilation custom command associated with an existing
-#  target and sets a dependancy on that new command.
+#  target and sets a dependency on that new command.
 #
 #  targetName : Name of the target to add a SYCL to.
 #  binaryDir : Intermediate directory to output the integration header.
diff --git a/cmake/FindEigen3.cmake b/cmake/FindEigen3.cmake
index 657440ba5..52efb4e15 100644
--- a/cmake/FindEigen3.cmake
+++ b/cmake/FindEigen3.cmake
@@ -15,7 +15,7 @@
 #  Eigen3::Eigen - The header-only Eigen library
 #
 # This module reads hints about search locations from 
-# the following enviroment variables:
+# the following environment variables:
 #
 # EIGEN3_ROOT
 # EIGEN3_ROOT_DIR
diff --git a/debug/msvc/eigen_autoexp_part.dat b/debug/msvc/eigen_autoexp_part.dat
index 07aa43739..35ef5807c 100644
--- a/debug/msvc/eigen_autoexp_part.dat
+++ b/debug/msvc/eigen_autoexp_part.dat
@@ -14,7 +14,7 @@
 ; * - Eigen::Matrix<*,-1,+,*,*,*>
 ; * - Eigen::Matrix<*,+,+,*,*,*>
 ; *
-; * Matrices are displayed properly independantly of the memory
+; * Matrices are displayed properly independently of the memory
 ; * alignment (RowMajor vs. ColMajor).
 ; *
 ; * This file is distributed WITHOUT ANY WARRANTY. Please ensure
diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in
index 2109978fe..49b9fba39 100644
--- a/doc/Doxyfile.in
+++ b/doc/Doxyfile.in
@@ -1764,7 +1764,7 @@ UML_LOOK               = YES
 # the class node. If there are many fields or methods and many nodes the
 # graph may become too big to be useful. The UML_LIMIT_NUM_FIELDS
 # threshold limits the number of items for each type to make the size more
-# managable. Set this to 0 for no limit. Note that the threshold may be
+# manageable. Set this to 0 for no limit. Note that the threshold may be
 # exceeded by 50% before the limit is enforced.
 
 UML_LIMIT_NUM_FIELDS   = 10
diff --git a/doc/FunctionsTakingEigenTypes.dox b/doc/FunctionsTakingEigenTypes.dox
index 152dda47d..e054714f9 100644
--- a/doc/FunctionsTakingEigenTypes.dox
+++ b/doc/FunctionsTakingEigenTypes.dox
@@ -133,7 +133,7 @@ In this special case, the example is fine and will be working because both param
 
 \section TopicPlainFunctionsFailing In which cases do functions taking a plain Matrix or Array argument fail?
 
-Here, we consider a slightly modified version of the function given above. This time, we do not want to return the result but pass an additional non-const paramter which allows us to store the result. A first naive implementation might look as follows.
+Here, we consider a slightly modified version of the function given above. This time, we do not want to return the result but pass an additional non-const parameter which allows us to store the result. A first naive implementation might look as follows.
 \code
 // Note: This code is flawed!
 void cov(const MatrixXf& x, const MatrixXf& y, MatrixXf& C)
@@ -176,7 +176,7 @@ The implementation above does now not only work with temporary expressions but i
 
 \section TopicResizingInGenericImplementations How to resize matrices in generic implementations?
 
-One might think we are done now, right? This is not completely true because in order for our covariance function to be generically applicable, we want the follwing code to work
+One might think we are done now, right? This is not completely true because in order for our covariance function to be generically applicable, we want the following code to work
 \code
 MatrixXf x = MatrixXf::Random(100,3);
 MatrixXf y = MatrixXf::Random(100,3);
diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox
index b6d08c700..b49f7d3cf 100644
--- a/doc/PreprocessorDirectives.dox
+++ b/doc/PreprocessorDirectives.dox
@@ -51,7 +51,7 @@ are doing.
 
 \section TopicPreprocessorDirectivesCppVersion C++ standard features
 
-By default, %Eigen strive to automatically detect and enable langage features at compile-time based on
+By default, %Eigen strive to automatically detect and enable language features at compile-time based on
 the information provided by the compiler.
 
  - \b EIGEN_MAX_CPP_VER - disables usage of C++ features requiring a version greater than EIGEN_MAX_CPP_VER.
diff --git a/doc/QuickStartGuide.dox b/doc/QuickStartGuide.dox
index ea32c3b3d..23bb2981b 100644
--- a/doc/QuickStartGuide.dox
+++ b/doc/QuickStartGuide.dox
@@ -68,7 +68,7 @@ The output is as follows:
 
 The second example starts by declaring a 3-by-3 matrix \c m which is initialized using the \link DenseBase::Random(Index,Index) Random() \endlink method with random values between -1 and 1. The next line applies a linear mapping such that the values are between 10 and 110. The function call \link DenseBase::Constant(Index,Index,const Scalar&) MatrixXd::Constant\endlink(3,3,1.2) returns a 3-by-3 matrix expression having all coefficients equal to 1.2. The rest is standard arithmetics.
 
-The next line of the \c main function introduces a new type: \c VectorXd. This represents a (column) vector of arbitrary size. Here, the vector \c v is created to contain \c 3 coefficients which are left unitialized. The one but last line uses the so-called comma-initializer, explained in \ref TutorialAdvancedInitialization, to set all coefficients of the vector \c v to be as follows:
+The next line of the \c main function introduces a new type: \c VectorXd. This represents a (column) vector of arbitrary size. Here, the vector \c v is created to contain \c 3 coefficients which are left uninitialized. The one but last line uses the so-called comma-initializer, explained in \ref TutorialAdvancedInitialization, to set all coefficients of the vector \c v to be as follows:
 
 \f[
 v =
diff --git a/doc/SparseQuickReference.dox b/doc/SparseQuickReference.dox
index a25622e80..81a73eec2 100644
--- a/doc/SparseQuickReference.dox
+++ b/doc/SparseQuickReference.dox
@@ -80,7 +80,7 @@ sm1.setZero();
 
 
 \section SparseBasicInfos Matrix properties
-Beyond the basic functions rows() and cols(), there are some useful functions that are available to easily get some informations from the matrix. 
+Beyond the basic functions rows() and cols(), there are some useful functions that are available to easily get some information from the matrix. 
 <table class="manual">
 <tr>
   <td> \code
diff --git a/doc/TemplateKeyword.dox b/doc/TemplateKeyword.dox
index b84cfdae9..fbf2c7081 100644
--- a/doc/TemplateKeyword.dox
+++ b/doc/TemplateKeyword.dox
@@ -76,7 +76,7 @@ point where the template is defined, without knowing the actual value of the tem
 and \c Derived2 in the example). That means that the compiler cannot know that <tt>dst.triangularView</tt> is
 a member template and that the following &lt; symbol is part of the delimiter for the template
 parameter. Another possibility would be that <tt>dst.triangularView</tt> is a member variable with the &lt;
-symbol refering to the <tt>operator&lt;()</tt> function. In fact, the compiler should choose the second
+symbol referring to the <tt>operator&lt;()</tt> function. In fact, the compiler should choose the second
 possibility, according to the standard. If <tt>dst.triangularView</tt> is a member template (as in our case),
 the programmer should specify this explicitly with the \c template keyword and write <tt>dst.template
 triangularView</tt>.
diff --git a/doc/TopicLazyEvaluation.dox b/doc/TopicLazyEvaluation.dox
index 101ef8c72..b7820e3e6 100644
--- a/doc/TopicLazyEvaluation.dox
+++ b/doc/TopicLazyEvaluation.dox
@@ -58,7 +58,7 @@ the product <tt>matrix3 * matrix4</tt> gets evaluated immediately into a tempora
 
 \code matrix1 = matrix2 * (matrix3 + matrix4); \endcode
 
-Here, provided the matrices have at least 2 rows and 2 columns, each coefficienct of the expression <tt>matrix3 + matrix4</tt> is going to be used several times in the matrix product. Instead of computing the sum everytime, it is much better to compute it once and store it in a temporary variable. Eigen understands this and evaluates <tt>matrix3 + matrix4</tt> into a temporary variable before evaluating the product.
+Here, provided the matrices have at least 2 rows and 2 columns, each coefficienct of the expression <tt>matrix3 + matrix4</tt> is going to be used several times in the matrix product. Instead of computing the sum every time, it is much better to compute it once and store it in a temporary variable. Eigen understands this and evaluates <tt>matrix3 + matrix4</tt> into a temporary variable before evaluating the product.
 
 */
 
diff --git a/doc/TopicLinearAlgebraDecompositions.dox b/doc/TopicLinearAlgebraDecompositions.dox
index 491470627..991f964cc 100644
--- a/doc/TopicLinearAlgebraDecompositions.dox
+++ b/doc/TopicLinearAlgebraDecompositions.dox
@@ -248,7 +248,7 @@ To get an overview of the true relative speed of the different decomposition, ch
   <dt><b>Blocking</b></dt>
     <dd>Means the algorithm can work per block, whence guaranteeing a good scaling of the performance for large matrices.</dd>
   <dt><b>Implicit Multi Threading (MT)</b></dt>
-    <dd>Means the algorithm can take advantage of multicore processors via OpenMP. "Implicit" means the algortihm itself is not parallelized, but that it relies on parallelized matrix-matrix product rountines.</dd>
+    <dd>Means the algorithm can take advantage of multicore processors via OpenMP. "Implicit" means the algortihm itself is not parallelized, but that it relies on parallelized matrix-matrix product routines.</dd>
   <dt><b>Explicit Multi Threading (MT)</b></dt>
     <dd>Means the algorithm is explicitly parallelized to take advantage of multicore processors via OpenMP.</dd>
   <dt><b>Meta-unroller</b></dt>
diff --git a/doc/TopicMultithreading.dox b/doc/TopicMultithreading.dox
index 47c9b261f..bc394f484 100644
--- a/doc/TopicMultithreading.dox
+++ b/doc/TopicMultithreading.dox
@@ -47,7 +47,7 @@ int main(int argc, char** argv)
 
 \warning note that all functions generating random matrices are \b not re-entrant nor thread-safe. Those include DenseBase::Random(), and DenseBase::setRandom() despite a call to Eigen::initParallel(). This is because these functions are based on std::rand which is not re-entrant. For thread-safe random generator, we recommend the use of boost::random or c++11 random feature.
 
-In the case your application is parallelized with OpenMP, you might want to disable Eigen's own parallization as detailed in the previous section.
+In the case your application is parallelized with OpenMP, you might want to disable Eigen's own parallelization as detailed in the previous section.
 
 */
 
diff --git a/doc/TutorialMapClass.dox b/doc/TutorialMapClass.dox
index f8fb0fd2f..caa2539d8 100644
--- a/doc/TutorialMapClass.dox
+++ b/doc/TutorialMapClass.dox
@@ -29,9 +29,9 @@ Map<const Vector4i> mi(pi);
 \endcode
 where \c pi is an \c int \c *. In this case the size does not have to be passed to the constructor, because it is already specified by the Matrix/Array type.
 
-Note that Map does not have a default constructor; you \em must pass a pointer to intialize the object. However, you can work around this requirement (see \ref TutorialMapPlacementNew).
+Note that Map does not have a default constructor; you \em must pass a pointer to initialize the object. However, you can work around this requirement (see \ref TutorialMapPlacementNew).
 
-Map is flexible enough to accomodate a variety of different data representations.  There are two other (optional) template parameters:
+Map is flexible enough to accommodate a variety of different data representations.  There are two other (optional) template parameters:
 \code
 Map<typename MatrixType,
     int MapOptions,
diff --git a/doc/TutorialSparse.dox b/doc/TutorialSparse.dox
index 352907408..350ea1139 100644
--- a/doc/TutorialSparse.dox
+++ b/doc/TutorialSparse.dox
@@ -57,7 +57,7 @@ The \c "_" indicates available free space to quickly insert new elements.
 Assuming no reallocation is needed, the insertion of a random element is therefore in O(nnz_j) where nnz_j is the number of nonzeros of the respective inner vector.
 On the other hand, inserting elements with increasing inner indices in a given inner vector is much more efficient since this only requires to increase the respective \c InnerNNZs entry that is a O(1) operation.
 
-The case where no empty space is available is a special case, and is refered as the \em compressed mode.
+The case where no empty space is available is a special case, and is referred as the \em compressed mode.
 It corresponds to the widely used Compressed Column (or Row) Storage schemes (CCS or CRS).
 Any SparseMatrix can be turned to this form by calling the SparseMatrix::makeCompressed() function.
 In this case, one can remark that the \c InnerNNZs array is redundant with \c OuterStarts because we the equality: \c InnerNNZs[j] = \c OuterStarts[j+1]-\c OuterStarts[j].
@@ -212,7 +212,7 @@ See the SparseMatrix::setFromTriplets() function and class Triplet for more deta
 
 
 In some cases, however, slightly higher performance, and lower memory consumption can be reached by directly inserting the non-zeros into the destination matrix.
-A typical scenario of this approach is illustrated bellow:
+A typical scenario of this approach is illustrated below:
 \code
 1: SparseMatrix<double> mat(rows,cols);         // default is column major
 2: mat.reserve(VectorXi::Constant(cols,6));
diff --git a/doc/UnalignedArrayAssert.dox b/doc/UnalignedArrayAssert.dox
index 0f7022973..8676faa1b 100644
--- a/doc/UnalignedArrayAssert.dox
+++ b/doc/UnalignedArrayAssert.dox
@@ -117,8 +117,8 @@ It doesn't disable 16-byte alignment, because that would mean that vectorized an
 
 \section checkmycode How can I check my code is safe regarding alignment issues?
 
-Unfortunately, there is no possibility in C++ to detect any of the aformentioned shortcoming at compile time (though static analysers are becoming more and more powerful and could detect some of them).
-Even at runtime, all we can do is to catch invalid unaligned allocation and trigger the explicit assertion mentioned at the begining of this page.
+Unfortunately, there is no possibility in C++ to detect any of the aforementioned shortcoming at compile time (though static analysers are becoming more and more powerful and could detect some of them).
+Even at runtime, all we can do is to catch invalid unaligned allocation and trigger the explicit assertion mentioned at the beginning of this page.
 Therefore, if your program runs fine on a given system with some given compilation flags, then this does not guarantee that your code is safe. For instance, on most 64 bits systems buffer are aligned on 16 bytes boundary and so, if you do not enable AVX instruction set, then your code will run fine. On the other hand, the same code may assert if moving to a more exotic platform, or enabling AVX instructions that required 32 bytes alignment by default.
 
 The situation is not hopeless though. Assuming your code is well covered by unit test, then you can check its alignment safety by linking it to a custom malloc library returning 8 bytes aligned buffers only. This way all alignment shortcomings should pop-up. To this end, you must also compile your program with \link TopicPreprocessorDirectivesPerformance EIGEN_MALLOC_ALREADY_ALIGNED=0 \endlink.
diff --git a/doc/UsingNVCC.dox b/doc/UsingNVCC.dox
index 9bcdf0bfc..36beb2ddd 100644
--- a/doc/UsingNVCC.dox
+++ b/doc/UsingNVCC.dox
@@ -5,7 +5,7 @@ namespace Eigen {
 
 Staring from CUDA 5.5 and Eigen 3.3, it is possible to use Eigen's matrices, vectors, and arrays for fixed size within CUDA kernels. This is especially useful when working on numerous but small problems. By default, when Eigen's headers are included within a .cu file compiled by nvcc most Eigen's functions and methods are prefixed by the \c __device__ \c __host__ keywords making them callable from both host and device code.
 This support can be disabled by defining \c EIGEN_NO_CUDA before including any Eigen's header.
-This might be usefull to disable some warnings when a .cu file makes use of Eigen on the host side only.
+This might be useful to disable some warnings when a .cu file makes use of Eigen on the host side only.
 However, in both cases, host's SIMD vectorization has to be disabled in .cu files.
 It is thus \b strongly \b recommended to properly move all costly host computation from your .cu files to regular .cpp files.
 
diff --git a/doc/eigendoxy.css b/doc/eigendoxy.css
index 6ce2b839b..b99d7914a 100644
--- a/doc/eigendoxy.css
+++ b/doc/eigendoxy.css
@@ -93,7 +93,7 @@ table th.inter {
 	border-color: #cccccc;
 }
 
-/** class for exemple / output tables **/
+/** class for example / output tables **/
 
 table.example {
 }
diff --git a/doc/special_examples/Tutorial_sparse_example.cpp b/doc/special_examples/Tutorial_sparse_example.cpp
index 830e196ea..89937b411 100644
--- a/doc/special_examples/Tutorial_sparse_example.cpp
+++ b/doc/special_examples/Tutorial_sparse_example.cpp
@@ -12,7 +12,7 @@ int main(int argc, char** argv)
   assert(argc==2);
   
   int n = 300;  // size of the image
-  int m = n*n;  // number of unknows (=number of pixels)
+  int m = n*n;  // number of unknowns (=number of pixels)
 
   // Assembly:
   std::vector<T> coefficients;            // list of non-zeros coefficients
diff --git a/lapack/CMakeLists.txt b/lapack/CMakeLists.txt
index 9883d4c72..52f18edfc 100644
--- a/lapack/CMakeLists.txt
+++ b/lapack/CMakeLists.txt
@@ -35,7 +35,7 @@ set(EigenLapack_SRCS  ${EigenLapack_SRCS}
   second_NONE.f dsecnd_NONE.f
 )
 
-option(EIGEN_ENABLE_LAPACK_TESTS OFF "Enbale the Lapack unit tests")
+option(EIGEN_ENABLE_LAPACK_TESTS OFF "Enable the Lapack unit tests")
 
 if(EIGEN_ENABLE_LAPACK_TESTS)
 
@@ -59,7 +59,7 @@ if(EIGEN_ENABLE_LAPACK_TESTS)
       message(STATUS "Setup lapack reference and lapack unit tests")
       execute_process(COMMAND tar xzf  "lapack_addons_3.4.1.tgz" WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
     else()
-      message(STATUS "Download of lapack_addons_3.4.1.tgz failed, LAPACK unit tests wont be enabled")
+      message(STATUS "Download of lapack_addons_3.4.1.tgz failed, LAPACK unit tests won't be enabled")
       set(EIGEN_ENABLE_LAPACK_TESTS false)
     endif()
                   
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 8bd086ce3..8bcf3f7c5 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -1,5 +1,5 @@
 # generate split test header file only if it does not yet exist
-# in order to prevent a rebuild everytime cmake is configured
+# in order to prevent a rebuild every time cmake is configured
 if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)  
   file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
   foreach(i RANGE 1 999)
diff --git a/test/bdcsvd.cpp b/test/bdcsvd.cpp
index 6c7b09696..109218766 100644
--- a/test/bdcsvd.cpp
+++ b/test/bdcsvd.cpp
@@ -104,7 +104,7 @@ void test_bdcsvd()
   CALL_SUBTEST_7( BDCSVD<MatrixXf>(10,10) );
 
   // Check that preallocation avoids subsequent mallocs
-  // Disbaled because not supported by BDCSVD
+  // Disabled because not supported by BDCSVD
   // CALL_SUBTEST_9( svd_preallocate<void>() );
 
   CALL_SUBTEST_2( svd_underoverflow<void>() );
diff --git a/test/eigensolver_complex.cpp b/test/eigensolver_complex.cpp
index 293b1b265..03d5774ef 100644
--- a/test/eigensolver_complex.cpp
+++ b/test/eigensolver_complex.cpp
@@ -47,7 +47,7 @@ template<typename MatrixType> bool find_pivot(typename MatrixType::Scalar tol, M
   return false;
 }
 
-/* Check that two column vectors are approximately equal upto permutations.
+/* Check that two column vectors are approximately equal up to permutations.
  * Initially, this method checked that the k-th power sums are equal for all k = 1, ..., vec1.rows(),
  * however this strategy is numerically inacurate because of numerical cancellation issues.
  */
diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp
index 8ee8fdb27..5854d39c5 100644
--- a/test/geo_quaternion.cpp
+++ b/test/geo_quaternion.cpp
@@ -241,7 +241,7 @@ template<typename Scalar> void mapQuaternion(void){
   const MQuaternionUA& cmq3(mq3);
   VERIFY( &cmq3.x() == &mq3.x() );
   // FIXME the following should be ok. The problem is that currently the LValueBit flag
-  // is used to determine wether we can return a coeff by reference or not, which is not enough for Map<const ...>.
+  // is used to determine whether we can return a coeff by reference or not, which is not enough for Map<const ...>.
   //const MCQuaternionUA& cmcq3(mcq3);
   //VERIFY( &cmcq3.x() == &mcq3.x() );
 }
diff --git a/test/main.h b/test/main.h
index 6079cbd06..14f4e3e7a 100644
--- a/test/main.h
+++ b/test/main.h
@@ -183,7 +183,7 @@ namespace Eigen
     };
   }
   // If EIGEN_DEBUG_ASSERTS is defined and if no assertion is triggered while
-  // one should have been, then the list of excecuted assertions is printed out.
+  // one should have been, then the list of executed assertions is printed out.
   //
   // EIGEN_DEBUG_ASSERTS is not enabled by default as it
   // significantly increases the compilation time
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 08b360340..3c11df7e8 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -28,7 +28,7 @@ template<typename T> T negate(const T& x) { return -x; }
 }
 }
 
-// NOTE: we disbale inlining for this function to workaround a GCC issue when using -O3 and the i387 FPU.
+// NOTE: we disable inlining for this function to workaround a GCC issue when using -O3 and the i387 FPU.
 template<typename Scalar> EIGEN_DONT_INLINE
 bool isApproxAbs(const Scalar& a, const Scalar& b, const typename NumTraits<Scalar>::Real& refvalue)
 {

From baf9a5a776b45a2d9f330afb4c716db16c2a3fb8 Mon Sep 17 00:00:00 2001
From: vhuber <anonymous@invalid.net>
Date: Fri, 30 Mar 2018 18:53:34 +0200
Subject: [PATCH 265/680] Add interface to umfpack_*l_* functions

---
 Eigen/src/UmfPackSupport/UmfPackSupport.h | 186 ++++++++++++++++++----
 1 file changed, 156 insertions(+), 30 deletions(-)

diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h
index 9568cc1d5..fa62636e9 100644
--- a/Eigen/src/UmfPackSupport/UmfPackSupport.h
+++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -12,47 +12,92 @@
 
 namespace Eigen {
 
+#define UF_long __int64
+
 /* TODO extract L, extract U, compute det, etc... */
 
 // generic double/complex<double> wrapper functions:
 
 
-inline void umfpack_defaults(double control[UMFPACK_CONTROL], double)
+ // Defaults
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], double, int)
 { umfpack_di_defaults(control); }
 
-inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>)
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>, int)
 { umfpack_zi_defaults(control); }
 
-inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double)
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], double, UF_long)
+{ umfpack_dl_defaults(control); }
+
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>, UF_long)
+{ umfpack_zl_defaults(control); }
+
+// Report info
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double, int)
 { umfpack_di_report_info(control, info);}
 
-inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>)
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>, int)
 { umfpack_zi_report_info(control, info);}
 
-inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double)
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double, UF_long)
+{ umfpack_dl_report_info(control, info);}
+
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>, UF_long)
+{ umfpack_zl_report_info(control, info);}
+
+// Report status
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double, int)
 { umfpack_di_report_status(control, status);}
 
-inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>)
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>, int)
 { umfpack_zi_report_status(control, status);}
 
-inline void umfpack_report_control(double control[UMFPACK_CONTROL], double)
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double, UF_long)
+{ umfpack_dl_report_status(control, status);}
+
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>, UF_long)
+{ umfpack_zl_report_status(control, status);}
+
+// report control
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], double, int)
 { umfpack_di_report_control(control);}
 
-inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>)
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>, int)
 { umfpack_zi_report_control(control);}
 
-inline void umfpack_free_numeric(void **Numeric, double)
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], double, UF_long)
+{ umfpack_dl_report_control(control);}
+
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>, UF_long)
+{ umfpack_zl_report_control(control);}
+
+// Free numeric
+inline void umfpack_free_numeric(void **Numeric, double, int)
 { umfpack_di_free_numeric(Numeric); *Numeric = 0; }
 
-inline void umfpack_free_numeric(void **Numeric, std::complex<double>)
+inline void umfpack_free_numeric(void **Numeric, std::complex<double>, int)
 { umfpack_zi_free_numeric(Numeric); *Numeric = 0; }
 
-inline void umfpack_free_symbolic(void **Symbolic, double)
+inline void umfpack_free_numeric(void **Numeric, double, UF_long)
+{ umfpack_dl_free_numeric(Numeric); *Numeric = 0; }
+
+inline void umfpack_free_numeric(void **Numeric, std::complex<double>, UF_long)
+{ umfpack_zl_free_numeric(Numeric); *Numeric = 0; }
+
+// Free symbolic
+inline void umfpack_free_symbolic(void **Symbolic, double, int)
 { umfpack_di_free_symbolic(Symbolic); *Symbolic = 0; }
 
-inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>)
+inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>, int)
 { umfpack_zi_free_symbolic(Symbolic); *Symbolic = 0; }
 
+inline void umfpack_free_symbolic(void **Symbolic, double, UF_long)
+{ umfpack_dl_free_symbolic(Symbolic); *Symbolic = 0; }
+
+inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>, UF_long)
+{ umfpack_zl_free_symbolic(Symbolic); *Symbolic = 0; }
+
+// Symbolic
 inline int umfpack_symbolic(int n_row,int n_col,
                             const int Ap[], const int Ai[], const double Ax[], void **Symbolic,
                             const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
@@ -66,7 +111,21 @@ inline int umfpack_symbolic(int n_row,int n_col,
 {
   return umfpack_zi_symbolic(n_row,n_col,Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Control,Info);
 }
+inline int umfpack_symbolic(UF_long n_row,UF_long n_col,
+                            const UF_long Ap[], const UF_long Ai[], const double Ax[], void **Symbolic,
+                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+{
+  return umfpack_dl_symbolic(n_row,n_col,Ap,Ai,Ax,Symbolic,Control,Info);
+}
 
+inline int umfpack_symbolic(UF_long n_row,UF_long n_col,
+                            const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[], void **Symbolic,
+                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+{
+  return umfpack_zl_symbolic(n_row,n_col,Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Control,Info);
+}
+
+// Numeric
 inline int umfpack_numeric( const int Ap[], const int Ai[], const double Ax[],
                             void *Symbolic, void **Numeric,
                             const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
@@ -80,7 +139,21 @@ inline int umfpack_numeric( const int Ap[], const int Ai[], const std::complex<d
 {
   return umfpack_zi_numeric(Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Numeric,Control,Info);
 }
+inline int umfpack_numeric( const UF_long Ap[], const UF_long Ai[], const double Ax[],
+                            void *Symbolic, void **Numeric,
+                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+{
+  return umfpack_dl_numeric(Ap,Ai,Ax,Symbolic,Numeric,Control,Info);
+}
 
+inline int umfpack_numeric( const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[],
+                            void *Symbolic, void **Numeric,
+                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+{
+  return umfpack_zl_numeric(Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Numeric,Control,Info);
+}
+
+// solve
 inline int umfpack_solve( int sys, const int Ap[], const int Ai[], const double Ax[],
                           double X[], const double B[], void *Numeric,
                           const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
@@ -95,6 +168,21 @@ inline int umfpack_solve( int sys, const int Ap[], const int Ai[], const std::co
   return umfpack_zi_solve(sys,Ap,Ai,&numext::real_ref(Ax[0]),0,&numext::real_ref(X[0]),0,&numext::real_ref(B[0]),0,Numeric,Control,Info);
 }
 
+inline int umfpack_solve( int sys, const UF_long Ap[], const UF_long Ai[], const double Ax[],
+                          double X[], const double B[], void *Numeric,
+                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+{
+  return umfpack_dl_solve(sys,Ap,Ai,Ax,X,B,Numeric,Control,Info);
+}
+
+inline int umfpack_solve( int sys, const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[],
+                          std::complex<double> X[], const std::complex<double> B[], void *Numeric,
+                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+{
+  return umfpack_zl_solve(sys,Ap,Ai,&numext::real_ref(Ax[0]),0,&numext::real_ref(X[0]),0,&numext::real_ref(B[0]),0,Numeric,Control,Info);
+}
+
+// Get Lunz
 inline int umfpack_get_lunz(int *lnz, int *unz, int *n_row, int *n_col, int *nz_udiag, void *Numeric, double)
 {
   return umfpack_di_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
@@ -105,6 +193,17 @@ inline int umfpack_get_lunz(int *lnz, int *unz, int *n_row, int *n_col, int *nz_
   return umfpack_zi_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
 }
 
+inline int umfpack_get_lunz(UF_long *lnz, UF_long *unz, UF_long *n_row, UF_long *n_col, UF_long *nz_udiag, void *Numeric, double)
+{
+  return umfpack_dl_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
+}
+
+inline int umfpack_get_lunz(UF_long *lnz, UF_long *unz, UF_long *n_row, UF_long *n_col, UF_long *nz_udiag, void *Numeric, std::complex<double>)
+{
+  return umfpack_zl_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
+}
+
+// Get Numeric
 inline int umfpack_get_numeric(int Lp[], int Lj[], double Lx[], int Up[], int Ui[], double Ux[],
                                int P[], int Q[], double Dx[], int *do_recip, double Rs[], void *Numeric)
 {
@@ -120,18 +219,45 @@ inline int umfpack_get_numeric(int Lp[], int Lj[], std::complex<double> Lx[], in
   return umfpack_zi_get_numeric(Lp,Lj,Lx?&lx0_real:0,0,Up,Ui,Ux?&ux0_real:0,0,P,Q,
                                 Dx?&dx0_real:0,0,do_recip,Rs,Numeric);
 }
+inline int umfpack_get_numeric(UF_long Lp[], UF_long Lj[], double Lx[], UF_long Up[], UF_long Ui[], double Ux[],
+                               UF_long P[], UF_long Q[], double Dx[], UF_long *do_recip, double Rs[], void *Numeric)
+{
+  return umfpack_dl_get_numeric(Lp,Lj,Lx,Up,Ui,Ux,P,Q,Dx,do_recip,Rs,Numeric);
+}
 
-inline int umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO])
+inline int umfpack_get_numeric(UF_long Lp[], UF_long Lj[], std::complex<double> Lx[], UF_long Up[], UF_long Ui[], std::complex<double> Ux[],
+                               UF_long P[], UF_long Q[], std::complex<double> Dx[], UF_long *do_recip, double Rs[], void *Numeric)
+{
+  double& lx0_real = numext::real_ref(Lx[0]);
+  double& ux0_real = numext::real_ref(Ux[0]);
+  double& dx0_real = numext::real_ref(Dx[0]);
+  return umfpack_zl_get_numeric(Lp,Lj,Lx?&lx0_real:0,0,Up,Ui,Ux?&ux0_real:0,0,P,Q,
+                                Dx?&dx0_real:0,0,do_recip,Rs,Numeric);
+}
+
+// Get Determinant
+inline int umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], int)
 {
   return umfpack_di_get_determinant(Mx,Ex,NumericHandle,User_Info);
 }
 
-inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO])
+inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], int)
 {
   double& mx_real = numext::real_ref(*Mx);
   return umfpack_zi_get_determinant(&mx_real,0,Ex,NumericHandle,User_Info);
 }
 
+inline int umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], UF_long)
+{
+  return umfpack_dl_get_determinant(Mx,Ex,NumericHandle,User_Info);
+}
+
+inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], UF_long)
+{
+  double& mx_real = numext::real_ref(*Mx);
+  return umfpack_zl_get_determinant(&mx_real,0,Ex,NumericHandle,User_Info);
+}
+
 
 /** \ingroup UmfPackSupport_Module
   * \brief A sparse LU factorization and solver based on UmfPack
@@ -164,7 +290,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
     typedef Matrix<int, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
     typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
     typedef SparseMatrix<Scalar> LUMatrixType;
-    typedef SparseMatrix<Scalar,ColMajor,int> UmfpackMatrixType;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> UmfpackMatrixType;
     typedef Ref<const UmfpackMatrixType, StandardCompressedFormat> UmfpackMatrixRef;
     enum {
       ColsAtCompileTime = MatrixType::ColsAtCompileTime,
@@ -192,8 +318,8 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
 
     ~UmfPackLU()
     {
-      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
-      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar(), StorageIndex());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar(), StorageIndex());
     }
 
     inline Index rows() const { return mp_matrix.rows(); }
@@ -241,8 +367,8 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
     template<typename InputMatrixType>
     void compute(const InputMatrixType& matrix)
     {
-      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
-      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar(),StorageIndex());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar(),StorageIndex());
       grab(matrix.derived());
       analyzePattern_impl();
       factorize_impl();
@@ -257,8 +383,8 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
     template<typename InputMatrixType>
     void analyzePattern(const InputMatrixType& matrix)
     {
-      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
-      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar(),StorageIndex());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar(),StorageIndex());
 
       grab(matrix.derived());
 
@@ -309,7 +435,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
     {
       eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
       if(m_numeric)
-        umfpack_free_numeric(&m_numeric,Scalar());
+        umfpack_free_numeric(&m_numeric,Scalar(),StorageIndex());
 
       grab(matrix.derived());
 
@@ -322,7 +448,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
       */
     void printUmfpackControl()
     {
-      umfpack_report_control(m_control.data(), Scalar());
+      umfpack_report_control(m_control.data(), Scalar(),StorageIndex());
     }
 
     /** Prints statistics collected by UmfPack.
@@ -332,7 +458,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
     void printUmfpackInfo()
     {
       eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
-      umfpack_report_info(m_control.data(), m_umfpackInfo.data(), Scalar());
+      umfpack_report_info(m_control.data(), m_umfpackInfo.data(), Scalar(),StorageIndex());
     }
 
     /** Prints the status of the previous factorization operation performed by UmfPack (symbolic or numerical factorization).
@@ -341,7 +467,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
       */
     void printUmfpackStatus() {
       eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
-      umfpack_report_status(m_control.data(), m_fact_errorCode, Scalar());
+      umfpack_report_status(m_control.data(), m_fact_errorCode, Scalar(),StorageIndex());
     }
 
     /** \internal */
@@ -362,13 +488,13 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
       m_symbolic              = 0;
       m_extractedDataAreDirty = true;
 
-      umfpack_defaults(m_control.data(), Scalar());
+      umfpack_defaults(m_control.data(), Scalar(),StorageIndex());
     }
 
     void analyzePattern_impl()
     {
-      m_fact_errorCode = umfpack_symbolic(internal::convert_index<int>(mp_matrix.rows()),
-                                          internal::convert_index<int>(mp_matrix.cols()),
+      m_fact_errorCode = umfpack_symbolic(internal::convert_index<StorageIndex>(mp_matrix.rows()),
+                                          internal::convert_index<StorageIndex>(mp_matrix.cols()),
                                           mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
                                           &m_symbolic, m_control.data(), m_umfpackInfo.data());
 
@@ -438,7 +564,7 @@ void UmfPackLU<MatrixType>::extractData() const
   if (m_extractedDataAreDirty)
   {
     // get size of the data
-    int lnz, unz, rows, cols, nz_udiag;
+    StorageIndex lnz, unz, rows, cols, nz_udiag;
     umfpack_get_lunz(&lnz, &unz, &rows, &cols, &nz_udiag, m_numeric, Scalar());
 
     // allocate data
@@ -464,7 +590,7 @@ template<typename MatrixType>
 typename UmfPackLU<MatrixType>::Scalar UmfPackLU<MatrixType>::determinant() const
 {
   Scalar det;
-  umfpack_get_determinant(&det, 0, m_numeric, 0);
+  umfpack_get_determinant(&det, 0, m_numeric, 0, StorageIndex());
   return det;
 }
 

From 267a144da5145d735934aab0a249418d39ce3f9b Mon Sep 17 00:00:00 2001
From: vhuber <anonymous@invalid.net>
Date: Fri, 30 Mar 2018 23:04:53 +0200
Subject: [PATCH 266/680] Remove unnecessary define

---
 Eigen/src/UmfPackSupport/UmfPackSupport.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h
index fa62636e9..ca2f73dbe 100644
--- a/Eigen/src/UmfPackSupport/UmfPackSupport.h
+++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -12,8 +12,6 @@
 
 namespace Eigen {
 
-#define UF_long __int64
-
 /* TODO extract L, extract U, compute det, etc... */
 
 // generic double/complex<double> wrapper functions:

From 524119d32a2ac354a384717da38526087a242eb0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 10:56:10 +0200
Subject: [PATCH 267/680] Fix uninitialized output argument.

---
 unsupported/Eigen/src/SparseExtra/MarketIO.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/src/SparseExtra/MarketIO.h b/unsupported/Eigen/src/SparseExtra/MarketIO.h
index fc70a24d8..6d57ab2e9 100644
--- a/unsupported/Eigen/src/SparseExtra/MarketIO.h
+++ b/unsupported/Eigen/src/SparseExtra/MarketIO.h
@@ -109,6 +109,7 @@ namespace internal
 inline bool getMarketHeader(const std::string& filename, int& sym, bool& iscomplex, bool& isvector)
 {
   sym = 0; 
+  iscomplex = false;
   isvector = false;
   std::ifstream in(filename.c_str(),std::ios::in);
   if(!in)

From 6719409cd92b19acabc4544f9ac5571a2ff9a88f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:11:56 +0200
Subject: [PATCH 268/680] AVX512: add missing pinsertfirst and pinsertlast,
 implement pblend for Packet8d, fix compilation without AVX512DQ

---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 51 ++++++++++++++++++++++---
 1 file changed, 46 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 12b897572..abece01bc 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -54,6 +54,7 @@ template<> struct packet_traits<float>  : default_packet_traits
     AlignedOnScalar = 1,
     size = 16,
     HasHalfPacket = 1,
+    HasBlend = 0,
 #if EIGEN_GNUC_AT_LEAST(5, 3)
 #ifdef EIGEN_VECTORIZE_AVX512DQ
     HasLog = 1,
@@ -470,6 +471,8 @@ EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) {
   __m512 pairs = _mm512_permute_ps(even_elements, _MM_SHUFFLE(2, 2, 0, 0));
   return pairs;
 }
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
 // Loads 4 doubles from memory a returns the packet {a0, a0  a1, a1, a2, a2, a3,
 // a3}
 template <>
@@ -481,6 +484,17 @@ EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) {
   x = _mm512_insertf64x2(x, _mm_loaddup_pd(&from[3]), 3);
   return x;
 }
+#else
+template <>
+EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) {
+  __m512d x = _mm512_setzero_pd();
+  x = _mm512_mask_broadcastsd_pd(x, 0x3<<0, _mm_load_sd(from+0));
+  x = _mm512_mask_broadcastsd_pd(x, 0x3<<2, _mm_load_sd(from+1));
+  x = _mm512_mask_broadcastsd_pd(x, 0x3<<4, _mm_load_sd(from+2));
+  x = _mm512_mask_broadcastsd_pd(x, 0x3<<6, _mm_load_sd(from+3));
+  return x;
+}
+#endif
 
 // Loads 4 floats from memory a returns the packet
 // {a0, a0  a0, a0, a1, a1, a1, a1, a2, a2, a2, a2, a3, a3, a3, a3}
@@ -1272,11 +1286,38 @@ EIGEN_STRONG_INLINE Packet16f pblend(const Selector<16>& /*ifPacket*/,
   return Packet16f();
 }
 template <>
-EIGEN_STRONG_INLINE Packet8d pblend(const Selector<8>& /*ifPacket*/,
-                                    const Packet8d& /*thenPacket*/,
-                                    const Packet8d& /*elsePacket*/) {
-  assert(false && "To be implemented");
-  return Packet8d();
+EIGEN_STRONG_INLINE Packet8d pblend(const Selector<8>& ifPacket,
+                                    const Packet8d& thenPacket,
+                                    const Packet8d& elsePacket) {
+  __mmask8 m = (ifPacket.select[0]   )
+             | (ifPacket.select[1]<<1)
+             | (ifPacket.select[2]<<2)
+             | (ifPacket.select[3]<<3)
+             | (ifPacket.select[4]<<4)
+             | (ifPacket.select[5]<<5)
+             | (ifPacket.select[6]<<6)
+             | (ifPacket.select[7]<<7);
+  return _mm512_mask_blend_pd(m, elsePacket, thenPacket);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f pinsertfirst(const Packet16f& a, float b)
+{
+  return _mm512_mask_broadcastss_ps(a, (1), _mm_load_ss(&b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d pinsertfirst(const Packet8d& a, double b)
+{
+  return _mm512_mask_broadcastsd_pd(a, (1), _mm_load_sd(&b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f pinsertlast(const Packet16f& a, float b)
+{
+  return _mm512_mask_broadcastss_ps(a, (1<<15), _mm_load_ss(&b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d pinsertlast(const Packet8d& a, double b)
+{
+  return _mm512_mask_broadcastsd_pd(a, (1<<7), _mm_load_sd(&b));
 }
 
 } // end namespace internal

From 7b0630315f343422b37f62f40a039c9e725fe9e1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:12:50 +0200
Subject: [PATCH 269/680] AVX512: fix psqrt and prsqrt

---
 Eigen/src/Core/arch/AVX512/MathFunctions.h | 22 ++++++++--------------
 1 file changed, 8 insertions(+), 14 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 399be0ee4..4695fbc81 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -266,8 +266,7 @@ psqrt<Packet16f>(const Packet16f& _x) {
   // select only the inverse sqrt of positive normal inputs (denormals are
   // flushed to zero and cause infs as well).
   __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ);
-  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x),
-                                     _mm512_setzero_ps());
+  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_setzero_ps(), _mm512_rsqrt14_ps(_x));
 
   // Do a single step of Newton's iteration.
   x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
@@ -289,8 +288,7 @@ psqrt<Packet8d>(const Packet8d& _x) {
   // select only the inverse sqrt of positive normal inputs (denormals are
   // flushed to zero and cause infs as well).
   __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ);
-  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x),
-                                    _mm512_setzero_pd());
+  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_setzero_pd(), _mm512_rsqrt14_pd(_x));
 
   // Do a first step of Newton's iteration.
   x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
@@ -333,20 +331,18 @@ prsqrt<Packet16f>(const Packet16f& _x) {
   // select only the inverse sqrt of positive normal inputs (denormals are
   // flushed to zero and cause infs as well).
   __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ);
-  Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(),
-                                     _mm512_rsqrt14_ps(_x));
+  Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_rsqrt14_ps(_x), _mm512_setzero_ps());
 
   // Fill in NaNs and Infs for the negative/zero entries.
   __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ);
   Packet16f infs_and_nans = _mm512_mask_blend_ps(
-      neg_mask, p16f_nan,
-      _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps()));
+      neg_mask, _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(), p16f_inf), p16f_nan);
 
   // Do a single step of Newton's iteration.
   x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
 
   // Insert NaNs and Infs in all the right places.
-  return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x);
+  return _mm512_mask_blend_ps(le_zero_mask, x, infs_and_nans);
 }
 
 template <>
@@ -363,14 +359,12 @@ prsqrt<Packet8d>(const Packet8d& _x) {
   // select only the inverse sqrt of positive normal inputs (denormals are
   // flushed to zero and cause infs as well).
   __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ);
-  Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(),
-                                    _mm512_rsqrt14_pd(_x));
+  Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_rsqrt14_pd(_x), _mm512_setzero_pd());
 
   // Fill in NaNs and Infs for the negative/zero entries.
   __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ);
   Packet8d infs_and_nans = _mm512_mask_blend_pd(
-      neg_mask, p8d_nan,
-      _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd()));
+      neg_mask, _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(), p8d_inf), p8d_nan);
 
   // Do a first step of Newton's iteration.
   x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
@@ -379,7 +373,7 @@ prsqrt<Packet8d>(const Packet8d& _x) {
   x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
 
   // Insert NaNs and Infs in all the right places.
-  return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x);
+  return _mm512_mask_blend_pd(le_zero_mask, x, infs_and_nans);
 }
 #else
 template <>

From d43b2f01f48d67507b921b8f9ee1fca071ac2ee2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:14:00 +0200
Subject: [PATCH 270/680] Fix unit testing of predux_downto4 (bad name), and
 add unit testing of prsqrt

---
 test/packetmath.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 3c11df7e8..31a524af3 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -248,12 +248,13 @@ template<typename Scalar> void packetmath()
   VERIFY(isApproxAbs(ref[0], internal::predux(internal::pload<Packet>(data1)), refvalue) && "internal::predux");
 
   {
-    for (int i=0; i<4; ++i)
+    int HalfPacketSize = PacketSize>4 ? PacketSize/2 : PacketSize;
+    for (int i=0; i<HalfPacketSize; ++i)
       ref[i] = 0;
     for (int i=0; i<PacketSize; ++i)
-      ref[i%4] += data1[i];
+      ref[i%HalfPacketSize] += data1[i];
     internal::pstore(data2, internal::predux_downto4(internal::pload<Packet>(data1)));
-    VERIFY(areApprox(ref, data2, PacketSize>4?PacketSize/2:PacketSize) && "internal::predux_downto4");
+    VERIFY(areApprox(ref, data2, HalfPacketSize) && "internal::predux_downto4");
   }
 
   ref[0] = 1;
@@ -436,6 +437,7 @@ template<typename Scalar> void packetmath_real()
   if(internal::random<float>(0,1)<0.1f)
     data1[internal::random<int>(0, PacketSize)] = 0;
   CHECK_CWISE1_IF(PacketTraits::HasSqrt, std::sqrt, internal::psqrt);
+  CHECK_CWISE1_IF(PacketTraits::HasSqrt, Scalar(1)/std::sqrt, internal::prsqrt);
   CHECK_CWISE1_IF(PacketTraits::HasLog, std::log, internal::plog);
 #if EIGEN_HAS_C99_MATH && (__cplusplus > 199711L)
   CHECK_CWISE1_IF(PacketTraits::HasExpm1, std::expm1, internal::pexpm1);

From 67bac6368cc6cdf6c1f420c53cff73878b3d86db Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:19:04 +0200
Subject: [PATCH 271/680] protect calls to isnan

---
 unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 7deca3e12..97b2e5a91 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -535,7 +535,7 @@ struct igammac_impl {
       return nan;
     }
 
-    if (numext::isnan(a) || numext::isnan(x)) { // propagate nans
+    if ((numext::isnan)(a) || (numext::isnan)(x)) { // propagate nans
       return nan;
     }
 
@@ -728,7 +728,7 @@ struct igamma_impl {
       return nan;
     }
 
-    if (numext::isnan(a) || numext::isnan(x)) { // propagate nans
+    if ((numext::isnan)(a) || (numext::isnan)(x)) { // propagate nans
       return nan;
     }
 

From 584951ca4d3f822470895370c2de087baa5ca751 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:28:38 +0200
Subject: [PATCH 272/680] Rename predux_downto4 to be more accurate on its
 semantic.

---
 Eigen/src/Core/GenericPacketMath.h                | 4 ++--
 Eigen/src/Core/arch/AVX/PacketMath.h              | 2 +-
 Eigen/src/Core/arch/AVX512/PacketMath.h           | 4 ++--
 Eigen/src/Core/products/GeneralBlockPanelKernel.h | 6 +++---
 test/packetmath.cpp                               | 4 ++--
 5 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 30878eda6..7979c3aff 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -324,13 +324,13 @@ preduxp(const Packet* vecs) { return vecs[0]; }
 template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a)
 { return a; }
 
-/** \internal \returns the sum of the elements of \a a by block of 4 elements.
+/** \internal \returns the sum of the elements of upper and lower half of \a a if \a a is larger than 4.
   * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7}
   * For packet-size smaller or equal to 4, this boils down to a noop.
   */
 template<typename Packet> EIGEN_DEVICE_FUNC inline
 typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
-predux_downto4(const Packet& a)
+predux_half_dowto4(const Packet& a)
 { return a; }
 
 /** \internal \returns the product of the elements of \a a*/
diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 636230944..af31023f0 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -412,7 +412,7 @@ template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a)
   return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1))));
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a)
+template<> EIGEN_STRONG_INLINE Packet4f predux_half_dowto4<Packet8f>(const Packet8f& a)
 {
   return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
 }
diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index abece01bc..eb5de43d4 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -888,7 +888,7 @@ EIGEN_STRONG_INLINE double predux<Packet8d>(const Packet8d& a) {
 }
 
 template <>
-EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) {
+EIGEN_STRONG_INLINE Packet8f predux_half_dowto4<Packet16f>(const Packet16f& a) {
 #ifdef EIGEN_VECTORIZE_AVX512DQ
   __m256 lane0 = _mm512_extractf32x8_ps(a, 0);
   __m256 lane1 = _mm512_extractf32x8_ps(a, 1);
@@ -904,7 +904,7 @@ EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) {
 #endif
 }
 template <>
-EIGEN_STRONG_INLINE Packet4d predux_downto4<Packet8d>(const Packet8d& a) {
+EIGEN_STRONG_INLINE Packet4d predux_half_dowto4<Packet8d>(const Packet8d& a) {
   __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
   __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
   __m256d res = _mm256_add_pd(lane0, lane1);
diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 9072d0ff3..dd4f366ee 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -580,7 +580,7 @@ DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Pack
 }
 
 template<typename Packet>
-const DoublePacket<Packet>& predux_downto4(const DoublePacket<Packet> &a)
+const DoublePacket<Packet>& predux_half_dowto4(const DoublePacket<Packet> &a)
 {
   return a;
 }
@@ -1596,13 +1596,13 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
                 SRhsPacketHalf b0;
                 straits.loadLhsUnaligned(blB, a0);
                 straits.loadRhs(blA, b0);
-                SAccPacketHalf c0 = predux_downto4(C0);
+                SAccPacketHalf c0 = predux_half_dowto4(C0);
                 straits.madd(a0,b0,c0,b0);
                 straits.acc(c0, alphav, R);
               }
               else
               {
-                straits.acc(predux_downto4(C0), alphav, R);
+                straits.acc(predux_half_dowto4(C0), alphav, R);
               }
               res.scatterPacket(i, j2, R);
             }
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 31a524af3..53838920e 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -253,8 +253,8 @@ template<typename Scalar> void packetmath()
       ref[i] = 0;
     for (int i=0; i<PacketSize; ++i)
       ref[i%HalfPacketSize] += data1[i];
-    internal::pstore(data2, internal::predux_downto4(internal::pload<Packet>(data1)));
-    VERIFY(areApprox(ref, data2, HalfPacketSize) && "internal::predux_downto4");
+    internal::pstore(data2, internal::predux_half_dowto4(internal::pload<Packet>(data1)));
+    VERIFY(areApprox(ref, data2, HalfPacketSize) && "internal::predux_half_dowto4");
   }
 
   ref[0] = 1;

From 40b4bf3d32265152ae26c2f630f6f70e5d676f3d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 14:36:27 +0200
Subject: [PATCH 273/680] AVX512: _mm512_rsqrt28_ps is available for AVX512ER
 only

---
 Eigen/Core                                 | 3 +++
 Eigen/src/Core/arch/AVX512/MathFunctions.h | 2 +-
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/Eigen/Core b/Eigen/Core
index a9bbfe276..c74340b81 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -183,6 +183,9 @@
       #ifdef __AVX512DQ__
         #define EIGEN_VECTORIZE_AVX512DQ
       #endif
+      #ifdef __AVX512ER__
+        #define EIGEN_VECTORIZE_AVX512ER
+      #endif
     #endif
 
     // include files
diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 4695fbc81..61434a33e 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -375,7 +375,7 @@ prsqrt<Packet8d>(const Packet8d& _x) {
   // Insert NaNs and Infs in all the right places.
   return _mm512_mask_blend_pd(le_zero_mask, x, infs_and_nans);
 }
-#else
+#elif defined(EIGEN_VECTORIZE_AVX512ER)
 template <>
 EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
   return _mm512_rsqrt28_ps(x);

From 407e3e2621077a6cd768042c88c652ad75a085ae Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 15:59:30 +0200
Subject: [PATCH 274/680] bug #1532: disable stl::*_negate in C++17 (they are
 deprecated)

---
 Eigen/src/Core/functors/StlFunctors.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/Eigen/src/Core/functors/StlFunctors.h b/Eigen/src/Core/functors/StlFunctors.h
index 6df3fa501..9c1d75850 100644
--- a/Eigen/src/Core/functors/StlFunctors.h
+++ b/Eigen/src/Core/functors/StlFunctors.h
@@ -83,13 +83,17 @@ struct functor_traits<std::binder1st<T> >
 { enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
 #endif
 
+#if (__cplusplus < 201703L) && (EIGEN_COMP_MSVC < 1910)
+// std::unary_negate is deprecated since c++17 and will be removed in c++20
 template<typename T>
 struct functor_traits<std::unary_negate<T> >
 { enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
 
+// std::binary_negate is deprecated since c++17 and will be removed in c++20
 template<typename T>
 struct functor_traits<std::binary_negate<T> >
 { enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
+#endif
 
 #ifdef EIGEN_STDEXT_SUPPORT
 

From 8d0ffe36552aeeb5f46d9c652edc45c68e536cdd Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 16:15:43 +0200
Subject: [PATCH 275/680] bug #1516: add assertion for out-of-range diagonal
 index in MatrixBase::diagonal(i)

---
 Eigen/src/Core/Diagonal.h | 5 ++++-
 test/diagonal.cpp         | 9 +++++++--
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h
index c62f5ff21..563135fb2 100644
--- a/Eigen/src/Core/Diagonal.h
+++ b/Eigen/src/Core/Diagonal.h
@@ -70,7 +70,10 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal)
 
     EIGEN_DEVICE_FUNC
-    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
+    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index)
+    {
+      eigen_assert( a_index <= m_matrix.cols() && -a_index <= m_matrix.rows() );
+    }
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal)
 
diff --git a/test/diagonal.cpp b/test/diagonal.cpp
index c1546e97d..0b5ae8280 100644
--- a/test/diagonal.cpp
+++ b/test/diagonal.cpp
@@ -66,6 +66,9 @@ template<typename MatrixType> void diagonal(const MatrixType& m)
     m2.diagonal(N2).coeffRef(0) = Scalar(2)*s1;
     VERIFY_IS_APPROX(m2.diagonal(N2).coeff(0), Scalar(2)*s1);
   }
+
+  VERIFY( m1.diagonal( cols).size()==0 );
+  VERIFY( m1.diagonal(-rows).size()==0 );
 }
 
 template<typename MatrixType> void diagonal_assert(const MatrixType& m) {
@@ -81,6 +84,9 @@ template<typename MatrixType> void diagonal_assert(const MatrixType& m) {
     VERIFY_RAISES_ASSERT( m1.array() *= m1.diagonal().array() );
     VERIFY_RAISES_ASSERT( m1.array() /= m1.diagonal().array() );
   }
+
+  VERIFY_RAISES_ASSERT( m1.diagonal(cols+1) );
+  VERIFY_RAISES_ASSERT( m1.diagonal(-(rows+1)) );
 }
 
 void test_diagonal()
@@ -95,7 +101,6 @@ void test_diagonal()
     CALL_SUBTEST_2( diagonal(MatrixXcd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_1( diagonal(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_1( diagonal(Matrix<float,Dynamic,4>(3, 4)) );
+    CALL_SUBTEST_1( diagonal_assert(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
   }
-
-  CALL_SUBTEST_1( diagonal_assert(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
 }

From c5b56f1fb27fb5b85eefef6b93dd71f4edb400db Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 16:49:35 +0200
Subject: [PATCH 276/680] bug #1528: better use numeric_limits::min() instead
 of 1/highest() that with underflow.

---
 Eigen/src/Cholesky/LDLT.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h
index 5be58377b..2dfeac333 100644
--- a/Eigen/src/Cholesky/LDLT.h
+++ b/Eigen/src/Cholesky/LDLT.h
@@ -569,13 +569,14 @@ void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) cons
   // more precisely, use pseudo-inverse of D (see bug 241)
   using std::abs;
   const typename Diagonal<const MatrixType>::RealReturnType vecD(vectorD());
-  // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
-  // as motivated by LAPACK's xGELSS:
+  // In some previous versions, tolerance was set to the max of 1/highest (or rather numeric_limits::min())
+  // and the maximal diagonal entry * epsilon as motivated by LAPACK's xGELSS:
   // RealScalar tolerance = numext::maxi(vecD.array().abs().maxCoeff() * NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
   // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
   // diagonal element is not well justified and leads to numerical issues in some cases.
   // Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
-  RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
+  // Using numeric_limits::min() gives us more robustness to denormals.
+  RealScalar tolerance = (std::numeric_limits<RealScalar>::min)();
 
   for (Index i = 0; i < vecD.size(); ++i)
   {

From dd4cc6bd9e60d1ab2cef4c66ee922d3d84f3d0b2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 17:11:15 +0200
Subject: [PATCH 277/680] bug #1527: fix support for MKL's VML (destination was
 not properly resized)

---
 Eigen/src/Core/Assign_MKL.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/Assign_MKL.h b/Eigen/src/Core/Assign_MKL.h
index 6c2ab9264..6866095bf 100755
--- a/Eigen/src/Core/Assign_MKL.h
+++ b/Eigen/src/Core/Assign_MKL.h
@@ -84,7 +84,8 @@ class vml_assign_traits
   struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE,EIGENTYPE>,   \
                    Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {              \
     typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                            \
-    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                   \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &func) {                       \
+      resize_if_allowed(dst, src, func);                                                                                        \
       eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                       \
       if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) {                                              \
         VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(),                                                        \
@@ -144,7 +145,8 @@ EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil,  Ceil,   _)
                    Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {            \
     typedef CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested,                                           \
                     const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> > SrcXprType;                         \
-    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                 \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &func) {                     \
+      resize_if_allowed(dst, src, func);                                                                                      \
       eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                     \
       VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.rhs().functor().m_other);                                       \
       if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal)                                              \

From 8c7b5158a10af19bb98f2c2b5c57d44fd8899fe8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 17:15:38 +0200
Subject: [PATCH 278/680] commit 45e9c9996da790b55ed9c4b0dfeae49492ac5c46 (HEAD
 -> memory_fix) Author: George Burgess IV <gbiv@google.com> Date:   Thu Mar 1
 11:20:24 2018 -0800

    Prefer `::operator new` to `new`

    The C++ standard allows compilers much flexibility with `new`
    expressions, including eliding them entirely
    (https://godbolt.org/g/yS6i91). However, calls to `operator new` are
    required to be treated like opaque function calls.

    Since we're calling `new` for side-effects other than allocating heap
    memory, we should prefer the less flexible version.

    Signed-off-by: George Burgess IV <gbiv@google.com>
---
 Eigen/src/Core/util/Memory.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 006b0bfba..53300c388 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -70,7 +70,7 @@ inline void throw_std_bad_alloc()
     throw std::bad_alloc();
   #else
     std::size_t huge = static_cast<std::size_t>(-1);
-    new int[huge];
+    ::operator new(huge);
   #endif
 }
 

From a1292395d6feefa5d78ed9101c2c85648ba8eea9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 23:06:44 +0200
Subject: [PATCH 279/680] Fix compilation of product with inverse
 transpositions (e.g., mat * Transpositions().inverse())

---
 Eigen/src/Core/Transpositions.h |  2 +-
 test/permutationmatrices.cpp    | 12 ++++++++++++
 2 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h
index 8798deca5..81a4a5855 100644
--- a/Eigen/src/Core/Transpositions.h
+++ b/Eigen/src/Core/Transpositions.h
@@ -384,7 +384,7 @@ class Transpose<TranspositionsBase<TranspositionsDerived> >
     const Product<OtherDerived, Transpose, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt)
     {
-      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt);
     }
 
     /** \returns the \a matrix with the inverse transpositions applied to the rows.
diff --git a/test/permutationmatrices.cpp b/test/permutationmatrices.cpp
index db1266579..0ed7a4b2b 100644
--- a/test/permutationmatrices.cpp
+++ b/test/permutationmatrices.cpp
@@ -19,9 +19,11 @@ template<typename MatrixType> void permutationmatrices(const MatrixType& m)
   enum { Rows = MatrixType::RowsAtCompileTime, Cols = MatrixType::ColsAtCompileTime,
          Options = MatrixType::Options };
   typedef PermutationMatrix<Rows> LeftPermutationType;
+  typedef Transpositions<Rows> LeftTranspositionsType;
   typedef Matrix<int, Rows, 1> LeftPermutationVectorType;
   typedef Map<LeftPermutationType> MapLeftPerm;
   typedef PermutationMatrix<Cols> RightPermutationType;
+  typedef Transpositions<Cols> RightTranspositionsType;
   typedef Matrix<int, Cols, 1> RightPermutationVectorType;
   typedef Map<RightPermutationType> MapRightPerm;
 
@@ -35,6 +37,8 @@ template<typename MatrixType> void permutationmatrices(const MatrixType& m)
   RightPermutationVectorType rv;
   randomPermutationVector(rv, cols);
   RightPermutationType rp(rv);
+  LeftTranspositionsType lt(lv);
+  RightTranspositionsType rt(rv);
   MatrixType m_permuted = MatrixType::Random(rows,cols);
   
   VERIFY_EVALUATION_COUNT(m_permuted = lp * m_original * rp, 1); // 1 temp for sub expression "lp * m_original"
@@ -115,6 +119,14 @@ template<typename MatrixType> void permutationmatrices(const MatrixType& m)
     Matrix<Scalar, Cols, Cols> B = rp.transpose();
     VERIFY_IS_APPROX(A, B.transpose());
   }
+
+  m_permuted = m_original;
+  lp = lt;
+  rp = rt;
+  VERIFY_EVALUATION_COUNT(m_permuted = lt * m_permuted * rt, 1);
+  VERIFY_IS_APPROX(m_permuted, lp*m_original*rp);
+  
+  VERIFY_IS_APPROX(lt.inverse()*m_permuted*rt.inverse(), m_original);
 }
 
 template<typename T>

From 112c8993049ea0c4ef8e3e4f257880626f8b39e1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Apr 2018 23:16:43 +0200
Subject: [PATCH 280/680] comment unreachable code

---
 Eigen/src/Core/arch/AVX/PacketMath.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index af31023f0..34b89c790 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -343,9 +343,12 @@ template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a)
 {
    __m256d tmp = _mm256_shuffle_pd(a,a,5);
   return _mm256_permute2f128_pd(tmp, tmp, 1);
-
+  #if 0
+  // This version is unlikely to be faster as _mm256_shuffle_ps and _mm256_permute_pd
+  // exhibit the same latency/throughput, but it is here for future reference/benchmarking...
   __m256d swap_halves = _mm256_permute2f128_pd(a,a,1);
     return _mm256_permute_pd(swap_halves,5);
+  #endif
 }
 
 // pabs should be ok

From e91e314347c14774206307a91d1b427e49f9b3e2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 11:39:19 +0200
Subject: [PATCH 281/680] bug #1494: makes pmin/pmax behave on Altivec/VSX as
 on x86 regading NaNs

---
 Eigen/src/Core/arch/AltiVec/PacketMath.h | 36 +++++++++++++++++++++---
 1 file changed, 32 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index 31bb896ca..ad0c6df45 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -388,10 +388,28 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, co
 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); }
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; }
 
-template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  #ifdef __VSX__
+  Packet4f ret;
+  __asm__ ("xvcmpgesp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
+  return ret;
+  #else
+  return vec_min(a, b);
+  #endif
+}
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
 
-template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  #ifdef __VSX__
+  Packet4f ret;
+  __asm__ ("xvcmpgtsp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
+  return ret;
+  #else
+  return vec_max(a, b);
+  #endif
+}
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
@@ -910,9 +928,19 @@ template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const
 // for some weird raisons, it has to be overloaded for packet of integers
 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
 
-template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  Packet2d ret;
+  __asm__ ("xvcmpgedp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
+  return ret;
+ }
 
-template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  Packet2d ret;
+  __asm__ ("xvcmpgtdp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
+  return ret;
+}
 
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
 

From 13f5df9f6763f2dc900b74df06a3b1e67bacdaaf Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 13:10:38 +0200
Subject: [PATCH 282/680] Add a note on vec_min vs asm

---
 Eigen/src/Core/arch/AltiVec/PacketMath.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index ad0c6df45..6d7190a56 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -391,6 +391,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i&
 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
   #ifdef __VSX__
+  // NOTE: about 10% slower than vec_min, but consistent with std::min and SSE regarding NaN
   Packet4f ret;
   __asm__ ("xvcmpgesp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
   return ret;
@@ -403,6 +404,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
   #ifdef __VSX__
+  // NOTE: about 10% slower than vec_max, but consistent with std::max and SSE regarding NaN
   Packet4f ret;
   __asm__ ("xvcmpgtsp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
   return ret;
@@ -930,6 +932,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d&
 
 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b)
 {
+  // NOTE: about 10% slower than vec_min, but consistent with std::min and SSE regarding NaN
   Packet2d ret;
   __asm__ ("xvcmpgedp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
   return ret;
@@ -937,6 +940,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const
 
 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b)
 {
+  // NOTE: about 10% slower than vec_max, but consistent with std::max and SSE regarding NaN
   Packet2d ret;
   __asm__ ("xvcmpgtdp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b));
   return ret;

From 73729025a4740d344bae98dfbae36a69405e217f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 13:45:34 +0200
Subject: [PATCH 283/680] bug #1521: add unit test dedicated to numbest::hypos

---
 test/stable_norm.cpp | 39 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp
index 3c02474b8..e4b97a6da 100644
--- a/test/stable_norm.cpp
+++ b/test/stable_norm.cpp
@@ -190,13 +190,52 @@ template<typename MatrixType> void stable_norm(const MatrixType& m)
   }
 }
 
+template<typename Scalar>
+void test_hypot()
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  Scalar factor = internal::random<Scalar>();
+  while(numext::abs2(factor)<RealScalar(1e-4))
+    factor = internal::random<Scalar>();
+  Scalar big = factor * ((std::numeric_limits<RealScalar>::max)() * RealScalar(1e-4));
+  
+  factor = internal::random<Scalar>();
+  while(numext::abs2(factor)<RealScalar(1e-4))
+    factor = internal::random<Scalar>();
+  Scalar small = factor * ((std::numeric_limits<RealScalar>::min)() * RealScalar(1e4));
+
+  std::cout << big << " " << small << "\n";
+
+  Scalar  one   (1),
+          zero  (0),
+          sqrt2 (std::sqrt(2)),
+          nan   (std::numeric_limits<RealScalar>::quiet_NaN());
+
+  Scalar a = internal::random<Scalar>(-1,1);
+  Scalar b = internal::random<Scalar>(-1,1);
+  VERIFY_IS_APPROX(numext::hypot(a,b),std::sqrt(numext::abs2(a)+numext::abs2(b)));
+  VERIFY_IS_EQUAL(numext::hypot(zero,zero), zero);
+  VERIFY_IS_APPROX(numext::hypot(one, one), sqrt2);
+  VERIFY_IS_APPROX(numext::hypot(big,big), sqrt2*numext::abs(big));
+  VERIFY_IS_APPROX(numext::hypot(small,small), sqrt2*numext::abs(small));
+  VERIFY_IS_APPROX(numext::hypot(small,big), numext::abs(big));
+  VERIFY((numext::isnan)(numext::hypot(nan,a)));
+  VERIFY((numext::isnan)(numext::hypot(a,nan)));
+}
+
 void test_stable_norm()
 {
   for(int i = 0; i < g_repeat; i++) {
+    CALL_SUBTEST_3( test_hypot<double>() );
+    CALL_SUBTEST_4( test_hypot<float>() );
+    CALL_SUBTEST_5( test_hypot<std::complex<double> >() );
+    CALL_SUBTEST_6( test_hypot<std::complex<float> >() );
+
     CALL_SUBTEST_1( stable_norm(Matrix<float, 1, 1>()) );
     CALL_SUBTEST_2( stable_norm(Vector4d()) );
     CALL_SUBTEST_3( stable_norm(VectorXd(internal::random<int>(10,2000))) );
     CALL_SUBTEST_4( stable_norm(VectorXf(internal::random<int>(10,2000))) );
     CALL_SUBTEST_5( stable_norm(VectorXcd(internal::random<int>(10,2000))) );
+    CALL_SUBTEST_6( stable_norm(VectorXcf(internal::random<int>(10,2000))) );
   }
 }

From e116f6847e3cd9d16ea3be14c2b7efbd0a1c2b0b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 13:47:23 +0200
Subject: [PATCH 284/680] bug #1521: avoid signalling NaN in hypot and make it
 std::complex<> friendly.

---
 Eigen/src/Core/MathFunctions.h     | 26 +-------------------------
 Eigen/src/Core/MathFunctionsImpl.h | 19 +++++++++++++++++++
 2 files changed, 20 insertions(+), 25 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index e981129b2..84f9d0cd5 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -347,31 +347,7 @@ struct norm1_retval
 * Implementation of hypot                                                *
 ****************************************************************************/
 
-template<typename Scalar>
-struct hypot_impl
-{
-  typedef typename NumTraits<Scalar>::Real RealScalar;
-  static inline RealScalar run(const Scalar& x, const Scalar& y)
-  {
-    EIGEN_USING_STD_MATH(abs);
-    EIGEN_USING_STD_MATH(sqrt);
-    RealScalar _x = abs(x);
-    RealScalar _y = abs(y);
-    Scalar p, qp;
-    if(_x>_y)
-    {
-      p = _x;
-      qp = _y / p;
-    }
-    else
-    {
-      p = _y;
-      qp = _x / p;
-    }
-    if(p==RealScalar(0)) return RealScalar(0);
-    return p * sqrt(RealScalar(1) + qp*qp);
-  }
-};
+template<typename Scalar> struct hypot_impl;
 
 template<typename Scalar>
 struct hypot_retval
diff --git a/Eigen/src/Core/MathFunctionsImpl.h b/Eigen/src/Core/MathFunctionsImpl.h
index ae1386b4c..034cfad7b 100644
--- a/Eigen/src/Core/MathFunctionsImpl.h
+++ b/Eigen/src/Core/MathFunctionsImpl.h
@@ -66,6 +66,25 @@ T generic_fast_tanh_float(const T& a_x)
   return pdiv(p, q);
 }
 
+
+template<typename Scalar>
+struct hypot_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  static inline RealScalar run(const Scalar& x, const Scalar& y)
+  {
+    EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
+    RealScalar _x = abs(x);
+    RealScalar _y = abs(y);
+    RealScalar p, qp;
+    p = numext::maxi(_x,_y);
+    if(p==RealScalar(0)) return RealScalar(0);
+    qp = numext::mini(_y,_x) / p;    
+    return p * sqrt(RealScalar(1) + qp*qp);
+  }
+};
+
 } // end namespace internal
 
 } // end namespace Eigen

From 368dd4cd9d18cf904b2ce3c4de18020269dcbba1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 15:09:21 +0200
Subject: [PATCH 285/680] Make innerVector() and innerVectors() methods
 available to all expressions supported by Block.

Before, only SparseBase exposed such methods.
---
 Eigen/src/SparseCore/SparseBlock.h      | 40 ----------------------
 Eigen/src/SparseCore/SparseMatrixBase.h | 12 -------
 Eigen/src/plugins/BlockMethods.h        | 44 +++++++++++++++++++++++++
 3 files changed, 44 insertions(+), 52 deletions(-)

diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h
index 511e92b2f..5ed7c437b 100644
--- a/Eigen/src/SparseCore/SparseBlock.h
+++ b/Eigen/src/SparseCore/SparseBlock.h
@@ -326,46 +326,6 @@ private:
 
 //----------
 
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major).
-  */
-template<typename Derived>
-typename SparseMatrixBase<Derived>::InnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer)
-{ return InnerVectorReturnType(derived(), outer); }
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major). Read-only.
-  */
-template<typename Derived>
-const typename SparseMatrixBase<Derived>::ConstInnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer) const
-{ return ConstInnerVectorReturnType(derived(), outer); }
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major).
-  */
-template<typename Derived>
-typename SparseMatrixBase<Derived>::InnerVectorsReturnType
-SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize)
-{
-  return Block<Derived,Dynamic,Dynamic,true>(derived(),
-                                             IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
-                                             IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
-
-}
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major). Read-only.
-  */
-template<typename Derived>
-const typename SparseMatrixBase<Derived>::ConstInnerVectorsReturnType
-SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize) const
-{
-  return Block<const Derived,Dynamic,Dynamic,true>(derived(),
-                                                  IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
-                                                  IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
-
-}
-
 /** Generic implementation of sparse Block expression.
   * Real-only.
   */
diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h
index c6b548f11..dac2d8c62 100644
--- a/Eigen/src/SparseCore/SparseMatrixBase.h
+++ b/Eigen/src/SparseCore/SparseMatrixBase.h
@@ -350,18 +350,6 @@ template<typename Derived> class SparseMatrixBase
     const ConstTransposeReturnType transpose() const { return ConstTransposeReturnType(derived()); }
     const AdjointReturnType adjoint() const { return AdjointReturnType(transpose()); }
 
-    // inner-vector
-    typedef Block<Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true>       InnerVectorReturnType;
-    typedef Block<const Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true> ConstInnerVectorReturnType;
-    InnerVectorReturnType innerVector(Index outer);
-    const ConstInnerVectorReturnType innerVector(Index outer) const;
-
-    // set of inner-vectors
-    typedef Block<Derived,Dynamic,Dynamic,true> InnerVectorsReturnType;
-    typedef Block<const Derived,Dynamic,Dynamic,true> ConstInnerVectorsReturnType;
-    InnerVectorsReturnType innerVectors(Index outerStart, Index outerSize);
-    const ConstInnerVectorsReturnType innerVectors(Index outerStart, Index outerSize) const;
-
     DenseMatrixType toDense() const
     {
       return DenseMatrixType(derived());
diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h
index 5caf14469..5fcd13cc4 100644
--- a/Eigen/src/plugins/BlockMethods.h
+++ b/Eigen/src/plugins/BlockMethods.h
@@ -40,6 +40,14 @@ typedef const VectorBlock<const Derived> ConstSegmentReturnType;
 template<int Size> struct FixedSegmentReturnType { typedef VectorBlock<Derived, Size> Type; };
 template<int Size> struct ConstFixedSegmentReturnType { typedef const VectorBlock<const Derived, Size> Type; };
 
+/// \internal inner-vector
+typedef Block<Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true>       InnerVectorReturnType;
+typedef Block<const Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true> ConstInnerVectorReturnType;
+
+/// \internal set of inner-vectors
+typedef Block<Derived,Dynamic,Dynamic,true> InnerVectorsReturnType;
+typedef Block<const Derived,Dynamic,Dynamic,true> ConstInnerVectorsReturnType;
+
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 /// \returns an expression of a block in \c *this with either dynamic or fixed sizes.
@@ -1354,3 +1362,39 @@ inline typename ConstFixedSegmentReturnType<N>::Type tail(Index n = N) const
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return typename ConstFixedSegmentReturnType<N>::Type(derived(), size() - n);
 }
+
+/// \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+/// is col-major (resp. row-major).
+///
+InnerVectorReturnType innerVector(Index outer)
+{ return InnerVectorReturnType(derived(), outer); }
+
+/// \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+/// is col-major (resp. row-major). Read-only.
+///
+const ConstInnerVectorReturnType innerVector(Index outer) const
+{ return ConstInnerVectorReturnType(derived(), outer); }
+
+/// \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+/// is col-major (resp. row-major).
+///
+InnerVectorsReturnType
+innerVectors(Index outerStart, Index outerSize)
+{
+  return Block<Derived,Dynamic,Dynamic,true>(derived(),
+                                             IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
+                                             IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
+
+}
+
+/// \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+/// is col-major (resp. row-major). Read-only.
+///
+const ConstInnerVectorsReturnType
+innerVectors(Index outerStart, Index outerSize) const
+{
+  return Block<const Derived,Dynamic,Dynamic,true>(derived(),
+                                                  IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
+                                                  IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
+
+}
\ No newline at end of file

From 4213b63f5ce33d3f904674ee7b0cabd6934dda6b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 15:12:43 +0200
Subject: [PATCH 286/680] Factories code between numext::hypot and
 scalar_hyot_op functor.

---
 Eigen/src/Core/MathFunctionsImpl.h       | 20 +++++++++++--------
 Eigen/src/Core/functors/BinaryFunctors.h | 25 +++++++++---------------
 2 files changed, 21 insertions(+), 24 deletions(-)

diff --git a/Eigen/src/Core/MathFunctionsImpl.h b/Eigen/src/Core/MathFunctionsImpl.h
index 034cfad7b..cc2ac0700 100644
--- a/Eigen/src/Core/MathFunctionsImpl.h
+++ b/Eigen/src/Core/MathFunctionsImpl.h
@@ -66,6 +66,17 @@ T generic_fast_tanh_float(const T& a_x)
   return pdiv(p, q);
 }
 
+template<typename RealScalar>
+EIGEN_STRONG_INLINE
+RealScalar positive_real_hypot(const RealScalar& x, const RealScalar& y)
+{
+  EIGEN_USING_STD_MATH(sqrt);
+  RealScalar p, qp;
+  p = numext::maxi(x,y);
+  if(p==RealScalar(0)) return RealScalar(0);
+  qp = numext::mini(y,x) / p;    
+  return p * sqrt(RealScalar(1) + qp*qp);
+}
 
 template<typename Scalar>
 struct hypot_impl
@@ -74,14 +85,7 @@ struct hypot_impl
   static inline RealScalar run(const Scalar& x, const Scalar& y)
   {
     EIGEN_USING_STD_MATH(abs);
-    EIGEN_USING_STD_MATH(sqrt);
-    RealScalar _x = abs(x);
-    RealScalar _y = abs(y);
-    RealScalar p, qp;
-    p = numext::maxi(_x,_y);
-    if(p==RealScalar(0)) return RealScalar(0);
-    qp = numext::mini(_y,_x) / p;    
-    return p * sqrt(RealScalar(1) + qp*qp);
+    return positive_real_hypot(abs(x), abs(y));
   }
 };
 
diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h
index 96747bac7..3eae6b8ca 100644
--- a/Eigen/src/Core/functors/BinaryFunctors.h
+++ b/Eigen/src/Core/functors/BinaryFunctors.h
@@ -255,7 +255,7 @@ struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_NEQ> : binary_op_base<LhsScalar,Rh
 
 
 /** \internal
-  * \brief Template functor to compute the hypot of two scalars
+  * \brief Template functor to compute the hypot of two \b positive \b and \b real scalars
   *
   * \sa MatrixBase::stableNorm(), class Redux
   */
@@ -263,22 +263,15 @@ template<typename Scalar>
 struct scalar_hypot_op<Scalar,Scalar> : binary_op_base<Scalar,Scalar>
 {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op)
-//   typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar &x, const Scalar &y) const
   {
-    EIGEN_USING_STD_MATH(sqrt)
-    Scalar p, qp;
-    if(_x>_y)
-    {
-      p = _x;
-      qp = _y / p;
-    }
-    else
-    {
-      p = _y;
-      qp = _x / p;
-    }
-    return p * sqrt(Scalar(1) + qp*qp);
+    // This functor is used by hypotNorm only for which it is faster to first apply abs
+    // on all coefficients prior to reduction through hypot.
+    // This way we avoid calling abs on positive and real entries, and this also permits
+    // to seamlessly handle complexes. Otherwise we would have to handle both real and complexes
+    // through the same functor...
+    return internal::positive_real_hypot(x,y);
   }
 };
 template<typename Scalar>

From 403f09ccef73d347e7d1a6966dfb04eb291cc8b5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 15:13:31 +0200
Subject: [PATCH 287/680] Make stableNorm and blueNorm compatible with 2D
 matrices.

---
 Eigen/src/Core/StableNorm.h | 115 +++++++++++++++++++++++++-----------
 test/stable_norm.cpp        |   3 +-
 2 files changed, 80 insertions(+), 38 deletions(-)

diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h
index e0e7a0c8f..77ea3c261 100644
--- a/Eigen/src/Core/StableNorm.h
+++ b/Eigen/src/Core/StableNorm.h
@@ -50,6 +50,71 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc
     ssq += (bl*invScale).squaredNorm();
 }
 
+template<typename VectorType, typename RealScalar>
+void stable_norm_impl_inner_step(const VectorType &vec, RealScalar& ssq, RealScalar& scale, RealScalar& invScale)
+{
+  typedef typename VectorType::Scalar Scalar;
+  const Index blockSize = 4096;
+  
+  typedef typename internal::nested_eval<VectorType,2>::type VectorTypeCopy;
+  typedef typename internal::remove_all<VectorTypeCopy>::type VectorTypeCopyClean;
+  const VectorTypeCopy copy(vec);
+  
+  enum {
+    CanAlign = (   (int(VectorTypeCopyClean::Flags)&DirectAccessBit)
+                || (int(internal::evaluator<VectorTypeCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
+               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
+                 && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
+  };
+  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<VectorTypeCopyClean>::Alignment>,
+                                                   typename VectorTypeCopyClean::ConstSegmentReturnType>::type SegmentWrapper;
+  Index n = vec.size();
+  
+  Index bi = internal::first_default_aligned(copy);
+  if (bi>0)
+    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
+  for (; bi<n; bi+=blockSize)
+    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
+}
+
+template<typename VectorType>
+typename VectorType::RealScalar
+stable_norm_impl(const VectorType &vec, typename enable_if<VectorType::IsVectorAtCompileTime>::type* = 0 )
+{
+  using std::sqrt;
+  using std::abs;
+
+  Index n = vec.size();
+
+  if(n==1)
+    return abs(vec.coeff(0));
+
+  typedef typename VectorType::RealScalar RealScalar;
+  RealScalar scale(0);
+  RealScalar invScale(1);
+  RealScalar ssq(0); // sum of squares
+
+  stable_norm_impl_inner_step(vec, ssq, scale, invScale);
+  
+  return scale * sqrt(ssq);
+}
+
+template<typename MatrixType>
+typename MatrixType::RealScalar
+stable_norm_impl(const MatrixType &mat, typename enable_if<!MatrixType::IsVectorAtCompileTime>::type* = 0 )
+{
+  using std::sqrt;
+
+  typedef typename MatrixType::RealScalar RealScalar;
+  RealScalar scale(0);
+  RealScalar invScale(1);
+  RealScalar ssq(0); // sum of squares
+
+  for(Index j=0; j<mat.outerSize(); ++j)
+    stable_norm_impl_inner_step(mat.innerVector(j), ssq, scale, invScale);
+  return scale * sqrt(ssq);
+}
+
 template<typename Derived>
 inline typename NumTraits<typename traits<Derived>::Scalar>::Real
 blueNorm_impl(const EigenBase<Derived>& _vec)
@@ -98,12 +163,16 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
   RealScalar asml = RealScalar(0);
   RealScalar amed = RealScalar(0);
   RealScalar abig = RealScalar(0);
-  for(typename Derived::InnerIterator it(vec, 0); it; ++it)
+
+  for(Index j=0; j<vec.outerSize(); ++j)
   {
-    RealScalar ax = abs(it.value());
-    if(ax > ab2)     abig += numext::abs2(ax*s2m);
-    else if(ax < b1) asml += numext::abs2(ax*s1m);
-    else             amed += numext::abs2(ax);
+    for(typename Derived::InnerIterator it(vec, j); it; ++it)
+    {
+      RealScalar ax = abs(it.value());
+      if(ax > ab2)     abig += numext::abs2(ax*s2m);
+      else if(ax < b1) asml += numext::abs2(ax*s1m);
+      else             amed += numext::abs2(ax);
+    }
   }
   if(amed!=amed)
     return amed;  // we got a NaN
@@ -156,36 +225,7 @@ template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 MatrixBase<Derived>::stableNorm() const
 {
-  using std::sqrt;
-  using std::abs;
-  const Index blockSize = 4096;
-  RealScalar scale(0);
-  RealScalar invScale(1);
-  RealScalar ssq(0); // sum of square
-  
-  typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
-  typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
-  const DerivedCopy copy(derived());
-  
-  enum {
-    CanAlign = (   (int(DerivedCopyClean::Flags)&DirectAccessBit)
-                || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
-               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
-                 && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
-  };
-  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
-                                                   typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper;
-  Index n = size();
-  
-  if(n==1)
-    return abs(this->coeff(0));
-  
-  Index bi = internal::first_default_aligned(copy);
-  if (bi>0)
-    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
-  for (; bi<n; bi+=blockSize)
-    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
-  return scale * sqrt(ssq);
+  return internal::stable_norm_impl(derived());
 }
 
 /** \returns the \em l2 norm of \c *this using the Blue's algorithm.
@@ -213,7 +253,10 @@ template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 MatrixBase<Derived>::hypotNorm() const
 {
-  return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
+  if(size()==1)
+    return numext::abs(coeff(0,0));
+  else
+    return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
 }
 
 } // end namespace Eigen
diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp
index e4b97a6da..0dcf072fe 100644
--- a/test/stable_norm.cpp
+++ b/test/stable_norm.cpp
@@ -204,8 +204,6 @@ void test_hypot()
     factor = internal::random<Scalar>();
   Scalar small = factor * ((std::numeric_limits<RealScalar>::min)() * RealScalar(1e4));
 
-  std::cout << big << " " << small << "\n";
-
   Scalar  one   (1),
           zero  (0),
           sqrt2 (std::sqrt(2)),
@@ -234,6 +232,7 @@ void test_stable_norm()
     CALL_SUBTEST_1( stable_norm(Matrix<float, 1, 1>()) );
     CALL_SUBTEST_2( stable_norm(Vector4d()) );
     CALL_SUBTEST_3( stable_norm(VectorXd(internal::random<int>(10,2000))) );
+    CALL_SUBTEST_3( stable_norm(MatrixXd(internal::random<int>(10,200), internal::random<int>(10,200))) );
     CALL_SUBTEST_4( stable_norm(VectorXf(internal::random<int>(10,2000))) );
     CALL_SUBTEST_5( stable_norm(VectorXcd(internal::random<int>(10,2000))) );
     CALL_SUBTEST_6( stable_norm(VectorXcf(internal::random<int>(10,2000))) );

From b903fa74fdaad01dbf87233d03deb3b69624ae5b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 15:14:09 +0200
Subject: [PATCH 288/680] Extend list of MSVC versions

---
 Eigen/src/Core/util/Macros.h | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 5872ade26..9c68ecb7d 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -73,12 +73,17 @@
 
 // For the record, here is a table summarizing the possible values for EIGEN_COMP_MSVC:
 //  name  ver   MSC_VER
-//  2008    9      1500
-//  2010   10      1600
-//  2012   11      1700
-//  2013   12      1800
-//  2015   14      1900
-//  "15"   15      1900
+//  2008         9      1500
+//  2010        10      1600
+//  2012        11      1700
+//  2013        12      1800
+//  2015        14      1900
+//  "15"        15      1900
+//  2017-14.1   15.0    1910
+//  2017-14.11  15.3    1911
+//  2017-14.12  15.5    1912
+//  2017-14.13  15.6    1913
+//  2017-14.14  15.7    1914
 
 /// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC or clang-cl
 #if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC || EIGEN_COMP_LLVM || EIGEN_COMP_CLANG)

From 2f833b1c6438fb379e05f12e3111c30c0b2f3416 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 4 Apr 2018 15:47:46 +0200
Subject: [PATCH 289/680] bug #1509: fix computeInverseWithCheck for complexes

---
 Eigen/src/LU/InverseImpl.h | 2 +-
 test/inverse.cpp           | 4 +++-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/LU/InverseImpl.h b/Eigen/src/LU/InverseImpl.h
index 018f99b58..f49f23360 100644
--- a/Eigen/src/LU/InverseImpl.h
+++ b/Eigen/src/LU/InverseImpl.h
@@ -404,7 +404,7 @@ inline void MatrixBase<Derived>::computeInverseWithCheck(
     const RealScalar& absDeterminantThreshold
   ) const
 {
-  RealScalar determinant;
+  Scalar determinant;
   // i'd love to put some static assertions there, but SFINAE means that they have no effect...
   eigen_assert(rows() == cols());
   computeInverseAndDetWithCheck(inverse,determinant,invertible,absDeterminantThreshold);
diff --git a/test/inverse.cpp b/test/inverse.cpp
index 5c6777a18..97fe6ff92 100644
--- a/test/inverse.cpp
+++ b/test/inverse.cpp
@@ -47,7 +47,7 @@ template<typename MatrixType> void inverse(const MatrixType& m)
   //computeInverseAndDetWithCheck tests
   //First: an invertible matrix
   bool invertible;
-  RealScalar det;
+  Scalar det;
 
   m2.setZero();
   m1.computeInverseAndDetWithCheck(m2, det, invertible);
@@ -113,5 +113,7 @@ void test_inverse()
     
     CALL_SUBTEST_7( inverse(Matrix4d()) );
     CALL_SUBTEST_7( inverse(Matrix<double,4,4,DontAlign>()) );
+
+    CALL_SUBTEST_8( inverse(Matrix4cd()) );
   }
 }

From c2624c03189562208a8e839b8a60291b763c8f44 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 7 Apr 2018 08:45:19 +0200
Subject: [PATCH 290/680] Fix cmake scripts with no fortran compiler

---
 cmake/language_support.cmake | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cmake/language_support.cmake b/cmake/language_support.cmake
index 2f14f30b8..ddba50945 100644
--- a/cmake/language_support.cmake
+++ b/cmake/language_support.cmake
@@ -26,7 +26,7 @@ function(workaround_9220 language language_works)
     cmake_minimum_required(VERSION 2.8.0)
     set (CMAKE_Fortran_FLAGS \"${CMAKE_Fortran_FLAGS}\")
     set (CMAKE_EXE_LINKER_FLAGS \"${CMAKE_EXE_LINKER_FLAGS}\")
-    enable_language(${language} OPTIONAL)
+    enable_language(${language})
   ")
   file(REMOVE_RECURSE ${CMAKE_BINARY_DIR}/language_tests/${language})
   file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/language_tests/${language})

From 04b1628e5546732785419560e94b77c1299f4f28 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Apr 2018 13:28:31 +0200
Subject: [PATCH 291/680] Add missing empty line.

---
 Eigen/src/plugins/BlockMethods.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h
index 5fcd13cc4..0a53db19f 100644
--- a/Eigen/src/plugins/BlockMethods.h
+++ b/Eigen/src/plugins/BlockMethods.h
@@ -1397,4 +1397,4 @@ innerVectors(Index outerStart, Index outerSize) const
                                                   IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
                                                   IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
 
-}
\ No newline at end of file
+}

From add15924ac7a358001a02c94295d17a1623444cf Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Apr 2018 13:29:26 +0200
Subject: [PATCH 292/680] Fix MKL backend for symmetric eigenvalues on
 row-major matrices.

---
 .../SelfAdjointEigenSolver_LAPACKE.h          | 23 ++++++++-----------
 1 file changed, 10 insertions(+), 13 deletions(-)

diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h
index 3891cf883..b0c947dc0 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h
@@ -37,7 +37,7 @@ namespace Eigen {
 
 /** \internal Specialization for the data types supported by LAPACKe */
 
-#define EIGEN_LAPACKE_EIG_SELFADJ(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, EIGCOLROW, LAPACKE_COLROW ) \
+#define EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, EIGCOLROW ) \
 template<> template<typename InputType> inline \
 SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
 SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, int options) \
@@ -47,7 +47,7 @@ SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(c
           && (options&EigVecMask)!=EigVecMask \
           && "invalid option parameter"); \
   bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; \
-  lapack_int n = internal::convert_index<lapack_int>(matrix.cols()), lda, matrix_order, info; \
+  lapack_int n = internal::convert_index<lapack_int>(matrix.cols()), lda, info; \
   m_eivalues.resize(n,1); \
   m_subdiag.resize(n-1); \
   m_eivec = matrix; \
@@ -63,27 +63,24 @@ SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(c
   } \
 \
   lda = internal::convert_index<lapack_int>(m_eivec.outerStride()); \
-  matrix_order=LAPACKE_COLROW; \
   char jobz, uplo='L'/*, range='A'*/; \
   jobz = computeEigenvectors ? 'V' : 'N'; \
 \
-  info = LAPACKE_##LAPACKE_NAME( matrix_order, jobz, uplo, n, (LAPACKE_TYPE*)m_eivec.data(), lda, (LAPACKE_RTYPE*)m_eivalues.data() ); \
+  info = LAPACKE_##LAPACKE_NAME( LAPACK_COL_MAJOR, jobz, uplo, n, (LAPACKE_TYPE*)m_eivec.data(), lda, (LAPACKE_RTYPE*)m_eivalues.data() ); \
   m_info = (info==0) ? Success : NoConvergence; \
   m_isInitialized = true; \
   m_eigenvectorsOk = computeEigenvectors; \
   return *this; \
 }
 
+#define EIGEN_LAPACKE_EIG_SELFADJ(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME )              \
+        EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, ColMajor )  \
+        EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, RowMajor ) 
 
-EIGEN_LAPACKE_EIG_SELFADJ(double,   double,                double, dsyev, ColMajor, LAPACK_COL_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(float,    float,                 float,  ssyev, ColMajor, LAPACK_COL_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, ColMajor, LAPACK_COL_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float,  float,  cheev, ColMajor, LAPACK_COL_MAJOR)
-
-EIGEN_LAPACKE_EIG_SELFADJ(double,   double,                double, dsyev, RowMajor, LAPACK_ROW_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(float,    float,                 float,  ssyev, RowMajor, LAPACK_ROW_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, RowMajor, LAPACK_ROW_MAJOR)
-EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float,  float,  cheev, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(double,   double,                double, dsyev)
+EIGEN_LAPACKE_EIG_SELFADJ(float,    float,                 float,  ssyev)
+EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev)
+EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float,  float,  cheev)
 
 } // end namespace Eigen
 

From 08008f67e1c8b3eb5fa6e846f2760aae38cf82c7 Mon Sep 17 00:00:00 2001
From: vhuber <anonymous@invalid.net>
Date: Mon, 9 Apr 2018 17:07:46 +0200
Subject: [PATCH 293/680] Add unitTest

---
 test/umfpack_support.cpp | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/test/umfpack_support.cpp b/test/umfpack_support.cpp
index 37ab11f0b..3e9b33ba4 100644
--- a/test/umfpack_support.cpp
+++ b/test/umfpack_support.cpp
@@ -12,10 +12,10 @@
 
 #include <Eigen/UmfPackSupport>
 
-template<typename T> void test_umfpack_support_T()
+template<typename T1, typename T2> void test_umfpack_support_T()
 {
-  UmfPackLU<SparseMatrix<T, ColMajor> > umfpack_colmajor;
-  UmfPackLU<SparseMatrix<T, RowMajor> > umfpack_rowmajor;
+  UmfPackLU<SparseMatrix<T1, ColMajor, T2> > umfpack_colmajor;
+  UmfPackLU<SparseMatrix<T1, RowMajor, T2> > umfpack_rowmajor;
   
   check_sparse_square_solving(umfpack_colmajor);
   check_sparse_square_solving(umfpack_rowmajor);
@@ -26,7 +26,9 @@ template<typename T> void test_umfpack_support_T()
 
 void test_umfpack_support()
 {
-  CALL_SUBTEST_1(test_umfpack_support_T<double>());
-  CALL_SUBTEST_2(test_umfpack_support_T<std::complex<double> >());
+  CALL_SUBTEST_1((test_umfpack_support_T<double, int>()));
+  CALL_SUBTEST_2((test_umfpack_support_T<std::complex<double>, int >()));
+  CALL_SUBTEST_3((test_umfpack_support_T<double, UF_long >()));
+  CALL_SUBTEST_4((test_umfpack_support_T<std::complex<double>,  UF_long >()));
 }
 

From 8c1652055aa65a0e7bbad1834fd603c8574ebaa2 Mon Sep 17 00:00:00 2001
From: Guillaume Jacob <guillaume.jacob@easymile.com>
Date: Mon, 9 Apr 2018 17:23:59 +0200
Subject: [PATCH 294/680] Fix code sample output in block(int, int, int, int)
 doxygen

---
 Eigen/src/plugins/BlockMethods.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h
index 0a53db19f..a5748525b 100644
--- a/Eigen/src/plugins/BlockMethods.h
+++ b/Eigen/src/plugins/BlockMethods.h
@@ -1044,7 +1044,7 @@ inline const typename ConstFixedBlockXpr<NRows,NCols>::Type block(Index startRow
 /// \a NRows is \a Dynamic, and the same for the number of columns.
 ///
 /// Example: \include MatrixBase_template_int_int_block_int_int_int_int.cpp
-/// Output: \verbinclude MatrixBase_template_int_int_block_int_int_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_block_int_int_int_int.out
 ///
 /// \note The usage of of this overload is discouraged from %Eigen 3.4, better used the generic
 /// block(Index,Index,NRowsType,NColsType), here is the one-to-one complete equivalence:

From c91906b065ddfd80997204e3072bb66bc9297bcd Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 09:59:59 +0200
Subject: [PATCH 295/680] Umfpack: UF_long has been removed in recent versions
 of suitesparse, and fix a few long-to-int conversions issues.

---
 Eigen/src/UmfPackSupport/UmfPackSupport.h | 99 +++++++++++++----------
 test/umfpack_support.cpp                  |  6 +-
 2 files changed, 59 insertions(+), 46 deletions(-)

diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h
index 4de7ccf4c..ba10a9318 100644
--- a/Eigen/src/UmfPackSupport/UmfPackSupport.h
+++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -10,6 +10,16 @@
 #ifndef EIGEN_UMFPACKSUPPORT_H
 #define EIGEN_UMFPACKSUPPORT_H
 
+// for compatibility with super old version of umfpack,
+// not sure this is really needed, but this is harmless.
+#ifndef SuiteSparse_long
+#ifdef UF_long
+#define SuiteSparse_long UF_long
+#else
+#error neither SuiteSparse_long nor UF_long are defined
+#endif
+#endif
+
 namespace Eigen {
 
 /* TODO extract L, extract U, compute det, etc... */
@@ -24,10 +34,10 @@ inline void umfpack_defaults(double control[UMFPACK_CONTROL], double, int)
 inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>, int)
 { umfpack_zi_defaults(control); }
 
-inline void umfpack_defaults(double control[UMFPACK_CONTROL], double, UF_long)
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], double, SuiteSparse_long)
 { umfpack_dl_defaults(control); }
 
-inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>, UF_long)
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>, SuiteSparse_long)
 { umfpack_zl_defaults(control); }
 
 // Report info
@@ -37,10 +47,10 @@ inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMF
 inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>, int)
 { umfpack_zi_report_info(control, info);}
 
-inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double, UF_long)
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double, SuiteSparse_long)
 { umfpack_dl_report_info(control, info);}
 
-inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>, UF_long)
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>, SuiteSparse_long)
 { umfpack_zl_report_info(control, info);}
 
 // Report status
@@ -50,10 +60,10 @@ inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, d
 inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>, int)
 { umfpack_zi_report_status(control, status);}
 
-inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double, UF_long)
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double, SuiteSparse_long)
 { umfpack_dl_report_status(control, status);}
 
-inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>, UF_long)
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>, SuiteSparse_long)
 { umfpack_zl_report_status(control, status);}
 
 // report control
@@ -63,10 +73,10 @@ inline void umfpack_report_control(double control[UMFPACK_CONTROL], double, int)
 inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>, int)
 { umfpack_zi_report_control(control);}
 
-inline void umfpack_report_control(double control[UMFPACK_CONTROL], double, UF_long)
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], double, SuiteSparse_long)
 { umfpack_dl_report_control(control);}
 
-inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>, UF_long)
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>, SuiteSparse_long)
 { umfpack_zl_report_control(control);}
 
 // Free numeric
@@ -76,10 +86,10 @@ inline void umfpack_free_numeric(void **Numeric, double, int)
 inline void umfpack_free_numeric(void **Numeric, std::complex<double>, int)
 { umfpack_zi_free_numeric(Numeric); *Numeric = 0; }
 
-inline void umfpack_free_numeric(void **Numeric, double, UF_long)
+inline void umfpack_free_numeric(void **Numeric, double, SuiteSparse_long)
 { umfpack_dl_free_numeric(Numeric); *Numeric = 0; }
 
-inline void umfpack_free_numeric(void **Numeric, std::complex<double>, UF_long)
+inline void umfpack_free_numeric(void **Numeric, std::complex<double>, SuiteSparse_long)
 { umfpack_zl_free_numeric(Numeric); *Numeric = 0; }
 
 // Free symbolic
@@ -89,10 +99,10 @@ inline void umfpack_free_symbolic(void **Symbolic, double, int)
 inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>, int)
 { umfpack_zi_free_symbolic(Symbolic); *Symbolic = 0; }
 
-inline void umfpack_free_symbolic(void **Symbolic, double, UF_long)
+inline void umfpack_free_symbolic(void **Symbolic, double, SuiteSparse_long)
 { umfpack_dl_free_symbolic(Symbolic); *Symbolic = 0; }
 
-inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>, UF_long)
+inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>, SuiteSparse_long)
 { umfpack_zl_free_symbolic(Symbolic); *Symbolic = 0; }
 
 // Symbolic
@@ -109,16 +119,16 @@ inline int umfpack_symbolic(int n_row,int n_col,
 {
   return umfpack_zi_symbolic(n_row,n_col,Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Control,Info);
 }
-inline int umfpack_symbolic(UF_long n_row,UF_long n_col,
-                            const UF_long Ap[], const UF_long Ai[], const double Ax[], void **Symbolic,
-                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+inline SuiteSparse_long umfpack_symbolic( SuiteSparse_long n_row,SuiteSparse_long n_col,
+                                          const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const double Ax[], void **Symbolic,
+                                          const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
 {
   return umfpack_dl_symbolic(n_row,n_col,Ap,Ai,Ax,Symbolic,Control,Info);
 }
 
-inline int umfpack_symbolic(UF_long n_row,UF_long n_col,
-                            const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[], void **Symbolic,
-                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+inline SuiteSparse_long umfpack_symbolic( SuiteSparse_long n_row,SuiteSparse_long n_col,
+                                          const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const std::complex<double> Ax[], void **Symbolic,
+                                          const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
 {
   return umfpack_zl_symbolic(n_row,n_col,Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Control,Info);
 }
@@ -137,16 +147,16 @@ inline int umfpack_numeric( const int Ap[], const int Ai[], const std::complex<d
 {
   return umfpack_zi_numeric(Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Numeric,Control,Info);
 }
-inline int umfpack_numeric( const UF_long Ap[], const UF_long Ai[], const double Ax[],
-                            void *Symbolic, void **Numeric,
-                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+inline SuiteSparse_long umfpack_numeric(const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const double Ax[],
+                                        void *Symbolic, void **Numeric,
+                                        const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
 {
   return umfpack_dl_numeric(Ap,Ai,Ax,Symbolic,Numeric,Control,Info);
 }
 
-inline int umfpack_numeric( const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[],
-                            void *Symbolic, void **Numeric,
-                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+inline SuiteSparse_long umfpack_numeric(const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const std::complex<double> Ax[],
+                                        void *Symbolic, void **Numeric,
+                                        const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
 {
   return umfpack_zl_numeric(Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Numeric,Control,Info);
 }
@@ -166,16 +176,16 @@ inline int umfpack_solve( int sys, const int Ap[], const int Ai[], const std::co
   return umfpack_zi_solve(sys,Ap,Ai,&numext::real_ref(Ax[0]),0,&numext::real_ref(X[0]),0,&numext::real_ref(B[0]),0,Numeric,Control,Info);
 }
 
-inline int umfpack_solve( int sys, const UF_long Ap[], const UF_long Ai[], const double Ax[],
-                          double X[], const double B[], void *Numeric,
-                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+inline SuiteSparse_long umfpack_solve(int sys, const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const double Ax[],
+                                      double X[], const double B[], void *Numeric,
+                                      const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
 {
   return umfpack_dl_solve(sys,Ap,Ai,Ax,X,B,Numeric,Control,Info);
 }
 
-inline int umfpack_solve( int sys, const UF_long Ap[], const UF_long Ai[], const std::complex<double> Ax[],
-                          std::complex<double> X[], const std::complex<double> B[], void *Numeric,
-                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+inline SuiteSparse_long umfpack_solve(int sys, const SuiteSparse_long Ap[], const SuiteSparse_long Ai[], const std::complex<double> Ax[],
+                                      std::complex<double> X[], const std::complex<double> B[], void *Numeric,
+                                      const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
 {
   return umfpack_zl_solve(sys,Ap,Ai,&numext::real_ref(Ax[0]),0,&numext::real_ref(X[0]),0,&numext::real_ref(B[0]),0,Numeric,Control,Info);
 }
@@ -191,12 +201,14 @@ inline int umfpack_get_lunz(int *lnz, int *unz, int *n_row, int *n_col, int *nz_
   return umfpack_zi_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
 }
 
-inline int umfpack_get_lunz(UF_long *lnz, UF_long *unz, UF_long *n_row, UF_long *n_col, UF_long *nz_udiag, void *Numeric, double)
+inline SuiteSparse_long umfpack_get_lunz( SuiteSparse_long *lnz, SuiteSparse_long *unz, SuiteSparse_long *n_row, SuiteSparse_long *n_col,
+                                          SuiteSparse_long *nz_udiag, void *Numeric, double)
 {
   return umfpack_dl_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
 }
 
-inline int umfpack_get_lunz(UF_long *lnz, UF_long *unz, UF_long *n_row, UF_long *n_col, UF_long *nz_udiag, void *Numeric, std::complex<double>)
+inline SuiteSparse_long umfpack_get_lunz( SuiteSparse_long *lnz, SuiteSparse_long *unz, SuiteSparse_long *n_row, SuiteSparse_long *n_col,
+                                          SuiteSparse_long *nz_udiag, void *Numeric, std::complex<double>)
 {
   return umfpack_zl_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
 }
@@ -217,14 +229,14 @@ inline int umfpack_get_numeric(int Lp[], int Lj[], std::complex<double> Lx[], in
   return umfpack_zi_get_numeric(Lp,Lj,Lx?&lx0_real:0,0,Up,Ui,Ux?&ux0_real:0,0,P,Q,
                                 Dx?&dx0_real:0,0,do_recip,Rs,Numeric);
 }
-inline int umfpack_get_numeric(UF_long Lp[], UF_long Lj[], double Lx[], UF_long Up[], UF_long Ui[], double Ux[],
-                               UF_long P[], UF_long Q[], double Dx[], UF_long *do_recip, double Rs[], void *Numeric)
+inline SuiteSparse_long umfpack_get_numeric(SuiteSparse_long Lp[], SuiteSparse_long Lj[], double Lx[], SuiteSparse_long Up[], SuiteSparse_long Ui[], double Ux[],
+                                            SuiteSparse_long P[], SuiteSparse_long Q[], double Dx[], SuiteSparse_long *do_recip, double Rs[], void *Numeric)
 {
   return umfpack_dl_get_numeric(Lp,Lj,Lx,Up,Ui,Ux,P,Q,Dx,do_recip,Rs,Numeric);
 }
 
-inline int umfpack_get_numeric(UF_long Lp[], UF_long Lj[], std::complex<double> Lx[], UF_long Up[], UF_long Ui[], std::complex<double> Ux[],
-                               UF_long P[], UF_long Q[], std::complex<double> Dx[], UF_long *do_recip, double Rs[], void *Numeric)
+inline SuiteSparse_long umfpack_get_numeric(SuiteSparse_long Lp[], SuiteSparse_long Lj[], std::complex<double> Lx[], SuiteSparse_long Up[], SuiteSparse_long Ui[], std::complex<double> Ux[],
+                                            SuiteSparse_long P[], SuiteSparse_long Q[], std::complex<double> Dx[], SuiteSparse_long *do_recip, double Rs[], void *Numeric)
 {
   double& lx0_real = numext::real_ref(Lx[0]);
   double& ux0_real = numext::real_ref(Ux[0]);
@@ -245,12 +257,12 @@ inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *N
   return umfpack_zi_get_determinant(&mx_real,0,Ex,NumericHandle,User_Info);
 }
 
-inline int umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], UF_long)
+inline SuiteSparse_long umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], SuiteSparse_long)
 {
   return umfpack_dl_get_determinant(Mx,Ex,NumericHandle,User_Info);
 }
 
-inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], UF_long)
+inline SuiteSparse_long umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO], SuiteSparse_long)
 {
   double& mx_real = numext::real_ref(*Mx);
   return umfpack_zl_get_determinant(&mx_real,0,Ex,NumericHandle,User_Info);
@@ -532,7 +544,7 @@ class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
 
     // cached data to reduce reallocation, etc.
     mutable LUMatrixType m_l;
-    int m_fact_errorCode;
+    StorageIndex m_fact_errorCode;
     UmfpackControl m_control;
     mutable UmfpackInfo m_umfpackInfo;
 
@@ -601,7 +613,7 @@ bool UmfPackLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBas
   eigen_assert((XDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major result yet");
   eigen_assert(b.derived().data() != x.derived().data() && " Umfpack does not support inplace solve");
 
-  int errorCode;
+  StorageIndex errorCode;
   Scalar* x_ptr = 0;
   Matrix<Scalar,Dynamic,1> x_tmp;
   if(x.innerStride()!=1)
@@ -613,9 +625,10 @@ bool UmfPackLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBas
   {
     if(x.innerStride()==1)
       x_ptr = &x.col(j).coeffRef(0);
-    errorCode = umfpack_solve(UMFPACK_A,
-        mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
-        x_ptr, &b.const_cast_derived().col(j).coeffRef(0), m_numeric, m_control.data(), m_umfpackInfo.data());
+      errorCode = umfpack_solve(UMFPACK_A,
+                                mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
+                                x_ptr, &b.const_cast_derived().col(j).coeffRef(0),
+                                m_numeric, m_control.data(), m_umfpackInfo.data());
     if(x.innerStride()!=1)
       x.col(j) = x_tmp;
     if (errorCode!=0)
diff --git a/test/umfpack_support.cpp b/test/umfpack_support.cpp
index 3e9b33ba4..0378801f4 100644
--- a/test/umfpack_support.cpp
+++ b/test/umfpack_support.cpp
@@ -27,8 +27,8 @@ template<typename T1, typename T2> void test_umfpack_support_T()
 void test_umfpack_support()
 {
   CALL_SUBTEST_1((test_umfpack_support_T<double, int>()));
-  CALL_SUBTEST_2((test_umfpack_support_T<std::complex<double>, int >()));
-  CALL_SUBTEST_3((test_umfpack_support_T<double, UF_long >()));
-  CALL_SUBTEST_4((test_umfpack_support_T<std::complex<double>,  UF_long >()));
+  CALL_SUBTEST_2((test_umfpack_support_T<std::complex<double>, int>()));
+  CALL_SUBTEST_3((test_umfpack_support_T<double, long >()));
+  CALL_SUBTEST_4((test_umfpack_support_T<std::complex<double>, long>()));
 }
 

From e798466871ceef80a5bd78eba460735fca829a8c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 10:46:11 +0200
Subject: [PATCH 296/680] bug #1538: update manual pages regarding BDCSVD.

---
 doc/LeastSquares.dox                     |  2 +-
 doc/TopicLinearAlgebraDecompositions.dox | 14 +++++++++++-
 doc/TutorialLinearAlgebra.dox            | 28 ++++++++++++++++++++----
 doc/examples/TutorialLinAlgSVDSolve.cpp  |  2 +-
 4 files changed, 39 insertions(+), 7 deletions(-)

diff --git a/doc/LeastSquares.dox b/doc/LeastSquares.dox
index e2191a22f..24dfe4b4f 100644
--- a/doc/LeastSquares.dox
+++ b/doc/LeastSquares.dox
@@ -16,7 +16,7 @@ equations is the fastest but least accurate, and the QR decomposition is in betw
 
 \section LeastSquaresSVD Using the SVD decomposition
 
-The \link JacobiSVD::solve() solve() \endlink method in the JacobiSVD class can be directly used to
+The \link BDCSVD::solve() solve() \endlink method in the BDCSVD class can be directly used to
 solve linear squares systems. It is not enough to compute only the singular values (the default for
 this class); you also need the singular vectors but the thin SVD decomposition suffices for
 computing least squares solutions:
diff --git a/doc/TopicLinearAlgebraDecompositions.dox b/doc/TopicLinearAlgebraDecompositions.dox
index 991f964cc..0965da872 100644
--- a/doc/TopicLinearAlgebraDecompositions.dox
+++ b/doc/TopicLinearAlgebraDecompositions.dox
@@ -4,7 +4,7 @@ namespace Eigen {
 
 This page presents a catalogue of the dense matrix decompositions offered by Eigen.
 For an introduction on linear solvers and decompositions, check this \link TutorialLinearAlgebra page \endlink.
-To get an overview of the true relative speed of the different decomposition, check this \link DenseDecompositionBenchmark benchmark \endlink.
+To get an overview of the true relative speed of the different decompositions, check this \link DenseDecompositionBenchmark benchmark \endlink.
 
 \section TopicLinAlgBigTable Catalogue of decompositions offered by Eigen
 
@@ -113,6 +113,18 @@ To get an overview of the true relative speed of the different decomposition, ch
 
     <tr><th class="inter" colspan="9">\n Singular values and eigenvalues decompositions</th></tr>
 
+    <tr>
+        <td>BDCSVD (divide \& conquer)</td>
+        <td>-</td>
+        <td>One of the fastest SVD algorithms</td>
+        <td>Excellent</td>
+        <td>Yes</td>
+        <td>Singular values/vectors, least squares</td>
+        <td>Yes (and does least squares)</td>
+        <td>Excellent</td>
+        <td>Blocked bidiagonalization</td>
+    </tr>
+
     <tr>
         <td>JacobiSVD (two-sided)</td>
         <td>-</td>
diff --git a/doc/TutorialLinearAlgebra.dox b/doc/TutorialLinearAlgebra.dox
index cb92ceeae..a72724143 100644
--- a/doc/TutorialLinearAlgebra.dox
+++ b/doc/TutorialLinearAlgebra.dox
@@ -73,7 +73,7 @@ depending on your matrix and the trade-off you want to make:
         <td>ColPivHouseholderQR</td>
         <td>colPivHouseholderQr()</td>
         <td>None</td>
-        <td>++</td>
+        <td>+</td>
         <td>-</td>
         <td>+++</td>
     </tr>
@@ -85,6 +85,14 @@ depending on your matrix and the trade-off you want to make:
         <td>- -</td>
         <td>+++</td>
     </tr>
+    <tr class="alt">
+        <td>CompleteOrthogonalDecomposition</td>
+        <td>completeOrthogonalDecomposition()</td>
+        <td>None</td>
+        <td>+</td>
+        <td>-</td>
+        <td>+++</td>
+    </tr>
     <tr class="alt">
         <td>LLT</td>
         <td>llt()</td>
@@ -101,15 +109,24 @@ depending on your matrix and the trade-off you want to make:
         <td>+</td>
         <td>++</td>
     </tr>
+    <tr class="alt">
+        <td>BDCSVD</td>
+        <td>bdcSvd()</td>
+        <td>None</td>
+        <td>-</td>
+        <td>-</td>
+        <td>+++</td>
+    </tr>
     <tr class="alt">
         <td>JacobiSVD</td>
         <td>jacobiSvd()</td>
         <td>None</td>
-        <td>- -</td>
+        <td>-</td>
         <td>- - -</td>
         <td>+++</td>
     </tr>
 </table>
+To get an overview of the true relative speed of the different decompositions, check this \link DenseDecompositionBenchmark benchmark \endlink.
 
 All of these decompositions offer a solve() method that works as in the above example.
 
@@ -183,8 +200,11 @@ Here is an example:
 
 \section TutorialLinAlgLeastsquares Least squares solving
 
-The most accurate method to do least squares solving is with a SVD decomposition. Eigen provides one
-as the JacobiSVD class, and its solve() is doing least-squares solving.
+The most accurate method to do least squares solving is with a SVD decomposition.
+Eigen provides two implementations.
+The recommended one is the BDCSVD class, which scale well for large problems
+and automatically fall-back to the JacobiSVD class for smaller problems.
+For both classes, their solve() method is doing least-squares solving.
 
 Here is an example:
 <table class="example">
diff --git a/doc/examples/TutorialLinAlgSVDSolve.cpp b/doc/examples/TutorialLinAlgSVDSolve.cpp
index 9fbc031de..f109f04e5 100644
--- a/doc/examples/TutorialLinAlgSVDSolve.cpp
+++ b/doc/examples/TutorialLinAlgSVDSolve.cpp
@@ -11,5 +11,5 @@ int main()
    VectorXf b = VectorXf::Random(3);
    cout << "Here is the right hand side b:\n" << b << endl;
    cout << "The least-squares solution is:\n"
-        << A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b) << endl;
+        << A.bdcSvd(ComputeThinU | ComputeThinV).solve(b) << endl;
 }

From 9c8decffbf13604d419d5ea530169cc91afcd45a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 11:30:14 +0200
Subject: [PATCH 297/680] Fix javascript hacks for oxygen 1.8.13

---
 doc/eigen_navtree_hacks.js | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/doc/eigen_navtree_hacks.js b/doc/eigen_navtree_hacks.js
index bd7e02b38..39c59f73c 100644
--- a/doc/eigen_navtree_hacks.js
+++ b/doc/eigen_navtree_hacks.js
@@ -65,6 +65,10 @@ function getNode(o, po)
 function resizeHeight() 
 {
   var toc = $("#nav-toc");
+  var header = $("#header");
+  var content = $("#doc-content");
+  var navtree = $("#nav-path");
+  var sidenav = $("#side-nav");
   var tocHeight = toc.height();  // <- we added this line
   var headerHeight = header.height();
   var footerHeight = footer.height();

From 426052ef6e10714f6eb1179eee405bfa77ce0cc0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 11:30:34 +0200
Subject: [PATCH 298/680] Update header/footer for doxygen 1.8.13

---
 doc/eigendoxy_footer.html.in |  4 ++--
 doc/eigendoxy_header.html.in | 18 ++++++++----------
 2 files changed, 10 insertions(+), 12 deletions(-)

diff --git a/doc/eigendoxy_footer.html.in b/doc/eigendoxy_footer.html.in
index 878244a19..9ac0596cb 100644
--- a/doc/eigendoxy_footer.html.in
+++ b/doc/eigendoxy_footer.html.in
@@ -5,14 +5,14 @@
     $navpath
     <li class="footer">$generatedby
     <a href="http://www.doxygen.org/index.html">
-    <img class="footer" src="$relpath$doxygen.png" alt="doxygen"/></a> $doxygenversion </li>
+    <img class="footer" src="$relpath^doxygen.png" alt="doxygen"/></a> $doxygenversion </li>
   </ul>
 </div>
 <!--END GENERATE_TREEVIEW-->
 <!--BEGIN !GENERATE_TREEVIEW-->
 <hr class="footer"/><address class="footer"><small>
 $generatedby &#160;<a href="http://www.doxygen.org/index.html">
-<img class="footer" src="$relpath$doxygen.png" alt="doxygen"/>
+<img class="footer" src="$relpath^doxygen.png" alt="doxygen"/>
 </a> $doxygenversion
 </small></address>
 <!--END !GENERATE_TREEVIEW-->
diff --git a/doc/eigendoxy_header.html.in b/doc/eigendoxy_header.html.in
index 0f3859f40..bb149f8f0 100644
--- a/doc/eigendoxy_header.html.in
+++ b/doc/eigendoxy_header.html.in
@@ -4,25 +4,23 @@
 <meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
 <meta http-equiv="X-UA-Compatible" content="IE=9"/>
 <meta name="generator" content="Doxygen $doxygenversion"/>
+<meta name="viewport" content="width=device-width, initial-scale=1"/>
 <!--BEGIN PROJECT_NAME--><title>$projectname: $title</title><!--END PROJECT_NAME-->
 <!--BEGIN !PROJECT_NAME--><title>$title</title><!--END !PROJECT_NAME-->
-<link href="$relpath$tabs.css" rel="stylesheet" type="text/css"/>
-<script type="text/javascript" src="$relpath$jquery.js"></script>
-<script type="text/javascript" src="$relpath$dynsections.js"></script>
+<link href="$relpath^tabs.css" rel="stylesheet" type="text/css"/>
+<script type="text/javascript" src="$relpath^jquery.js"></script>
+<script type="text/javascript" src="$relpath^dynsections.js"></script>
 $treeview
 $search
 $mathjax
-<link href="$relpath$$stylesheet"   rel="stylesheet" type="text/css" />
+<link href="$relpath^$stylesheet" rel="stylesheet" type="text/css" />
 <link href="$relpath$eigendoxy.css" rel="stylesheet" type="text/css">
 <!-- $extrastylesheet -->
 <script type="text/javascript" src="$relpath$eigen_navtree_hacks.js"></script>
-<!-- <script type="text/javascript"> -->
-<!-- </script> -->
 
 </head>
 <body>
 <div id="top"><!-- do not remove this div, it is closed by doxygen! -->
-<!-- <a name="top"></a> -->
 
 <!--BEGIN TITLEAREA-->
 <div id="titlearea">
@@ -30,10 +28,10 @@ $mathjax
  <tbody>
  <tr style="height: 56px;">
   <!--BEGIN PROJECT_LOGO-->
-  <td id="projectlogo"><img alt="Logo" src="$relpath$$projectlogo"/></td>
+  <td id="projectlogo"><img alt="Logo" src="$relpath^$projectlogo"/></td>
   <!--END PROJECT_LOGO-->
   <!--BEGIN PROJECT_NAME-->
-  <td style="padding-left: 0.5em;">
+  <td id="projectalign" style="padding-left: 0.5em;">
    <div id="projectname"><a href="http://eigen.tuxfamily.org">$projectname</a>
    <!--BEGIN PROJECT_NUMBER-->&#160;<span id="projectnumber">$projectnumber</span><!--END PROJECT_NUMBER-->
    </div>
@@ -42,7 +40,7 @@ $mathjax
   <!--END PROJECT_NAME-->
   <!--BEGIN !PROJECT_NAME-->
    <!--BEGIN PROJECT_BRIEF-->
-    <td style="padding-left: 0.5em;">
+    <td id="projectalign" style="padding-left: 0.5em;">
     <div id="projectbrief">$projectbrief</div>
     </td>
    <!--END PROJECT_BRIEF-->

From 79266fec75a78a31126efbbfd603d8459660350d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 11:31:17 +0200
Subject: [PATCH 299/680] extend doxygen splitter for huge screens

---
 doc/eigendoxy.css | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/doc/eigendoxy.css b/doc/eigendoxy.css
index b99d7914a..427b128ba 100644
--- a/doc/eigendoxy.css
+++ b/doc/eigendoxy.css
@@ -219,3 +219,8 @@ h3.version {
 td.width20em p.endtd {
   width:  20em;
 }
+
+/* needed for huge screens */
+.ui-resizable-e {
+  background-repeat: repeat-y;
+}
\ No newline at end of file

From b0eda3cb9fc08d7b0bd39d99a1874b9575439e2c Mon Sep 17 00:00:00 2001
From: Weiming Zhao <weiming.zhao@alibaba-inc.com>
Date: Wed, 11 Apr 2018 11:37:06 +0200
Subject: [PATCH 300/680] Avoid using memcpy for non-POD elements

---
 unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index 2264be391..fe62ff1ea 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -55,7 +55,7 @@ struct TensorEvaluator
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* dest) {
-    if (dest) {
+    if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization && dest) {
       m_device.memcpy((void*)dest, m_data, sizeof(Scalar) * m_dims.TotalSize());
       return false;
     }

From e43ca0320d58587736d8bab7c6ed7212aa6b8d9d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Apr 2018 15:24:13 +0200
Subject: [PATCH 301/680] bug #1520: workaround some -Wfloat-equal warnings by
 calling std::equal_to

---
 Eigen/src/Core/MathFunctions.h  |  6 +++---
 Eigen/src/Core/arch/CUDA/Half.h |  4 ++--
 Eigen/src/Core/util/Meta.h      | 20 ++++++++++++++++++++
 3 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 84f9d0cd5..ac1b7a6a1 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -473,11 +473,11 @@ namespace std_fallback {
 
     EIGEN_USING_STD_MATH(exp);
     Scalar u = exp(x);
-    if (u == Scalar(1)) {
+    if (numext::equal_strict(u, Scalar(1))) {
       return x;
     }
     Scalar um1 = u - RealScalar(1);
-    if (um1 == Scalar(-1)) {
+    if (numext::equal_strict(um1, Scalar(-1))) {
       return RealScalar(-1);
     }
 
@@ -519,7 +519,7 @@ namespace std_fallback {
     typedef typename NumTraits<Scalar>::Real RealScalar;
     EIGEN_USING_STD_MATH(log);
     Scalar x1p = RealScalar(1) + x;
-    return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
+    return numext::equal_strict(x1p, Scalar(1)) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
   }
 }
 
diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h
index 2bb3d0d9b..c10550050 100644
--- a/Eigen/src/Core/arch/CUDA/Half.h
+++ b/Eigen/src/Core/arch/CUDA/Half.h
@@ -299,10 +299,10 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b)
   return a;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
-  return float(a) == float(b);
+  return numext::equal_strict(float(a),float(b));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
-  return float(a) != float(b);
+  return numext::not_equal_strict(float(a), float(b));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
   return float(a) < float(b);
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 998b8921a..e7729fdc8 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -543,6 +543,26 @@ T div_ceil(const T &a, const T &b)
   return (a+b-1) / b;
 }
 
+// The aim of the following functions is to bypass -Wfloat-equal warnings
+// when we really want a strict equality comparison on floating points.
+template<typename X, typename Y>
+bool equal_strict(const X& x,const Y& y) { return x == y; }
+
+template<>
+bool equal_strict(const float& x,const float& y) { return std::equal_to<float>()(x,y); }
+
+template<>
+bool equal_strict(const double& x,const double& y) { return std::equal_to<double>()(x,y); }
+
+template<typename X, typename Y>
+bool not_equal_strict(const X& x,const Y& y) { return x != y; }
+
+template<>
+bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to<float>()(x,y); }
+
+template<>
+bool not_equal_strict(const double& x,const double& y) { return std::not_equal_to<double>()(x,y); }
+
 } // end namespace numext
 
 } // end namespace Eigen

From 7a9089c33cdb7f9d6cea0380ca75a44969c3ef78 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Apr 2018 08:51:47 +0200
Subject: [PATCH 302/680] fix linking issue

---
 Eigen/src/Core/util/Meta.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index e7729fdc8..1180f0c13 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -545,22 +545,22 @@ T div_ceil(const T &a, const T &b)
 
 // The aim of the following functions is to bypass -Wfloat-equal warnings
 // when we really want a strict equality comparison on floating points.
-template<typename X, typename Y>
+template<typename X, typename Y> EIGEN_STRONG_INLINE
 bool equal_strict(const X& x,const Y& y) { return x == y; }
 
-template<>
+template<> EIGEN_STRONG_INLINE
 bool equal_strict(const float& x,const float& y) { return std::equal_to<float>()(x,y); }
 
-template<>
+template<> EIGEN_STRONG_INLINE
 bool equal_strict(const double& x,const double& y) { return std::equal_to<double>()(x,y); }
 
-template<typename X, typename Y>
+template<typename X, typename Y> EIGEN_STRONG_INLINE
 bool not_equal_strict(const X& x,const Y& y) { return x != y; }
 
-template<>
+template<> EIGEN_STRONG_INLINE
 bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to<float>()(x,y); }
 
-template<>
+template<> EIGEN_STRONG_INLINE
 bool not_equal_strict(const double& x,const double& y) { return std::not_equal_to<double>()(x,y); }
 
 } // end namespace numext

From 072e111ec08be523f8bf5b4c742b5cbcb8ea0983 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 19:00:34 +0200
Subject: [PATCH 303/680] SelfAdjointView<...,Mode> causes a static assert
 since commit d820ab9edc0b38af4cdb3d545714a0c9083e5a78

---
 Eigen/src/Core/SelfAdjointView.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h
index be219726b..2cf3fa1ef 100644
--- a/Eigen/src/Core/SelfAdjointView.h
+++ b/Eigen/src/Core/SelfAdjointView.h
@@ -191,7 +191,7 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
                                    TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type(tmp2);
     }
 
-    typedef SelfAdjointView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    typedef SelfAdjointView<const MatrixConjugateReturnType,UpLo> ConjugateReturnType;
     /** \sa MatrixBase::conjugate() const */
     EIGEN_DEVICE_FUNC
     inline const ConjugateReturnType conjugate() const

From 4d392d93aaf35447dff7b4522dcf587e49a44b8f Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 19:01:37 +0200
Subject: [PATCH 304/680] Make hypot_impl compile again for types with
 expression-templates (e.g., boost::multiprecision)

---
 Eigen/src/Core/MathFunctionsImpl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/MathFunctionsImpl.h b/Eigen/src/Core/MathFunctionsImpl.h
index cc2ac0700..28737c15e 100644
--- a/Eigen/src/Core/MathFunctionsImpl.h
+++ b/Eigen/src/Core/MathFunctionsImpl.h
@@ -85,7 +85,7 @@ struct hypot_impl
   static inline RealScalar run(const Scalar& x, const Scalar& y)
   {
     EIGEN_USING_STD_MATH(abs);
-    return positive_real_hypot(abs(x), abs(y));
+    return positive_real_hypot<RealScalar>(abs(x), abs(y));
   }
 };
 

From 84dcd998a92e4f90f963d2849252f4338adaf398 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 19:10:23 +0200
Subject: [PATCH 305/680] Recent Adolc versions require C++11

---
 unsupported/test/CMakeLists.txt | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 22647cadd..42b790d55 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -41,11 +41,16 @@ else(GOOGLEHASH_FOUND)
   ei_add_property(EIGEN_MISSING_BACKENDS  "GoogleHash, ")
 endif(GOOGLEHASH_FOUND)
 
+
 find_package(Adolc)
 if(ADOLC_FOUND)
   include_directories(${ADOLC_INCLUDES})
   ei_add_property(EIGEN_TESTED_BACKENDS "Adolc, ")
-  ei_add_test(forward_adolc "" ${ADOLC_LIBRARIES})
+  if(EIGEN_TEST_CXX11)
+    ei_add_test(forward_adolc "" ${ADOLC_LIBRARIES})
+  else()
+    message(STATUS "Adolc found, but tests require C++11 mode")
+  endif()
 else(ADOLC_FOUND)
   ei_add_property(EIGEN_MISSING_BACKENDS "Adolc, ")
 endif(ADOLC_FOUND)

From 2cbb00b18ee3d02d3b747bcb1775d7f851e46d7e Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 19:14:25 +0200
Subject: [PATCH 306/680] No need to make noise, if KLU is found

---
 test/CMakeLists.txt | 1 -
 1 file changed, 1 deletion(-)

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 8bcf3f7c5..f8deafc07 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -309,7 +309,6 @@ if(UMFPACK_FOUND)
 endif()
 
 if(KLU_FOUND OR SuiteSparse_FOUND)
-  message("ADDING KLU TEST")
   ei_add_test(klu_support "" "${KLU_ALL_LIBS}")
 endif()
 

From c8b19702bc25b64b5ac9ab8f590a4706c3f46208 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 20:36:58 +0200
Subject: [PATCH 307/680] Limit test size for sparse Cholesky solvers to
 EIGEN_TEST_MAX_SIZE

---
 test/simplicial_cholesky.cpp | 4 ++--
 test/sparse_solver.h         | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/test/simplicial_cholesky.cpp b/test/simplicial_cholesky.cpp
index 649c817b4..5bb1e0d6c 100644
--- a/test/simplicial_cholesky.cpp
+++ b/test/simplicial_cholesky.cpp
@@ -35,8 +35,8 @@ template<typename T, typename I> void test_simplicial_cholesky_T()
   check_sparse_spd_determinant(ldlt_colmajor_lower_amd);
   check_sparse_spd_determinant(ldlt_colmajor_upper_amd);
   
-  check_sparse_spd_solving(ldlt_colmajor_lower_nat, 300, 1000);
-  check_sparse_spd_solving(ldlt_colmajor_upper_nat, 300, 1000);
+  check_sparse_spd_solving(ldlt_colmajor_lower_nat, (std::min)(300,EIGEN_TEST_MAX_SIZE), 1000);
+  check_sparse_spd_solving(ldlt_colmajor_upper_nat, (std::min)(300,EIGEN_TEST_MAX_SIZE), 1000);
 }
 
 void test_simplicial_cholesky()
diff --git a/test/sparse_solver.h b/test/sparse_solver.h
index 5145bc3eb..19416ed5d 100644
--- a/test/sparse_solver.h
+++ b/test/sparse_solver.h
@@ -266,7 +266,7 @@ std::string solver_stats(const SparseSolverBase<Derived> &/*solver*/)
 }
 #endif
 
-template<typename Solver> void check_sparse_spd_solving(Solver& solver, int maxSize = 300, int maxRealWorldSize = 100000)
+template<typename Solver> void check_sparse_spd_solving(Solver& solver, int maxSize = (std::min)(300,EIGEN_TEST_MAX_SIZE), int maxRealWorldSize = 100000)
 {
   typedef typename Solver::MatrixType Mat;
   typedef typename Mat::Scalar Scalar;

From c9ecfff2e6ab3e68bc4569e05f3e29503e756da2 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 13 Apr 2018 21:05:28 +0000
Subject: [PATCH 308/680] Add links where to make PRs and report bugs into
 README.md

---
 README.md | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/README.md b/README.md
index 4654a81c3..99c9e2933 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,7 @@
 **Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.**
 
 For more information go to http://eigen.tuxfamily.org/.
+
+For ***pull request*** please only use the official repository at https://bitbucket.org/eigen/eigen.
+
+For ***bug reports*** and ***feature requests*** go to http://eigen.tuxfamily.org/bz.

From 42715533f1bff346d024edce4f79c5fdef7dba04 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sun, 15 Apr 2018 10:15:28 +0200
Subject: [PATCH 309/680] bug #1493: Make representation of HouseholderSequence
 consistent and working for complex numbers. Made corresponding unit test
 actually test that. Also simplify implementation of QR decompositions

---
 .../src/Eigenvalues/HessenbergDecomposition.h |  2 +-
 Eigen/src/Householder/Householder.h           |  8 ++---
 Eigen/src/Householder/HouseholderSequence.h   | 36 +++++++++++++++----
 Eigen/src/QR/ColPivHouseholderQR.h            |  6 +---
 .../src/QR/CompleteOrthogonalDecomposition.h  |  7 ++--
 Eigen/src/QR/HouseholderQR.h                  |  6 +---
 Eigen/src/SVD/UpperBidiagonalization.h        |  2 +-
 test/householder.cpp                          | 29 ++++++++++-----
 test/qr.cpp                                   |  2 +-
 9 files changed, 60 insertions(+), 38 deletions(-)

diff --git a/Eigen/src/Eigenvalues/HessenbergDecomposition.h b/Eigen/src/Eigenvalues/HessenbergDecomposition.h
index f647f69b0..d947dac4e 100644
--- a/Eigen/src/Eigenvalues/HessenbergDecomposition.h
+++ b/Eigen/src/Eigenvalues/HessenbergDecomposition.h
@@ -315,7 +315,7 @@ void HessenbergDecomposition<MatrixType>::_compute(MatrixType& matA, CoeffVector
 
     // A = A H'
     matA.rightCols(remainingSize)
-        .applyHouseholderOnTheRight(matA.col(i).tail(remainingSize-1).conjugate(), numext::conj(h), &temp.coeffRef(0));
+        .applyHouseholderOnTheRight(matA.col(i).tail(remainingSize-1), numext::conj(h), &temp.coeffRef(0));
   }
 }
 
diff --git a/Eigen/src/Householder/Householder.h b/Eigen/src/Householder/Householder.h
index 80de2c305..a5f336d18 100644
--- a/Eigen/src/Householder/Householder.h
+++ b/Eigen/src/Householder/Householder.h
@@ -103,7 +103,7 @@ void MatrixBase<Derived>::makeHouseholder(
   * \param essential the essential part of the vector \c v
   * \param tau the scaling factor of the Householder transformation
   * \param workspace a pointer to working space with at least
-  *                  this->cols() * essential.size() entries
+  *                  this->cols() entries
   *
   * \sa MatrixBase::makeHouseholder(), MatrixBase::makeHouseholderInPlace(), 
   *     MatrixBase::applyHouseholderOnTheRight()
@@ -140,7 +140,7 @@ void MatrixBase<Derived>::applyHouseholderOnTheLeft(
   * \param essential the essential part of the vector \c v
   * \param tau the scaling factor of the Householder transformation
   * \param workspace a pointer to working space with at least
-  *                  this->cols() * essential.size() entries
+  *                  this->rows() entries
   *
   * \sa MatrixBase::makeHouseholder(), MatrixBase::makeHouseholderInPlace(), 
   *     MatrixBase::applyHouseholderOnTheLeft()
@@ -160,10 +160,10 @@ void MatrixBase<Derived>::applyHouseholderOnTheRight(
   {
     Map<typename internal::plain_col_type<PlainObject>::type> tmp(workspace,rows());
     Block<Derived, Derived::RowsAtCompileTime, EssentialPart::SizeAtCompileTime> right(derived(), 0, 1, rows(), cols()-1);
-    tmp.noalias() = right * essential.conjugate();
+    tmp.noalias() = right * essential;
     tmp += this->col(0);
     this->col(0) -= tau * tmp;
-    right.noalias() -= tau * tmp * essential.transpose();
+    right.noalias() -= tau * tmp * essential.adjoint();
   }
 }
 
diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h
index 3ce0a693d..13030c664 100644
--- a/Eigen/src/Householder/HouseholderSequence.h
+++ b/Eigen/src/Householder/HouseholderSequence.h
@@ -140,6 +140,22 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       Side
     > ConjugateReturnType;
 
+    typedef HouseholderSequence<
+      VectorsType,
+      typename internal::conditional<NumTraits<Scalar>::IsComplex,
+        typename internal::remove_all<typename CoeffsType::ConjugateReturnType>::type,
+        CoeffsType>::type,
+      Side
+    > AdjointReturnType;
+
+    typedef HouseholderSequence<
+      typename internal::conditional<NumTraits<Scalar>::IsComplex,
+        typename internal::remove_all<typename VectorsType::ConjugateReturnType>::type,
+        VectorsType>::type,
+      CoeffsType,
+      Side
+    > TransposeReturnType;
+
     /** \brief Constructor.
       * \param[in]  v      %Matrix containing the essential parts of the Householder vectors
       * \param[in]  h      Vector containing the Householder coefficients
@@ -206,9 +222,12 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     }
 
     /** \brief %Transpose of the Householder sequence. */
-    HouseholderSequence transpose() const
+    TransposeReturnType transpose() const
     {
-      return HouseholderSequence(*this).setTrans(!m_trans);
+      return TransposeReturnType(m_vectors.conjugate(), m_coeffs)
+              .setTrans(!m_trans)
+              .setLength(m_length)
+              .setShift(m_shift);
     }
 
     /** \brief Complex conjugate of the Householder sequence. */
@@ -221,13 +240,16 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     }
 
     /** \brief Adjoint (conjugate transpose) of the Householder sequence. */
-    ConjugateReturnType adjoint() const
+    AdjointReturnType adjoint() const
     {
-      return conjugate().setTrans(!m_trans);
+      return AdjointReturnType(m_vectors, m_coeffs.conjugate())
+              .setTrans(!m_trans)
+              .setLength(m_length)
+              .setShift(m_shift);
     }
 
     /** \brief Inverse of the Householder sequence (equals the adjoint). */
-    ConjugateReturnType inverse() const { return adjoint(); }
+    AdjointReturnType inverse() const { return adjoint(); }
 
     /** \internal */
     template<typename DestType> inline void evalTo(DestType& dst) const
@@ -273,10 +295,10 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
           Index cornerSize = rows() - k - m_shift;
           if(m_trans)
             dst.bottomRightCorner(cornerSize, cornerSize)
-               .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), &workspace.coeffRef(0));
+               .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), workspace.data());
           else
             dst.bottomRightCorner(cornerSize, cornerSize)
-               .applyHouseholderOnTheLeft(essentialVector(k), m_coeffs.coeff(k), &workspace.coeffRef(0));
+               .applyHouseholderOnTheLeft(essentialVector(k), m_coeffs.coeff(k), workspace.data());
         }
       }
     }
diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h
index ed47b05e3..1faa3442e 100644
--- a/Eigen/src/QR/ColPivHouseholderQR.h
+++ b/Eigen/src/QR/ColPivHouseholderQR.h
@@ -595,11 +595,7 @@ void ColPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &
 
   typename RhsType::PlainObject c(rhs);
 
-  // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T
-  c.applyOnTheLeft(householderSequence(m_qr, m_hCoeffs)
-                    .setLength(nonzero_pivots)
-                    .transpose()
-    );
+  c.applyOnTheLeft(householderQ().setLength(nonzero_pivots).adjoint() );
 
   m_qr.topLeftCorner(nonzero_pivots, nonzero_pivots)
       .template triangularView<Upper>()
diff --git a/Eigen/src/QR/CompleteOrthogonalDecomposition.h b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
index 880becb25..03017a375 100644
--- a/Eigen/src/QR/CompleteOrthogonalDecomposition.h
+++ b/Eigen/src/QR/CompleteOrthogonalDecomposition.h
@@ -452,7 +452,7 @@ void CompleteOrthogonalDecomposition<MatrixType>::computeInPlace()
         // Apply Z(k) to the first k rows of X_k
         m_cpqr.m_qr.topRightCorner(k, cols - rank + 1)
             .applyHouseholderOnTheRight(
-                m_cpqr.m_qr.row(k).tail(cols - rank).transpose(), m_zCoeffs(k),
+                m_cpqr.m_qr.row(k).tail(cols - rank).adjoint(), m_zCoeffs(k),
                 &m_temp(0));
       }
       if (k != rank - 1) {
@@ -500,11 +500,8 @@ void CompleteOrthogonalDecomposition<_MatrixType>::_solve_impl(
   }
 
   // Compute c = Q^* * rhs
-  // Note that the matrix Q = H_0^* H_1^*... so its inverse is
-  // Q^* = (H_0 H_1 ...)^T
   typename RhsType::PlainObject c(rhs);
-  c.applyOnTheLeft(
-      householderSequence(matrixQTZ(), hCoeffs()).setLength(rank).transpose());
+  c.applyOnTheLeft(matrixQ().setLength(rank).adjoint());
 
   // Solve T z = c(1:rank, :)
   dst.topRows(rank) = matrixT()
diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h
index 762b21c36..13c02685d 100644
--- a/Eigen/src/QR/HouseholderQR.h
+++ b/Eigen/src/QR/HouseholderQR.h
@@ -353,11 +353,7 @@ void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) c
 
   typename RhsType::PlainObject c(rhs);
 
-  // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T
-  c.applyOnTheLeft(householderSequence(
-    m_qr.leftCols(rank),
-    m_hCoeffs.head(rank)).transpose()
-  );
+  c.applyOnTheLeft(householderQ().setLength(rank).adjoint() );
 
   m_qr.topLeftCorner(rank, rank)
       .template triangularView<Upper>()
diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h
index 0526ac931..997defc47 100644
--- a/Eigen/src/SVD/UpperBidiagonalization.h
+++ b/Eigen/src/SVD/UpperBidiagonalization.h
@@ -127,7 +127,7 @@ void upperbidiagonalization_inplace_unblocked(MatrixType& mat,
        .makeHouseholderInPlace(mat.coeffRef(k,k+1), upper_diagonal[k]);
     // apply householder transform to remaining part of mat on the left
     mat.bottomRightCorner(remainingRows-1, remainingCols)
-       .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData);
+       .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).adjoint(), mat.coeff(k,k+1), tempData);
   }
 }
 
diff --git a/test/householder.cpp b/test/householder.cpp
index c5f6b5e4f..78d44d48a 100644
--- a/test/householder.cpp
+++ b/test/householder.cpp
@@ -49,6 +49,17 @@ template<typename MatrixType> void householder(const MatrixType& m)
   v1.applyHouseholderOnTheLeft(essential,beta,tmp);
   VERIFY_IS_APPROX(v1.norm(), v2.norm());
 
+  // reconstruct householder matrix:
+  SquareMatrixType id, H1, H2;
+  id.setIdentity(rows, rows);
+  H1 = H2 = id;
+  VectorType vv(rows);
+  vv << Scalar(1), essential;
+  H1.applyHouseholderOnTheLeft(essential, beta, tmp);
+  H2.applyHouseholderOnTheRight(essential, beta, tmp);
+  VERIFY_IS_APPROX(H1, H2);
+  VERIFY_IS_APPROX(H1, id - beta * vv*vv.adjoint());
+
   MatrixType m1(rows, cols),
              m2(rows, cols);
 
@@ -69,7 +80,7 @@ template<typename MatrixType> void householder(const MatrixType& m)
   m3.rowwise() = v1.transpose();
   m4 = m3;
   m3.row(0).makeHouseholder(essential, beta, alpha);
-  m3.applyHouseholderOnTheRight(essential,beta,tmp);
+  m3.applyHouseholderOnTheRight(essential.conjugate(),beta,tmp);
   VERIFY_IS_APPROX(m3.norm(), m4.norm());
   if(rows>=2) VERIFY_IS_MUCH_SMALLER_THAN(m3.block(0,1,rows,rows-1).norm(), m3.norm());
   VERIFY_IS_MUCH_SMALLER_THAN(numext::imag(m3(0,0)), numext::real(m3(0,0)));
@@ -104,14 +115,14 @@ template<typename MatrixType> void householder(const MatrixType& m)
   VERIFY_IS_APPROX(hseq_mat.adjoint(),    hseq_mat_adj);
   VERIFY_IS_APPROX(hseq_mat.conjugate(),  hseq_mat_conj);
   VERIFY_IS_APPROX(hseq_mat.transpose(),  hseq_mat_trans);
-  VERIFY_IS_APPROX(hseq_mat * m6,             hseq_mat * m6);
-  VERIFY_IS_APPROX(hseq_mat.adjoint() * m6,   hseq_mat_adj * m6);
-  VERIFY_IS_APPROX(hseq_mat.conjugate() * m6, hseq_mat_conj * m6);
-  VERIFY_IS_APPROX(hseq_mat.transpose() * m6, hseq_mat_trans * m6);
-  VERIFY_IS_APPROX(m6 * hseq_mat,             m6 * hseq_mat);
-  VERIFY_IS_APPROX(m6 * hseq_mat.adjoint(),   m6 * hseq_mat_adj);
-  VERIFY_IS_APPROX(m6 * hseq_mat.conjugate(), m6 * hseq_mat_conj);
-  VERIFY_IS_APPROX(m6 * hseq_mat.transpose(), m6 * hseq_mat_trans);
+  VERIFY_IS_APPROX(hseq * m6,             hseq_mat * m6);
+  VERIFY_IS_APPROX(hseq.adjoint() * m6,   hseq_mat_adj * m6);
+  VERIFY_IS_APPROX(hseq.conjugate() * m6, hseq_mat_conj * m6);
+  VERIFY_IS_APPROX(hseq.transpose() * m6, hseq_mat_trans * m6);
+  VERIFY_IS_APPROX(m6 * hseq,             m6 * hseq_mat);
+  VERIFY_IS_APPROX(m6 * hseq.adjoint(),   m6 * hseq_mat_adj);
+  VERIFY_IS_APPROX(m6 * hseq.conjugate(), m6 * hseq_mat_conj);
+  VERIFY_IS_APPROX(m6 * hseq.transpose(), m6 * hseq_mat_trans);
 
   // test householder sequence on the right with a shift
 
diff --git a/test/qr.cpp b/test/qr.cpp
index dfcc1e8f9..00ef99ef9 100644
--- a/test/qr.cpp
+++ b/test/qr.cpp
@@ -85,7 +85,7 @@ template<typename MatrixType> void qr_invertible()
   qr.compute(m1);
   VERIFY_IS_APPROX(log(absdet), qr.logAbsDeterminant());
   // This test is tricky if the determinant becomes too small.
-  // Since we generate random numbers with magnitude rrange [0,1], the average determinant is 0.5^size
+  // Since we generate random numbers with magnitude range [0,1], the average determinant is 0.5^size
   VERIFY_IS_MUCH_SMALLER_THAN( abs(absdet-qr.absDeterminant()), numext::maxi(RealScalar(pow(0.5,size)),numext::maxi<RealScalar>(abs(absdet),abs(qr.absDeterminant()))) );
   
 }

From 775766d175f52d4aa86268c6e6ff43cee374a56a Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sun, 15 Apr 2018 18:43:56 +0200
Subject: [PATCH 310/680] Add parenthesis to fix compiler warnings

---
 test/main.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/main.h b/test/main.h
index 14f4e3e7a..0fcd6cb76 100644
--- a/test/main.h
+++ b/test/main.h
@@ -265,7 +265,7 @@ namespace Eigen
       {                                       \
         Eigen::no_more_assert = true;         \
         if(report_on_cerr_on_assert_failure)  \
-          eigen_plain_assert(a && #MSG);      \
+          eigen_plain_assert((a) && #MSG);      \
         else                                  \
           EIGEN_THROW_X(Eigen::eigen_static_assert_exception()); \
       }

From 39c2cba810a573ae4d0efd2b0b80e08c934b99b3 Mon Sep 17 00:00:00 2001
From: nicolov <nicolo.valigi@gmail.com>
Date: Fri, 13 Apr 2018 22:29:10 +0000
Subject: [PATCH 311/680] Add a specialization of Eigen::numext::conj for
 std::complex<T> to be used when compiling a cuda kernel. This fixes the
 compilation of TensorFlow 1.4 with clang 6.0 used as CUDA compiler with
 libc++.

This follows the previous change in https://bitbucket.org/eigen/eigen/commits/2a69290ddb165b7103c87ba8f5b257eca23f62aa
, which mentions OSX (I guess because it uses libc++ too).
---
 Eigen/src/Core/MathFunctions.h | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index ac1b7a6a1..05462c5e1 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -238,7 +238,7 @@ struct imag_ref_retval
 ****************************************************************************/
 
 template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
-struct conj_impl
+struct conj_default_impl
 {
   EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
@@ -248,7 +248,7 @@ struct conj_impl
 };
 
 template<typename Scalar>
-struct conj_impl<Scalar,true>
+struct conj_default_impl<Scalar,true>
 {
   EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
@@ -258,6 +258,20 @@ struct conj_impl<Scalar,true>
   }
 };
 
+template<typename Scalar> struct conj_impl : conj_default_impl<Scalar> {};
+
+#ifdef EIGEN_CUDA_ARCH
+template<typename T>
+struct conj_impl<std::complex<T> >
+{
+  EIGEN_DEVICE_FUNC
+  static inline std::complex<T> run(const std::complex<T>& x)
+  {
+    return std::complex<T>(x.real(), -x.imag());
+  }
+};
+#endif
+
 template<typename Scalar>
 struct conj_retval
 {

From 50633d1a83eb830803e63ae12596fad9f28db866 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 17 Apr 2018 11:30:27 +0200
Subject: [PATCH 312/680] Renamed .trans() et al. to .reverseFlag() et at.
 Adapted documentation of .setReverseFlag()

---
 Eigen/src/Householder/HouseholderSequence.h | 44 +++++++++++----------
 1 file changed, 23 insertions(+), 21 deletions(-)

diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h
index 13030c664..fad1d5ab6 100644
--- a/Eigen/src/Householder/HouseholderSequence.h
+++ b/Eigen/src/Householder/HouseholderSequence.h
@@ -174,7 +174,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       * \sa setLength(), setShift()
       */
     HouseholderSequence(const VectorsType& v, const CoeffsType& h)
-      : m_vectors(v), m_coeffs(h), m_trans(false), m_length(v.diagonalSize()),
+      : m_vectors(v), m_coeffs(h), m_reverse(false), m_length(v.diagonalSize()),
         m_shift(0)
     {
     }
@@ -183,7 +183,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     HouseholderSequence(const HouseholderSequence& other)
       : m_vectors(other.m_vectors),
         m_coeffs(other.m_coeffs),
-        m_trans(other.m_trans),
+        m_reverse(other.m_reverse),
         m_length(other.m_length),
         m_shift(other.m_shift)
     {
@@ -225,7 +225,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     TransposeReturnType transpose() const
     {
       return TransposeReturnType(m_vectors.conjugate(), m_coeffs)
-              .setTrans(!m_trans)
+              .setReverseFlag(!m_reverse)
               .setLength(m_length)
               .setShift(m_shift);
     }
@@ -234,7 +234,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     ConjugateReturnType conjugate() const
     {
       return ConjugateReturnType(m_vectors.conjugate(), m_coeffs.conjugate())
-             .setTrans(m_trans)
+             .setReverseFlag(m_reverse)
              .setLength(m_length)
              .setShift(m_shift);
     }
@@ -243,7 +243,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     AdjointReturnType adjoint() const
     {
       return AdjointReturnType(m_vectors, m_coeffs.conjugate())
-              .setTrans(!m_trans)
+              .setReverseFlag(!m_reverse)
               .setLength(m_length)
               .setShift(m_shift);
     }
@@ -273,7 +273,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
         for(Index k = vecs-1; k >= 0; --k)
         {
           Index cornerSize = rows() - k - m_shift;
-          if(m_trans)
+          if(m_reverse)
             dst.bottomRightCorner(cornerSize, cornerSize)
                .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), workspace.data());
           else
@@ -293,7 +293,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
         for(Index k = vecs-1; k >= 0; --k)
         {
           Index cornerSize = rows() - k - m_shift;
-          if(m_trans)
+          if(m_reverse)
             dst.bottomRightCorner(cornerSize, cornerSize)
                .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), workspace.data());
           else
@@ -317,7 +317,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       workspace.resize(dst.rows());
       for(Index k = 0; k < m_length; ++k)
       {
-        Index actual_k = m_trans ? m_length-k-1 : k;
+        Index actual_k = m_reverse ? m_length-k-1 : k;
         dst.rightCols(rows()-m_shift-actual_k)
            .applyHouseholderOnTheRight(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data());
       }
@@ -340,8 +340,8 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       {
         for(Index i = 0; i < m_length; i+=BlockSize)
         {
-          Index end = m_trans ? (std::min)(m_length,i+BlockSize) : m_length-i;
-          Index k = m_trans ? i : (std::max)(Index(0),end-BlockSize);
+          Index end = m_reverse ? (std::min)(m_length,i+BlockSize) : m_length-i;
+          Index k = m_reverse ? i : (std::max)(Index(0),end-BlockSize);
           Index bs = end-k;
           Index start = k + m_shift;
           
@@ -352,7 +352,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
                                                                    Side==OnTheRight ? m_vectors.cols()-start : bs);
           typename internal::conditional<Side==OnTheRight, Transpose<SubVectorsType>, SubVectorsType&>::type sub_vecs(sub_vecs1);
           Block<Dest,Dynamic,Dynamic> sub_dst(dst,dst.rows()-rows()+m_shift+k,0, rows()-m_shift-k,dst.cols());
-          apply_block_householder_on_the_left(sub_dst, sub_vecs, m_coeffs.segment(k, bs), !m_trans);
+          apply_block_householder_on_the_left(sub_dst, sub_vecs, m_coeffs.segment(k, bs), !m_reverse);
         }
       }
       else
@@ -360,7 +360,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
         workspace.resize(dst.cols());
         for(Index k = 0; k < m_length; ++k)
         {
-          Index actual_k = m_trans ? k : m_length-k-1;
+          Index actual_k = m_reverse ? k : m_length-k-1;
           dst.bottomRows(rows()-m_shift-actual_k)
             .applyHouseholderOnTheLeft(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data());
         }
@@ -425,25 +425,27 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
 
   protected:
 
-    /** \brief Sets the transpose flag.
-      * \param [in]  trans  New value of the transpose flag.
+    /** \internal
+      * \brief Sets the reverse flag.
+      * \param [in]  reverse  New value of the reverse flag.
       *
-      * By default, the transpose flag is not set. If the transpose flag is set, then this object represents 
-      * \f$ H^T = H_{n-1}^T \ldots H_1^T H_0^T \f$ instead of \f$ H = H_0 H_1 \ldots H_{n-1} \f$.
+      * By default, the reverse flag is not set. If the reverse flag is set, then this object represents
+      * \f$ H^r = H_{n-1} \ldots H_1 H_0 \f$ instead of \f$ H = H_0 H_1 \ldots H_{n-1} \f$.
+      * \note For real valued HouseholderSequence this is equivalent to transposing \f$ H \f$.
       *
-      * \sa trans()
+      * \sa reverseFlag(), transpose(), adjoint()
       */
-    HouseholderSequence& setTrans(bool trans)
+    HouseholderSequence& setReverseFlag(bool reverse)
     {
-      m_trans = trans;
+      m_reverse = reverse;
       return *this;
     }
 
-    bool trans() const { return m_trans; }     /**< \brief Returns the transpose flag. */
+    bool reverseFlag() const { return m_reverse; }     /**< \internal \brief Returns the reverse flag. */
 
     typename VectorsType::Nested m_vectors;
     typename CoeffsType::Nested m_coeffs;
-    bool m_trans;
+    bool m_reverse;
     Index m_length;
     Index m_shift;
 };

From 02d2f1cb4ad6fdcc38d691c31a4b0de043e17654 Mon Sep 17 00:00:00 2001
From: Dmitriy Korchemkin <dmitriy.korchemkin@gmail.com>
Date: Wed, 18 Apr 2018 13:52:46 +0300
Subject: [PATCH 313/680] Cast zeros to Scalar in RealSchur

---
 Eigen/src/Eigenvalues/RealSchur.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h
index 9e71f3040..b72799e5b 100644
--- a/Eigen/src/Eigenvalues/RealSchur.h
+++ b/Eigen/src/Eigenvalues/RealSchur.h
@@ -303,7 +303,7 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa
   Scalar exshift(0);   // sum of exceptional shifts
   Scalar norm = computeNormOfT();
 
-  if(norm!=0)
+  if(norm!=Scalar(0))
   {
     while (iu >= 0)
     {
@@ -327,7 +327,7 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa
       else // No convergence yet
       {
         // The firstHouseholderVector vector has to be initialized to something to get rid of a silly GCC warning (-O1 -Wall -DNDEBUG )
-        Vector3s firstHouseholderVector(0,0,0), shiftInfo;
+        Vector3s firstHouseholderVector = Vector3s::Zero(), shiftInfo;
         computeShift(iu, iter, exshift, shiftInfo);
         iter = iter + 1;
         totalIter = totalIter + 1;

From 686fb57233736583c7ff4462aac403591ba762b9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Apr 2018 18:46:34 +0200
Subject: [PATCH 314/680] fix const cast in NEON

---
 Eigen/src/Core/arch/NEON/Complex.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h
index ef50ba303..110371ba8 100644
--- a/Eigen/src/Core/arch/NEON/Complex.h
+++ b/Eigen/src/Core/arch/NEON/Complex.h
@@ -67,7 +67,7 @@ template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type;
 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
 {
   float32x2_t r64;
-  r64 = vld1_f32((float *)&from);
+  r64 = vld1_f32((const float *)&from);
 
   return Packet2cf(vcombine_f32(r64, r64));
 }
@@ -142,7 +142,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf
   to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3));
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { EIGEN_ARM_PREFETCH((float *)addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { EIGEN_ARM_PREFETCH((const float *)addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
 {
@@ -383,7 +383,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ARM_PREFETCH((double *)addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ARM_PREFETCH((const double *)addr); }
 
 template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
 {

From b7b868d1c43b1ac0c7367fedb338dfd726a69ee3 Mon Sep 17 00:00:00 2001
From: Jayaram Bobba <jayaram.bobba@intel.com>
Date: Fri, 20 Apr 2018 13:39:18 -0700
Subject: [PATCH 315/680] fix AVX512 plog

---
 Eigen/src/Core/arch/AVX512/MathFunctions.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 61434a33e..9c1717f76 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -88,9 +88,9 @@ plog<Packet16f>(const Packet16f& _x) {
   //     x = x + x - 1.0;
   //   } else { x = x - 1.0; }
   __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ);
-  Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps());
+  Packet16f tmp = _mm512_mask_blend_ps(mask, _mm512_setzero_ps(), x);
   x = psub(x, p16f_1);
-  e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps()));
+  e = psub(e, _mm512_mask_blend_ps(mask, _mm512_setzero_ps(), p16f_1));
   x = padd(x, tmp);
 
   Packet16f x2 = pmul(x, x);
@@ -119,8 +119,9 @@ plog<Packet16f>(const Packet16f& _x) {
   x = padd(x, y2);
 
   // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
-  return _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf,
-                              _mm512_mask_blend_ps(invalid_mask, p16f_nan, x));
+  return _mm512_mask_blend_ps(iszero_mask,
+                              _mm512_mask_blend_ps(invalid_mask, x, p16f_nan),
+                              p16f_minus_inf);
 }
 #endif
 

From 34e499ad36363692615b800b5dbe6cd496341339 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 21 Apr 2018 22:08:26 +0200
Subject: [PATCH 316/680] Disable -Wshadow when compiling with g++

---
 CMakeLists.txt                               | 6 +-----
 Eigen/src/Core/util/DisableStupidWarnings.h  | 8 ++++++--
 Eigen/src/Core/util/ReenableStupidWarnings.h | 2 +-
 3 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8a068739b..5245b2930 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -153,11 +153,7 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-Wdouble-promotion")
 #  ei_add_cxx_compiler_flag("-Wconversion")
   
-  # -Wshadow is insanely too strict with gcc, hopefully it will become usable with gcc 6
-  # if(NOT CMAKE_COMPILER_IS_GNUCXX OR (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER "5.0.0"))
-  if(NOT CMAKE_COMPILER_IS_GNUCXX)
-    ei_add_cxx_compiler_flag("-Wshadow")
-  endif()
+  ei_add_cxx_compiler_flag("-Wshadow")
   
   ei_add_cxx_compiler_flag("-Wno-psabi")
   ei_add_cxx_compiler_flag("-Wno-variadic-macros")
diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index 8ef0f3594..af832ae77 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -45,12 +45,16 @@
     #pragma clang diagnostic ignored "-Wabsolute-value"
   #endif
 
-#elif defined __GNUC__ && __GNUC__>=6
+#elif defined __GNUC__
 
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma GCC diagnostic push
   #endif
-  #pragma GCC diagnostic ignored "-Wignored-attributes"
+  // g++ warns about local variables shadowing member functions, which is too strict
+  #pragma GCC diagnostic ignored "-Wshadow"
+  #if __GNUC__>=6
+    #pragma GCC diagnostic ignored "-Wignored-attributes"
+  #endif
 
 #endif
 
diff --git a/Eigen/src/Core/util/ReenableStupidWarnings.h b/Eigen/src/Core/util/ReenableStupidWarnings.h
index 86b60f52f..e9353a623 100644
--- a/Eigen/src/Core/util/ReenableStupidWarnings.h
+++ b/Eigen/src/Core/util/ReenableStupidWarnings.h
@@ -8,7 +8,7 @@
     #pragma warning pop
   #elif defined __clang__
     #pragma clang diagnostic pop
-  #elif defined __GNUC__ && __GNUC__>=6
+  #elif defined __GNUC__
     #pragma GCC diagnostic pop
   #endif
 

From 35b31353abebbbd823457c58d00d85c9f65e8a30 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 22 Apr 2018 22:49:08 +0200
Subject: [PATCH 317/680] Fix unit test

---
 test/permutationmatrices.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/permutationmatrices.cpp b/test/permutationmatrices.cpp
index 0ed7a4b2b..b2229cfe9 100644
--- a/test/permutationmatrices.cpp
+++ b/test/permutationmatrices.cpp
@@ -124,7 +124,7 @@ template<typename MatrixType> void permutationmatrices(const MatrixType& m)
   lp = lt;
   rp = rt;
   VERIFY_EVALUATION_COUNT(m_permuted = lt * m_permuted * rt, 1);
-  VERIFY_IS_APPROX(m_permuted, lp*m_original*rp);
+  VERIFY_IS_APPROX(m_permuted, lp*m_original*rp.transpose());
   
   VERIFY_IS_APPROX(lt.inverse()*m_permuted*rt.inverse(), m_original);
 }

From 5679e439e0c87e4b93ad7870f463d5168edcd57d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Apr 2018 14:40:16 +0200
Subject: [PATCH 318/680] bug #1543: fix linear indexing in generic block
 evaluation (this completes the fix in commit
 12efc7d41b80259b996be5781bf596c249c90d3f )

---
 Eigen/src/Core/CoreEvaluators.h | 11 ++++++-----
 test/diagonalmatrices.cpp       |  6 ++++++
 2 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 8419798c1..6231f2edd 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1087,7 +1087,8 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
   enum {
-    RowsAtCompileTime = XprType::RowsAtCompileTime
+    RowsAtCompileTime = XprType::RowsAtCompileTime,
+    ForwardLinearAccess = InnerPanel && bool(evaluator<ArgType>::Flags&LinearAccessBit)
   };
  
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -1099,7 +1100,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   CoeffReturnType coeff(Index index) const
   { 
-    if (InnerPanel)
+    if (ForwardLinearAccess)
       return m_argImpl.coeff(m_linear_offset.value() + index); 
     else
       return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
@@ -1114,7 +1115,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   Scalar& coeffRef(Index index)
   { 
-    if (InnerPanel)
+    if (ForwardLinearAccess)
       return m_argImpl.coeffRef(m_linear_offset.value() + index); 
     else
       return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
@@ -1131,7 +1132,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_STRONG_INLINE
   PacketType packet(Index index) const 
   { 
-    if (InnerPanel)
+    if (ForwardLinearAccess)
       return m_argImpl.template packet<LoadMode,PacketType>(m_linear_offset.value() + index);
     else
       return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
@@ -1149,7 +1150,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketType& x) 
   {
-    if (InnerPanel)
+    if (ForwardLinearAccess)
       return m_argImpl.template writePacket<StoreMode,PacketType>(m_linear_offset.value() + index, x);
     else
       return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp
index cd6dc8cf0..a2092c43e 100644
--- a/test/diagonalmatrices.cpp
+++ b/test/diagonalmatrices.cpp
@@ -99,6 +99,12 @@ template<typename MatrixType> void diagonalmatrices(const MatrixType& m)
   VERIFY_IS_APPROX( (sq_m1 += (s1*v1).asDiagonal()), sq_m2 += (s1*v1).asDiagonal().toDenseMatrix() );
   VERIFY_IS_APPROX( (sq_m1 -= (s1*v1).asDiagonal()), sq_m2 -= (s1*v1).asDiagonal().toDenseMatrix() );
   VERIFY_IS_APPROX( (sq_m1 = (s1*v1).asDiagonal()), (s1*v1).asDiagonal().toDenseMatrix() );
+
+  sq_m1.setRandom();
+  sq_m2 = v1.asDiagonal();
+  sq_m2 = sq_m1 * sq_m2;
+  VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).col(i), sq_m2.col(i) );
+  VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).row(i), sq_m2.row(i) );
 }
 
 template<int>

From 11123175dbe49a9b17335f67f093ef04a2779d49 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Apr 2018 15:30:35 +0200
Subject: [PATCH 319/680] typo in doc

---
 Eigen/src/Core/VectorBlock.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/VectorBlock.h b/Eigen/src/Core/VectorBlock.h
index d72fbf7e9..0ede5d58e 100644
--- a/Eigen/src/Core/VectorBlock.h
+++ b/Eigen/src/Core/VectorBlock.h
@@ -35,7 +35,7 @@ struct traits<VectorBlock<VectorType, Size> >
   * It is the return type of DenseBase::segment(Index,Index) and DenseBase::segment<int>(Index) and
   * most of the time this is the only way it is used.
   *
-  * However, if you want to directly maniputate sub-vector expressions,
+  * However, if you want to directly manipulate sub-vector expressions,
   * for instance if you want to write a function returning such an expression, you
   * will need to use this class.
   *

From a57e6e5f0fa8ca5cb621bb07a2a72cbbdf98e300 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Apr 2018 15:31:51 +0200
Subject: [PATCH 320/680] workaround MSVC 2013 compilation issue (ambiguous
 call)

---
 Eigen/src/SparseCore/SparseSelfAdjointView.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h
index 5ab64f1a8..65611b3d4 100644
--- a/Eigen/src/SparseCore/SparseSelfAdjointView.h
+++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h
@@ -311,7 +311,7 @@ inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, cons
         while (i && i.index()<j) ++i;
         if(i && i.index()==j)
         {
-          res(j,k) += alpha * i.value() * rhs(j,k);
+          res.coeffRef(j,k) += alpha * i.value() * rhs.coeff(j,k);
           ++i;
         }
       }
@@ -324,14 +324,14 @@ inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, cons
       {
         LhsScalar lhs_ij = i.value();
         if(!LhsIsRowMajor) lhs_ij = numext::conj(lhs_ij);
-        res_j += lhs_ij * rhs(i.index(),k);
+        res_j += lhs_ij * rhs.coeff(i.index(),k);
         res(i.index(),k) += numext::conj(lhs_ij) * rhs_j;
       }
-      res(j,k) += alpha * res_j;
+      res.coeffRef(j,k) += alpha * res_j;
 
       // handle diagonal coeff
       if (ProcessFirstHalf && i && (i.index()==j))
-        res(j,k) += alpha * i.value() * rhs(j,k);
+        res.coeffRef(j,k) += alpha * i.value() * rhs.coeff(j,k);
     }
   }
 }

From e9da464e202ead443e2429e5ed1a74631e2f6832 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Apr 2018 16:26:29 +0200
Subject: [PATCH 321/680] Add specializations of is_arithmetic for long long in
 c++11

---
 Eigen/src/Core/util/Meta.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 1180f0c13..0c7b8e709 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -98,6 +98,8 @@ template<> struct is_arithmetic<signed long>   { enum { value = true }; };
 template<> struct is_arithmetic<unsigned long> { enum { value = true }; };
 
 #if EIGEN_HAS_CXX11
+template<> struct is_arithmetic<signed long long>   { enum { value = true }; };
+template<> struct is_arithmetic<unsigned long long> { enum { value = true }; };
 using std::is_integral;
 #else
 template<typename T> struct is_integral               { enum { value = false }; };

From e8ca5166a921af2ed7df1ddda12fbc9316b2b2cf Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 24 Apr 2018 11:19:49 +0200
Subject: [PATCH 322/680] bug #1428: atempt to make NEON vectorization
 compilable by MSVC. The workaround is to wrap NEON packet types to make them
 different c++ types.

---
 Eigen/src/Core/arch/NEON/PacketMath.h | 31 +++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)

diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 6283b6cfa..010739380 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -36,12 +36,43 @@ namespace internal {
 #endif
 #endif
 
+#if EIGEN_COMP_MSVC
+
+// In MSVC's arm_neon.h header file, all NEON vector types
+// are aliases to the same underlying type __n128.
+// We thus have to wrap them to make them different C++ types.
+// (See also bug 1428)
+
+template<typename T,int unique_id>
+struct eigen_packet_wrapper
+{
+  operator T&() { return m_val; }
+  operator const T&() const { return m_val; }
+  eigen_packet_wrapper() {}
+  eigen_packet_wrapper(const T &v) : m_val(v) {}
+  eigen_packet_wrapper& operator=(const T &v) {
+    m_val = v;
+    return *this;
+  }
+
+  T m_val;
+};
+typedef eigen_packet_wrapper<float32x2_t,0> Packet2f;
+typedef eigen_packet_wrapper<float32x4_t,1> Packet4f;
+typedef eigen_packet_wrapper<int32x4_t  ,2> Packet4i;
+typedef eigen_packet_wrapper<int32x2_t  ,3> Packet2i;
+typedef eigen_packet_wrapper<uint32x4_t ,4> Packet4ui;
+
+#else
+
 typedef float32x2_t Packet2f;
 typedef float32x4_t Packet4f;
 typedef int32x4_t   Packet4i;
 typedef int32x2_t   Packet2i;
 typedef uint32x4_t  Packet4ui;
 
+#endif // EIGEN_COMP_MSVC
+
 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
   const Packet4f p4f_##NAME = pset1<Packet4f>(X)
 

From 3ffd449ef5f3e2e3d08dc3b683be3a2f291b6e70 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 24 Apr 2018 17:11:51 +0200
Subject: [PATCH 323/680] Workaround warning

---
 unsupported/test/autodiff_scalar.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp
index 9cf11280c..a917ec344 100644
--- a/unsupported/test/autodiff_scalar.cpp
+++ b/unsupported/test/autodiff_scalar.cpp
@@ -81,6 +81,9 @@ void check_limits_specialization()
   typedef std::numeric_limits<AD> A;
   typedef std::numeric_limits<Scalar> B;
 
+  // workaround "unsed typedef" warning:
+  VERIFY(!bool(internal::is_same<B, A>::value));
+
 #if EIGEN_HAS_CXX11
   VERIFY(bool(std::is_base_of<B, A>::value));
 #endif

From 2f3287da7df977e1e5faae40cf0276e83369da97 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 24 Apr 2018 17:17:25 +0200
Subject: [PATCH 324/680] Fix "used uninitialized" warnings

---
 unsupported/test/autodiff.cpp | 22 +++++++++++++---------
 1 file changed, 13 insertions(+), 9 deletions(-)

diff --git a/unsupported/test/autodiff.cpp b/unsupported/test/autodiff.cpp
index 85743137e..1c5e0dc66 100644
--- a/unsupported/test/autodiff.cpp
+++ b/unsupported/test/autodiff.cpp
@@ -306,6 +306,8 @@ double bug_1222() {
   return denom.value();
 }
 
+#ifdef EIGEN_TEST_PART_5
+
 double bug_1223() {
   using std::min;
   typedef Eigen::AutoDiffScalar<Eigen::Vector3d> AD;
@@ -326,8 +328,8 @@ double bug_1223() {
 
 // regression test for some compilation issues with specializations of ScalarBinaryOpTraits
 void bug_1260() {
-  Matrix4d A;
-  Vector4d v;
+  Matrix4d A = Matrix4d::Ones();
+  Vector4d v = Vector4d::Ones();
   A*v;
 }
 
@@ -336,7 +338,7 @@ double bug_1261() {
   typedef AutoDiffScalar<Matrix2d> AD;
   typedef Matrix<AD,2,1> VectorAD;
 
-  VectorAD v;
+  VectorAD v(0.,0.);
   const AD maxVal = v.maxCoeff();
   const AD minVal = v.minCoeff();
   return maxVal.value() + minVal.value();
@@ -344,12 +346,14 @@ double bug_1261() {
 
 double bug_1264() {
   typedef AutoDiffScalar<Vector2d> AD;
-  const AD s;
-  const Matrix<AD, 3, 1> v1;
+  const AD s = 0.;
+  const Matrix<AD, 3, 1> v1(0.,0.,0.);
   const Matrix<AD, 3, 1> v2 = (s + 3.0) * v1;
   return v2(0).value();
 }
 
+#endif
+
 void test_autodiff()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -359,9 +363,9 @@ void test_autodiff()
     CALL_SUBTEST_4( test_autodiff_hessian<1>() );
   }
 
-  bug_1222();
-  bug_1223();
-  bug_1260();
-  bug_1261();
+  CALL_SUBTEST_5( bug_1222() );
+  CALL_SUBTEST_5( bug_1223() );
+  CALL_SUBTEST_5( bug_1260() );
+  CALL_SUBTEST_5( bug_1261() );
 }
 

From 8810baaed40f554ff59db65d5f5025ef5247c3c2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 25 Apr 2018 10:14:48 +0200
Subject: [PATCH 325/680] Add multi-threading for sparse-row-major *
 dense-row-major

---
 Eigen/src/SparseCore/SparseDenseProduct.h | 38 ++++++++++++++++++-----
 1 file changed, 30 insertions(+), 8 deletions(-)

diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h
index 0547db596..f005a18a1 100644
--- a/Eigen/src/SparseCore/SparseDenseProduct.h
+++ b/Eigen/src/SparseCore/SparseDenseProduct.h
@@ -88,10 +88,11 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, A
   typedef typename internal::remove_all<SparseLhsType>::type Lhs;
   typedef typename internal::remove_all<DenseRhsType>::type Rhs;
   typedef typename internal::remove_all<DenseResType>::type Res;
-  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename LhsEval::InnerIterator LhsInnerIterator;
   static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
   {
-    evaluator<Lhs> lhsEval(lhs);
+    LhsEval lhsEval(lhs);
     for(Index c=0; c<rhs.cols(); ++c)
     {
       for(Index j=0; j<lhs.outerSize(); ++j)
@@ -111,16 +112,37 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, t
   typedef typename internal::remove_all<SparseLhsType>::type Lhs;
   typedef typename internal::remove_all<DenseRhsType>::type Rhs;
   typedef typename internal::remove_all<DenseResType>::type Res;
-  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename LhsEval::InnerIterator LhsInnerIterator;
   static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
   {
-    evaluator<Lhs> lhsEval(lhs);
-    for(Index j=0; j<lhs.outerSize(); ++j)
+    Index n = lhs.rows();
+    LhsEval lhsEval(lhs);
+
+#ifdef EIGEN_HAS_OPENMP
+    Eigen::initParallel();
+    Index threads = Eigen::nbThreads();
+    // This 20000 threshold has been found experimentally on 2D and 3D Poisson problems.
+    // It basically represents the minimal amount of work to be done to be worth it.
+    if(threads>1 && lhsEval.nonZerosEstimate()*rhs.cols() > 20000)
     {
-      typename Res::RowXpr res_j(res.row(j));
-      for(LhsInnerIterator it(lhsEval,j); it ;++it)
-        res_j += (alpha*it.value()) * rhs.row(it.index());
+      #pragma omp parallel for schedule(dynamic,(n+threads*4-1)/(threads*4)) num_threads(threads)
+      for(Index i=0; i<n; ++i)
+        processRow(lhsEval,rhs,res,alpha,i);
     }
+    else
+#endif
+    {
+      for(Index i=0; i<n; ++i)
+        processRow(lhsEval, rhs, res, alpha, i);
+    }
+  }
+
+  static void processRow(const LhsEval& lhsEval, const DenseRhsType& rhs, Res& res, const typename Res::Scalar& alpha, Index i)
+  {
+    typename Res::RowXpr res_i(res.row(i));
+    for(LhsInnerIterator it(lhsEval,i); it ;++it)
+      res_i += (alpha*it.value()) * rhs.row(it.index());
   }
 };
 

From 097dd4616d7d1c8b41b9026b9fec16faf57f4df1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 26 Apr 2018 10:47:39 +0200
Subject: [PATCH 326/680] Fix unit test for SIMD engine not supporting sqrt

---
 test/packetmath.cpp | 56 +++++++++++++++++++++++++--------------------
 1 file changed, 31 insertions(+), 25 deletions(-)

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 53838920e..4b19be92d 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -451,35 +451,41 @@ template<typename Scalar> void packetmath_real()
   {
     data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
     data1[1] = std::numeric_limits<Scalar>::epsilon();
-    packet_helper<PacketTraits::HasLog,Packet> h;
-    h.store(data2, internal::plog(h.load(data1)));
-    VERIFY((numext::isnan)(data2[0]));
-    VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::epsilon()), data2[1]);
+    {
+      packet_helper<PacketTraits::HasLog,Packet> h;
+      h.store(data2, internal::plog(h.load(data1)));
+      VERIFY((numext::isnan)(data2[0]));
+      VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::epsilon()), data2[1]);
 
-    data1[0] = -std::numeric_limits<Scalar>::epsilon();
-    data1[1] = 0;
-    h.store(data2, internal::plog(h.load(data1)));
-    VERIFY((numext::isnan)(data2[0]));
-    VERIFY_IS_EQUAL(std::log(Scalar(0)), data2[1]);
+      data1[0] = -std::numeric_limits<Scalar>::epsilon();
+      data1[1] = 0;
+      h.store(data2, internal::plog(h.load(data1)));
+      VERIFY((numext::isnan)(data2[0]));
+      VERIFY_IS_EQUAL(std::log(Scalar(0)), data2[1]);
 
-    data1[0] = (std::numeric_limits<Scalar>::min)();
-    data1[1] = -(std::numeric_limits<Scalar>::min)();
-    h.store(data2, internal::plog(h.load(data1)));
-    VERIFY_IS_EQUAL(std::log((std::numeric_limits<Scalar>::min)()), data2[0]);
-    VERIFY((numext::isnan)(data2[1]));
+      data1[0] = (std::numeric_limits<Scalar>::min)();
+      data1[1] = -(std::numeric_limits<Scalar>::min)();
+      h.store(data2, internal::plog(h.load(data1)));
+      VERIFY_IS_EQUAL(std::log((std::numeric_limits<Scalar>::min)()), data2[0]);
+      VERIFY((numext::isnan)(data2[1]));
 
-    data1[0] = std::numeric_limits<Scalar>::denorm_min();
-    data1[1] = -std::numeric_limits<Scalar>::denorm_min();
-    h.store(data2, internal::plog(h.load(data1)));
-    // VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
-    VERIFY((numext::isnan)(data2[1]));
+      data1[0] = std::numeric_limits<Scalar>::denorm_min();
+      data1[1] = -std::numeric_limits<Scalar>::denorm_min();
+      h.store(data2, internal::plog(h.load(data1)));
+      // VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
+      VERIFY((numext::isnan)(data2[1]));
 
-    data1[0] = Scalar(-1.0f);
-    h.store(data2, internal::plog(h.load(data1)));
-    VERIFY((numext::isnan)(data2[0]));
-    h.store(data2, internal::psqrt(h.load(data1)));
-    VERIFY((numext::isnan)(data2[0]));
-    VERIFY((numext::isnan)(data2[1]));
+      data1[0] = Scalar(-1.0f);
+      h.store(data2, internal::plog(h.load(data1)));
+      VERIFY((numext::isnan)(data2[0]));
+    }
+    {
+      packet_helper<PacketTraits::HasSqrt,Packet> h;
+      data1[0] = Scalar(-1.0f);
+      h.store(data2, internal::psqrt(h.load(data1)));
+      VERIFY((numext::isnan)(data2[0]));
+      VERIFY((numext::isnan)(data2[1]));
+    }
   }
 }
 

From 6e7118265dc47e07107859c0cdba7eb37e2200e7 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 26 Apr 2018 10:50:41 +0200
Subject: [PATCH 327/680] Fix compilation with NEON+MSVC

---
 Eigen/src/Core/arch/NEON/Complex.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h
index 110371ba8..306a309be 100644
--- a/Eigen/src/Core/arch/NEON/Complex.h
+++ b/Eigen/src/Core/arch/NEON/Complex.h
@@ -277,7 +277,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
   s = vmulq_f32(b.v, b.v);
   rev_s = vrev64q_f32(s);
 
-  return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s)));
+  return Packet2cf(pdiv<Packet4f>(res.v, vaddq_f32(s,rev_s)));
 }
 
 EIGEN_DEVICE_FUNC inline void

From 6118c6ff4f898dad99564ffdeb99036beb2ff0ea Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 30 Apr 2018 11:28:12 -0700
Subject: [PATCH 328/680] Enable RawAccess to tensor slices whenever possinle.
 Avoid 32-bit integer overflow in TensorSlicingOp

---
 unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 329655817..e59074506 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -398,7 +398,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
       const MemcpyTriggerForSlicing<Index, Device> trigger(m_device);
       if (trigger(contiguous_values)) {
         Scalar* src = (Scalar*)m_impl.data();
-        for (int i = 0; i < internal::array_prod(dimensions()); i += contiguous_values) {
+        for (Index i = 0; i < internal::array_prod(dimensions()); i += contiguous_values) {
           Index offset = srcCoeff(i);
           m_device.memcpy((void*)(data+i), src+offset, contiguous_values * sizeof(Scalar));
         }
@@ -559,7 +559,7 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
-    RawAccess = false
+    RawAccess = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)

From b8c8e5f436743ac6a6a5ed0ad7ec5cce7dd00248 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 14 May 2018 16:07:13 -0700
Subject: [PATCH 329/680] Add vectorized clip functor for Eigen Tensors.

---
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  6 ++++++
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   | 19 +++++++++++++++++
 unsupported/test/cxx11_tensor_expr.cpp        | 21 +++++++++++++++++++
 3 files changed, 46 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 1d459a3a0..c8dc3bbc5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -209,6 +209,12 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return unaryExpr(internal::scalar_abs_op<Scalar>());
     }
 
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clip_op<Scalar>, const Derived>
+    clip(Scalar min, Scalar max) const {
+      return unaryExpr(internal::scalar_clip_op<Scalar>(min, max));
+    }
+
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, const Derived>
     conjugate() const {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 5dcc3794c..4cee38339 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -487,6 +487,25 @@ struct functor_traits<GaussianGenerator<T, Index, NumDims> > {
   };
 };
 
+template <typename Scalar>
+struct scalar_clip_op {
+  EIGEN_DEVICE_FUNC inline scalar_clip_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar
+  operator()(const Scalar& x) const {
+    return numext::mini(numext::maxi(x, m_min), m_max);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet
+  packetOp(const Packet& x) const {
+    return internal::pmin(internal::pmax(x, pset1<Packet>(m_min)), pset1<Packet>(m_max));
+  }
+  const Scalar m_min;
+  const Scalar m_max;
+};
+template<typename Scalar>
+struct functor_traits<scalar_clip_op<Scalar> >
+{ enum { Cost = 2 * NumTraits<Scalar>::AddCost, PacketAccess = (packet_traits<Scalar>::HasMin && packet_traits<Scalar>::HasMax)}; };
+
 } // end namespace internal
 } // end namespace Eigen
 
diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp
index 129b4e659..30e5a9860 100644
--- a/unsupported/test/cxx11_tensor_expr.cpp
+++ b/unsupported/test/cxx11_tensor_expr.cpp
@@ -340,6 +340,26 @@ void test_minmax_nan_propagation_templ() {
   }
 }
 
+static void test_clip()
+{
+  Tensor<float, 1> vec(6);
+  vec(0) = 4.0;
+  vec(1) = 8.0;
+  vec(2) = 15.0;
+  vec(3) = 16.0;
+  vec(4) = 23.0;
+  vec(5) = 42.0;
+
+  float kMin = 20;
+  float kMax = 30;
+
+  Tensor<float, 1> vec_clipped(6);
+  vec_clipped = vec.clip(kMin, kMax);
+  for (int i = 0; i < 6; ++i) {
+    VERIFY_IS_EQUAL(vec_clipped(i), std::min(std::max(vec(i), kMin), kMax));
+  }
+}
+
 static void test_minmax_nan_propagation()
 {
   test_minmax_nan_propagation_templ<float>();
@@ -356,5 +376,6 @@ void test_cxx11_tensor_expr()
   CALL_SUBTEST(test_functors());
   CALL_SUBTEST(test_type_casting());
   CALL_SUBTEST(test_select());
+  CALL_SUBTEST(test_clip());
   CALL_SUBTEST(test_minmax_nan_propagation());
 }

From afec3021f731d053254fe64983fc0623fff507f9 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 14 May 2018 16:35:39 -0700
Subject: [PATCH 330/680] Use numext::maxi & numext::mini.

---
 unsupported/test/cxx11_tensor_expr.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp
index 30e5a9860..d64301318 100644
--- a/unsupported/test/cxx11_tensor_expr.cpp
+++ b/unsupported/test/cxx11_tensor_expr.cpp
@@ -356,7 +356,7 @@ static void test_clip()
   Tensor<float, 1> vec_clipped(6);
   vec_clipped = vec.clip(kMin, kMax);
   for (int i = 0; i < 6; ++i) {
-    VERIFY_IS_EQUAL(vec_clipped(i), std::min(std::max(vec(i), kMin), kMax));
+    VERIFY_IS_EQUAL(vec_clipped(i), numext::mini(numext::maxi(vec(i), kMin), kMax));
   }
 }
 

From 0272f2451a1c85e0d74235ac8f5c468776171580 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 15 May 2018 19:35:53 +0200
Subject: [PATCH 331/680] Fix "suggest parentheses around comparison" warning

---
 Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
index 9176a1382..49565c070 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
@@ -37,7 +37,7 @@ namespace Eigen {
 
 namespace internal {
 
-template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int  UpLo>
+template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int UpLo>
 struct general_matrix_matrix_rankupdate :
        general_matrix_matrix_triangular_product<
          Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,UpLo,BuiltIn> {};
@@ -52,7 +52,7 @@ struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,Con
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \
                           const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \
   { \
-    if ( lhs==rhs && ((UpLo&(Lower|Upper)==UpLo)) ) { \
+    if ( lhs==rhs && ((UpLo&(Lower|Upper))==UpLo) ) { \
       general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
       ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
     } else { \

From 812480baa340f2d8b0bc21b3f2d227168280ad06 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 16 May 2018 09:49:24 -0700
Subject: [PATCH 332/680] Rename scalar_clip_op to scalar_clip2_op to prevent
 collision with existing functor in TensorFlow.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h     | 4 ++--
 unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h | 6 +++---
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index c8dc3bbc5..497e2a797 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -210,9 +210,9 @@ class TensorBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clip_op<Scalar>, const Derived>
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clip2_op<Scalar>, const Derived>
     clip(Scalar min, Scalar max) const {
-      return unaryExpr(internal::scalar_clip_op<Scalar>(min, max));
+      return unaryExpr(internal::scalar_clip2_op<Scalar>(min, max));
     }
 
     EIGEN_DEVICE_FUNC
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 4cee38339..adadb5afc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -488,8 +488,8 @@ struct functor_traits<GaussianGenerator<T, Index, NumDims> > {
 };
 
 template <typename Scalar>
-struct scalar_clip_op {
-  EIGEN_DEVICE_FUNC inline scalar_clip_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
+struct scalar_clip2_op {
+  EIGEN_DEVICE_FUNC inline scalar_clip2_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar
   operator()(const Scalar& x) const {
     return numext::mini(numext::maxi(x, m_min), m_max);
@@ -503,7 +503,7 @@ struct scalar_clip_op {
   const Scalar m_max;
 };
 template<typename Scalar>
-struct functor_traits<scalar_clip_op<Scalar> >
+struct functor_traits<scalar_clip2_op<Scalar> >
 { enum { Cost = 2 * NumTraits<Scalar>::AddCost, PacketAccess = (packet_traits<Scalar>::HasMin && packet_traits<Scalar>::HasMax)}; };
 
 } // end namespace internal

From b8d36774faafa89ffc016c2c35f7e906035b2fa5 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 16 May 2018 14:04:48 -0700
Subject: [PATCH 333/680] Rename clip2 to clamp.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h     | 4 ++--
 unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h | 6 +++---
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 497e2a797..745b16d2c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -210,9 +210,9 @@ class TensorBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clip2_op<Scalar>, const Derived>
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clamp_op<Scalar>, const Derived>
     clip(Scalar min, Scalar max) const {
-      return unaryExpr(internal::scalar_clip2_op<Scalar>(min, max));
+      return unaryExpr(internal::scalar_clamp_op<Scalar>(min, max));
     }
 
     EIGEN_DEVICE_FUNC
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index adadb5afc..0efcd6253 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -488,8 +488,8 @@ struct functor_traits<GaussianGenerator<T, Index, NumDims> > {
 };
 
 template <typename Scalar>
-struct scalar_clip2_op {
-  EIGEN_DEVICE_FUNC inline scalar_clip2_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
+struct scalar_clamp_op {
+  EIGEN_DEVICE_FUNC inline scalar_clamp_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar
   operator()(const Scalar& x) const {
     return numext::mini(numext::maxi(x, m_min), m_max);
@@ -503,7 +503,7 @@ struct scalar_clip2_op {
   const Scalar m_max;
 };
 template<typename Scalar>
-struct functor_traits<scalar_clip2_op<Scalar> >
+struct functor_traits<scalar_clamp_op<Scalar> >
 { enum { Cost = 2 * NumTraits<Scalar>::AddCost, PacketAccess = (packet_traits<Scalar>::HasMin && packet_traits<Scalar>::HasMax)}; };
 
 } // end namespace internal

From d6559009530b87f80fc5aaa864c012ae5109e848 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 17 May 2018 19:17:01 +0200
Subject: [PATCH 334/680] bug #1544: Generate correct Q matrix in complex case.
 Original patch was by Jeff Trull in PR-386.

---
 Eigen/src/SparseQR/SparseQR.h | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h
index f7111fe2e..278d775aa 100644
--- a/Eigen/src/SparseQR/SparseQR.h
+++ b/Eigen/src/SparseQR/SparseQR.h
@@ -52,7 +52,7 @@ namespace internal {
   * rank-revealing permutations. Use colsPermutation() to get it.
   * 
   * Q is the orthogonal matrix represented as products of Householder reflectors. 
-  * Use matrixQ() to get an expression and matrixQ().transpose() to get the transpose.
+  * Use matrixQ() to get an expression and matrixQ().adjoint() to get the adjoint.
   * You can then apply it to a vector.
   * 
   * R is the sparse triangular or trapezoidal matrix. The later occurs when A is rank-deficient.
@@ -65,6 +65,7 @@ namespace internal {
   * \implsparsesolverconcept
   *
   * \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()).
+  * \warning For complex matrices matrixQ().transpose() will actually return the adjoint matrix.
   * 
   */
 template<typename _MatrixType, typename _OrderingType>
@@ -196,9 +197,9 @@ class SparseQR : public SparseSolverBase<SparseQR<_MatrixType,_OrderingType> >
 
       Index rank = this->rank();
       
-      // Compute Q^T * b;
+      // Compute Q^* * b;
       typename Dest::PlainObject y, b;
-      y = this->matrixQ().transpose() * B; 
+      y = this->matrixQ().adjoint() * B;
       b = y;
       
       // Solve with the triangular matrix R
@@ -641,7 +642,7 @@ struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived
           Scalar tau = Scalar(0);
           tau = m_qr.m_Q.col(k).dot(res.col(j));
           if(tau==Scalar(0)) continue;
-          tau = tau * m_qr.m_hcoeffs(k);
+          tau = tau * numext::conj(m_qr.m_hcoeffs(k));
           res.col(j) -= tau * m_qr.m_Q.col(k);
         }
       }
@@ -650,7 +651,7 @@ struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived
   
   const SparseQRType& m_qr;
   const Derived& m_other;
-  bool m_transpose;
+  bool m_transpose; // TODO this actually means adjoint
 };
 
 template<typename SparseQRType>
@@ -668,13 +669,14 @@ struct SparseQRMatrixQReturnType : public EigenBase<SparseQRMatrixQReturnType<Sp
   {
     return SparseQR_QProduct<SparseQRType,Derived>(m_qr,other.derived(),false);
   }
+  // To use for operations with the adjoint of Q
   SparseQRMatrixQTransposeReturnType<SparseQRType> adjoint() const
   {
     return SparseQRMatrixQTransposeReturnType<SparseQRType>(m_qr);
   }
   inline Index rows() const { return m_qr.rows(); }
   inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); }
-  // To use for operations with the transpose of Q
+  // To use for operations with the transpose of Q FIXME this is the same as adjoint at the moment
   SparseQRMatrixQTransposeReturnType<SparseQRType> transpose() const
   {
     return SparseQRMatrixQTransposeReturnType<SparseQRType>(m_qr);
@@ -682,6 +684,7 @@ struct SparseQRMatrixQReturnType : public EigenBase<SparseQRMatrixQReturnType<Sp
   const SparseQRType& m_qr;
 };
 
+// TODO this actually represents the adjoint of Q
 template<typename SparseQRType>
 struct SparseQRMatrixQTransposeReturnType
 {

From 9f0c5c3669a985a0dfa44db2ab05306edf1130c1 Mon Sep 17 00:00:00 2001
From: Jeff Trull <edaskel@att.net>
Date: Thu, 15 Feb 2018 15:00:31 -0800
Subject: [PATCH 335/680] Make sparse QR result sizes consistent with dense QR,
 with the following rules:

1) Q is always square
2) Q*R*P' is valid and recovers the original matrix

This implies that the size of Q is the number of rows in the original matrix, square,
and that the size of R is the size of the original matrix.
---
 Eigen/src/SparseQR/SparseQR.h | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h
index 278d775aa..7409fcae9 100644
--- a/Eigen/src/SparseQR/SparseQR.h
+++ b/Eigen/src/SparseQR/SparseQR.h
@@ -328,7 +328,7 @@ void SparseQR<MatrixType,OrderingType>::analyzePattern(const MatrixType& mat)
   internal::coletree(matCpy, m_etree, m_firstRowElt, m_outputPerm_c.indices().data());
   m_isEtreeOk = true;
   
-  m_R.resize(diagSize, n);
+  m_R.resize(m, n);
   m_Q.resize(m, diagSize);
   
   // Allocate space for nonzero elements : rough estimation
@@ -605,7 +605,7 @@ struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived
   // Get the references 
   SparseQR_QProduct(const SparseQRType& qr, const Derived& other, bool transpose) : 
   m_qr(qr),m_other(other),m_transpose(transpose) {}
-  inline Index rows() const { return m_transpose ? m_qr.rows() : m_qr.cols(); }
+  inline Index rows() const { return m_qr.matrixQ().rows(); }
   inline Index cols() const { return m_other.cols(); }
   
   // Assign to a vector
@@ -633,7 +633,10 @@ struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived
     }
     else
     {
-      eigen_assert(m_qr.m_Q.rows() == m_other.rows() && "Non conforming object sizes");
+      eigen_assert(m_qr.matrixQ().cols() == m_other.rows() && "Non conforming object sizes");
+
+      res.conservativeResize(rows(), cols());
+
       // Compute res = Q * other column by column
       for(Index j = 0; j < res.cols(); j++)
       {
@@ -675,7 +678,7 @@ struct SparseQRMatrixQReturnType : public EigenBase<SparseQRMatrixQReturnType<Sp
     return SparseQRMatrixQTransposeReturnType<SparseQRType>(m_qr);
   }
   inline Index rows() const { return m_qr.rows(); }
-  inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); }
+  inline Index cols() const { return m_qr.rows(); }
   // To use for operations with the transpose of Q FIXME this is the same as adjoint at the moment
   SparseQRMatrixQTransposeReturnType<SparseQRType> transpose() const
   {
@@ -715,7 +718,7 @@ struct Assignment<DstXprType, SparseQRMatrixQReturnType<SparseQRType>, internal:
   typedef typename DstXprType::StorageIndex StorageIndex;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &/*func*/)
   {
-    typename DstXprType::PlainObject idMat(src.m_qr.rows(), src.m_qr.rows());
+    typename DstXprType::PlainObject idMat(src.rows(), src.cols());
     idMat.setIdentity();
     // Sort the sparse householder reflectors if needed
     const_cast<SparseQRType *>(&src.m_qr)->_sort_matrix_Q();

From b2053990d078e15017610d526445ce78f2632edf Mon Sep 17 00:00:00 2001
From: Robert Lukierski <robert@lukierski.eu>
Date: Wed, 14 Mar 2018 16:19:43 +0000
Subject: [PATCH 336/680] Adding EIGEN_DEVICE_FUNC to Products, especially
 Dense2Dense Assignment specializations. Otherwise causes problems with small
 fixed size matrix multiplication (call to 0x00 in call_assignment_no_alias in
 debug mode or trap in release with CUDA 9.1).

---
 Eigen/src/Core/ProductEvaluators.h | 50 +++++++++++++++---------------
 1 file changed, 25 insertions(+), 25 deletions(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index fb637191d..5d2737d01 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -137,7 +137,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scal
   typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
 {
   typedef Product<Lhs,Rhs,Options> SrcXprType;
-  static EIGEN_STRONG_INLINE
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
     Index dstRows = src.rows();
@@ -155,7 +155,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::add_assign_op<
   typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
 {
   typedef Product<Lhs,Rhs,Options> SrcXprType;
-  static EIGEN_STRONG_INLINE
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,Scalar> &)
   {
     eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
@@ -170,7 +170,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::sub_assign_op<
   typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
 {
   typedef Product<Lhs,Rhs,Options> SrcXprType;
-  static EIGEN_STRONG_INLINE
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,Scalar> &)
   {
     eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
@@ -190,7 +190,7 @@ struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBi
   typedef CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>,
                         const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>,
                         const Product<Lhs,Rhs,DefaultProduct> > SrcXprType;
-  static EIGEN_STRONG_INLINE
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func)
   {
     call_assignment_no_alias(dst, (src.lhs().functor().m_other * src.rhs().lhs())*src.rhs().rhs(), func);
@@ -217,7 +217,7 @@ template<typename DstXprType, typename OtherXpr, typename ProductType, typename
 struct assignment_from_xpr_op_product
 {
   template<typename SrcXprType, typename InitialFunc>
-  static EIGEN_STRONG_INLINE
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/)
   {
     call_assignment_no_alias(dst, src.lhs(), Func1());
@@ -246,19 +246,19 @@ template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
 {
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum();
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); }
 };
 
@@ -269,7 +269,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
 
 // Column major result
 template<typename Dst, typename Lhs, typename Rhs, typename Func>
-void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
+void EIGEN_DEVICE_FUNC outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
 {
   evaluator<Rhs> rhsEval(rhs);
   typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs);
@@ -282,7 +282,7 @@ void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const
 
 // Row major result
 template<typename Dst, typename Lhs, typename Rhs, typename Func>
-void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
+void EIGEN_DEVICE_FUNC outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
 {
   evaluator<Lhs> lhsEval(lhs);
   typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs);
@@ -300,37 +300,37 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose
-  struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
-  struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
-  struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
+  struct set  { template<typename Dst, typename Src> EIGEN_DEVICE_FUNC void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
+  struct add  { template<typename Dst, typename Src> EIGEN_DEVICE_FUNC void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
+  struct sub  { template<typename Dst, typename Src> EIGEN_DEVICE_FUNC void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
   struct adds {
     Scalar m_scale;
     explicit adds(const Scalar& s) : m_scale(s) {}
-    template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
+    template<typename Dst, typename Src> void EIGEN_DEVICE_FUNC operator()(const Dst& dst, const Src& src) const {
       dst.const_cast_derived() += m_scale * src;
     }
   };
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>());
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
     internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>());
   }
@@ -345,15 +345,15 @@ struct generic_product_impl_base
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   { dst.setZero(); scaleAndAddTo(dst, lhs, rhs, Scalar(1)); }
 
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   { scaleAndAddTo(dst,lhs, rhs, Scalar(1)); }
 
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); }
   
   template<typename Dst>
@@ -373,7 +373,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct>
   typedef typename internal::remove_all<typename internal::conditional<int(Side)==OnTheRight,LhsNested,RhsNested>::type>::type MatrixType;
 
   template<typename Dest>
-  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
     LhsNested actual_lhs(lhs);
     RhsNested actual_rhs(rhs);
@@ -390,7 +390,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     // Same as: dst.noalias() = lhs.lazyProduct(rhs);
     // but easier on the compiler side
@@ -398,14 +398,14 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     // dst.noalias() += lhs.lazyProduct(rhs);
     call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::add_assign_op<typename Dst::Scalar,Scalar>());
   }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     // dst.noalias() -= lhs.lazyProduct(rhs);
     call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>());

From e7147f69ae3282b7ae9feb80ffdcc4ae432c68e0 Mon Sep 17 00:00:00 2001
From: Jeff Trull <edaskel@att.net>
Date: Sat, 21 Apr 2018 10:26:30 -0700
Subject: [PATCH 337/680] Add tests for sparseQR results (value and size)
 covering bugs #1522 and #1544

---
 test/sparseqr.cpp | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/test/sparseqr.cpp b/test/sparseqr.cpp
index e8605fd21..f0e721fce 100644
--- a/test/sparseqr.cpp
+++ b/test/sparseqr.cpp
@@ -54,6 +54,28 @@ template<typename Scalar> void test_sparseqr_scalar()
   
   b = dA * DenseVector::Random(A.cols());
   solver.compute(A);
+
+  // Q should be MxM
+  VERIFY_IS_EQUAL(solver.matrixQ().rows(), A.rows());
+  VERIFY_IS_EQUAL(solver.matrixQ().cols(), A.rows());
+
+  // R should be MxN
+  VERIFY_IS_EQUAL(solver.matrixR().rows(), A.rows());
+  VERIFY_IS_EQUAL(solver.matrixR().cols(), A.cols());
+
+  // Q and R can be multiplied
+  DenseMat recoveredA = solver.matrixQ()
+                      * DenseMat(solver.matrixR().template triangularView<Upper>())
+                      * solver.colsPermutation().transpose();
+  VERIFY_IS_EQUAL(recoveredA.rows(), A.rows());
+  VERIFY_IS_EQUAL(recoveredA.cols(), A.cols());
+
+  // and in the full rank case the original matrix is recovered
+  if (solver.rank() == A.cols())
+  {
+      VERIFY_IS_APPROX(A, recoveredA);
+  }
+
   if(internal::random<float>(0,1)>0.5f)
     solver.factorize(A);  // this checks that calling analyzePattern is not needed if the pattern do not change.
   if (solver.info() != Success)

From 7134fa7a2eea469f35ea12899e693a633b5b42e5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 09:33:10 +0200
Subject: [PATCH 338/680] Fix compilation with MSVC by reverting to char* for
 _mm_prefetch except for PGI (the later being the one that has the wrong
 prototype).

---
 Eigen/src/Core/arch/AVX/PacketMath.h    |  6 +++---
 Eigen/src/Core/arch/AVX512/PacketMath.h |  6 +++---
 Eigen/src/Core/arch/SSE/Complex.h       |  4 ++--
 Eigen/src/Core/arch/SSE/PacketMath.h    | 12 +++++++++---
 4 files changed, 17 insertions(+), 11 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 59aebd912..774e64981 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -318,9 +318,9 @@ template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a)
 }
 
 #ifndef EIGEN_VECTORIZE_AVX512
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
 #endif
 
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet8f>(const Packet8f& a) {
diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index eb5de43d4..4e2e916de 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -604,9 +604,9 @@ EIGEN_STRONG_INLINE void pstore1<Packet16i>(int* to, const int& a) {
   pstore(to, pa);
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
 
 template <>
 EIGEN_STRONG_INLINE float pfirst<Packet16f>(const Packet16f& a) {
diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h
index c618cfaaa..d075043ce 100644
--- a/Eigen/src/Core/arch/SSE/Complex.h
+++ b/Eigen/src/Core/arch/SSE/Complex.h
@@ -128,7 +128,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf
                                      _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
 {
@@ -324,7 +324,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
 
 template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
 {
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index d1a7c65be..c944d2c0e 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -461,10 +461,16 @@ template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double&
   pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
 }
 
+#if EIGEN_COMP_PGI
+typedef const void * SsePrefetchPtrType;
+#else
+typedef const char * SsePrefetchPtrType;
+#endif
+
 #ifndef EIGEN_VECTORIZE_AVX
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
 #endif
 
 #if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64

From 6293ad3f392a7b97ebb9f9f874682505c1391f2d Mon Sep 17 00:00:00 2001
From: Vamsi Sripathi <vamsi.sripathi@intel.com>
Date: Wed, 23 May 2018 14:02:05 -0700
Subject: [PATCH 339/680] Performance improvements to tensor broadcast
 operation   1. Added new packet functions using SIMD for NByOne, OneByN cases
   2. Modified existing packet functions to reduce index calculations when
 input stride is non-SIMD   3. Added 4 test cases to cover the new packet
 functions

---
 .../CXX11/src/Tensor/TensorBroadcasting.h     | 110 +++++++++++++++++-
 .../test/cxx11_tensor_broadcasting.cpp        |  62 ++++++++++
 2 files changed, 168 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index b6c93aff9..9ab6b3565 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -105,6 +105,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  bool nByOne = false, oneByN = false;
 
   enum {
     IsAligned = true,
@@ -142,6 +143,24 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
         m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
       }
     }
+
+    if (input_dims[0] == 1) {
+      oneByN = true;
+      for (int i = 1; i < NumDims; ++i) {
+        if (broadcast[i] != 1) {
+          oneByN = false;
+          break;
+        }
+      }
+    } else if (input_dims[NumDims-1] == 1) {
+      nByOne = true;
+      for (int i = 0; i < NumDims-1; ++i) {
+        if (broadcast[i] != 1) {
+          nByOne = false;
+          break;
+        }
+      }
+    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -237,9 +256,84 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     }
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      return packetColMajor<LoadMode>(index);
+      if (oneByN) {
+        return packetNByOne<LoadMode>(index);
+      } else if (nByOne) {
+        return packetOneByN<LoadMode>(index);
+      } else {
+        return packetColMajor<LoadMode>(index);
+      }
     } else {
-      return packetRowMajor<LoadMode>(index);
+      if (oneByN) {
+        return packetOneByN<LoadMode>(index);
+      } else if (nByOne) {
+        return packetNByOne<LoadMode>(index);
+      } else {
+        return packetRowMajor<LoadMode>(index);
+      }
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByN(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    Index dim, inputIndex;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      dim = NumDims - 1;
+    } else {
+      dim = 0;
+    }
+
+    inputIndex = index % m_inputStrides[dim];
+    if (inputIndex + PacketSize <= m_inputStrides[dim]) {
+      return m_impl.template packet<Unaligned>(inputIndex);
+    } else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      for (int i = 0; i < PacketSize; ++i) {
+        if (inputIndex > m_inputStrides[dim]-1) {
+          inputIndex = 0;
+        }
+        values[i] = m_impl.coeff(inputIndex++);
+      }
+      return internal::pload<PacketReturnType>(values);
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetNByOne(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    Index dim, inputIndex, outputOffset;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      dim = 1;
+    } else {
+      dim = NumDims - 2;
+    }
+
+    inputIndex   = index / m_outputStrides[dim];
+    outputOffset = index % m_outputStrides[dim];
+    if (outputOffset + PacketSize <= m_outputStrides[dim]) {
+      values[0] = m_impl.coeff(inputIndex);
+      return internal::pload1<PacketReturnType>(values);
+    } else {
+      for (int i = 0, cur = 0; i < PacketSize; ++i, ++cur) {
+        if (outputOffset + cur < m_outputStrides[dim]) {
+          values[i] = m_impl.coeff(inputIndex);
+        } else {
+          values[i] = m_impl.coeff(++inputIndex);
+          outputOffset = 0;
+          cur = 0;
+        }
+      }
+      return internal::pload<PacketReturnType>(values);
     }
   }
 
@@ -290,7 +384,11 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
       values[0] = m_impl.coeff(inputIndex);
       for (int i = 1; i < PacketSize; ++i) {
-        values[i] = coeffColMajor(originalIndex+i);
+        if (innermostLoc + i < m_impl.dimensions()[0]) {
+          values[i] = m_impl.coeff(inputIndex+i);
+        } else {
+          values[i] = coeffColMajor(originalIndex+i);
+        }
       }
       PacketReturnType rslt = internal::pload<PacketReturnType>(values);
       return rslt;
@@ -342,7 +440,11 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
       values[0] = m_impl.coeff(inputIndex);
       for (int i = 1; i < PacketSize; ++i) {
-        values[i] = coeffRowMajor(originalIndex+i);
+        if (innermostLoc + i < m_impl.dimensions()[NumDims-1]) {
+          values[i] = m_impl.coeff(inputIndex+i);
+        } else {
+          values[i] = coeffRowMajor(originalIndex+i);
+        }
       }
       PacketReturnType rslt = internal::pload<PacketReturnType>(values);
       return rslt;
diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp
index 5c0ea5889..a9d268ea6 100644
--- a/unsupported/test/cxx11_tensor_broadcasting.cpp
+++ b/unsupported/test/cxx11_tensor_broadcasting.cpp
@@ -180,6 +180,64 @@ static void test_fixed_size_broadcasting()
 #endif
 }
 
+template <int DataLayout>
+static void test_simple_broadcasting_one_by_n()
+{
+  Tensor<float, 4, DataLayout> tensor(1,13,5,7);
+  tensor.setRandom();
+  array<ptrdiff_t, 4> broadcasts;
+  broadcasts[0] = 9;
+  broadcasts[1] = 1;
+  broadcasts[2] = 1;
+  broadcasts[3] = 1;
+  Tensor<float, 4, DataLayout> broadcast;
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 9);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 13);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 5);
+  VERIFY_IS_EQUAL(broadcast.dimension(3), 7);
+
+  for (int i = 0; i < 9; ++i) {
+    for (int j = 0; j < 13; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        for (int l = 0; l < 7; ++l) {
+          VERIFY_IS_EQUAL(tensor(i%1,j%13,k%5,l%7), broadcast(i,j,k,l));
+        }
+      }
+    }
+  }
+}
+
+template <int DataLayout>
+static void test_simple_broadcasting_n_by_one()
+{
+  Tensor<float, 4, DataLayout> tensor(7,3,5,1);
+  tensor.setRandom();
+  array<ptrdiff_t, 4> broadcasts;
+  broadcasts[0] = 1;
+  broadcasts[1] = 1;
+  broadcasts[2] = 1;
+  broadcasts[3] = 19;
+  Tensor<float, 4, DataLayout> broadcast;
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 7);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 3);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 5);
+  VERIFY_IS_EQUAL(broadcast.dimension(3), 19);
+
+  for (int i = 0; i < 7; ++i) {
+    for (int j = 0; j < 3; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        for (int l = 0; l < 19; ++l) {
+          VERIFY_IS_EQUAL(tensor(i%7,j%3,k%5,l%1), broadcast(i,j,k,l));
+        }
+      }
+    }
+  }
+}
+
 
 void test_cxx11_tensor_broadcasting()
 {
@@ -191,4 +249,8 @@ void test_cxx11_tensor_broadcasting()
   CALL_SUBTEST(test_static_broadcasting<RowMajor>());
   CALL_SUBTEST(test_fixed_size_broadcasting<ColMajor>());
   CALL_SUBTEST(test_fixed_size_broadcasting<RowMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n<RowMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_n_by_one<RowMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n<ColMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_n_by_one<ColMajor>());
 }

From 405859f18dac56f324e1d93ca8721d5f7fd22c62 Mon Sep 17 00:00:00 2001
From: Mark D Ryan <mark.d.ryan@intel.com>
Date: Thu, 17 May 2018 17:04:00 +0100
Subject: [PATCH 340/680] Set EIGEN_IDEAL_MAX_ALIGN_BYTES correctly for AVX512
 builds

bug #1548

The macro EIGEN_IDEAL_MAX_ALIGN_BYTES is being incorrectly set to 32
on AVX512 builds.  It should be set to 64.  In the current code it is
only set to 64 if the macro EIGEN_VECTORIZE_AVX512 is defined.  This
macro does get defined in AVX512 builds in Core, but only after Macros.h,
the file that defines EIGEN_IDEAL_MAX_ALIGN_BYTES, has been included.
This commit fixes the issue by setting EIGEN_IDEAL_MAX_ALIGN_BYTES to
64 if __AVX512F__ is defined.
---
 Eigen/src/Core/util/Macros.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 9c68ecb7d..729fb3b1d 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -702,7 +702,7 @@ namespace Eigen {
 // If the user explicitly disable vectorization, then we also disable alignment
 #if defined(EIGEN_DONT_VECTORIZE)
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
-#elif defined(EIGEN_VECTORIZE_AVX512)
+#elif defined(__AVX512F__)
   // 64 bytes static alignment is preferred only if really required
   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
 #elif defined(__AVX__)

From 345c0ab450d623b79868ea60253db9d0d8bdd5c7 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 18 May 2018 13:46:46 +0200
Subject: [PATCH 341/680] check that all integer types are properly handled by
 mat(i,j)

---
 test/indexed_view.cpp | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 8b3082cea..71de60d84 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -371,6 +371,21 @@ void check_indexed_view()
   a(X) = 1;
   A(X,Y) = 1;
 
+  // Check compilation of varying integer types as index types:
+  Index i = n/2;
+  short i_short(i);
+  std::size_t i_sizet(i);
+  VERIFY_IS_EQUAL( a(i), a.coeff(i_short) );
+  VERIFY_IS_EQUAL( a(i), a.coeff(i_sizet) );
+
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i_short, i_short) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i_short, i) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i, i_short) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i, i_sizet) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i_sizet, i) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(i_sizet, i_short) );
+  VERIFY_IS_EQUAL( A(i,i), A.coeff(5, i_sizet) );
+
 }
 
 void test_indexed_view()

From 4dd767f455f1adfea7cb2febeab3efaa81246845 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 18 May 2018 13:59:55 +0200
Subject: [PATCH 342/680] add some internal checks

---
 Eigen/src/Core/util/XprHelper.h        |  6 ++++++
 Eigen/src/plugins/IndexedViewMethods.h | 11 +++--------
 test/indexed_view.cpp                  | 15 +++++++++++++++
 3 files changed, 24 insertions(+), 8 deletions(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 10328be0d..1217404d6 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -47,6 +47,12 @@ template<typename T> struct is_valid_index_type
   };
 };
 
+// true if both types are not valid index types
+template<typename RowIndices, typename ColIndices>
+struct valid_indexed_view_overload {
+  enum { value = !(internal::is_valid_index_type<RowIndices>::value && internal::is_valid_index_type<ColIndices>::value) };
+};
+
 // promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
 //    expression * scalar
 // Its role is to determine how the type T of the scalar operand should be promoted given the scalar type ExprScalar of the given expression.
diff --git a/Eigen/src/plugins/IndexedViewMethods.h b/Eigen/src/plugins/IndexedViewMethods.h
index 9ad2d9aee..d435aa597 100644
--- a/Eigen/src/plugins/IndexedViewMethods.h
+++ b/Eigen/src/plugins/IndexedViewMethods.h
@@ -53,11 +53,6 @@ ivcSize(const Indices& indices) const {
   return internal::makeIndexedViewCompatible(indices, internal::variable_if_dynamic<Index,SizeAtCompileTime>(derived().size()),Specialized);
 }
 
-template<typename RowIndices, typename ColIndices>
-struct valid_indexed_view_overload {
-  enum { value = !(internal::is_valid_index_type<RowIndices>::value && internal::is_valid_index_type<ColIndices>::value) };
-};
-
 public:
 
 #endif
@@ -72,7 +67,7 @@ struct EIGEN_INDEXED_VIEW_METHOD_TYPE {
 // This is the generic version
 
 template<typename RowIndices, typename ColIndices>
-typename internal::enable_if<valid_indexed_view_overload<RowIndices,ColIndices>::value
+typename internal::enable_if<internal::valid_indexed_view_overload<RowIndices,ColIndices>::value
   && internal::traits<typename EIGEN_INDEXED_VIEW_METHOD_TYPE<RowIndices,ColIndices>::type>::ReturnAsIndexedView,
   typename EIGEN_INDEXED_VIEW_METHOD_TYPE<RowIndices,ColIndices>::type >::type
 operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_INDEXED_VIEW_METHOD_CONST
@@ -84,7 +79,7 @@ operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_IND
 // The following overload returns a Block<> object
 
 template<typename RowIndices, typename ColIndices>
-typename internal::enable_if<valid_indexed_view_overload<RowIndices,ColIndices>::value
+typename internal::enable_if<internal::valid_indexed_view_overload<RowIndices,ColIndices>::value
   && internal::traits<typename EIGEN_INDEXED_VIEW_METHOD_TYPE<RowIndices,ColIndices>::type>::ReturnAsBlock,
   typename internal::traits<typename EIGEN_INDEXED_VIEW_METHOD_TYPE<RowIndices,ColIndices>::type>::BlockType>::type
 operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_INDEXED_VIEW_METHOD_CONST
@@ -102,7 +97,7 @@ operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_IND
 // The following overload returns a Scalar
 
 template<typename RowIndices, typename ColIndices>
-typename internal::enable_if<valid_indexed_view_overload<RowIndices,ColIndices>::value
+typename internal::enable_if<internal::valid_indexed_view_overload<RowIndices,ColIndices>::value
   && internal::traits<typename EIGEN_INDEXED_VIEW_METHOD_TYPE<RowIndices,ColIndices>::type>::ReturnAsScalar,
   CoeffReturnType >::type
 operator()(const RowIndices& rowIndices, const ColIndices& colIndices) EIGEN_INDEXED_VIEW_METHOD_CONST
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 71de60d84..033d8833f 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -395,4 +395,19 @@ void test_indexed_view()
     CALL_SUBTEST_2( check_indexed_view() );
     CALL_SUBTEST_3( check_indexed_view() );
 //   }
+
+  // static checks of some internals:
+
+  #define STATIC_CHECK( COND ) \
+    EIGEN_STATIC_ASSERT( (COND) , EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT )
+
+  STATIC_CHECK(( internal::is_valid_index_type<int>::value ));
+  STATIC_CHECK(( internal::is_valid_index_type<unsigned int>::value ));
+  STATIC_CHECK(( internal::is_valid_index_type<short>::value ));
+  STATIC_CHECK(( internal::is_valid_index_type<std::ptrdiff_t>::value ));
+  STATIC_CHECK(( internal::is_valid_index_type<std::size_t>::value ));
+  STATIC_CHECK(( !internal::valid_indexed_view_overload<int,int>::value ));
+  STATIC_CHECK(( !internal::valid_indexed_view_overload<int,std::ptrdiff_t>::value ));
+  STATIC_CHECK(( !internal::valid_indexed_view_overload<std::ptrdiff_t,int>::value ));
+  STATIC_CHECK(( !internal::valid_indexed_view_overload<std::size_t,int>::value ));
 }

From a382bc9364b5d5a049678f8cc138670ca1c7a4fa Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 18 May 2018 17:02:27 +0200
Subject: [PATCH 343/680] is_convertible<T,Index> does not seems to work well
 with MSVC 2013, so let's rather use __is_enum(T) for old MSVC versions

---
 Eigen/src/Core/util/XprHelper.h | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 1217404d6..9ffb0474b 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -39,10 +39,12 @@ template<typename T> struct is_valid_index_type
 {
   enum { value =
 #if EIGEN_HAS_TYPE_TRAITS
-   internal::is_integral<T>::value || std::is_enum<T>::value
+    internal::is_integral<T>::value || std::is_enum<T>::value
+#else if EIGEN_COMP_MSVC >= 1500
+    internal::is_integral<T>::value || __is_enum(T)
 #else
-  // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
-  internal::is_convertible<T,Index>::value
+    // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
+    internal::is_convertible<T,Index>::value
 #endif
   };
 };

From 725bd929038300c89c079325d1989e542beaf23c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 18 May 2018 17:46:43 +0200
Subject: [PATCH 344/680] fix stupid typo

---
 Eigen/src/Core/util/XprHelper.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 9ffb0474b..16714fa5c 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -40,7 +40,7 @@ template<typename T> struct is_valid_index_type
   enum { value =
 #if EIGEN_HAS_TYPE_TRAITS
     internal::is_integral<T>::value || std::is_enum<T>::value
-#else if EIGEN_COMP_MSVC >= 1500
+#elif EIGEN_COMP_MSVC
     internal::is_integral<T>::value || __is_enum(T)
 #else
     // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.

From 36e413a534e105ec3864b1a163e0ec3234f866a1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 May 2018 18:51:35 +0200
Subject: [PATCH 345/680] Workaround a MSVC 2013 compilation issue with
 MatrixBase(Index,int)

---
 Eigen/src/SparseLU/SparseLU.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h
index 383a203b4..3bc06f4e3 100644
--- a/Eigen/src/SparseLU/SparseLU.h
+++ b/Eigen/src/SparseLU/SparseLU.h
@@ -747,8 +747,9 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator
       }
       else
       {
+        // FIXME: the following lines should use Block expressions and not Map!
         Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > A( &(m_mapL.valuePtr()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
-        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
+        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X.coeffRef(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
         U = A.template triangularView<Upper>().solve(U);
       }
 

From f0862b062fcb613ee6c60745631d90a43d54a6d4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 22 May 2018 19:29:51 +0200
Subject: [PATCH 346/680] Fix internal::is_integral<size_t/ptrdiff_t> with MSVC
 2013 and older.

---
 Eigen/src/Core/util/Meta.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 0c7b8e709..6e5af35c0 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -113,6 +113,10 @@ template<> struct is_integral<signed int>             { enum { value = true }; }
 template<> struct is_integral<unsigned int>           { enum { value = true }; };
 template<> struct is_integral<signed long>            { enum { value = true }; };
 template<> struct is_integral<unsigned long>          { enum { value = true }; };
+#if EIGEN_COMP_MSVC
+template<> struct is_integral<signed __int64>         { enum { value = true }; };
+template<> struct is_integral<unsigned __int64>       { enum { value = true }; };
+#endif
 #endif
 
 

From d06a753d1017255f2a10a5b00a830865c30fe4df Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 22 May 2018 20:29:17 +0200
Subject: [PATCH 347/680] Make qr_fullpivoting unit test run for fixed-sized
 matrices

---
 test/qr_fullpivoting.cpp | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/test/qr_fullpivoting.cpp b/test/qr_fullpivoting.cpp
index 70e89c198..3d9accea6 100644
--- a/test/qr_fullpivoting.cpp
+++ b/test/qr_fullpivoting.cpp
@@ -15,11 +15,12 @@ template<typename MatrixType> void qr()
 {
   typedef typename MatrixType::Index Index;
 
+  static const int Rows = MatrixType::RowsAtCompileTime, Cols = MatrixType::ColsAtCompileTime;
   Index max_size = EIGEN_TEST_MAX_SIZE;
   Index min_size = numext::maxi(1,EIGEN_TEST_MAX_SIZE/10);
-  Index rows  = internal::random<Index>(min_size,max_size),
-        cols  = internal::random<Index>(min_size,max_size),
-        cols2 = internal::random<Index>(min_size,max_size),
+  Index rows  = Rows == Dynamic ? internal::random<Index>(min_size,max_size) : Rows,
+        cols  = Cols == Dynamic ? internal::random<Index>(min_size,max_size) : Cols,
+        cols2 = Cols == Dynamic ? internal::random<Index>(min_size,max_size) : Cols,
         rank  = internal::random<Index>(1, (std::min)(rows, cols)-1);
 
   typedef typename MatrixType::Scalar Scalar;
@@ -128,9 +129,10 @@ template<typename MatrixType> void qr_verify_assert()
 
 void test_qr_fullpivoting()
 {
- for(int i = 0; i < 1; i++) {
-    // FIXME : very weird bug here
-//     CALL_SUBTEST(qr(Matrix2f()) );
+  for(int i = 0; i < 1; i++) {
+    CALL_SUBTEST_5( qr<Matrix3f>() );
+    CALL_SUBTEST_6( qr<Matrix3d>() );
+    CALL_SUBTEST_8( qr<Matrix2f>() );
     CALL_SUBTEST_1( qr<MatrixXf>() );
     CALL_SUBTEST_2( qr<MatrixXd>() );
     CALL_SUBTEST_3( qr<MatrixXcd>() );

From 750af063629cd366f6d01985b37a21ebc37a5af3 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 22 May 2018 21:04:32 +0200
Subject: [PATCH 348/680] Add an option to test with external BLAS library

---
 cmake/EigenTesting.cmake | 4 ++++
 test/CMakeLists.txt      | 9 +++++++++
 2 files changed, 13 insertions(+)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 16d6d279f..7d2d63722 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -679,6 +679,10 @@ macro(ei_set_build_string)
     set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-${LOCAL_COMPILER_FLAGS})
   endif()
 
+  if(EIGEN_TEST_EXTERNAL_BLAS)
+    set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-external_blas)
+  endif()
+
   ei_is_64bit_env(IS_64BIT_ENV)
   if(NOT IS_64BIT_ENV)
     set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-32bit)
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index f8deafc07..073effd81 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -30,6 +30,15 @@ if(NOT EIGEN_Fortran_COMPILER_WORKS)
   find_package(LAPACK QUIET)
 endif()
 
+# TODO do the same for EXTERNAL_LAPACK
+option(EIGEN_TEST_EXTERNAL_BLAS "Use external BLAS library for testsuite" OFF)
+if(EIGEN_TEST_EXTERNAL_BLAS)
+  find_package(BLAS REQUIRED)
+  message(STATUS "BLAS_COMPILER_FLAGS: ${BLAS_COMPILER_FLAGS}")
+  add_definitions("-DEIGEN_USE_BLAS") # is adding  ${BLAS_COMPILER_FLAGS} necessary?
+  list(APPEND EXTERNAL_LIBS "${BLAS_LIBRARIES}")
+endif(EIGEN_TEST_EXTERNAL_BLAS)
+
 # configure blas/lapack (use Eigen's ones)
 set(EIGEN_BLAS_LIBRARIES eigen_blas)
 set(EIGEN_LAPACK_LIBRARIES eigen_lapack)

From 49262dfee6785bcf161ee48e619035bc5b30976c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 29 May 2018 15:09:31 +0200
Subject: [PATCH 349/680] Fix compilation and SSE support with PGI compiler

---
 Eigen/src/Core/arch/AVX/PacketMath.h |  6 +++---
 Eigen/src/Core/arch/SSE/Complex.h    |  4 ++--
 Eigen/src/Core/arch/SSE/PacketMath.h | 16 +++++++++++++---
 3 files changed, 18 insertions(+), 8 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 34b89c790..59aebd912 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -318,9 +318,9 @@ template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a)
 }
 
 #ifndef EIGEN_VECTORIZE_AVX512
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
 #endif
 
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet8f>(const Packet8f& a) {
diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h
index 23e717f28..c618cfaaa 100644
--- a/Eigen/src/Core/arch/SSE/Complex.h
+++ b/Eigen/src/Core/arch/SSE/Complex.h
@@ -128,7 +128,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf
                                      _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
 {
@@ -324,7 +324,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
 
 template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
 {
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 03c8a2c13..d1a7c65be 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -462,9 +462,9 @@ template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double&
 }
 
 #ifndef EIGEN_VECTORIZE_AVX
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const void*)(addr), _MM_HINT_T0); }
 #endif
 
 #if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
@@ -928,4 +928,14 @@ template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, co
 
 } // end namespace Eigen
 
+#if EIGEN_COMP_PGI
+// PGI++ does not define the following intrinsics in C++ mode.
+static inline __m128  _mm_castpd_ps   (__m128d x) { return reinterpret_cast<__m128&>(x);  }
+static inline __m128i _mm_castpd_si128(__m128d x) { return reinterpret_cast<__m128i&>(x); }
+static inline __m128d _mm_castps_pd   (__m128  x) { return reinterpret_cast<__m128d&>(x); }
+static inline __m128i _mm_castps_si128(__m128  x) { return reinterpret_cast<__m128i&>(x); }
+static inline __m128  _mm_castsi128_ps(__m128i x) { return reinterpret_cast<__m128&>(x);  }
+static inline __m128d _mm_castsi128_pd(__m128i x) { return reinterpret_cast<__m128d&>(x); }
+#endif
+
 #endif // EIGEN_PACKET_MATH_SSE_H

From 647b724a364f4492f39f607d7007b7feac838bf8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 29 May 2018 20:46:46 +0200
Subject: [PATCH 350/680] Define pcast<> for SSE types even when AVX is
 enabled. (otherwise float are silently reinterpreted as int instead of being
 converted)

---
 Eigen/Core                            |  1 +
 Eigen/src/Core/arch/SSE/TypeCasting.h | 28 +++++++++++++--------------
 2 files changed, 15 insertions(+), 14 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index c74340b81..f6bc18a08 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -413,6 +413,7 @@ using std::ptrdiff_t;
   #include "src/Core/arch/AVX/MathFunctions.h"
   #include "src/Core/arch/AVX/Complex.h"
   #include "src/Core/arch/AVX/TypeCasting.h"
+  #include "src/Core/arch/SSE/TypeCasting.h"
 #elif defined EIGEN_VECTORIZE_SSE
   #include "src/Core/arch/SSE/PacketMath.h"
   #include "src/Core/arch/SSE/MathFunctions.h"
diff --git a/Eigen/src/Core/arch/SSE/TypeCasting.h b/Eigen/src/Core/arch/SSE/TypeCasting.h
index c84893230..c6ca8c716 100644
--- a/Eigen/src/Core/arch/SSE/TypeCasting.h
+++ b/Eigen/src/Core/arch/SSE/TypeCasting.h
@@ -14,6 +14,7 @@ namespace Eigen {
 
 namespace internal {
 
+#ifndef EIGEN_VECTORIZE_AVX
 template <>
 struct type_casting_traits<float, int> {
   enum {
@@ -23,11 +24,6 @@ struct type_casting_traits<float, int> {
   };
 };
 
-template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
-  return _mm_cvttps_epi32(a);
-}
-
-
 template <>
 struct type_casting_traits<int, float> {
   enum {
@@ -37,11 +33,6 @@ struct type_casting_traits<int, float> {
   };
 };
 
-template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
-  return _mm_cvtepi32_ps(a);
-}
-
-
 template <>
 struct type_casting_traits<double, float> {
   enum {
@@ -51,10 +42,6 @@ struct type_casting_traits<double, float> {
   };
 };
 
-template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) {
-  return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
-}
-
 template <>
 struct type_casting_traits<float, double> {
   enum {
@@ -63,6 +50,19 @@ struct type_casting_traits<float, double> {
     TgtCoeffRatio = 2
   };
 };
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
+  return _mm_cvttps_epi32(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
+  return _mm_cvtepi32_ps(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) {
+  return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
+}
 
 template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) {
   // Simply discard the second half of the input

From eef4b7bd8768a79078ba6d4adc9f85b3f5937e39 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 29 May 2018 20:49:06 +0200
Subject: [PATCH 351/680] Fix handling of path names containing spaces and the
 likes.

---
 test/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 073effd81..12c12b06d 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -47,7 +47,7 @@ set(EIGEN_TEST_MATRIX_DIR "" CACHE STRING "Enable testing of realword sparse mat
 if(EIGEN_TEST_MATRIX_DIR)
   if(NOT WIN32)
     message(STATUS "Test realworld sparse matrices: ${EIGEN_TEST_MATRIX_DIR}")
-    add_definitions( -DTEST_REAL_CASES="${EIGEN_TEST_MATRIX_DIR}" )
+    add_definitions( -DTEST_REAL_CASES=${EIGEN_TEST_MATRIX_DIR} )
   else(NOT WIN32)
     message(STATUS "REAL CASES CAN NOT BE CURRENTLY TESTED ON WIN32")
   endif(NOT WIN32)

From 999b552c16fe48ba6ca2ea0fe8d8788775f3c58d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 29 May 2018 20:49:25 +0200
Subject: [PATCH 352/680] Search for sequential Pastix.

---
 test/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 12c12b06d..e1eef086e 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -100,7 +100,7 @@ else()
 endif()
 
 
-find_package(PASTIX QUIET COMPONENTS METIS SCOTCH)
+find_package(PASTIX QUIET COMPONENTS METIS SEQ)
 # check that the PASTIX found is a version without MPI
 find_path(PASTIX_pastix_nompi.h_INCLUDE_DIRS
   NAMES pastix_nompi.h

From ea94543190354e9cac55784b9e4f17549a14906f Mon Sep 17 00:00:00 2001
From: Katrin Leinweber <katrin.leinweber@uni-konstanz.de>
Date: Thu, 24 May 2018 18:55:40 +0200
Subject: [PATCH 353/680] Hyperlink DOIs against preferred resolver

---
 unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h | 2 +-
 unsupported/Eigen/src/IterativeSolvers/DGMRES.h   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
index fb6454623..bf5c532ff 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
@@ -21,7 +21,7 @@ namespace Eigen {
   * \brief Fast integer division by a constant.
   *
   * See the paper from Granlund and Montgomery for explanation.
-  *   (at http://dx.doi.org/10.1145/773473.178249)
+  *   (at https://doi.org/10.1145/773473.178249)
   *
   * \sa Tensor
   */
diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index bae04fc30..d603ba336 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -89,7 +89,7 @@ void sortWithPermutation (VectorType& vec, IndexType& perm, typename IndexType::
  * [1] D. NUENTSA WAKAM and F. PACULL, Memory Efficient Hybrid
  *  Algebraic Solvers for Linear Systems Arising from Compressible
  *  Flows, Computers and Fluids, In Press,
- *  http://dx.doi.org/10.1016/j.compfluid.2012.03.023   
+ *  https://doi.org/10.1016/j.compfluid.2012.03.023   
  * [2] K. Burrage and J. Erhel, On the performance of various 
  * adaptive preconditioned GMRES strategies, 5(1998), 101-121.
  * [3] J. Erhel, K. Burrage and B. Pohl, Restarted GMRES 

From 6af1433cb50af7423a1a69afc24c098af9c76bb1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 29 May 2018 22:37:47 +0200
Subject: [PATCH 354/680] Doc: add aliasing in common pitfaffs.

---
 doc/Pitfalls.dox | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/doc/Pitfalls.dox b/doc/Pitfalls.dox
index cf42effef..3f395053d 100644
--- a/doc/Pitfalls.dox
+++ b/doc/Pitfalls.dox
@@ -2,10 +2,16 @@ namespace Eigen {
 
 /** \page TopicPitfalls Common pitfalls
 
+
 \section TopicPitfalls_template_keyword Compilation error with template methods
 
 See this \link TopicTemplateKeyword page \endlink.
 
+\section TopicPitfalls_aliasing Aliasing
+
+Don't miss this \link TopicAliasing page \endlink on aliasing,
+especially if you got wrong results in statements where the destination appears on the right hand side of the expression.
+
 \section TopicPitfalls_auto_keyword C++11 and the auto keyword
 
 In short: do not use the auto keywords with Eigen's expressions, unless you are 100% sure about what you are doing. In particular, do not use the auto keyword as a replacement for a Matrix<> type. Here is an example:

From f216854453887f31ac02ffefb7a7a569dc3fa54d Mon Sep 17 00:00:00 2001
From: Michael Figurnov <mfigurnov@google.com>
Date: Thu, 31 May 2018 15:34:53 +0100
Subject: [PATCH 355/680] Exponentially scaled modified Bessel functions of
 order zero and one.

The functions are conventionally called i0e and i1e. The exponentially scaled version is more numerically stable. The standard Bessel functions can be obtained as i0(x) = exp(|x|) i0e(x)

The code is ported from Cephes and tested against SciPy.
---
 Eigen/src/Core/GenericPacketMath.h            |   2 +
 Eigen/src/Core/arch/CUDA/PacketMath.h         |   4 +
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  12 +
 unsupported/Eigen/SpecialFunctions            |   2 +
 .../SpecialFunctionsArrayAPI.h                |  46 ++
 .../SpecialFunctionsFunctors.h                |  54 +++
 .../SpecialFunctions/SpecialFunctionsHalf.h   |   8 +
 .../SpecialFunctions/SpecialFunctionsImpl.h   | 409 ++++++++++++++++++
 .../SpecialFunctionsPacketMath.h              |  12 +
 .../arch/CUDA/CudaSpecialFunctions.h          |  26 ++
 unsupported/test/cxx11_tensor_cuda.cu         | 116 +++++
 unsupported/test/special_functions.cpp        |  40 ++
 12 files changed, 731 insertions(+)

diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 7979c3aff..888a3f7ea 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -82,6 +82,8 @@ struct default_packet_traits
     HasPolygamma = 0,
     HasErf = 0,
     HasErfc = 0,
+    HasI0e = 0,
+    HasI1e = 0,
     HasIGamma = 0,
     HasIGammac = 0,
     HasBetaInc = 0,
diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index 97a8abe59..704a4e0d9 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -44,6 +44,8 @@ template<> struct packet_traits<float> : default_packet_traits
     HasPolygamma = 1,
     HasErf = 1,
     HasErfc = 1,
+    HasI0e = 1,
+    HasI1e = 1,
     HasIGamma = 1,
     HasIGammac = 1,
     HasBetaInc = 1,
@@ -73,6 +75,8 @@ template<> struct packet_traits<double> : default_packet_traits
     HasPolygamma = 1,
     HasErf = 1,
     HasErfc = 1,
+    HasI0e = 1,
+    HasI1e = 1,
     HasIGamma = 1,
     HasIGammac = 1,
     HasBetaInc = 1,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 745b16d2c..a942c98dd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -133,6 +133,18 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return unaryExpr(internal::scalar_digamma_op<Scalar>());
     }
 
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_i0e_op<Scalar>, const Derived>
+    i0e() const {
+      return unaryExpr(internal::scalar_i0e_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_i1e_op<Scalar>, const Derived>
+    i1e() const {
+      return unaryExpr(internal::scalar_i1e_op<Scalar>());
+    }
+
     // igamma(a = this, x = other)
     template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorCwiseBinaryOp<internal::scalar_igamma_op<Scalar>, const Derived, const OtherDerived>
diff --git a/unsupported/Eigen/SpecialFunctions b/unsupported/Eigen/SpecialFunctions
index a2ad4925e..482ec6e6f 100644
--- a/unsupported/Eigen/SpecialFunctions
+++ b/unsupported/Eigen/SpecialFunctions
@@ -34,6 +34,8 @@ namespace Eigen {
   * - polygamma
   * - zeta
   * - betainc
+  * - i0e
+  * - i1e
   *
   * \code
   * #include <unsupported/Eigen/SpecialFunctions>
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
index ed415db99..b7a9d035b 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
@@ -119,6 +119,52 @@ zeta(const Eigen::ArrayBase<DerivedX>& x, const Eigen::ArrayBase<DerivedQ>& q)
   );
 }
 
+/** \returns an expression of the coefficient-wise i0e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i0e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::i0e()
+  */
+template <typename Derived>
+inline const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_i0e_op<typename Derived::Scalar>, const Derived>
+i0e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_i0e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise i1e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i1e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::i1e()
+  */
+template <typename Derived>
+inline const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_i1e_op<typename Derived::Scalar>, const Derived>
+i1e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_i1e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_SPECIALFUNCTIONS_ARRAYAPI_H
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
index d8f2363be..8420f0174 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
@@ -229,6 +229,60 @@ struct functor_traits<scalar_erfc_op<Scalar> >
   };
 };
 
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of order zero
+ * \sa class CwiseUnaryOp, Cwise::i0e()
+ */
+template <typename Scalar>
+struct scalar_i0e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_i0e_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& x) const {
+    using numext::i0e;
+    return i0e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const {
+    return internal::pi0e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_i0e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions.
+    Cost = 20 * NumTraits<Scalar>::MulCost + 40 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasI0e
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of order zero
+ * \sa class CwiseUnaryOp, Cwise::i1e()
+ */
+template <typename Scalar>
+struct scalar_i1e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_i1e_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& x) const {
+    using numext::i1e;
+    return i1e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const {
+    return internal::pi1e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_i1e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions.
+    Cost = 20 * NumTraits<Scalar>::MulCost + 40 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasI1e
+  };
+};
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
index 553bcda6a..c5867002e 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
@@ -39,6 +39,14 @@ template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igammac(const Eigen
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half betainc(const Eigen::half& a, const Eigen::half& b, const Eigen::half& x) {
   return Eigen::half(Eigen::numext::betainc(static_cast<float>(a), static_cast<float>(b), static_cast<float>(x)));
 }
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half i0e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::i0e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half i1e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::i1e(static_cast<float>(x)));
+}
 #endif
 
 }  // end namespace numext
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 97b2e5a91..6c7ac3f3b 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -95,6 +95,75 @@ struct polevl<Scalar, 0> {
   }
 };
 
+/* chbevl (modified for Eigen)
+ *
+ *     Evaluate Chebyshev series
+ *
+ *
+ *
+ * SYNOPSIS:
+ *
+ * int N;
+ * Scalar x, y, coef[N], chebevl();
+ *
+ * y = chbevl( x, coef, N );
+ *
+ *
+ *
+ * DESCRIPTION:
+ *
+ * Evaluates the series
+ *
+ *        N-1
+ *         - '
+ *  y  =   >   coef[i] T (x/2)
+ *         -            i
+ *        i=0
+ *
+ * of Chebyshev polynomials Ti at argument x/2.
+ *
+ * Coefficients are stored in reverse order, i.e. the zero
+ * order term is last in the array.  Note N is the number of
+ * coefficients, not the order.
+ *
+ * If coefficients are for the interval a to b, x must
+ * have been transformed to x -> 2(2x - b - a)/(b-a) before
+ * entering the routine.  This maps x from (a, b) to (-1, 1),
+ * over which the Chebyshev polynomials are defined.
+ *
+ * If the coefficients are for the inverted interval, in
+ * which (a, b) is mapped to (1/b, 1/a), the transformation
+ * required is x -> 2(2ab/x - b - a)/(b-a).  If b is infinity,
+ * this becomes x -> 4a/x - 1.
+ *
+ *
+ *
+ * SPEED:
+ *
+ * Taking advantage of the recurrence properties of the
+ * Chebyshev polynomials, the routine requires one more
+ * addition per loop than evaluating a nested polynomial of
+ * the same degree.
+ *
+ */
+template <typename Scalar, int N>
+struct chebevl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(Scalar x, const Scalar coef[]) {
+    Scalar b0 = coef[0];
+    Scalar b1 = 0;
+    Scalar b2;
+
+    for (int i = 1; i < N; i++) {
+      b2 = b1;
+      b1 = b0;
+      b0 = x * b1 - b2 + coef[i];
+    }
+
+    return Scalar(0.5) * (b0 - b2);
+  }
+};
+
 }  // end namespace cephes
 
 /****************************************************************************
@@ -1505,6 +1574,334 @@ struct betainc_impl<double> {
   }
 };
 
+/****************************************************************************
+ * Implementation of Bessel function, based on Cephes                       *
+ ****************************************************************************/
+
+template <typename Scalar>
+struct i0e_retval {
+  typedef Scalar type;
+};
+
+template <typename Scalar>
+struct i0e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template <>
+struct i0e_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(float x) {
+    /*  i0ef.c
+     *
+     *  Modified Bessel function of order zero,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, i0ef();
+     *
+     * y = i0ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order zero of the argument.
+     *
+     * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        100000      3.7e-7      7.0e-8
+     * See i0f().
+     *
+     */
+    const float A[] = {-1.30002500998624804212E-8f, 6.04699502254191894932E-8f,
+                       -2.67079385394061173391E-7f, 1.11738753912010371815E-6f,
+                       -4.41673835845875056359E-6f, 1.64484480707288970893E-5f,
+                       -5.75419501008210370398E-5f, 1.88502885095841655729E-4f,
+                       -5.76375574538582365885E-4f, 1.63947561694133579842E-3f,
+                       -4.32430999505057594430E-3f, 1.05464603945949983183E-2f,
+                       -2.37374148058994688156E-2f, 4.93052842396707084878E-2f,
+                       -9.49010970480476444210E-2f, 1.71620901522208775349E-1f,
+                       -3.04682672343198398683E-1f, 6.76795274409476084995E-1f};
+
+    const float B[] = {3.39623202570838634515E-9f, 2.26666899049817806459E-8f,
+                       2.04891858946906374183E-7f, 2.89137052083475648297E-6f,
+                       6.88975834691682398426E-5f, 3.36911647825569408990E-3f,
+                       8.04490411014108831608E-1f};
+    if (x < 0.0f) {
+      x = -x;
+    }
+
+    if (x <= 8.0f) {
+      float y = 0.5f * x - 2.0f;
+      return cephes::chebevl<float, 18>::run(y, A);
+    }
+
+    return cephes::chebevl<float, 7>::run(32.0f / x - 2.0f, B) / numext::sqrt(x);
+  }
+};
+
+template <>
+struct i0e_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(double x) {
+    /*  i0e.c
+     *
+     *  Modified Bessel function of order zero,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, i0e();
+     *
+     * y = i0e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order zero of the argument.
+     *
+     * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        30000       5.4e-16     1.2e-16
+     * See i0().
+     *
+     */
+    const double A[] = {-4.41534164647933937950E-18, 3.33079451882223809783E-17,
+                        -2.43127984654795469359E-16, 1.71539128555513303061E-15,
+                        -1.16853328779934516808E-14, 7.67618549860493561688E-14,
+                        -4.85644678311192946090E-13, 2.95505266312963983461E-12,
+                        -1.72682629144155570723E-11, 9.67580903537323691224E-11,
+                        -5.18979560163526290666E-10, 2.65982372468238665035E-9,
+                        -1.30002500998624804212E-8,  6.04699502254191894932E-8,
+                        -2.67079385394061173391E-7,  1.11738753912010371815E-6,
+                        -4.41673835845875056359E-6,  1.64484480707288970893E-5,
+                        -5.75419501008210370398E-5,  1.88502885095841655729E-4,
+                        -5.76375574538582365885E-4,  1.63947561694133579842E-3,
+                        -4.32430999505057594430E-3,  1.05464603945949983183E-2,
+                        -2.37374148058994688156E-2,  4.93052842396707084878E-2,
+                        -9.49010970480476444210E-2,  1.71620901522208775349E-1,
+                        -3.04682672343198398683E-1,  6.76795274409476084995E-1};
+    const double B[] = {
+        -7.23318048787475395456E-18, -4.83050448594418207126E-18,
+        4.46562142029675999901E-17,  3.46122286769746109310E-17,
+        -2.82762398051658348494E-16, -3.42548561967721913462E-16,
+        1.77256013305652638360E-15,  3.81168066935262242075E-15,
+        -9.55484669882830764870E-15, -4.15056934728722208663E-14,
+        1.54008621752140982691E-14,  3.85277838274214270114E-13,
+        7.18012445138366623367E-13,  -1.79417853150680611778E-12,
+        -1.32158118404477131188E-11, -3.14991652796324136454E-11,
+        1.18891471078464383424E-11,  4.94060238822496958910E-10,
+        3.39623202570838634515E-9,   2.26666899049817806459E-8,
+        2.04891858946906374183E-7,   2.89137052083475648297E-6,
+        6.88975834691682398426E-5,   3.36911647825569408990E-3,
+        8.04490411014108831608E-1};
+
+    if (x < 0.0) {
+      x = -x;
+    }
+
+    if (x <= 8.0) {
+      double y = (x / 2.0) - 2.0;
+      return cephes::chebevl<double, 30>::run(y, A);
+    }
+
+    return cephes::chebevl<double, 25>::run(32.0 / x - 2.0, B) /
+           numext::sqrt(x);
+  }
+};
+
+template <typename Scalar>
+struct i1e_retval {
+  typedef Scalar type;
+};
+
+template <typename Scalar>
+struct i1e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template <>
+struct i1e_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(float x) {
+    /* i1ef.c
+     *
+     *  Modified Bessel function of order one,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, i1ef();
+     *
+     * y = i1ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order one of the argument.
+     *
+     * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       1.5e-6      1.5e-7
+     * See i1().
+     *
+     */
+    const float A[] = {9.38153738649577178388E-9f, -4.44505912879632808065E-8f,
+                       2.00329475355213526229E-7f, -8.56872026469545474066E-7f,
+                       3.47025130813767847674E-6f, -1.32731636560394358279E-5f,
+                       4.78156510755005422638E-5f, -1.61760815825896745588E-4f,
+                       5.12285956168575772895E-4f, -1.51357245063125314899E-3f,
+                       4.15642294431288815669E-3f, -1.05640848946261981558E-2f,
+                       2.47264490306265168283E-2f, -5.29459812080949914269E-2f,
+                       1.02643658689847095384E-1f, -1.76416518357834055153E-1f,
+                       2.52587186443633654823E-1f};
+
+    const float B[] = {-3.83538038596423702205E-9f, -2.63146884688951950684E-8f,
+                       -2.51223623787020892529E-7f, -3.88256480887769039346E-6f,
+                       -1.10588938762623716291E-4f, -9.76109749136146840777E-3f,
+                       7.78576235018280120474E-1f};
+
+    float z = numext::abs(x);
+
+    if (z <= 8.0f) {
+      float y = 0.5f * z - 2.0f;
+      z = cephes::chebevl<float, 17>::run(y, A) * z;
+    } else {
+      z = cephes::chebevl<float, 7>::run(32.0f / z - 2.0f, B) / numext::sqrt(z);
+    }
+
+    if (x < 0.0f) {
+      z = -z;
+    }
+
+    return z;
+  }
+};
+
+template <>
+struct i1e_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(double x) {
+    /*  i1e.c
+     *
+     *  Modified Bessel function of order one,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, i1e();
+     *
+     * y = i1e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order one of the argument.
+     *
+     * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       2.0e-15     2.0e-16
+     * See i1().
+     *
+     */
+    const double A[] = {2.77791411276104639959E-18, -2.11142121435816608115E-17,
+                        1.55363195773620046921E-16, -1.10559694773538630805E-15,
+                        7.60068429473540693410E-15, -5.04218550472791168711E-14,
+                        3.22379336594557470981E-13, -1.98397439776494371520E-12,
+                        1.17361862988909016308E-11, -6.66348972350202774223E-11,
+                        3.62559028155211703701E-10, -1.88724975172282928790E-9,
+                        9.38153738649577178388E-9,  -4.44505912879632808065E-8,
+                        2.00329475355213526229E-7,  -8.56872026469545474066E-7,
+                        3.47025130813767847674E-6,  -1.32731636560394358279E-5,
+                        4.78156510755005422638E-5,  -1.61760815825896745588E-4,
+                        5.12285956168575772895E-4,  -1.51357245063125314899E-3,
+                        4.15642294431288815669E-3,  -1.05640848946261981558E-2,
+                        2.47264490306265168283E-2,  -5.29459812080949914269E-2,
+                        1.02643658689847095384E-1,  -1.76416518357834055153E-1,
+                        2.52587186443633654823E-1};
+    const double B[] = {
+        7.51729631084210481353E-18,  4.41434832307170791151E-18,
+        -4.65030536848935832153E-17, -3.20952592199342395980E-17,
+        2.96262899764595013876E-16,  3.30820231092092828324E-16,
+        -1.88035477551078244854E-15, -3.81440307243700780478E-15,
+        1.04202769841288027642E-14,  4.27244001671195135429E-14,
+        -2.10154184277266431302E-14, -4.08355111109219731823E-13,
+        -7.19855177624590851209E-13, 2.03562854414708950722E-12,
+        1.41258074366137813316E-11,  3.25260358301548823856E-11,
+        -1.89749581235054123450E-11, -5.58974346219658380687E-10,
+        -3.83538038596423702205E-9,  -2.63146884688951950684E-8,
+        -2.51223623787020892529E-7,  -3.88256480887769039346E-6,
+        -1.10588938762623716291E-4,  -9.76109749136146840777E-3,
+        7.78576235018280120474E-1};
+
+    double z = numext::abs(x);
+
+    if (z <= 8.0) {
+      double y = (z / 2.0) - 2.0;
+      z = cephes::chebevl<double, 29>::run(y, A) * z;
+    } else {
+      z = cephes::chebevl<double, 25>::run(32.0 / z - 2.0, B) / numext::sqrt(z);
+    }
+
+    if (x < 0.0) {
+      z = -z;
+    }
+
+    return z;
+  }
+};
+
 #endif  // EIGEN_HAS_C99_MATH
 
 }  // end namespace internal
@@ -1565,6 +1962,18 @@ EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(betainc, Scalar)
   return EIGEN_MATHFUNC_IMPL(betainc, Scalar)::run(a, b, x);
 }
 
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(i0e, Scalar)
+    i0e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(i0e, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(i1e, Scalar)
+    i1e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(i1e, Scalar)::run(x);
+}
+
 }  // end namespace numext
 
 
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
index 46d60d323..4c176716b 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
@@ -50,6 +50,18 @@ Packet pigammac(const Packet& a, const Packet& x) { using numext::igammac; retur
 template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 Packet pbetainc(const Packet& a, const Packet& b,const Packet& x) { using numext::betainc; return betainc(a, b, x); }
 
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero i0e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pi0e(const Packet& x) { using numext::i0e; return i0e(x); }
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one i1e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pi1e(const Packet& x) { using numext::i1e; return i1e(x); }
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
index e0e3a8be6..c25fea0b3 100644
--- a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
+++ b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
@@ -156,6 +156,32 @@ double2 pbetainc<double2>(const double2& a, const double2& b, const double2& x)
   return make_double2(betainc(a.x, b.x, x.x), betainc(a.y, b.y, x.y));
 }
 
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pi0e<float4>(const float4& x) {
+  using numext::i0e;
+  return make_float4(i0e(x.x), i0e(x.y), i0e(x.z), i0e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pi0e<double2>(const double2& x) {
+  using numext::i0e;
+  return make_double2(i0e(x.x), i0e(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pi1e<float4>(const float4& x) {
+  using numext::i1e;
+  return make_float4(i1e(x.x), i1e(x.y), i1e(x.z), i1e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pi1e<double2>(const double2& x) {
+  using numext::i1e;
+  return make_double2(i1e(x.x), i1e(x.y));
+}
+
 #endif
 
 } // end namespace internal
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index 9584a539f..63d0a345a 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -1208,6 +1208,116 @@ void test_cuda_betainc()
   cudaFree(d_out);
 }
 
+template <typename Scalar>
+void test_cuda_i0e()
+{
+  Tensor<Scalar, 1> in_x(21);
+  Tensor<Scalar, 1> out(21);
+  Tensor<Scalar, 1> expected_out(21);
+  out.setZero();
+
+  Array<Scalar, 1, Dynamic> in_x_array(21);
+  Array<Scalar, 1, Dynamic> expected_out_array(21);
+
+  in_x_array << -20.0, -18.0, -16.0, -14.0, -12.0, -10.0, -8.0, -6.0, -4.0,
+      -2.0, 0.0, 2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0;
+
+  expected_out_array << 0.0897803118848, 0.0947062952128, 0.100544127361,
+      0.107615251671, 0.116426221213, 0.127833337163, 0.143431781857,
+      0.16665743264, 0.207001921224, 0.308508322554, 1.0, 0.308508322554,
+      0.207001921224, 0.16665743264, 0.143431781857, 0.127833337163,
+      0.116426221213, 0.107615251671, 0.100544127361, 0.0947062952128,
+      0.0897803118848;
+
+  for (int i = 0; i < 21; ++i) {
+    in_x(i) = in_x_array(i);
+    expected_out(i) = expected_out_array(i);
+  }
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  cudaMalloc((void**)(&d_in), bytes);
+  cudaMalloc((void**)(&d_out), bytes);
+
+  cudaMemcpy(d_in, in_x.data(), bytes, cudaMemcpyHostToDevice);
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 21);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 21);
+
+  gpu_out.device(gpu_device) = gpu_in.i0e();
+
+  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
+                         gpu_device.stream()) == cudaSuccess);
+  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+
+  for (int i = 0; i < 21; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+
+  cudaFree(d_in);
+  cudaFree(d_out);
+}
+
+template <typename Scalar>
+void test_cuda_i1e()
+{
+  Tensor<Scalar, 1> in_x(21);
+  Tensor<Scalar, 1> out(21);
+  Tensor<Scalar, 1> expected_out(21);
+  out.setZero();
+
+  Array<Scalar, 1, Dynamic> in_x_array(21);
+  Array<Scalar, 1, Dynamic> expected_out_array(21);
+
+  in_x_array << -20.0, -18.0, -16.0, -14.0, -12.0, -10.0, -8.0, -6.0, -4.0,
+      -2.0, 0.0, 2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0;
+
+  expected_out_array << -0.0875062221833, -0.092036796872, -0.0973496147565,
+      -0.103697667463, -0.11146429929, -0.121262681384, -0.134142493293,
+      -0.152051459309, -0.178750839502, -0.215269289249, 0.0, 0.215269289249,
+      0.178750839502, 0.152051459309, 0.134142493293, 0.121262681384,
+      0.11146429929, 0.103697667463, 0.0973496147565, 0.092036796872,
+      0.0875062221833;
+
+  for (int i = 0; i < 21; ++i) {
+    in_x(i) = in_x_array(i);
+    expected_out(i) = expected_out_array(i);
+  }
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  cudaMalloc((void**)(&d_in), bytes);
+  cudaMalloc((void**)(&d_out), bytes);
+
+  cudaMemcpy(d_in, in_x.data(), bytes, cudaMemcpyHostToDevice);
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 21);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 21);
+
+  gpu_out.device(gpu_device) = gpu_in.i1e();
+
+  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
+                         gpu_device.stream()) == cudaSuccess);
+  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+
+  for (int i = 0; i < 21; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+
+  cudaFree(d_in);
+  cudaFree(d_out);
+}
+
 
 void test_cxx11_tensor_cuda()
 {
@@ -1280,5 +1390,11 @@ void test_cxx11_tensor_cuda()
 
   CALL_SUBTEST_6(test_cuda_betainc<float>());
   CALL_SUBTEST_6(test_cuda_betainc<double>());
+
+  CALL_SUBTEST_6(test_cuda_i0e<float>());
+  CALL_SUBTEST_6(test_cuda_i0e<double>());
+
+  CALL_SUBTEST_6(test_cuda_i1e<float>());
+  CALL_SUBTEST_6(test_cuda_i1e<double>());
 #endif
 }
diff --git a/unsupported/test/special_functions.cpp b/unsupported/test/special_functions.cpp
index 057fb3e92..48d0db95e 100644
--- a/unsupported/test/special_functions.cpp
+++ b/unsupported/test/special_functions.cpp
@@ -335,6 +335,46 @@ template<typename ArrayType> void array_special_functions()
         ArrayType test = betainc(a, b + one, x) + eps;
         verify_component_wise(test, expected););
   }
+
+  // Test Bessel function i0e. Reference results obtained with SciPy.
+  {
+    ArrayType x(21);
+    ArrayType expected(21);
+    ArrayType res(21);
+
+    x << -20.0, -18.0, -16.0, -14.0, -12.0, -10.0, -8.0, -6.0, -4.0, -2.0, 0.0,
+        2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0;
+
+    expected << 0.0897803118848, 0.0947062952128, 0.100544127361,
+        0.107615251671, 0.116426221213, 0.127833337163, 0.143431781857,
+        0.16665743264, 0.207001921224, 0.308508322554, 1.0, 0.308508322554,
+        0.207001921224, 0.16665743264, 0.143431781857, 0.127833337163,
+        0.116426221213, 0.107615251671, 0.100544127361, 0.0947062952128,
+        0.0897803118848;
+
+    CALL_SUBTEST(res = i0e(x);
+                 verify_component_wise(res, expected););
+  }
+
+  // Test Bessel function i1e. Reference results obtained with SciPy.
+  {
+    ArrayType x(21);
+    ArrayType expected(21);
+    ArrayType res(21);
+
+    x << -20.0, -18.0, -16.0, -14.0, -12.0, -10.0, -8.0, -6.0, -4.0, -2.0, 0.0,
+        2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0;
+
+    expected << -0.0875062221833, -0.092036796872, -0.0973496147565,
+        -0.103697667463, -0.11146429929, -0.121262681384, -0.134142493293,
+        -0.152051459309, -0.178750839502, -0.215269289249, 0.0, 0.215269289249,
+        0.178750839502, 0.152051459309, 0.134142493293, 0.121262681384,
+        0.11146429929, 0.103697667463, 0.0973496147565, 0.092036796872,
+        0.0875062221833;
+
+    CALL_SUBTEST(res = i1e(x);
+                 verify_component_wise(res, expected););
+  }
 #endif
 }
 

From 84868da904ac8d07342983b9bb78cf3360142363 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 31 May 2018 21:21:57 +0200
Subject: [PATCH 356/680] Don't run hg on non mercurial clone

---
 CMakeLists.txt | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5245b2930..91545bdb0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -41,10 +41,13 @@ string(REGEX MATCH "define[ \t]+EIGEN_MINOR_VERSION[ \t]+([0-9]+)" _eigen_minor_
 set(EIGEN_MINOR_VERSION "${CMAKE_MATCH_1}")
 set(EIGEN_VERSION_NUMBER ${EIGEN_WORLD_VERSION}.${EIGEN_MAJOR_VERSION}.${EIGEN_MINOR_VERSION})
 
-# if the mercurial program is absent, this will leave the EIGEN_HG_CHANGESET string empty,
-# but won't stop CMake.
-execute_process(COMMAND hg tip -R ${CMAKE_SOURCE_DIR} OUTPUT_VARIABLE EIGEN_HGTIP_OUTPUT)
-execute_process(COMMAND hg branch -R ${CMAKE_SOURCE_DIR} OUTPUT_VARIABLE EIGEN_BRANCH_OUTPUT)
+# if we are not in a mercurial clone
+if(IS_DIRECTORY ${CMAKE_SOURCE_DIR}/.hg)
+  # if the mercurial program is absent or this will leave the EIGEN_HG_CHANGESET string empty,
+  # but won't stop CMake.
+  execute_process(COMMAND hg tip -R ${CMAKE_SOURCE_DIR} OUTPUT_VARIABLE EIGEN_HGTIP_OUTPUT)
+  execute_process(COMMAND hg branch -R ${CMAKE_SOURCE_DIR} OUTPUT_VARIABLE EIGEN_BRANCH_OUTPUT)
+endif()
 
 # if this is the default (aka development) branch, extract the mercurial changeset number from the hg tip output...
 if(EIGEN_BRANCH_OUTPUT MATCHES "default")

From e2ed0cf8abfe275bd147f3228e4c59f5242993f6 Mon Sep 17 00:00:00 2001
From: Penporn Koanantakool <penporn@google.com>
Date: Sat, 2 Jun 2018 12:07:49 -0700
Subject: [PATCH 357/680] Add a ThreadPoolInterface* getter for
 ThreadPoolDevice.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index ec6802e85..ca9ba402e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -169,7 +169,7 @@ struct ThreadPoolDevice {
 
   // parallelFor executes f with [0, n) arguments in parallel and waits for
   // completion. F accepts a half-open interval [first, last).
-  // Block size is choosen based on the iteration cost and resulting parallel
+  // Block size is chosen based on the iteration cost and resulting parallel
   // efficiency. If block_align is not nullptr, it is called to round up the
   // block size.
   void parallelFor(Index n, const TensorOpCost& cost,
@@ -261,6 +261,9 @@ struct ThreadPoolDevice {
     parallelFor(n, cost, nullptr, std::move(f));
   }
 
+  // Thread pool accessor.
+  ThreadPoolInterface* getPool() const { return pool_; }
+
  private:
   ThreadPoolInterface* pool_;
   int num_threads_;

From 8fbd47052bcafea612b8ae2841c1de5db738f042 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 6 Jun 2018 10:12:58 -0400
Subject: [PATCH 358/680] Adding support for using Eigen in HIP kernels.

This commit enables the use of Eigen on HIP kernels / AMD GPUs. Support has been added along the same lines as what already exists for using Eigen in CUDA kernels / NVidia GPUs.

Application code needs to explicitly define EIGEN_USE_HIP when using Eigen in HIP kernels. This is because some of the CUDA headers get picked up by default during Eigen compile (irrespective of whether or not the underlying compiler is CUDACC/NVCC, for e.g. Eigen/src/Core/arch/CUDA/Half.h). In order to maintain this behavior, the EIGEN_USE_HIP macro is used to switch to using the HIP version of those header files (see Eigen/Core and unsupported/Eigen/CXX11/Tensor)


Use the "-DEIGEN_TEST_HIP" cmake option to enable the HIP specific unit tests.
---
 Eigen/Core                                    |   69 +-
 Eigen/src/Core/GenericPacketMath.h            |    8 +-
 Eigen/src/Core/MathFunctions.h                |   61 +-
 Eigen/src/Core/ProductEvaluators.h            |    6 +
 Eigen/src/Core/arch/HIP/hcc/Half.h            |  705 ++++++++
 Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h  | 1019 +++++++++++
 Eigen/src/Core/arch/HIP/hcc/TypeCasting.h     |  212 +++
 Eigen/src/Core/arch/HIP/hcc/math_constants.h  |   23 +
 Eigen/src/Core/functors/BinaryFunctors.h      |    6 +
 Eigen/src/Core/util/BlasUtil.h                |    5 +-
 Eigen/src/Core/util/Macros.h                  |    5 +-
 Eigen/src/Core/util/Memory.h                  |   38 +
 Eigen/src/Core/util/Meta.h                    |   52 +-
 .../src/Eigenvalues/SelfAdjointEigenSolver.h  |    1 +
 Eigen/src/SVD/BDCSVD.h                        |    2 +-
 cmake/EigenTesting.cmake                      |    9 +-
 test/CMakeLists.txt                           |   42 +
 test/hip_basic.cu                             |  172 ++
 test/hip_common.h                             |  103 ++
 test/main.h                                   |   22 +-
 unsupported/Eigen/CXX11/Tensor                |   40 +-
 .../CXX11/src/Tensor/TensorContraction.h      |   10 +-
 .../src/Tensor/TensorContractionBlocking.h    |    5 +-
 .../CXX11/src/Tensor/TensorContractionHip.h   | 1521 +++++++++++++++++
 .../CXX11/src/Tensor/TensorConvolutionHip.h   | 1119 ++++++++++++
 .../CXX11/src/Tensor/TensorDeviceDefault.h    |   15 +-
 .../Eigen/CXX11/src/Tensor/TensorDeviceHip.h  |  352 ++++
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |   16 +-
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |    5 +-
 .../Eigen/CXX11/src/Tensor/TensorIndexList.h  |    6 +-
 .../Eigen/CXX11/src/Tensor/TensorMacros.h     |    2 +-
 .../Eigen/CXX11/src/Tensor/TensorMeta.h       |    2 +-
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |    5 +-
 .../Eigen/CXX11/src/Tensor/TensorRandom.h     |    2 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |   24 +-
 .../CXX11/src/Tensor/TensorReductionHip.h     |  815 +++++++++
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |    9 +-
 unsupported/Eigen/CXX11/src/util/CXX11Meta.h  |   12 +
 .../Eigen/CXX11/src/util/EmulateArray.h       |    2 +-
 .../SpecialFunctions/SpecialFunctionsImpl.h   |    4 +-
 unsupported/test/CMakeLists.txt               |   52 +
 unsupported/test/cxx11_tensor_argmax_hip.cu   |  251 +++
 .../test/cxx11_tensor_cast_float16_hip.cu     |   79 +
 unsupported/test/cxx11_tensor_contract_hip.cu |  215 +++
 unsupported/test/cxx11_tensor_device_hip.cu   |  389 +++++
 unsupported/test/cxx11_tensor_hip.cu          | 1296 ++++++++++++++
 .../test/cxx11_tensor_of_float16_hip.cu       |  498 ++++++
 unsupported/test/cxx11_tensor_random_hip.cu   |   85 +
 .../test/cxx11_tensor_reduction_hip.cu        |  154 ++
 unsupported/test/cxx11_tensor_scan_hip.cu     |   76 +
 50 files changed, 9527 insertions(+), 94 deletions(-)
 create mode 100644 Eigen/src/Core/arch/HIP/hcc/Half.h
 create mode 100644 Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
 create mode 100644 Eigen/src/Core/arch/HIP/hcc/TypeCasting.h
 create mode 100644 Eigen/src/Core/arch/HIP/hcc/math_constants.h
 create mode 100644 test/hip_basic.cu
 create mode 100644 test/hip_common.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
 create mode 100644 unsupported/test/cxx11_tensor_argmax_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_cast_float16_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_contract_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_device_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_of_float16_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_random_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_reduction_hip.cu
 create mode 100644 unsupported/test/cxx11_tensor_scan_hip.cu

diff --git a/Eigen/Core b/Eigen/Core
index f6bc18a08..c72d5468a 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -22,6 +22,17 @@
   #define EIGEN_CUDA_ARCH __CUDA_ARCH__
 #endif
 
+#if defined(__HIPCC__) && !defined(EIGEN_NO_HIP)
+  // analogous to EIGEN_CUDACC, but for HIP
+  #define EIGEN_HIPCC __HIPCC__
+#endif
+
+// NVCC is not supported as the target platform for HIPCC
+// Note that this also makes EIGEN_CUDACC and EIGEN_HIPCC mutually exclusive
+#if defined(__NVCC__) && defined(__HIPCC__)
+  #error "NVCC as the target platform for HIPCC is currently not supported."
+#endif
+
 // Starting with CUDA 9 the composite __CUDACC_VER__ is not available.
 #if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
 #define EIGEN_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
@@ -32,8 +43,8 @@
 #endif
 
 // Handle NVCC/CUDA/SYCL
-#if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__)
-  // Do not try asserts on CUDA and SYCL!
+#if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__) || defined(EIGEN_HIPCC)
+  // Do not try asserts on CUDA, HIP and SYCL!
   #ifndef EIGEN_NO_DEBUG
   #define EIGEN_NO_DEBUG
   #endif
@@ -57,6 +68,26 @@
     // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
     // works properly on the device side
     #include <cuda_runtime.h>
+
+  #elif defined(EIGEN_HIPCC)
+    // Do not try to vectorize on HIP
+    #ifndef EIGEN_DONT_VECTORIZE
+    #define EIGEN_DONT_VECTORIZE
+    #endif
+
+    #define EIGEN_DEVICE_FUNC __host__ __device__
+    // We need hip_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
+    // works properly on the device side
+    #include <hip/hip_runtime.h>
+    
+   #if defined(__HIP_DEVICE_COMPILE__) && !defined(EIGEN_NO_HIP)
+     // analogous to EIGEN_CUDA_ARCH, but for HIP
+     #define EIGEN_HIP_DEVICE_COMPILE __HIP_DEVICE_COMPILE__
+     // Note this check needs to come after we include hip_runtime.h since
+     // hip_runtime.h includes hip_common.h which in turn has the define
+     // for __HIP_DEVICE_COMPILE__
+   #endif
+
   #else
     #define EIGEN_DEVICE_FUNC
   #endif
@@ -68,16 +99,16 @@
 #define EIGEN_DONT_VECTORIZE
 #endif
 
-// When compiling CUDA device code with NVCC, pull in math functions from the
-// global namespace.  In host mode, and when device doee with clang, use the
-// std versions.
-#if defined(EIGEN_CUDA_ARCH) && defined(__NVCC__)
+// When compiling CUDA device code with NVCC, or HIP device code with HIPCC
+// pull in math functions from the global namespace.  In host mode, and when
+// device doee with clang, use the std versions.
+#if (defined(EIGEN_CUDA_ARCH) && defined(__NVCC__)) || (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIPCC__))
   #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
 #else
   #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
 #endif
 
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL)
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL) && !defined(EIGEN_HIP_DEVICE_COMPILE)
   #define EIGEN_EXCEPTIONS
 #endif
 
@@ -270,6 +301,17 @@
   #include <cuda_fp16.h>
 #endif
 
+#if defined(EIGEN_HIPCC) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  #define EIGEN_HAS_HIP_FP16
+  #include <hip/hip_fp16.h>
+  #define HIP_PATCH_WITH_NEW_FP16 18215
+  #if (HIP_VERSION_PATCH < HIP_PATCH_WITH_NEW_FP16)
+    #define EIGEN_HAS_OLD_HIP_FP16
+    // Old HIP implementation does not have a explicit typedef for "half2"
+    typedef __half2 half2;
+  #endif
+#endif
+
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
   #define EIGEN_HAS_OPENMP
 #endif
@@ -390,7 +432,6 @@ using std::ptrdiff_t;
 #include "src/Core/util/IntegralConstant.h"
 #include "src/Core/util/SymbolicIndex.h"
 
-
 #include "src/Core/NumTraits.h"
 #include "src/Core/MathFunctions.h"
 #include "src/Core/GenericPacketMath.h"
@@ -434,9 +475,15 @@ using std::ptrdiff_t;
 #endif
 
 // Half float support
-#include "src/Core/arch/CUDA/Half.h"
-#include "src/Core/arch/CUDA/PacketMathHalf.h"
-#include "src/Core/arch/CUDA/TypeCasting.h"
+#if defined EIGEN_USE_HIP
+  #include "src/Core/arch/HIP/hcc/Half.h"
+  #include "src/Core/arch/HIP/hcc/PacketMathHalf.h"
+  #include "src/Core/arch/HIP/hcc/TypeCasting.h"
+#else
+  #include "src/Core/arch/CUDA/Half.h"
+  #include "src/Core/arch/CUDA/PacketMathHalf.h"
+  #include "src/Core/arch/CUDA/TypeCasting.h"
+#endif
 
 #if defined EIGEN_VECTORIZE_CUDA
   #include "src/Core/arch/CUDA/PacketMath.h"
diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 888a3f7ea..0903c3a6e 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -299,7 +299,11 @@ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu
  { pstore(to, from); }
 
 /** \internal tries to do cache prefetching of \a addr */
-template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
+template<typename Scalar>
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+  inline void prefetch(const Scalar* addr)
 {
 #ifdef EIGEN_CUDA_ARCH
 #if defined(__LP64__)
@@ -528,7 +532,7 @@ inline void palign(PacketType& first, const PacketType& second)
 ***************************************************************************/
 
 // Eigen+CUDA does not support complexes.
-#ifndef EIGEN_CUDACC
+#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
 
 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
 { return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 05462c5e1..6beef5def 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -96,7 +96,7 @@ struct real_default_impl<Scalar,true>
 
 template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
 
-#ifdef EIGEN_CUDA_ARCH
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 template<typename T>
 struct real_impl<std::complex<T> >
 {
@@ -144,7 +144,7 @@ struct imag_default_impl<Scalar,true>
 
 template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
 
-#ifdef EIGEN_CUDA_ARCH
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 template<typename T>
 struct imag_impl<std::complex<T> >
 {
@@ -260,7 +260,7 @@ struct conj_default_impl<Scalar,true>
 
 template<typename Scalar> struct conj_impl : conj_default_impl<Scalar> {};
 
-#ifdef EIGEN_CUDA_ARCH
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 template<typename T>
 struct conj_impl<std::complex<T> >
 {
@@ -435,7 +435,12 @@ struct round_retval
   struct arg_impl {
     static inline Scalar run(const Scalar& x)
     {
+      #if defined(EIGEN_HIP_DEVICE_COMPILE)
+      // HIP does not seem to have a native device side implementation for the math routine "arg"
+      using std::arg;
+      #else 		  
       EIGEN_USING_STD_MATH(arg);
+      #endif
       return arg(x);
     }
   };
@@ -768,7 +773,9 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isfinite_impl(const T& x)
 {
-  #ifdef EIGEN_CUDA_ARCH
+  #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    return isfinite(x);
+  #elif defined(EIGEN_CUDA_ARCH)
     return (::isfinite)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isfinite;
@@ -783,7 +790,9 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isinf_impl(const T& x)
 {
-  #ifdef EIGEN_CUDA_ARCH
+  #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    return isinf(x);
+  #elif defined(EIGEN_CUDA_ARCH)
     return (::isinf)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isinf;
@@ -798,7 +807,9 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isnan_impl(const T& x)
 {
-  #ifdef EIGEN_CUDA_ARCH
+  #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    return isnan(x);
+  #elif defined(EIGEN_CUDA_ARCH)
     return (::isnan)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isnan;
@@ -864,7 +875,7 @@ template<typename T> T generic_fast_tanh_float(const T& a_x);
 
 namespace numext {
 
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE) && !defined(__SYCL_DEVICE_ONLY__)
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
@@ -1078,7 +1089,7 @@ EIGEN_ALWAYS_INLINE float   log1p(float x) { return cl::sycl::log1p(x); }
 EIGEN_ALWAYS_INLINE double  log1p(double x) { return cl::sycl::log1p(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log1p(const float &x) { return ::log1pf(x); }
 
@@ -1136,7 +1147,7 @@ EIGEN_ALWAYS_INLINE float   floor(float x) { return cl::sycl::floor(x); }
 EIGEN_ALWAYS_INLINE double  floor(double x) { return cl::sycl::floor(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float floor(const float &x) { return ::floorf(x); }
 
@@ -1157,7 +1168,7 @@ EIGEN_ALWAYS_INLINE float   ceil(float x) { return cl::sycl::ceil(x); }
 EIGEN_ALWAYS_INLINE double  ceil(double x) { return cl::sycl::ceil(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float ceil(const float &x) { return ::ceilf(x); }
 
@@ -1215,7 +1226,7 @@ EIGEN_ALWAYS_INLINE double  log(double x) { return cl::sycl::log(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log(const float &x) { return ::logf(x); }
 
@@ -1243,7 +1254,7 @@ EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
 EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float abs(const float &x) { return ::fabsf(x); }
 
@@ -1273,7 +1284,7 @@ EIGEN_ALWAYS_INLINE float   exp(float x) { return cl::sycl::exp(x); }
 EIGEN_ALWAYS_INLINE double  exp(double x) { return cl::sycl::exp(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float exp(const float &x) { return ::expf(x); }
 
@@ -1309,7 +1320,7 @@ EIGEN_ALWAYS_INLINE float   expm1(float x) { return cl::sycl::expm1(x); }
 EIGEN_ALWAYS_INLINE double  expm1(double x) { return cl::sycl::expm1(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float expm1(const float &x) { return ::expm1f(x); }
 
@@ -1329,7 +1340,7 @@ EIGEN_ALWAYS_INLINE float   cos(float x) { return cl::sycl::cos(x); }
 EIGEN_ALWAYS_INLINE double  cos(double x) { return cl::sycl::cos(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cos(const float &x) { return ::cosf(x); }
 
@@ -1349,7 +1360,7 @@ EIGEN_ALWAYS_INLINE float   sin(float x) { return cl::sycl::sin(x); }
 EIGEN_ALWAYS_INLINE double  sin(double x) { return cl::sycl::sin(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sin(const float &x) { return ::sinf(x); }
 
@@ -1369,7 +1380,7 @@ EIGEN_ALWAYS_INLINE float   tan(float x) { return cl::sycl::tan(x); }
 EIGEN_ALWAYS_INLINE double  tan(double x) { return cl::sycl::tan(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tan(const float &x) { return ::tanf(x); }
 
@@ -1400,7 +1411,7 @@ EIGEN_ALWAYS_INLINE float   acosh(float x) { return cl::sycl::acosh(x); }
 EIGEN_ALWAYS_INLINE double  acosh(double x) { return cl::sycl::acosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float acos(const float &x) { return ::acosf(x); }
 
@@ -1431,7 +1442,7 @@ EIGEN_ALWAYS_INLINE float   asinh(float x) { return cl::sycl::asinh(x); }
 EIGEN_ALWAYS_INLINE double  asinh(double x) { return cl::sycl::asinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float asin(const float &x) { return ::asinf(x); }
 
@@ -1462,7 +1473,7 @@ EIGEN_ALWAYS_INLINE float   atanh(float x) { return cl::sycl::atanh(x); }
 EIGEN_ALWAYS_INLINE double  atanh(double x) { return cl::sycl::atanh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float atan(const float &x) { return ::atanf(x); }
 
@@ -1483,7 +1494,7 @@ EIGEN_ALWAYS_INLINE float   cosh(float x) { return cl::sycl::cosh(x); }
 EIGEN_ALWAYS_INLINE double  cosh(double x) { return cl::sycl::cosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cosh(const float &x) { return ::coshf(x); }
 
@@ -1503,7 +1514,7 @@ EIGEN_ALWAYS_INLINE float   sinh(float x) { return cl::sycl::sinh(x); }
 EIGEN_ALWAYS_INLINE double  sinh(double x) { return cl::sycl::sinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sinh(const float &x) { return ::sinhf(x); }
 
@@ -1521,12 +1532,12 @@ T tanh(const T &x) {
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   tanh(float x) { return cl::sycl::tanh(x); }
 EIGEN_ALWAYS_INLINE double  tanh(double x) { return cl::sycl::tanh(x); }
-#elif (!defined(EIGEN_CUDACC)) && EIGEN_FAST_MATH
+#elif (!defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)) && EIGEN_FAST_MATH
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(float x) { return internal::generic_fast_tanh_float(x); }
 #endif
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(const float &x) { return ::tanhf(x); }
 
@@ -1546,7 +1557,7 @@ EIGEN_ALWAYS_INLINE float   fmod(float x, float y) { return cl::sycl::fmod(x, y)
 EIGEN_ALWAYS_INLINE double  fmod(double x, double y) { return cl::sycl::fmod(x, y); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#ifdef EIGEN_CUDACC
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template <>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float fmod(const float& a, const float& b) {
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index fb637191d..cc75fbce3 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -137,6 +137,9 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scal
   typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
 {
   typedef Product<Lhs,Rhs,Options> SrcXprType;
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif  
   static EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
@@ -390,6 +393,9 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   template<typename Dst>
+  #if defined(EIGEN_HIPCC)  
+  EIGEN_DEVICE_FUNC
+  #endif  
   static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     // Same as: dst.noalias() = lhs.lazyProduct(rhs);
diff --git a/Eigen/src/Core/arch/HIP/hcc/Half.h b/Eigen/src/Core/arch/HIP/hcc/Half.h
new file mode 100644
index 000000000..2ce8a412c
--- /dev/null
+++ b/Eigen/src/Core/arch/HIP/hcc/Half.h
@@ -0,0 +1,705 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+//
+// The conversion routines are Copyright (c) Fabian Giesen, 2016.
+// The original license follows:
+//
+// Copyright (c) Fabian Giesen, 2016
+// All rights reserved.
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted.
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+// Standard 16-bit float type, mostly useful for GPUs. Defines a new
+// type Eigen::half (inheriting from HIP's __half struct) with
+// operator overloads such that it behaves basically as an arithmetic
+// type. It will be quite slow on CPUs (so it is recommended to stay
+// in fp32 for CPUs, except for simple parameter conversions, I/O
+// to disk and the likes), but fast on GPUs.
+
+
+#ifndef EIGEN_HALF_HIP_H
+#define EIGEN_HALF_HIP_H
+
+#if __cplusplus > 199711L
+#define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
+#else
+#define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type()
+#endif
+
+
+namespace Eigen {
+
+struct half;
+
+namespace half_impl {
+
+#if !defined(EIGEN_HAS_HIP_FP16)
+// Make our own __half_raw definition that is similar to CUDA's.
+struct __half_raw {
+  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
+  unsigned short x;
+};
+#elif defined(EIGEN_HAS_OLD_HIP_FP16)
+// Make a __half_raw definition that is
+// ++ compatible with that of Eigen and
+// ++ add a implcit conversion to the native __half of the old HIP implementation.
+//
+// Keeping ".x" as "unsigned short" keeps the interface the same between the Eigen and HIP implementation.
+//
+// In the old HIP implementation,
+//   ++ __half is a typedef of __fp16
+//   ++ the "__h*" routines take "__half" arguments
+// so we need to implicitly convert "__half_raw" to "__half" to avoid having to explicitly make 
+// that conversiion in each call to a "__h*" routine...that is why we have "operator __half" routine
+struct __half_raw {
+  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
+  union {
+    unsigned short x;
+    __half data;
+  };
+  operator __half(void) const { return data; }
+};
+#endif
+ 
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
+
+struct half_base : public __half_raw {
+  EIGEN_DEVICE_FUNC half_base() {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
+#if defined(EIGEN_HAS_HIP_FP16)
+  #if defined(EIGEN_HAS_OLD_HIP_FP16)
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(__half_as_ushort(h)) {}
+  #else
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
+  #endif
+#endif
+};
+
+} // namespace half_impl
+
+// Class definition.
+struct half : public half_impl::half_base {
+  #if !defined(EIGEN_HAS_HIP_FP16) || defined(EIGEN_HAS_OLD_HIP_FP16)
+    typedef half_impl::__half_raw __half_raw;
+  #endif
+
+  EIGEN_DEVICE_FUNC half() {}
+
+  EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
+#if defined(EIGEN_HAS_HIP_FP16)
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+#endif
+
+  explicit EIGEN_DEVICE_FUNC half(bool b)
+      : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
+  template<class T>
+  explicit EIGEN_DEVICE_FUNC half(const T& val)
+      : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
+  explicit EIGEN_DEVICE_FUNC half(float f)
+      : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
+    // +0.0 and -0.0 become false, everything else becomes true.
+    return (x & 0x7fff) != 0;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
+    return static_cast<signed char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
+    return static_cast<unsigned char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
+    return static_cast<short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
+    return static_cast<unsigned short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
+    return static_cast<int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
+    return static_cast<unsigned int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
+    return static_cast<long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
+    return static_cast<unsigned long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
+    return static_cast<long long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
+    return static_cast<unsigned long long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
+    return half_impl::half_to_float(*this);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
+    return static_cast<double>(half_impl::half_to_float(*this));
+  }
+
+  EIGEN_DEVICE_FUNC half& operator=(const half& other) {
+    x = other.x;
+    return *this;
+  }
+};
+
+} // end namespace Eigen
+
+namespace std {
+template<>
+struct numeric_limits<Eigen::half> {
+  static const bool is_specialized = true;
+  static const bool is_signed = true;
+  static const bool is_integer = false;
+  static const bool is_exact = false;
+  static const bool has_infinity = true;
+  static const bool has_quiet_NaN = true;
+  static const bool has_signaling_NaN = true;
+  static const float_denorm_style has_denorm = denorm_present;
+  static const bool has_denorm_loss = false;
+  static const std::float_round_style round_style = std::round_to_nearest;
+  static const bool is_iec559 = false;
+  static const bool is_bounded = false;
+  static const bool is_modulo = false;
+  static const int digits = 11;
+  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int radix = 2;
+  static const int min_exponent = -13;
+  static const int min_exponent10 = -4;
+  static const int max_exponent = 16;
+  static const int max_exponent10 = 4;
+  static const bool traps = true;
+  static const bool tinyness_before = false;
+
+  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
+  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
+  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
+  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
+  static Eigen::half round_error() { return Eigen::half(0.5); }
+  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
+  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
+};
+
+// If std::numeric_limits<T> is specialized, should also specialize
+// std::numeric_limits<const T>, std::numeric_limits<volatile T>, and
+// std::numeric_limits<const volatile T>
+// https://stackoverflow.com/a/16519653/
+template<>
+struct numeric_limits<const Eigen::half> : numeric_limits<Eigen::half> {};
+template<>
+struct numeric_limits<volatile Eigen::half> : numeric_limits<Eigen::half> {};
+template<>
+struct numeric_limits<const volatile Eigen::half> : numeric_limits<Eigen::half> {};
+} // end namespace std
+
+namespace Eigen {
+
+namespace half_impl {
+
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+
+// Intrinsics for native fp16 support. Note that on current hardware,
+// these are no faster than fp32 arithmetic (you need to use the half2
+// versions to get the ALU speed increased), but you do save the
+// conversion steps back and forth.
+
+EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
+  return __hadd(a, b);
+}
+EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
+  return __hmul(a, b);
+}
+EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
+  return __hsub(a, b);
+}
+EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
+  float num = __half2float(a);
+  float denom = __half2float(b);
+  return __float2half(num / denom);
+}
+EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
+  return __hneg(a);
+}
+EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) {
+  a = a + b;
+  return a;
+}
+EIGEN_STRONG_INLINE __device__ half& operator *= (half& a, const half& b) {
+  a = a * b;
+  return a;
+}
+EIGEN_STRONG_INLINE __device__ half& operator -= (half& a, const half& b) {
+  a = a - b;
+  return a;
+}
+EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) {
+  a = a / b;
+  return a;
+}
+EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) {
+  return __heq(a, b);
+}
+EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) {
+  return __hne(a, b);
+}
+EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) {
+  return __hlt(a, b);
+}
+EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) {
+  return __hle(a, b);
+}
+EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) {
+  return __hgt(a, b);
+}
+EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
+  return __hge(a, b);
+}
+
+#else  // Emulate support for half floats
+
+// Definitions for CPUs mostly working through conversion to/from fp32.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
+  return half(float(a) + float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
+  return half(float(a) * float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
+  return half(float(a) - float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
+  return half(float(a) / float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
+  half result;
+  result.x = a.x ^ 0x8000;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
+  a = half(float(a) + float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
+  a = half(float(a) * float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
+  a = half(float(a) - float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
+  a = half(float(a) / float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
+  return numext::equal_strict(float(a),float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
+  return numext::not_equal_strict(float(a), float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
+  return float(a) < float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
+  return float(a) <= float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
+  return float(a) > float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
+  return float(a) >= float(b);
+}
+
+#endif  // Emulate support for half floats
+
+// Division by an index. Do it in full float precision to avoid accuracy
+// issues in converting the denominator to half.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
+  return half(static_cast<float>(a) / static_cast<float>(b));
+}
+
+// Conversion routines, including fallbacks for the host or older CUDA.
+// Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
+// these in hardware. If we need more performance on older/other CPUs, they are
+// also possible to vectorize directly.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x) {
+  __half_raw h;
+  h.x = x;
+  return h;
+}
+
+union FP32 {
+  unsigned int u;
+  float f;
+};
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  __half tmp_ff = __float2half(ff);
+  #if defined(EIGEN_HAS_OLD_HIP_FP16)
+    __half_raw h;
+    h.data = tmp_ff;
+    return h;
+  #else
+    return *(__half_raw*)&tmp_ff;
+  #endif
+
+#elif defined(EIGEN_HAS_FP16_C)
+  __half_raw h;
+  h.x = _cvtss_sh(ff, 0);
+  return h;
+
+#else
+  FP32 f; f.f = ff;
+
+  const FP32 f32infty = { 255 << 23 };
+  const FP32 f16max = { (127 + 16) << 23 };
+  const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
+  unsigned int sign_mask = 0x80000000u;
+  __half_raw o;
+  o.x = static_cast<unsigned short>(0x0u);
+
+  unsigned int sign = f.u & sign_mask;
+  f.u ^= sign;
+
+  // NOTE all the integer compares in this function can be safely
+  // compiled into signed compares since all operands are below
+  // 0x80000000. Important if you want fast straight SSE2 code
+  // (since there's no unsigned PCMPGTD).
+
+  if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
+    o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
+  } else {  // (De)normalized number or zero
+    if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
+      // use a magic value to align our 10 mantissa bits at the bottom of
+      // the float. as long as FP addition is round-to-nearest-even this
+      // just works.
+      f.f += denorm_magic.f;
+
+      // and one integer subtract of the bias later, we have our final float!
+      o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
+    } else {
+      unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
+
+      // update exponent, rounding bias part 1
+      f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
+      // rounding bias part 2
+      f.u += mant_odd;
+      // take the bits!
+      o.x = static_cast<unsigned short>(f.u >> 13);
+    }
+  }
+
+  o.x |= static_cast<unsigned short>(sign >> 16);
+  return o;
+#endif
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return __half2float(h);
+
+#elif defined(EIGEN_HAS_FP16_C)
+  return _cvtsh_ss(h.x);
+
+#else
+  const FP32 magic = { 113 << 23 };
+  const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
+  FP32 o;
+
+  o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
+  unsigned int exp = shifted_exp & o.u;   // just the exponent
+  o.u += (127 - 15) << 23;                // exponent adjust
+
+  // handle exponent special cases
+  if (exp == shifted_exp) {     // Inf/NaN?
+    o.u += (128 - 16) << 23;    // extra exp adjust
+  } else if (exp == 0) {        // Zero/Denormal?
+    o.u += 1 << 23;             // extra exp adjust
+    o.f -= magic.f;             // renormalize
+  }
+
+  o.u |= (h.x & 0x8000) << 16;    // sign bit
+  return o.f;
+#endif
+}
+
+// --- standard functions ---
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
+  return (a.x & 0x7fff) == 0x7c00;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return __hisnan(a);
+#else
+  return (a.x & 0x7fff) > 0x7c00;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
+  return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
+  half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return half(hexp(a));
+#else
+   return half(::expf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
+  return half(numext::expm1(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return half(hlog(a));
+#else
+  return half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
+  return half(numext::log1p(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
+  return half(::log10f(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return half(hsqrt(a));
+#else
+    return half(::sqrtf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
+  return half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
+  return half(::sinf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
+  return half(::cosf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
+  return half(::tanf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
+  return half(::tanhf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return half(hfloor(a));
+#else
+  return half(::floorf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return half(hceil(a));
+#else
+  return half(::ceilf(float(a)));
+#endif
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return __hlt(b, a) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f2 < f1 ? b : a;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+  return __hlt(a, b) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f1 < f2 ? b : a;
+#endif
+}
+
+EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
+  os << static_cast<float>(v);
+  return os;
+}
+
+} // end namespace half_impl
+
+// import Eigen::half_impl::half into Eigen namespace
+// using half_impl::half;
+
+namespace internal {
+
+template<>
+struct random_default_impl<half, false, false>
+{
+  static inline half run(const half& x, const half& y)
+  {
+    return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
+  }
+  static inline half run()
+  {
+    return run(half(-1.f), half(1.f));
+  }
+};
+
+template<> struct is_arithmetic<half> { enum { value = true }; };
+
+} // end namespace internal
+
+template<> struct NumTraits<Eigen::half>
+    : GenericNumTraits<Eigen::half>
+{
+  enum {
+    IsSigned = true,
+    IsInteger = false,
+    IsComplex = false,
+    RequireInitialization = false
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
+    return half_impl::raw_uint16_to_half(0x0800);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
+    return half_impl::raw_uint16_to_half(0x7bff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
+    return half_impl::raw_uint16_to_half(0xfbff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
+    return half_impl::raw_uint16_to_half(0x7c00);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
+    return half_impl::raw_uint16_to_half(0x7c01);
+  }
+};
+
+} // end namespace Eigen
+
+// C-like standard mathematical functions and trancendentals.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) {
+  Eigen::half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
+  return Eigen::half(::expf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+    return Eigen::half(hlog(a));
+#else
+  return Eigen::half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
+  return Eigen::half(::sqrtf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) {
+  return Eigen::half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) {
+  return Eigen::half(::floorf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) {
+  return Eigen::half(::ceilf(float(a)));
+}
+
+namespace std {
+
+#if __cplusplus > 199711L
+template <>
+struct hash<Eigen::half> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
+    return static_cast<std::size_t>(a.x);
+  }
+};
+#endif
+
+} // end namespace std
+
+
+// Add the missing shfl_xor intrinsic
+#if defined(EIGEN_HAS_HIP_FP16) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
+__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+  // FIXME
+  //return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
+  return var;
+}
+#endif
+
+// ldg() has an overload for __half, but we also need one for Eigen::half.
+#if defined(EIGEN_HAS_HIP_FP16) && \
+    defined(__HIP_ARCH_HAS_WARP_FUNNEL_SHIFT__) && defined(__HIP_ARCH_HAS_DYNAMIC_PARALLEL__)
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
+  // FIXME
+  //return Eigen::half_impl::raw_uint16_to_half(
+  //    __ldg(reinterpret_cast<const unsigned short*>(ptr)));
+  return *ptr;
+}
+#endif
+
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+namespace Eigen {
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isnan)(const Eigen::half& h) {
+  return (half_impl::isnan)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isinf)(const Eigen::half& h) {
+  return (half_impl::isinf)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isfinite)(const Eigen::half& h) {
+  return (half_impl::isfinite)(h);
+}
+
+} // namespace Eigen
+}  // namespace numext
+#endif
+
+#endif // EIGEN_HALF_HIP_H
diff --git a/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h b/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
new file mode 100644
index 000000000..29c3f4671
--- /dev/null
+++ b/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
@@ -0,0 +1,1019 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_HALF_HIP_H
+#define EIGEN_PACKET_MATH_HALF_HIP_H
+
+
+namespace Eigen {
+namespace internal {
+
+// Most of the following operations require arch >= 3.0
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+
+template<> struct is_arithmetic<half2> { enum { value = true }; };
+
+template<> struct packet_traits<Eigen::half> : default_packet_traits
+{
+  typedef half2 type;
+  typedef half2 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+    HasAdd    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasSqrt   = 1,
+    HasRsqrt  = 1,
+    HasExp    = 1,
+    HasExpm1  = 1,
+    HasLog    = 1,
+    HasLog1p  = 1
+  };
+};
+
+template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return half2half2(from);
+#else
+  return __half2half2(from);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
+  return *reinterpret_cast<const half2*>(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
+  return __halves2half2(from[0], from[1]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
+  return __halves2half2(from[0], from[0]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
+  *reinterpret_cast<half2*>(to) = from;
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
+  to[0] = __low2half(from);
+  to[1] = __high2half(from);
+}
+
+template<>
+ __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return __halves2half2((*(from+0)), (*(from+1)));
+#else
+  return __ldg((const half2*)from);
+#endif
+}
+
+template<>
+__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return __halves2half2((*(from+0)), (*(from+1)));
+#else
+  return __halves2half2(__ldg(from+0), __ldg(from+1));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
+  return __halves2half2(from[0*stride], from[1*stride]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
+  to[stride*0] = __low2half(from);
+  to[stride*1] = __high2half(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
+  return __low2half(a); 
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
+  __half x = __ushort_as_half(__half_as_ushort(__low2half(a)) & 0x7FFF);
+  __half y = __ushort_as_half(__half_as_ushort(__high2half(a)) & 0x7FFF);
+  return __halves2half2(x, y);
+}
+
+
+__device__ EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<half2,2>& kernel) {
+  __half a1 = __low2half(kernel.packet[0]);
+  __half a2 = __high2half(kernel.packet[0]);
+  __half b1 = __low2half(kernel.packet[1]);
+  __half b2 = __high2half(kernel.packet[1]);
+  kernel.packet[0] = __halves2half2(a1, b1);
+  kernel.packet[1] = __halves2half2(a2, b2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+  return __halves2half2(a, __hadd(a, __float2half(1.0f)));
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+  return __hadd2(a, b);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+  return __hsub2(a, b);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+  return __hneg2(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+  return __hmul2(a, b);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+   return __hfma2(a, b, c);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return h2div(a, b);
+#else
+  return __h2div(a, b);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 < b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 < b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 > b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 > b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+  return __hadd(__low2half(a), __high2half(a));
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hgt(first, second) ? first : second;
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hlt(first, second) ? first : second;
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+  return __hmul(__low2half(a), __high2half(a));
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = log1pf(a1);
+  float r2 = log1pf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = expm1f(a1);
+  float r2 = expm1f(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<>  __device__ EIGEN_STRONG_INLINE
+half2 plog<half2>(const half2& a) {
+  return h2log(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 pexp<half2>(const half2& a) {
+  return h2exp(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 psqrt<half2>(const half2& a) {
+  return h2sqrt(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 prsqrt<half2>(const half2& a) {
+  return h2rsqrt(a);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX512
+
+typedef struct {
+  __m256i x;
+} Packet16h;
+
+
+template<> struct is_arithmetic<Packet16h> { enum { value = true }; };
+
+template <>
+struct packet_traits<half> : default_packet_traits {
+  typedef Packet16h type;
+  // There is no half-size packet for Packet16h.
+  typedef Packet16h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; };
+
+template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) {
+  Packet16h result;
+  result.x = _mm256_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_store_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_storeu_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h
+ploadquad(const Eigen::half* from) {
+  Packet16h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm512_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+  float f8(aux[8]);
+  float f9(aux[9]);
+  float fa(aux[10]);
+  float fb(aux[11]);
+  float fc(aux[12]);
+  float fd(aux[13]);
+  float fe(aux[14]);
+  float ff(aux[15]);
+
+  return _mm512_set_ps(
+      ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet16h result;
+  result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN64 float aux[16];
+  pstore(aux, a);
+  half h0(aux[0]);
+  half h1(aux[1]);
+  half h2(aux[2]);
+  half h3(aux[3]);
+  half h4(aux[4]);
+  half h5(aux[5]);
+  half h6(aux[6]);
+  half h7(aux[7]);
+  half h8(aux[8]);
+  half h9(aux[9]);
+  half ha(aux[10]);
+  half hb(aux[11]);
+  half hc(aux[12]);
+  half hd(aux[13]);
+  half he(aux[14]);
+  half hf(aux[15]);
+
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x,
+      h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
+  Packet16f from_float = half2float(from);
+  return half(predux(from_float));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
+{
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x,
+      from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x,
+      from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x,
+      from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride)
+{
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+  to[stride*8].x = aux[8].x;
+  to[stride*9].x = aux[9].x;
+  to[stride*10].x = aux[10].x;
+  to[stride*11].x = aux[11].x;
+  to[stride*12].x = aux[12].x;
+  to[stride*13].x = aux[13].x;
+  to[stride*14].x = aux[14].x;
+  to[stride*15].x = aux[15].x;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,16>& kernel) {
+  __m256i a = kernel.packet[0].x;
+  __m256i b = kernel.packet[1].x;
+  __m256i c = kernel.packet[2].x;
+  __m256i d = kernel.packet[3].x;
+  __m256i e = kernel.packet[4].x;
+  __m256i f = kernel.packet[5].x;
+  __m256i g = kernel.packet[6].x;
+  __m256i h = kernel.packet[7].x;
+  __m256i i = kernel.packet[8].x;
+  __m256i j = kernel.packet[9].x;
+  __m256i k = kernel.packet[10].x;
+  __m256i l = kernel.packet[11].x;
+  __m256i m = kernel.packet[12].x;
+  __m256i n = kernel.packet[13].x;
+  __m256i o = kernel.packet[14].x;
+  __m256i p = kernel.packet[15].x;
+
+  __m256i ab_07 = _mm256_unpacklo_epi16(a, b);
+  __m256i cd_07 = _mm256_unpacklo_epi16(c, d);
+  __m256i ef_07 = _mm256_unpacklo_epi16(e, f);
+  __m256i gh_07 = _mm256_unpacklo_epi16(g, h);
+  __m256i ij_07 = _mm256_unpacklo_epi16(i, j);
+  __m256i kl_07 = _mm256_unpacklo_epi16(k, l);
+  __m256i mn_07 = _mm256_unpacklo_epi16(m, n);
+  __m256i op_07 = _mm256_unpacklo_epi16(o, p);
+
+  __m256i ab_8f = _mm256_unpackhi_epi16(a, b);
+  __m256i cd_8f = _mm256_unpackhi_epi16(c, d);
+  __m256i ef_8f = _mm256_unpackhi_epi16(e, f);
+  __m256i gh_8f = _mm256_unpackhi_epi16(g, h);
+  __m256i ij_8f = _mm256_unpackhi_epi16(i, j);
+  __m256i kl_8f = _mm256_unpackhi_epi16(k, l);
+  __m256i mn_8f = _mm256_unpackhi_epi16(m, n);
+  __m256i op_8f = _mm256_unpackhi_epi16(o, p);
+
+  __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07);
+  __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07);
+  __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07);
+  __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07);
+  __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07);
+  __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07);
+  __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07);
+  __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07);
+
+  __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f);
+  __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f);
+  __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f);
+  __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f);
+  __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f);
+  __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f);
+  __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f);
+  __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f);
+
+  __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03);
+  __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03);
+  __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03);
+  __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03);
+  __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47);
+  __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47);
+  __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47);
+  __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47);
+  __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b);
+  __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b);
+  __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b);
+  __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b);
+  __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf);
+  __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf);
+  __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf);
+  __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf);
+
+  // NOTE: no unpacklo/hi instr in this case, so using permute instr.
+  __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
+  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
+  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
+  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
+  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
+  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
+  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
+  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
+  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
+  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
+  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
+  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
+  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
+  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
+  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
+
+  kernel.packet[0].x = a_p_0;
+  kernel.packet[1].x = a_p_1;
+  kernel.packet[2].x = a_p_2;
+  kernel.packet[3].x = a_p_3;
+  kernel.packet[4].x = a_p_4;
+  kernel.packet[5].x = a_p_5;
+  kernel.packet[6].x = a_p_6;
+  kernel.packet[7].x = a_p_7;
+  kernel.packet[8].x = a_p_8;
+  kernel.packet[9].x = a_p_9;
+  kernel.packet[10].x = a_p_a;
+  kernel.packet[11].x = a_p_b;
+  kernel.packet[12].x = a_p_c;
+  kernel.packet[13].x = a_p_d;
+  kernel.packet[14].x = a_p_e;
+  kernel.packet[15].x = a_p_f;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,8>& kernel) {
+  EIGEN_ALIGN64 half in[8][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+  pstore<half>(in[4], kernel.packet[4]);
+  pstore<half>(in[5], kernel.packet[5]);
+  pstore<half>(in[6], kernel.packet[6]);
+  pstore<half>(in[7], kernel.packet[7]);
+
+  EIGEN_ALIGN64 half out[8][16];
+
+  for (int i = 0; i < 8; ++i) {
+    for (int j = 0; j < 8; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 8; ++j) {
+      out[i][j+8] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+  kernel.packet[4] = pload<Packet16h>(out[4]);
+  kernel.packet[5] = pload<Packet16h>(out[5]);
+  kernel.packet[6] = pload<Packet16h>(out[6]);
+  kernel.packet[7] = pload<Packet16h>(out[7]);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,4>& kernel) {
+  EIGEN_ALIGN64 half in[4][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN64 half out[4][16];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][4*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][4*i+1];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+8] = in[j][4*i+2];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+12] = in[j][4*i+3];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+}
+
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+typedef struct {
+  __m128i x;
+} Packet8h;
+
+
+template<> struct is_arithmetic<Packet8h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet8h type;
+  // There is no half-size packet for Packet8h.
+  typedef Packet8h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 8,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; };
+
+template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) {
+  Packet8h result;
+  result.x = _mm_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet8h>(const Packet8h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pload<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_load_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h ploadu<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_store_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h
+ploadquad<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  result.x = _mm_set_epi16(b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm256_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+
+  return _mm256_set_ps(f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet8h result;
+  result.x = _mm256_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN32 float aux[8];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+  Eigen::half h4(aux[4]);
+  Eigen::half h5(aux[5]);
+  Eigen::half h6(aux[6]);
+  Eigen::half h7(aux[7]);
+
+  Packet8h result;
+  result.x = _mm_set_epi16(h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride)
+{
+  Packet8h result;
+  result.x = _mm_set_epi16(from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half* to, const Packet8h& from, Index stride)
+{
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_max<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_min<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_mul<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,8>& kernel) {
+  __m128i a = kernel.packet[0].x;
+  __m128i b = kernel.packet[1].x;
+  __m128i c = kernel.packet[2].x;
+  __m128i d = kernel.packet[3].x;
+  __m128i e = kernel.packet[4].x;
+  __m128i f = kernel.packet[5].x;
+  __m128i g = kernel.packet[6].x;
+  __m128i h = kernel.packet[7].x;
+
+  __m128i a03b03 = _mm_unpacklo_epi16(a, b);
+  __m128i c03d03 = _mm_unpacklo_epi16(c, d);
+  __m128i e03f03 = _mm_unpacklo_epi16(e, f);
+  __m128i g03h03 = _mm_unpacklo_epi16(g, h);
+  __m128i a47b47 = _mm_unpackhi_epi16(a, b);
+  __m128i c47d47 = _mm_unpackhi_epi16(c, d);
+  __m128i e47f47 = _mm_unpackhi_epi16(e, f);
+  __m128i g47h47 = _mm_unpackhi_epi16(g, h);
+
+  __m128i a01b01c01d01 = _mm_unpacklo_epi32(a03b03, c03d03);
+  __m128i a23b23c23d23 = _mm_unpackhi_epi32(a03b03, c03d03);
+  __m128i e01f01g01h01 = _mm_unpacklo_epi32(e03f03, g03h03);
+  __m128i e23f23g23h23 = _mm_unpackhi_epi32(e03f03, g03h03);
+  __m128i a45b45c45d45 = _mm_unpacklo_epi32(a47b47, c47d47);
+  __m128i a67b67c67d67 = _mm_unpackhi_epi32(a47b47, c47d47);
+  __m128i e45f45g45h45 = _mm_unpacklo_epi32(e47f47, g47h47);
+  __m128i e67f67g67h67 = _mm_unpackhi_epi32(e47f47, g47h47);
+
+  __m128i a0b0c0d0e0f0g0h0 = _mm_unpacklo_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a1b1c1d1e1f1g1h1 = _mm_unpackhi_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a2b2c2d2e2f2g2h2 = _mm_unpacklo_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a3b3c3d3e3f3g3h3 = _mm_unpackhi_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a4b4c4d4e4f4g4h4 = _mm_unpacklo_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a5b5c5d5e5f5g5h5 = _mm_unpackhi_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a6b6c6d6e6f6g6h6 = _mm_unpacklo_epi64(a67b67c67d67, e67f67g67h67);
+  __m128i a7b7c7d7e7f7g7h7 = _mm_unpackhi_epi64(a67b67c67d67, e67f67g67h67);
+
+  kernel.packet[0].x = a0b0c0d0e0f0g0h0;
+  kernel.packet[1].x = a1b1c1d1e1f1g1h1;
+  kernel.packet[2].x = a2b2c2d2e2f2g2h2;
+  kernel.packet[3].x = a3b3c3d3e3f3g3h3;
+  kernel.packet[4].x = a4b4c4d4e4f4g4h4;
+  kernel.packet[5].x = a5b5c5d5e5f5g5h5;
+  kernel.packet[6].x = a6b6c6d6e6f6g6h6;
+  kernel.packet[7].x = a7b7c7d7e7f7g7h7;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,4>& kernel) {
+  EIGEN_ALIGN32 Eigen::half in[4][8];
+  pstore<Eigen::half>(in[0], kernel.packet[0]);
+  pstore<Eigen::half>(in[1], kernel.packet[1]);
+  pstore<Eigen::half>(in[2], kernel.packet[2]);
+  pstore<Eigen::half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN32 Eigen::half out[4][8];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet8h>(out[0]);
+  kernel.packet[1] = pload<Packet8h>(out[1]);
+  kernel.packet[2] = pload<Packet8h>(out[2]);
+  kernel.packet[3] = pload<Packet8h>(out[3]);
+}
+
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+typedef struct {
+  __m64 x;
+} Packet4h;
+
+
+template<> struct is_arithmetic<Packet4h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet4h type;
+  // There is no half-size packet for Packet4h.
+  typedef Packet4h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; };
+
+template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) {
+  Packet4h result;
+  result.x = _mm_set1_pi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha + hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha * hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h
+ploadquad<Packet4h>(const Eigen::half* from) {
+  return pset1<Packet4h>(*from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(const Eigen::half* from, Index stride)
+{
+  Packet4h result;
+  result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4h>(Eigen::half* to, const Packet4h& from, Index stride)
+{
+  __int64_t a = _mm_cvtm64_si64(from.x);
+  to[stride*0].x = static_cast<unsigned short>(a);
+  to[stride*1].x = static_cast<unsigned short>(a >> 16);
+  to[stride*2].x = static_cast<unsigned short>(a >> 32);
+  to[stride*3].x = static_cast<unsigned short>(a >> 48);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet4h,4>& kernel) {
+  __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x);
+  __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x);
+
+  kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1);
+  kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1);
+  kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3);
+  kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3);
+}
+
+#endif
+
+}
+}
+
+#endif // EIGEN_PACKET_MATH_HALF_HIP_H
diff --git a/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h b/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h
new file mode 100644
index 000000000..915266a9d
--- /dev/null
+++ b/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h
@@ -0,0 +1,212 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_HIP_H
+#define EIGEN_TYPE_CASTING_HIP_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<>
+struct scalar_cast_op<float, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
+    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+      return __float2half(a);
+    #else
+      return Eigen::half(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<float, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<int, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
+    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+    return __float2half(static_cast<float>(a));
+    #else
+      return Eigen::half(static_cast<float>(a));
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<int, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<Eigen::half, float> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef float result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
+    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+      return __half2float(a);
+    #else
+      return static_cast<float>(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<Eigen::half, float> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+
+#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
+  float2 r1 = __half22float2(a);
+  float2 r2 = __half22float2(b);
+  return make_float4(r1.x, r1.y, r2.x, r2.y);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
+  // Simply discard the second half of the input
+  return __floats2half2_rn(a.x, a.y);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX512
+template <>
+struct type_casting_traits<half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) {
+  return float2half(a);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
+  return float2half(a);
+}
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4h, Packet4f>(const Packet4h& a) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  Eigen::half h = raw_uint16_to_half(static_cast<unsigned short>(a64));
+  float f1 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  float f2 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  float f3 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  float f4 = static_cast<float>(h);
+  return _mm_set_ps(f4, f3, f2, f1);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f& a) {
+  EIGEN_ALIGN16 float aux[4];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+
+  Packet4h result;
+  result.x = _mm_set_pi16(h3.x, h2.x, h1.x, h0.x);
+  return result;
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_HIP_H
diff --git a/Eigen/src/Core/arch/HIP/hcc/math_constants.h b/Eigen/src/Core/arch/HIP/hcc/math_constants.h
new file mode 100644
index 000000000..25375a0a4
--- /dev/null
+++ b/Eigen/src/Core/arch/HIP/hcc/math_constants.h
@@ -0,0 +1,23 @@
+/*
+ * math_constants.h - 
+ *  HIP equivalent of the CUDA header of the same name
+ */
+
+#ifndef __MATH_CONSTANTS_H__
+#define __MATH_CONSTANTS_H__
+
+/* single precision constants */
+
+#define HIPRT_INF_F        __int_as_float(0x7f800000)
+#define HIPRT_NAN_F        __int_as_float(0x7fffffff)
+#define HIPRT_MIN_DENORM_F __int_as_float(0x00000001)
+#define HIPRT_MAX_NORMAL_F __int_as_float(0x7f7fffff)
+#define HIPRT_NEG_ZERO_F   __int_as_float(0x80000000)
+#define HIPRT_ZERO_F       0.0f
+#define HIPRT_ONE_F        1.0f
+
+/* double precision constants */
+#define HIPRT_INF          __hiloint2double(0x7ff00000, 0x00000000)
+#define HIPRT_NAN          __hiloint2double(0xfff80000, 0x00000000)
+
+#endif
diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h
index 3eae6b8ca..e269140bd 100644
--- a/Eigen/src/Core/functors/BinaryFunctors.h
+++ b/Eigen/src/Core/functors/BinaryFunctors.h
@@ -436,6 +436,9 @@ template<typename BinaryOp> struct bind1st_op : BinaryOp {
   typedef typename BinaryOp::second_argument_type second_argument_type;
   typedef typename BinaryOp::result_type          result_type;
 
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC explicit
+  #endif
   bind1st_op(const first_argument_type &val) : m_value(val) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const second_argument_type& b) const { return BinaryOp::operator()(m_value,b); }
@@ -455,6 +458,9 @@ template<typename BinaryOp> struct bind2nd_op : BinaryOp {
   typedef typename BinaryOp::second_argument_type second_argument_type;
   typedef typename BinaryOp::result_type          result_type;
 
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC explicit
+  #endif
   bind2nd_op(const second_argument_type &val) : m_value(val) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const first_argument_type& a) const { return BinaryOp::operator()(a,m_value); }
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index b1791fb3a..a4cde6d95 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -163,7 +163,10 @@ class BlasLinearMapper {
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+  EIGEN_ALWAYS_INLINE void prefetch(int i) const {
     internal::prefetch(&operator()(i));
   }
 
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 9c68ecb7d..c6e27f6af 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1008,9 +1008,12 @@ namespace Eigen {
 #  define EIGEN_TRY try
 #  define EIGEN_CATCH(X) catch (X)
 #else
-#  ifdef EIGEN_CUDA_ARCH
+#  if defined(EIGEN_CUDA_ARCH)
 #    define EIGEN_THROW_X(X) asm("trap;")
 #    define EIGEN_THROW asm("trap;")
+#  elif defined(EIGEN_HIP_DEVICE_COMPILE)
+#    define EIGEN_THROW_X(X) asm("s_trap 0")
+#    define EIGEN_THROW asm("s_trap 0")
 #  else
 #    define EIGEN_THROW_X(X) std::abort()
 #    define EIGEN_THROW std::abort()
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 53300c388..87fcc30f5 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -70,7 +70,20 @@ inline void throw_std_bad_alloc()
     throw std::bad_alloc();
   #else
     std::size_t huge = static_cast<std::size_t>(-1);
+    #if defined(EIGEN_HIPCC)
+    //
+    // calls to "::operator new" are to be treated as opaque function calls (i.e no inlining),
+    // and as a consequence the code in the #else block triggers the hipcc warning :
+    // "no overloaded function has restriction specifiers that are compatible with the ambient context"
+    //
+    // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
+    // the same on "operator new" 
+    // Reverting code back to the old version in this #if block for the hipcc compiler
+    //
+    new int[huge];
+    #else
     ::operator new(huge);
+    #endif
   #endif
 }
 
@@ -156,7 +169,13 @@ EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
 
   void *result;
   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
+  
+    #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    result = aligned_malloc(size);
+    #else
     result = std::malloc(size);
+    #endif
+    
     #if EIGEN_DEFAULT_ALIGN_BYTES==16
     eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator.");
     #endif
@@ -174,7 +193,13 @@ EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
 EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
 {
   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
+
+    #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    aligned_free(ptr);
+    #else
     std::free(ptr);
+    #endif
+    
   #else
     handmade_aligned_free(ptr);
   #endif
@@ -218,7 +243,12 @@ template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std:
 {
   check_that_malloc_is_allowed();
 
+  #if defined(EIGEN_HIP_DEVICE_COMPILE)
+  void *result = aligned_malloc(size);
+  #else
   void *result = std::malloc(size);
+  #endif
+  
   if(!result && size)
     throw_std_bad_alloc();
   return result;
@@ -232,7 +262,11 @@ template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void
 
 template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
 {
+  #if defined(EIGEN_HIP_DEVICE_COMPILE)
+  aligned_free(ptr);
+  #else
   std::free(ptr);
+  #endif
 }
 
 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
@@ -493,7 +527,11 @@ template<typename T> struct smart_copy_helper<T,true> {
     IntPtr size = IntPtr(end)-IntPtr(start);
     if(size==0) return;
     eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    ::memcpy(target, start, size);
+    #else
     std::memcpy(target, start, size);
+    #endif
   }
 };
 
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 6e5af35c0..ca6fa6ce9 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -16,6 +16,12 @@
 #include <math_constants.h>
 #endif
 
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+#include <cfloat>
+#include "Eigen/src/Core/arch/HIP/hcc/math_constants.h"
+#endif
+
+
 #if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
 #include <cstdint>
 #endif
@@ -175,7 +181,7 @@ template<bool Condition, typename T=void> struct enable_if;
 template<typename T> struct enable_if<true,T>
 { typedef T type; };
 
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 #if !defined(__FLT_EPSILON__)
 #define __FLT_EPSILON__ FLT_EPSILON
 #define __DBL_EPSILON__ DBL_EPSILON
@@ -197,13 +203,31 @@ template<> struct numeric_limits<float>
   EIGEN_DEVICE_FUNC
   static float epsilon() { return __FLT_EPSILON__; }
   EIGEN_DEVICE_FUNC
-  static float (max)() { return CUDART_MAX_NORMAL_F; }
+  static float (max)() {
+  #if defined(EIGEN_CUDA_ARCH)
+    return CUDART_MAX_NORMAL_F;
+  #else
+    return HIPRT_MAX_NORMAL_F;
+  #endif
+  }
   EIGEN_DEVICE_FUNC
   static float (min)() { return FLT_MIN; }
   EIGEN_DEVICE_FUNC
-  static float infinity() { return CUDART_INF_F; }
+  static float infinity() {
+  #if defined(EIGEN_CUDA_ARCH)
+    return CUDART_INF_F;
+  #else
+    return HIPRT_INF_F;
+  #endif
+  }
   EIGEN_DEVICE_FUNC
-  static float quiet_NaN() { return CUDART_NAN_F; }
+  static float quiet_NaN() {
+  #if defined(EIGEN_CUDA_ARCH)
+    return CUDART_NAN_F;
+  #else
+    return HIPRT_NAN_F;
+  #endif
+  }
 };
 template<> struct numeric_limits<double>
 {
@@ -214,9 +238,21 @@ template<> struct numeric_limits<double>
   EIGEN_DEVICE_FUNC
   static double (min)() { return DBL_MIN; }
   EIGEN_DEVICE_FUNC
-  static double infinity() { return CUDART_INF; }
+  static double infinity() {
+  #if defined(EIGEN_CUDA_ARCH)
+    return CUDART_INF;
+  #else
+    return HIPRT_INF;
+  #endif
+  }
   EIGEN_DEVICE_FUNC
-  static double quiet_NaN() { return CUDART_NAN; }
+  static double quiet_NaN() {
+  #if defined(EIGEN_CUDA_ARCH)
+    return CUDART_NAN;
+  #else
+    return HIPRT_NAN;
+  #endif
+  }
 };
 template<> struct numeric_limits<int>
 {
@@ -529,13 +565,13 @@ template<typename T, typename U> struct scalar_product_traits
 
 namespace numext {
   
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
 #else
 template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
 #endif
 
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
 using internal::device::numeric_limits;
 #else
 using std::numeric_limits;
diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
index 040f8d3bb..bf28edc0e 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -354,6 +354,7 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
     static const int m_maxIterations = 30;
 
   protected:
+    EIGEN_DEVICE_FUNC 
     static void check_template_parameters()
     {
       EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index a24deb96a..e977b9623 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -1299,7 +1299,7 @@ void BDCSVD<MatrixType>::deflation(Eigen::Index firstCol, Eigen::Index lastCol,
 #endif
 }//end deflation
 
-#ifndef EIGEN_CUDACC
+#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
 /** \svd_module
   *
   * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 7d2d63722..f818ae840 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -19,7 +19,9 @@ macro(ei_add_test_internal testname testname_with_suffix)
   endif()
 
   if(EIGEN_ADD_TEST_FILENAME_EXTENSION STREQUAL cu)
-    if(EIGEN_TEST_CUDA_CLANG)
+    if(EIGEN_TEST_HIP)
+      hip_add_executable(${targetname} ${filename} HIPCC_OPTIONS "-DEIGEN_USE_HIP")
+    elseif(EIGEN_TEST_CUDA_CLANG)
       set_source_files_properties(${filename} PROPERTIES LANGUAGE CXX)
       if(CUDA_64_BIT_DEVICE_CODE)
         link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib64")
@@ -491,6 +493,11 @@ macro(ei_testing_print_summary)
     else()
       message(STATUS "CUDA:              OFF")
     endif()
+    if(EIGEN_TEST_HIP)
+      message(STATUS "HIP:               ON (using hipcc)")
+    else()
+      message(STATUS "HIP:               OFF")
+    endif()
 
   endif() # vectorization / alignment options
 
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index e1eef086e..4a5c1d36d 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -407,6 +407,48 @@ endif(CUDA_FOUND)
 endif(EIGEN_TEST_CUDA)
 
 
+# HIP unit tests
+option(EIGEN_TEST_HIP "Add HIP support." OFF)
+if (EIGEN_TEST_HIP)
+
+  set(HIP_PATH "/opt/rocm/hip" CACHE STRING "Path to the HIP installation.")
+
+  if (EXISTS ${HIP_PATH})
+    
+    list(APPEND CMAKE_MODULE_PATH ${HIP_PATH}/cmake) 
+
+    find_package(HIP REQUIRED)
+    if (HIP_FOUND)
+
+      execute_process(COMMAND ${HIP_PATH}/bin/hipconfig --platform OUTPUT_VARIABLE HIP_PLATFORM)
+
+      if (${HIP_PLATFORM} STREQUAL "hcc")
+
+	include_directories(${CMAKE_CURRENT_BINARY_DIR})
+	include_directories(${HIP_PATH}/include)
+
+	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
+	ei_add_test(hip_basic)
+	unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
+	
+      elseif (${HIP_PLATFORM} STREQUAL "nvcc")
+	message(FATAL_ERROR "HIP_PLATFORM = nvcc is not supported within Eigen")
+      else ()
+	message(FATAL_ERROR "Unknown HIP_PLATFORM = ${HIP_PLATFORM}")
+      endif()
+      
+    endif(HIP_FOUND)
+    
+  else ()
+
+    message(FATAL_ERROR "EIGEN_TEST_HIP is ON, but the specified HIP_PATH (${HIP_PATH}) does not exist")
+    
+  endif()
+    
+endif(EIGEN_TEST_HIP)
+
+
+
 file(MAKE_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/failtests)    
 add_test(NAME failtests WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/failtests COMMAND ${CMAKE_COMMAND} ${Eigen_SOURCE_DIR} -G "${CMAKE_GENERATOR}" -DEIGEN_FAILTEST=ON)
 
diff --git a/test/hip_basic.cu b/test/hip_basic.cu
new file mode 100644
index 000000000..2e1bf94a4
--- /dev/null
+++ b/test/hip_basic.cu
@@ -0,0 +1,172 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// workaround issue between gcc >= 4.7 and cuda 5.5
+#if (defined __GNUC__) && (__GNUC__>4 || __GNUC_MINOR__>=7)
+  #undef _GLIBCXX_ATOMIC_BUILTINS
+  #undef _GLIBCXX_USE_INT128
+#endif
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC hip_basic
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+
+#include <hip/hip_runtime.h>
+
+#include "main.h"
+#include "hip_common.h"
+
+// Check that dense modules can be properly parsed by hipcc
+#include <Eigen/Dense>
+
+// struct Foo{
+//   EIGEN_DEVICE_FUNC
+//   void operator()(int i, const float* mats, float* vecs) const {
+//     using namespace Eigen;
+//   //   Matrix3f M(data);
+//   //   Vector3f x(data+9);
+//   //   Map<Vector3f>(data+9) = M.inverse() * x;
+//     Matrix3f M(mats+i/16);
+//     Vector3f x(vecs+i*3);
+//   //   using std::min;
+//   //   using std::sqrt;
+//     Map<Vector3f>(vecs+i*3) << x.minCoeff(), 1, 2;// / x.dot(x);//(M.inverse() *  x) / x.x();
+//     //x = x*2 + x.y() * x + x * x.maxCoeff() - x / x.sum();
+//   }
+// };
+
+template<typename T>
+struct coeff_wise {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    T x1(in+i);
+    T x2(in+i+1);
+    T x3(in+i+2);
+    Map<T> res(out+i*T::MaxSizeAtCompileTime);
+    
+    res.array() += (in[0] * x1 + x2).array() * x3.array();
+  }
+};
+
+template<typename T>
+struct replicate {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    T x1(in+i);
+    int step   = x1.size() * 4;
+    int stride = 3 * step;
+    
+    typedef Map<Array<typename T::Scalar,Dynamic,Dynamic> > MapType;
+    MapType(out+i*stride+0*step, x1.rows()*2, x1.cols()*2) = x1.replicate(2,2);
+    MapType(out+i*stride+1*step, x1.rows()*3, x1.cols()) = in[i] * x1.colwise().replicate(3);
+    MapType(out+i*stride+2*step, x1.rows(), x1.cols()*3) = in[i] * x1.rowwise().replicate(3);
+  }
+};
+
+template<typename T>
+struct redux {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    int N = 10;
+    T x1(in+i);
+    out[i*N+0] = x1.minCoeff();
+    out[i*N+1] = x1.maxCoeff();
+    out[i*N+2] = x1.sum();
+    out[i*N+3] = x1.prod();
+    out[i*N+4] = x1.matrix().squaredNorm();
+    out[i*N+5] = x1.matrix().norm();
+    out[i*N+6] = x1.colwise().sum().maxCoeff();
+    out[i*N+7] = x1.rowwise().maxCoeff().sum();
+    out[i*N+8] = x1.matrix().colwise().squaredNorm().sum();
+  }
+};
+
+template<typename T1, typename T2>
+struct prod_test {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const
+  {
+    using namespace Eigen;
+    typedef Matrix<typename T1::Scalar, T1::RowsAtCompileTime, T2::ColsAtCompileTime> T3;
+    T1 x1(in+i);
+    T2 x2(in+i+1);
+    Map<T3> res(out+i*T3::MaxSizeAtCompileTime);
+    res += in[i] * x1 * x2;
+  }
+};
+
+template<typename T1, typename T2>
+struct diagonal {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const
+  {
+    using namespace Eigen;
+    T1 x1(in+i);
+    Map<T2> res(out+i*T2::MaxSizeAtCompileTime);
+    res += x1.diagonal();
+  }
+};
+
+template<typename T>
+struct eigenvalues {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    typedef Matrix<typename T::Scalar, T::RowsAtCompileTime, 1> Vec;
+    T M(in+i);
+    Map<Vec> res(out+i*Vec::MaxSizeAtCompileTime);
+    T A = M*M.adjoint();
+    SelfAdjointEigenSolver<T> eig;
+    eig.computeDirect(M);
+    res = eig.eigenvalues();
+  }
+};
+
+void test_hip_basic()
+{
+  ei_test_init_hip();
+  
+  int nthreads = 100;
+  Eigen::VectorXf in, out;
+  
+  #ifndef __HIP_DEVICE_COMPILE__ 
+  int data_size = nthreads * 512;
+  in.setRandom(data_size);
+  out.setRandom(data_size);
+  #endif
+  
+  CALL_SUBTEST( run_and_compare_to_hip(coeff_wise<Vector3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_hip(coeff_wise<Array44f>(), nthreads, in, out) );
+
+  // FIXME compile fails when we uncomment the followig two tests
+  // CALL_SUBTEST( run_and_compare_to_hip(replicate<Array4f>(), nthreads, in, out) );
+  // CALL_SUBTEST( run_and_compare_to_hip(replicate<Array33f>(), nthreads, in, out) );
+  
+  CALL_SUBTEST( run_and_compare_to_hip(redux<Array4f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_hip(redux<Matrix3f>(), nthreads, in, out) );
+  
+  CALL_SUBTEST( run_and_compare_to_hip(prod_test<Matrix3f,Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_hip(prod_test<Matrix4f,Vector4f>(), nthreads, in, out) );
+  
+  CALL_SUBTEST( run_and_compare_to_hip(diagonal<Matrix3f,Vector3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_hip(diagonal<Matrix4f,Vector4f>(), nthreads, in, out) );
+
+  // FIXME : Runtime failure occurs when we uncomment the following two tests
+  // CALL_SUBTEST( run_and_compare_to_hip(eigenvalues<Matrix3f>(), nthreads, in, out) );
+  // CALL_SUBTEST( run_and_compare_to_hip(eigenvalues<Matrix2f>(), nthreads, in, out) );
+
+}
diff --git a/test/hip_common.h b/test/hip_common.h
new file mode 100644
index 000000000..251585c52
--- /dev/null
+++ b/test/hip_common.h
@@ -0,0 +1,103 @@
+
+#ifndef EIGEN_TEST_HIP_COMMON_H
+#define EIGEN_TEST_HIP_COMMON_H
+
+#include "hip/hip_runtime.h"
+#include "hip/hip_runtime_api.h"
+#include <iostream>
+
+#ifndef __HIPCC__
+dim3 threadIdx, blockDim, blockIdx;
+#endif
+
+template<typename Kernel, typename Input, typename Output>
+void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out)
+{
+  for(int i=0; i<n; i++)
+    ker(i, in.data(), out.data());
+}
+
+
+template<typename Kernel, typename Input, typename Output>
+__global__ __attribute__((used))
+void run_on_hip_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
+{
+  int i = hipThreadIdx_x + hipBlockIdx_x*hipBlockDim_x;
+  if(i<n) {
+    ker(i, in, out);
+  }
+}
+
+
+template<typename Kernel, typename Input, typename Output>
+void run_on_hip(const Kernel& ker, int n, const Input& in, Output& out)
+{
+  typename Input::Scalar*  d_in;
+  typename Output::Scalar* d_out;
+  std::ptrdiff_t in_bytes  = in.size()  * sizeof(typename Input::Scalar);
+  std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
+  
+  hipMalloc((void**)(&d_in),  in_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+  
+  hipMemcpy(d_in,  in.data(),  in_bytes,  hipMemcpyHostToDevice);
+  hipMemcpy(d_out, out.data(), out_bytes, hipMemcpyHostToDevice);
+  
+  // Simple and non-optimal 1D mapping assuming n is not too large
+  // That's only for unit testing!
+  dim3 Blocks(128);
+  dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
+
+  hipDeviceSynchronize();
+  hipLaunchKernelGGL(HIP_KERNEL_NAME(run_on_hip_meta_kernel<Kernel,
+                                                         typename std::decay<decltype(*d_in)>::type,
+                                                         typename std::decay<decltype(*d_out)>::type>), 
+                  dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
+  hipDeviceSynchronize();
+  
+  // check inputs have not been modified
+  hipMemcpy(const_cast<typename Input::Scalar*>(in.data()),  d_in,  in_bytes,  hipMemcpyDeviceToHost);
+  hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost);
+  
+  hipFree(d_in);
+  hipFree(d_out);
+}
+
+
+template<typename Kernel, typename Input, typename Output>
+void run_and_compare_to_hip(const Kernel& ker, int n, const Input& in, Output& out)
+{
+  Input  in_ref,  in_hip;
+  Output out_ref, out_hip;
+  #ifndef __HIP_DEVICE_COMPILE__
+  in_ref = in_hip = in;
+  out_ref = out_hip = out;
+  #endif
+  run_on_cpu (ker, n, in_ref,  out_ref);
+  run_on_hip(ker, n, in_hip, out_hip);
+  #ifndef __HIP_DEVICE_COMPILE__
+  VERIFY_IS_APPROX(in_ref, in_hip);
+  VERIFY_IS_APPROX(out_ref, out_hip);
+  #endif
+}
+
+
+void ei_test_init_hip()
+{
+  int device = 0;
+  hipDeviceProp_t deviceProp;
+  hipGetDeviceProperties(&deviceProp, device);
+  std::cout << "HIP device info:\n";
+  std::cout << "  name:                        " << deviceProp.name << "\n";
+  std::cout << "  capability:                  " << deviceProp.major << "." << deviceProp.minor << "\n";
+  std::cout << "  multiProcessorCount:         " << deviceProp.multiProcessorCount << "\n";
+  std::cout << "  maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n";
+  std::cout << "  warpSize:                    " << deviceProp.warpSize << "\n";
+  std::cout << "  regsPerBlock:                " << deviceProp.regsPerBlock << "\n";
+  std::cout << "  concurrentKernels:           " << deviceProp.concurrentKernels << "\n";
+  std::cout << "  clockRate:                   " << deviceProp.clockRate << "\n";
+  std::cout << "  canMapHostMemory:            " << deviceProp.canMapHostMemory << "\n";
+  std::cout << "  computeMode:                 " << deviceProp.computeMode << "\n";
+}
+
+#endif // EIGEN_TEST_HIP_COMMON_H
diff --git a/test/main.h b/test/main.h
index 0fcd6cb76..79717a532 100644
--- a/test/main.h
+++ b/test/main.h
@@ -67,11 +67,17 @@
 // protected by parenthesis against macro expansion, the min()/max() macros
 // are defined here and any not-parenthesized min/max call will cause a
 // compiler error.
-#define min(A,B) please_protect_your_min_with_parentheses
-#define max(A,B) please_protect_your_max_with_parentheses
-#define isnan(X) please_protect_your_isnan_with_parentheses
-#define isinf(X) please_protect_your_isinf_with_parentheses
-#define isfinite(X) please_protect_your_isfinite_with_parentheses
+#if !defined(__HIPCC__)
+  // HIP headers include the <thread> header which contains not-parenthesized
+  // calls to "max", triggering the following check and causing the compile to fail
+  // so disabling the following checks for HIP
+  #define min(A,B) please_protect_your_min_with_parentheses
+  #define max(A,B) please_protect_your_max_with_parentheses
+  #define isnan(X) please_protect_your_isnan_with_parentheses
+  #define isinf(X) please_protect_your_isinf_with_parentheses
+  #define isfinite(X) please_protect_your_isfinite_with_parentheses
+#endif
+
 #ifdef M_PI
 #undef M_PI
 #endif
@@ -154,7 +160,7 @@ namespace Eigen
 
 #define EIGEN_DEFAULT_IO_FORMAT IOFormat(4, 0, "  ", "\n", "", "", "", "")
 
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__)
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
   #define EIGEN_EXCEPTIONS
 #endif
 
@@ -233,7 +239,7 @@ namespace Eigen
       }
     #endif //EIGEN_EXCEPTIONS
 
-  #elif !defined(__CUDACC__) // EIGEN_DEBUG_ASSERTS
+  #elif !defined(__CUDACC__) && !defined(__HIPCC__)// EIGEN_DEBUG_ASSERTS
     // see bug 89. The copy_bool here is working around a bug in gcc <= 4.3
     #define eigen_assert(a) \
       if( (!Eigen::internal::copy_bool(a)) && (!no_more_assert) )\
@@ -290,7 +296,7 @@ namespace Eigen
     std::cout << "Can't VERIFY_RAISES_STATIC_ASSERT( " #a " ) with exceptions disabled\n";
 #endif
     
-  #if !defined(__CUDACC__)
+  #if !defined(__CUDACC__) && !defined(__HIPCC__)
   #define EIGEN_USE_CUSTOM_ASSERT
   #endif
 
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index d243fe035..4b7c7d724 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -80,12 +80,16 @@ typedef unsigned __int64 uint64_t;
 #endif
 
 #ifdef EIGEN_USE_GPU
-#include <iostream>
-#include <cuda_runtime.h>
-#if __cplusplus >= 201103L
-#include <atomic>
-#include <unistd.h>
-#endif
+  #include <iostream>
+  #if defined(EIGEN_USE_HIP)
+    #include <hip/hip_runtime.h>
+  #else
+    #include <cuda_runtime.h>
+  #endif
+  #if __cplusplus >= 201103L
+    #include <atomic>
+    #include <unistd.h>
+  #endif
 #endif
 
 #include "src/Tensor/TensorMacros.h"
@@ -95,7 +99,11 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorCostModel.h"
 #include "src/Tensor/TensorDeviceDefault.h"
 #include "src/Tensor/TensorDeviceThreadPool.h"
-#include "src/Tensor/TensorDeviceCuda.h"
+#if defined(EIGEN_USE_HIP)
+  #include "src/Tensor/TensorDeviceHip.h"
+#else
+  #include "src/Tensor/TensorDeviceCuda.h"
+#endif
 #include "src/Tensor/TensorDeviceSycl.h"
 #include "src/Tensor/TensorIndexList.h"
 #include "src/Tensor/TensorDimensionList.h"
@@ -112,16 +120,28 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorEvaluator.h"
 #include "src/Tensor/TensorExpr.h"
 #include "src/Tensor/TensorReduction.h"
-#include "src/Tensor/TensorReductionCuda.h"
+#if defined(EIGEN_USE_HIP)
+  #include "src/Tensor/TensorReductionHip.h"
+#else
+  #include "src/Tensor/TensorReductionCuda.h"
+#endif
 #include "src/Tensor/TensorArgMax.h"
 #include "src/Tensor/TensorConcatenation.h"
 #include "src/Tensor/TensorContractionMapper.h"
 #include "src/Tensor/TensorContractionBlocking.h"
 #include "src/Tensor/TensorContraction.h"
 #include "src/Tensor/TensorContractionThreadPool.h"
-#include "src/Tensor/TensorContractionCuda.h"
+#if defined(EIGEN_USE_HIP)
+  #include "src/Tensor/TensorContractionHip.h"
+#else
+  #include "src/Tensor/TensorContractionCuda.h"
+#endif
 #include "src/Tensor/TensorConversion.h"
-#include "src/Tensor/TensorConvolution.h"
+#if defined(EIGEN_USE_HIP)
+  #include "src/Tensor/TensorConvolutionHip.h"
+#else
+  #include "src/Tensor/TensorConvolution.h"
+#endif   
 #include "src/Tensor/TensorFFT.h"
 #include "src/Tensor/TensorPatch.h"
 #include "src/Tensor/TensorImagePatch.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index e72ddb4a9..979fcf4d9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -448,7 +448,10 @@ struct TensorContractionEvaluatorBase
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  EIGEN_DEVICE_FUNC void evalGemv(Scalar* buffer) const {
+  #if !defined(EIGEN_HIPCC)    
+  EIGEN_DEVICE_FUNC
+  #endif
+  void evalGemv(Scalar* buffer) const {
     const Index rows = m_i_size;
     const Index cols = m_k_size;
 
@@ -489,7 +492,10 @@ struct TensorContractionEvaluatorBase
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  EIGEN_DEVICE_FUNC void evalGemm(Scalar* buffer) const {
+  #if !defined(EIGEN_HIPCC)    
+  EIGEN_DEVICE_FUNC
+  #endif
+  void evalGemm(Scalar* buffer) const {
     #if defined(EIGEN_VECTORIZE_AVX) && defined(EIGEN_USE_LIBXSMM)
     if (m_can_use_xsmm) {
       evalGemmXSMM(buffer);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
index d34f9caee..4853dd37b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -28,7 +28,10 @@ class TensorContractionBlocking {
   typedef typename LhsMapper::Scalar LhsScalar;
   typedef typename RhsMapper::Scalar RhsScalar;
 
-  EIGEN_DEVICE_FUNC TensorContractionBlocking(Index k, Index m, Index n, Index num_threads = 1) :
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+ TensorContractionBlocking(Index k, Index m, Index n, Index num_threads = 1) :
       kc_(k), mc_(m), nc_(n)
   {
     if (ShardingType == ShardByCol) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
new file mode 100644
index 000000000..7561846a3
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
@@ -0,0 +1,1521 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014-2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2015 Navdeep Jaitly <ndjaitly@google.com>
+// Copyright (C) 2014 Eric Martin <eric@ericmart.in>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
+
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
+
+namespace Eigen {
+
+template<typename Scalar, typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool needs_edge_check>
+__device__ EIGEN_STRONG_INLINE void
+EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
+                               const OutputMapper output, Scalar* lhs_shmem, Scalar* rhs_shmem,
+                       const Index m_size, const Index n_size, const Index k_size) {
+
+  const Index m_block_idx = hipBlockIdx_x;
+  const Index n_block_idx = hipBlockIdx_y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  // declare and initialize 64 registers for output 8x8 block
+
+  // prefetch registers
+  Scalar lhs_pf0;
+  Scalar lhs_pf1;
+  Scalar lhs_pf2;
+  Scalar lhs_pf3;
+  Scalar lhs_pf4;
+  Scalar lhs_pf5;
+  Scalar lhs_pf6;
+  Scalar lhs_pf7;
+
+  Scalar rhs_pf0;
+  Scalar rhs_pf1;
+  Scalar rhs_pf2;
+  Scalar rhs_pf3;
+  Scalar rhs_pf4;
+  Scalar rhs_pf5;
+  Scalar rhs_pf6;
+  Scalar rhs_pf7;
+
+  // shared memory is formatted
+  // (contract idx in block, nocontract idx in block, block idx)
+  // where block idx is column major. This transposition limits the number of
+  // bank conflicts when reading the LHS. The core idea is that since the contracting
+  // index is shared by both sides, then the contracting index should be in hipThreadIdx_x.
+
+  // On the LHS, we pad each row inside of each block with an extra element. This makes
+  // each block 8 rows of 9 elements, which is 72 elements. This gives no bank conflicts
+  // on writes and very few 2-way conflicts on reads. There is an 8x8 grid of these blocks.
+
+  // On the RHS we just add 8 padding elements to the end of each block. This gives no bank
+  // conflicts on writes and also none on reads.
+
+  // storage indices
+  const Index lhs_store_idx_base = hipThreadIdx_y * 72 + hipThreadIdx_x * 9 + hipThreadIdx_z;
+  const Index rhs_store_idx_base = hipThreadIdx_y * 72 + hipThreadIdx_z * 8 + hipThreadIdx_x;
+
+  const Index lhs_store_idx_0 = lhs_store_idx_base + 576 * 0;
+  const Index lhs_store_idx_1 = lhs_store_idx_base + 576 * 1;
+  const Index lhs_store_idx_2 = lhs_store_idx_base + 576 * 2;
+  const Index lhs_store_idx_3 = lhs_store_idx_base + 576 * 3;
+  const Index lhs_store_idx_4 = lhs_store_idx_base + 576 * 4;
+  const Index lhs_store_idx_5 = lhs_store_idx_base + 576 * 5;
+  const Index lhs_store_idx_6 = lhs_store_idx_base + 576 * 6;
+  const Index lhs_store_idx_7 = lhs_store_idx_base + 576 * 7;
+
+  const Index rhs_store_idx_0 = rhs_store_idx_base + 576 * 0;
+  const Index rhs_store_idx_1 = rhs_store_idx_base + 576 * 1;
+  const Index rhs_store_idx_2 = rhs_store_idx_base + 576 * 2;
+  const Index rhs_store_idx_3 = rhs_store_idx_base + 576 * 3;
+  const Index rhs_store_idx_4 = rhs_store_idx_base + 576 * 4;
+  const Index rhs_store_idx_5 = rhs_store_idx_base + 576 * 5;
+  const Index rhs_store_idx_6 = rhs_store_idx_base + 576 * 6;
+  const Index rhs_store_idx_7 = rhs_store_idx_base + 576 * 7;
+
+  // in the loading code, the following variables are important:
+  // hipThreadIdx_x: the vertical position in an 8x8 block
+  // hipThreadIdx_y: the vertical index of the 8x8 block in the grid
+  // hipThreadIdx_z: the horizontal position in an 8x8 block
+  // k: the horizontal index of the 8x8 block in the grid
+  //
+  // The k parameter is implicit (it was the loop counter for a loop that went
+  // from 0 to <8, but now that loop is unrolled in the below code.
+
+  const Index load_idx_vert = hipThreadIdx_x + 8 * hipThreadIdx_y;
+  const Index lhs_vert = base_m + load_idx_vert;
+
+#define prefetchIntoRegisters(base_k)                           \
+  {                                                             \
+    lhs_pf0 = conv(0);                                          \
+    lhs_pf1 = conv(0);                                          \
+    lhs_pf2 = conv(0);                                          \
+    lhs_pf3 = conv(0);                                          \
+    lhs_pf4 = conv(0);                                          \
+    lhs_pf5 = conv(0);                                          \
+    lhs_pf6 = conv(0);                                          \
+    lhs_pf7 = conv(0);                                          \
+                                                                \
+    rhs_pf0 = conv(0);                                          \
+    rhs_pf1 = conv(0);                                          \
+    rhs_pf2 = conv(0);                                          \
+    rhs_pf3 = conv(0);                                          \
+    rhs_pf4 = conv(0);                                          \
+    rhs_pf5 = conv(0);                                          \
+    rhs_pf6 = conv(0);                                          \
+    rhs_pf7 = conv(0);                                          \
+                                                                \
+    if (!needs_edge_check || lhs_vert < m_size) {               \
+      const Index lhs_horiz_0 = base_k + hipThreadIdx_z + 0 * 8;   \
+      const Index lhs_horiz_1 = base_k + hipThreadIdx_z + 1 * 8;   \
+      const Index lhs_horiz_2 = base_k + hipThreadIdx_z + 2 * 8;   \
+      const Index lhs_horiz_3 = base_k + hipThreadIdx_z + 3 * 8;   \
+      const Index lhs_horiz_4 = base_k + hipThreadIdx_z + 4 * 8;   \
+      const Index lhs_horiz_5 = base_k + hipThreadIdx_z + 5 * 8;   \
+      const Index lhs_horiz_6 = base_k + hipThreadIdx_z + 6 * 8;   \
+      const Index lhs_horiz_7 = base_k + hipThreadIdx_z + 7 * 8;   \
+                                                                \
+      if (!needs_edge_check || lhs_horiz_7 < k_size) {          \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
+        lhs_pf7 = lhs(lhs_vert, lhs_horiz_7);                   \
+      } else if (lhs_horiz_6 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
+      } else if (lhs_horiz_5 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+      } else if (lhs_horiz_4 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+      } else if (lhs_horiz_3 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+      } else if (lhs_horiz_2 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+      } else if (lhs_horiz_1 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+      } else if (lhs_horiz_0 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+      }                                                         \
+    }                                                           \
+                                                                \
+    const Index rhs_vert = base_k + load_idx_vert;              \
+    if (!needs_edge_check || rhs_vert < k_size) {               \
+      const Index rhs_horiz_0 = base_n + hipThreadIdx_z + 0 * 8;   \
+      const Index rhs_horiz_1 = base_n + hipThreadIdx_z + 1 * 8;   \
+      const Index rhs_horiz_2 = base_n + hipThreadIdx_z + 2 * 8;   \
+      const Index rhs_horiz_3 = base_n + hipThreadIdx_z + 3 * 8;   \
+      const Index rhs_horiz_4 = base_n + hipThreadIdx_z + 4 * 8;   \
+      const Index rhs_horiz_5 = base_n + hipThreadIdx_z + 5 * 8;   \
+      const Index rhs_horiz_6 = base_n + hipThreadIdx_z + 6 * 8;   \
+      const Index rhs_horiz_7 = base_n + hipThreadIdx_z + 7 * 8;   \
+                                                                \
+      if (rhs_horiz_7 < n_size) {                               \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
+        rhs_pf7 = rhs(rhs_vert, rhs_horiz_7);                   \
+      } else if (rhs_horiz_6 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
+      } else if (rhs_horiz_5 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+      } else if (rhs_horiz_4 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+      } else if (rhs_horiz_3 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+      } else if (rhs_horiz_2 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+      } else if (rhs_horiz_1 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+      } else if (rhs_horiz_0 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+      }                                                         \
+    }                                                           \
+  }                                                             \
+
+#define writeRegToShmem(_)                      \
+  lhs_shmem[lhs_store_idx_0] = lhs_pf0;         \
+  rhs_shmem[rhs_store_idx_0] = rhs_pf0;         \
+                                                \
+  lhs_shmem[lhs_store_idx_1] = lhs_pf1;         \
+  rhs_shmem[rhs_store_idx_1] = rhs_pf1;         \
+                                                \
+  lhs_shmem[lhs_store_idx_2] = lhs_pf2;         \
+  rhs_shmem[rhs_store_idx_2] = rhs_pf2;         \
+                                                \
+  lhs_shmem[lhs_store_idx_3] = lhs_pf3;         \
+  rhs_shmem[rhs_store_idx_3] = rhs_pf3;         \
+                                                \
+  lhs_shmem[lhs_store_idx_4] = lhs_pf4;         \
+  rhs_shmem[rhs_store_idx_4] = rhs_pf4;         \
+                                                \
+  lhs_shmem[lhs_store_idx_5] = lhs_pf5;         \
+  rhs_shmem[rhs_store_idx_5] = rhs_pf5;         \
+                                                \
+  lhs_shmem[lhs_store_idx_6] = lhs_pf6;         \
+  rhs_shmem[rhs_store_idx_6] = rhs_pf6;         \
+                                                \
+  lhs_shmem[lhs_store_idx_7] = lhs_pf7;         \
+  rhs_shmem[rhs_store_idx_7] = rhs_pf7;         \
+
+  // declare and initialize result array
+#define res(i, j) _res_##i##j
+#define initResultRow(i)                        \
+  Scalar res(i, 0) = conv(0);                   \
+  Scalar res(i, 1) = conv(0);                   \
+  Scalar res(i, 2) = conv(0);                   \
+  Scalar res(i, 3) = conv(0);                   \
+  Scalar res(i, 4) = conv(0);                   \
+  Scalar res(i, 5) = conv(0);                   \
+  Scalar res(i, 6) = conv(0);                   \
+  Scalar res(i, 7) = conv(0);                   \
+
+  internal::scalar_cast_op<int, Scalar> conv;
+  initResultRow(0);
+  initResultRow(1);
+  initResultRow(2);
+  initResultRow(3);
+  initResultRow(4);
+  initResultRow(5);
+  initResultRow(6);
+  initResultRow(7);
+#undef initResultRow
+
+  for (Index base_k = 0; base_k < k_size; base_k += 64) {
+    // wait for previous iteration to finish with shmem. Despite common sense,
+    // the code is a bit faster with this here then at bottom of loop
+    __syncthreads();
+
+    prefetchIntoRegisters(base_k);
+    writeRegToShmem();
+
+    #undef prefetchIntoRegisters
+    #undef writeRegToShmem
+
+    // wait for shared mem packing to be done before starting computation
+    __syncthreads();
+
+    // compute 8x8 matrix product by outer product. This involves packing one column
+    // of LHS and one row of RHS into registers (takes 16 registers).
+
+#define lcol(i) _lcol##i
+    Scalar lcol(0);
+    Scalar lcol(1);
+    Scalar lcol(2);
+    Scalar lcol(3);
+    Scalar lcol(4);
+    Scalar lcol(5);
+    Scalar lcol(6);
+    Scalar lcol(7);
+
+#define rrow(j) _rrow##j
+    Scalar rrow(0);
+    Scalar rrow(1);
+    Scalar rrow(2);
+    Scalar rrow(3);
+    Scalar rrow(4);
+    Scalar rrow(5);
+    Scalar rrow(6);
+    Scalar rrow(7);
+
+    // Now x corresponds to k, y to m, and z to n
+    const Scalar* lhs_block = &lhs_shmem[hipThreadIdx_x + 9 * hipThreadIdx_y];
+    const Scalar* rhs_block = &rhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_z];
+
+#define lhs_element(i, j) lhs_block[72 * ((i) + 8 * (j))]
+#define rhs_element(i, j) rhs_block[72 * ((i) + 8 * (j))]
+
+#define loadData(i, j)                          \
+    lcol(0) = lhs_element(0, j);               \
+    rrow(0) = rhs_element(i, 0);               \
+    lcol(1) = lhs_element(1, j);               \
+    rrow(1) = rhs_element(i, 1);               \
+    lcol(2) = lhs_element(2, j);               \
+    rrow(2) = rhs_element(i, 2);               \
+    lcol(3) = lhs_element(3, j);               \
+    rrow(3) = rhs_element(i, 3);               \
+    lcol(4) = lhs_element(4, j);               \
+    rrow(4) = rhs_element(i, 4);               \
+    lcol(5) = lhs_element(5, j);               \
+    rrow(5) = rhs_element(i, 5);               \
+    lcol(6) = lhs_element(6, j);               \
+    rrow(6) = rhs_element(i, 6);               \
+    lcol(7) = lhs_element(7, j);               \
+    rrow(7) = rhs_element(i, 7);               \
+
+#define computeCol(j)                           \
+    res(0, j) += lcol(0) * rrow(j);             \
+    res(1, j) += lcol(1) * rrow(j);             \
+    res(2, j) += lcol(2) * rrow(j);             \
+    res(3, j) += lcol(3) * rrow(j);             \
+    res(4, j) += lcol(4) * rrow(j);             \
+    res(5, j) += lcol(5) * rrow(j);             \
+    res(6, j) += lcol(6) * rrow(j);             \
+    res(7, j) += lcol(7) * rrow(j);             \
+
+#define computePass(i)                          \
+    loadData(i, i);                             \
+                                                \
+    computeCol(0);                              \
+    computeCol(1);                              \
+    computeCol(2);                              \
+    computeCol(3);                              \
+    computeCol(4);                              \
+    computeCol(5);                              \
+    computeCol(6);                              \
+    computeCol(7);                              \
+
+    computePass(0);
+    computePass(1);
+    computePass(2);
+    computePass(3);
+    computePass(4);
+    computePass(5);
+    computePass(6);
+    computePass(7);
+
+#undef lcol
+#undef rrow
+#undef lhs_element
+#undef rhs_element
+#undef loadData
+#undef computeCol
+#undef computePass
+  } // end loop over k
+
+  // we've now iterated over all of the large (ie width 64) k blocks and
+  // accumulated results in registers. At this point thread (x, y, z) contains
+  // the sum across all big k blocks of the product of little k block of index (x, y)
+  // with block of index (y, z). To compute the final output, we need to reduce
+  // the 8 threads over y by summation.
+#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
+
+#define reduceRow(i, mask)                      \
+  shuffleInc(i, 0, mask);                       \
+  shuffleInc(i, 1, mask);                       \
+  shuffleInc(i, 2, mask);                       \
+  shuffleInc(i, 3, mask);                       \
+  shuffleInc(i, 4, mask);                       \
+  shuffleInc(i, 5, mask);                       \
+  shuffleInc(i, 6, mask);                       \
+  shuffleInc(i, 7, mask);                       \
+
+#define reduceMatrix(mask)                      \
+  reduceRow(0, mask);                           \
+  reduceRow(1, mask);                           \
+  reduceRow(2, mask);                           \
+  reduceRow(3, mask);                           \
+  reduceRow(4, mask);                           \
+  reduceRow(5, mask);                           \
+  reduceRow(6, mask);                           \
+  reduceRow(7, mask);                           \
+
+  // actually perform the reduction, now each thread of index (_, y, z)
+  // contains the correct values in its registers that belong in the output
+  // block
+  reduceMatrix(1);
+  reduceMatrix(2);
+  reduceMatrix(4);
+
+#undef shuffleInc
+#undef reduceRow
+#undef reduceMatrix
+
+  // now we need to copy the 64 values into main memory. We can't split work
+  // among threads because all variables are in registers. There's 2 ways
+  // to do this:
+  // (1) have 1 thread do 64 writes from registers into global memory
+  // (2) have 1 thread do 64 writes into shared memory, and then 8 threads
+  //     each do 8 writes into global memory. We can just overwrite the shared
+  //     memory from the problem we just solved.
+  // (2) is slightly faster than (1) due to less branching and more ILP
+
+  // TODO: won't yield much gain, but could just use currently unused shared mem
+  //       and then we won't have to sync
+  // wait for shared mem to be out of use
+  __syncthreads();
+
+#define writeResultShmem(i, j)                                          \
+  lhs_shmem[i + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * j] = res(i, j); \
+
+#define writeRow(i)                             \
+  writeResultShmem(i, 0);                       \
+  writeResultShmem(i, 1);                       \
+  writeResultShmem(i, 2);                       \
+  writeResultShmem(i, 3);                       \
+  writeResultShmem(i, 4);                       \
+  writeResultShmem(i, 5);                       \
+  writeResultShmem(i, 6);                       \
+  writeResultShmem(i, 7);                       \
+
+  if (hipThreadIdx_x == 0) {
+    writeRow(0);
+    writeRow(1);
+    writeRow(2);
+    writeRow(3);
+    writeRow(4);
+    writeRow(5);
+    writeRow(6);
+    writeRow(7);
+  }
+#undef writeResultShmem
+#undef writeRow
+
+  const int max_i_write = numext::mini((int)((m_size - base_m - hipThreadIdx_y + 7) / 8), 8);
+  const int max_j_write = numext::mini((int)((n_size - base_n - hipThreadIdx_z + 7) / 8), 8);
+
+  if (hipThreadIdx_x < max_i_write) {
+    if (max_j_write == 8) {
+      // TODO: can i trade bank conflicts for coalesced writes?
+      Scalar val0 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 0];
+      Scalar val1 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 1];
+      Scalar val2 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 2];
+      Scalar val3 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 3];
+      Scalar val4 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 4];
+      Scalar val5 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 5];
+      Scalar val6 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 6];
+      Scalar val7 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 7];
+
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 0) = val0;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 1) = val1;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 2) = val2;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 3) = val3;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 4) = val4;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 5) = val5;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 6) = val6;
+      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 7) = val7;
+    } else {
+#pragma unroll 7
+      for (int j = 0; j < max_j_write; j++) {
+        Scalar val = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * j];
+        output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * j) = val;
+      }
+    }
+  }
+#undef res
+}
+
+
+template<typename Scalar, typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+__launch_bounds__(512, 1)
+EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ Scalar lhs_shmem[72 * 64];
+  __shared__ Scalar rhs_shmem[72 * 64];
+
+  const Index m_block_idx = hipBlockIdx_x;
+  const Index n_block_idx = hipBlockIdx_y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  if (base_m + 63 < m_size && base_n + 63 < n_size) {
+    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
+  } else {
+    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
+         bool CHECK_RHS_BOUNDARY>
+__device__ EIGEN_STRONG_INLINE void
+EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output, float2 lhs_shmem2[][16],
+                       float2 rhs_shmem2[][8], const Index m_size,
+                       const Index n_size, const Index k_size,
+                       const Index base_m, const Index base_n) {
+
+  // prefetch registers
+  float4 lhs_pf0, rhs_pf0;
+
+  float4 results[4];
+  for (int i=0; i < 4; i++) {
+    results[i].x = results[i].y = results[i].z = results[i].w = 0;
+  }
+
+
+#define prefetch_lhs(reg, row, col)                   \
+    if (!CHECK_LHS_BOUNDARY) {                        \
+      if (col < k_size) {                             \
+        /*reg = lhs.template loadPacket<Unaligned>(row, col);*/     \
+        reg.x =lhs(row + 0, col);                     \
+        reg.y =lhs(row + 1, col);                     \
+        reg.z =lhs(row + 2, col);                     \
+        reg.w =lhs(row + 3, col);                     \
+      }                                               \
+    } else {                                          \
+      if (col < k_size) {                             \
+        if (row + 3 < m_size) {                       \
+          /*reg =lhs.template loadPacket<Unaligned>(row, col);*/   \
+          reg.x =lhs(row + 0, col);                   \
+          reg.y =lhs(row + 1, col);                   \
+          reg.z =lhs(row + 2, col);                   \
+          reg.w =lhs(row + 3, col);                   \
+        } else if (row + 2 < m_size) {                \
+          reg.x =lhs(row + 0, col);                   \
+          reg.y =lhs(row + 1, col);                   \
+          reg.z =lhs(row + 2, col);                   \
+        } else if (row + 1 < m_size) {                \
+          reg.x =lhs(row + 0, col);                   \
+          reg.y =lhs(row + 1, col);                   \
+        } else if (row  < m_size) {                   \
+          reg.x =lhs(row + 0, col);                   \
+        }                                             \
+      }                                               \
+    }                                                 \
+
+
+  Index lhs_vert = base_m+hipThreadIdx_x*4;
+
+  for (Index k = 0; k < k_size; k += 16) {
+    //lhs_pf0 = internal::pset1<float4>(0);
+    //rhs_pf0 = internal::pset1<float4>(0);
+    lhs_pf0 = make_float4(0, 0, 0, 0);
+    rhs_pf0 = make_float4(0, 0, 0, 0);
+
+    Index lhs_horiz = hipThreadIdx_y+k;
+    prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
+
+    Index rhs_vert = k+(hipThreadIdx_x%4)*4;
+    Index rhs_horiz0 = (hipThreadIdx_x>>2)+hipThreadIdx_y*4+base_n;
+
+    if (!CHECK_RHS_BOUNDARY) {
+      if ((rhs_vert + 3) < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
+      } else if (rhs_vert + 2 < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+      } else if (rhs_vert + 1 < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+      } else if (rhs_vert  < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+      }
+    } else {
+      if (rhs_horiz0 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
+        } else if ((rhs_vert + 2) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        } else if ((rhs_vert + 1) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        } else if (rhs_vert  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        }
+      }
+    }
+    float x1, x2 ;
+    // the following can be a bitwise operation..... some day.
+    if((hipThreadIdx_x%8) < 4) {
+      x1 = rhs_pf0.y;
+      x2 = rhs_pf0.w;
+    } else {
+      x1 = rhs_pf0.x;
+      x2 = rhs_pf0.z;
+    }
+    x1 = __shfl_xor(x1, 4);
+    x2 = __shfl_xor(x2, 4);
+    if((hipThreadIdx_x%8) < 4) {
+      rhs_pf0.y = x1;
+      rhs_pf0.w = x2;
+    } else {
+      rhs_pf0.x = x1;
+      rhs_pf0.z = x2;
+    }
+
+    // We have 64 features.
+    // Row 0 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 0, 1.
+    // Row 1 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 2, 3.
+    // ...
+    // Row 31 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 62, 63
+    // Row 32 -> times (2, 6, 10, 14, 3, 7, 11, 15) for features 0, 1
+    // ...
+    rhs_shmem2[(hipThreadIdx_x>>3)+ hipThreadIdx_y*2][hipThreadIdx_x%8] = make_float2(rhs_pf0.x, rhs_pf0.y);
+    rhs_shmem2[(hipThreadIdx_x>>3)+ hipThreadIdx_y*2+32][hipThreadIdx_x%8] = make_float2(rhs_pf0.z, rhs_pf0.w);
+
+    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // ...
+    // Row 15 (time 15) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // Row 16 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63)
+    // ...
+
+    lhs_shmem2[hipThreadIdx_y][hipThreadIdx_x] = make_float2(lhs_pf0.x, lhs_pf0.y);
+    lhs_shmem2[hipThreadIdx_y+16][hipThreadIdx_x] = make_float2(lhs_pf0.z, lhs_pf0.w);
+
+
+#define add_vals(fl1, fl2, fr1, fr2)\
+    results[0].x += fl1.x * fr1.x;\
+    results[0].y += fl1.y * fr1.x;\
+    results[0].z += fl2.x * fr1.x;\
+    results[0].w += fl2.y * fr1.x;\
+\
+    results[1].x += fl1.x * fr1.y;\
+    results[1].y += fl1.y * fr1.y;\
+    results[1].z += fl2.x * fr1.y;\
+    results[1].w += fl2.y * fr1.y;\
+\
+    results[2].x += fl1.x * fr2.x;\
+    results[2].y += fl1.y * fr2.x;\
+    results[2].z += fl2.x * fr2.x;\
+    results[2].w += fl2.y * fr2.x;\
+\
+    results[3].x += fl1.x * fr2.y;\
+    results[3].y += fl1.y * fr2.y;\
+    results[3].z += fl2.x * fr2.y;\
+    results[3].w += fl2.y * fr2.y;\
+
+    __syncthreads();
+
+    // Do the multiplies.
+    #pragma unroll
+    for (int koff = 0; koff < 16; koff ++) {
+      // 32 x threads.
+      float2 fl1 = lhs_shmem2[koff][hipThreadIdx_x];
+      float2 fl2 = lhs_shmem2[koff + 16][hipThreadIdx_x];
+
+      int start_feature = hipThreadIdx_y * 4;
+      float2 fr1 = rhs_shmem2[(start_feature>>1) + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
+      float2 fr2 = rhs_shmem2[(start_feature>>1) + 1 + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
+
+      add_vals(fl1, fl2, fr1, fr2)
+    }
+    __syncthreads();
+  }
+
+#undef prefetch_lhs
+#undef add_vals
+
+  Index horiz_base = hipThreadIdx_y*4+base_n;
+  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
+    for (int i = 0; i < 4; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }
+  } else if (!CHECK_RHS_BOUNDARY) {
+    // CHECK LHS
+    if (lhs_vert + 3 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    } else if (lhs_vert + 2 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      }
+    } else if (lhs_vert + 1 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      }
+    } else if (lhs_vert  < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+      }
+    }
+  } else if (!CHECK_LHS_BOUNDARY) {
+    // CHECK RHS
+    /*
+    int ncols_rem = fminf(n_size- horiz_base, 4);
+    for (int i = 0; i < ncols_rem; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }*/
+    for (int i = 0; i < 4; i++) {
+      if (horiz_base+i < n_size) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+       }
+    }
+  } else {
+    // CHECK both boundaries.
+    for (int i = 0; i < 4; i++) {
+      if (horiz_base+i < n_size) {
+        if (lhs_vert < m_size)
+          output(lhs_vert, horiz_base + i) = results[i].x;
+        if (lhs_vert + 1 < m_size)
+          output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        if (lhs_vert + 2 < m_size)
+          output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        if (lhs_vert + 3 < m_size)
+          output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
+         bool CHECK_RHS_BOUNDARY>
+__device__ EIGEN_STRONG_INLINE void
+EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output, float2 lhs_shmem2[][32],
+                       float2 rhs_shmem2[][8], const Index m_size,
+                       const Index n_size, const Index k_size,
+                       const Index base_m, const Index base_n) {
+
+  // prefetch registers
+  float4 lhs_pf0, lhs_pf1, lhs_pf2, lhs_pf3;
+  float4 rhs_pf0, rhs_pf1;
+
+  float4 results[8];
+  for (int i=0; i < 8; i++) {
+    results[i].x = results[i].y = results[i].z = results[i].w = 0;
+  }
+
+
+  Index lhs_vert = base_m+hipThreadIdx_x*4+(hipThreadIdx_y%4)*32;
+  for (Index k = 0; k < k_size; k += 32) {
+    /*lhs_pf0 = internal::pset1<float4>(0);
+    lhs_pf1 = internal::pset1<float4>(0);
+    lhs_pf2 = internal::pset1<float4>(0);
+    lhs_pf3 = internal::pset1<float4>(0);
+
+    rhs_pf0 = internal::pset1<float4>(0);
+    rhs_pf1 = internal::pset1<float4>(0);*/
+
+
+    lhs_pf0 = make_float4(0, 0, 0, 0);
+    lhs_pf1 = make_float4(0, 0, 0, 0);
+    lhs_pf2 = make_float4(0, 0, 0, 0);
+    lhs_pf3 = make_float4(0, 0, 0, 0);
+
+    rhs_pf0 = make_float4(0, 0, 0, 0);
+    rhs_pf1 = make_float4(0, 0, 0, 0);
+
+     if (!CHECK_LHS_BOUNDARY) {
+      if ((hipThreadIdx_y/4+k+24) < k_size) {
+        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+        //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
+        //lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+24));
+        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+        lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
+        lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
+        lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
+        lhs_pf3.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
+      } else if ((hipThreadIdx_y/4+k+16) < k_size) {
+        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+        //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
+        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+        lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+      } else if ((hipThreadIdx_y/4+k+8) < k_size) {
+        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+      } else if ((hipThreadIdx_y/4+k) < k_size) {
+        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+      }
+    } else {
+      // just CHECK_LHS_BOUNDARY
+      if (lhs_vert + 3 < m_size) {
+        if ((hipThreadIdx_y/4+k+24) < k_size) {
+          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+          //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
+          //lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+24));
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
+        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
+          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+          //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
+          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        } else if ((hipThreadIdx_y/4+k) < k_size) {
+          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        }
+      } else if (lhs_vert + 2 < m_size) {
+        if ((hipThreadIdx_y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
+        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
+        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
+        } else if ((hipThreadIdx_y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
+        }
+      } else if (lhs_vert + 1 < m_size) {
+        if ((hipThreadIdx_y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
+          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
+        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
+        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
+        } else if ((hipThreadIdx_y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
+        }
+      } else if (lhs_vert < m_size) {
+        if ((hipThreadIdx_y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
+        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
+        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
+        } else if ((hipThreadIdx_y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
+        }
+      }
+    }
+    __syncthreads();
+    Index rhs_vert = k+hipThreadIdx_x*4;
+    Index rhs_horiz0 = hipThreadIdx_y*2+base_n;
+    Index rhs_horiz1 = hipThreadIdx_y*2+1+base_n;
+    if (!CHECK_RHS_BOUNDARY) {
+      if ((rhs_vert + 3) < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        //rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+        rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+        rhs_pf1.w = rhs(rhs_vert + 3, rhs_horiz1);
+      } else if (rhs_vert + 2 < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+        rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+      } else if (rhs_vert + 1 < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+      } else if (rhs_vert  < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+      }
+    } else {
+      if (rhs_horiz1 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          //rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+          rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+          rhs_pf1.w = rhs(rhs_vert + 3, rhs_horiz1);
+        } else if (rhs_vert + 2 < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+          rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+        } else if (k+hipThreadIdx_x*4 + 1 < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+        } else if (k+hipThreadIdx_x*4  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        }
+      } else if (rhs_horiz0 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
+        } else if ((rhs_vert + 2) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        } else if ((rhs_vert + 1) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        } else if (rhs_vert  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        }
+      }
+    }
+    __syncthreads();
+    // Loaded. Do computation
+    // Row 0 -> times (0, 4, 8, .. 28) for features 0, 1.
+    // Row 1 -> times (0, 4, 8, .. 28) for features 2, 3.
+    // ..
+    // Row 31 -> times (0, 4, 8, .. 28) for features 62, 63
+    rhs_shmem2[hipThreadIdx_y][hipThreadIdx_x] = make_float2(rhs_pf0.x, rhs_pf1.x);
+    // Row 32 -> times (1, 5, 9, .. 29) for features 0, 1.
+    // Row 33 -> times (1, 5, 9, .. 29) for features 2, 3.
+    // ..
+    rhs_shmem2[hipThreadIdx_y+32][hipThreadIdx_x] = make_float2(rhs_pf0.y, rhs_pf1.y);
+    // Row 64 -> times (2, 6, 10, .. 30) for features 0, 1.
+    // Row 65 -> times (2, 6, 10, .. 30) for features 2, 3.
+    rhs_shmem2[hipThreadIdx_y+64][hipThreadIdx_x] = make_float2(rhs_pf0.z, rhs_pf1.z);
+    // Row 96 -> times (3, 7, 11, .. 31) for features 0, 1.
+    // Row 97 -> times (3, 7, 11, .. 31) for features 2, 3.
+    rhs_shmem2[hipThreadIdx_y+96][hipThreadIdx_x] = make_float2(rhs_pf0.w, rhs_pf1.w);
+
+    // LHS.
+    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
+    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
+    // ...
+    // Row 8 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
+    // Row 15 (time 7) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
+
+
+#define add_vals(a_feat1, a_feat2, f1, f2, f3, f4)\
+      results[0].x += a_feat1.x * f1.x;\
+      results[1].x += a_feat1.x * f1.y;\
+      results[2].x += a_feat1.x * f2.x;\
+      results[3].x += a_feat1.x * f2.y;\
+      results[4].x += a_feat1.x * f3.x;\
+      results[5].x += a_feat1.x * f3.y;\
+      results[6].x += a_feat1.x * f4.x;\
+      results[7].x += a_feat1.x * f4.y;\
+\
+      results[0].y += a_feat1.y * f1.x;\
+      results[1].y += a_feat1.y * f1.y;\
+      results[2].y += a_feat1.y * f2.x;\
+      results[3].y += a_feat1.y * f2.y;\
+      results[4].y += a_feat1.y * f3.x;\
+      results[5].y += a_feat1.y * f3.y;\
+      results[6].y += a_feat1.y * f4.x;\
+      results[7].y += a_feat1.y * f4.y;\
+\
+      results[0].z += a_feat2.x * f1.x;\
+      results[1].z += a_feat2.x * f1.y;\
+      results[2].z += a_feat2.x * f2.x;\
+      results[3].z += a_feat2.x * f2.y;\
+      results[4].z += a_feat2.x * f3.x;\
+      results[5].z += a_feat2.x * f3.y;\
+      results[6].z += a_feat2.x * f4.x;\
+      results[7].z += a_feat2.x * f4.y;\
+\
+      results[0].w += a_feat2.y * f1.x;\
+      results[1].w += a_feat2.y * f1.y;\
+      results[2].w += a_feat2.y * f2.x;\
+      results[3].w += a_feat2.y * f2.y;\
+      results[4].w += a_feat2.y * f3.x;\
+      results[5].w += a_feat2.y * f3.y;\
+      results[6].w += a_feat2.y * f4.x;\
+      results[7].w += a_feat2.y * f4.y;\
+
+    lhs_shmem2[hipThreadIdx_y/4][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf0.x, lhs_pf0.y);
+    lhs_shmem2[hipThreadIdx_y/4+8][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf1.x, lhs_pf1.y);
+    lhs_shmem2[hipThreadIdx_y/4+16][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf2.x, lhs_pf2.y);
+    lhs_shmem2[hipThreadIdx_y/4+24][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf3.x, lhs_pf3.y);
+
+    lhs_shmem2[hipThreadIdx_y/4 + 32][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf0.z, lhs_pf0.w);
+    lhs_shmem2[hipThreadIdx_y/4 + 40][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf1.z, lhs_pf1.w);
+    lhs_shmem2[hipThreadIdx_y/4 + 48][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf2.z, lhs_pf2.w);
+    lhs_shmem2[hipThreadIdx_y/4 + 56][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf3.z, lhs_pf3.w);
+
+    __syncthreads();
+
+    // Do the multiplies.
+    #pragma unroll
+    for (int koff = 0; koff < 32; koff ++) {
+      float2 a3 = lhs_shmem2[koff][hipThreadIdx_x + (hipThreadIdx_y % 4) * 8];
+      float2 a4 = lhs_shmem2[koff + 32][hipThreadIdx_x + (hipThreadIdx_y % 4) * 8];
+
+      // first feature is at (hipThreadIdx_y/4) * 8 last is at start + 8.
+      int start_feature = (hipThreadIdx_y / 4) * 8;
+
+      float2 br1 = rhs_shmem2[start_feature/2 +     (koff % 4) * 32][koff/4];
+      float2 br2 = rhs_shmem2[start_feature/2 + 1 + (koff % 4) * 32][koff/4];
+      float2 br3 = rhs_shmem2[start_feature/2 + 2 + (koff % 4) * 32][koff/4];
+      float2 br4 = rhs_shmem2[start_feature/2 + 3 + (koff % 4) * 32][koff/4];
+
+      add_vals(a3, a4, br1, br2, br3, br4)
+    }
+    __syncthreads();
+  } // end loop over k
+
+
+  __syncthreads();
+  Index horiz_base = (hipThreadIdx_y/4)*8+base_n;
+  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
+    for (int i = 0; i < 8; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }
+  } else if (!CHECK_RHS_BOUNDARY) {
+    if (lhs_vert + 3 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    } else if (lhs_vert + 2 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      }
+    } else if (lhs_vert + 1 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      }
+    } else if (lhs_vert  < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+      }
+    }
+  } else if (!CHECK_LHS_BOUNDARY) {
+    // CHECK BOUNDARY_B
+    for (int i = 0; i < 8; i++) {
+      if (horiz_base + i < n_size) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  } else {
+    // CHECK both boundaries.
+    for (int i = 0; i < 8; i++) {
+      if (horiz_base + i < n_size) {
+        if (lhs_vert < m_size)
+          output(lhs_vert, horiz_base + i) = results[i].x;
+        if (lhs_vert + 1 < m_size)
+          output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        if (lhs_vert + 2 < m_size)
+          output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        if (lhs_vert + 3 < m_size)
+          output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+__launch_bounds__(256, 1)
+EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ float2 lhs_shmem[64*32];
+  __shared__ float2 rhs_shmem[128*8];
+
+  typedef float2 LHS_MEM[64][32];
+  typedef float2 RHS_MEM[128][8];
+
+  typedef float2 LHS_MEM16x16[32][16];
+  typedef float2 RHS_MEM16x16[64][8];
+
+  const Index m_block_idx = hipBlockIdx_x;
+  const Index n_block_idx = hipBlockIdx_y;
+
+  const Index base_m = 128 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  bool check_rhs = (base_n + 63) >= n_size;
+  bool check_lhs128 = (base_m + 127) >= m_size;
+
+  if (!check_rhs) {
+    if (!check_lhs128) {
+      // >= 128 rows left
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    }
+  } else {
+    if (!check_lhs128) {
+      // >= 128 rows left
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    }
+  }
+}
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+__launch_bounds__(256, 1)
+EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ float2 lhs_shmem[32][16];
+  __shared__ float2 rhs_shmem[64][8];
+
+  const Index m_block_idx = hipBlockIdx_x;
+  const Index n_block_idx = hipBlockIdx_y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  if (base_m + 63 < m_size) {
+    if (base_n + 63 < n_size) {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    }
+  } else {
+    if (base_n + 63 < n_size) {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    }
+  }
+}
+
+
+template<typename Indices, typename LeftArgType, typename RightArgType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> > {
+
+  typedef GpuDevice Device;
+
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
+  typedef TensorContractionEvaluatorBase<Self> Base;
+
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
+
+  enum {
+    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
+  };
+
+  // Most of the code is assuming that both input tensors are ColMajor. If the
+  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
+  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
+  // will pretend B is LHS and A is RHS.
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
+
+  static const int LDims =
+      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
+  static const int RDims =
+      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
+  static const int ContractDims = internal::array_size<Indices>::value;
+
+  typedef array<Index, LDims> left_dim_mapper_t;
+  typedef array<Index, RDims> right_dim_mapper_t;
+
+  typedef array<Index, ContractDims> contract_t;
+  typedef array<Index, LDims - ContractDims> left_nocontract_t;
+  typedef array<Index, RDims - ContractDims> right_nocontract_t;
+
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  // typedefs needed in evalTo
+  typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
+  typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
+
+  typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
+  typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
+
+  typedef typename LeftEvaluator::Dimensions LeftDimensions;
+  typedef typename RightEvaluator::Dimensions RightDimensions;
+
+  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) :
+      Base(op, device) {}
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
+
+  // We need to redefine this method to make hipcc happy
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
+    this->m_leftImpl.evalSubExprsIfNeeded(NULL);
+    this->m_rightImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      this->m_result = static_cast<Scalar *>(this->m_device.allocate(this->dimensions().TotalSize() * sizeof(Scalar)));
+      evalTo(this->m_result);
+      return true;
+    }
+  }
+
+  void evalTo(Scalar* buffer) const {
+    if (this->m_lhs_inner_dim_contiguous) {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, true, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<true, true, false, Unaligned>(buffer);
+        }
+      }
+      else {
+       if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, false, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<true, false, false, Unaligned>(buffer);
+        }
+      }
+    }
+    else {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, true, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<false, true, false, Unaligned>(buffer);
+        }
+      }
+      else {
+       if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, false, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<false, false, false, Unaligned>(buffer);
+        }
+      }
+    }
+  }
+
+  template <typename LhsScalar, typename RhsScalar, typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels {
+    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
+    const Index m_blocks = (m + 63) / 64;
+    const Index n_blocks = (n + 63) / 64;
+    const dim3 num_blocks(m_blocks, n_blocks, 1);
+    const dim3 block_size(8, 8, 8);
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>),
+                    dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
+    }
+  };
+
+  template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels<float, float, Index, LhsMapper, RhsMapper, OutputMapper> {
+    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
+      if (m < 768 || n < 768) {
+        const Index m_blocks = (m + 63) / 64;
+        const Index n_blocks = (n + 63) / 64;
+        const dim3 num_blocks(m_blocks, n_blocks, 1);
+        const dim3 block_size(16, 16, 1);
+        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>),
+                        dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
+      } else {
+        const Index m_blocks = (m + 127) / 128;
+        const Index n_blocks = (n + 63) / 64;
+        const dim3 num_blocks(m_blocks, n_blocks, 1);
+        const dim3 block_size(8, 32, 1);
+        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>),
+                        dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
+      }
+    }
+  };
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
+  void evalTyped(Scalar* buffer) const {
+    // columns in left side, rows in right side
+    const Index k = this->m_k_size;
+    EIGEN_UNUSED_VARIABLE(k)
+
+    // rows in left side
+    const Index m = this->m_i_size;
+
+    // columns in right side
+    const Index n = this->m_j_size;
+
+    // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar)
+    this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
+
+    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
+                                                   LeftEvaluator, left_nocontract_t,
+                                                   contract_t, 4,
+                                                   lhs_inner_dim_contiguous,
+                                                   false, Unaligned> LhsMapper;
+
+    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
+                                                   RightEvaluator, right_nocontract_t,
+                                                   contract_t, 4,
+                                                   rhs_inner_dim_contiguous,
+                                                   rhs_inner_dim_reordered, Unaligned> RhsMapper;
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+
+
+    // initialize data mappers
+    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
+                  this->m_left_contracting_strides, this->m_k_strides);
+
+    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
+                  this->m_right_contracting_strides, this->m_k_strides);
+
+    OutputMapper output(buffer, m);
+
+    setHipSharedMemConfig(hipSharedMemBankSizeEightByte);
+    LaunchKernels<LhsScalar, RhsScalar, Index, LhsMapper, RhsMapper, OutputMapper>::Run(lhs, rhs, output,  m, n, k, this->m_device);
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_USE_GPU and EIGEN_HIPCC
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
new file mode 100644
index 000000000..ba9971050
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
@@ -0,0 +1,1119 @@
+//#include "hip/hip_runtime.h"
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
+
+namespace Eigen {
+
+/** \class TensorConvolution
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor convolution class.
+  *
+  *
+  */
+namespace internal {
+
+template <typename Index, typename InputDims, int NumKernelDims, int Layout>
+class IndexMapper {
+ public:
+  IndexMapper(const InputDims& input_dims, const array<Index, NumKernelDims>& kernel_dims,
+              const array<Index, NumKernelDims>& indices) {
+
+    array<Index, NumDims> dimensions = input_dims;
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = indices[i];
+      const Index input_dim = input_dims[index];
+      const Index kernel_dim = kernel_dims[i];
+      const Index result_dim = input_dim - kernel_dim + 1;
+      dimensions[index] = result_dim;
+    }
+
+    array<Index, NumDims> inputStrides;
+    array<Index, NumDims> outputStrides;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      inputStrides[0] = 1;
+      outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        inputStrides[i] = inputStrides[i-1] * input_dims[i-1];
+        outputStrides[i] = outputStrides[i-1] * dimensions[i-1];
+      }
+    } else {
+      inputStrides[NumDims - 1] = 1;
+      outputStrides[NumDims - 1] = 1;
+      for (int i = static_cast<int>(NumDims) - 2; i >= 0; --i) {
+        inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
+        outputStrides[i] = outputStrides[i + 1] * dimensions[i + 1];
+      }
+    }
+
+    array<Index, NumDims> hipInputDimensions;
+    array<Index, NumDims> hipOutputDimensions;
+    array<Index, NumDims> tmp = dimensions;
+    array<Index, NumDims> ordering;
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = i + offset;
+      ordering[index] = indices[i];
+      tmp[indices[i]] = -1;
+      hipInputDimensions[index] = input_dims[indices[i]];
+      hipOutputDimensions[index] = dimensions[indices[i]];
+    }
+
+    int written = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                      ? NumKernelDims
+                      : 0;
+    for (int i = 0; i < NumDims; ++i) {
+      if (tmp[i] >= 0) {
+        ordering[written] = i;
+        hipInputDimensions[written] = input_dims[i];
+        hipOutputDimensions[written] = dimensions[i];
+        ++written;
+      }
+    }
+
+    for (int i = 0; i < NumDims; ++i) {
+      m_inputStrides[i] = inputStrides[ordering[i]];
+      m_outputStrides[i] = outputStrides[ordering[i]];
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumDims; ++i) {
+        if (i > NumKernelDims) {
+          m_hipInputStrides[i] =
+              m_hipInputStrides[i - 1] * hipInputDimensions[i - 1];
+          m_hipOutputStrides[i] =
+              m_hipOutputStrides[i - 1] * hipOutputDimensions[i - 1];
+        } else {
+          m_hipInputStrides[i] = 1;
+          m_hipOutputStrides[i] = 1;
+        }
+      }
+    } else {
+      for (int i = NumDims - 1; i >= 0; --i) {
+        if (i + 1 < offset) {
+          m_hipInputStrides[i] =
+              m_hipInputStrides[i + 1] * hipInputDimensions[i + 1];
+          m_hipOutputStrides[i] =
+              m_hipOutputStrides[i + 1] * hipOutputDimensions[i + 1];
+        } else {
+          m_hipInputStrides[i] = 1;
+          m_hipOutputStrides[i] = 1;
+        }
+      }
+    }
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputPlaneToTensorInputOffset(Index p) const {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int d = NumDims - 1; d > NumKernelDims; --d) {
+        const Index idx = p / m_hipInputStrides[d];
+        inputIndex += idx * m_inputStrides[d];
+        p -= idx * m_hipInputStrides[d];
+      }
+      inputIndex += p * m_inputStrides[NumKernelDims];
+    } else {
+      std::ptrdiff_t limit = 0;
+      if (NumKernelDims < NumDims) {
+        limit = NumDims - NumKernelDims - 1;
+      }
+      for (int d = 0; d < limit; ++d) {
+        const Index idx = p / m_hipInputStrides[d];
+        inputIndex += idx * m_inputStrides[d];
+        p -= idx * m_hipInputStrides[d];
+      }
+      inputIndex += p * m_inputStrides[limit];
+    }
+    return inputIndex;
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputPlaneToTensorOutputOffset(Index p) const {
+    Index outputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int d = NumDims - 1; d > NumKernelDims; --d) {
+        const Index idx = p / m_hipOutputStrides[d];
+        outputIndex += idx * m_outputStrides[d];
+        p -= idx * m_hipOutputStrides[d];
+      }
+      outputIndex += p * m_outputStrides[NumKernelDims];
+    } else {
+      std::ptrdiff_t limit = 0;
+      if (NumKernelDims < NumDims) {
+        limit = NumDims - NumKernelDims - 1;
+      }
+      for (int d = 0; d < limit; ++d) {
+        const Index idx = p / m_hipOutputStrides[d];
+        outputIndex += idx * m_outputStrides[d];
+        p -= idx * m_hipOutputStrides[d];
+      }
+      outputIndex += p * m_outputStrides[limit];
+    }
+    return outputIndex;
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i, Index j) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i, Index j) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1] +
+           k * m_inputStrides[offset + 2];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1] +
+           k * m_outputStrides[offset + 2];
+  }
+
+ private:
+  static const int NumDims = internal::array_size<InputDims>::value;
+  array<Index, NumDims> m_inputStrides;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_hipInputStrides;
+  array<Index, NumDims> m_hipOutputStrides;
+};
+
+
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct traits<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename promote_storage_type<typename InputXprType::Scalar,
+                                        typename KernelXprType::Scalar>::ret Scalar;
+  typedef typename promote_storage_type<typename traits<InputXprType>::StorageKind,
+                                        typename traits<KernelXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<InputXprType>::Index,
+                                      typename traits<KernelXprType>::Index>::type Index;
+  typedef typename InputXprType::Nested LhsNested;
+  typedef typename KernelXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = traits<InputXprType>::NumDimensions;
+  static const int Layout = traits<InputXprType>::Layout;
+
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, Eigen::Dense>
+{
+  typedef const TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>& type;
+};
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct nested<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, 1, typename eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >::type>
+{
+  typedef TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Indices, typename InputXprType, typename KernelXprType>
+class TensorConvolutionOp : public TensorBase<TensorConvolutionOp<Indices, InputXprType, KernelXprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::promote_storage_type<typename InputXprType::CoeffReturnType,
+                                                  typename KernelXprType::CoeffReturnType>::ret CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorConvolutionOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConvolutionOp(const InputXprType& input, const KernelXprType& kernel, const Indices& dims)
+      : m_input_xpr(input), m_kernel_xpr(kernel), m_indices(dims) {}
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Indices& indices() const { return m_indices; }
+
+    /** \returns the nested expressions */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename InputXprType::Nested>::type&
+    inputExpression() const { return m_input_xpr; }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename KernelXprType::Nested>::type&
+    kernelExpression() const { return m_kernel_xpr; }
+
+  protected:
+    typename InputXprType::Nested m_input_xpr;
+    typename KernelXprType::Nested m_kernel_xpr;
+    const Indices m_indices;
+};
+
+
+template<typename Indices, typename InputArgType, typename KernelArgType, typename Device>
+struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, Device>
+{
+  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
+
+  static const int NumDims = internal::array_size<typename TensorEvaluator<InputArgType, Device>::Dimensions>::value;
+  static const int NumKernelDims = internal::array_size<Indices>::value;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+
+  enum {
+    IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
+    PacketAccess = TensorEvaluator<InputArgType, Device>::PacketAccess & TensorEvaluator<KernelArgType, Device>::PacketAccess,
+    Layout = TensorEvaluator<InputArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_kernel(NULL), m_local_kernel(false), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<InputArgType, Device>::Dimensions& input_dims = m_inputImpl.dimensions();
+    const typename TensorEvaluator<KernelArgType, Device>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStride[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_inputStride[i] = m_inputStride[i - 1] * input_dims[i - 1];
+      }
+    } else {
+      m_inputStride[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_inputStride[i] = m_inputStride[i + 1] * input_dims[i + 1];
+      }
+    }
+
+    m_dimensions = m_inputImpl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumKernelDims; ++i) {
+        const Index index = op.indices()[i];
+        const Index input_dim = input_dims[index];
+        const Index kernel_dim = kernel_dims[i];
+        const Index result_dim = input_dim - kernel_dim + 1;
+        m_dimensions[index] = result_dim;
+        if (i > 0) {
+          m_kernelStride[i] = m_kernelStride[i - 1] * kernel_dims[i - 1];
+        } else {
+          m_kernelStride[0] = 1;
+        }
+        m_indexStride[i] = m_inputStride[index];
+      }
+
+      m_outputStride[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStride[i] = m_outputStride[i - 1] * m_dimensions[i - 1];
+      }
+    } else {
+      for (int i = NumKernelDims - 1; i >= 0; --i) {
+        const Index index = op.indices()[i];
+        const Index input_dim = input_dims[index];
+        const Index kernel_dim = kernel_dims[i];
+        const Index result_dim = input_dim - kernel_dim + 1;
+        m_dimensions[index] = result_dim;
+        if (i < NumKernelDims - 1) {
+          m_kernelStride[i] = m_kernelStride[i + 1] * kernel_dims[i + 1];
+        } else {
+          m_kernelStride[NumKernelDims - 1] = 1;
+        }
+        m_indexStride[i] = m_inputStride[index];
+      }
+
+      m_outputStride[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStride[i] = m_outputStride[i + 1] * m_dimensions[i + 1];
+      }
+    }
+  }
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
+    m_inputImpl.evalSubExprsIfNeeded(NULL);
+    preloadKernel();
+    return true;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
+    m_inputImpl.cleanup();
+    if (m_local_kernel) {
+      m_device.deallocate((void*)m_kernel);
+      m_local_kernel = false;
+    }
+    m_kernel = NULL;
+  }
+
+  void evalTo(typename XprType::Scalar* buffer) {
+    evalSubExprsIfNeeded(NULL);
+    for (int i = 0; i < dimensions().TotalSize(); ++i) {
+      buffer[i] += coeff(i);
+    }
+    cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    CoeffReturnType result = CoeffReturnType(0);
+    convolve(firstInput(index), 0, NumKernelDims-1, result);
+    return result;
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(const Index index) const
+  {
+    Index indices[2] = {index, index+PacketSize-1};
+    Index startInputs[2] = {0, 0};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / m_outputStride[i];
+        const Index idx1 = indices[1] / m_outputStride[i];
+        startInputs[0] += idx0 * m_inputStride[i];
+        startInputs[1] += idx1 * m_inputStride[i];
+        indices[0] -= idx0 * m_outputStride[i];
+        indices[1] -= idx1 * m_outputStride[i];
+      }
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / m_outputStride[i];
+        const Index idx1 = indices[1] / m_outputStride[i];
+        startInputs[0] += idx0 * m_inputStride[i];
+        startInputs[1] += idx1 * m_inputStride[i];
+        indices[0] -= idx0 * m_outputStride[i];
+        indices[1] -= idx1 * m_outputStride[i];
+      }
+    }
+    startInputs[0] += indices[0];
+    startInputs[1] += indices[1];
+
+    if (startInputs[1]-startInputs[0] == PacketSize-1) {
+      PacketReturnType result = internal::pset1<PacketReturnType>(0);
+      convolvePacket(startInputs[0], 0, NumKernelDims-1, result);
+      return result;
+    } else {
+      EIGEN_ALIGN_MAX Scalar data[PacketSize];
+      data[0] = Scalar(0);
+      convolve(startInputs[0], 0, NumKernelDims-1, data[0]);
+      for (int i = 1; i < PacketSize-1; ++i) {
+        data[i] = Scalar(0);
+        convolve(firstInput(index+i), 0, NumKernelDims-1, data[i]);
+      }
+      data[PacketSize-1] = Scalar(0);
+      convolve(startInputs[1], 0, NumKernelDims-1, data[PacketSize-1]);
+      return internal::pload<PacketReturnType>(data);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
+    // We ignore the use of fused multiply-add.
+    const double convolve_compute_cost =
+        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
+    const double firstIndex_compute_cost =
+        NumDims *
+        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
+         TensorOpCost::DivCost<Index>());
+    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
+           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
+                          m_kernelImpl.costPerCoeff(vectorized) +
+                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
+                                       PacketSize));
+  }
+
+  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
+
+ private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
+    Index startInput = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStride[i];
+        startInput += idx * m_inputStride[i];
+        index -= idx * m_outputStride[i];
+      }
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStride[i];
+        startInput += idx * m_inputStride[i];
+        index -= idx * m_outputStride[i];
+      }
+    }
+    startInput += index;
+    return startInput;
+  }
+
+  EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const {
+    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
+      const Index input = firstIndex + j * m_indexStride[DimIndex];
+      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
+      if (DimIndex > 0) {
+        convolve(input, kernel, DimIndex-1, accum);
+      } else {
+        accum += m_inputImpl.coeff(input) * m_kernel[kernel];
+      }
+    }
+  }
+
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC void convolvePacket(Index firstIndex, Index firstKernel, int DimIndex, Packet& accum) const {
+    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
+      const Index input = firstIndex + j * m_indexStride[DimIndex];
+      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
+      if (DimIndex > 0) {
+        convolvePacket(input, kernel, DimIndex-1, accum);
+      } else {
+        accum = internal::pmadd<Packet>(m_inputImpl.template packet<Unaligned>(input), internal::pset1<Packet>(m_kernel[kernel]), accum);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void preloadKernel() {
+    // Don't make a local copy of the kernel unless we have to (i.e. it's an
+    // expression that needs to be evaluated)
+    const Scalar* in_place = m_kernelImpl.data();
+    if (in_place) {
+      m_kernel = in_place;
+      m_local_kernel = false;
+    } else {
+      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
+      typedef TensorEvalToOp<const KernelArgType> EvalTo;
+      EvalTo evalToTmp(local, m_kernelArg);
+      const bool PacketAccess = internal::IsVectorizable<Device, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, Device, PacketAccess>::run(evalToTmp, m_device);
+
+      m_kernel = local;
+      m_local_kernel = true;
+    }
+  }
+
+  array<Index, NumDims> m_inputStride;
+  array<Index, NumDims> m_outputStride;
+
+  array<Index, NumKernelDims> m_indexStride;
+  array<Index, NumKernelDims> m_kernelStride;
+  TensorEvaluator<InputArgType, Device> m_inputImpl;
+  TensorEvaluator<KernelArgType, Device> m_kernelImpl;
+  Dimensions m_dimensions;
+
+  KernelArgType m_kernelArg;
+  const Scalar* m_kernel;
+  bool m_local_kernel;
+  const Device& m_device;
+};
+
+
+
+
+// Use an optimized implementation of the evaluation code for GPUs whenever possible.
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
+
+template <int StaticKernelSize>
+struct GetKernelSize {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int /*kernelSize*/) const {
+    return StaticKernelSize;
+  }
+};
+template <>
+struct GetKernelSize<Dynamic> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int kernelSize) const {
+    return kernelSize;
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims,
+          int StaticKernelSize>
+__global__ void EigenConvolutionKernel1D(
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 1, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const int numPlanes, const int numX,
+    const int maxX, const int kernelSize, float* buffer) {
+  HIP_DYNAMIC_SHARED( float, s)
+
+  const int first_x = hipBlockIdx_x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSize>()(kernelSize);
+  const int num_x_output = last_x - first_x + 1;
+
+  const int first_plane = hipBlockIdx_y * hipBlockDim_y;
+  const int plane_stride = hipBlockDim_y * hipGridDim_y;
+
+  for (int p = first_plane + hipThreadIdx_y; p < numPlanes; p += plane_stride) {
+    // Load inputs to shared memory
+    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = hipThreadIdx_y * num_x_input;
+    #pragma unroll
+    for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
+      const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x);
+      s[i + plane_kernel_offset] = eval.coeff(tensor_index);
+    }
+
+    __syncthreads();
+
+    // Compute the convolution
+    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
+
+    #pragma unroll
+    for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
+      const int kernel_offset = plane_kernel_offset + i;
+      float result = 0.0f;
+      #pragma unroll
+      for (int k = 0; k < GetKernelSize<StaticKernelSize>()(kernelSize); ++k) {
+        result += s[k + kernel_offset] * kernel[k];
+      }
+      const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x);
+      buffer[tensor_index] = result;
+    }
+    __syncthreads();
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims,
+          int StaticKernelSizeX, int StaticKernelSizeY>
+__global__ void EigenConvolutionKernel2D(
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 2, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const int numPlanes, const int numX,
+    const int maxX, const int numY, const int maxY, const int kernelSizeX,
+    const int kernelSizeY, float* buffer) {
+  HIP_DYNAMIC_SHARED( float, s)
+
+  const int first_x = hipBlockIdx_x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSizeX>()(kernelSizeX);
+  const int num_x_output = last_x - first_x + 1;
+
+  const int first_y = hipBlockIdx_y * maxY;
+  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
+  const int num_y_input = last_y - first_y + GetKernelSize<StaticKernelSizeY>()(kernelSizeY);
+  const int num_y_output = last_y - first_y + 1;
+
+  const int first_plane = hipBlockIdx_z * hipBlockDim_z;
+  const int plane_stride = hipBlockDim_z * hipGridDim_z;
+
+  for (int p = first_plane + hipThreadIdx_z; p < numPlanes; p += plane_stride) {
+
+    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = hipThreadIdx_z * num_y_input;
+
+    // Load inputs to shared memory
+    #pragma unroll
+    for (int j = hipThreadIdx_y; j < num_y_input; j += hipBlockDim_y) {
+      const int input_offset = num_x_input * (j + plane_kernel_offset);
+      #pragma unroll
+      for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
+        const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x, j+first_y);
+        s[i + input_offset] = eval.coeff(tensor_index);
+      }
+    }
+
+    __syncthreads();
+
+    // Convolution
+    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
+
+    #pragma unroll
+    for (int j = hipThreadIdx_y; j < num_y_output; j += hipBlockDim_y) {
+      #pragma unroll
+      for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
+        float result = 0.0f;
+        #pragma unroll
+        for (int l = 0; l < GetKernelSize<StaticKernelSizeY>()(kernelSizeY); ++l) {
+          const int kernel_offset = kernelSizeX * l;
+          const int input_offset = i + num_x_input * (j + l + plane_kernel_offset);
+          #pragma unroll
+          for (int k = 0; k < GetKernelSize<StaticKernelSizeX>()(kernelSizeX); ++k) {
+            result += s[k + input_offset] * kernel[k + kernel_offset];
+          }
+        }
+        const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x, j+first_y);
+        buffer[tensor_index] = result;
+      }
+    }
+
+    __syncthreads();
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims>
+__global__ void EigenConvolutionKernel3D( 
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 3, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const size_t numPlanes, const size_t numX,
+    const size_t maxX, const size_t numY, const size_t maxY, const size_t numZ,
+    const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY,
+    const size_t kernelSizeZ, float* buffer) {
+  HIP_DYNAMIC_SHARED( float, s)
+
+  // Load inputs to shared memory
+  const int first_x = hipBlockIdx_x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + kernelSizeX;
+
+  const int first_y = hipBlockIdx_y * maxY;
+  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
+  const int num_y_input = last_y - first_y + kernelSizeY;
+
+  const int first_z = hipBlockIdx_z * maxZ;
+  const int last_z = (first_z + maxZ < numZ ? first_z + maxZ : numZ) - 1;
+  const int num_z_input = last_z - first_z + kernelSizeZ;
+
+  for (int p = 0; p < numPlanes; ++p) {
+
+    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = 0;
+
+    for (int k = hipThreadIdx_z; k < num_z_input; k += hipBlockDim_z) {
+      for (int j = hipThreadIdx_y; j < num_y_input; j += hipBlockDim_y) {
+        for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
+          const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z);
+          s[i + num_x_input * (j + num_y_input * (k + plane_kernel_offset))] = eval.coeff(tensor_index);
+        }
+      }
+    }
+
+    __syncthreads();
+
+    // Convolution
+    const int num_z_output = last_z - first_z + 1;
+    const int num_y_output = last_y - first_y + 1;
+    const int num_x_output = last_x - first_x + 1;
+    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
+
+    for (int k = hipThreadIdx_z; k < num_z_output; k += hipBlockDim_z) {
+      for (int j = hipThreadIdx_y; j < num_y_output; j += hipBlockDim_y) {
+        for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
+          float result = 0.0f;
+          for (int n = 0; n < kernelSizeZ; ++n) {
+            for (int m = 0; m < kernelSizeY; ++m) {
+              for (int l = 0; l < kernelSizeX; ++l) {
+                result += s[i + l + num_x_input * (j + m + num_y_input * (k + n + plane_kernel_offset))] * kernel[l + kernelSizeX * (m + kernelSizeY * n)];
+              }
+            }
+          }
+          const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z);
+          buffer[tensor_index] = result;
+        }
+      }
+    }
+    __syncthreads();
+  }
+};
+
+
+
+template<typename Indices, typename InputArgType, typename KernelArgType>
+struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, GpuDevice>
+{
+  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
+
+  static const int NumDims =  internal::array_size<typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions>::value;
+  static const int NumKernelDims = internal::array_size<Indices>::value;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions KernelDimensions;
+
+  enum {
+    IsAligned = TensorEvaluator<InputArgType, GpuDevice>::IsAligned & TensorEvaluator<KernelArgType, GpuDevice>::IsAligned,
+    PacketAccess = false,
+    Layout = TensorEvaluator<InputArgType, GpuDevice>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const GpuDevice& device)
+      : m_inputImpl(op.inputExpression(), device), m_kernelArg(op.kernelExpression()), m_kernelImpl(op.kernelExpression(), device), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, GpuDevice>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, GpuDevice>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions& input_dims = m_inputImpl.dimensions();
+    const typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
+
+    m_dimensions = m_inputImpl.dimensions();
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = op.indices()[i];
+      const Index input_dim = input_dims[index];
+      const Index kernel_dim = kernel_dims[i];
+      const Index result_dim = input_dim - kernel_dim + 1;
+      m_dimensions[index] = result_dim;
+    }
+  }
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
+  typedef typename InputArgType::Scalar Scalar;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
+    preloadKernel();
+    m_inputImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      executeEval(data);
+      return false;
+    } else {
+      m_buf = (Scalar*)m_device.allocate(dimensions().TotalSize() * sizeof(Scalar));
+      executeEval(m_buf);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_inputImpl.cleanup();
+    if (m_buf) {
+      m_device.deallocate(m_buf);
+      m_buf = NULL;
+    }
+    if (m_local_kernel) {
+      m_device.deallocate((void*)m_kernel);
+      m_local_kernel = false;
+    }
+    m_kernel = NULL;
+  }
+
+  EIGEN_STRONG_INLINE void preloadKernel() {
+    // Don't make a local copy of the kernel unless we have to (i.e. it's an
+    // expression that needs to be evaluated)
+    const Scalar* in_place = m_kernelImpl.data();
+    if (in_place) {
+      m_kernel = in_place;
+      m_local_kernel = false;
+    } else {
+      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
+      typedef TensorEvalToOp<const KernelArgType> EvalTo;
+      EvalTo evalToTmp(local, m_kernelArg);
+      const bool PacketAccess = internal::IsVectorizable<GpuDevice, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, GpuDevice, PacketAccess>::run(evalToTmp, m_device);
+
+      m_kernel = local;
+      m_local_kernel = true;
+    }
+  }
+
+  static unsigned int ceil(unsigned int num, unsigned int denom) {
+    const unsigned int rounded_toward_zero = num / denom;
+    if (num > rounded_toward_zero * denom) {
+      return rounded_toward_zero + 1;
+    }
+    return rounded_toward_zero;
+  }
+
+  void executeEval(Scalar* data) const {
+    typedef typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions InputDims;
+
+    const int maxSharedMem = m_device.sharedMemPerBlock();
+    const int maxThreadsPerBlock = m_device.maxHipThreadsPerBlock();
+    const int maxBlocksPerProcessor = m_device.maxHipThreadsPerMultiProcessor() / maxThreadsPerBlock;
+    const int numMultiProcessors = m_device.getNumHipMultiProcessors();
+    const int hipWarpSize = 32;
+
+    switch (NumKernelDims) {
+      case 1: {
+        const int kernel_size = m_kernelImpl.dimensions().TotalSize();
+
+        const int numX = dimensions()[m_indices[0]];
+        const int numP = dimensions().TotalSize() / numX;
+        int maxX;
+        dim3 block_size;
+
+        const int single_stride_dim =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                ? 0
+                : m_inputImpl.dimensions().rank() - 1;
+        if (m_indices[0] == single_stride_dim) {
+          // Maximum the reuse
+          const int inner_dim = ((maxSharedMem / (sizeof(Scalar)) - kernel_size + 1 + 31) / 32) * 32;
+          maxX = numext::mini<int>(inner_dim, numX);
+          const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size - 1 + maxX) * sizeof(Scalar)), numP);
+          block_size.x = numext::mini(maxThreadsPerBlock, maxX);
+          block_size.y = numext::mini<int>(maxThreadsPerBlock / block_size.x, maxP);
+        }
+        else {
+          // Read as much as possible alongside the inner most dimension, that is the plane
+          const int inner_dim = maxSharedMem / ((hipWarpSize + kernel_size) * sizeof(Scalar));
+          const int maxP = numext::mini<int>(inner_dim, numP);
+          maxX = numext::mini<int>(maxSharedMem / (inner_dim * sizeof(Scalar)) - kernel_size + 1, numX);
+
+          block_size.x = numext::mini(hipWarpSize, maxX);
+          block_size.y = numext::mini<int>(maxThreadsPerBlock/block_size.x, maxP);
+        }
+
+        const int shared_mem = block_size.y * (maxX + kernel_size - 1) * sizeof(Scalar);
+        assert(shared_mem <= maxSharedMem);
+
+        const int num_x_blocks = ceil(numX, maxX);
+        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
+        const int num_y_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks);
+
+        dim3 num_blocks(num_x_blocks, numext::mini<int>(num_y_blocks, ceil(numP, block_size.y)));
+
+
+        //cout << "launching 1D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " maxX: " << maxX << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+
+        const array<Index, 1> indices(m_indices[0]);
+        const array<Index, 1> kernel_dims(m_kernelImpl.dimensions()[0]);
+        internal::IndexMapper<Index, InputDims, 1, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+        switch(kernel_size) {
+          case 4: {
+            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4>),
+                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data);
+            break;
+          }
+          case 7: {
+            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7>),
+                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data);
+            break;
+          }
+          default: {
+            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic>),
+                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data);
+          }
+        }
+        break;
+      }
+
+      case 2: {
+        const int idxX =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 1;
+        const int idxY =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 0;
+        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
+        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
+
+        const int numX = dimensions()[m_indices[idxX]];
+        const int numY = dimensions()[m_indices[idxY]];
+        const int numP = dimensions().TotalSize() / (numX*numY);
+
+        const float scaling_factor = sqrtf(static_cast<float>(maxSharedMem) / (sizeof(Scalar) * kernel_size_y * kernel_size_x));
+
+        // Snap maxX to warp size
+        int inner_dim = ((static_cast<int>(scaling_factor * kernel_size_x) - kernel_size_x + 1 + 32) / 32) * 32;
+        const int maxX = numext::mini<int>(inner_dim, numX);
+        const int maxY = numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1)) - kernel_size_y + 1, numY);
+        const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size_x - 1 + maxX) * (kernel_size_y - 1 + maxY) * sizeof(Scalar)), numP);
+
+        dim3 block_size;
+        block_size.x = numext::mini(1024, maxX);
+        block_size.y = numext::mini<int>(1024/block_size.x, maxY);
+        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxP);
+
+        const int shared_mem = block_size.z * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * sizeof(Scalar);
+        assert(shared_mem <= maxSharedMem);
+
+        const int num_x_blocks = ceil(numX, maxX);
+        const int num_y_blocks = ceil(numY, maxY);
+        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
+        const int num_z_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks * num_y_blocks);
+
+        dim3 num_blocks(num_x_blocks, num_y_blocks, numext::mini<int>(num_z_blocks, ceil(numP, block_size.z)));
+
+
+        //cout << "launching 2D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z << " maxX: " << maxX << " maxY: " << maxY << " maxP: " << maxP << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+
+        const array<Index, 2> indices(m_indices[idxX], m_indices[idxY]);
+        const array<Index, 2> kernel_dims(m_kernelImpl.dimensions()[idxX],
+                                          m_kernelImpl.dimensions()[idxY]);
+        internal::IndexMapper<Index, InputDims, 2, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+        switch (kernel_size_x) {
+          case 4: {
+            switch (kernel_size_y) {
+              case 7: {
+                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, 7>),
+                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data);
+                break;
+              }
+              default: {
+                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, Dynamic>),
+                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data);
+                break;
+              }
+            }
+            break;
+          }
+          case 7: {
+            switch (kernel_size_y) {
+              case 4: {
+                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, 4>),
+                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data);
+                break;
+              }
+              default: {
+                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, Dynamic>),
+                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data);
+                break;
+              }
+            }
+            break;
+          }
+          default: {
+            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic, Dynamic>),
+                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data);
+            break;
+          }
+        }
+        break;
+      }
+
+      case 3: {
+        const int idxX =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 2;
+        const int idxY =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 1;
+        const int idxZ =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 2 : 0;
+
+        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
+        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
+        const int kernel_size_z = m_kernelImpl.dimensions()[idxZ];
+
+        const int numX = dimensions()[m_indices[idxX]];
+        const int numY = dimensions()[m_indices[idxY]];
+        const int numZ = dimensions()[m_indices[idxZ]];
+        const int numP = dimensions().TotalSize() / (numX*numY*numZ);
+
+        const int maxX = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * kernel_size_y * kernel_size_z) - kernel_size_x + 1, numX));
+        const int maxY = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * kernel_size_z) - kernel_size_y + 1, numY));
+        const int maxZ = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1)) - kernel_size_z + 1, numZ));
+
+        dim3 block_size;
+        block_size.x = numext::mini(32, maxX);
+        block_size.y = numext::mini(32, maxY);
+        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxZ);
+        dim3 num_blocks(ceil(numX, maxX), ceil(numY, maxY), ceil(numZ, maxZ));
+
+        const int shared_mem = (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * (maxZ + kernel_size_z - 1) * sizeof(Scalar);
+        assert(shared_mem <= maxSharedMem);
+
+        //cout << "launching 3D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z  << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+        const array<Index, 3> indices(m_indices[idxX], m_indices[idxY],
+                                      m_indices[idxZ]);
+        const array<Index, 3> kernel_dims(m_kernelImpl.dimensions()[idxX],
+                                          m_kernelImpl.dimensions()[idxY],
+                                          m_kernelImpl.dimensions()[idxZ]);
+        internal::IndexMapper<Index, InputDims, 3, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+
+        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel3D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims>),
+                        dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel,
+                        numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data);
+        break;
+      }
+
+      default: {
+        EIGEN_STATIC_ASSERT((NumKernelDims >= 1 && NumKernelDims <= 3), THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    eigen_assert(m_buf);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return m_buf[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(const Index index) const
+  {
+    eigen_assert(m_buf);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return internal::ploadt<PacketReturnType, LoadMode>(m_buf+index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): FIXME: For now, this is just a copy of the CPU cost
+    // model.
+    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
+    // We ignore the use of fused multiply-add.
+    const double convolve_compute_cost =
+        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
+    const double firstIndex_compute_cost =
+        NumDims *
+        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
+         TensorOpCost::DivCost<Index>());
+    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
+           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
+                          m_kernelImpl.costPerCoeff(vectorized) +
+                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
+                                       PacketSize));
+  }
+
+ private:
+  // No assignment (copies are needed by the kernels)
+  TensorEvaluator& operator = (const TensorEvaluator&);
+
+  TensorEvaluator<InputArgType, GpuDevice> m_inputImpl;
+  TensorEvaluator<KernelArgType, GpuDevice> m_kernelImpl;
+  KernelArgType m_kernelArg;
+  Indices m_indices;
+  Dimensions m_dimensions;
+  Scalar* m_buf;
+  const Scalar* m_kernel;
+  bool m_local_kernel;
+
+  const GpuDevice& m_device;
+};
+#endif
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
index 341889e88..e94e577fc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
@@ -35,9 +35,12 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const {
-#ifndef EIGEN_CUDA_ARCH
+#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     // Running on the host CPU
     return 1;
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return 64;
 #else
     // Running on a CUDA device
     return 32;
@@ -45,7 +48,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     // Running on the host CPU
     return l1CacheSize();
 #else
@@ -55,7 +58,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     // Running single threaded on the host CPU
     return l3CacheSize();
 #else
@@ -65,10 +68,14 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
-#ifndef EIGEN_CUDA_ARCH
+#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     // Running single threaded on the host CPU
     // Should return an enum that encodes the ISA supported by the CPU
     return 1;
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    // return 1 as major for HIP
+    return 1;
 #else
     // Running on a CUDA device
     return EIGEN_CUDA_ARCH / 100;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
new file mode 100644
index 000000000..c0e110987
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
@@ -0,0 +1,352 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H)
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H
+
+#if defined(EIGEN_HIPCC)
+#include "hip/hip_runtime.h"
+#include "hip/hip_runtime_api.h"
+#endif
+#include <unistd.h> //for sleep function
+
+namespace Eigen {
+
+static const int kHipScratchSize = 1024;
+
+// This defines an interface that GPUDevice can take to use
+// HIP streams underneath.
+class StreamInterface {
+ public:
+  virtual ~StreamInterface() {}
+
+  virtual const hipStream_t& stream() const = 0;
+  virtual const hipDeviceProp_t& deviceProperties() const = 0;
+
+  // Allocate memory on the actual device where the computation will run
+  virtual void* allocate(size_t num_bytes) const = 0;
+  virtual void deallocate(void* buffer) const = 0;
+
+  // Return a scratchpad buffer of size 1k
+  virtual void* scratchpad() const = 0;
+
+  // Return a semaphore. The semaphore is initially initialized to 0, and
+  // each kernel using it is responsible for resetting to 0 upon completion
+  // to maintain the invariant that the semaphore is always equal to 0 upon
+  // each kernel start.
+  virtual unsigned int* semaphore() const = 0;
+};
+
+static hipDeviceProp_t* m_deviceProperties;
+static bool m_devicePropInitialized = false;
+
+static void initializeDeviceProp() {
+  if (!m_devicePropInitialized) {
+    // Attempts to ensure proper behavior in the case of multiple threads
+    // calling this function simultaneously. This would be trivial to
+    // implement if we could use std::mutex, but unfortunately mutex don't
+    // compile with nvcc, so we resort to atomics and thread fences instead.
+    // Note that if the caller uses a compiler that doesn't support c++11 we
+    // can't ensure that the initialization is thread safe.
+#if 0 && __cplusplus >= 201103L
+    static std::atomic<bool> first(true);
+    if (first.exchange(false)) {
+#else
+    static bool first = true;
+    if (first) {
+      first = false;
+#endif
+      // We're the first thread to reach this point.
+      int num_devices;
+      hipError_t status = hipGetDeviceCount(&num_devices);
+      if (status != hipSuccess) {
+        std::cerr << "Failed to get the number of HIP devices: "
+                  << hipGetErrorString(status)
+                  << std::endl;
+        assert(status == hipSuccess);
+      }
+      m_deviceProperties = new hipDeviceProp_t[num_devices];
+      for (int i = 0; i < num_devices; ++i) {
+        status = hipGetDeviceProperties(&m_deviceProperties[i], i);
+        if (status != hipSuccess) {
+          std::cerr << "Failed to initialize HIP device #"
+                    << i
+                    << ": "
+                    << hipGetErrorString(status)
+                    << std::endl;
+          assert(status == hipSuccess);
+        }
+      }
+
+#if 0 && __cplusplus >= 201103L
+      std::atomic_thread_fence(std::memory_order_release);
+#endif
+      m_devicePropInitialized = true;
+    } else {
+      // Wait for the other thread to inititialize the properties.
+      while (!m_devicePropInitialized) {
+#if 0 && __cplusplus >= 201103L
+        std::atomic_thread_fence(std::memory_order_acquire);
+#endif
+        sleep(1);
+      }
+    }
+  }
+}
+
+static const hipStream_t default_stream = 0x00;//TODO: Use hipStreamDefault instead of 0x00;
+
+class HipStreamDevice : public StreamInterface {
+ public:
+  // Use the default stream on the current device
+  HipStreamDevice() : stream_(&default_stream), scratch_(NULL), semaphore_(NULL) {
+    hipGetDevice(&device_);
+    initializeDeviceProp();
+  }
+  // Use the default stream on the specified device
+  HipStreamDevice(int device) : stream_(&default_stream), device_(device), scratch_(NULL), semaphore_(NULL) {
+    initializeDeviceProp();
+  }
+  // Use the specified stream. Note that it's the
+  // caller responsibility to ensure that the stream can run on
+  // the specified device. If no device is specified the code
+  // assumes that the stream is associated to the current gpu device.
+  HipStreamDevice(const hipStream_t* stream, int device = -1)
+      : stream_(stream), device_(device), scratch_(NULL), semaphore_(NULL) {
+    if (device < 0) {
+      hipGetDevice(&device_);
+    } else {
+      int num_devices;
+      hipError_t err = hipGetDeviceCount(&num_devices);
+      EIGEN_UNUSED_VARIABLE(err)
+      assert(err == hipSuccess);
+      assert(device < num_devices);
+      device_ = device;
+    }
+    initializeDeviceProp();
+  }
+
+  virtual ~HipStreamDevice() {
+    if (scratch_) {
+      deallocate(scratch_);
+    }
+  }
+
+  const hipStream_t& stream() const { return *stream_; }
+  const hipDeviceProp_t& deviceProperties() const {
+    return m_deviceProperties[device_];
+  }
+  virtual void* allocate(size_t num_bytes) const {
+    hipError_t err = hipSetDevice(device_);
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+    void* result;
+    err = hipMalloc(&result, num_bytes);
+    assert(err == hipSuccess);
+    assert(result != NULL);
+    return result;
+  }
+  virtual void deallocate(void* buffer) const {
+    hipError_t err = hipSetDevice(device_);
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+    assert(buffer != NULL);
+    err = hipFree(buffer);
+    assert(err == hipSuccess);
+  }
+
+  virtual void* scratchpad() const {
+    if (scratch_ == NULL) {
+      scratch_ = allocate(kHipScratchSize + sizeof(unsigned int));
+    }
+    return scratch_;
+  }
+
+  virtual unsigned int* semaphore() const {
+    if (semaphore_ == NULL) {
+      char* scratch = static_cast<char*>(scratchpad()) + kHipScratchSize;
+      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
+      //hipError_t err = hipMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
+      hipError_t err = hipMemset(semaphore_, 0, sizeof(unsigned int));
+      EIGEN_UNUSED_VARIABLE(err)
+      assert(err == hipSuccess);
+    }
+    return semaphore_;
+  }
+
+ private:
+  const hipStream_t* stream_;
+  int device_;
+  mutable void* scratch_;
+  mutable unsigned int* semaphore_;
+};
+
+struct GpuDevice {
+  // The StreamInterface is not owned: the caller is
+  // responsible for its initialization and eventual destruction.
+  explicit GpuDevice(const StreamInterface* stream) : stream_(stream), max_blocks_(INT_MAX) {
+    eigen_assert(stream);
+  }
+  explicit GpuDevice(const StreamInterface* stream, int num_blocks) : stream_(stream), max_blocks_(num_blocks) {
+    eigen_assert(stream);
+  }
+  // TODO(bsteiner): This is an internal API, we should not expose it.
+  EIGEN_STRONG_INLINE const hipStream_t& stream() const {
+    return stream_->stream();
+  }
+
+  EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    return stream_->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
+    stream_->deallocate(buffer);
+  }
+
+  EIGEN_STRONG_INLINE void* scratchpad() const {
+    return stream_->scratchpad();
+  }
+
+  EIGEN_STRONG_INLINE unsigned int* semaphore() const {
+    return stream_->semaphore();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
+#if !defined(EIGEN_HIP_DEVICE_COMPILE)
+    hipError_t err = hipMemcpyAsync(dst, src, n, hipMemcpyDeviceToDevice,
+                                      stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+#else
+  eigen_assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
+    hipError_t err =
+        hipMemcpyAsync(dst, src, n, hipMemcpyHostToDevice, stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+  }
+
+   EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
+    hipError_t err =
+        hipMemcpyAsync(dst, src, n, hipMemcpyDeviceToHost, stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
+#if !defined(EIGEN_HIP_DEVICE_COMPILE)
+    //TODO:hipError_t err = hipMemsetAsync(buffer, c, n, stream_->stream());
+    hipError_t err = hipMemset(buffer, c, n);
+    EIGEN_UNUSED_VARIABLE(err)
+    assert(err == hipSuccess);
+#else
+  eigen_assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE size_t numThreads() const {
+    // FIXME
+    return 32;
+  }
+
+  EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
+    // FIXME
+    return 48*1024;
+  }
+
+  EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
+    // We won't try to take advantage of the l2 cache for the time being, and
+    // there is no l3 cache on hip devices.
+    return firstLevelCacheSize();
+  }
+
+// FIXME - this will move into HIP
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+#undef assert
+#define assert(COND)
+#endif
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void synchronize() const {
+#if defined(EIGEN_HIPCC) && \
+    !defined(EIGEN_HIP_DEVICE_COMPILE)
+    hipError_t err = hipStreamSynchronize(stream_->stream());
+    if (err != hipSuccess) {
+      std::cerr << "Error detected in HIP stream: "
+                << hipGetErrorString(err)
+                << std::endl;
+      assert(err == hipSuccess);
+    }
+#else
+    assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE int getNumHipMultiProcessors() const {
+    return stream_->deviceProperties().multiProcessorCount;
+  }
+  EIGEN_STRONG_INLINE int maxHipThreadsPerBlock() const {
+    return stream_->deviceProperties().maxThreadsPerBlock;
+  }
+  EIGEN_STRONG_INLINE int maxHipThreadsPerMultiProcessor() const {
+    return stream_->deviceProperties().maxThreadsPerMultiProcessor;
+  }
+  EIGEN_STRONG_INLINE int sharedMemPerBlock() const {
+    return stream_->deviceProperties().sharedMemPerBlock;
+  }
+   EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+    return stream_->deviceProperties().major;
+  }
+  EIGEN_STRONG_INLINE int minorDeviceVersion() const {
+    return stream_->deviceProperties().minor;
+  }
+
+  EIGEN_STRONG_INLINE int maxBlocks() const {
+    return max_blocks_;
+  }
+
+  // This function checks if the HIP runtime recorded an error for the
+  // underlying stream device.
+  inline bool ok() const {
+#if defined(EIGEN_HIPCC)
+    hipError_t error = hipStreamQuery(stream_->stream());
+    return (error == hipSuccess) || (error == hipErrorNotReady);
+#else
+    return false;
+#endif
+  }
+
+ private:
+  const StreamInterface* stream_;
+  int max_blocks_;
+};
+
+#define LAUNCH_HIP_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
+  hipLaunchKernelGGL(HIP_KERNEL_NAME(kernel), dim3(gridsize), dim3(blocksize), (sharedmem), (device).stream(), (__VA_ARGS__)); \
+  assert(hipGetLastError() == hipSuccess);
+
+
+// FIXME: Should be device and kernel specific.
+#if defined(EIGEN_HIPCC)
+static EIGEN_DEVICE_FUNC inline void setHipSharedMemConfig(hipSharedMemConfig config) {
+#if !defined(EIGEN_HIP_DEVICE_COMPILE)
+  hipError_t status = hipDeviceSetSharedMemConfig(config);
+  EIGEN_UNUSED_VARIABLE(status)
+  assert(status == hipSuccess);
+#else
+  EIGEN_UNUSED_VARIABLE(config)
+#endif
+}
+#endif
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 0ffe68ab3..24a57970a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -201,7 +201,7 @@ class TensorExecutor<Expression, GpuDevice, Vectorizable> {
 };
 
 
-#if defined(EIGEN_CUDACC)
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
 template <typename Evaluator, typename Index, bool Vectorizable>
 struct EigenMetaKernelEval {
   static __device__ EIGEN_ALWAYS_INLINE
@@ -250,6 +250,17 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
   TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
   const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
   if (needs_assign) {
+#if defined(EIGEN_HIPCC)
+    const int block_size = device.maxHipThreadsPerBlock();
+    const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / block_size;
+    const Index size = array_prod(evaluator.dimensions());
+    // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
+    const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
+ 
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),
+        dim3(num_blocks), dim3(block_size), 0, device.stream(), evaluator, size);
+#else
     const int block_size = device.maxCudaThreadsPerBlock();
     const int max_blocks = device.getNumCudaMultiProcessors() *
                            device.maxCudaThreadsPerMultiProcessor() / block_size;
@@ -260,11 +271,12 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
     LAUNCH_CUDA_KERNEL(
         (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),
         num_blocks, block_size, 0, device, evaluator, size);
+#endif
   }
   evaluator.cleanup();
 }
 
-#endif  // EIGEN_CUDACC
+#endif  // EIGEN_CUDACC || EIGEN_HIPCC
 #endif  // EIGEN_USE_GPU
 
 // SYCL Executor policy
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index c015ce196..b8f0bc798 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -109,7 +109,10 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
     const Index numValues =  internal::array_prod(m_impl.dimensions());
     m_buffer = (CoeffReturnType*)m_device.allocate(numValues * sizeof(CoeffReturnType));
     // Should initialize the memory in case we're dealing with non POD types.
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index 3209fecd3..835efbf72 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -350,7 +350,11 @@ struct IndexPairList : internal::IndexTuple<FirstType, OtherTypes...> {
 
 namespace internal {
 
-template<typename FirstType, typename... OtherTypes> size_t array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
+template<typename FirstType, typename... OtherTypes>
+  #if defined(EIGEN_HIPCC)
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
+  #endif
+  size_t array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
   size_t result = 1;
   for (int i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
     result *= sizes[i];
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
index c9e61f359..8e1ba486d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
@@ -27,7 +27,7 @@
  */
 
 // SFINAE requires variadic templates
-#ifndef EIGEN_CUDACC
+#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   // SFINAE doesn't work for gcc <= 4.7
   #ifdef EIGEN_COMP_GNUC
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index 5431eb740..de1075cc1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -52,7 +52,7 @@ struct PacketType : internal::packet_traits<Scalar> {
 };
 
 // For CUDA packet types when using a GpuDevice
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC) && defined(EIGEN_HAS_CUDA_FP16)
+#if defined(EIGEN_USE_GPU) && ((defined(EIGEN_CUDACC) && defined(EIGEN_HAS_CUDA_FP16)) || (defined(EIGEN_HIPCC) && defined(EIGEN_HAS_HIP_FP16)))
 template <>
 struct PacketType<half, GpuDevice> {
   typedef half2 type;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index e59074506..2a979845b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -858,7 +858,10 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
     }
     return inputIndex;
   }
-
+  
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
   static EIGEN_STRONG_INLINE Index clamp(Index value, Index min, Index max) {
 #ifndef __SYCL_DEVICE_ONLY__
     return numext::maxi(min, numext::mini(max,value));
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
index 230915db2..71536a4b9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
@@ -16,7 +16,7 @@ namespace internal {
 namespace {
 
 EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
-#ifdef EIGEN_CUDA_ARCH
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
   // We don't support 3d kernels since we currently only use 1 and
   // 2d kernels.
   assert(threadIdx.z == 0);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index da0ffe728..d2fb3fd32 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -334,12 +334,12 @@ struct OuterReducer {
 };
 
 
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
 template <int B, int N, typename S, typename R, typename I>
 __global__ void FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
 
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
 template <typename S, typename R, typename I>
 __global__ void ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
 template <int B, int N, typename S, typename R, typename I>
@@ -495,7 +495,11 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(typename MakePointer_<CoeffReturnType>::Type data) {
+  EIGEN_STRONG_INLINE
+    #if !defined(EIGEN_HIPCC)
+    EIGEN_DEVICE_FUNC
+    #endif
+    bool evalSubExprsIfNeeded(typename MakePointer_<CoeffReturnType>::Type data) {
     m_impl.evalSubExprsIfNeeded(NULL);
 
     // Use the FullReducer if possible.
@@ -694,9 +698,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 #ifdef EIGEN_USE_THREADS
   template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
 #endif
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
   template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
-#ifdef EIGEN_HAS_CUDA_FP16
+#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
   template <typename S, typename R, typename I> KERNEL_FRIEND void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
   template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
   template <int NPT, typename S, typename R, typename I> KERNEL_FRIEND void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
@@ -774,14 +778,22 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   // Indexed by reduced dimensions.
   array<Index, NumReducedDims> m_reducedDims;
 
+#if defined(EIGEN_HIPCC)
+ public:
+#endif
+  
   // Evaluator for the input expression.
   TensorEvaluator<ArgType, Device> m_impl;
 
+#if defined(EIGEN_HIPCC)
+ private:
+#endif
+  
   // Operation to apply for computing the reduction.
   Op m_reducer;
 
   // For full reductions
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
   static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
   static const bool RunningOnSycl = false;
 #elif defined(EIGEN_USE_SYCL)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
new file mode 100644
index 000000000..5304a22c5
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
@@ -0,0 +1,815 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+#include "Eigen/src/Core/arch/HIP/hcc/math_constants.h"
+#endif
+
+#if defined(EIGEN_HIPCC)
+#define HIP_WARP_SIZE 64
+#endif
+
+namespace Eigen {
+namespace internal {
+
+
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
+// Full reducers for GPU, don't vectorize for now
+
+// Reducer function that enables multiple hip thread to safely accumulate at the same
+// output address. It basically reads the current value of the output variable, and
+// attempts to update it with the new value. If in the meantime another hip thread
+// updated the content of the output address it will try again.
+template <typename T, typename R>
+__device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
+  if (sizeof(T) == 4)
+  {
+    unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
+    unsigned int newval = oldval;
+    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+    unsigned int readback;
+    while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
+      oldval = readback;
+      newval = oldval;
+      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+      if (newval == oldval) {
+        return;
+      }
+    }
+  }
+  else if (sizeof(T) == 8) {
+    unsigned long long oldval = *reinterpret_cast<unsigned long long*>(output);
+    unsigned long long newval = oldval;
+    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+    unsigned long long readback;
+    while ((readback = atomicCAS((unsigned long long*)output, oldval, newval)) != oldval) {
+      oldval = readback;
+      newval = oldval;
+      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+      if (newval == oldval) {
+        return;
+      }
+    }
+  }
+  else {
+    assert(0 && "Wordsize not supported");
+  }
+#else
+  assert(0 && "Shouldn't be called on unsupported device");
+#endif
+}
+
+// We extend atomicExch to support extra data types
+template <typename Type>
+__device__ inline Type atomicExchCustom(Type* address, Type val) {
+  return atomicExch(address, val);
+}
+
+template <>
+__device__ inline double atomicExchCustom(double* address, double val) {
+  unsigned long long int* address_as_ull = reinterpret_cast<unsigned long long int*>(address);
+  return __longlong_as_double(atomicExch(address_as_ull, __double_as_longlong(val)));
+}
+
+#if defined(EIGEN_HAS_HIP_FP16)
+template <template <typename T> class R>
+__device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer) {
+  unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
+  unsigned int newval = oldval;
+  reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
+  if (newval == oldval) {
+    return;
+  }
+  unsigned int readback;
+  while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
+    oldval = readback;
+    newval = oldval;
+    reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+  }
+}
+#endif
+
+template <>
+__device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
+    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
+  atomicAdd(output, accum);
+#else
+  assert(0 && "Shouldn't be called on unsupported device");
+#endif
+}
+
+
+template <typename CoeffType, typename Index>
+__global__ void ReductionInitKernel(const CoeffType val, Index num_preserved_coeffs, CoeffType* output) {
+  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+  const Index num_threads = hipBlockDim_x * hipGridDim_x;
+  for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+    output[i] = val;
+  }
+}
+
+
+template <int BlockSize, int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void FullReductionKernel(const Self input, Index num_coeffs,
+                                    typename Self::CoeffReturnType* output, unsigned int* semaphore, Reducer reducer) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
+    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
+  // Initialize the output value
+  const Index first_index = hipBlockIdx_x * BlockSize * NumPerThread + hipThreadIdx_x;
+  if (hipGridDim_x == 1) {
+    if (first_index == 0) {
+      *output = reducer.initialize();
+    }
+  }
+  else {
+    if (hipThreadIdx_x == 0) {
+      unsigned int block = atomicCAS(semaphore, 0u, 1u);
+      if (block == 0) {
+        // We're the first block to run, initialize the output value
+        atomicExchCustom(output, reducer.initialize());
+        __threadfence();
+        atomicExch(semaphore, 2u);
+      }
+      else {
+        // Wait for the first block to initialize the output value.
+        // Use atomicCAS here to ensure that the reads aren't cached
+        unsigned int val;
+        do {
+          val = atomicCAS(semaphore, 2u, 2u);
+        }
+        while (val < 2u);
+      }
+    }
+  }
+
+  __syncthreads();
+
+  eigen_assert(hipGridDim_x == 1 || *semaphore >= 2u);
+
+  typename Self::CoeffReturnType accum = reducer.initialize();
+  Index max_iter = numext::mini<Index>(num_coeffs - first_index, NumPerThread*BlockSize);
+  for (Index i = 0; i < max_iter; i+=BlockSize) {
+    const Index index = first_index + i;
+    eigen_assert(index < num_coeffs);
+    typename Self::CoeffReturnType val = input.m_impl.coeff(index);
+    reducer.reduce(val, &accum);
+  }
+
+#pragma unroll
+  for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
+    // XXX use std::is_floating_point to determine the type of accum
+    if (std::is_floating_point<typename Self::CoeffReturnType>::value) {
+      reducer.reduce(__shfl_down(static_cast<float>(accum), offset, HIP_WARP_SIZE), &accum);
+    } else {
+      reducer.reduce(__shfl_down(static_cast<int>(accum), offset, HIP_WARP_SIZE), &accum);
+    }
+  }
+
+  if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
+    atomicReduce(output, accum, reducer);
+  }
+
+  if (hipGridDim_x > 1 && hipThreadIdx_x == 0) {
+    // Let the last block reset the semaphore
+    atomicInc(semaphore, hipGridDim_x + 1);
+    __threadfence_system();
+  }
+
+#else
+  assert(0 && "Shouldn't be called on unsupported device");
+#endif
+}
+
+
+#if defined(EIGEN_HAS_HIP_FP16)
+template <typename Self,
+          typename Reducer, typename Index>
+__global__ void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half2* scratch) {
+  eigen_assert(hipBlockDim_x == 1);
+  eigen_assert(hipGridDim_x == 1);
+  if (num_coeffs % 2 != 0) {
+    half last = input.m_impl.coeff(num_coeffs-1);
+    *scratch = __halves2half2(last, reducer.initialize());
+  } else {
+    *scratch = reducer.template initializePacket<half2>();
+  }
+}
+
+template <typename Self,
+          typename Reducer, typename Index>
+__global__ void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output) {
+  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+  const Index num_threads = hipBlockDim_x * hipGridDim_x;
+  const Index num_packets = num_coeffs / 2;
+  for (Index i = thread_id; i < num_packets; i += num_threads) {
+    ((half2*)output)[i] = reducer.template initializePacket<half2>();
+  }
+
+  if (thread_id == 0 && num_coeffs % 2 != 0) {
+    output[num_coeffs-1] = reducer.initialize();
+  }
+}
+
+template <int BlockSize, int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs,
+                                    half* output, half2* scratch) {
+  eigen_assert(NumPerThread % 2 == 0);
+
+  const Index first_index = hipBlockIdx_x * BlockSize * NumPerThread + 2*hipThreadIdx_x;
+
+  // Initialize the output value if it wasn't initialized by the ReductionInitKernel
+  if (hipGridDim_x == 1 && first_index == 0) {
+    if (num_coeffs % 2 != 0) {
+      half last = input.m_impl.coeff(num_coeffs-1);
+      *scratch = __halves2half2(last, reducer.initialize());
+    } else {
+      *scratch = reducer.template initializePacket<half2>();
+    }
+    __syncthreads();
+  }
+
+  half2 accum = reducer.template initializePacket<half2>();
+  const Index max_iter = numext::mini<Index>((num_coeffs - first_index) / 2, NumPerThread*BlockSize / 2);
+  for (Index i = 0; i < max_iter; i += BlockSize) {
+    const Index index = first_index + 2*i;
+    eigen_assert(index + 1 < num_coeffs);
+    half2 val = input.m_impl.template packet<Unaligned>(index);
+    reducer.reducePacket(val, &accum);
+  }
+
+#pragma unroll
+  for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
+    // FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+    union { int i; half2 h; } wka_in, wka_out;
+    wka_in.h = accum;
+    wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
+    reducer.reducePacket(wka_out.h, &accum);
+  }
+
+  if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
+    atomicReduce(scratch, accum, reducer);
+  }
+
+  __syncthreads();
+
+  if (hipGridDim_x == 1 && first_index == 0) {
+    half tmp = __low2half(*scratch);
+    reducer.reduce(__high2half(*scratch), &tmp);
+    *output = tmp;
+  }
+}
+
+template <typename Op>
+__global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2* scratch) {
+  eigen_assert(hipThreadIdx_x == 1);
+  half tmp = __low2half(*scratch);
+  reducer.reduce(__high2half(*scratch), &tmp);
+  *output = tmp;
+}
+
+#endif
+
+template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
+struct FullReductionLauncher {
+  static void run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index) {
+    assert(false && "Should only be called on doubles, floats and half floats");
+  }
+};
+
+namespace {
+  std::mutex __eigen_reduction_hip_mutex;
+}
+
+// Specialization for float and double
+template <typename Self, typename Op, typename OutputType, bool PacketAccess>
+struct FullReductionLauncher<
+    Self, Op, OutputType, PacketAccess,
+    typename internal::enable_if<
+      internal::is_same<float, OutputType>::value ||
+      internal::is_same<double, OutputType>::value,
+    void>::type> {
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) {
+    // guard FullReductionLauncher with a mutex so only 1 FullReductionKernel
+    // is dispatched at a time
+    std::lock_guard<std::mutex> lock(__eigen_reduction_hip_mutex);
+
+    typedef typename Self::Index Index;
+    typedef typename Self::CoeffReturnType Scalar;
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+
+    unsigned int* semaphore = NULL;
+    if (num_blocks > 1) {
+      semaphore = device.semaphore();
+
+      unsigned int semaphore_host = 0xFF;
+      hipMemcpy(&semaphore_host, semaphore, sizeof(unsigned int), hipMemcpyDeviceToHost);
+      if (semaphore_host != 0) {
+        std::cerr << "[WARN][EIGEN][FullReductionLauncher] incorrect semaphore value: "
+                  << semaphore_host << "\n";
+        // wait for all commands on the device to complete so semaphore value
+        // is reset to 0
+        hipDeviceSynchronize();
+
+        // read again
+        hipMemcpy(&semaphore_host, semaphore, sizeof(unsigned int), hipMemcpyDeviceToHost);
+        if (semaphore_host != 0) {
+          std::cerr << "[ERROR][EIGEN][FullReductionLauncher] CRITICAL incorrect semaphore value: "
+                    << semaphore_host << ", apply manual override to 0\n";
+
+          // force set semaphore value to be 0
+          semaphore_host = 0;
+          hipMemcpy(semaphore, &semaphore_host, sizeof(unsigned int), hipMemcpyHostToDevice);
+        }
+      }
+    }
+
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
+                       dim3(num_blocks), dim3(block_size), 0, device.stream(), self, num_coeffs, output, semaphore, reducer);
+  }
+};
+
+#if defined(EIGEN_HAS_HIP_FP16)
+template <typename Self, typename Op>
+struct FullReductionLauncher<Self, Op, Eigen::half, false> {
+  static void run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index) {
+    assert(false && "Should not be called since there is no packet accessor");
+  }
+};
+
+template <typename Self, typename Op>
+struct FullReductionLauncher<Self, Op, Eigen::half, true> {
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs) {
+    typedef typename Self::Index Index;
+
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    half2* scratch = static_cast<half2*>(device.scratchpad());
+
+    if (num_blocks > 1) {
+      // We initialize the output and the scrathpad outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitFullReduxKernelHalfFloat<Self, Op, Index>),
+                         dim3(1), dim3(1), 0, device.stream(), reducer, self, num_coeffs, scratch);
+    }
+
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(FullReductionKernelHalfFloat<block_size, num_per_thread, Self, Op, Index>),
+                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs, output, scratch);
+
+    if (num_blocks > 1) {
+      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionCleanupKernelHalfFloat<Op>),
+                         dim3(1), dim3(1), 0, device.stream(), reducer, output, scratch);
+    }
+  }
+};
+#endif
+
+
+template <typename Self, typename Op, bool Vectorizable>
+struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple cases
+  // of doubles, floats and half floats
+#if defined(EIGEN_HAS_HIP_FP16)
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+       internal::is_same<typename Self::CoeffReturnType, double>::value ||
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
+#else
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+                                                (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                 internal::is_same<typename Self::CoeffReturnType, double>::value);
+#endif
+
+  template <typename OutputType>
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
+    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    // Don't crash when we're called with an input tensor of size 0.
+    if (num_coeffs == 0) {
+      return;
+    }
+
+    FullReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs);
+  }
+};
+
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                         typename Self::CoeffReturnType* output) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
+    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
+  typedef typename Self::CoeffReturnType Type;
+  eigen_assert(hipBlockDim_y == 1);
+  eigen_assert(hipBlockDim_z == 1);
+  eigen_assert(hipGridDim_y == 1);
+  eigen_assert(hipGridDim_z == 1);
+
+  const int unroll_times = 16;
+  eigen_assert(NumPerThread % unroll_times == 0);
+
+  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, hipBlockDim_x * NumPerThread);
+  const Index num_input_blocks = input_col_blocks * num_preserved_coeffs;
+
+  const Index num_threads = hipBlockDim_x * hipGridDim_x;
+  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (hipGridDim_x == 1) {
+    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  for (Index i = hipBlockIdx_x; i < num_input_blocks; i += hipGridDim_x) {
+    const Index row = i / input_col_blocks;
+
+    if (row < num_preserved_coeffs) {
+      const Index col_block = i % input_col_blocks;
+      const Index col_begin = col_block * hipBlockDim_x * NumPerThread + hipThreadIdx_x;
+
+      Type reduced_val = reducer.initialize();
+
+      for (Index j = 0; j < NumPerThread; j += unroll_times) {
+        const Index last_col = col_begin + hipBlockDim_x * (j + unroll_times - 1);
+        if (last_col >= num_coeffs_to_reduce) {
+          for (Index col = col_begin + hipBlockDim_x * j; col < num_coeffs_to_reduce; col += hipBlockDim_x) {
+            const Type val = input.m_impl.coeff(row * num_coeffs_to_reduce + col);
+            reducer.reduce(val, &reduced_val);
+          }
+          break;
+        } else {
+          // Faster version of the loop with no branches after unrolling.
+#pragma unroll
+          for (int k = 0; k < unroll_times; ++k) {
+            const Index col = col_begin + hipBlockDim_x * (j + k);
+            reducer.reduce(input.m_impl.coeff(row * num_coeffs_to_reduce + col), &reduced_val);
+          }
+        }
+      }
+
+#pragma unroll
+      for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
+        // XXX use std::is_floating_point to determine the type of reduced_val
+        if (std::is_floating_point<Type>::value) {
+          reducer.reduce(__shfl_down(static_cast<float>(reduced_val), offset), &reduced_val);
+        } else {
+          reducer.reduce(__shfl_down(static_cast<int>(reduced_val), offset), &reduced_val);
+        }
+      }
+
+      if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
+        atomicReduce(&(output[row]), reduced_val, reducer);
+      }
+    }
+  }
+#else
+  assert(0 && "Shouldn't be called on unsupported device");
+#endif
+}
+
+#if defined(EIGEN_HAS_HIP_FP16)
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                              half* output) {
+  eigen_assert(hipBlockDim_y == 1);
+  eigen_assert(hipBlockDim_z == 1);
+  eigen_assert(hipGridDim_y == 1);
+  eigen_assert(hipGridDim_z == 1);
+
+  const int unroll_times = 16;
+  eigen_assert(NumPerThread % unroll_times == 0);
+  eigen_assert(unroll_times % 2 == 0);
+
+  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, hipBlockDim_x * NumPerThread * 2);
+  const Index num_input_blocks = divup<Index>(input_col_blocks * num_preserved_coeffs, 2);
+
+  const Index num_threads = hipBlockDim_x * hipGridDim_x;
+  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (hipGridDim_x == 1) {
+    Index i = 2*thread_id;
+    for (; i + 1 < num_preserved_coeffs; i += 2*num_threads) {
+      half* loc = output + i;
+      *((half2*)loc) = reducer.template initializePacket<half2>();
+    }
+    if (i < num_preserved_coeffs) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  for (Index i = hipBlockIdx_x; i < num_input_blocks; i += hipGridDim_x) {
+    const Index row = 2 * (i / input_col_blocks);
+
+    if (row + 1 < num_preserved_coeffs) {
+      const Index col_block = i % input_col_blocks;
+      const Index col_begin = 2 * (col_block * hipBlockDim_x * NumPerThread + hipThreadIdx_x);
+
+      half2 reduced_val1 = reducer.template initializePacket<half2>();
+      half2 reduced_val2 = reducer.template initializePacket<half2>();
+
+      for (Index j = 0; j < NumPerThread; j += unroll_times) {
+        const Index last_col = col_begin + hipBlockDim_x * (j + unroll_times - 1) * 2;
+        if (last_col >= num_coeffs_to_reduce) {
+          Index col = col_begin + hipBlockDim_x * j;
+          for (; col + 1 < num_coeffs_to_reduce; col += hipBlockDim_x) {
+            const half2 val1 = input.m_impl.template packet<Unaligned>(row * num_coeffs_to_reduce + col);
+            reducer.reducePacket(val1, &reduced_val1);
+            const half2 val2 = input.m_impl.template packet<Unaligned>((row+1) * num_coeffs_to_reduce + col);
+            reducer.reducePacket(val2, &reduced_val2);
+          }
+          if (col < num_coeffs_to_reduce) {
+            // Peel;
+            const half last1 = input.m_impl.coeff(row * num_coeffs_to_reduce + col);
+            const half2 val1 = __halves2half2(last1, reducer.initialize());
+            reducer.reducePacket(val1, &reduced_val1);
+            const half last2 = input.m_impl.coeff((row+1) * num_coeffs_to_reduce + col);
+            const half2 val2 = __halves2half2(last2, reducer.initialize());
+            reducer.reducePacket(val2, &reduced_val2);
+          }
+          break;
+        } else {
+          // Faster version of the loop with no branches after unrolling.
+#pragma unroll
+          for (int k = 0; k < unroll_times; ++k) {
+            const Index col = col_begin + hipBlockDim_x * (j + k) * 2;
+            reducer.reducePacket(input.m_impl.template packet<Unaligned>(row * num_coeffs_to_reduce + col), &reduced_val1);
+            reducer.reducePacket(input.m_impl.template packet<Unaligned>((row + 1)* num_coeffs_to_reduce + col), &reduced_val2);
+          }
+        }
+      }
+
+#pragma unroll
+      for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
+	// FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+	union { int i; half2 h; } wka_in, wka_out;
+
+	wka_in.h = reduced_val1;
+	wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
+        reducer.reducePacket(wka_out.h, &reduced_val1);
+	
+	wka_in.h = reduced_val2;
+	wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
+        reducer.reducePacket(wka_out.h, &reduced_val2);
+      }
+
+      half val1 =  __low2half(reduced_val1);
+      reducer.reduce(__high2half(reduced_val1), &val1);
+      half val2 =  __low2half(reduced_val2);
+      reducer.reduce(__high2half(reduced_val2), &val2);
+      half2 val = __halves2half2(val1, val2);
+
+      if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
+        half* loc = output + row;
+        atomicReduce((half2*)loc, val, reducer);
+      }
+    }
+  }
+}
+
+#endif
+
+template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
+struct InnerReductionLauncher {
+  static bool run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index, typename Self::Index) {
+    assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device");
+    return true;
+  }
+};
+
+// Specialization for float and double
+template <typename Self, typename Op, typename OutputType, bool PacketAccess>
+struct InnerReductionLauncher<
+  Self, Op, OutputType, PacketAccess,
+  typename internal::enable_if<
+    internal::is_same<float, OutputType>::value ||
+    internal::is_same<double, OutputType>::value,
+  void>::type> {
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
+      const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / 1024;
+      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernel<OutputType, Index>),
+                         dim3(num_blocks), dim3(1024), 0, device.stream(),
+                         reducer.initialize(), num_preserved_vals, output);
+    }
+
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(InnerReductionKernel<num_per_thread, Self, Op, Index>),
+                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self,
+                       num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+
+#if defined(EIGEN_HAS_HIP_FP16)
+template <typename Self, typename Op>
+struct InnerReductionLauncher<Self, Op, Eigen::half, false> {
+  static bool run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index, typename Self::Index) {
+    assert(false && "Should not be called since there is no packet accessor");
+    return true;
+  }
+};
+
+template <typename Self, typename Op>
+struct InnerReductionLauncher<Self, Op, Eigen::half, true> {
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    if (num_preserved_vals % 2 != 0) {
+      // Not supported yet, revert to the slower code path
+      return true;
+    }
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = /*256*/128;
+    const int num_per_thread = /*128*/64;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
+      const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / 1024;
+      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernelHalfFloat<Self, Op, Index>),
+                         dim3(1), dim3(1), 0, device.stream(), reducer, self, num_preserved_vals, output);
+    }
+
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(InnerReductionKernelHalfFloat<num_per_thread, Self, Op, Index>),
+                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+#endif
+
+
+template <typename Self, typename Op>
+struct InnerReducer<Self, Op, GpuDevice> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple case
+  // of floats and half floats.
+#if defined(EIGEN_HAS_HIP_FP16)
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+       internal::is_same<typename Self::CoeffReturnType, double>::value ||
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
+#else
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
+#endif
+
+  template <typename OutputType>
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    // Don't crash when we're called with an input tensor of size 0.
+    if (num_coeffs == 0) {
+      return true;
+    }
+    // It's faster to use the usual code.
+    if (num_coeffs_to_reduce <= 128) {
+      return true;
+    }
+
+    return InnerReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals);
+  }
+};
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                     typename Self::CoeffReturnType* output) {
+  const Index num_threads = hipBlockDim_x * hipGridDim_x;
+  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (hipGridDim_x == 1) {
+    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  // Do the reduction.
+  const Index max_iter = num_preserved_coeffs * divup<Index>(num_coeffs_to_reduce, NumPerThread);
+  for (Index i = thread_id; i < max_iter; i += num_threads) {
+    const Index input_col = i % num_preserved_coeffs;
+    const Index input_row = (i / num_preserved_coeffs) * NumPerThread;
+    typename Self::CoeffReturnType reduced_val = reducer.initialize();
+    const Index max_row = numext::mini(input_row + NumPerThread, num_coeffs_to_reduce);
+    for (Index j = input_row; j < max_row; j++) {
+      typename Self::CoeffReturnType val = input.m_impl.coeff(j * num_preserved_coeffs + input_col);
+      reducer.reduce(val, &reduced_val);
+    }
+    atomicReduce(&(output[input_col]), reduced_val, reducer);
+  }
+}
+
+
+template <typename Self, typename Op>
+struct OuterReducer<Self, Op, GpuDevice> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple case
+  // of floats.
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
+  template <typename Device, typename OutputType>
+  static bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
+    assert(false && "Should only be called to reduce doubles or floats on a gpu device");
+    return true;
+  }
+
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    // It's faster to use the usual code.
+    if (num_coeffs_to_reduce <= 32) {
+      return true;
+    }
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = 256;
+    const int num_per_thread = 16;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumHipMultiProcessors() *
+                           device.maxHipThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs in the reduction kernel itself when we don't have to worry
+      // about race conditions between multiple thread blocks.
+      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
+      const int max_blocks = device.getNumHipMultiProcessors() *
+                             device.maxHipThreadsPerMultiProcessor() / 1024;
+      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernel<float, Index>),
+                         dim3(num_blocks), dim3(1024), 0, device.stream(),
+                         reducer.initialize(), num_preserved_vals, output);
+    }
+
+    hipLaunchKernelGGL(HIP_KERNEL_NAME(OuterReductionKernel<num_per_thread, Self, Op, Index>),
+                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+
+#endif
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 1f545ef1a..174a6a064 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -242,7 +242,7 @@ struct ScanLauncher {
   }
 };
 
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
 
 // GPU implementation of scan
 // TODO(ibab) This placeholder implementation performs multiple scans in
@@ -278,10 +278,15 @@ struct ScanLauncher<Self, Reducer, GpuDevice> {
      Index total_size = internal::array_prod(self.dimensions());
      Index num_blocks = (total_size / self.size() + 63) / 64;
      Index block_size = 64;
+#if defined(EIGEN_HIPCC)     
+     hipLaunchKernelGGL(HIP_KERNEL_NAME(ScanKernel<Self, Reducer>), dim3(num_blocks),
+			dim3(block_size), 0, self.device().stream(), self, total_size, data);
+#else
      LAUNCH_CUDA_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
+#endif     
   }
 };
-#endif  // EIGEN_USE_GPU && EIGEN_CUDACC
+#endif  // EIGEN_USE_GPU && (EIGEN_CUDACC || EIGEN_HIPCC)
 
 }  // end namespace Eigen
 
diff --git a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
index 49d315a66..bb584e3f9 100644
--- a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
+++ b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
@@ -268,6 +268,9 @@ template<
   typename Reducer
 > struct reduce<Reducer>
 {
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
   constexpr static inline int run() { return Reducer::Identity; }
 };
 
@@ -276,6 +279,9 @@ template<
   typename A
 > struct reduce<Reducer, A>
 {
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
   constexpr static inline A run(A a) { return a; }
 };
 
@@ -285,6 +291,9 @@ template<
   typename... Ts
 > struct reduce<Reducer, A, Ts...>
 {
+  #if defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
   constexpr static inline auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
     return Reducer::run(a, reduce<Reducer, Ts...>::run(ts...));
   }
@@ -324,6 +333,9 @@ struct greater_equal_zero_op { template<typename A> constexpr static inline auto
 // together in front... (13.0 doesn't work with array_prod/array_reduce/... anyway, but 13.1
 // does...
 template<typename... Ts>
+#if defined(EIGEN_HIPCC)
+EIGEN_DEVICE_FUNC
+#endif
 constexpr inline decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
 {
   return reduce<product_op, Ts...>::run(ts...);
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 96b3a8261..18b76350b 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -15,7 +15,7 @@
 // The array class is only available starting with cxx11. Emulate our own here
 // if needed. Beware, msvc still doesn't advertise itself as a c++11 compiler!
 // Moreover, CUDA doesn't support the STL containers, so we use our own instead.
-#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_CUDACC) || defined(EIGEN_AVOID_STL_ARRAY)
+#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC) || defined(EIGEN_AVOID_STL_ARRAY)
 
 namespace Eigen {
 template <typename T, size_t n> class array {
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 6c7ac3f3b..5ab67dc1f 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -190,7 +190,7 @@ template <>
 struct lgamma_impl<float> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE float run(float x) {
-#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
+#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     int dummy;
     return ::lgammaf_r(x, &dummy);
 #else
@@ -203,7 +203,7 @@ template <>
 struct lgamma_impl<double> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE double run(double x) {
-#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__)
+#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
     int dummy;
     return ::lgamma_r(x, &dummy);
 #else
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 42b790d55..c3dba5f2a 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -297,3 +297,55 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
 
   unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 endif()
+
+# Add HIP specific tests 
+if (EIGEN_TEST_HIP)
+
+  set(HIP_PATH "/opt/rocm/hip" CACHE STRING "Path to the HIP installation.")
+
+  if (EXISTS ${HIP_PATH})
+    
+    list(APPEND CMAKE_MODULE_PATH ${HIP_PATH}/cmake) 
+
+    find_package(HIP REQUIRED)
+    if (HIP_FOUND)
+
+      execute_process(COMMAND ${HIP_PATH}/bin/hipconfig --platform OUTPUT_VARIABLE HIP_PLATFORM)
+
+      if (${HIP_PLATFORM} STREQUAL "hcc")
+
+	include_directories(${CMAKE_CURRENT_BINARY_DIR})
+	include_directories(${HIP_PATH}/include)
+
+	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
+
+	# ei_add_test(cxx11_tensor_complex_hip)
+	# ei_add_test(cxx11_tensor_complex_cwise_ops_hip)
+	ei_add_test(cxx11_tensor_reduction_hip)
+	ei_add_test(cxx11_tensor_argmax_hip)
+	ei_add_test(cxx11_tensor_cast_float16_hip)
+	ei_add_test(cxx11_tensor_scan_hip)
+	ei_add_test(cxx11_tensor_device_hip)
+	ei_add_test(cxx11_tensor_hip)
+	ei_add_test(cxx11_tensor_contract_hip)
+	ei_add_test(cxx11_tensor_of_float16_hip)
+	ei_add_test(cxx11_tensor_random_hip)
+	
+	unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
+	
+      elseif (${HIP_PLATFORM} STREQUAL "nvcc")
+	message(FATAL_ERROR "HIP_PLATFORM = nvcc is not supported within Eigen")
+      else ()
+	message(FATAL_ERROR "Unknown HIP_PLATFORM = ${HIP_PLATFORM}")
+      endif()
+      
+    endif(HIP_FOUND)
+
+  else ()
+
+    message(FATAL_ERROR "EIGEN_TEST_HIP is ON, but the specified HIP_PATH (${HIP_PATH}) does not exist")
+    
+  endif()
+    
+endif(EIGEN_TEST_HIP)
+
diff --git a/unsupported/test/cxx11_tensor_argmax_hip.cu b/unsupported/test/cxx11_tensor_argmax_hip.cu
new file mode 100644
index 000000000..57d6ca000
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_argmax_hip.cu
@@ -0,0 +1,251 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_FUNC cxx11_tensor_hip
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+template <int Layout>
+void test_hip_simple_argmax()
+{
+  Tensor<double, 3, Layout> in(Eigen::array<DenseIndex, 3>(72,53,97));
+  Tensor<DenseIndex, 1, Layout> out_max(Eigen::array<DenseIndex, 1>(1));
+  Tensor<DenseIndex, 1, Layout> out_min(Eigen::array<DenseIndex, 1>(1));
+  in.setRandom();
+  in *= in.constant(100.0);
+  in(0, 0, 0) = -1000.0;
+  in(71, 52, 96) = 1000.0;
+
+  std::size_t in_bytes = in.size() * sizeof(double);
+  std::size_t out_bytes = out_max.size() * sizeof(DenseIndex);
+
+  double* d_in;
+  DenseIndex* d_out_max;
+  DenseIndex* d_out_min;
+  hipMalloc((void**)(&d_in), in_bytes);
+  hipMalloc((void**)(&d_out_max), out_bytes);
+  hipMalloc((void**)(&d_out_min), out_bytes);
+
+  hipMemcpy(d_in, in.data(), in_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<double, 3, Layout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 3>(72,53,97));
+  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_max(d_out_max, Eigen::array<DenseIndex, 1>(1));
+  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_min(d_out_min, Eigen::array<DenseIndex, 1>(1));
+
+  gpu_out_max.device(gpu_device) = gpu_in.argmax();
+  gpu_out_min.device(gpu_device) = gpu_in.argmin();
+
+  assert(hipMemcpyAsync(out_max.data(), d_out_max, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipMemcpyAsync(out_min.data(), d_out_min, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  VERIFY_IS_EQUAL(out_max(Eigen::array<DenseIndex, 1>(0)), 72*53*97 - 1);
+  VERIFY_IS_EQUAL(out_min(Eigen::array<DenseIndex, 1>(0)), 0);
+
+  hipFree(d_in);
+  hipFree(d_out_max);
+  hipFree(d_out_min);
+}
+
+template <int DataLayout>
+void test_hip_argmax_dim()
+{
+  Tensor<float, 4, DataLayout> tensor(2,3,5,7);
+  std::vector<int> dims;
+  dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
+
+  for (int dim = 0; dim < 4; ++dim) {
+    tensor.setRandom();
+    tensor = (tensor + tensor.constant(0.5)).log();
+
+    array<DenseIndex, 3> out_shape;
+    for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
+
+    Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
+
+    array<DenseIndex, 4> ix;
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 3; ++j) {
+        for (int k = 0; k < 5; ++k) {
+          for (int l = 0; l < 7; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            if (ix[dim] != 0) continue;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
+            tensor(ix) = 10.0;
+          }
+        }
+      }
+    }
+
+    std::size_t in_bytes = tensor.size() * sizeof(float);
+    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
+
+    float* d_in;
+    DenseIndex* d_out;
+    hipMalloc((void**)(&d_in), in_bytes);
+    hipMalloc((void**)(&d_out), out_bytes);
+
+    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
+
+    Eigen::HipStreamDevice stream;
+    Eigen::GpuDevice gpu_device(&stream);
+
+    Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
+    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
+
+    gpu_out.device(gpu_device) = gpu_in.argmax(dim);
+
+    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+    VERIFY_IS_EQUAL(tensor_arg.size(),
+                    size_t(2*3*5*7 / tensor.dimension(dim)));
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the first index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
+    }
+
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 3; ++j) {
+        for (int k = 0; k < 5; ++k) {
+          for (int l = 0; l < 7; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            if (ix[dim] != tensor.dimension(dim) - 1) continue;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
+            tensor(ix) = 20.0;
+          }
+        }
+      }
+    }
+
+    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
+
+    gpu_out.device(gpu_device) = gpu_in.argmax(dim);
+
+    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the last index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
+    }
+
+    hipFree(d_in);
+    hipFree(d_out);
+  }
+}
+
+template <int DataLayout>
+void test_hip_argmin_dim()
+{
+  Tensor<float, 4, DataLayout> tensor(2,3,5,7);
+  std::vector<int> dims;
+  dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
+
+  for (int dim = 0; dim < 4; ++dim) {
+    tensor.setRandom();
+    tensor = (tensor + tensor.constant(0.5)).log();
+
+    array<DenseIndex, 3> out_shape;
+    for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
+
+    Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
+
+    array<DenseIndex, 4> ix;
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 3; ++j) {
+        for (int k = 0; k < 5; ++k) {
+          for (int l = 0; l < 7; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            if (ix[dim] != 0) continue;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
+            tensor(ix) = -10.0;
+          }
+        }
+      }
+    }
+
+    std::size_t in_bytes = tensor.size() * sizeof(float);
+    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
+
+    float* d_in;
+    DenseIndex* d_out;
+    hipMalloc((void**)(&d_in), in_bytes);
+    hipMalloc((void**)(&d_out), out_bytes);
+
+    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
+
+    Eigen::HipStreamDevice stream;
+    Eigen::GpuDevice gpu_device(&stream);
+
+    Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
+    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
+
+    gpu_out.device(gpu_device) = gpu_in.argmin(dim);
+
+    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+    VERIFY_IS_EQUAL(tensor_arg.size(),
+                    2*3*5*7 / tensor.dimension(dim));
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect min to be in the first index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
+    }
+
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 3; ++j) {
+        for (int k = 0; k < 5; ++k) {
+          for (int l = 0; l < 7; ++l) {
+            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
+            if (ix[dim] != tensor.dimension(dim) - 1) continue;
+            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
+            tensor(ix) = -20.0;
+          }
+        }
+      }
+    }
+
+    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
+
+    gpu_out.device(gpu_device) = gpu_in.argmin(dim);
+
+    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
+      // Expect max to be in the last index of the reduced dimension
+      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
+    }
+
+    hipFree(d_in);
+    hipFree(d_out);
+  }
+}
+
+void test_cxx11_tensor_hip()
+{
+  CALL_SUBTEST(test_hip_simple_argmax<RowMajor>());
+  CALL_SUBTEST(test_hip_simple_argmax<ColMajor>());
+  CALL_SUBTEST(test_hip_argmax_dim<RowMajor>());
+  CALL_SUBTEST(test_hip_argmax_dim<ColMajor>());
+  CALL_SUBTEST(test_hip_argmin_dim<RowMajor>());
+  CALL_SUBTEST(test_hip_argmin_dim<ColMajor>());
+}
diff --git a/unsupported/test/cxx11_tensor_cast_float16_hip.cu b/unsupported/test/cxx11_tensor_cast_float16_hip.cu
new file mode 100644
index 000000000..bf6a49df4
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_cast_float16_hip.cu
@@ -0,0 +1,79 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+void test_hip_conversion() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  Tensor<float, 1> floats(num_elem);
+  floats.setRandom();
+
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
+      d_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
+      d_conv, num_elem);
+
+  gpu_device.memcpyHostToDevice(d_float, floats.data(), num_elem*sizeof(float));
+
+  gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
+  gpu_conv.device(gpu_device) = gpu_half.cast<float>();
+
+  Tensor<float, 1> initial(num_elem);
+  Tensor<float, 1> final(num_elem);
+  gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    VERIFY_IS_APPROX(initial(i), final(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_half);
+  gpu_device.deallocate(d_conv);
+}
+
+
+void test_fallback_conversion() {
+  int num_elem = 101;
+  Tensor<float, 1> floats(num_elem);
+  floats.setRandom();
+
+  Eigen::Tensor<Eigen::half, 1> halfs = floats.cast<Eigen::half>();
+  Eigen::Tensor<float, 1> conv = halfs.cast<float>();
+
+  for (int i = 0; i < num_elem; ++i) {
+    VERIFY_IS_APPROX(floats(i), conv(i));
+  }
+}
+
+
+void test_cxx11_tensor_cast_float16_hip()
+{
+  CALL_SUBTEST(test_hip_conversion());
+  CALL_SUBTEST(test_fallback_conversion());
+}
diff --git a/unsupported/test/cxx11_tensor_contract_hip.cu b/unsupported/test/cxx11_tensor_contract_hip.cu
new file mode 100644
index 000000000..652af0ab0
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_contract_hip.cu
@@ -0,0 +1,215 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2014 Navdeep Jaitly <ndjaitly@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+
+using Eigen::Tensor;
+typedef Tensor<float, 1>::DimensionPair DimPair;
+
+template<int DataLayout>
+void test_hip_contraction(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  // with these dimensions, the output has 300 * 140 elements, which is
+  // more than 30 * 1024, which is the number of threads in blocks on
+  // a 15 SM GK110 GPU
+  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
+  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
+  Tensor<float, 2, DataLayout> t_result(m_size, n_size);
+  Tensor<float, 2, DataLayout> t_result_gpu(m_size, n_size);
+  Eigen::array<DimPair, 1> dims(DimPair(1, 0));
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
+  std::size_t t_right_bytes = t_right.size() * sizeof(float);
+  std::size_t t_result_bytes = t_result.size() * sizeof(float);
+
+  float* d_t_left;
+  float* d_t_right;
+  float* d_t_result;
+
+  hipMalloc((void**)(&d_t_left), t_left_bytes);
+  hipMalloc((void**)(&d_t_right), t_right_bytes);
+  hipMalloc((void**)(&d_t_result), t_result_bytes);
+
+  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_left(d_t_left, Eigen::array<int, 2>(m_size, k_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_right(d_t_right, Eigen::array<int, 2>(k_size, n_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_result(d_t_result, Eigen::array<int, 2>(m_size, n_size));
+
+
+  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
+  t_result = t_left.contract(t_right, dims);
+
+  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
+  for (DenseIndex i = 0; i < t_result.size(); i++) {
+    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
+      continue;
+    }
+    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
+      continue;
+    }
+    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
+              << " vs " <<  t_result_gpu(i) << std::endl;
+    assert(false);
+  }
+
+  hipFree((void*)d_t_left);
+  hipFree((void*)d_t_right);
+  hipFree((void*)d_t_result);
+}
+
+
+template<int DataLayout>
+void test_scalar(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  // with these dimensions, the output has 300 * 140 elements, which is
+  // more than 30 * 1024, which is the number of threads in blocks on
+  // a 15 SM GK110 GPU
+  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
+  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
+  Tensor<float, 0, DataLayout> t_result;
+  Tensor<float, 0, DataLayout> t_result_gpu;
+  Eigen::array<DimPair, 2> dims(DimPair(0, 0), DimPair(1, 1));
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
+  std::size_t t_right_bytes = t_right.size() * sizeof(float);
+  std::size_t t_result_bytes = sizeof(float);
+
+  float* d_t_left;
+  float* d_t_right;
+  float* d_t_result;
+
+  hipMalloc((void**)(&d_t_left), t_left_bytes);
+  hipMalloc((void**)(&d_t_right), t_right_bytes);
+  hipMalloc((void**)(&d_t_result), t_result_bytes);
+
+  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_left(d_t_left, m_size, k_size);
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_right(d_t_right, k_size, n_size);
+  Eigen::TensorMap<Eigen::Tensor<float, 0, DataLayout> >
+      gpu_t_result(d_t_result);
+
+  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
+  t_result = t_left.contract(t_right, dims);
+
+  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
+  if (fabs(t_result() - t_result_gpu()) > 1e-4f &&
+      !Eigen::internal::isApprox(t_result(), t_result_gpu(), 1e-4f)) {
+    std::cout << "mismatch detected: " << t_result()
+              << " vs " <<  t_result_gpu() << std::endl;
+    assert(false);
+  }
+
+  hipFree((void*)d_t_left);
+  hipFree((void*)d_t_right);
+  hipFree((void*)d_t_result);
+}
+
+
+template<int DataLayout>
+void test_hip_contraction_m() {
+  for (int k = 32; k < 256; k++) {
+    test_hip_contraction<ColMajor>(k, 128, 128);
+    test_hip_contraction<RowMajor>(k, 128, 128);
+  }
+}
+
+template<int DataLayout>
+void test_hip_contraction_k() {
+  for (int k = 32; k < 256; k++) {
+    test_hip_contraction<ColMajor>(128, k, 128);
+    test_hip_contraction<RowMajor>(128, k, 128);
+  }
+}
+
+template<int DataLayout>
+void test_hip_contraction_n() {
+  for (int k = 32; k < 256; k++) {
+    test_hip_contraction<ColMajor>(128, 128, k);
+    test_hip_contraction<RowMajor>(128, 128, k);
+  }
+}
+
+
+template<int DataLayout>
+void test_hip_contraction_sizes() {
+  int m_sizes[] = { 31,  39,   63,   64,   65,
+                   127, 129,  255,  257 , 511,
+                   512, 513, 1023, 1024, 1025};
+
+  int n_sizes[] = { 31,  39,   63,   64,   65,
+                   127, 129,  255,  257,  511,
+                   512, 513, 1023, 1024, 1025};
+
+  int k_sizes[] = {  31,   39,  63,  64,   65,
+                     95,   96, 127, 129,  255,
+                    257,  511, 512, 513, 1023,
+                   1024, 1025};
+
+  for (int i = 0; i < 15; i++) {
+    for (int j = 0; j < 15; j++) {
+      for (int k = 0; k < 17; k++) {
+        test_hip_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
+      }
+    }
+  }
+}
+
+void test_cxx11_tensor_hip()
+{
+  CALL_SUBTEST(test_hip_contraction<ColMajor>(128, 128, 128));
+  CALL_SUBTEST(test_hip_contraction<RowMajor>(128, 128, 128));
+
+  CALL_SUBTEST(test_scalar<ColMajor>(128, 128, 128));
+  CALL_SUBTEST(test_scalar<RowMajor>(128, 128, 128));
+
+  CALL_SUBTEST(test_hip_contraction_m<ColMajor>());
+  CALL_SUBTEST(test_hip_contraction_m<RowMajor>());
+
+  CALL_SUBTEST(test_hip_contraction_k<ColMajor>());
+  CALL_SUBTEST(test_hip_contraction_k<RowMajor>());
+
+  CALL_SUBTEST(test_hip_contraction_n<ColMajor>());
+  CALL_SUBTEST(test_hip_contraction_n<RowMajor>());
+
+  // Commenting out these tests due to long runtimes
+  // CALL_SUBTEST(test_hip_contraction_sizes<ColMajor>());
+  // CALL_SUBTEST(test_hip_contraction_sizes<RowMajor>());
+}
diff --git a/unsupported/test/cxx11_tensor_device_hip.cu b/unsupported/test/cxx11_tensor_device_hip.cu
new file mode 100644
index 000000000..b98c481ff
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_device_hip.cu
@@ -0,0 +1,389 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_device
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+using Eigen::RowMajor;
+
+// Context for evaluation on cpu
+struct CPUContext {
+  CPUContext(const Eigen::Tensor<float, 3>& in1, Eigen::Tensor<float, 3>& in2, Eigen::Tensor<float, 3>& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(2,2), kernel_3d_(2,2,2) {
+    kernel_1d_(0) = 3.14f;
+    kernel_1d_(1) = 2.7f;
+
+    kernel_2d_(0,0) = 3.14f;
+    kernel_2d_(1,0) = 2.7f;
+    kernel_2d_(0,1) = 0.2f;
+    kernel_2d_(1,1) = 7.0f;
+
+    kernel_3d_(0,0,0) = 3.14f;
+    kernel_3d_(0,1,0) = 2.7f;
+    kernel_3d_(0,0,1) = 0.2f;
+    kernel_3d_(0,1,1) = 7.0f;
+    kernel_3d_(1,0,0) = -1.0f;
+    kernel_3d_(1,1,0) = -0.3f;
+    kernel_3d_(1,0,1) = -0.7f;
+    kernel_3d_(1,1,1) = -0.5f;
+  }
+
+  const Eigen::DefaultDevice& device() const { return cpu_device_; }
+
+  const Eigen::Tensor<float, 3>& in1() const { return in1_; }
+  const Eigen::Tensor<float, 3>& in2() const { return in2_; }
+  Eigen::Tensor<float, 3>& out() { return out_; }
+  const Eigen::Tensor<float, 1>& kernel1d() const { return kernel_1d_; }
+  const Eigen::Tensor<float, 2>& kernel2d() const { return kernel_2d_; }
+  const Eigen::Tensor<float, 3>& kernel3d() const { return kernel_3d_; }
+
+ private:
+  const Eigen::Tensor<float, 3>& in1_;
+  const Eigen::Tensor<float, 3>& in2_;
+  Eigen::Tensor<float, 3>& out_;
+
+  Eigen::Tensor<float, 1> kernel_1d_;
+  Eigen::Tensor<float, 2> kernel_2d_;
+  Eigen::Tensor<float, 3> kernel_3d_;
+
+  Eigen::DefaultDevice cpu_device_;
+};
+
+
+// Context for evaluation on GPU
+struct GPUContext {
+  GPUContext(const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1, Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2, Eigen::TensorMap<Eigen::Tensor<float, 3> >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) {
+    assert(hipMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == hipSuccess);
+    float kernel_1d_val[] = {3.14f, 2.7f};
+    assert(hipMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
+
+    assert(hipMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == hipSuccess);
+    float kernel_2d_val[] = {3.14f, 2.7f, 0.2f, 7.0f};
+    assert(hipMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
+
+    assert(hipMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == hipSuccess);
+    float kernel_3d_val[] = {3.14f, -1.0f, 2.7f, -0.3f, 0.2f, -0.7f, 7.0f, -0.5f};
+    assert(hipMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
+  }
+  ~GPUContext() {
+    assert(hipFree(kernel_1d_) == hipSuccess);
+    assert(hipFree(kernel_2d_) == hipSuccess);
+    assert(hipFree(kernel_3d_) == hipSuccess);
+  }
+
+  const Eigen::GpuDevice& device() const { return gpu_device_; }
+
+  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1() const { return in1_; }
+  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2() const { return in2_; }
+  Eigen::TensorMap<Eigen::Tensor<float, 3> >& out() { return out_; }
+  Eigen::TensorMap<Eigen::Tensor<float, 1> > kernel1d() const { return Eigen::TensorMap<Eigen::Tensor<float, 1> >(kernel_1d_, 2); }
+  Eigen::TensorMap<Eigen::Tensor<float, 2> > kernel2d() const { return Eigen::TensorMap<Eigen::Tensor<float, 2> >(kernel_2d_, 2, 2); }
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > kernel3d() const { return Eigen::TensorMap<Eigen::Tensor<float, 3> >(kernel_3d_, 2, 2, 2); }
+
+ private:
+  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1_;
+  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2_;
+  Eigen::TensorMap<Eigen::Tensor<float, 3> >& out_;
+
+  float* kernel_1d_;
+  float* kernel_2d_;
+  float* kernel_3d_;
+
+  Eigen::HipStreamDevice stream_;
+  Eigen::GpuDevice gpu_device_;
+};
+
+
+// The actual expression to evaluate
+template <typename Context>
+void test_contextual_eval(Context* context)
+{
+  context->out().device(context->device()) = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f);
+}
+
+template <typename Context>
+void test_forced_contextual_eval(Context* context)
+{
+  context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f);
+}
+
+template <typename Context>
+void test_compound_assignment(Context* context)
+{
+  context->out().device(context->device()) = context->in1().constant(2.718f);
+  context->out().device(context->device()) += context->in1() + context->in2() * 3.14f;
+}
+
+
+template <typename Context>
+void test_contraction(Context* context)
+{
+  Eigen::array<std::pair<int, int>, 2> dims;
+  dims[0] = std::make_pair(1, 1);
+  dims[1] = std::make_pair(2, 2);
+
+  Eigen::array<int, 2> shape(40, 50*70);
+
+  Eigen::DSizes<int, 2> indices(0,0);
+  Eigen::DSizes<int, 2> sizes(40,40);
+
+  context->out().reshape(shape).slice(indices, sizes).device(context->device()) = context->in1().contract(context->in2(), dims);
+}
+
+
+template <typename Context>
+void test_1d_convolution(Context* context)
+{
+  Eigen::DSizes<int, 3> indices(0,0,0);
+  Eigen::DSizes<int, 3> sizes(40,49,70);
+
+  Eigen::array<int, 1> dims(1);
+  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel1d(), dims);
+}
+
+template <typename Context>
+void test_2d_convolution(Context* context)
+{
+  Eigen::DSizes<int, 3> indices(0,0,0);
+  Eigen::DSizes<int, 3> sizes(40,49,69);
+
+  Eigen::array<int, 2> dims(1,2);
+  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel2d(), dims);
+}
+
+template <typename Context>
+void test_3d_convolution(Context* context)
+{
+  Eigen::DSizes<int, 3> indices(0,0,0);
+  Eigen::DSizes<int, 3> sizes(39,49,69);
+
+  Eigen::array<int, 3> dims(0,1,2);
+  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel3d(), dims);
+}
+
+
+void test_cpu() {
+  Eigen::Tensor<float, 3> in1(40,50,70);
+  Eigen::Tensor<float, 3> in2(40,50,70);
+  Eigen::Tensor<float, 3> out(40,50,70);
+
+  in1 = in1.random() + in1.constant(10.0f);
+  in2 = in2.random() + in2.constant(10.0f);
+
+  CPUContext context(in1, in2, out);
+  test_contextual_eval(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_forced_contextual_eval(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_compound_assignment(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_contraction(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 40; ++j) {
+      const float result = out(i,j,0);
+      float expected = 0;
+      for (int k = 0; k < 50; ++k) {
+        for (int l = 0; l < 70; ++l) {
+          expected += in1(i, k, l) * in2(j, k, l);
+        }
+      }
+      VERIFY_IS_APPROX(expected, result);
+    }
+  }
+
+  test_1d_convolution(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
+      }
+    }
+  }
+
+  test_2d_convolution(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 69; ++k) {
+        const float result = out(i,j,k);
+        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f) +
+                               (in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
+        if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
+          continue;
+        }
+        VERIFY_IS_APPROX(expected, result);
+      }
+    }
+  }
+
+  test_3d_convolution(&context);
+  for (int i = 0; i < 39; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 69; ++k) {
+        const float result = out(i,j,k);
+        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
+                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f) +
+                               (in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
+                                in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
+        if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
+          continue;
+        }
+        VERIFY_IS_APPROX(expected, result);
+      }
+    }
+  }
+}
+
+void test_gpu() {
+  Eigen::Tensor<float, 3> in1(40,50,70);
+  Eigen::Tensor<float, 3> in2(40,50,70);
+  Eigen::Tensor<float, 3> out(40,50,70);
+  in1 = in1.random() + in1.constant(10.0f);
+  in2 = in2.random() + in2.constant(10.0f);
+
+  std::size_t in1_bytes = in1.size() * sizeof(float);
+  std::size_t in2_bytes = in2.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_in1;
+  float* d_in2;
+  float* d_out;
+  hipMalloc((void**)(&d_in1), in1_bytes);
+  hipMalloc((void**)(&d_in2), in2_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, 40,50,70);
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, 40,50,70);
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, 40,50,70);
+
+  GPUContext context(gpu_in1, gpu_in2, gpu_out);
+  test_contextual_eval(&context);
+  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_forced_contextual_eval(&context);
+  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_compound_assignment(&context);
+  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
+      }
+    }
+  }
+
+  test_contraction(&context);
+  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 40; ++j) {
+      const float result = out(i,j,0);
+      float expected = 0;
+      for (int k = 0; k < 50; ++k) {
+        for (int l = 0; l < 70; ++l) {
+          expected += in1(i, k, l) * in2(j, k, l);
+        }
+      }
+      VERIFY_IS_APPROX(expected, result);
+    }
+  }
+
+  test_1d_convolution(&context);
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
+  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
+      }
+    }
+  }
+
+  test_2d_convolution(&context);
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
+  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 69; ++k) {
+        const float result = out(i,j,k);
+        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
+                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
+        VERIFY_IS_APPROX(expected, result);
+      }
+    }
+  }
+
+  /*
+  test_3d_convolution(&context);
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
+  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
+  for (int i = 0; i < 39; ++i) {
+    for (int j = 0; j < 49; ++j) {
+      for (int k = 0; k < 69; ++k) {
+       const float result = out(i,j,k);
+        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
+                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f +
+                                in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
+                                in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
+        VERIFY_IS_APPROX(expected, result);
+      }
+    }
+  }
+  */
+}
+
+
+void test_cxx11_tensor_device()
+{
+  CALL_SUBTEST(test_cpu());
+  CALL_SUBTEST(test_gpu());
+}
diff --git a/unsupported/test/cxx11_tensor_hip.cu b/unsupported/test/cxx11_tensor_hip.cu
new file mode 100644
index 000000000..f1794aa87
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_hip.cu
@@ -0,0 +1,1296 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_hip
+#define EIGEN_USE_GPU
+
+#ifdef __NVCC__
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+#include <cuda_fp16.h>
+#endif
+#endif
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+void test_hip_nullary() {
+  Tensor<float, 1, 0, int> in1(2);
+  Tensor<float, 1, 0, int> in2(2);
+  in1.setRandom();
+  in2.setRandom();
+
+  std::size_t tensor_bytes = in1.size() * sizeof(float);
+
+  float* d_in1;
+  float* d_in2;
+  hipMalloc((void**)(&d_in1), tensor_bytes);
+  hipMalloc((void**)(&d_in2), tensor_bytes);
+  hipMemcpy(d_in1, in1.data(), tensor_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in2, in2.data(), tensor_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in1(
+      d_in1, 2);
+  Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in2(
+      d_in2, 2);
+
+  gpu_in1.device(gpu_device) = gpu_in1.constant(3.14f);
+  gpu_in2.device(gpu_device) = gpu_in2.random();
+
+  Tensor<float, 1, 0, int> new1(2);
+  Tensor<float, 1, 0, int> new2(2);
+
+  assert(hipMemcpyAsync(new1.data(), d_in1, tensor_bytes, hipMemcpyDeviceToHost,
+                         gpu_device.stream()) == hipSuccess);
+  assert(hipMemcpyAsync(new2.data(), d_in2, tensor_bytes, hipMemcpyDeviceToHost,
+                         gpu_device.stream()) == hipSuccess);
+
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 2; ++i) {
+    VERIFY_IS_APPROX(new1(i), 3.14f);
+    VERIFY_IS_NOT_EQUAL(new2(i), in2(i));
+  }
+
+  hipFree(d_in1);
+  hipFree(d_in2);
+}
+
+void test_hip_elementwise_small() {
+  Tensor<float, 1> in1(Eigen::array<Eigen::DenseIndex, 1>{2});
+  Tensor<float, 1> in2(Eigen::array<Eigen::DenseIndex, 1>{2});
+  Tensor<float, 1> out(Eigen::array<Eigen::DenseIndex, 1>{2});
+  in1.setRandom();
+  in2.setRandom();
+
+  std::size_t in1_bytes = in1.size() * sizeof(float);
+  std::size_t in2_bytes = in2.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_in1;
+  float* d_in2;
+  float* d_out;
+  hipMalloc((void**)(&d_in1), in1_bytes);
+  hipMalloc((void**)(&d_in2), in2_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
+      d_in1, Eigen::array<Eigen::DenseIndex, 1>{2});
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in2(
+      d_in2, Eigen::array<Eigen::DenseIndex, 1>{2});
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_out(
+      d_out, Eigen::array<Eigen::DenseIndex, 1>{2});
+
+  gpu_out.device(gpu_device) = gpu_in1 + gpu_in2;
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost,
+                         gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 2; ++i) {
+    VERIFY_IS_APPROX(
+        out(Eigen::array<Eigen::DenseIndex, 1>{i}),
+        in1(Eigen::array<Eigen::DenseIndex, 1>{i}) + in2(Eigen::array<Eigen::DenseIndex, 1>{i}));
+  }
+
+  hipFree(d_in1);
+  hipFree(d_in2);
+  hipFree(d_out);
+}
+
+void test_hip_elementwise()
+{
+  Tensor<float, 3> in1(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Tensor<float, 3> in2(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Tensor<float, 3> in3(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Tensor<float, 3> out(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  in1.setRandom();
+  in2.setRandom();
+  in3.setRandom();
+
+  std::size_t in1_bytes = in1.size() * sizeof(float);
+  std::size_t in2_bytes = in2.size() * sizeof(float);
+  std::size_t in3_bytes = in3.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_in1;
+  float* d_in2;
+  float* d_in3;
+  float* d_out;
+  hipMalloc((void**)(&d_in1), in1_bytes);
+  hipMalloc((void**)(&d_in2), in2_bytes);
+  hipMalloc((void**)(&d_in3), in3_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in3, in3.data(), in3_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in3(d_in3, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
+
+  gpu_out.device(gpu_device) = gpu_in1 + gpu_in2 * gpu_in3;
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 72; ++i) {
+    for (int j = 0; j < 53; ++j) {
+      for (int k = 0; k < 97; ++k) {
+        VERIFY_IS_APPROX(out(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}), in1(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}) + in2(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}) * in3(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}));
+      }
+    }
+  }
+
+  hipFree(d_in1);
+  hipFree(d_in2);
+  hipFree(d_in3);
+  hipFree(d_out);
+}
+
+void test_hip_props() {
+  Tensor<float, 1> in1(200);
+  Tensor<bool, 1> out(200);
+  in1.setRandom();
+
+  std::size_t in1_bytes = in1.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(bool);
+
+  float* d_in1;
+  bool* d_out;
+  hipMalloc((void**)(&d_in1), in1_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
+      d_in1, 200);
+  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_out(
+      d_out, 200);
+
+  gpu_out.device(gpu_device) = (gpu_in1.isnan)();
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost,
+                         gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 200; ++i) {
+    VERIFY_IS_EQUAL(out(i), (std::isnan)(in1(i)));
+  }
+
+  hipFree(d_in1);
+  hipFree(d_out);
+}
+
+void test_hip_reduction()
+{
+  Tensor<float, 4> in1(72,53,97,113);
+  Tensor<float, 2> out(72,97);
+  in1.setRandom();
+
+  std::size_t in1_bytes = in1.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_in1;
+  float* d_out;
+  hipMalloc((void**)(&d_in1), in1_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4> > gpu_in1(d_in1, 72,53,97,113);
+  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
+
+  array<Eigen::DenseIndex, 2> reduction_axis;
+  reduction_axis[0] = 1;
+  reduction_axis[1] = 3;
+
+  gpu_out.device(gpu_device) = gpu_in1.maximum(reduction_axis);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 72; ++i) {
+    for (int j = 0; j < 97; ++j) {
+      float expected = 0;
+      for (int k = 0; k < 53; ++k) {
+        for (int l = 0; l < 113; ++l) {
+          expected =
+              std::max<float>(expected, in1(i, k, j, l));
+        }
+      }
+      VERIFY_IS_APPROX(out(i,j), expected);
+    }
+  }
+
+  hipFree(d_in1);
+  hipFree(d_out);
+}
+
+template<int DataLayout>
+void test_hip_contraction()
+{
+  // with these dimensions, the output has 300 * 140 elements, which is
+  // more than 30 * 1024, which is the number of threads in blocks on
+  // a 15 SM GK110 GPU
+  Tensor<float, 4, DataLayout> t_left(6, 50, 3, 31);
+  Tensor<float, 5, DataLayout> t_right(Eigen::array<Eigen::DenseIndex, 5>{3, 31, 7, 20, 1});
+  Tensor<float, 5, DataLayout> t_result(Eigen::array<Eigen::DenseIndex, 5>{6, 50, 7, 20, 1});
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
+  std::size_t t_right_bytes = t_right.size() * sizeof(float);
+  std::size_t t_result_bytes = t_result.size() * sizeof(float);
+
+  float* d_t_left;
+  float* d_t_right;
+  float* d_t_result;
+
+  hipMalloc((void**)(&d_t_left), t_left_bytes);
+  hipMalloc((void**)(&d_t_right), t_right_bytes);
+  hipMalloc((void**)(&d_t_result), t_result_bytes);
+
+  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_t_left(d_t_left, 6, 50, 3, 31);
+  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_t_right(d_t_right, 3, 31, 7, 20, 1);
+  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_t_result(d_t_result, 6, 50, 7, 20, 1);
+
+  typedef Eigen::Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> > MapXf;
+  MapXf m_left(t_left.data(), 300, 93);
+  MapXf m_right(t_right.data(), 93, 140);
+  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(300, 140);
+
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+  Eigen::array<DimPair, 2> dims;
+  dims[0] = DimPair(2, 0);
+  dims[1] = DimPair(3, 1);
+
+  m_result = m_left * m_right;
+  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
+
+  hipMemcpy(t_result.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
+
+  for (DenseIndex i = 0; i < t_result.size(); i++) {
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4f) {
+      std::cout << "mismatch detected at index " << i << ": " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+
+  hipFree(d_t_left);
+  hipFree(d_t_right);
+  hipFree(d_t_result);
+}
+
+
+template<int DataLayout>
+void test_hip_convolution_1d()
+{
+  Tensor<float, 4, DataLayout> input(74,37,11,137);
+  Tensor<float, 1, DataLayout> kernel(4);
+  Tensor<float, 4, DataLayout> out(74,34,11,137);
+  input = input.constant(10.0f) + input.random();
+  kernel = kernel.constant(7.0f) + kernel.random();
+
+  std::size_t input_bytes = input.size() * sizeof(float);
+  std::size_t kernel_bytes = kernel.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_input;
+  float* d_kernel;
+  float* d_out;
+  hipMalloc((void**)(&d_input), input_bytes);
+  hipMalloc((void**)(&d_kernel), kernel_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input, 74,37,11,137);
+  Eigen::TensorMap<Eigen::Tensor<float, 1, DataLayout> > gpu_kernel(d_kernel, 4);
+  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out, 74,34,11,137);
+
+  Eigen::array<Eigen::DenseIndex, 1> dims{1};
+  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 74; ++i) {
+    for (int j = 0; j < 34; ++j) {
+      for (int k = 0; k < 11; ++k) {
+        for (int l = 0; l < 137; ++l) {
+          const float result = out(i,j,k,l);
+          const float expected = input(i,j+0,k,l) * kernel(0) + input(i,j+1,k,l) * kernel(1) +
+                                 input(i,j+2,k,l) * kernel(2) + input(i,j+3,k,l) * kernel(3);
+          VERIFY_IS_APPROX(result, expected);
+        }
+      }
+    }
+  }
+
+  hipFree(d_input);
+  hipFree(d_kernel);
+  hipFree(d_out);
+}
+
+void test_hip_convolution_inner_dim_col_major_1d()
+{
+  Tensor<float, 4, ColMajor> input(74,9,11,7);
+  Tensor<float, 1, ColMajor> kernel(4);
+  Tensor<float, 4, ColMajor> out(71,9,11,7);
+  input = input.constant(10.0f) + input.random();
+  kernel = kernel.constant(7.0f) + kernel.random();
+
+  std::size_t input_bytes = input.size() * sizeof(float);
+  std::size_t kernel_bytes = kernel.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_input;
+  float* d_kernel;
+  float* d_out;
+  hipMalloc((void**)(&d_input), input_bytes);
+  hipMalloc((void**)(&d_kernel), kernel_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_input(d_input,74,9,11,7);
+  Eigen::TensorMap<Eigen::Tensor<float, 1, ColMajor> > gpu_kernel(d_kernel,4);
+  Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_out(d_out,71,9,11,7);
+
+  Eigen::array<Eigen::DenseIndex, 1> dims{0};
+  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 71; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 11; ++k) {
+        for (int l = 0; l < 7; ++l) {
+          const float result = out(i,j,k,l);
+          const float expected = input(i+0,j,k,l) * kernel(0) + input(i+1,j,k,l) * kernel(1) +
+                                 input(i+2,j,k,l) * kernel(2) + input(i+3,j,k,l) * kernel(3);
+          VERIFY_IS_APPROX(result, expected);
+        }
+      }
+    }
+  }
+
+  hipFree(d_input);
+  hipFree(d_kernel);
+  hipFree(d_out);
+}
+
+void test_hip_convolution_inner_dim_row_major_1d()
+{
+  Tensor<float, 4, RowMajor> input(7,9,11,74);
+  Tensor<float, 1, RowMajor> kernel(4);
+  Tensor<float, 4, RowMajor> out(7,9,11,71);
+  input = input.constant(10.0f) + input.random();
+  kernel = kernel.constant(7.0f) + kernel.random();
+
+  std::size_t input_bytes = input.size() * sizeof(float);
+  std::size_t kernel_bytes = kernel.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_input;
+  float* d_kernel;
+  float* d_out;
+  hipMalloc((void**)(&d_input), input_bytes);
+  hipMalloc((void**)(&d_kernel), kernel_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_input(d_input, 7,9,11,74);
+  Eigen::TensorMap<Eigen::Tensor<float, 1, RowMajor> > gpu_kernel(d_kernel, 4);
+  Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_out(d_out, 7,9,11,71);
+
+  Eigen::array<Eigen::DenseIndex, 1> dims{3};
+  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 7; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 11; ++k) {
+        for (int l = 0; l < 71; ++l) {
+          const float result = out(i,j,k,l);
+          const float expected = input(i,j,k,l+0) * kernel(0) + input(i,j,k,l+1) * kernel(1) +
+                                 input(i,j,k,l+2) * kernel(2) + input(i,j,k,l+3) * kernel(3);
+          VERIFY_IS_APPROX(result, expected);
+        }
+      }
+    }
+  }
+
+  hipFree(d_input);
+  hipFree(d_kernel);
+  hipFree(d_out);
+}
+
+template<int DataLayout>
+void test_hip_convolution_2d()
+{
+  Tensor<float, 4, DataLayout> input(74,37,11,137);
+  Tensor<float, 2, DataLayout> kernel(3,4);
+  Tensor<float, 4, DataLayout> out(74,35,8,137);
+  input = input.constant(10.0f) + input.random();
+  kernel = kernel.constant(7.0f) + kernel.random();
+
+  std::size_t input_bytes = input.size() * sizeof(float);
+  std::size_t kernel_bytes = kernel.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_input;
+  float* d_kernel;
+  float* d_out;
+  hipMalloc((void**)(&d_input), input_bytes);
+  hipMalloc((void**)(&d_kernel), kernel_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input,74,37,11,137);
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> > gpu_kernel(d_kernel,3,4);
+  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out,74,35,8,137);
+
+  Eigen::array<Eigen::DenseIndex, 2> dims{1,2};
+  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 74; ++i) {
+    for (int j = 0; j < 35; ++j) {
+      for (int k = 0; k < 8; ++k) {
+        for (int l = 0; l < 137; ++l) {
+          const float result = out(i,j,k,l);
+          const float expected = input(i,j+0,k+0,l) * kernel(0,0) +
+                                 input(i,j+1,k+0,l) * kernel(1,0) +
+                                 input(i,j+2,k+0,l) * kernel(2,0) +
+                                 input(i,j+0,k+1,l) * kernel(0,1) +
+                                 input(i,j+1,k+1,l) * kernel(1,1) +
+                                 input(i,j+2,k+1,l) * kernel(2,1) +
+                                 input(i,j+0,k+2,l) * kernel(0,2) +
+                                 input(i,j+1,k+2,l) * kernel(1,2) +
+                                 input(i,j+2,k+2,l) * kernel(2,2) +
+                                 input(i,j+0,k+3,l) * kernel(0,3) +
+                                 input(i,j+1,k+3,l) * kernel(1,3) +
+                                 input(i,j+2,k+3,l) * kernel(2,3);
+          VERIFY_IS_APPROX(result, expected);
+        }
+      }
+    }
+  }
+
+  hipFree(d_input);
+  hipFree(d_kernel);
+  hipFree(d_out);
+}
+
+template<int DataLayout>
+void test_hip_convolution_3d()
+{
+  Tensor<float, 5, DataLayout> input(Eigen::array<Eigen::DenseIndex, 5>{74,37,11,137,17});
+  Tensor<float, 3, DataLayout> kernel(3,4,2);
+  Tensor<float, 5, DataLayout> out(Eigen::array<Eigen::DenseIndex, 5>{74,35,8,136,17});
+  input = input.constant(10.0f) + input.random();
+  kernel = kernel.constant(7.0f) + kernel.random();
+
+  std::size_t input_bytes = input.size() * sizeof(float);
+  std::size_t kernel_bytes = kernel.size() * sizeof(float);
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_input;
+  float* d_kernel;
+  float* d_out;
+  hipMalloc((void**)(&d_input), input_bytes);
+  hipMalloc((void**)(&d_kernel), kernel_bytes);
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;    
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_input(d_input,74,37,11,137,17);
+  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> > gpu_kernel(d_kernel,3,4,2);
+  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_out(d_out,74,35,8,136,17);
+
+  Eigen::array<Eigen::DenseIndex, 3> dims{1,2,3};
+  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 74; ++i) {
+    for (int j = 0; j < 35; ++j) {
+      for (int k = 0; k < 8; ++k) {
+        for (int l = 0; l < 136; ++l) {
+          for (int m = 0; m < 17; ++m) {
+            const float result = out(i,j,k,l,m);
+            const float expected = input(i,j+0,k+0,l+0,m) * kernel(0,0,0) +
+                                   input(i,j+1,k+0,l+0,m) * kernel(1,0,0) +
+                                   input(i,j+2,k+0,l+0,m) * kernel(2,0,0) +
+                                   input(i,j+0,k+1,l+0,m) * kernel(0,1,0) +
+                                   input(i,j+1,k+1,l+0,m) * kernel(1,1,0) +
+                                   input(i,j+2,k+1,l+0,m) * kernel(2,1,0) +
+                                   input(i,j+0,k+2,l+0,m) * kernel(0,2,0) +
+                                   input(i,j+1,k+2,l+0,m) * kernel(1,2,0) +
+                                   input(i,j+2,k+2,l+0,m) * kernel(2,2,0) +
+                                   input(i,j+0,k+3,l+0,m) * kernel(0,3,0) +
+                                   input(i,j+1,k+3,l+0,m) * kernel(1,3,0) +
+                                   input(i,j+2,k+3,l+0,m) * kernel(2,3,0) +
+                                   input(i,j+0,k+0,l+1,m) * kernel(0,0,1) +
+                                   input(i,j+1,k+0,l+1,m) * kernel(1,0,1) +
+                                   input(i,j+2,k+0,l+1,m) * kernel(2,0,1) +
+                                   input(i,j+0,k+1,l+1,m) * kernel(0,1,1) +
+                                   input(i,j+1,k+1,l+1,m) * kernel(1,1,1) +
+                                   input(i,j+2,k+1,l+1,m) * kernel(2,1,1) +
+                                   input(i,j+0,k+2,l+1,m) * kernel(0,2,1) +
+                                   input(i,j+1,k+2,l+1,m) * kernel(1,2,1) +
+                                   input(i,j+2,k+2,l+1,m) * kernel(2,2,1) +
+                                   input(i,j+0,k+3,l+1,m) * kernel(0,3,1) +
+                                   input(i,j+1,k+3,l+1,m) * kernel(1,3,1) +
+                                   input(i,j+2,k+3,l+1,m) * kernel(2,3,1);
+            VERIFY_IS_APPROX(result, expected);
+          }
+        }
+      }
+    }
+  }
+
+  hipFree(d_input);
+  hipFree(d_kernel);
+  hipFree(d_out);
+}
+
+
+template <typename Scalar>
+void test_hip_lgamma(const Scalar stddev)
+{
+  Tensor<Scalar, 2> in(72,97);
+  in.setRandom();
+  in *= in.constant(stddev);
+  Tensor<Scalar, 2> out(72,97);
+  out.setZero();
+
+  std::size_t bytes = in.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
+
+  gpu_out.device(gpu_device) = gpu_in.lgamma();
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 72; ++i) {
+    for (int j = 0; j < 97; ++j) {
+      VERIFY_IS_APPROX(out(i,j), (std::lgamma)(in(i,j)));
+    }
+  }
+
+  hipFree(d_in);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_digamma()
+{
+  Tensor<Scalar, 1> in(7);
+  Tensor<Scalar, 1> out(7);
+  Tensor<Scalar, 1> expected_out(7);
+  out.setZero();
+
+  in(0) = Scalar(1);
+  in(1) = Scalar(1.5);
+  in(2) = Scalar(4);
+  in(3) = Scalar(-10.5);
+  in(4) = Scalar(10000.5);
+  in(5) = Scalar(0);
+  in(6) = Scalar(-1);
+
+  expected_out(0) = Scalar(-0.5772156649015329);
+  expected_out(1) = Scalar(0.03648997397857645);
+  expected_out(2) = Scalar(1.2561176684318);
+  expected_out(3) = Scalar(2.398239129535781);
+  expected_out(4) = Scalar(9.210340372392849);
+  expected_out(5) = std::numeric_limits<Scalar>::infinity();
+  expected_out(6) = std::numeric_limits<Scalar>::infinity();
+
+  std::size_t bytes = in.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 7);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 7);
+
+  gpu_out.device(gpu_device) = gpu_in.digamma();
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 5; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+  for (int i = 5; i < 7; ++i) {
+    VERIFY_IS_EQUAL(out(i), expected_out(i));
+  }
+
+  hipFree(d_in);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_zeta()
+{
+  Tensor<Scalar, 1> in_x(6);
+  Tensor<Scalar, 1> in_q(6);
+  Tensor<Scalar, 1> out(6);
+  Tensor<Scalar, 1> expected_out(6);
+  out.setZero();
+
+  in_x(0) = Scalar(1);
+  in_x(1) = Scalar(1.5);
+  in_x(2) = Scalar(4);
+  in_x(3) = Scalar(-10.5);
+  in_x(4) = Scalar(10000.5);
+  in_x(5) = Scalar(3);
+  
+  in_q(0) = Scalar(1.2345);
+  in_q(1) = Scalar(2);
+  in_q(2) = Scalar(1.5);
+  in_q(3) = Scalar(3);
+  in_q(4) = Scalar(1.0001);
+  in_q(5) = Scalar(-2.5);
+
+  expected_out(0) = std::numeric_limits<Scalar>::infinity();
+  expected_out(1) = Scalar(1.61237534869);
+  expected_out(2) = Scalar(0.234848505667);
+  expected_out(3) = Scalar(1.03086757337e-5);
+  expected_out(4) = Scalar(0.367879440865);
+  expected_out(5) = Scalar(0.054102025820864097);
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_in_x;
+  Scalar* d_in_q;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in_x), bytes);
+  hipMalloc((void**)(&d_in_q), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in_q, in_q.data(), bytes, hipMemcpyHostToDevice);
+  
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_q(d_in_q, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 6);
+
+  gpu_out.device(gpu_device) = gpu_in_x.zeta(gpu_in_q);
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  VERIFY_IS_EQUAL(out(0), expected_out(0));
+  VERIFY((std::isnan)(out(3)));
+
+  for (int i = 1; i < 6; ++i) {
+    if (i != 3) {
+      VERIFY_IS_APPROX(out(i), expected_out(i));
+    }
+  }
+
+  hipFree(d_in_x);
+  hipFree(d_in_q);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_polygamma()
+{
+  Tensor<Scalar, 1> in_x(7);
+  Tensor<Scalar, 1> in_n(7);
+  Tensor<Scalar, 1> out(7);
+  Tensor<Scalar, 1> expected_out(7);
+  out.setZero();
+
+  in_n(0) = Scalar(1);
+  in_n(1) = Scalar(1);
+  in_n(2) = Scalar(1);
+  in_n(3) = Scalar(17);
+  in_n(4) = Scalar(31);
+  in_n(5) = Scalar(28);
+  in_n(6) = Scalar(8);
+  
+  in_x(0) = Scalar(2);
+  in_x(1) = Scalar(3);
+  in_x(2) = Scalar(25.5);
+  in_x(3) = Scalar(4.7);
+  in_x(4) = Scalar(11.8);
+  in_x(5) = Scalar(17.7);
+  in_x(6) = Scalar(30.2);
+
+  expected_out(0) = Scalar(0.644934066848);
+  expected_out(1) = Scalar(0.394934066848);
+  expected_out(2) = Scalar(0.0399946696496);
+  expected_out(3) = Scalar(293.334565435);
+  expected_out(4) = Scalar(0.445487887616);
+  expected_out(5) = Scalar(-2.47810300902e-07);
+  expected_out(6) = Scalar(-8.29668781082e-09);
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_in_x;
+  Scalar* d_in_n;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in_x), bytes);
+  hipMalloc((void**)(&d_in_n), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in_n, in_n.data(), bytes, hipMemcpyHostToDevice);
+  
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 7);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_n(d_in_n, 7);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 7);
+
+  gpu_out.device(gpu_device) = gpu_in_n.polygamma(gpu_in_x);
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 7; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+
+  hipFree(d_in_x);
+  hipFree(d_in_n);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_igamma()
+{
+  Tensor<Scalar, 2> a(6, 6);
+  Tensor<Scalar, 2> x(6, 6);
+  Tensor<Scalar, 2> out(6, 6);
+  out.setZero();
+
+  Scalar a_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
+  Scalar x_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
+
+  for (int i = 0; i < 6; ++i) {
+    for (int j = 0; j < 6; ++j) {
+      a(i, j) = a_s[i];
+      x(i, j) = x_s[j];
+    }
+  }
+
+  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
+  Scalar igamma_s[][6] = {{0.0, nan, nan, nan, nan, nan},
+                          {0.0, 0.6321205588285578, 0.7768698398515702,
+                           0.9816843611112658, 9.999500016666262e-05, 1.0},
+                          {0.0, 0.4275932955291202, 0.608374823728911,
+                           0.9539882943107686, 7.522076445089201e-07, 1.0},
+                          {0.0, 0.01898815687615381, 0.06564245437845008,
+                           0.5665298796332909, 4.166333347221828e-18, 1.0},
+                          {0.0, 0.9999780593618628, 0.9999899967080838,
+                           0.9999996219837988, 0.9991370418689945, 1.0},
+                          {0.0, 0.0, 0.0, 0.0, 0.0, 0.5042041932513908}};
+
+
+
+  std::size_t bytes = a.size() * sizeof(Scalar);
+
+  Scalar* d_a;
+  Scalar* d_x;
+  Scalar* d_out;
+  assert(hipMalloc((void**)(&d_a), bytes) == hipSuccess);
+  assert(hipMalloc((void**)(&d_x), bytes) == hipSuccess);
+  assert(hipMalloc((void**)(&d_out), bytes) == hipSuccess);
+
+  hipMemcpy(d_a, a.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_x, x.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_x(d_x, 6, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 6, 6);
+
+  gpu_out.device(gpu_device) = gpu_a.igamma(gpu_x);
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 6; ++i) {
+    for (int j = 0; j < 6; ++j) {
+      if ((std::isnan)(igamma_s[i][j])) {
+        VERIFY((std::isnan)(out(i, j)));
+      } else {
+        VERIFY_IS_APPROX(out(i, j), igamma_s[i][j]);
+      }
+    }
+  }
+
+  hipFree(d_a);
+  hipFree(d_x);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_igammac()
+{
+  Tensor<Scalar, 2> a(6, 6);
+  Tensor<Scalar, 2> x(6, 6);
+  Tensor<Scalar, 2> out(6, 6);
+  out.setZero();
+
+  Scalar a_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
+  Scalar x_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
+
+  for (int i = 0; i < 6; ++i) {
+    for (int j = 0; j < 6; ++j) {
+      a(i, j) = a_s[i];
+      x(i, j) = x_s[j];
+    }
+  }
+
+  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
+  Scalar igammac_s[][6] = {{nan, nan, nan, nan, nan, nan},
+                           {1.0, 0.36787944117144233, 0.22313016014842982,
+                            0.018315638888734182, 0.9999000049998333, 0.0},
+                           {1.0, 0.5724067044708798, 0.3916251762710878,
+                            0.04601170568923136, 0.9999992477923555, 0.0},
+                           {1.0, 0.9810118431238462, 0.9343575456215499,
+                            0.4334701203667089, 1.0, 0.0},
+                           {1.0, 2.1940638138146658e-05, 1.0003291916285e-05,
+                            3.7801620118431334e-07, 0.0008629581310054535,
+                            0.0},
+                           {1.0, 1.0, 1.0, 1.0, 1.0, 0.49579580674813944}};
+
+  std::size_t bytes = a.size() * sizeof(Scalar);
+
+  Scalar* d_a;
+  Scalar* d_x;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_a), bytes);
+  hipMalloc((void**)(&d_x), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_a, a.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_x, x.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_x(d_x, 6, 6);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 6, 6);
+
+  gpu_out.device(gpu_device) = gpu_a.igammac(gpu_x);
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 6; ++i) {
+    for (int j = 0; j < 6; ++j) {
+      if ((std::isnan)(igammac_s[i][j])) {
+        VERIFY((std::isnan)(out(i, j)));
+      } else {
+        VERIFY_IS_APPROX(out(i, j), igammac_s[i][j]);
+      }
+    }
+  }
+
+  hipFree(d_a);
+  hipFree(d_x);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_erf(const Scalar stddev)
+{
+  Tensor<Scalar, 2> in(72,97);
+  in.setRandom();
+  in *= in.constant(stddev);
+  Tensor<Scalar, 2> out(72,97);
+  out.setZero();
+
+  std::size_t bytes = in.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  assert(hipMalloc((void**)(&d_in), bytes) == hipSuccess);
+  assert(hipMalloc((void**)(&d_out), bytes) == hipSuccess);
+
+  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
+
+  gpu_out.device(gpu_device) = gpu_in.erf();
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 72; ++i) {
+    for (int j = 0; j < 97; ++j) {
+      VERIFY_IS_APPROX(out(i,j), (std::erf)(in(i,j)));
+    }
+  }
+
+  hipFree(d_in);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_erfc(const Scalar stddev)
+{
+  Tensor<Scalar, 2> in(72,97);
+  in.setRandom();
+  in *= in.constant(stddev);
+  Tensor<Scalar, 2> out(72,97);
+  out.setZero();
+
+  std::size_t bytes = in.size() * sizeof(Scalar);
+
+  Scalar* d_in;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
+
+  gpu_out.device(gpu_device) = gpu_in.erfc();
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 0; i < 72; ++i) {
+    for (int j = 0; j < 97; ++j) {
+      VERIFY_IS_APPROX(out(i,j), (std::erfc)(in(i,j)));
+    }
+  }
+
+  hipFree(d_in);
+  hipFree(d_out);
+}
+
+template <typename Scalar>
+void test_hip_betainc()
+{
+  Tensor<Scalar, 1> in_x(125);
+  Tensor<Scalar, 1> in_a(125);
+  Tensor<Scalar, 1> in_b(125);
+  Tensor<Scalar, 1> out(125);
+  Tensor<Scalar, 1> expected_out(125);
+  out.setZero();
+
+  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
+
+  Array<Scalar, 1, Dynamic> x(125);
+  Array<Scalar, 1, Dynamic> a(125);
+  Array<Scalar, 1, Dynamic> b(125);
+  Array<Scalar, 1, Dynamic> v(125);
+
+  a << 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
+      0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999,
+      0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999,
+      0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
+      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999,
+      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999,
+      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999;
+
+  b << 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
+      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
+      999.999, 999.999, 999.999, 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 999.999, 999.999, 999.999, 999.999, 999.999, 0.0, 0.0,
+      0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
+      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
+      999.999, 999.999, 999.999, 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 999.999, 999.999, 999.999, 999.999, 999.999, 0.0, 0.0,
+      0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
+      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
+      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
+      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
+      999.999, 999.999, 999.999;
+
+  x << -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8,
+      1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5,
+      0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2,
+      0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1,
+      0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1,
+      -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8,
+      1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5,
+      0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2,
+      0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1;
+
+  v << nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
+      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
+      nan, nan, 0.47972119876364683, 0.5, 0.5202788012363533, nan, nan,
+      0.9518683957740043, 0.9789663010413743, 0.9931729188073435, nan, nan,
+      0.999995949033062, 0.9999999999993698, 0.9999999999999999, nan, nan,
+      0.9999999999999999, 0.9999999999999999, 0.9999999999999999, nan, nan, nan,
+      nan, nan, nan, nan, 0.006827081192655869, 0.0210336989586256,
+      0.04813160422599567, nan, nan, 0.20014344256217678, 0.5000000000000001,
+      0.7998565574378232, nan, nan, 0.9991401428435834, 0.999999999698403,
+      0.9999999999999999, nan, nan, 0.9999999999999999, 0.9999999999999999,
+      0.9999999999999999, nan, nan, nan, nan, nan, nan, nan,
+      1.0646600232370887e-25, 6.301722877826246e-13, 4.050966937974938e-06, nan,
+      nan, 7.864342668429763e-23, 3.015969667594166e-10, 0.0008598571564165444,
+      nan, nan, 6.031987710123844e-08, 0.5000000000000007, 0.9999999396801229,
+      nan, nan, 0.9999999999999999, 0.9999999999999999, 0.9999999999999999, nan,
+      nan, nan, nan, nan, nan, nan, 0.0, 7.029920380986636e-306,
+      2.2450728208591345e-101, nan, nan, 0.0, 9.275871147869727e-302,
+      1.2232913026152827e-97, nan, nan, 0.0, 3.0891393081932924e-252,
+      2.9303043666183996e-60, nan, nan, 2.248913486879199e-196,
+      0.5000000000004947, 0.9999999999999999, nan;
+
+  for (int i = 0; i < 125; ++i) {
+    in_x(i) = x(i);
+    in_a(i) = a(i);
+    in_b(i) = b(i);
+    expected_out(i) = v(i);
+  }
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_in_x;
+  Scalar* d_in_a;
+  Scalar* d_in_b;
+  Scalar* d_out;
+  hipMalloc((void**)(&d_in_x), bytes);
+  hipMalloc((void**)(&d_in_a), bytes);
+  hipMalloc((void**)(&d_in_b), bytes);
+  hipMalloc((void**)(&d_out), bytes);
+
+  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in_a, in_a.data(), bytes, hipMemcpyHostToDevice);
+  hipMemcpy(d_in_b, in_b.data(), bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 125);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_a(d_in_a, 125);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_b(d_in_b, 125);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 125);
+
+  gpu_out.device(gpu_device) = betainc(gpu_in_a, gpu_in_b, gpu_in_x);
+
+  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  for (int i = 1; i < 125; ++i) {
+    if ((std::isnan)(expected_out(i))) {
+      VERIFY((std::isnan)(out(i)));
+    } else {
+      VERIFY_IS_APPROX(out(i), expected_out(i));
+    }
+  }
+
+  hipFree(d_in_x);
+  hipFree(d_in_a);
+  hipFree(d_in_b);
+  hipFree(d_out);
+}
+
+
+void test_cxx11_tensor_hip()
+{
+  CALL_SUBTEST(test_hip_nullary());
+  CALL_SUBTEST(test_hip_elementwise_small());
+  CALL_SUBTEST(test_hip_elementwise());
+  CALL_SUBTEST(test_hip_props());
+  CALL_SUBTEST(test_hip_reduction());
+  // FIXME : uncommenting following tests results in compile failure
+  // CALL_SUBTEST(test_hip_contraction<ColMajor>());
+  // CALL_SUBTEST(test_hip_contraction<RowMajor>());
+  CALL_SUBTEST(test_hip_convolution_1d<ColMajor>());
+  CALL_SUBTEST(test_hip_convolution_1d<RowMajor>());
+  CALL_SUBTEST(test_hip_convolution_inner_dim_col_major_1d());
+  CALL_SUBTEST(test_hip_convolution_inner_dim_row_major_1d());
+  CALL_SUBTEST(test_hip_convolution_2d<ColMajor>());
+  CALL_SUBTEST(test_hip_convolution_2d<RowMajor>());
+  // The following two tests commented out due to long runtime
+  // CALL_SUBTEST(test_hip_convolution_3d<ColMajor>());
+  // CALL_SUBTEST(test_hip_convolution_3d<RowMajor>());
+
+#if __cplusplus > 199711L
+  // std::erf, std::erfc, and so on where only added in c++11. We use them
+  // as a golden reference to validate the results produced by Eigen. Therefore
+  // we can only run these tests if we use a c++11 compiler.
+
+  CALL_SUBTEST(test_hip_lgamma<float>(1.0f));
+  CALL_SUBTEST(test_hip_lgamma<float>(100.0f));
+  CALL_SUBTEST(test_hip_lgamma<float>(0.01f));
+  CALL_SUBTEST(test_hip_lgamma<float>(0.001f));
+
+  CALL_SUBTEST(test_hip_lgamma<double>(1.0));
+  CALL_SUBTEST(test_hip_lgamma<double>(100.0));
+  CALL_SUBTEST(test_hip_lgamma<double>(0.01));
+  CALL_SUBTEST(test_hip_lgamma<double>(0.001));
+
+  CALL_SUBTEST(test_hip_erf<float>(1.0f));
+  CALL_SUBTEST(test_hip_erf<float>(100.0f));
+  CALL_SUBTEST(test_hip_erf<float>(0.01f));
+  CALL_SUBTEST(test_hip_erf<float>(0.001f));
+
+  CALL_SUBTEST(test_hip_erfc<float>(1.0f));
+  // CALL_SUBTEST(test_hip_erfc<float>(100.0f)); // HIP erfc lacks precision for large inputs
+  CALL_SUBTEST(test_hip_erfc<float>(5.0f)); 
+  CALL_SUBTEST(test_hip_erfc<float>(0.01f));
+  CALL_SUBTEST(test_hip_erfc<float>(0.001f));
+
+  CALL_SUBTEST(test_hip_erf<double>(1.0));
+  CALL_SUBTEST(test_hip_erf<double>(100.0));
+  CALL_SUBTEST(test_hip_erf<double>(0.01));
+  CALL_SUBTEST(test_hip_erf<double>(0.001));
+
+  CALL_SUBTEST(test_hip_erfc<double>(1.0));
+  // CALL_SUBTEST(test_hip_erfc<double>(100.0)); // HIP erfc lacks precision for large inputs
+  CALL_SUBTEST(test_hip_erfc<double>(5.0)); 
+  CALL_SUBTEST(test_hip_erfc<double>(0.01));
+  CALL_SUBTEST(test_hip_erfc<double>(0.001));
+
+  // Following tests have functional failures on some seeds
+  // CALL_SUBTEST(test_hip_digamma<float>());
+  // CALL_SUBTEST(test_hip_digamma<double>());
+
+  // Following tests have functional failures on some seeds
+  // CALL_SUBTEST(test_hip_polygamma<float>());
+  // CALL_SUBTEST(test_hip_polygamma<double>());
+
+  // Following tests have functional failures on some seeds
+  // CALL_SUBTEST(test_hip_zeta<float>());
+  // CALL_SUBTEST(test_hip_zeta<double>());
+
+  CALL_SUBTEST(test_hip_igamma<float>());
+  CALL_SUBTEST(test_hip_igammac<float>());
+
+  CALL_SUBTEST(test_hip_igamma<double>());
+  CALL_SUBTEST(test_hip_igammac<double>());
+
+  CALL_SUBTEST(test_hip_betainc<float>());
+  CALL_SUBTEST(test_hip_betainc<double>());
+#endif
+}
diff --git a/unsupported/test/cxx11_tensor_of_float16_hip.cu b/unsupported/test/cxx11_tensor_of_float16_hip.cu
new file mode 100644
index 000000000..6c1a401bf
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_of_float16_hip.cu
@@ -0,0 +1,498 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+
+using Eigen::Tensor;
+
+template<typename>
+void test_hip_numext() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  bool* d_res_half = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
+  bool* d_res_float = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_half(
+      d_res_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_float(
+      d_res_float, num_elem);
+
+  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
+  gpu_res_float.device(gpu_device) = gpu_float.unaryExpr(Eigen::internal::scalar_isnan_op<float>());
+  gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().unaryExpr(Eigen::internal::scalar_isnan_op<Eigen::half>());
+
+  Tensor<bool, 1> half_prec(num_elem);
+  Tensor<bool, 1> full_prec(num_elem);
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(bool));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(bool));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking numext " << i << std::endl;
+    VERIFY_IS_EQUAL(full_prec(i), half_prec(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+
+#ifdef EIGEN_HAS_HIP_FP16
+
+template<typename>
+void test_hip_conversion() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+  
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
+      d_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
+      d_conv, num_elem);
+
+  gpu_float.device(gpu_device) = gpu_float.random();
+  gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
+  gpu_conv.device(gpu_device) = gpu_half.cast<float>();
+
+  Tensor<float, 1> initial(num_elem);
+  Tensor<float, 1> final(num_elem);
+  gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
+
+  for (int i = 0; i < num_elem; ++i) {
+    VERIFY_IS_APPROX(initial(i), final(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_half);
+  gpu_device.deallocate(d_conv);
+}
+
+template<typename>
+void test_hip_unary() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
+      d_res_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
+      d_res_float, num_elem);
+
+  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
+  gpu_res_float.device(gpu_device) = gpu_float.abs();
+  gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().cast<float>();
+
+  Tensor<float, 1> half_prec(num_elem);
+  Tensor<float, 1> full_prec(num_elem);
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking unary " << i << std::endl;
+    VERIFY_IS_APPROX(full_prec(i), half_prec(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+template<typename>
+void test_hip_elementwise() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(
+      d_float1, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(
+      d_float2, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
+      d_res_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
+      d_res_float, num_elem);
+
+  gpu_float1.device(gpu_device) = gpu_float1.random();
+  gpu_float2.device(gpu_device) = gpu_float2.random();
+  gpu_res_float.device(gpu_device) = (gpu_float1 + gpu_float2) * gpu_float1;
+  gpu_res_half.device(gpu_device) = ((gpu_float1.cast<Eigen::half>() + gpu_float2.cast<Eigen::half>()) * gpu_float1.cast<Eigen::half>()).cast<float>();
+
+  Tensor<float, 1> half_prec(num_elem);
+  Tensor<float, 1> full_prec(num_elem);
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking elemwise " << i << ": full prec = " << full_prec(i) << " vs half prec = " << half_prec(i) << std::endl;
+    VERIFY_IS_APPROX(static_cast<Eigen::half>(full_prec(i)), static_cast<Eigen::half>(half_prec(i)));
+  }
+
+  gpu_device.deallocate(d_float1);
+  gpu_device.deallocate(d_float2);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+template<typename>
+void test_hip_trancendental() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float3 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_res1_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res1_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res2_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res2_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res3_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res3_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(d_float1, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(d_float2, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float3(d_float3, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half(d_res1_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float(d_res1_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half(d_res2_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float(d_res2_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_half(d_res3_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_float(d_res3_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res4_half(d_res3_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res4_float(d_res3_float, num_elem);
+
+  gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
+  gpu_float2.device(gpu_device) = gpu_float2.random() + gpu_float1.constant(0.5f);
+  gpu_float3.device(gpu_device) = gpu_float3.random();
+  gpu_res1_float.device(gpu_device) = gpu_float1.exp().cast<Eigen::half>();
+  gpu_res2_float.device(gpu_device) = gpu_float2.log().cast<Eigen::half>();
+  gpu_res3_float.device(gpu_device) = gpu_float3.log1p().cast<Eigen::half>();
+  gpu_res4_float.device(gpu_device) = gpu_float3.expm1().cast<Eigen::half>();
+
+  gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>();
+  gpu_res1_half.device(gpu_device) = gpu_res1_half.exp();
+
+  gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>();
+  gpu_res2_half.device(gpu_device) = gpu_res2_half.log();
+
+  gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>();
+  gpu_res3_half.device(gpu_device) = gpu_res3_half.log1p();
+
+  gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>();
+  gpu_res3_half.device(gpu_device) = gpu_res3_half.expm1();
+
+  Tensor<float, 1> input1(num_elem);
+  Tensor<Eigen::half, 1> half_prec1(num_elem);
+  Tensor<Eigen::half, 1> full_prec1(num_elem);
+  Tensor<float, 1> input2(num_elem);
+  Tensor<Eigen::half, 1> half_prec2(num_elem);
+  Tensor<Eigen::half, 1> full_prec2(num_elem);
+  Tensor<float, 1> input3(num_elem);
+  Tensor<Eigen::half, 1> half_prec3(num_elem);
+  Tensor<Eigen::half, 1> full_prec3(num_elem);
+  gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem*sizeof(Eigen::half));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i) << " half = " << half_prec1(i) << std::endl;
+    VERIFY_IS_APPROX(full_prec1(i), half_prec1(i));
+  }
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
+    if(std::abs(input2(i)-1.f)<0.05f) // log lacks accurary nearby 1
+      VERIFY_IS_APPROX(full_prec2(i)+Eigen::half(0.1f), half_prec2(i)+Eigen::half(0.1f));
+    else
+      VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
+  }
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i) << " half = " << half_prec3(i) << std::endl;
+    VERIFY_IS_APPROX(full_prec3(i), half_prec3(i));
+  }
+  gpu_device.deallocate(d_float1);
+  gpu_device.deallocate(d_float2);
+  gpu_device.deallocate(d_float3);
+  gpu_device.deallocate(d_res1_half);
+  gpu_device.deallocate(d_res1_float);
+  gpu_device.deallocate(d_res2_half);
+  gpu_device.deallocate(d_res2_float);
+  gpu_device.deallocate(d_res3_float);
+  gpu_device.deallocate(d_res3_half);
+}
+
+template<typename>
+void test_hip_contractions() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int rows = 23;
+  int cols = 23;
+  int num_elem = rows*cols;
+
+  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
+      d_float1, rows, cols);
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
+      d_float2, rows, cols);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(
+      d_res_half, rows, cols);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(
+      d_res_float, rows, cols);
+
+  gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
+  gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float2.constant(0.5f);
+
+  typedef Tensor<float, 2>::DimensionPair DimPair;
+  Eigen::array<DimPair, 1> dims(DimPair(1, 0));
+  gpu_res_float.device(gpu_device) = gpu_float1.contract(gpu_float2, dims).cast<Eigen::half>();
+  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().contract(gpu_float2.cast<Eigen::half>(), dims);
+
+  Tensor<Eigen::half, 2> half_prec(rows, cols);
+  Tensor<Eigen::half, 2> full_prec(rows, cols);
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(Eigen::half));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < rows; ++i) {
+    for (int j = 0; j < cols; ++j) {
+      std::cout << "Checking contract " << i << " " << j << full_prec(i, j) << " " << half_prec(i, j) << std::endl;
+      if (numext::abs(full_prec(i, j) - half_prec(i, j)) > Eigen::half(1e-2f)) {
+        VERIFY_IS_APPROX(full_prec(i, j), half_prec(i, j));
+      }
+    }
+  }
+
+  gpu_device.deallocate(d_float1);
+  gpu_device.deallocate(d_float2);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+template<typename>
+void test_hip_reductions(int size1, int size2, int redux) {
+
+   std::cout << "Reducing " << size1 << " by " << size2
+             << " tensor along dim " << redux << std::endl; 
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = size1*size2;
+  int result_size = (redux == 1 ? size1 : size2);
+
+  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
+  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
+      d_float1, size1, size2);
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
+      d_float2, size1, size2);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(
+      d_res_half, result_size);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(
+      d_res_float, result_size);
+
+  gpu_float1.device(gpu_device) = gpu_float1.random() * 2.0f;
+  gpu_float2.device(gpu_device) = gpu_float2.random() * 2.0f;
+
+  Eigen::array<int, 1> redux_dim(redux);
+  gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim).cast<Eigen::half>();
+  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim);
+
+  Tensor<Eigen::half, 1> half_prec(result_size);
+  Tensor<Eigen::half, 1> full_prec(result_size);
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, result_size*sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, result_size*sizeof(Eigen::half));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < result_size; ++i) {
+    std::cout << "EXPECTED " << full_prec(i) << " GOT " << half_prec(i) << std::endl;
+    VERIFY_IS_APPROX(full_prec(i), half_prec(i));
+  }
+
+  gpu_device.deallocate(d_float1);
+  gpu_device.deallocate(d_float2);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+template<typename>
+void test_hip_reductions() {
+  test_hip_reductions<void>(13, 13, 0);
+  test_hip_reductions<void>(13, 13, 1);
+
+  test_hip_reductions<void>(35, 36, 0);
+  test_hip_reductions<void>(35, 36, 1);
+
+  test_hip_reductions<void>(36, 35, 0);
+  test_hip_reductions<void>(36, 35, 1);
+}
+
+template<typename>
+void test_hip_full_reductions() {
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int size = 13;
+  int num_elem = size*size;
+
+  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
+  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
+      d_float1, size, size);
+  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
+      d_float2, size, size);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(
+      d_res_half);
+  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(
+      d_res_float);
+
+  gpu_float1.device(gpu_device) = gpu_float1.random();
+  gpu_float2.device(gpu_device) = gpu_float2.random();
+
+  gpu_res_float.device(gpu_device) = gpu_float1.sum().cast<Eigen::half>();
+  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum();
+
+  Tensor<Eigen::half, 0> half_prec;
+  Tensor<Eigen::half, 0> full_prec;
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
+  gpu_device.synchronize();
+
+  VERIFY_IS_APPROX(full_prec(), half_prec());
+
+  gpu_res_float.device(gpu_device) = gpu_float1.maximum().cast<Eigen::half>();
+  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().maximum();
+  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
+  gpu_device.synchronize();
+
+  VERIFY_IS_APPROX(full_prec(), half_prec());
+
+  gpu_device.deallocate(d_float1);
+  gpu_device.deallocate(d_float2);
+  gpu_device.deallocate(d_res_half);
+  gpu_device.deallocate(d_res_float);
+}
+
+template<typename>
+void test_hip_forced_evals() {
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_half1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_half2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half1(
+      d_res_half1, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Unaligned> gpu_res_half2(
+      d_res_half2, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
+      d_res_float, num_elem);
+
+  Eigen::array<int, 1> no_bcast;
+  no_bcast[0] = 1;
+
+  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
+  gpu_res_float.device(gpu_device) = gpu_float.abs();
+  gpu_res_half1.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().eval().cast<float>();
+  gpu_res_half2.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().broadcast(no_bcast).eval().cast<float>();
+
+  Tensor<float, 1> half_prec1(num_elem);
+  Tensor<float, 1> half_prec2(num_elem);
+  Tensor<float, 1> full_prec(num_elem);
+  gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res_half1, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res_half1, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
+  gpu_device.synchronize();
+
+  for (int i = 0; i < num_elem; ++i) {
+    std::cout << "Checking forced eval " << i << full_prec(i) << " vs " << half_prec1(i) << " vs " << half_prec2(i) << std::endl;
+    VERIFY_IS_APPROX(full_prec(i), half_prec1(i));
+    VERIFY_IS_APPROX(full_prec(i), half_prec2(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_res_half1);
+  gpu_device.deallocate(d_res_half2);
+  gpu_device.deallocate(d_res_float);
+}
+#endif
+
+
+void test_cxx11_tensor_of_float16_hip()
+{
+  CALL_SUBTEST(test_hip_numext<void>());
+
+#ifdef EIGEN_HAS_HIP_FP16
+  CALL_SUBTEST(test_hip_conversion<void>());
+  CALL_SUBTEST(test_hip_unary<void>());
+  CALL_SUBTEST(test_hip_elementwise<void>());
+  CALL_SUBTEST(test_hip_trancendental<void>());
+  CALL_SUBTEST(test_hip_contractions<void>());
+  CALL_SUBTEST(test_hip_reductions<void>());
+  CALL_SUBTEST(test_hip_full_reductions<void>());
+  CALL_SUBTEST(test_hip_forced_evals<void>());
+#else
+  std::cout << "Half floats are not supported by this version of hip: skipping the test" << std::endl;
+#endif
+}
diff --git a/unsupported/test/cxx11_tensor_random_hip.cu b/unsupported/test/cxx11_tensor_random_hip.cu
new file mode 100644
index 000000000..7d7e72ca2
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_random_hip.cu
@@ -0,0 +1,85 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_random_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <Eigen/CXX11/Tensor>
+
+
+void test_hip_random_uniform()
+{
+  Tensor<float, 2> out(72,97);
+  out.setZero();
+
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_out;
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
+
+  gpu_out.device(gpu_device) = gpu_out.random();
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+
+  // For now we just check thes code doesn't crash.
+  // TODO: come up with a valid test of randomness
+}
+
+
+void test_hip_random_normal()
+{
+  Tensor<float, 2> out(72,97);
+  out.setZero();
+
+  std::size_t out_bytes = out.size() * sizeof(float);
+
+  float* d_out;
+  hipMalloc((void**)(&d_out), out_bytes);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
+
+  Eigen::internal::NormalRandomGenerator<float> gen(true);
+  gpu_out.device(gpu_device) = gpu_out.random(gen);
+
+  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
+  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
+}
+
+static void test_complex()
+{
+  Tensor<std::complex<float>, 1> vec(6);
+  vec.setRandom();
+
+  // Fixme: we should check that the generated numbers follow a uniform
+  // distribution instead.
+  for (int i = 1; i < 6; ++i) {
+    VERIFY_IS_NOT_EQUAL(vec(i), vec(i-1));
+  }
+}
+
+
+void test_cxx11_tensor_random_hip()
+{
+  CALL_SUBTEST(test_hip_random_uniform());
+  CALL_SUBTEST(test_hip_random_normal());
+  CALL_SUBTEST(test_complex());
+}
diff --git a/unsupported/test/cxx11_tensor_reduction_hip.cu b/unsupported/test/cxx11_tensor_reduction_hip.cu
new file mode 100644
index 000000000..c5aad05be
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_reduction_hip.cu
@@ -0,0 +1,154 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_reduction_hip
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+
+template<typename Type, int DataLayout>
+static void test_full_reductions() {
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  const int num_rows = internal::random<int>(1024, 5*1024);
+  const int num_cols = internal::random<int>(1024, 5*1024);
+
+  Tensor<Type, 2, DataLayout> in(num_rows, num_cols);
+  in.setRandom();
+
+  Tensor<Type, 0, DataLayout> full_redux;
+  full_redux = in.sum();
+
+  std::size_t in_bytes = in.size() * sizeof(Type);
+  std::size_t out_bytes = full_redux.size() * sizeof(Type);
+  Type* gpu_in_ptr = static_cast<Type*>(gpu_device.allocate(in_bytes));
+  Type* gpu_out_ptr = static_cast<Type*>(gpu_device.allocate(out_bytes));
+  gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
+
+  TensorMap<Tensor<Type, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols);
+  TensorMap<Tensor<Type, 0, DataLayout> > out_gpu(gpu_out_ptr);
+
+  out_gpu.device(gpu_device) = in_gpu.sum();
+
+  Tensor<Type, 0, DataLayout> full_redux_gpu;
+  gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
+  gpu_device.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  VERIFY_IS_APPROX(full_redux(), full_redux_gpu());
+
+  gpu_device.deallocate(gpu_in_ptr);
+  gpu_device.deallocate(gpu_out_ptr);
+}
+
+template<typename Type, int DataLayout>
+static void test_first_dim_reductions() {
+  int dim_x = 33;
+  int dim_y = 1;
+  int dim_z = 128;
+
+  Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z);
+  in.setRandom();
+
+  Eigen::array<int, 1> red_axis;
+  red_axis[0] = 0;
+  Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
+
+  // Create device
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice dev(&stream);
+  
+  // Create data(T)
+  Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type));
+  Type* out_data = (Type*)dev.allocate(dim_z*dim_y*sizeof(Type));
+  Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z);
+  Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_y, dim_z);
+  
+  // Perform operation
+  dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type));
+  gpu_out.device(dev) = gpu_in.sum(red_axis);
+  gpu_out.device(dev) += gpu_in.sum(red_axis);
+  Tensor<Type, 2, DataLayout> redux_gpu(dim_y, dim_z);
+  dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type));
+  dev.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  for (int i = 0; i < gpu_out.size(); ++i) {
+    VERIFY_IS_APPROX(2*redux(i), redux_gpu(i));
+  }
+
+  dev.deallocate(in_data);
+  dev.deallocate(out_data);
+}
+
+template<typename Type, int DataLayout>
+static void test_last_dim_reductions() {
+  int dim_x = 128;
+  int dim_y = 1;
+  int dim_z = 33;
+
+  Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z);
+  in.setRandom();
+
+  Eigen::array<int, 1> red_axis;
+  red_axis[0] = 2;
+  Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
+
+  // Create device
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice dev(&stream);
+  
+  // Create data
+  Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type));
+  Type* out_data = (Type*)dev.allocate(dim_x*dim_y*sizeof(Type));
+  Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z);
+  Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_x, dim_y);
+  
+  // Perform operation
+  dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type));
+  gpu_out.device(dev) = gpu_in.sum(red_axis);
+  gpu_out.device(dev) += gpu_in.sum(red_axis);
+  Tensor<Type, 2, DataLayout> redux_gpu(dim_x, dim_y);
+  dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type));
+  dev.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  for (int i = 0; i < gpu_out.size(); ++i) {
+    VERIFY_IS_APPROX(2*redux(i), redux_gpu(i));
+  }
+
+  dev.deallocate(in_data);
+  dev.deallocate(out_data);
+}
+
+
+void test_cxx11_tensor_reduction_hip() {
+  CALL_SUBTEST((test_full_reductions<float, ColMajor>()));
+  CALL_SUBTEST((test_full_reductions<double, ColMajor>()));
+  CALL_SUBTEST((test_full_reductions<float, RowMajor>()));
+  CALL_SUBTEST((test_full_reductions<double, RowMajor>()));
+  
+  CALL_SUBTEST((test_first_dim_reductions<float, ColMajor>()));
+  CALL_SUBTEST((test_first_dim_reductions<double, ColMajor>()));
+  CALL_SUBTEST((test_first_dim_reductions<float, RowMajor>()));
+// Outer reductions of doubles aren't supported just yet.  					      
+//  CALL_SUBTEST((test_first_dim_reductions<double, RowMajor>()))
+
+  CALL_SUBTEST((test_last_dim_reductions<float, ColMajor>()));
+// Outer reductions of doubles aren't supported just yet.  					      
+//  CALL_SUBTEST((test_last_dim_reductions<double, ColMajor>()));
+  CALL_SUBTEST((test_last_dim_reductions<float, RowMajor>()));
+  CALL_SUBTEST((test_last_dim_reductions<double, RowMajor>()));
+}
diff --git a/unsupported/test/cxx11_tensor_scan_hip.cu b/unsupported/test/cxx11_tensor_scan_hip.cu
new file mode 100644
index 000000000..f4c4f59b9
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_scan_hip.cu
@@ -0,0 +1,76 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_scan_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+typedef Tensor<float, 1>::DimensionPair DimPair;
+
+template<int DataLayout>
+void test_hip_cumsum(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size);
+  Tensor<float, 3, DataLayout> t_result(m_size, k_size, n_size);
+  Tensor<float, 3, DataLayout> t_result_gpu(m_size, k_size, n_size);
+
+  t_input.setRandom();
+
+  std::size_t t_input_bytes = t_input.size()  * sizeof(float);
+  std::size_t t_result_bytes = t_result.size() * sizeof(float);
+
+  float* d_t_input;
+  float* d_t_result;
+
+  hipMalloc((void**)(&d_t_input), t_input_bytes);
+  hipMalloc((void**)(&d_t_result), t_result_bytes);
+
+  hipMemcpy(d_t_input, t_input.data(), t_input_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
+      gpu_t_input(d_t_input, Eigen::array<int, 3>(m_size, k_size, n_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
+      gpu_t_result(d_t_result, Eigen::array<int, 3>(m_size, k_size, n_size));
+
+  gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1);
+  t_result = t_input.cumsum(1);
+
+  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
+  for (DenseIndex i = 0; i < t_result.size(); i++) {
+    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
+      continue;
+    }
+    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
+      continue;
+    }
+    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
+              << " vs " <<  t_result_gpu(i) << std::endl;
+    assert(false);
+  }
+
+  hipFree((void*)d_t_input);
+  hipFree((void*)d_t_result);
+}
+
+
+void test_cxx11_tensor_scan_hip()
+{
+  CALL_SUBTEST(test_hip_cumsum<ColMajor>(128, 128, 128));
+  CALL_SUBTEST(test_hip_cumsum<RowMajor>(128, 128, 128));
+}

From 4bd158fa37b4bba74e6421575d5c69eeea547172 Mon Sep 17 00:00:00 2001
From: Michael Figurnov <mfigurnov@google.com>
Date: Wed, 6 Jun 2018 18:49:26 +0100
Subject: [PATCH 359/680] Derivative of the incomplete Gamma function and the
 sample of a Gamma random variable.

In addition to igamma(a, x), this code implements:
* igamma_der_a(a, x) = d igamma(a, x) / da -- derivative of igamma with respect to the parameter
* gamma_sample_der_alpha(alpha, sample) -- reparameterization derivative of a Gamma(alpha, 1) random variable sample with respect to the alpha parameter

The derivatives are computed by forward mode differentiation of the igamma(a, x) code. Although gamma_sample_der_alpha can be implemented via igamma_der_a, a separate function is more accurate and efficient due to analytical cancellation of some terms. All three functions are implemented by a method parameterized with "mode" that always computes the derivatives, but does not return them unless required by the mode. The compiler is expected to (and, based on benchmarks, does) skip the unnecessary computations depending on the mode.
---
 Eigen/src/Core/GenericPacketMath.h            |   2 +
 Eigen/src/Core/arch/CUDA/PacketMath.h         |   4 +
 .../Eigen/CXX11/src/Tensor/TensorBase.h       |  14 +
 unsupported/Eigen/SpecialFunctions            |   2 +
 .../SpecialFunctionsArrayAPI.h                |  42 ++
 .../SpecialFunctionsFunctors.h                |  54 ++
 .../SpecialFunctions/SpecialFunctionsHalf.h   |   8 +
 .../SpecialFunctions/SpecialFunctionsImpl.h   | 573 +++++++++++-------
 .../SpecialFunctionsPacketMath.h              |  15 +
 .../arch/CUDA/CudaSpecialFunctions.h          |  35 ++
 unsupported/test/cxx11_tensor_cuda.cu         | 157 +++++
 unsupported/test/special_functions.cpp        |  92 +++
 12 files changed, 791 insertions(+), 207 deletions(-)

diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 888a3f7ea..55b6a89e2 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -85,6 +85,8 @@ struct default_packet_traits
     HasI0e = 0,
     HasI1e = 0,
     HasIGamma = 0,
+    HasIGammaDerA = 0,
+    HasGammaSampleDerAlpha = 0,
     HasIGammac = 0,
     HasBetaInc = 0,
 
diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h
index 704a4e0d9..ab8e477f4 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMath.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -47,6 +47,8 @@ template<> struct packet_traits<float> : default_packet_traits
     HasI0e = 1,
     HasI1e = 1,
     HasIGamma = 1,
+    HasIGammaDerA = 1,
+    HasGammaSampleDerAlpha = 1,
     HasIGammac = 1,
     HasBetaInc = 1,
 
@@ -78,6 +80,8 @@ template<> struct packet_traits<double> : default_packet_traits
     HasI0e = 1,
     HasI1e = 1,
     HasIGamma = 1,
+    HasIGammaDerA = 1,
+    HasGammaSampleDerAlpha = 1,
     HasIGammac = 1,
     HasBetaInc = 1,
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index a942c98dd..88b0af0d3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -152,6 +152,20 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return binaryExpr(other.derived(), internal::scalar_igamma_op<Scalar>());
     }
 
+    // igamma_der_a(a = this, x = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_igamma_der_a_op<Scalar>, const Derived, const OtherDerived>
+    igamma_der_a(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_igamma_der_a_op<Scalar>());
+    }
+
+    // gamma_sample_der_alpha(alpha = this, sample = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_gamma_sample_der_alpha_op<Scalar>, const Derived, const OtherDerived>
+    gamma_sample_der_alpha(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_gamma_sample_der_alpha_op<Scalar>());
+    }
+
     // igammac(a = this, x = other)
     template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorCwiseBinaryOp<internal::scalar_igammac_op<Scalar>, const Derived, const OtherDerived>
diff --git a/unsupported/Eigen/SpecialFunctions b/unsupported/Eigen/SpecialFunctions
index 482ec6e6f..9441ba8f5 100644
--- a/unsupported/Eigen/SpecialFunctions
+++ b/unsupported/Eigen/SpecialFunctions
@@ -29,6 +29,8 @@ namespace Eigen {
   * - erfc
   * - lgamma
   * - igamma
+  * - igamma_der_a
+  * - gamma_sample_der_alpha
   * - igammac
   * - digamma
   * - polygamma
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
index b7a9d035b..30cdf4751 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
@@ -33,6 +33,48 @@ igamma(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerive
   );
 }
 
+/** \cpp11 \returns an expression of the coefficient-wise igamma_der_a(\a a, \a x) to the given arrays.
+  *
+  * This function computes the coefficient-wise derivative of the incomplete
+  * gamma function with respect to the parameter a.
+  *
+  * \note This function supports only float and double scalar types in c++11
+  * mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations
+  * of igamma_der_a(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igamma(), Eigen::lgamma()
+  */
+template <typename Derived, typename ExponentDerived>
+inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+igamma_der_a(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x) {
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
+    a.derived(),
+    x.derived());
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise gamma_sample_der_alpha(\a alpha, \a sample) to the given arrays.
+  *
+  * This function computes the coefficient-wise derivative of the sample
+  * of a Gamma(alpha, 1) random variable with respect to the parameter alpha.
+  *
+  * \note This function supports only float and double scalar types in c++11
+  * mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations
+  * of gamma_sample_der_alpha(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igamma(), Eigen::lgamma()
+  */
+template <typename AlphaDerived, typename SampleDerived>
+inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>
+gamma_sample_der_alpha(const Eigen::ArrayBase<AlphaDerived>& alpha, const Eigen::ArrayBase<SampleDerived>& sample) {
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>(
+      alpha.derived(),
+      sample.derived());
+}
+
 /** \cpp11 \returns an expression of the coefficient-wise igammac(\a a, \a x) to the given arrays.
   *
   * This function computes the coefficient-wise complementary incomplete gamma function.
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
index 8420f0174..3a63dcdd6 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
@@ -41,6 +41,60 @@ struct functor_traits<scalar_igamma_op<Scalar> > {
   };
 };
 
+/** \internal
+  * \brief Template functor to compute the derivative of the incomplete gamma
+  * function igamma_der_a(a, x)
+  *
+  * \sa class CwiseBinaryOp, Cwise::igamma_der_a
+  */
+template <typename Scalar>
+struct scalar_igamma_der_a_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_igamma_der_a_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& a, const Scalar& x) const {
+    using numext::igamma_der_a;
+    return igamma_der_a(a, x);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& x) const {
+    return internal::pigamma_der_a(a, x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_igamma_der_a_op<Scalar> > {
+  enum {
+    // 2x the cost of igamma
+    Cost = 40 * NumTraits<Scalar>::MulCost + 20 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasIGammaDerA
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the derivative of the sample
+  * of a Gamma(alpha, 1) random variable with respect to the parameter alpha
+  * gamma_sample_der_alpha(alpha, sample)
+  *
+  * \sa class CwiseBinaryOp, Cwise::gamma_sample_der_alpha
+  */
+template <typename Scalar>
+struct scalar_gamma_sample_der_alpha_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_gamma_sample_der_alpha_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& alpha, const Scalar& sample) const {
+    using numext::gamma_sample_der_alpha;
+    return gamma_sample_der_alpha(alpha, sample);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& alpha, const Packet& sample) const {
+    return internal::pgamma_sample_der_alpha(alpha, sample);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_gamma_sample_der_alpha_op<Scalar> > {
+  enum {
+    // 2x the cost of igamma, minus the lgamma cost (the lgamma cancels out)
+    Cost = 30 * NumTraits<Scalar>::MulCost + 15 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasGammaSampleDerAlpha
+  };
+};
 
 /** \internal
   * \brief Template functor to compute the complementary incomplete gamma function igammac(a, x)
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
index c5867002e..fbdfd299e 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsHalf.h
@@ -33,6 +33,14 @@ template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half erfc(const Eigen::h
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igamma(const Eigen::half& a, const Eigen::half& x) {
   return Eigen::half(Eigen::numext::igamma(static_cast<float>(a), static_cast<float>(x)));
 }
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igamma_der_a(const Eigen::half& a, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::igamma_der_a(static_cast<float>(a), static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half gamma_sample_der_alpha(const Eigen::half& alpha, const Eigen::half& sample) {
+  return Eigen::half(Eigen::numext::gamma_sample_der_alpha(static_cast<float>(alpha), static_cast<float>(sample)));
+}
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igammac(const Eigen::half& a, const Eigen::half& x) {
   return Eigen::half(Eigen::numext::igammac(static_cast<float>(a), static_cast<float>(x)));
 }
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 6c7ac3f3b..b24df2a95 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -521,6 +521,198 @@ struct cephes_helper<double> {
   }
 };
 
+enum IgammaComputationMode { VALUE, DERIVATIVE, SAMPLE_DERIVATIVE };
+
+template <typename Scalar, IgammaComputationMode mode>
+EIGEN_DEVICE_FUNC
+int igamma_num_iterations() {
+  /* Returns the maximum number of internal iterations for igamma computation.
+   */
+  if (mode == VALUE) {
+    return 2000;
+  }
+
+  if (internal::is_same<Scalar, float>::value) {
+    return 200;
+  } else if (internal::is_same<Scalar, double>::value) {
+    return 500;
+  } else {
+    return 2000;
+  }
+}
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igammac_cf_impl {
+  /* Computes igamc(a, x) or derivative (depending on the mode)
+   * using the continued fraction expansion of the complementary
+   * incomplete Gamma function.
+   *
+   * Preconditions:
+   *   a > 0
+   *   x >= 1
+   *   x >= a
+   */
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar two = 2;
+    const Scalar machep = cephes_helper<Scalar>::machep();
+    const Scalar big = cephes_helper<Scalar>::big();
+    const Scalar biginv = cephes_helper<Scalar>::biginv();
+
+    if ((numext::isinf)(x)) {
+      return zero;
+    }
+
+    // continued fraction
+    Scalar y = one - a;
+    Scalar z = x + y + one;
+    Scalar c = zero;
+    Scalar pkm2 = one;
+    Scalar qkm2 = x;
+    Scalar pkm1 = x + one;
+    Scalar qkm1 = z * x;
+    Scalar ans = pkm1 / qkm1;
+
+    Scalar dpkm2_da = zero;
+    Scalar dqkm2_da = zero;
+    Scalar dpkm1_da = zero;
+    Scalar dqkm1_da = -x;
+    Scalar dans_da = (dpkm1_da - ans * dqkm1_da) / qkm1;
+
+    for (int i = 0; i < igamma_num_iterations<Scalar, mode>(); i++) {
+      c += one;
+      y += one;
+      z += two;
+
+      Scalar yc = y * c;
+      Scalar pk = pkm1 * z - pkm2 * yc;
+      Scalar qk = qkm1 * z - qkm2 * yc;
+
+      Scalar dpk_da = dpkm1_da * z - pkm1 - dpkm2_da * yc + pkm2 * c;
+      Scalar dqk_da = dqkm1_da * z - qkm1 - dqkm2_da * yc + qkm2 * c;
+
+      if (qk != zero) {
+        Scalar ans_prev = ans;
+        ans = pk / qk;
+
+        Scalar dans_da_prev = dans_da;
+        dans_da = (dpk_da - ans * dqk_da) / qk;
+
+        if (mode == VALUE) {
+          if (numext::abs(ans_prev - ans) <= machep * numext::abs(ans)) {
+            break;
+          }
+        } else {
+          if (numext::abs(dans_da - dans_da_prev) <=
+              machep * numext::abs(dans_da)) {
+            break;
+          }
+        }
+      }
+
+      pkm2 = pkm1;
+      pkm1 = pk;
+      qkm2 = qkm1;
+      qkm1 = qk;
+
+      dpkm2_da = dpkm1_da;
+      dpkm1_da = dpk_da;
+      dqkm2_da = dqkm1_da;
+      dqkm1_da = dqk_da;
+
+      if (numext::abs(pk) > big) {
+        pkm2 *= biginv;
+        pkm1 *= biginv;
+        qkm2 *= biginv;
+        qkm1 *= biginv;
+
+        dpkm2_da *= biginv;
+        dpkm1_da *= biginv;
+        dqkm2_da *= biginv;
+        dqkm1_da *= biginv;
+      }
+    }
+
+    /* Compute  x**a * exp(-x) / gamma(a)  */
+    Scalar logax = a * numext::log(x) - x - lgamma_impl<Scalar>::run(a);
+    Scalar dlogax_da = numext::log(x) - digamma_impl<Scalar>::run(a);
+    Scalar ax = numext::exp(logax);
+    Scalar dax_da = ax * dlogax_da;
+
+    switch (mode) {
+      case VALUE:
+        return ans * ax;
+      case DERIVATIVE:
+        return ans * dax_da + dans_da * ax;
+      case SAMPLE_DERIVATIVE:
+        return -(dans_da + ans * dlogax_da) * x;
+    }
+  }
+};
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_series_impl {
+  /* Computes igam(a, x) or its derivative (depending on the mode)
+   * using the series expansion of the incomplete Gamma function.
+   *
+   * Preconditions:
+   *   x > 0
+   *   a > 0
+   *   !(x > 1 && x > a)
+   */
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar machep = cephes_helper<Scalar>::machep();
+
+    /* power series */
+    Scalar r = a;
+    Scalar c = one;
+    Scalar ans = one;
+
+    Scalar dc_da = zero;
+    Scalar dans_da = zero;
+
+    for (int i = 0; i < igamma_num_iterations<Scalar, mode>(); i++) {
+      r += one;
+      Scalar term = x / r;
+      Scalar dterm_da = -x / (r * r);
+      dc_da = term * dc_da + dterm_da * c;
+      dans_da += dc_da;
+      c *= term;
+      ans += c;
+
+      if (mode == VALUE) {
+        if (c <= machep * ans) {
+          break;
+        }
+      } else {
+        if (numext::abs(dc_da) <= machep * numext::abs(dans_da)) {
+          break;
+        }
+      }
+    }
+
+    /* Compute  x**a * exp(-x) / gamma(a + 1)  */
+    Scalar logax = a * numext::log(x) - x - lgamma_impl<Scalar>::run(a + one);
+    Scalar dlogax_da = numext::log(x) - digamma_impl<Scalar>::run(a + one);
+    Scalar ax = numext::exp(logax);
+    Scalar dax_da = ax * dlogax_da;
+
+    switch (mode) {
+      case VALUE:
+        return ans * ax;
+      case DERIVATIVE:
+        return ans * dax_da + dans_da * ax;
+      case SAMPLE_DERIVATIVE:
+        return -(dans_da + ans * dlogax_da) * x / a;
+    }
+  }
+};
+
 #if !EIGEN_HAS_C99_MATH
 
 template <typename Scalar>
@@ -535,8 +727,6 @@ struct igammac_impl {
 
 #else
 
-template <typename Scalar> struct igamma_impl;  // predeclare igamma_impl
-
 template <typename Scalar>
 struct igammac_impl {
   EIGEN_DEVICE_FUNC
@@ -604,97 +794,15 @@ struct igammac_impl {
       return nan;
     }
 
-    if ((numext::isnan)(a) || (numext::isnan)(x)) { // propagate nans
+    if ((numext::isnan)(a) || (numext::isnan)(x)) {  // propagate nans
       return nan;
     }
 
     if ((x < one) || (x < a)) {
-      /* The checks above ensure that we meet the preconditions for
-       * igamma_impl::Impl(), so call it, rather than igamma_impl::Run().
-       * Calling Run() would also work, but in that case the compiler may not be
-       * able to prove that igammac_impl::Run and igamma_impl::Run are not
-       * mutually recursive.  This leads to worse code, particularly on
-       * platforms like nvptx, where recursion is allowed only begrudgingly.
-       */
-      return (one - igamma_impl<Scalar>::Impl(a, x));
+      return (one - igamma_series_impl<Scalar, VALUE>::run(a, x));
     }
 
-    return Impl(a, x);
-  }
-
- private:
-  /* igamma_impl calls igammac_impl::Impl. */
-  friend struct igamma_impl<Scalar>;
-
-  /* Actually computes igamc(a, x).
-   *
-   * Preconditions:
-   *   a > 0
-   *   x >= 1
-   *   x >= a
-   */
-  EIGEN_DEVICE_FUNC static Scalar Impl(Scalar a, Scalar x) {
-    const Scalar zero = 0;
-    const Scalar one = 1;
-    const Scalar two = 2;
-    const Scalar machep = cephes_helper<Scalar>::machep();
-    const Scalar maxlog = numext::log(NumTraits<Scalar>::highest());
-    const Scalar big = cephes_helper<Scalar>::big();
-    const Scalar biginv = cephes_helper<Scalar>::biginv();
-    const Scalar inf = NumTraits<Scalar>::infinity();
-
-    Scalar ans, ax, c, yc, r, t, y, z;
-    Scalar pk, pkm1, pkm2, qk, qkm1, qkm2;
-
-    if (x == inf) return zero;  // std::isinf crashes on CUDA
-
-    /* Compute  x**a * exp(-x) / gamma(a)  */
-    ax = a * numext::log(x) - x - lgamma_impl<Scalar>::run(a);
-    if (ax < -maxlog) {  // underflow
-      return zero;
-    }
-    ax = numext::exp(ax);
-
-    // continued fraction
-    y = one - a;
-    z = x + y + one;
-    c = zero;
-    pkm2 = one;
-    qkm2 = x;
-    pkm1 = x + one;
-    qkm1 = z * x;
-    ans = pkm1 / qkm1;
-
-    for (int i = 0; i < 2000; i++) {
-      c += one;
-      y += one;
-      z += two;
-      yc = y * c;
-      pk = pkm1 * z - pkm2 * yc;
-      qk = qkm1 * z - qkm2 * yc;
-      if (qk != zero) {
-        r = pk / qk;
-        t = numext::abs((ans - r) / r);
-        ans = r;
-      } else {
-        t = one;
-      }
-      pkm2 = pkm1;
-      pkm1 = pk;
-      qkm2 = qkm1;
-      qkm1 = qk;
-      if (numext::abs(pk) > big) {
-        pkm2 *= biginv;
-        pkm1 *= biginv;
-        qkm2 *= biginv;
-        qkm1 *= biginv;
-      }
-      if (t <= machep) {
-        break;
-      }
-    }
-
-    return (ans * ax);
+    return igammac_cf_impl<Scalar, VALUE>::run(a, x);
   }
 };
 
@@ -704,15 +812,10 @@ struct igammac_impl {
  * Implementation of igamma (incomplete gamma integral), based on Cephes but requires C++11/C99 *
  ************************************************************************************************/
 
-template <typename Scalar>
-struct igamma_retval {
-  typedef Scalar type;
-};
-
 #if !EIGEN_HAS_C99_MATH
 
-template <typename Scalar>
-struct igamma_impl {
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_generic_impl {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Scalar run(Scalar a, Scalar x) {
     EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
@@ -723,69 +826,17 @@ struct igamma_impl {
 
 #else
 
-template <typename Scalar>
-struct igamma_impl {
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_generic_impl {
   EIGEN_DEVICE_FUNC
   static Scalar run(Scalar a, Scalar x) {
-    /*	igam()
-     *	Incomplete gamma integral
-     *
-     *
-     *
-     * SYNOPSIS:
-     *
-     * double a, x, y, igam();
-     *
-     * y = igam( a, x );
-     *
-     * DESCRIPTION:
-     *
-     * The function is defined by
-     *
-     *                           x
-     *                            -
-     *                   1       | |  -t  a-1
-     *  igam(a,x)  =   -----     |   e   t   dt.
-     *                  -      | |
-     *                 | (a)    -
-     *                           0
-     *
-     *
-     * In this implementation both arguments must be positive.
-     * The integral is evaluated by either a power series or
-     * continued fraction expansion, depending on the relative
-     * values of a and x.
-     *
-     * ACCURACY (double):
-     *
-     *                      Relative error:
-     * arithmetic   domain     # trials      peak         rms
-     *    IEEE      0,30       200000       3.6e-14     2.9e-15
-     *    IEEE      0,100      300000       9.9e-14     1.5e-14
-     *
-     *
-     * ACCURACY (float):
-     *
-     *                      Relative error:
-     * arithmetic   domain     # trials      peak         rms
-     *    IEEE      0,30        20000       7.8e-6      5.9e-7
-     *
-     */
-    /*
-      Cephes Math Library Release 2.2: June, 1992
-      Copyright 1985, 1987, 1992 by Stephen L. Moshier
-      Direct inquiries to 30 Frost Street, Cambridge, MA 02140
-    */
-
-
-    /* left tail of incomplete gamma function:
-     *
-     *          inf.      k
-     *   a  -x   -       x
-     *  x  e     >   ----------
-     *           -     -
-     *          k=0   | (a+k+1)
+    /* Depending on the mode, returns
+     * - VALUE: incomplete Gamma function igamma(a, x)
+     * - DERIVATIVE: derivative of incomplete Gamma function d/da igamma(a, x)
+     * - SAMPLE_DERIVATIVE: implicit derivative of a Gamma random variable
+     * x ~ Gamma(x | a, 1), dx/da = -1 / Gamma(x | a, 1) * d igamma(a, x) / dx
      *
+     * Derivatives are implemented by forward-mode differentiation.
      */
     const Scalar zero = 0;
     const Scalar one = 1;
@@ -797,71 +848,167 @@ struct igamma_impl {
       return nan;
     }
 
-    if ((numext::isnan)(a) || (numext::isnan)(x)) { // propagate nans
+    if ((numext::isnan)(a) || (numext::isnan)(x)) {  // propagate nans
       return nan;
     }
 
     if ((x > one) && (x > a)) {
-      /* The checks above ensure that we meet the preconditions for
-       * igammac_impl::Impl(), so call it, rather than igammac_impl::Run().
-       * Calling Run() would also work, but in that case the compiler may not be
-       * able to prove that igammac_impl::Run and igamma_impl::Run are not
-       * mutually recursive.  This leads to worse code, particularly on
-       * platforms like nvptx, where recursion is allowed only begrudgingly.
-       */
-      return (one - igammac_impl<Scalar>::Impl(a, x));
-    }
-
-    return Impl(a, x);
-  }
-
- private:
-  /* igammac_impl calls igamma_impl::Impl. */
-  friend struct igammac_impl<Scalar>;
-
-  /* Actually computes igam(a, x).
-   *
-   * Preconditions:
-   *   x > 0
-   *   a > 0
-   *   !(x > 1 && x > a)
-   */
-  EIGEN_DEVICE_FUNC static Scalar Impl(Scalar a, Scalar x) {
-    const Scalar zero = 0;
-    const Scalar one = 1;
-    const Scalar machep = cephes_helper<Scalar>::machep();
-    const Scalar maxlog = numext::log(NumTraits<Scalar>::highest());
-
-    Scalar ans, ax, c, r;
-
-    /* Compute  x**a * exp(-x) / gamma(a)  */
-    ax = a * numext::log(x) - x - lgamma_impl<Scalar>::run(a);
-    if (ax < -maxlog) {
-      // underflow
-      return zero;
-    }
-    ax = numext::exp(ax);
-
-    /* power series */
-    r = a;
-    c = one;
-    ans = one;
-
-    for (int i = 0; i < 2000; i++) {
-      r += one;
-      c *= x/r;
-      ans += c;
-      if (c/ans <= machep) {
-        break;
+      Scalar ret = igammac_cf_impl<Scalar, mode>::run(a, x);
+      if (mode == VALUE) {
+        return one - ret;
+      } else {
+        return -ret;
       }
     }
 
-    return (ans * ax / a);
+    return igamma_series_impl<Scalar, mode>::run(a, x);
   }
 };
 
 #endif  // EIGEN_HAS_C99_MATH
 
+template <typename Scalar>
+struct igamma_retval {
+  typedef Scalar type;
+};
+
+template <typename Scalar>
+struct igamma_impl : igamma_generic_impl<Scalar, VALUE> {
+  /* igam()
+   * Incomplete gamma integral.
+   *
+   * The CDF of Gamma(a, 1) random variable at the point x.
+   *
+   * Accuracy estimation. For each a in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables x ~ Gamma(x | a, 1), a total of 300 points.
+   * The ground truth is computed by mpmath. Mean absolute error:
+   * float: 1.26713e-05
+   * double: 2.33606e-12
+   *
+   * Cephes documentation below.
+   *
+   * SYNOPSIS:
+   *
+   * double a, x, y, igam();
+   *
+   * y = igam( a, x );
+   *
+   * DESCRIPTION:
+   *
+   * The function is defined by
+   *
+   *                           x
+   *                            -
+   *                   1       | |  -t  a-1
+   *  igam(a,x)  =   -----     |   e   t   dt.
+   *                  -      | |
+   *                 | (a)    -
+   *                           0
+   *
+   *
+   * In this implementation both arguments must be positive.
+   * The integral is evaluated by either a power series or
+   * continued fraction expansion, depending on the relative
+   * values of a and x.
+   *
+   * ACCURACY (double):
+   *
+   *                      Relative error:
+   * arithmetic   domain     # trials      peak         rms
+   *    IEEE      0,30       200000       3.6e-14     2.9e-15
+   *    IEEE      0,100      300000       9.9e-14     1.5e-14
+   *
+   *
+   * ACCURACY (float):
+   *
+   *                      Relative error:
+   * arithmetic   domain     # trials      peak         rms
+   *    IEEE      0,30        20000       7.8e-6      5.9e-7
+   *
+   */
+  /*
+    Cephes Math Library Release 2.2: June, 1992
+    Copyright 1985, 1987, 1992 by Stephen L. Moshier
+    Direct inquiries to 30 Frost Street, Cambridge, MA 02140
+  */
+
+  /* left tail of incomplete gamma function:
+   *
+   *          inf.      k
+   *   a  -x   -       x
+   *  x  e     >   ----------
+   *           -     -
+   *          k=0   | (a+k+1)
+   *
+   */
+};
+
+template <typename Scalar>
+struct igamma_der_a_retval : igamma_retval<Scalar> {};
+
+template <typename Scalar>
+struct igamma_der_a_impl : igamma_generic_impl<Scalar, DERIVATIVE> {
+  /* Derivative of the incomplete Gamma function with respect to a.
+   *
+   * Computes d/da igamma(a, x) by forward differentiation of the igamma code.
+   *
+   * Accuracy estimation. For each a in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables x ~ Gamma(x | a, 1), a total of 300 points.
+   * The ground truth is computed by mpmath. Mean absolute error:
+   * float: 6.27648e-07
+   * double: 4.60455e-12
+   *
+   * Reference:
+   * R. Moore. "Algorithm AS 187: Derivatives of the incomplete gamma
+   * integral". Journal of the Royal Statistical Society. 1982
+   */
+};
+
+template <typename Scalar>
+struct gamma_sample_der_alpha_retval : igamma_retval<Scalar> {};
+
+template <typename Scalar>
+struct gamma_sample_der_alpha_impl
+    : igamma_generic_impl<Scalar, SAMPLE_DERIVATIVE> {
+  /* Derivative of a Gamma random variable sample with respect to alpha.
+   *
+   * Consider a sample of a Gamma random variable with the concentration
+   * parameter alpha: sample ~ Gamma(alpha, 1). The reparameterization
+   * derivative that we want to compute is dsample / dalpha =
+   * d igammainv(alpha, u) / dalpha, where u = igamma(alpha, sample).
+   * However, this formula is numerically unstable and expensive, so instead
+   * we use implicit differentiation:
+   *
+   * igamma(alpha, sample) = u, where u ~ Uniform(0, 1).
+   * Apply d / dalpha to both sides:
+   * d igamma(alpha, sample) / dalpha
+   *     + d igamma(alpha, sample) / dsample * dsample/dalpha  = 0
+   * d igamma(alpha, sample) / dalpha
+   *     + Gamma(sample | alpha, 1) dsample / dalpha = 0
+   * dsample/dalpha = - (d igamma(alpha, sample) / dalpha)
+   *                   / Gamma(sample | alpha, 1)
+   *
+   * Here Gamma(sample | alpha, 1) is the PDF of the Gamma distribution
+   * (note that the derivative of the CDF w.r.t. sample is the PDF).
+   * See the reference below for more details.
+   *
+   * The derivative of igamma(alpha, sample) is computed by forward
+   * differentiation of the igamma code. Division by the Gamma PDF is performed
+   * in the same code, increasing the accuracy and speed due to cancellation
+   * of some terms.
+   *
+   * Accuracy estimation. For each alpha in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables sample ~ Gamma(sample | alpha, 1), a total of 300
+   * points. The ground truth is computed by mpmath. Mean absolute error:
+   * float: 1.0993e-06
+   * double: 1.47631e-12
+   *
+   * Reference:
+   * M. Figurnov, S. Mohamed, A. Mnih "Implicit Reparameterization Gradients".
+   * 2018
+   */
+};
+
 /*****************************************************************************
  * Implementation of Riemann zeta function of two arguments, based on Cephes *
  *****************************************************************************/
@@ -1950,6 +2097,18 @@ EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igamma, Scalar)
   return EIGEN_MATHFUNC_IMPL(igamma, Scalar)::run(a, x);
 }
 
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igamma_der_a, Scalar)
+    igamma_der_a(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(igamma_der_a, Scalar)::run(a, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(gamma_sample_der_alpha, Scalar)
+    gamma_sample_der_alpha(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(gamma_sample_der_alpha, Scalar)::run(a, x);
+}
+
 template <typename Scalar>
 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igammac, Scalar)
     igammac(const Scalar& a, const Scalar& x) {
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
index 4c176716b..465f41d54 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
@@ -42,6 +42,21 @@ Packet perfc(const Packet& a) { using numext::erfc; return erfc(a); }
 template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 Packet pigamma(const Packet& a, const Packet& x) { using numext::igamma; return igamma(a, x); }
 
+/** \internal \returns the derivative of the incomplete gamma function
+ * igamma_der_a(\a a, \a x) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet pigamma_der_a(const Packet& a, const Packet& x) {
+  using numext::igamma_der_a; return igamma_der_a(a, x);
+}
+
+/** \internal \returns compute the derivative of the sample
+  * of Gamma(alpha, 1) random variable with respect to the parameter a
+  * gamma_sample_der_alpha(\a alpha, \a sample) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet pgamma_sample_der_alpha(const Packet& alpha, const Packet& sample) {
+  using numext::gamma_sample_der_alpha; return gamma_sample_der_alpha(alpha, sample);
+}
+
 /** \internal \returns the complementary incomplete gamma function igammac(\a a, \a x) */
 template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 Packet pigammac(const Packet& a, const Packet& x) { using numext::igammac; return igammac(a, x); }
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
index c25fea0b3..020ac1b62 100644
--- a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
+++ b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
@@ -120,6 +120,41 @@ double2 pigamma<double2>(const double2& a, const double2& x)
   return make_double2(igamma(a.x, x.x), igamma(a.y, x.y));
 }
 
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pigamma_der_a<float4>(
+    const float4& a, const float4& x) {
+  using numext::igamma_der_a;
+  return make_float4(igamma_der_a(a.x, x.x), igamma_der_a(a.y, x.y),
+                     igamma_der_a(a.z, x.z), igamma_der_a(a.w, x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pigamma_der_a<double2>(const double2& a, const double2& x) {
+  using numext::igamma_der_a;
+  return make_double2(igamma_der_a(a.x, x.x), igamma_der_a(a.y, x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pgamma_sample_der_alpha<float4>(
+    const float4& alpha, const float4& sample) {
+  using numext::gamma_sample_der_alpha;
+  return make_float4(
+      gamma_sample_der_alpha(alpha.x, sample.x),
+      gamma_sample_der_alpha(alpha.y, sample.y),
+      gamma_sample_der_alpha(alpha.z, sample.z),
+      gamma_sample_der_alpha(alpha.w, sample.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pgamma_sample_der_alpha<double2>(const double2& alpha, const double2& sample) {
+  using numext::gamma_sample_der_alpha;
+  return make_double2(
+      gamma_sample_der_alpha(alpha.x, sample.x),
+      gamma_sample_der_alpha(alpha.y, sample.y));
+}
+
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 pigammac<float4>(const float4& a, const float4& x)
 {
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu
index 63d0a345a..f238ed5be 100644
--- a/unsupported/test/cxx11_tensor_cuda.cu
+++ b/unsupported/test/cxx11_tensor_cuda.cu
@@ -1318,6 +1318,157 @@ void test_cuda_i1e()
   cudaFree(d_out);
 }
 
+template <typename Scalar>
+void test_cuda_igamma_der_a()
+{
+  Tensor<Scalar, 1> in_x(30);
+  Tensor<Scalar, 1> in_a(30);
+  Tensor<Scalar, 1> out(30);
+  Tensor<Scalar, 1> expected_out(30);
+  out.setZero();
+
+  Array<Scalar, 1, Dynamic> in_a_array(30);
+  Array<Scalar, 1, Dynamic> in_x_array(30);
+  Array<Scalar, 1, Dynamic> expected_out_array(30);
+
+  // See special_functions.cpp for the Python code that generates the test data.
+
+  in_a_array << 0.01, 0.01, 0.01, 0.01, 0.01, 0.1, 0.1, 0.1, 0.1, 0.1, 1.0, 1.0,
+      1.0, 1.0, 1.0, 10.0, 10.0, 10.0, 10.0, 10.0, 100.0, 100.0, 100.0, 100.0,
+      100.0, 1000.0, 1000.0, 1000.0, 1000.0, 1000.0;
+
+  in_x_array << 1.25668890405e-26, 1.17549435082e-38, 1.20938905072e-05,
+      1.17549435082e-38, 1.17549435082e-38, 5.66572070696e-16, 0.0132865061065,
+      0.0200034203853, 6.29263709118e-17, 1.37160367764e-06, 0.333412038288,
+      1.18135687766, 0.580629033777, 0.170631439426, 0.786686768458,
+      7.63873279537, 13.1944344379, 11.896042354, 10.5830172417, 10.5020942233,
+      92.8918587747, 95.003720371, 86.3715926467, 96.0330217672, 82.6389930677,
+      968.702906754, 969.463546828, 1001.79726022, 955.047416547, 1044.27458568;
+
+  expected_out_array << -32.7256441441, -36.4394150514, -9.66467612263,
+      -36.4394150514, -36.4394150514, -1.0891900302, -2.66351229645,
+      -2.48666868596, -0.929700494428, -3.56327722764, -0.455320135314,
+      -0.391437214323, -0.491352055991, -0.350454834292, -0.471773162921,
+      -0.104084440522, -0.0723646747909, -0.0992828975532, -0.121638215446,
+      -0.122619605294, -0.0317670267286, -0.0359974812869, -0.0154359225363,
+      -0.0375775365921, -0.00794899153653, -0.00777303219211, -0.00796085782042,
+      -0.0125850719397, -0.00455500206958, -0.00476436993148;
+
+  for (int i = 0; i < 30; ++i) {
+    in_x(i) = in_x_array(i);
+    in_a(i) = in_a_array(i);
+    expected_out(i) = expected_out_array(i);
+  }
+
+  std::size_t bytes = in_x.size() * sizeof(Scalar);
+
+  Scalar* d_a;
+  Scalar* d_x;
+  Scalar* d_out;
+  cudaMalloc((void**)(&d_a), bytes);
+  cudaMalloc((void**)(&d_x), bytes);
+  cudaMalloc((void**)(&d_out), bytes);
+
+  cudaMemcpy(d_a, in_a.data(), bytes, cudaMemcpyHostToDevice);
+  cudaMemcpy(d_x, in_x.data(), bytes, cudaMemcpyHostToDevice);
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_a(d_a, 30);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_x(d_x, 30);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 30);
+
+  gpu_out.device(gpu_device) = gpu_a.igamma_der_a(gpu_x);
+
+  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
+                         gpu_device.stream()) == cudaSuccess);
+  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+
+  for (int i = 0; i < 30; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+
+  cudaFree(d_a);
+  cudaFree(d_x);
+  cudaFree(d_out);
+}
+
+template <typename Scalar>
+void test_cuda_gamma_sample_der_alpha()
+{
+  Tensor<Scalar, 1> in_alpha(30);
+  Tensor<Scalar, 1> in_sample(30);
+  Tensor<Scalar, 1> out(30);
+  Tensor<Scalar, 1> expected_out(30);
+  out.setZero();
+
+  Array<Scalar, 1, Dynamic> in_alpha_array(30);
+  Array<Scalar, 1, Dynamic> in_sample_array(30);
+  Array<Scalar, 1, Dynamic> expected_out_array(30);
+
+  // See special_functions.cpp for the Python code that generates the test data.
+
+  in_alpha_array << 0.01, 0.01, 0.01, 0.01, 0.01, 0.1, 0.1, 0.1, 0.1, 0.1, 1.0,
+      1.0, 1.0, 1.0, 1.0, 10.0, 10.0, 10.0, 10.0, 10.0, 100.0, 100.0, 100.0,
+      100.0, 100.0, 1000.0, 1000.0, 1000.0, 1000.0, 1000.0;
+
+  in_sample_array << 1.25668890405e-26, 1.17549435082e-38, 1.20938905072e-05,
+      1.17549435082e-38, 1.17549435082e-38, 5.66572070696e-16, 0.0132865061065,
+      0.0200034203853, 6.29263709118e-17, 1.37160367764e-06, 0.333412038288,
+      1.18135687766, 0.580629033777, 0.170631439426, 0.786686768458,
+      7.63873279537, 13.1944344379, 11.896042354, 10.5830172417, 10.5020942233,
+      92.8918587747, 95.003720371, 86.3715926467, 96.0330217672, 82.6389930677,
+      968.702906754, 969.463546828, 1001.79726022, 955.047416547, 1044.27458568;
+
+  expected_out_array << 7.42424742367e-23, 1.02004297287e-34, 0.0130155240738,
+      1.02004297287e-34, 1.02004297287e-34, 1.96505168277e-13, 0.525575786243,
+      0.713903991771, 2.32077561808e-14, 0.000179348049886, 0.635500453302,
+      1.27561284917, 0.878125852156, 0.41565819538, 1.03606488534,
+      0.885964824887, 1.16424049334, 1.10764479598, 1.04590810812,
+      1.04193666963, 0.965193152414, 0.976217589464, 0.93008035061,
+      0.98153216096, 0.909196397698, 0.98434963993, 0.984738050206,
+      1.00106492525, 0.97734200649, 1.02198794179;
+
+  for (int i = 0; i < 30; ++i) {
+    in_alpha(i) = in_alpha_array(i);
+    in_sample(i) = in_sample_array(i);
+    expected_out(i) = expected_out_array(i);
+  }
+
+  std::size_t bytes = in_alpha.size() * sizeof(Scalar);
+
+  Scalar* d_alpha;
+  Scalar* d_sample;
+  Scalar* d_out;
+  cudaMalloc((void**)(&d_alpha), bytes);
+  cudaMalloc((void**)(&d_sample), bytes);
+  cudaMalloc((void**)(&d_out), bytes);
+
+  cudaMemcpy(d_alpha, in_alpha.data(), bytes, cudaMemcpyHostToDevice);
+  cudaMemcpy(d_sample, in_sample.data(), bytes, cudaMemcpyHostToDevice);
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_alpha(d_alpha, 30);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_sample(d_sample, 30);
+  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 30);
+
+  gpu_out.device(gpu_device) = gpu_alpha.gamma_sample_der_alpha(gpu_sample);
+
+  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
+                         gpu_device.stream()) == cudaSuccess);
+  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+
+  for (int i = 0; i < 30; ++i) {
+    VERIFY_IS_APPROX(out(i), expected_out(i));
+  }
+
+  cudaFree(d_alpha);
+  cudaFree(d_sample);
+  cudaFree(d_out);
+}
 
 void test_cxx11_tensor_cuda()
 {
@@ -1396,5 +1547,11 @@ void test_cxx11_tensor_cuda()
 
   CALL_SUBTEST_6(test_cuda_i1e<float>());
   CALL_SUBTEST_6(test_cuda_i1e<double>());
+
+  CALL_SUBTEST_6(test_cuda_igamma_der_a<float>());
+  CALL_SUBTEST_6(test_cuda_igamma_der_a<double>());
+
+  CALL_SUBTEST_6(test_cuda_gamma_sample_der_alpha<float>());
+  CALL_SUBTEST_6(test_cuda_gamma_sample_der_alpha<double>());
 #endif
 }
diff --git a/unsupported/test/special_functions.cpp b/unsupported/test/special_functions.cpp
index 48d0db95e..29ba6203a 100644
--- a/unsupported/test/special_functions.cpp
+++ b/unsupported/test/special_functions.cpp
@@ -375,6 +375,98 @@ template<typename ArrayType> void array_special_functions()
     CALL_SUBTEST(res = i1e(x);
                  verify_component_wise(res, expected););
   }
+
+  /* Code to generate the data for the following two test cases.
+    N = 5
+    np.random.seed(3)
+
+    a = np.logspace(-2, 3, 6)
+    a = np.ravel(np.tile(np.reshape(a, [-1, 1]), [1, N]))
+    x = np.random.gamma(a, 1.0)
+    x = np.maximum(x, np.finfo(np.float32).tiny)
+
+    def igamma(a, x):
+      return mpmath.gammainc(a, 0, x, regularized=True)
+
+    def igamma_der_a(a, x):
+      res = mpmath.diff(lambda a_prime: igamma(a_prime, x), a)
+      return np.float64(res)
+
+    def gamma_sample_der_alpha(a, x):
+      igamma_x = igamma(a, x)
+      def igammainv_of_igamma(a_prime):
+        return mpmath.findroot(lambda x_prime: igamma(a_prime, x_prime) -
+            igamma_x, x, solver='newton')
+      return np.float64(mpmath.diff(igammainv_of_igamma, a))
+
+    v_igamma_der_a = np.vectorize(igamma_der_a)(a, x)
+    v_gamma_sample_der_alpha = np.vectorize(gamma_sample_der_alpha)(a, x)
+  */
+
+  // Test igamma_der_a
+  {
+    ArrayType a(30);
+    ArrayType x(30);
+    ArrayType res(30);
+    ArrayType v(30);
+
+    a << 0.01, 0.01, 0.01, 0.01, 0.01, 0.1, 0.1, 0.1, 0.1, 0.1, 1.0, 1.0, 1.0,
+        1.0, 1.0, 10.0, 10.0, 10.0, 10.0, 10.0, 100.0, 100.0, 100.0, 100.0,
+        100.0, 1000.0, 1000.0, 1000.0, 1000.0, 1000.0;
+
+    x << 1.25668890405e-26, 1.17549435082e-38, 1.20938905072e-05,
+        1.17549435082e-38, 1.17549435082e-38, 5.66572070696e-16,
+        0.0132865061065, 0.0200034203853, 6.29263709118e-17, 1.37160367764e-06,
+        0.333412038288, 1.18135687766, 0.580629033777, 0.170631439426,
+        0.786686768458, 7.63873279537, 13.1944344379, 11.896042354,
+        10.5830172417, 10.5020942233, 92.8918587747, 95.003720371,
+        86.3715926467, 96.0330217672, 82.6389930677, 968.702906754,
+        969.463546828, 1001.79726022, 955.047416547, 1044.27458568;
+
+    v << -32.7256441441, -36.4394150514, -9.66467612263, -36.4394150514,
+        -36.4394150514, -1.0891900302, -2.66351229645, -2.48666868596,
+        -0.929700494428, -3.56327722764, -0.455320135314, -0.391437214323,
+        -0.491352055991, -0.350454834292, -0.471773162921, -0.104084440522,
+        -0.0723646747909, -0.0992828975532, -0.121638215446, -0.122619605294,
+        -0.0317670267286, -0.0359974812869, -0.0154359225363, -0.0375775365921,
+        -0.00794899153653, -0.00777303219211, -0.00796085782042,
+        -0.0125850719397, -0.00455500206958, -0.00476436993148;
+
+    CALL_SUBTEST(res = igamma_der_a(a, x); verify_component_wise(res, v););
+  }
+
+  // Test gamma_sample_der_alpha
+  {
+    ArrayType alpha(30);
+    ArrayType sample(30);
+    ArrayType res(30);
+    ArrayType v(30);
+
+    alpha << 0.01, 0.01, 0.01, 0.01, 0.01, 0.1, 0.1, 0.1, 0.1, 0.1, 1.0, 1.0,
+        1.0, 1.0, 1.0, 10.0, 10.0, 10.0, 10.0, 10.0, 100.0, 100.0, 100.0, 100.0,
+        100.0, 1000.0, 1000.0, 1000.0, 1000.0, 1000.0;
+
+    sample << 1.25668890405e-26, 1.17549435082e-38, 1.20938905072e-05,
+        1.17549435082e-38, 1.17549435082e-38, 5.66572070696e-16,
+        0.0132865061065, 0.0200034203853, 6.29263709118e-17, 1.37160367764e-06,
+        0.333412038288, 1.18135687766, 0.580629033777, 0.170631439426,
+        0.786686768458, 7.63873279537, 13.1944344379, 11.896042354,
+        10.5830172417, 10.5020942233, 92.8918587747, 95.003720371,
+        86.3715926467, 96.0330217672, 82.6389930677, 968.702906754,
+        969.463546828, 1001.79726022, 955.047416547, 1044.27458568;
+
+    v << 7.42424742367e-23, 1.02004297287e-34, 0.0130155240738,
+        1.02004297287e-34, 1.02004297287e-34, 1.96505168277e-13, 0.525575786243,
+        0.713903991771, 2.32077561808e-14, 0.000179348049886, 0.635500453302,
+        1.27561284917, 0.878125852156, 0.41565819538, 1.03606488534,
+        0.885964824887, 1.16424049334, 1.10764479598, 1.04590810812,
+        1.04193666963, 0.965193152414, 0.976217589464, 0.93008035061,
+        0.98153216096, 0.909196397698, 0.98434963993, 0.984738050206,
+        1.00106492525, 0.97734200649, 1.02198794179;
+
+    CALL_SUBTEST(res = gamma_sample_der_alpha(alpha, sample);
+                 verify_component_wise(res, v););
+  }
 #endif
 }
 

From 5172a32849a310d8942273da307309721dc37c3f Mon Sep 17 00:00:00 2001
From: Michael Figurnov <mfigurnov@google.com>
Date: Thu, 7 Jun 2018 12:03:58 +0100
Subject: [PATCH 360/680] Updated the stopping criteria in igammac_cf_impl.

Previously, when computing the derivative, it used a relative error threshold. Now it uses an absolute error threshold. The behavior for computing the value is unchanged. This makes more sense since we do not expect the derivative to often be close to zero. This change makes the derivatives about 30% faster across the board. The error for the igamma_der_a is almost unchanged, while for gamma_sample_der_alpha it is a bit worse for float32 and unchanged for float64.
---
 .../Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index b24df2a95..78702905b 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -605,8 +605,7 @@ struct igammac_cf_impl {
             break;
           }
         } else {
-          if (numext::abs(dans_da - dans_da_prev) <=
-              machep * numext::abs(dans_da)) {
+          if (numext::abs(dans_da - dans_da_prev) <= machep) {
             break;
           }
         }
@@ -955,8 +954,8 @@ struct igamma_der_a_impl : igamma_generic_impl<Scalar, DERIVATIVE> {
    * Accuracy estimation. For each a in [10^-2, 10^-1...10^3] we sample
    * 50 Gamma random variables x ~ Gamma(x | a, 1), a total of 300 points.
    * The ground truth is computed by mpmath. Mean absolute error:
-   * float: 6.27648e-07
-   * double: 4.60455e-12
+   * float: 6.17992e-07
+   * double: 4.60453e-12
    *
    * Reference:
    * R. Moore. "Algorithm AS 187: Derivatives of the incomplete gamma
@@ -1000,8 +999,8 @@ struct gamma_sample_der_alpha_impl
    * Accuracy estimation. For each alpha in [10^-2, 10^-1...10^3] we sample
    * 50 Gamma random variables sample ~ Gamma(sample | alpha, 1), a total of 300
    * points. The ground truth is computed by mpmath. Mean absolute error:
-   * float: 1.0993e-06
-   * double: 1.47631e-12
+   * float: 2.1686e-06
+   * double: 1.4774e-12
    *
    * Reference:
    * M. Figurnov, S. Mohamed, A. Mnih "Implicit Reparameterization Gradients".

From b3fd93207bea08fc459be46f84fd6bbe18b19523 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 14:43:02 +0200
Subject: [PATCH 361/680] Fix typos found using codespell

---
 Eigen/src/Core/Solve.h                           |  2 +-
 Eigen/src/Core/products/Parallelizer.h           |  2 +-
 .../src/IterativeLinearSolvers/SolveWithGuess.h  |  2 +-
 .../src/SparseCholesky/SimplicialCholesky_impl.h |  2 +-
 test/indexed_view.cpp                            |  2 +-
 unsupported/Eigen/CXX11/src/Tensor/README.md     |  2 +-
 unsupported/Eigen/CXX11/src/Tensor/Tensor.h      |  2 +-
 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h  |  2 +-
 .../CXX11/src/Tensor/TensorContractionBlocking.h |  2 +-
 .../src/Tensor/TensorContractionThreadPool.h     |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorCostModel.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorDeviceSycl.h    | 12 ++++++------
 unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h   |  2 +-
 .../CXX11/src/Tensor/TensorForwardDeclarations.h |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h    |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorIntDiv.h        |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorReductionSycl.h |  4 ++--
 unsupported/Eigen/CXX11/src/Tensor/TensorRef.h   |  2 +-
 .../CXX11/src/Tensor/TensorSyclExtractFunctors.h |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorSyclFunctors.h  |  8 ++++----
 .../Eigen/CXX11/src/Tensor/TensorSyclTuple.h     |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorVolumePatch.h   |  2 +-
 .../TensorSymmetry/util/TemplateGroupTheory.h    |  2 +-
 .../Eigen/CXX11/src/ThreadPool/EventCount.h      |  8 ++++----
 .../Eigen/CXX11/src/ThreadPool/RunQueue.h        |  2 +-
 unsupported/Eigen/CXX11/src/util/EmulateArray.h  |  2 +-
 unsupported/Eigen/NonLinearOptimization          |  4 ++--
 unsupported/Eigen/OpenGLSupport                  |  2 +-
 unsupported/Eigen/src/BVH/KdBVH.h                |  2 +-
 .../Eigenvalues/ArpackSelfAdjointEigenSolver.h   |  2 +-
 unsupported/Eigen/src/EulerAngles/EulerSystem.h  |  2 +-
 .../src/IterativeSolvers/ConstrainedConjGrad.h   |  2 +-
 unsupported/Eigen/src/IterativeSolvers/DGMRES.h  |  4 ++--
 .../Eigen/src/LevenbergMarquardt/LMqrsolv.h      |  2 +-
 .../src/LevenbergMarquardt/LevenbergMarquardt.h  |  4 ++--
 .../src/MatrixFunctions/MatrixExponential.h      |  2 +-
 .../Eigen/src/MatrixFunctions/MatrixPower.h      |  2 +-
 .../Eigen/src/NonLinearOptimization/qrsolv.h     |  2 +-
 .../Eigen/src/NonLinearOptimization/r1updt.h     |  2 +-
 unsupported/Eigen/src/Polynomials/Companion.h    |  4 ++--
 unsupported/Eigen/src/Skyline/SkylineInplaceLU.h |  2 +-
 unsupported/Eigen/src/Skyline/SkylineMatrix.h    | 16 ++++++++--------
 .../Eigen/src/SparseExtra/DynamicSparseMatrix.h  |  8 ++++----
 unsupported/Eigen/src/SparseExtra/MarketIO.h     |  2 +-
 unsupported/Eigen/src/Splines/SplineFitting.h    |  2 +-
 unsupported/README.txt                           |  2 +-
 unsupported/bench/bench_svd.cpp                  |  2 +-
 unsupported/test/CMakeLists.txt                  |  2 +-
 unsupported/test/autodiff_scalar.cpp             |  2 +-
 unsupported/test/cxx11_tensor_inflation_sycl.cpp |  4 ++--
 unsupported/test/cxx11_tensor_of_float16_cuda.cu |  2 +-
 unsupported/test/cxx11_tensor_random_cuda.cu     |  2 +-
 unsupported/test/forward_adolc.cpp               |  2 +-
 unsupported/test/sparse_extra.cpp                |  2 +-
 54 files changed, 84 insertions(+), 84 deletions(-)

diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h
index a8daea511..2bf940a26 100644
--- a/Eigen/src/Core/Solve.h
+++ b/Eigen/src/Core/Solve.h
@@ -181,7 +181,7 @@ struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<t
   }
 };
 
-} // end namepsace internal
+} // end namespace internal
 
 } // end namespace Eigen
 
diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h
index e4d13103b..15b5c5f94 100644
--- a/Eigen/src/Core/products/Parallelizer.h
+++ b/Eigen/src/Core/products/Parallelizer.h
@@ -91,7 +91,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth,
   // FIXME the transpose variable is only needed to properly split
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole
-  // parallelizer mechanism has to be redisigned anyway.
+  // parallelizer mechanism has to be redesigned anyway.
   EIGEN_UNUSED_VARIABLE(depth);
   EIGEN_UNUSED_VARIABLE(transpose);
   func(0,rows, 0,cols);
diff --git a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h
index 0ace45177..79e1e4819 100644
--- a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h
+++ b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h
@@ -108,7 +108,7 @@ struct Assignment<DstXprType, SolveWithGuess<DecType,RhsType,GuessType>, interna
   }
 };
 
-} // end namepsace internal
+} // end namespace internal
 
 } // end namespace Eigen
 
diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
index 84a1bf2bd..0aa92f8bc 100644
--- a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
+++ b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
@@ -5,7 +5,7 @@
 
 /*
 
-NOTE: thes functions vave been adapted from the LDL library:
+NOTE: these functions have been adapted from the LDL library:
 
 LDL Copyright (c) 2005 by Timothy A. Davis.  All Rights Reserved.
 
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 033d8833f..2d46ffd6b 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -140,7 +140,7 @@ void check_indexed_view()
     "500  501  502  503  504  505  506  507  508  509")
   );
 
-  // takes the row numer 3, and repeat it 5 times
+  // take row number 3, and repeat it 5 times
   VERIFY( MATCH( A(seqN(3,5,0), all),
     "300  301  302  303  304  305  306  307  308  309\n"
     "300  301  302  303  304  305  306  307  308  309\n"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index 49cc33c5d..dfd7ab7c7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -581,7 +581,7 @@ is not initialized.
 
 Creates a tensor mapping an existing array of data. The data must not be freed
 until the TensorMap is discarded, and the size of the data must be large enough
-to accomodate of the coefficients of the tensor.
+to accommodate the coefficients of the tensor.
 
     float data[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
     Eigen::TensorMap<Tensor<float, 2>> a(data, 3, 4);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
index 1940a9692..4fd96448f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
@@ -48,7 +48,7 @@ namespace Eigen {
   *
   * <dl>
   * <dt><b>Relation to other parts of Eigen:</b></dt>
-  * <dd>The midterm developement goal for this class is to have a similar hierarchy as Eigen uses for matrices, so that
+  * <dd>The midterm development goal for this class is to have a similar hierarchy as Eigen uses for matrices, so that
   * taking blocks or using tensors in expressions is easily possible, including an interface with the vector/matrix code
   * by providing .asMatrix() and .asVector() (or similar) methods for rank 2 and 1 tensors. However, currently, the %Tensor
   * class does not provide any of these features and is only available as a stand-alone class that just allows for
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index a942c98dd..d88e0df71 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -20,7 +20,7 @@ namespace Eigen {
   * \brief The tensor base class.
   *
   * This class is the common parent of the Tensor and TensorMap class, thus
-  * making it possible to use either class interchangably in expressions.
+  * making it possible to use either class interchangeably in expressions.
   */
 
 template<typename Derived>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
index d34f9caee..639d99f9d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -75,7 +75,7 @@ class TensorXsmmContractionBlocking {
       outer_n_ = outer_n_ != 0 ? outer_n_ : n;
     }
 #else
-    // Defaults, possibly overriden per-platform.
+    // Defaults, possibly overridden per-platform.
     copyA_ = true;
     copyB_ = false;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index d30cc96ab..3c007b183 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -350,7 +350,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         // Normal number of notifications for k slice switch is
         // nm_ + nn_ + nm_ * nn_. However, first P - 1 slices will receive only
         // nm_ + nn_ notifications, because they will not receive notifications
-        // from preceeding kernels.
+        // from preceding kernels.
         state_switch_[x] =
             x == 0
                 ? 1
@@ -530,7 +530,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
     void kernel(Index m, Index n, Index k) {
       // Note: order of iteration matters here. Iteration over m is innermost
-      // because we want to reuse the same packed rhs in consequetive tasks
+      // because we want to reuse the same packed rhs in consecutive tasks
       // (rhs fits into L2$ while lhs only into L3$).
       const Index nend = n * gn_ + gn(n);
       const Index mend = m * gm_ + gm(m);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
index b148dae39..bb63baee2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
@@ -195,7 +195,7 @@ class TensorCostModel {
     // 11 is L2 cache latency on Haswell.
     // We don't know whether data is in L1, L2 or L3. But we are most interested
     // in single-threaded computational time around 100us-10ms (smaller time
-    // is too small for parallelization, larger time is not intersting
+    // is too small for parallelization, larger time is not interesting
     // either because we are probably using all available threads already).
     // And for the target time range, L2 seems to be what matters. Data set
     // fitting into L1 is too small to take noticeable time. Data set fitting
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 6158acbd9..e7beb2c82 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -286,7 +286,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
     tileSize =static_cast<Index>(m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
     auto s=  m_queue.get_device().template get_info<cl::sycl::info::device::vendor>();
     std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesn't allow to use max workgroup size
       tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
     }
     rng = n;
@@ -303,7 +303,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
     Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesn't allow to use max workgroup size
       max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
     }
     Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
@@ -331,7 +331,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
     Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesn't allow to use max workgroup size
       max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
     }
     Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
@@ -377,7 +377,7 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
 
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
-    // OpenCL doesnot have such concept
+    // OpenCL doesn't have such concept
     return 2;
   }
 
@@ -519,7 +519,7 @@ struct SyclDevice {
     return m_queue_stream->maxSyclThreadsPerBlock();
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
-    // OpenCL doesnot have such concept
+    // OpenCL doesn't have such concept
     return m_queue_stream->maxSyclThreadsPerMultiProcessor();
   //  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
   }
@@ -544,7 +544,7 @@ struct SyclDevice {
 };
 // This is used as a distingushable device inside the kernel as the sycl device class is not Standard layout.
 // This is internal and must not be used by user. This dummy device allow us to specialise the tensor evaluator
-// inside the kenrel. So we can have two types of eval for host and device. This is required for TensorArgMax operation
+// inside the kernel. So we can have two types of eval for host and device. This is required for TensorArgMax operation
 struct SyclKernelDevice:DefaultDevice{};
 
 }  // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
index f81da318c..d6ab4d997 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
@@ -274,7 +274,7 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
           }
         }
 
-        // processs the line
+        // process the line
         if (is_power_of_two) {
           processDataLineCooleyTukey(line_buf, line_len, log_len);
         }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 354bbe8d1..6c237bac3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -12,7 +12,7 @@
 
 namespace Eigen {
 
-// MakePointer class is used as a container of the adress space of the pointer
+// MakePointer class is used as a container of the address space of the pointer
 // on the host and on the device. From the host side it generates the T* pointer
 // and when EIGEN_USE_SYCL is used it construct a buffer with a map_allocator to
 // T* m_data on the host. It is always called on the device.
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 91d4ead28..f0f7c7826 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -272,8 +272,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
           break;
         default:
           eigen_assert(false && "unexpected padding");
-          m_outputCols=0; // silence the uninitialised warnig;
-          m_outputRows=0; //// silence the uninitialised warnig;
+          m_outputCols=0; // silence the uninitialised warning;
+          m_outputRows=0; //// silence the uninitialised warning;
       }
     }
     eigen_assert(m_outputRows > 0);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
index bf5c532ff..25ba2001e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
@@ -167,7 +167,7 @@ struct TensorIntDivisor {
     shift2 = log_div > 1 ? log_div-1 : 0;
   }
 
-  // Must have 0 <= numerator. On platforms that dont support the __uint128_t
+  // Must have 0 <= numerator. On platforms that don't support the __uint128_t
   // type numerator should also be less than 2^32-1.
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(const T numerator) const {
     eigen_assert(static_cast<typename UnsignedTraits<T>::type>(numerator) < NumTraits<UnsignedType>::highest()/2);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index 94899252b..a379f5a94 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -106,7 +106,7 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
     /// if the shared memory is less than the GRange, we set shared_mem size to the TotalSize and in this case one kernel would be created for recursion to reduce all to one.
     if (GRange < outTileSize) outTileSize=GRange;
     /// creating the shared memory for calculating reduction.
-    /// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can
+    /// This one is used to collect all the reduced value of shared memory as we don't have global barrier on GPU. Once it is saved we can
     /// recursively apply reduction on it in order to reduce the whole.
     auto temp_global_buffer =cl::sycl::buffer<CoeffReturnType, 1>(cl::sycl::range<1>(GRange));
     typedef typename Eigen::internal::remove_all<decltype(self.xprDims())>::type Dims;
@@ -150,7 +150,7 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
 
     // getting final out buffer at the moment the created buffer is true because there is no need for assign
     /// creating the shared memory for calculating reduction.
-    /// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can
+    /// This one is used to collect all the reduced value of shared memory as we don't have global barrier on GPU. Once it is saved we can
     /// recursively apply reduction on it in order to reduce the whole.
       dev.parallel_for_setup(num_coeffs_to_preserve, tileSize, range, GRange);
       dev.sycl_queue().submit([&](cl::sycl::handler &cgh) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
index 99245f778..b2b4fd8d3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
@@ -31,7 +31,7 @@ class TensorLazyBaseEvaluator {
   int refCount() const { return m_refcount; }
 
  private:
-  // No copy, no assigment;
+  // No copy, no assignment;
   TensorLazyBaseEvaluator(const TensorLazyBaseEvaluator& other);
   TensorLazyBaseEvaluator& operator = (const TensorLazyBaseEvaluator& other);
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
index a7905706d..a248e303b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -117,7 +117,7 @@ SYCLEXTRFUNCTERNARY()
 
 
 
-//TensorCustomOp must be specialised otherewise it will be captured by UnaryCategory while its action is different
+//TensorCustomOp must be specialised otherwise it will be captured by UnaryCategory while its action is different
 //from the UnaryCategory and it is similar to the general FunctorExtractor.
 /// specialisation of TensorCustomOp
 #define SYCLEXTRFUNCCUSTOMUNARYOP(CVQual)\
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
index e5b892f2e..a447c3f88 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclFunctors.h
@@ -80,7 +80,7 @@ template < typename HostExpr, typename FunctorExpr, typename Tuple_of_Acc, typen
     typedef typename ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
     auto device_expr = createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
     /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour
-    /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the
+    /// the first behaviour is when it is used as a root to launch the sub-kernel. The second one is when it is treated as a leafnode to pass the
     /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here.
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
@@ -121,7 +121,7 @@ class ReductionFunctor<HostExpr, FunctorExpr, Tuple_of_Acc, Dims, Eigen::interna
     typedef typename ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
     auto device_expr = createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
     /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour
-    /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the
+    /// the first behaviour is when it is used as a root to launch the sub-kernel. The second one is when it is treated as a leafnode to pass the
     /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here.
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
@@ -168,7 +168,7 @@ public:
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
     auto device_expr = TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
     /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour
-    /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the
+    /// the first behaviour is when it is used as a root to launch the sub-kernel. The second one is when it is treated as a leafnode to pass the
     /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here.
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, op);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
@@ -215,7 +215,7 @@ public:
     typedef typename TensorSycl::internal::ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
     auto device_expr = TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
     /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour
-    /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the
+    /// the first behaviour is when it is used as a root to launch the sub-kernel. The second one is when it is treated as a leafnode to pass the
     /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here.
     const auto device_self_expr= Eigen::TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, op);
     /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclTuple.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclTuple.h
index 58ab0f0d5..9e6c3e4fa 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclTuple.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclTuple.h
@@ -143,7 +143,7 @@ struct IndexList {};
 /// \brief Collects internal details for generating index ranges [MIN, MAX)
 /// Declare primary template for index range builder
 /// \tparam MIN is the starting index in the tuple
-/// \tparam N represents sizeof..(elemens)- sizeof...(Is)
+/// \tparam N represents sizeof..(elements)- sizeof...(Is)
 /// \tparam Is... are the list of generated index so far
 template <size_t MIN, size_t N, size_t... Is>
 struct RangeBuilder;
@@ -161,7 +161,7 @@ struct RangeBuilder<MIN, MIN, Is...> {
 /// in this case we are recursively subtracting N by one and adding one
 /// index to Is... list until MIN==N
 /// \tparam MIN is the starting index in the tuple
-/// \tparam N represents sizeof..(elemens)- sizeof...(Is)
+/// \tparam N represents sizeof..(elements)- sizeof...(Is)
 /// \tparam Is... are the list of generated index so far
 template <size_t MIN, size_t N, size_t... Is>
 struct RangeBuilder : public RangeBuilder<MIN, N - 1, N - 1, Is...> {};
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index 51c099591..ef199bfb6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -568,7 +568,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
 
   Dimensions m_dimensions;
 
-  // Parameters passed to the costructor.
+  // Parameters passed to the constructor.
   Index m_plane_strides;
   Index m_row_strides;
   Index m_col_strides;
diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
index 0fe0b7c46..04d6d6b23 100644
--- a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
+++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
@@ -241,7 +241,7 @@ struct dimino_first_step_elements
   * multiplying all elements in the given subgroup with the new
   * coset representative. Note that the first element of the
   * subgroup is always the identity element, so the first element of
-  * ther result of this template is going to be the coset
+  * the result of this template is going to be the coset
   * representative itself.
   *
   * Note that this template accepts an additional boolean parameter
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
index 71d55552d..0a7181102 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
@@ -33,10 +33,10 @@ namespace Eigen {
 //   ec.Notify(true);
 //
 // Notify is cheap if there are no waiting threads. Prewait/CommitWait are not
-// cheap, but they are executed only if the preceeding predicate check has
+// cheap, but they are executed only if the preceding predicate check has
 // failed.
 //
-// Algorihtm outline:
+// Algorithm outline:
 // There are two main variables: predicate (managed by user) and state_.
 // Operation closely resembles Dekker mutual algorithm:
 // https://en.wikipedia.org/wiki/Dekker%27s_algorithm
@@ -79,7 +79,7 @@ class EventCount {
     uint64_t state = state_.load(std::memory_order_seq_cst);
     for (;;) {
       if (int64_t((state & kEpochMask) - epoch) < 0) {
-        // The preceeding waiter has not decided on its fate. Wait until it
+        // The preceding waiter has not decided on its fate. Wait until it
         // calls either CancelWait or CommitWait, or is notified.
         EIGEN_THREAD_YIELD();
         state = state_.load(std::memory_order_seq_cst);
@@ -110,7 +110,7 @@ class EventCount {
     uint64_t state = state_.load(std::memory_order_relaxed);
     for (;;) {
       if (int64_t((state & kEpochMask) - epoch) < 0) {
-        // The preceeding waiter has not decided on its fate. Wait until it
+        // The preceding waiter has not decided on its fate. Wait until it
         // calls either CancelWait or CommitWait, or is notified.
         EIGEN_THREAD_YIELD();
         state = state_.load(std::memory_order_relaxed);
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h b/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
index 49d0cdc36..cb3690a2e 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
@@ -198,7 +198,7 @@ class RunQueue {
   };
   std::mutex mutex_;
   // Low log(kSize) + 1 bits in front_ and back_ contain rolling index of
-  // front/back, repsectively. The remaining bits contain modification counters
+  // front/back, respectively. The remaining bits contain modification counters
   // that are incremented on Push operations. This allows us to (1) distinguish
   // between empty and full conditions (if we would use log(kSize) bits for
   // position, these conditions would be indistinguishable); (2) obtain
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 96b3a8261..ddd54f4b3 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -219,7 +219,7 @@ template<class T, std::size_t N> struct array_size<const array<T,N>& > {
 
 #else
 
-// The compiler supports c++11, and we're not targetting cuda: use std::array as Eigen::array
+// The compiler supports c++11, and we're not targeting cuda: use std::array as Eigen::array
 #include <array>
 namespace Eigen {
 
diff --git a/unsupported/Eigen/NonLinearOptimization b/unsupported/Eigen/NonLinearOptimization
index 600ab4c12..c22b89054 100644
--- a/unsupported/Eigen/NonLinearOptimization
+++ b/unsupported/Eigen/NonLinearOptimization
@@ -35,7 +35,7 @@
   * a zero for the system (Powell hybrid "dogleg" method).
   *
   * This code is a port of minpack (http://en.wikipedia.org/wiki/MINPACK).
-  * Minpack is a very famous, old, robust and well-reknown package, written in 
+  * Minpack is a very famous, old, robust and well renowned package, written in
   * fortran. Those implementations have been carefully tuned, tested, and used
   * for several decades.
   *
@@ -63,7 +63,7 @@
   * Other tests were added by myself at the very beginning of the 
   * process and check the results for levenberg-marquardt using the reference data 
   * on http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml. Since then i've 
-  * carefully checked that the same results were obtained when modifiying the 
+  * carefully checked that the same results were obtained when modifying the
   * code. Please note that we do not always get the exact same decimals as they do,
   * but this is ok : they use 128bits float, and we do the tests using the C type 'double',
   * which is 64 bits on most platforms (x86 and amd64, at least).
diff --git a/unsupported/Eigen/OpenGLSupport b/unsupported/Eigen/OpenGLSupport
index 87f50947d..11d99567e 100644
--- a/unsupported/Eigen/OpenGLSupport
+++ b/unsupported/Eigen/OpenGLSupport
@@ -25,7 +25,7 @@ namespace Eigen {
   *
   * This module provides wrapper functions for a couple of OpenGL functions
   * which simplify the way to pass Eigen's object to openGL.
-  * Here is an exmaple:
+  * Here is an example:
   * 
   * \code
   * // You need to add path_to_eigen/unsupported to your include path.
diff --git a/unsupported/Eigen/src/BVH/KdBVH.h b/unsupported/Eigen/src/BVH/KdBVH.h
index 1b8d75865..13f792cd0 100644
--- a/unsupported/Eigen/src/BVH/KdBVH.h
+++ b/unsupported/Eigen/src/BVH/KdBVH.h
@@ -170,7 +170,7 @@ private:
   typedef internal::vector_int_pair<Scalar, Dim> VIPair;
   typedef std::vector<VIPair, aligned_allocator<VIPair> > VIPairList;
   typedef Matrix<Scalar, Dim, 1> VectorType;
-  struct VectorComparator //compares vectors, or, more specificall, VIPairs along a particular dimension
+  struct VectorComparator //compares vectors, or more specifically, VIPairs along a particular dimension
   {
     VectorComparator(int inDim) : dim(inDim) {}
     inline bool operator()(const VIPair &v1, const VIPair &v2) const { return v1.first[dim] < v2.first[dim]; }
diff --git a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
index 866a8a460..9f7bff764 100644
--- a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
+++ b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
@@ -300,7 +300,7 @@ public:
 
   /** \brief Reports whether previous computation was successful.
    *
-   * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+   * \returns \c Success if computation was successful, \c NoConvergence otherwise.
    */
   ComputationInfo info() const
   {
diff --git a/unsupported/Eigen/src/EulerAngles/EulerSystem.h b/unsupported/Eigen/src/EulerAngles/EulerSystem.h
index 28f52da61..65c2e94c7 100644
--- a/unsupported/Eigen/src/EulerAngles/EulerSystem.h
+++ b/unsupported/Eigen/src/EulerAngles/EulerSystem.h
@@ -12,7 +12,7 @@
 
 namespace Eigen
 {
-  // Forward declerations
+  // Forward declarations
   template <typename _Scalar, class _System>
   class EulerAngles;
   
diff --git a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
index dc0093eb9..37d5b4c6c 100644
--- a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
+++ b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
@@ -99,7 +99,7 @@ void pseudo_inverse(const CMatrix &C, CINVMatrix &CINV)
 /** \ingroup IterativeSolvers_Module
   * Constrained conjugate gradient
   *
-  * Computes the minimum of \f$ 1/2((Ax).x) - bx \f$ under the contraint \f$ Cx \le f \f$
+  * Computes the minimum of \f$ 1/2((Ax).x) - bx \f$ under the constraint \f$ Cx \le f \f$
   */
 template<typename TMatrix, typename CMatrix,
          typename VectorX, typename VectorB, typename VectorF>
diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index d603ba336..5cfe49302 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -214,7 +214,7 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
     void dgmresInitDeflation(Index& rows) const; 
     mutable DenseMatrix m_V; // Krylov basis vectors
     mutable DenseMatrix m_H; // Hessenberg matrix 
-    mutable DenseMatrix m_Hes; // Initial hessenberg matrix wihout Givens rotations applied
+    mutable DenseMatrix m_Hes; // Initial hessenberg matrix without Givens rotations applied
     mutable Index m_restart; // Maximum size of the Krylov subspace
     mutable DenseMatrix m_U; // Vectors that form the basis of the invariant subspace 
     mutable DenseMatrix m_MU; // matrix operator applied to m_U (for next cycles)
@@ -250,7 +250,7 @@ void DGMRES<_MatrixType, _Preconditioner>::dgmres(const MatrixType& mat,const Rh
   m_H.resize(m_restart+1, m_restart);
   m_Hes.resize(m_restart, m_restart);
   m_V.resize(n,m_restart+1);
-  //Initial residual vector and intial norm
+  //Initial residual vector and initial norm
   x = precond.solve(x);
   r0 = rhs - mat * x; 
   RealScalar beta = r0.norm(); 
diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h
index ae9d793b1..123485817 100644
--- a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h
+++ b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h
@@ -73,7 +73,7 @@ void lmqrsolv(
             qtbpj = -givens.s() * wa[k] + givens.c() * qtbpj;
             wa[k] = temp;
 
-            /*           accumulate the tranformation in the row of s. */
+            /*           accumulate the transformation in the row of s. */
             for (i = k+1; i<n; ++i) {
                 temp = givens.c() * s(i,k) + givens.s() * sdiag[i];
                 sdiag[i] = -givens.s() * s(i,k) + givens.c() * sdiag[i];
diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
index 995427978..5f64501be 100644
--- a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
+++ b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h
@@ -233,9 +233,9 @@ class LevenbergMarquardt : internal::no_assignment_operator
     
     /** 
      * \brief Reports whether the minimization was successful
-     * \returns \c Success if the minimization was succesful,
+     * \returns \c Success if the minimization was successful,
      *         \c NumericalIssue if a numerical problem arises during the 
-     *          minimization process, for exemple during the QR factorization
+     *          minimization process, for example during the QR factorization
      *         \c NoConvergence if the minimization did not converge after 
      *          the maximum number of function evaluation allowed
      *          \c InvalidInput if the input matrix is invalid
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
index 85ab3d97c..03356998b 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
@@ -313,7 +313,7 @@ struct matrix_exp_computeUV<MatrixType, long double>
       matrix_exp_pade17(A, U, V);
     }
   
-#elif LDBL_MANT_DIG <= 112  // quadruple precison
+#elif LDBL_MANT_DIG <= 112  // quadruple precision
   
     if (l1norm < 1.639394610288918690547467954466970e-005L) {
       matrix_exp_pade3(arg, U, V);
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
index ebc433d89..33609aea9 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
@@ -81,7 +81,7 @@ class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParen
  *
  * \note Currently this class is only used by MatrixPower. One may
  * insist that this be nested into MatrixPower. This class is here to
- * faciliate future development of triangular matrix functions.
+ * facilitate future development of triangular matrix functions.
  */
 template<typename MatrixType>
 class MatrixPowerAtomic : internal::noncopyable
diff --git a/unsupported/Eigen/src/NonLinearOptimization/qrsolv.h b/unsupported/Eigen/src/NonLinearOptimization/qrsolv.h
index feafd62a8..4f2f560b3 100644
--- a/unsupported/Eigen/src/NonLinearOptimization/qrsolv.h
+++ b/unsupported/Eigen/src/NonLinearOptimization/qrsolv.h
@@ -61,7 +61,7 @@ void qrsolv(
             qtbpj = -givens.s() * wa[k] + givens.c() * qtbpj;
             wa[k] = temp;
 
-            /*           accumulate the tranformation in the row of s. */
+            /*           accumulate the transformation in the row of s. */
             for (i = k+1; i<n; ++i) {
                 temp = givens.c() * s(i,k) + givens.s() * sdiag[i];
                 sdiag[i] = -givens.s() * s(i,k) + givens.c() * sdiag[i];
diff --git a/unsupported/Eigen/src/NonLinearOptimization/r1updt.h b/unsupported/Eigen/src/NonLinearOptimization/r1updt.h
index f28766061..09fc65255 100644
--- a/unsupported/Eigen/src/NonLinearOptimization/r1updt.h
+++ b/unsupported/Eigen/src/NonLinearOptimization/r1updt.h
@@ -22,7 +22,7 @@ void r1updt(
     Scalar temp;
     JacobiRotation<Scalar> givens;
 
-    // r1updt had a broader usecase, but we dont use it here. And, more
+    // r1updt had a broader usecase, but we don't use it here. And, more
     // importantly, we can not test it.
     eigen_assert(m==n);
     eigen_assert(u.size()==m);
diff --git a/unsupported/Eigen/src/Polynomials/Companion.h b/unsupported/Eigen/src/Polynomials/Companion.h
index e0af6ebe4..41a4efc2f 100644
--- a/unsupported/Eigen/src/Polynomials/Companion.h
+++ b/unsupported/Eigen/src/Polynomials/Companion.h
@@ -104,7 +104,7 @@ class companion
     /** Helper function for the balancing algorithm.
      * \returns true if the row and the column, having colNorm and rowNorm
      * as norms, are balanced, false otherwise.
-     * colB and rowB are repectively the multipliers for
+     * colB and rowB are respectively the multipliers for
      * the column and the row in order to balance them.
      * */
     bool balanced( RealScalar colNorm, RealScalar rowNorm,
@@ -113,7 +113,7 @@ class companion
     /** Helper function for the balancing algorithm.
      * \returns true if the row and the column, having colNorm and rowNorm
      * as norms, are balanced, false otherwise.
-     * colB and rowB are repectively the multipliers for
+     * colB and rowB are respectively the multipliers for
      * the column and the row in order to balance them.
      * */
     bool balancedR( RealScalar colNorm, RealScalar rowNorm,
diff --git a/unsupported/Eigen/src/Skyline/SkylineInplaceLU.h b/unsupported/Eigen/src/Skyline/SkylineInplaceLU.h
index a1f54ed35..bda057a85 100644
--- a/unsupported/Eigen/src/Skyline/SkylineInplaceLU.h
+++ b/unsupported/Eigen/src/Skyline/SkylineInplaceLU.h
@@ -41,7 +41,7 @@ public:
 
     /** Sets the relative threshold value used to prune zero coefficients during the decomposition.
      *
-     * Setting a value greater than zero speeds up computation, and yields to an imcomplete
+     * Setting a value greater than zero speeds up computation, and yields to an incomplete
      * factorization with fewer non zero coefficients. Such approximate factors are especially
      * useful to initialize an iterative solver.
      *
diff --git a/unsupported/Eigen/src/Skyline/SkylineMatrix.h b/unsupported/Eigen/src/Skyline/SkylineMatrix.h
index a2a8933ca..f77d79a04 100644
--- a/unsupported/Eigen/src/Skyline/SkylineMatrix.h
+++ b/unsupported/Eigen/src/Skyline/SkylineMatrix.h
@@ -206,26 +206,26 @@ public:
             if (col > row) //upper matrix
             {
                 const Index minOuterIndex = inner - m_data.upperProfile(inner);
-                eigen_assert(outer >= minOuterIndex && "you try to acces a coeff that do not exist in the storage");
+                eigen_assert(outer >= minOuterIndex && "You tried to access a coeff that does not exist in the storage");
                 return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
             }
             if (col < row) //lower matrix
             {
                 const Index minInnerIndex = outer - m_data.lowerProfile(outer);
-                eigen_assert(inner >= minInnerIndex && "you try to acces a coeff that do not exist in the storage");
+                eigen_assert(inner >= minInnerIndex && "You tried to access a coeff that does not exist in the storage");
                 return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
             }
         } else {
             if (outer > inner) //upper matrix
             {
                 const Index maxOuterIndex = inner + m_data.upperProfile(inner);
-                eigen_assert(outer <= maxOuterIndex && "you try to acces a coeff that do not exist in the storage");
+                eigen_assert(outer <= maxOuterIndex && "You tried to access a coeff that does not exist in the storage");
                 return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
             }
             if (outer < inner) //lower matrix
             {
                 const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
-                eigen_assert(inner <= maxInnerIndex && "you try to acces a coeff that do not exist in the storage");
+                eigen_assert(inner <= maxInnerIndex && "You tried to access a coeff that does not exist in the storage");
                 return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
             }
         }
@@ -300,11 +300,11 @@ public:
 
         if (IsRowMajor) {
             const Index minInnerIndex = outer - m_data.lowerProfile(outer);
-            eigen_assert(inner >= minInnerIndex && "you try to acces a coeff that do not exist in the storage");
+            eigen_assert(inner >= minInnerIndex && "You tried to access a coeff that does not exist in the storage");
             return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
         } else {
             const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
-            eigen_assert(inner <= maxInnerIndex && "you try to acces a coeff that do not exist in the storage");
+            eigen_assert(inner <= maxInnerIndex && "You tried to access a coeff that does not exist in the storage");
             return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
         }
     }
@@ -336,11 +336,11 @@ public:
 
         if (IsRowMajor) {
             const Index minOuterIndex = inner - m_data.upperProfile(inner);
-            eigen_assert(outer >= minOuterIndex && "you try to acces a coeff that do not exist in the storage");
+            eigen_assert(outer >= minOuterIndex && "You tried to access a coeff that does not exist in the storage");
             return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
         } else {
             const Index maxOuterIndex = inner + m_data.upperProfile(inner);
-            eigen_assert(outer <= maxOuterIndex && "you try to acces a coeff that do not exist in the storage");
+            eigen_assert(outer <= maxOuterIndex && "You tried to access a coeff that does not exist in the storage");
             return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
         }
     }
diff --git a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h
index 037a13f86..3f1ff14ad 100644
--- a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h
+++ b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h
@@ -187,7 +187,7 @@ template<typename _Scalar, int _Options, typename _StorageIndex>
     /** Does nothing: provided for compatibility with SparseMatrix */
     inline void finalize() {}
 
-    /** Suppress all nonzeros which are smaller than \a reference under the tolerence \a epsilon */
+    /** Suppress all nonzeros which are smaller than \a reference under the tolerance \a epsilon */
     void prune(Scalar reference, RealScalar epsilon = NumTraits<RealScalar>::dummy_precision())
     {
       for (Index j=0; j<outerSize(); ++j)
@@ -224,21 +224,21 @@ template<typename _Scalar, int _Options, typename _StorageIndex>
       }
     }
 
-    /** The class DynamicSparseMatrix is deprectaed */
+    /** The class DynamicSparseMatrix is deprecated */
     EIGEN_DEPRECATED inline DynamicSparseMatrix()
       : m_innerSize(0), m_data(0)
     {
       eigen_assert(innerSize()==0 && outerSize()==0);
     }
 
-    /** The class DynamicSparseMatrix is deprectaed */
+    /** The class DynamicSparseMatrix is deprecated */
     EIGEN_DEPRECATED inline DynamicSparseMatrix(Index rows, Index cols)
       : m_innerSize(0)
     {
       resize(rows, cols);
     }
 
-    /** The class DynamicSparseMatrix is deprectaed */
+    /** The class DynamicSparseMatrix is deprecated */
     template<typename OtherDerived>
     EIGEN_DEPRECATED explicit inline DynamicSparseMatrix(const SparseMatrixBase<OtherDerived>& other)
       : m_innerSize(0)
diff --git a/unsupported/Eigen/src/SparseExtra/MarketIO.h b/unsupported/Eigen/src/SparseExtra/MarketIO.h
index 6d57ab2e9..1618b09a8 100644
--- a/unsupported/Eigen/src/SparseExtra/MarketIO.h
+++ b/unsupported/Eigen/src/SparseExtra/MarketIO.h
@@ -104,7 +104,7 @@ namespace internal
     out << value.real << " " << value.imag()<< "\n"; 
   }
 
-} // end namepsace internal
+} // end namespace internal
 
 inline bool getMarketHeader(const std::string& filename, int& sym, bool& iscomplex, bool& isvector)
 {
diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h
index c761a9b3d..1a4b80a2f 100644
--- a/unsupported/Eigen/src/Splines/SplineFitting.h
+++ b/unsupported/Eigen/src/Splines/SplineFitting.h
@@ -181,7 +181,7 @@ namespace Eigen
    * \ingroup Splines_Module
    *
    * \param[in] pts The data points to which a spline should be fit.
-   * \param[out] chord_lengths The resulting chord lenggth vector.
+   * \param[out] chord_lengths The resulting chord length vector.
    *
    * \sa Les Piegl and Wayne Tiller, The NURBS book (2nd ed.), 1997, 9.2.1 Global Curve Interpolation to Point Data
    **/   
diff --git a/unsupported/README.txt b/unsupported/README.txt
index 83479ff0b..70793bf13 100644
--- a/unsupported/README.txt
+++ b/unsupported/README.txt
@@ -20,7 +20,7 @@ However, it:
  - must rely on Eigen,
  - must be highly related to math,
  - should have some general purpose in the sense that it could
-   potentially become an offical Eigen module (or be merged into another one).
+   potentially become an official Eigen module (or be merged into another one).
 
 In doubt feel free to contact us. For instance, if your addons is very too specific
 but it shows an interesting way of using Eigen, then it could be a nice demo.
diff --git a/unsupported/bench/bench_svd.cpp b/unsupported/bench/bench_svd.cpp
index 01d8231ae..e7028a2b9 100644
--- a/unsupported/bench/bench_svd.cpp
+++ b/unsupported/bench/bench_svd.cpp
@@ -70,7 +70,7 @@ void bench_svd(const MatrixType& a = MatrixType())
   std::cout<< std::endl;
   timerJacobi.reset();
   timerBDC.reset();
-  cout << " Computes rotaion matrix" <<endl;
+  cout << " Computes rotation matrix" <<endl;
   for (int k=1; k<=NUMBER_SAMPLE; ++k)
   {
     timerBDC.start();
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 42b790d55..e99eab0e3 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -1,5 +1,5 @@
 # generate split test header file only if it does not yet exist
-# in order to prevent a rebuild everytime cmake is configured
+# in order to prevent a rebuild every time cmake is configured
 if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
   file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
   foreach(i RANGE 1 999)
diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp
index a917ec344..1d4418203 100644
--- a/unsupported/test/autodiff_scalar.cpp
+++ b/unsupported/test/autodiff_scalar.cpp
@@ -81,7 +81,7 @@ void check_limits_specialization()
   typedef std::numeric_limits<AD> A;
   typedef std::numeric_limits<Scalar> B;
 
-  // workaround "unsed typedef" warning:
+  // workaround "unused typedef" warning:
   VERIFY(!bool(internal::is_same<B, A>::value));
 
 #if EIGEN_HAS_CXX11
diff --git a/unsupported/test/cxx11_tensor_inflation_sycl.cpp b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
index f2f87f7ed..cf3e29f4c 100644
--- a/unsupported/test/cxx11_tensor_inflation_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_inflation_sycl.cpp
@@ -22,10 +22,10 @@
 
 using Eigen::Tensor;
 
-// Inflation Defenition for each dimention the inflated val would be
+// Inflation Definition for each dimension the inflated val would be
 //((dim-1)*strid[dim] +1)
 
-// for 1 dimnention vector of size 3 with value (4,4,4) with the inflated stride value of 3 would be changed to
+// for 1 dimension vector of size 3 with value (4,4,4) with the inflated stride value of 3 would be changed to
 // tensor of size (2*3) +1 = 7 with the value of
 // (4, 0, 0, 4, 0, 0, 4).
 
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index 167b75d25..7a751ff02 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -247,7 +247,7 @@ void test_cuda_trancendental() {
   }
   for (int i = 0; i < num_elem; ++i) {
     std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
-    if(std::abs(input2(i)-1.f)<0.05f) // log lacks accurary nearby 1
+    if(std::abs(input2(i)-1.f)<0.05f) // log lacks accuracy nearby 1
       VERIFY_IS_APPROX(full_prec2(i)+Eigen::half(0.1f), half_prec2(i)+Eigen::half(0.1f));
     else
       VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu
index fa1a46732..389c0a8c2 100644
--- a/unsupported/test/cxx11_tensor_random_cuda.cu
+++ b/unsupported/test/cxx11_tensor_random_cuda.cu
@@ -37,7 +37,7 @@ void test_cuda_random_uniform()
   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
   assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
 
-  // For now we just check thes code doesn't crash.
+  // For now we just check this code doesn't crash.
   // TODO: come up with a valid test of randomness
 }
 
diff --git a/unsupported/test/forward_adolc.cpp b/unsupported/test/forward_adolc.cpp
index 866db8e86..6d0ae738d 100644
--- a/unsupported/test/forward_adolc.cpp
+++ b/unsupported/test/forward_adolc.cpp
@@ -132,7 +132,7 @@ void test_forward_adolc()
   }
 
   {
-    // simple instanciation tests
+    // simple instantiation tests
     Matrix<adtl::adouble,2,1> x;
     foo(x);
     Matrix<adtl::adouble,Dynamic,Dynamic> A(4,4);;
diff --git a/unsupported/test/sparse_extra.cpp b/unsupported/test/sparse_extra.cpp
index 4f6723d6d..7cf4a77c3 100644
--- a/unsupported/test/sparse_extra.cpp
+++ b/unsupported/test/sparse_extra.cpp
@@ -8,7 +8,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 
-// import basic and product tests for deprectaed DynamicSparseMatrix
+// import basic and product tests for deprecated DynamicSparseMatrix
 #define EIGEN_NO_DEPRECATED_WARNING
 #include "sparse_basic.cpp"
 #include "sparse_product.cpp"

From e5f9f4768f9118dcafa5d254cfce7696e36f3a16 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 7 Jun 2018 15:03:50 +0200
Subject: [PATCH 362/680] Avoid unnecessary C++11 dependency

---
 Eigen/src/Core/MathFunctions.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 05462c5e1..954863c39 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -1282,17 +1282,17 @@ double exp(const double &x) { return ::exp(x); }
 
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 std::complex<float> exp(const std::complex<float>& x) {
-  auto com = ::expf(x.real());
-  auto res_real = com * ::cosf(x.imag());
-  auto res_imag = com * ::sinf(x.imag());
+  float com = ::expf(x.real());
+  float res_real = com * ::cosf(x.imag());
+  float res_imag = com * ::sinf(x.imag());
   return std::complex<float>(res_real, res_imag);
 }
 
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 std::complex<double> exp(const std::complex<double>& x) {
-  auto com = ::exp(x.real());
-  auto res_real = com * ::cos(x.imag());
-  auto res_imag = com * ::sin(x.imag());
+  double com = ::exp(x.real());
+  double res_real = com * ::cos(x.imag());
+  double res_imag = com * ::sin(x.imag());
   return std::complex<double>(res_real, res_imag);
 }
 #endif

From 55774b48e4718763214ed9e52faa52dc2622aef0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 15:26:25 +0200
Subject: [PATCH 363/680] Fix short vs long

---
 test/main.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/main.h b/test/main.h
index 0fcd6cb76..a94e1ab41 100644
--- a/test/main.h
+++ b/test/main.h
@@ -359,7 +359,7 @@ template<> inline long double test_precision<std::complex<long double> >() { ret
 inline bool test_isApprox(const short& a, const short& b)
 { return internal::isApprox(a, b, test_precision<short>()); }
 inline bool test_isApprox(const unsigned short& a, const unsigned short& b)
-{ return internal::isApprox(a, b, test_precision<unsigned long>()); }
+{ return internal::isApprox(a, b, test_precision<unsigned short>()); }
 inline bool test_isApprox(const unsigned int& a, const unsigned int& b)
 { return internal::isApprox(a, b, test_precision<unsigned int>()); }
 inline bool test_isApprox(const long& a, const long& b)

From aa813d417bf89910d9f6944357314fa3a1280e56 Mon Sep 17 00:00:00 2001
From: Michael Figurnov <mfigurnov@google.com>
Date: Thu, 7 Jun 2018 14:35:07 +0100
Subject: [PATCH 364/680] Fix compilation of special functions without C99
 math.

The commit with Bessel functions i0e and i1e placed the ifdef/endif incorrectly,
causing i0e/i1e to be undefined when EIGEN_HAS_C99_MATH=0. These functions do not
actually require C99 math, so now they are always available.
---
 unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h | 4 ++--
 unsupported/test/special_functions.cpp                        | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 6c7ac3f3b..293b0597b 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -1574,6 +1574,8 @@ struct betainc_impl<double> {
   }
 };
 
+#endif  // EIGEN_HAS_C99_MATH
+
 /****************************************************************************
  * Implementation of Bessel function, based on Cephes                       *
  ****************************************************************************/
@@ -1902,8 +1904,6 @@ struct i1e_impl<double> {
   }
 };
 
-#endif  // EIGEN_HAS_C99_MATH
-
 }  // end namespace internal
 
 namespace numext {
diff --git a/unsupported/test/special_functions.cpp b/unsupported/test/special_functions.cpp
index 48d0db95e..0729dd4dc 100644
--- a/unsupported/test/special_functions.cpp
+++ b/unsupported/test/special_functions.cpp
@@ -335,6 +335,7 @@ template<typename ArrayType> void array_special_functions()
         ArrayType test = betainc(a, b + one, x) + eps;
         verify_component_wise(test, expected););
   }
+#endif  // EIGEN_HAS_C99_MATH
 
   // Test Bessel function i0e. Reference results obtained with SciPy.
   {
@@ -375,7 +376,6 @@ template<typename ArrayType> void array_special_functions()
     CALL_SUBTEST(res = i1e(x);
                  verify_component_wise(res, expected););
   }
-#endif
 }
 
 void test_special_functions()

From 7fe29aceeb9f7a7ec8bd6f9fa81b88d50e4819b6 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 15:36:20 +0200
Subject: [PATCH 365/680] Fix MSVC warning C4290: C++ exception specification
 ignored except to indicate a function is not __declspec(nothrow)

---
 Eigen/src/Core/util/Macros.h | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 729fb3b1d..8927bd404 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1030,7 +1030,13 @@ namespace Eigen {
 #   define EIGEN_NOEXCEPT
 #   define EIGEN_NOEXCEPT_IF(x)
 #   define EIGEN_NO_THROW throw()
-#   define EIGEN_EXCEPTION_SPEC(X) throw(X)
+#   if EIGEN_COMP_MSVC
+      // MSVC does not support exception specifications (warning C4290),
+      // and they are deprecated in c++11 anyway.
+#     define EIGEN_EXCEPTION_SPEC(X) throw()
+#   else
+#     define EIGEN_EXCEPTION_SPEC(X) throw(X)
+#   endif
 #endif
 
 #endif // EIGEN_MACROS_H

From af7c83b9a20b1903ae5524c5f1be9d4093026d86 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 15:45:24 +0200
Subject: [PATCH 366/680] Fix warning

---
 Eigen/src/Core/SelfCwiseBinaryOp.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h
index 50099df82..7c89c2e23 100644
--- a/Eigen/src/Core/SelfCwiseBinaryOp.h
+++ b/Eigen/src/Core/SelfCwiseBinaryOp.h
@@ -17,7 +17,6 @@ namespace Eigen {
 template<typename Derived>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 {
-  typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar,Scalar>());
   return derived();
 }
@@ -25,7 +24,6 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(co
 template<typename Derived>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
 {
-  typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar,Scalar>());
   return derived();
 }
@@ -33,7 +31,6 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(co
 template<typename Derived>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
 {
-  typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar,Scalar>());
   return derived();
 }
@@ -41,7 +38,6 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(co
 template<typename Derived>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
-  typedef typename Derived::PlainObject PlainObject;
   internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar,Scalar>());
   return derived();
 }

From c723ffd7635ea2067a514272c46e3c29184dc4eb Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 15:56:20 +0200
Subject: [PATCH 367/680] Fix warning

---
 doc/examples/matrixfree_cg.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/doc/examples/matrixfree_cg.cpp b/doc/examples/matrixfree_cg.cpp
index 6a205aea3..74699381c 100644
--- a/doc/examples/matrixfree_cg.cpp
+++ b/doc/examples/matrixfree_cg.cpp
@@ -67,6 +67,7 @@ namespace internal {
       // This method should implement "dst += alpha * lhs * rhs" inplace,
       // however, for iterative solvers, alpha is always equal to 1, so let's not bother about it.
       assert(alpha==Scalar(1) && "scaling is not implemented");
+      EIGEN_ONLY_USED_FOR_DEBUG(alpha);
 
       // Here we could simply call dst.noalias() += lhs.my_matrix() * rhs,
       // but let's do something fancier (and less efficient):

From 37348d03aee064e66c5969fb3f08080397d6a9c2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 15:56:43 +0200
Subject: [PATCH 368/680] Fix int versus Index

---
 .../Eigen/src/IterativeSolvers/DGMRES.h       | 60 +++++++++----------
 1 file changed, 29 insertions(+), 31 deletions(-)

diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index 5cfe49302..f40b80eda 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -39,7 +39,6 @@ template <typename VectorType, typename IndexType>
 void sortWithPermutation (VectorType& vec, IndexType& perm, typename IndexType::Scalar& ncut)
 {
   eigen_assert(vec.size() == perm.size());
-  typedef typename IndexType::Scalar Index; 
   bool flag; 
   for (Index k  = 0; k < ncut; k++)
   {
@@ -112,7 +111,6 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
     using Base::_solve_impl;
     typedef _MatrixType MatrixType;
     typedef typename MatrixType::Scalar Scalar;
-    typedef typename MatrixType::Index Index;
     typedef typename MatrixType::StorageIndex StorageIndex;
     typedef typename MatrixType::RealScalar RealScalar;
     typedef _Preconditioner Preconditioner;
@@ -146,7 +144,7 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
   void _solve_with_guess_impl(const Rhs& b, Dest& x) const
   {    
     bool failed = false;
-    for(int j=0; j<b.cols(); ++j)
+    for(Index j=0; j<b.cols(); ++j)
     {
       m_iterations = Base::maxIterations();
       m_error = Base::m_tolerance;
@@ -170,17 +168,17 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
   /** 
    * Get the restart value
     */
-  int restart() { return m_restart; }
+  Index restart() { return m_restart; }
   
   /** 
    * Set the restart value (default is 30)  
    */
-  void set_restart(const int restart) { m_restart=restart; }
+  Index set_restart(const Index restart) { m_restart=restart; }
   
   /** 
    * Set the number of eigenvalues to deflate at each restart 
    */
-  void setEigenv(const int neig) 
+  void setEigenv(const Index neig) 
   {
     m_neig = neig;
     if (neig+1 > m_maxNeig) m_maxNeig = neig+1; // To allow for complex conjugates
@@ -189,12 +187,12 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
   /** 
    * Get the size of the deflation subspace size
    */ 
-  int deflSize() {return m_r; }
+  Index deflSize() {return m_r; }
   
   /**
    * Set the maximum size of the deflation subspace
    */
-  void setMaxEigenv(const int maxNeig) { m_maxNeig = maxNeig; }
+  void setMaxEigenv(const Index maxNeig) { m_maxNeig = maxNeig; }
   
   protected:
     // DGMRES algorithm 
@@ -202,12 +200,12 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
     void dgmres(const MatrixType& mat,const Rhs& rhs, Dest& x, const Preconditioner& precond) const;
     // Perform one cycle of GMRES
     template<typename Dest>
-    int dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, int& nbIts) const; 
+    Index dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, Index& nbIts) const; 
     // Compute data to use for deflation 
-    int dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const;
+    Index dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const;
     // Apply deflation to a vector
     template<typename RhsType, typename DestType>
-    int dgmresApplyDeflation(const RhsType& In, DestType& Out) const; 
+    Index dgmresApplyDeflation(const RhsType& In, DestType& Out) const; 
     ComplexVector schurValues(const ComplexSchur<DenseMatrix>& schurofH) const;
     ComplexVector schurValues(const RealSchur<DenseMatrix>& schurofH) const;
     // Init data for deflation
@@ -221,8 +219,8 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
     mutable DenseMatrix m_T; /* T=U^T*M^{-1}*A*U */
     mutable PartialPivLU<DenseMatrix> m_luT; // LU factorization of m_T
     mutable StorageIndex m_neig; //Number of eigenvalues to extract at each restart
-    mutable int m_r; // Current number of deflated eigenvalues, size of m_U
-    mutable int m_maxNeig; // Maximum number of eigenvalues to deflate
+    mutable Index m_r; // Current number of deflated eigenvalues, size of m_U
+    mutable Index m_maxNeig; // Maximum number of eigenvalues to deflate
     mutable RealScalar m_lambdaN; //Modulus of the largest eigenvalue of A
     mutable bool m_isDeflAllocated;
     mutable bool m_isDeflInitialized;
@@ -244,9 +242,9 @@ void DGMRES<_MatrixType, _Preconditioner>::dgmres(const MatrixType& mat,const Rh
               const Preconditioner& precond) const
 {
   //Initialization
-  int n = mat.rows(); 
+  Index n = mat.rows(); 
   DenseVector r0(n); 
-  int nbIts = 0; 
+  Index nbIts = 0; 
   m_H.resize(m_restart+1, m_restart);
   m_Hes.resize(m_restart, m_restart);
   m_V.resize(n,m_restart+1);
@@ -284,7 +282,7 @@ void DGMRES<_MatrixType, _Preconditioner>::dgmres(const MatrixType& mat,const Rh
  */
 template< typename _MatrixType, typename _Preconditioner>
 template<typename Dest>
-int DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, int& nbIts) const
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, Index& nbIts) const
 {
   //Initialization 
   DenseVector g(m_restart+1); // Right hand side of the least square problem
@@ -293,8 +291,8 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, con
   m_V.col(0) = r0/beta; 
   m_info = NoConvergence; 
   std::vector<JacobiRotation<Scalar> >gr(m_restart); // Givens rotations
-  int it = 0; // Number of inner iterations 
-  int n = mat.rows();
+  Index it = 0; // Number of inner iterations 
+  Index n = mat.rows();
   DenseVector tv1(n), tv2(n);  //Temporary vectors
   while (m_info == NoConvergence && it < m_restart && nbIts < m_iterations)
   {    
@@ -312,7 +310,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, con
    
     // Orthogonalize it with the previous basis in the basis using modified Gram-Schmidt
     Scalar coef; 
-    for (int i = 0; i <= it; ++i)
+    for (Index i = 0; i <= it; ++i)
     { 
       coef = tv1.dot(m_V.col(i));
       tv1 = tv1 - coef * m_V.col(i); 
@@ -328,7 +326,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, con
     // FIXME Check for happy breakdown 
     
     // Update Hessenberg matrix with Givens rotations
-    for (int i = 1; i <= it; ++i) 
+    for (Index i = 1; i <= it; ++i) 
     {
       m_H.col(it).applyOnTheLeft(i-1,i,gr[i-1].adjoint());
     }
@@ -418,7 +416,7 @@ inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_Matr
 }
 
 template< typename _MatrixType, typename _Preconditioner>
-int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const
 {
   // First, find the Schur form of the Hessenberg matrix H
   typename internal::conditional<NumTraits<Scalar>::IsComplex, ComplexSchur<DenseMatrix>, RealSchur<DenseMatrix> >::type schurofH; 
@@ -433,8 +431,8 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
   
   // Reorder the absolute values of Schur values
   DenseRealVector modulEig(it); 
-  for (int j=0; j<it; ++j) modulEig(j) = std::abs(eig(j)); 
-  perm.setLinSpaced(it,0,it-1);
+  for (Index j=0; j<it; ++j) modulEig(j) = std::abs(eig(j)); 
+  perm.setLinSpaced(it,0,internal::convert_index<StorageIndex>(it-1));
   internal::sortWithPermutation(modulEig, perm, neig);
   
   if (!m_lambdaN)
@@ -442,7 +440,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
     m_lambdaN = (std::max)(modulEig.maxCoeff(), m_lambdaN);
   }
   //Count the real number of extracted eigenvalues (with complex conjugates)
-  int nbrEig = 0; 
+  Index nbrEig = 0; 
   while (nbrEig < neig)
   {
     if(eig(perm(it-nbrEig-1)).imag() == RealScalar(0)) nbrEig++; 
@@ -451,7 +449,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
   // Extract the  Schur vectors corresponding to the smallest Ritz values
   DenseMatrix Sr(it, nbrEig); 
   Sr.setZero();
-  for (int j = 0; j < nbrEig; j++)
+  for (Index j = 0; j < nbrEig; j++)
   {
     Sr.col(j) = schurofH.matrixU().col(perm(it-j-1));
   }
@@ -462,8 +460,8 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
   if (m_r)
   {
    // Orthogonalize X against m_U using modified Gram-Schmidt
-   for (int j = 0; j < nbrEig; j++)
-     for (int k =0; k < m_r; k++)
+   for (Index j = 0; j < nbrEig; j++)
+     for (Index k =0; k < m_r; k++)
       X.col(j) = X.col(j) - (m_U.col(k).dot(X.col(j)))*m_U.col(k); 
   }
   
@@ -473,7 +471,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
     dgmresInitDeflation(m); 
   DenseMatrix MX(m, nbrEig);
   DenseVector tv1(m);
-  for (int j = 0; j < nbrEig; j++)
+  for (Index j = 0; j < nbrEig; j++)
   {
     tv1 = mat * X.col(j);
     MX.col(j) = precond.solve(tv1);
@@ -488,8 +486,8 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
   }
   
   // Save X into m_U and m_MX in m_MU
-  for (int j = 0; j < nbrEig; j++) m_U.col(m_r+j) = X.col(j);
-  for (int j = 0; j < nbrEig; j++) m_MU.col(m_r+j) = MX.col(j);
+  for (Index j = 0; j < nbrEig; j++) m_U.col(m_r+j) = X.col(j);
+  for (Index j = 0; j < nbrEig; j++) m_MU.col(m_r+j) = MX.col(j);
   // Increase the size of the invariant subspace
   m_r += nbrEig; 
   
@@ -502,7 +500,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const Matri
 }
 template<typename _MatrixType, typename _Preconditioner>
 template<typename RhsType, typename DestType>
-int DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, DestType &y) const
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, DestType &y) const
 {
   DenseVector x1 = m_U.leftCols(m_r).transpose() * x; 
   y = x + m_U.leftCols(m_r) * ( m_lambdaN * m_luT.solve(x1) - x1);

From c25034710ede8eb7e177e8a7426817b6a29e5440 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 7 Jun 2018 16:09:22 +0200
Subject: [PATCH 369/680] Fiw some warnings in dox examples

---
 doc/examples/Cwise_lgamma.cpp                         | 2 +-
 doc/examples/Tutorial_simple_example_dynamic_size.cpp | 2 +-
 doc/special_examples/Tutorial_sparse_example.cpp      | 6 +++++-
 3 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/doc/examples/Cwise_lgamma.cpp b/doc/examples/Cwise_lgamma.cpp
index f1c4f503e..6bfaccbce 100644
--- a/doc/examples/Cwise_lgamma.cpp
+++ b/doc/examples/Cwise_lgamma.cpp
@@ -6,4 +6,4 @@ int main()
 {
   Array4d v(0.5,10,0,-1);
   std::cout << v.lgamma() << std::endl;
-}
\ No newline at end of file
+}
diff --git a/doc/examples/Tutorial_simple_example_dynamic_size.cpp b/doc/examples/Tutorial_simple_example_dynamic_size.cpp
index 0f0280e0e..defcb1ee4 100644
--- a/doc/examples/Tutorial_simple_example_dynamic_size.cpp
+++ b/doc/examples/Tutorial_simple_example_dynamic_size.cpp
@@ -10,7 +10,7 @@ int main()
     MatrixXi m(size,size+1);         // a (size)x(size+1)-matrix of int's
     for (int j=0; j<m.cols(); ++j)   // loop over columns
       for (int i=0; i<m.rows(); ++i) // loop over rows
-        m(i,j) = i+j*m.rows();       // to access matrix coefficients,
+        m(i,j) = i+j*size;           // to access matrix coefficients,
                                      // use operator()(int,int)
     std::cout << m << "\n\n";
   }
diff --git a/doc/special_examples/Tutorial_sparse_example.cpp b/doc/special_examples/Tutorial_sparse_example.cpp
index 89937b411..8850db052 100644
--- a/doc/special_examples/Tutorial_sparse_example.cpp
+++ b/doc/special_examples/Tutorial_sparse_example.cpp
@@ -1,5 +1,6 @@
 #include <Eigen/Sparse>
 #include <vector>
+#include <iostream>
 
 typedef Eigen::SparseMatrix<double> SpMat; // declares a column-major sparse matrix type of double
 typedef Eigen::Triplet<double> T;
@@ -9,7 +10,10 @@ void saveAsBitmap(const Eigen::VectorXd& x, int n, const char* filename);
 
 int main(int argc, char** argv)
 {
-  assert(argc==2);
+  if(argc!=2) {
+    std::cerr << "Error: expected one and only one argument.\n";
+    return -1;
+  }
   
   int n = 300;  // size of the image
   int m = n*n;  // number of unknowns (=number of pixels)

From 7d7bb91537e679c8246107936b5fd376bba1f5b0 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 7 Jun 2018 20:30:09 +0200
Subject: [PATCH 370/680] Missing line during manual rebase of PR-374

---
 Eigen/src/Core/ProductEvaluators.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 5d2737d01..0d5aec570 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -357,7 +357,7 @@ struct generic_product_impl_base
   { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); }
   
   template<typename Dst>
-  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   { Derived::scaleAndAddTo(dst,lhs,rhs,alpha); }
 
 };

From 522d3ca54dcb0d79b2c69acf0e1624746d3f41ec Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 7 Jun 2018 13:02:07 -0700
Subject: [PATCH 371/680] Don't use std::equal_to inside cuda kernels since
 it's not supported.

---
 Eigen/src/Core/util/Meta.h | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 6e5af35c0..0d0b8c43a 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -551,23 +551,27 @@ T div_ceil(const T &a, const T &b)
 
 // The aim of the following functions is to bypass -Wfloat-equal warnings
 // when we really want a strict equality comparison on floating points.
-template<typename X, typename Y> EIGEN_STRONG_INLINE
+template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const X& x,const Y& y) { return x == y; }
 
-template<> EIGEN_STRONG_INLINE
+#if !defined(EIGEN_CUDA_ARCH)
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const float& x,const float& y) { return std::equal_to<float>()(x,y); }
 
-template<> EIGEN_STRONG_INLINE
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const double& x,const double& y) { return std::equal_to<double>()(x,y); }
+#endif
 
-template<typename X, typename Y> EIGEN_STRONG_INLINE
+template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const X& x,const Y& y) { return x != y; }
 
-template<> EIGEN_STRONG_INLINE
+#if !defined(EIGEN_CUDA_ARCH)
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to<float>()(x,y); }
 
-template<> EIGEN_STRONG_INLINE
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const double& x,const double& y) { return std::not_equal_to<double>()(x,y); }
+#endif
 
 } // end namespace numext
 

From f4d1461874879f249f1f45c111ca238f2ee51831 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 09:55:46 +0200
Subject: [PATCH 372/680] Fix the way matrix folder is passed to the tests.

---
 test/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index e1eef086e..af4fa73e4 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -47,7 +47,7 @@ set(EIGEN_TEST_MATRIX_DIR "" CACHE STRING "Enable testing of realword sparse mat
 if(EIGEN_TEST_MATRIX_DIR)
   if(NOT WIN32)
     message(STATUS "Test realworld sparse matrices: ${EIGEN_TEST_MATRIX_DIR}")
-    add_definitions( -DTEST_REAL_CASES=${EIGEN_TEST_MATRIX_DIR} )
+    add_definitions( -DTEST_REAL_CASES="${EIGEN_TEST_MATRIX_DIR}" )
   else(NOT WIN32)
     message(STATUS "REAL CASES CAN NOT BE CURRENTLY TESTED ON WIN32")
   endif(NOT WIN32)

From 7933267c671c4166ca0f537a522168e398b94238 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 09:56:01 +0200
Subject: [PATCH 373/680] fix prototype

---
 unsupported/Eigen/src/IterativeSolvers/DGMRES.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index f40b80eda..be039e07f 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -173,7 +173,7 @@ class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
   /** 
    * Set the restart value (default is 30)  
    */
-  Index set_restart(const Index restart) { m_restart=restart; }
+  void set_restart(const Index restart) { m_restart=restart; }
   
   /** 
    * Set the number of eigenvalues to deflate at each restart 

From f05dea6b2326836e5e0243fbaffbece84b833d64 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 10:14:57 +0200
Subject: [PATCH 374/680] bug #1550: prevent avoidable memory allocation in
 RealSchur

---
 Eigen/src/Eigenvalues/RealSchur.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h
index b72799e5b..4c53344bb 100644
--- a/Eigen/src/Eigenvalues/RealSchur.h
+++ b/Eigen/src/Eigenvalues/RealSchur.h
@@ -284,13 +284,13 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa
   using std::abs;
 
   m_matT = matrixH;
+  m_workspaceVector.resize(m_matT.cols());
   if(computeU)
-    m_matU = matrixQ;
+    matrixQ.evalTo(m_matU, m_workspaceVector);
   
   Index maxIters = m_maxIters;
   if (maxIters == -1)
     maxIters = m_maxIterationsPerRow * matrixH.rows();
-  m_workspaceVector.resize(m_matT.cols());
   Scalar* workspace = &m_workspaceVector.coeffRef(0);
 
   // The matrix m_matT is divided in three parts. 

From 89d65bb9d6ed8ebc9ce297a3fad90d78ba6cbbbb Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 16:50:17 +0200
Subject: [PATCH 375/680] bug #1531: expose NumDimensions for compatibility
 with Tensor

---
 Eigen/src/Core/DenseBase.h |  5 ++++
 test/indexed_view.cpp      |  4 ----
 test/main.h                |  2 ++
 test/mapped_matrix.cpp     | 49 +++++++++++++++++++++++++++++++++++++-
 4 files changed, 55 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index 53b427b17..8b994f35c 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -157,6 +157,11 @@ template<typename Derived> class DenseBase
           * we are dealing with a column-vector (if there is only one column) or with
           * a row-vector (if there is only one row). */
 
+      NumDimensions = int(MaxSizeAtCompileTime) == 1 ? 0 : bool(IsVectorAtCompileTime) ? 1 : 2,
+        /**< This value is equal to Tensor::NumDimensions, i.e. 0 for scalars, 1 for vectors, 
+         * and 2 for matrices.
+         */
+
       Flags = internal::traits<Derived>::Flags,
         /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
           * constructed from this one. See the \ref flags "list of flags".
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 2d46ffd6b..551dc55b0 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -397,10 +397,6 @@ void test_indexed_view()
 //   }
 
   // static checks of some internals:
-
-  #define STATIC_CHECK( COND ) \
-    EIGEN_STATIC_ASSERT( (COND) , EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT )
-
   STATIC_CHECK(( internal::is_valid_index_type<int>::value ));
   STATIC_CHECK(( internal::is_valid_index_type<unsigned int>::value ));
   STATIC_CHECK(( internal::is_valid_index_type<short>::value ));
diff --git a/test/main.h b/test/main.h
index a94e1ab41..9c8148de2 100644
--- a/test/main.h
+++ b/test/main.h
@@ -339,6 +339,8 @@ inline void verify_impl(bool condition, const char *testname, const char *file,
 
 #define VERIFY_IS_UNITARY(a) VERIFY(test_isUnitary(a))
 
+#define STATIC_CHECK(COND) EIGEN_STATIC_ASSERT( (COND) , EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT )
+
 #define CALL_SUBTEST(FUNC) do { \
     g_test_stack.push_back(EI_PP_MAKE_STRING(FUNC)); \
     FUNC; \
diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp
index 6a84c5897..b9f36d8d6 100644
--- a/test/mapped_matrix.cpp
+++ b/test/mapped_matrix.cpp
@@ -181,6 +181,49 @@ void map_not_aligned_on_scalar()
   internal::aligned_delete(array1, (size+1)*(size+1)+1);
 }
 
+#if EIGEN_HAS_CXX11
+template<template <typename,int,int> class Object>
+void map_num_dimensions()
+{
+  typedef Object<double, 1, 1> ArrayScalarType;
+  typedef Object<double, 2, 1> ArrayVectorType;
+  typedef Object<double, 1, 2> TransposeArrayVectorType;
+  typedef Object<double, 2, 2> ArrayType;
+  typedef Object<double, Eigen::Dynamic, 1> DynamicArrayVectorType;
+  typedef Object<double, 1, Eigen::Dynamic> DynamicTransposeArrayVectorType;
+  typedef Object<double, Eigen::Dynamic, Eigen::Dynamic> DynamicArrayType;
+
+  STATIC_CHECK(ArrayScalarType::NumDimensions == 0);
+  STATIC_CHECK(ArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(TransposeArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(ArrayType::NumDimensions == 2);
+  STATIC_CHECK(DynamicArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(DynamicTransposeArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(DynamicArrayType::NumDimensions == 2);
+
+  typedef Eigen::Map<ArrayScalarType> ArrayScalarMap;
+  typedef Eigen::Map<ArrayVectorType> ArrayVectorMap;
+  typedef Eigen::Map<TransposeArrayVectorType> TransposeArrayVectorMap;
+  typedef Eigen::Map<ArrayType> ArrayMap;
+  typedef Eigen::Map<DynamicArrayVectorType> DynamicArrayVectorMap;
+  typedef Eigen::Map<DynamicTransposeArrayVectorType> DynamicTransposeArrayVectorMap;
+  typedef Eigen::Map<DynamicArrayType> DynamicArrayMap;
+
+  STATIC_CHECK(ArrayScalarMap::NumDimensions == 0);
+  STATIC_CHECK(ArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(TransposeArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(ArrayMap::NumDimensions == 2);
+  STATIC_CHECK(DynamicArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(DynamicTransposeArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(DynamicArrayMap::NumDimensions == 2);
+}
+
+template<typename Scalar, int Rows, int Cols>
+using TArray = Array<Scalar,Rows,Cols>;
+template<typename Scalar, int Rows, int Cols>
+using TMatrix = Matrix<Scalar,Rows,Cols>;
+#endif
+
 void test_mapped_matrix()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -205,7 +248,11 @@ void test_mapped_matrix()
     CALL_SUBTEST_8( map_static_methods(RowVector3d()) );
     CALL_SUBTEST_9( map_static_methods(VectorXcd(8)) );
     CALL_SUBTEST_10( map_static_methods(VectorXf(12)) );
-    
     CALL_SUBTEST_11( map_not_aligned_on_scalar<double>() );
   }
+
+  #if EIGEN_HAS_CXX11
+    CALL_SUBTEST_12( map_num_dimensions<TArray>() );
+    CALL_SUBTEST_12( map_num_dimensions<TMatrix>() );
+  #endif
 }

From d6813fb1c565e0150cad0332b5afa6ac2e1ed286 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 16:55:10 +0200
Subject: [PATCH 376/680] bug #1531: expose NumDimensions for solve and sparse
 expressions.

---
 Eigen/src/Core/SolverBase.h             | 3 ++-
 Eigen/src/SparseCore/SparseMatrixBase.h | 5 +++++
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/SolverBase.h b/Eigen/src/Core/SolverBase.h
index 8a4adc229..702a5485c 100644
--- a/Eigen/src/Core/SolverBase.h
+++ b/Eigen/src/Core/SolverBase.h
@@ -56,7 +56,8 @@ class SolverBase : public EigenBase<Derived>
       MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
                                                              internal::traits<Derived>::MaxColsAtCompileTime>::ret),
       IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1
-                           || internal::traits<Derived>::MaxColsAtCompileTime == 1
+                           || internal::traits<Derived>::MaxColsAtCompileTime == 1,
+      NumDimensions = int(MaxSizeAtCompileTime) == 1 ? 0 : bool(IsVectorAtCompileTime) ? 1 : 2
     };
 
     /** Default constructor */
diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h
index dac2d8c62..229449f02 100644
--- a/Eigen/src/SparseCore/SparseMatrixBase.h
+++ b/Eigen/src/SparseCore/SparseMatrixBase.h
@@ -87,6 +87,11 @@ template<typename Derived> class SparseMatrixBase
           * we are dealing with a column-vector (if there is only one column) or with
           * a row-vector (if there is only one row). */
 
+      NumDimensions = int(MaxSizeAtCompileTime) == 1 ? 0 : bool(IsVectorAtCompileTime) ? 1 : 2,
+        /**< This value is equal to Tensor::NumDimensions, i.e. 0 for scalars, 1 for vectors,
+         * and 2 for matrices.
+         */
+
       Flags = internal::traits<Derived>::Flags,
         /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
           * constructed from this one. See the \ref flags "list of flags".

From cb4c9a6a9459a1c9bee1b22cabb5e8aa154968d9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 8 Jun 2018 17:11:45 +0200
Subject: [PATCH 377/680] bug #1531: make dedicatd unit testing for
 NumDimensions

---
 test/CMakeLists.txt     |  1 +
 test/mapped_matrix.cpp  | 48 ----------------------
 test/num_dimensions.cpp | 90 +++++++++++++++++++++++++++++++++++++++++
 3 files changed, 91 insertions(+), 48 deletions(-)
 create mode 100644 test/num_dimensions.cpp

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index af4fa73e4..ab3ff4795 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -292,6 +292,7 @@ ei_add_test(mpl2only)
 ei_add_test(inplace_decomposition)
 ei_add_test(half_float)
 ei_add_test(array_of_string)
+ei_add_test(num_dimensions)
 
 add_executable(bug1213 bug1213.cpp bug1213_main.cpp)
 
diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp
index b9f36d8d6..7b9e1fe52 100644
--- a/test/mapped_matrix.cpp
+++ b/test/mapped_matrix.cpp
@@ -181,49 +181,6 @@ void map_not_aligned_on_scalar()
   internal::aligned_delete(array1, (size+1)*(size+1)+1);
 }
 
-#if EIGEN_HAS_CXX11
-template<template <typename,int,int> class Object>
-void map_num_dimensions()
-{
-  typedef Object<double, 1, 1> ArrayScalarType;
-  typedef Object<double, 2, 1> ArrayVectorType;
-  typedef Object<double, 1, 2> TransposeArrayVectorType;
-  typedef Object<double, 2, 2> ArrayType;
-  typedef Object<double, Eigen::Dynamic, 1> DynamicArrayVectorType;
-  typedef Object<double, 1, Eigen::Dynamic> DynamicTransposeArrayVectorType;
-  typedef Object<double, Eigen::Dynamic, Eigen::Dynamic> DynamicArrayType;
-
-  STATIC_CHECK(ArrayScalarType::NumDimensions == 0);
-  STATIC_CHECK(ArrayVectorType::NumDimensions == 1);
-  STATIC_CHECK(TransposeArrayVectorType::NumDimensions == 1);
-  STATIC_CHECK(ArrayType::NumDimensions == 2);
-  STATIC_CHECK(DynamicArrayVectorType::NumDimensions == 1);
-  STATIC_CHECK(DynamicTransposeArrayVectorType::NumDimensions == 1);
-  STATIC_CHECK(DynamicArrayType::NumDimensions == 2);
-
-  typedef Eigen::Map<ArrayScalarType> ArrayScalarMap;
-  typedef Eigen::Map<ArrayVectorType> ArrayVectorMap;
-  typedef Eigen::Map<TransposeArrayVectorType> TransposeArrayVectorMap;
-  typedef Eigen::Map<ArrayType> ArrayMap;
-  typedef Eigen::Map<DynamicArrayVectorType> DynamicArrayVectorMap;
-  typedef Eigen::Map<DynamicTransposeArrayVectorType> DynamicTransposeArrayVectorMap;
-  typedef Eigen::Map<DynamicArrayType> DynamicArrayMap;
-
-  STATIC_CHECK(ArrayScalarMap::NumDimensions == 0);
-  STATIC_CHECK(ArrayVectorMap::NumDimensions == 1);
-  STATIC_CHECK(TransposeArrayVectorMap::NumDimensions == 1);
-  STATIC_CHECK(ArrayMap::NumDimensions == 2);
-  STATIC_CHECK(DynamicArrayVectorMap::NumDimensions == 1);
-  STATIC_CHECK(DynamicTransposeArrayVectorMap::NumDimensions == 1);
-  STATIC_CHECK(DynamicArrayMap::NumDimensions == 2);
-}
-
-template<typename Scalar, int Rows, int Cols>
-using TArray = Array<Scalar,Rows,Cols>;
-template<typename Scalar, int Rows, int Cols>
-using TMatrix = Matrix<Scalar,Rows,Cols>;
-#endif
-
 void test_mapped_matrix()
 {
   for(int i = 0; i < g_repeat; i++) {
@@ -250,9 +207,4 @@ void test_mapped_matrix()
     CALL_SUBTEST_10( map_static_methods(VectorXf(12)) );
     CALL_SUBTEST_11( map_not_aligned_on_scalar<double>() );
   }
-
-  #if EIGEN_HAS_CXX11
-    CALL_SUBTEST_12( map_num_dimensions<TArray>() );
-    CALL_SUBTEST_12( map_num_dimensions<TMatrix>() );
-  #endif
 }
diff --git a/test/num_dimensions.cpp b/test/num_dimensions.cpp
new file mode 100644
index 000000000..f5209283d
--- /dev/null
+++ b/test/num_dimensions.cpp
@@ -0,0 +1,90 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+#include <Eigen/SparseCore>
+
+template<int ExpectedDim,typename Xpr>
+void check_dim(const Xpr& ) {
+  STATIC_CHECK( Xpr::NumDimensions == ExpectedDim );
+}
+
+#if EIGEN_HAS_CXX11
+template<template <typename,int,int> class Object>
+void map_num_dimensions()
+{
+  typedef Object<double, 1, 1> ArrayScalarType;
+  typedef Object<double, 2, 1> ArrayVectorType;
+  typedef Object<double, 1, 2> TransposeArrayVectorType;
+  typedef Object<double, 2, 2> ArrayType;
+  typedef Object<double, Eigen::Dynamic, 1> DynamicArrayVectorType;
+  typedef Object<double, 1, Eigen::Dynamic> DynamicTransposeArrayVectorType;
+  typedef Object<double, Eigen::Dynamic, Eigen::Dynamic> DynamicArrayType;
+
+  STATIC_CHECK(ArrayScalarType::NumDimensions == 0);
+  STATIC_CHECK(ArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(TransposeArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(ArrayType::NumDimensions == 2);
+  STATIC_CHECK(DynamicArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(DynamicTransposeArrayVectorType::NumDimensions == 1);
+  STATIC_CHECK(DynamicArrayType::NumDimensions == 2);
+
+  typedef Eigen::Map<ArrayScalarType> ArrayScalarMap;
+  typedef Eigen::Map<ArrayVectorType> ArrayVectorMap;
+  typedef Eigen::Map<TransposeArrayVectorType> TransposeArrayVectorMap;
+  typedef Eigen::Map<ArrayType> ArrayMap;
+  typedef Eigen::Map<DynamicArrayVectorType> DynamicArrayVectorMap;
+  typedef Eigen::Map<DynamicTransposeArrayVectorType> DynamicTransposeArrayVectorMap;
+  typedef Eigen::Map<DynamicArrayType> DynamicArrayMap;
+
+  STATIC_CHECK(ArrayScalarMap::NumDimensions == 0);
+  STATIC_CHECK(ArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(TransposeArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(ArrayMap::NumDimensions == 2);
+  STATIC_CHECK(DynamicArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(DynamicTransposeArrayVectorMap::NumDimensions == 1);
+  STATIC_CHECK(DynamicArrayMap::NumDimensions == 2);
+}
+
+template<typename Scalar, int Rows, int Cols>
+using TArray = Array<Scalar,Rows,Cols>;
+
+template<typename Scalar, int Rows, int Cols>
+using TMatrix = Matrix<Scalar,Rows,Cols>;
+
+#endif
+
+void test_num_dimensions()
+{
+  int n = 10;
+  ArrayXXd A(n,n);
+  CALL_SUBTEST( check_dim<2>(A) );
+  CALL_SUBTEST( check_dim<2>(A.block(1,1,2,2)) );
+  CALL_SUBTEST( check_dim<1>(A.col(1)) );
+  CALL_SUBTEST( check_dim<1>(A.row(1)) );
+
+  MatrixXd M(n,n);
+  CALL_SUBTEST( check_dim<0>(M.row(1)*M.col(1)) );
+
+  SparseMatrix<double> S(n,n);
+  CALL_SUBTEST( check_dim<2>(S) );
+  CALL_SUBTEST( check_dim<2>(S.block(1,1,2,2)) );
+  CALL_SUBTEST( check_dim<1>(S.col(1)) );
+  CALL_SUBTEST( check_dim<1>(S.row(1)) );
+
+  SparseVector<double> s(n);
+  CALL_SUBTEST( check_dim<1>(s) );
+  CALL_SUBTEST( check_dim<1>(s.head(2)) );
+  
+
+  #if EIGEN_HAS_CXX11
+  CALL_SUBTEST( map_num_dimensions<TArray>() );
+  CALL_SUBTEST( map_num_dimensions<TMatrix>() );
+  #endif
+}

From 5418154a45db637211e94f11ee04c6ae4dc8cf85 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 20 Jun 2018 17:51:48 -0700
Subject: [PATCH 378/680] Fix oversharding bug in parallelFor.

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h   | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index ca9ba402e..90fd99027 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -189,9 +189,11 @@ struct ThreadPoolDevice {
     // of blocks to be evenly dividable across threads.
 
     double block_size_f = 1.0 / CostModel::taskSize(1, cost);
-    Index block_size = numext::mini(n, numext::maxi<Index>(1, block_size_f));
-    const Index max_block_size =
-        numext::mini(n, numext::maxi<Index>(1, 2 * block_size_f));
+    const Index max_oversharding_factor = 4;
+    Index block_size = numext::mini(
+        n, numext::maxi<Index>(divup<Index>(n, max_oversharding_factor * numThreads()),
+                               block_size_f));
+    const Index max_block_size = numext::mini(n, 2 * block_size);
     if (block_align) {
       Index new_block_size = block_align(block_size);
       eigen_assert(new_block_size >= block_size);
@@ -205,7 +207,8 @@ struct ThreadPoolDevice {
         (divup<int>(block_count, numThreads()) * numThreads());
     // Now try to increase block size up to max_block_size as long as it
     // doesn't decrease parallel efficiency.
-    for (Index prev_block_count = block_count; prev_block_count > 1;) {
+    for (Index prev_block_count = block_count;
+         max_efficiency < 1.0 && prev_block_count > 1;) {
       // This is the next block size that divides size into a smaller number
       // of blocks than the current block_size.
       Index coarser_block_size = divup(n, prev_block_count - 1);

From 4cc32d80fd091559bc90a55946945fc1584fdd86 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 21 Jun 2018 10:28:38 +0200
Subject: [PATCH 379/680] bug #1555: compilation fix with XLC

---
 Eigen/src/Core/arch/AltiVec/PacketMath.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index 6d7190a56..7f6dbe62a 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -1054,7 +1054,7 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) {
 
 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
   Packet2l select = { ifPacket.select[0], ifPacket.select[1] };
-  Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
+  Packet2bl mask = reinterpret_cast<Packet2bl>( vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
   return vec_sel(elsePacket, thenPacket, mask);
 }
 #endif // __VSX__

From bda71ad3945d4cf2bf06cf567b0b5a8a4663fede Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 22 Jun 2018 15:04:35 -0700
Subject: [PATCH 380/680] Fix typo in pbend for AltiVec.

---
 Eigen/src/Core/arch/AltiVec/PacketMath.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index 6d7190a56..7f4e90f75 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -103,7 +103,7 @@ static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4u
 static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
 static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
 #else
-static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; 
+static Packet16uc p16uc_FORWARD = p16uc_REVERSE32;
 static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
 static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
 static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
@@ -784,7 +784,7 @@ typedef __vector __bool long         Packet2bl;
 
 static Packet2l  p2l_ONE  = { 1, 1 };
 static Packet2l  p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO);
-static Packet2d  p2d_ONE  = { 1.0, 1.0 }; 
+static Packet2d  p2d_ONE  = { 1.0, 1.0 };
 static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
 static Packet2d  p2d_MZERO = { -0.0, -0.0 };
 
@@ -1001,7 +1001,7 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
   Packet2d v[2], sum;
   v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8));
   v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8));
- 
+
 #ifdef _BIG_ENDIAN
   sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8));
 #else
@@ -1054,7 +1054,7 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) {
 
 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
   Packet2l select = { ifPacket.select[0], ifPacket.select[1] };
-  Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
+  Packet2bl mask = reinterpret_cast<Packet2bl>( vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE)) );
   return vec_sel(elsePacket, thenPacket, mask);
 }
 #endif // __VSX__

From ee5864f72e83830f536ad91dc38d574c02a08348 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 25 Jun 2018 10:30:12 +0200
Subject: [PATCH 381/680] bug #1560 fix product with a 1x1 diagonal matrix

---
 Eigen/src/Core/ProductEvaluators.h | 11 +++++++++--
 test/diagonalmatrices.cpp          | 21 +++++++++++++++++++++
 2 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 0d5aec570..60de771b2 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -785,7 +785,11 @@ public:
     _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))),
     _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
     Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0),
-    Alignment = evaluator<MatrixType>::Alignment
+    Alignment = evaluator<MatrixType>::Alignment,
+
+    AsScalarProduct =     (DiagonalType::SizeAtCompileTime==1)
+                      ||  (DiagonalType::SizeAtCompileTime==Dynamic && MatrixType::RowsAtCompileTime==1 && ProductOrder==OnTheLeft)
+                      ||  (DiagonalType::SizeAtCompileTime==Dynamic && MatrixType::ColsAtCompileTime==1 && ProductOrder==OnTheRight)
   };
   
   diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag)
@@ -797,7 +801,10 @@ public:
   
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
   {
-    return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx);
+    if(AsScalarProduct)
+      return m_diagImpl.coeff(0) * m_matImpl.coeff(idx);
+    else
+      return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx);
   }
   
 protected:
diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp
index a2092c43e..73b4cd58c 100644
--- a/test/diagonalmatrices.cpp
+++ b/test/diagonalmatrices.cpp
@@ -30,6 +30,7 @@ template<typename MatrixType> void diagonalmatrices(const MatrixType& m)
              v2 = VectorType::Random(rows);
   RowVectorType rv1 = RowVectorType::Random(cols),
              rv2 = RowVectorType::Random(cols);
+
   LeftDiagonalMatrix ldm1(v1), ldm2(v2);
   RightDiagonalMatrix rdm1(rv1), rdm2(rv2);
   
@@ -105,6 +106,25 @@ template<typename MatrixType> void diagonalmatrices(const MatrixType& m)
   sq_m2 = sq_m1 * sq_m2;
   VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).col(i), sq_m2.col(i) );
   VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).row(i), sq_m2.row(i) );
+
+  if(v1.size()==1)
+  {
+    typedef Matrix<Scalar, Dynamic, 1> DynVectorType;
+    typedef Matrix<Scalar, 1, Dynamic> DynRowVectorType;
+    Index depth = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
+    DynVectorType     dv1 = DynVectorType::Random(depth);
+    DynRowVectorType  drv1 = DynRowVectorType::Random(depth);
+    DynMatrixType dm1 = dv1;
+    DynMatrixType drm1 = drv1;
+    
+    Scalar s = v1(0);
+
+    VERIFY_IS_APPROX( v1.asDiagonal() * drv1, s*drv1 );
+    VERIFY_IS_APPROX( dv1 * v1.asDiagonal(), dv1*s );
+
+    VERIFY_IS_APPROX( v1.asDiagonal() * drm1, s*drm1 );
+    VERIFY_IS_APPROX( dm1 * v1.asDiagonal(), dm1*s );
+  }
 }
 
 template<int>
@@ -130,6 +150,7 @@ void test_diagonalmatrices()
     CALL_SUBTEST_7( diagonalmatrices(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_8( diagonalmatrices(Matrix<double,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_9( diagonalmatrices(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_9( diagonalmatrices(MatrixXf(1,1)) );
   }
   CALL_SUBTEST_10( bug987<0>() );
 }

From f9d337780d49825765cbb8ea51843a905c0e5253 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 25 Jun 2018 14:26:51 +0200
Subject: [PATCH 382/680] First step towards a generic vectorised quaternion
 product

---
 Eigen/src/Geometry/arch/Geometry_SSE.h | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/Eigen/src/Geometry/arch/Geometry_SSE.h b/Eigen/src/Geometry/arch/Geometry_SSE.h
index f68cab583..d4346aa1c 100644
--- a/Eigen/src/Geometry/arch/Geometry_SSE.h
+++ b/Eigen/src/Geometry/arch/Geometry_SSE.h
@@ -26,17 +26,17 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, float>
   static inline Quaternion<float> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
   {
     Quaternion<float> res;
-    const __m128 mask = _mm_setr_ps(0.f,0.f,0.f,-0.f);
-    __m128 a = _a.coeffs().template packet<AAlignment>(0);
-    __m128 b = _b.coeffs().template packet<BAlignment>(0);
-    __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
-    __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
+    const Packet4f mask = _mm_setr_ps(0.f,0.f,0.f,-0.f);
+    Packet4f a = _a.coeffs().template packet<AAlignment>(0);
+    Packet4f b = _b.coeffs().template packet<BAlignment>(0);
+    Packet4f s1 = pmul(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
+    Packet4f s2 = pmul(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
     pstoret<float,Packet4f,ResAlignment>(
               &res.x(),
-              _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
-                                    _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0),
+              padd(psub(pmul(a,vec4f_swizzle1(b,3,3,3,3)),
+                                    pmul(vec4f_swizzle1(a,2,0,1,0),
                                                vec4f_swizzle1(b,1,2,0,0))),
-                         _mm_xor_ps(mask,_mm_add_ps(s1,s2))));
+                         pxor(mask,padd(s1,s2))));
     
     return res;
   }

From b7689bded9e87c95affb5963ad565a68974aa7f7 Mon Sep 17 00:00:00 2001
From: Jonathan Liu <net147@gmail.com>
Date: Thu, 28 Jun 2018 00:32:37 +1000
Subject: [PATCH 383/680] Use std::complex constructor instead of assignment
 from scalar

Fixes GCC conversion to non-scalar type requested compile error when
using boost::multiprecision::cpp_dec_float_50 as scalar type.
---
 unsupported/Eigen/src/FFT/ei_kissfft_impl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/src/FFT/ei_kissfft_impl.h b/unsupported/Eigen/src/FFT/ei_kissfft_impl.h
index be51b4e6f..079e88602 100644
--- a/unsupported/Eigen/src/FFT/ei_kissfft_impl.h
+++ b/unsupported/Eigen/src/FFT/ei_kissfft_impl.h
@@ -316,8 +316,8 @@ struct kissfft_impl
 
         // use optimized mode for even real
         fwd( dst, reinterpret_cast<const Complex*> (src), ncfft);
-        Complex dc = dst[0].real() +  dst[0].imag();
-        Complex nyquist = dst[0].real() -  dst[0].imag();
+        Complex dc(dst[0].real() +  dst[0].imag());
+        Complex nyquist(dst[0].real() -  dst[0].imag());
         int k;
         for ( k=1;k <= ncfft2 ; ++k ) {
           Complex fpk = dst[k];

From 9a81de1d3543d89d2cc5d32294d96b3c58730b21 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 28 Jun 2018 00:20:59 +0200
Subject: [PATCH 384/680] Fix order of EIGEN_DEVICE_FUNC and returned type

---
 Eigen/src/Core/ArrayWrapper.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h
index 688aadd62..757b31825 100644
--- a/Eigen/src/Core/ArrayWrapper.h
+++ b/Eigen/src/Core/ArrayWrapper.h
@@ -90,8 +90,8 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const { dst = m_expression; }
 
-    const typename internal::remove_all<NestedExpressionType>::type& 
     EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<NestedExpressionType>::type& 
     nestedExpression() const 
     {
       return m_expression;

From 0cdacf3fa49ff52fac3058beae1d6beaa0ebec46 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 29 Jun 2018 11:28:36 +0200
Subject: [PATCH 385/680] update comment

---
 Eigen/src/QR/HouseholderQR.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h
index 13c02685d..33cb9c8ff 100644
--- a/Eigen/src/QR/HouseholderQR.h
+++ b/Eigen/src/QR/HouseholderQR.h
@@ -291,7 +291,7 @@ template<typename MatrixQR, typename HCoeffs,
   bool InnerStrideIsOne = (MatrixQR::InnerStrideAtCompileTime == 1 && HCoeffs::InnerStrideAtCompileTime == 1)>
 struct householder_qr_inplace_blocked
 {
-  // This is specialized for MKL-supported Scalar types in HouseholderQR_MKL.h
+  // This is specialized for LAPACK-supported Scalar types in HouseholderQR_LAPACKE.h
   static void run(MatrixQR& mat, HCoeffs& hCoeffs, Index maxBlockSize=32,
       typename MatrixQR::Scalar* tempData = 0)
   {

From a7b313a16cf5b64981dd953f150327638379e68b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 1 Jul 2018 22:45:47 +0200
Subject: [PATCH 386/680] Fix unit test

---
 test/diagonalmatrices.cpp | 43 ++++++++++++++++++++++++---------------
 1 file changed, 27 insertions(+), 16 deletions(-)

diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp
index 73b4cd58c..a4ff10239 100644
--- a/test/diagonalmatrices.cpp
+++ b/test/diagonalmatrices.cpp
@@ -106,25 +106,32 @@ template<typename MatrixType> void diagonalmatrices(const MatrixType& m)
   sq_m2 = sq_m1 * sq_m2;
   VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).col(i), sq_m2.col(i) );
   VERIFY_IS_APPROX( (sq_m1*v1.asDiagonal()).row(i), sq_m2.row(i) );
+}
 
-  if(v1.size()==1)
-  {
-    typedef Matrix<Scalar, Dynamic, 1> DynVectorType;
-    typedef Matrix<Scalar, 1, Dynamic> DynRowVectorType;
-    Index depth = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
-    DynVectorType     dv1 = DynVectorType::Random(depth);
-    DynRowVectorType  drv1 = DynRowVectorType::Random(depth);
-    DynMatrixType dm1 = dv1;
-    DynMatrixType drm1 = drv1;
-    
-    Scalar s = v1(0);
+template<typename MatrixType> void as_scalar_product(const MatrixType& m)
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
+  typedef Matrix<Scalar, Dynamic, Dynamic> DynMatrixType;
+  typedef Matrix<Scalar, Dynamic, 1> DynVectorType;
+  typedef Matrix<Scalar, 1, Dynamic> DynRowVectorType;
 
-    VERIFY_IS_APPROX( v1.asDiagonal() * drv1, s*drv1 );
-    VERIFY_IS_APPROX( dv1 * v1.asDiagonal(), dv1*s );
+  Index rows = m.rows();
+  Index depth = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
 
-    VERIFY_IS_APPROX( v1.asDiagonal() * drm1, s*drm1 );
-    VERIFY_IS_APPROX( dm1 * v1.asDiagonal(), dm1*s );
-  }
+  VectorType v1 = VectorType::Random(rows);  
+  DynVectorType     dv1  = DynVectorType::Random(depth);
+  DynRowVectorType  drv1 = DynRowVectorType::Random(depth);
+  DynMatrixType     dm1  = dv1;
+  DynMatrixType     drm1 = drv1;
+  
+  Scalar s = v1(0);
+
+  VERIFY_IS_APPROX( v1.asDiagonal() * drv1, s*drv1 );
+  VERIFY_IS_APPROX( dv1 * v1.asDiagonal(), dv1*s );
+
+  VERIFY_IS_APPROX( v1.asDiagonal() * drm1, s*drm1 );
+  VERIFY_IS_APPROX( dm1 * v1.asDiagonal(), dm1*s );
 }
 
 template<int>
@@ -142,15 +149,19 @@ void test_diagonalmatrices()
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( diagonalmatrices(Matrix<float, 1, 1>()) );
+    CALL_SUBTEST_1( as_scalar_product(Matrix<float, 1, 1>()) );
+
     CALL_SUBTEST_2( diagonalmatrices(Matrix3f()) );
     CALL_SUBTEST_3( diagonalmatrices(Matrix<double,3,3,RowMajor>()) );
     CALL_SUBTEST_4( diagonalmatrices(Matrix4d()) );
     CALL_SUBTEST_5( diagonalmatrices(Matrix<float,4,4,RowMajor>()) );
     CALL_SUBTEST_6( diagonalmatrices(MatrixXcf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_6( as_scalar_product(MatrixXcf(1,1)) );
     CALL_SUBTEST_7( diagonalmatrices(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_8( diagonalmatrices(Matrix<double,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_9( diagonalmatrices(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_9( diagonalmatrices(MatrixXf(1,1)) );
+    CALL_SUBTEST_9( as_scalar_product(MatrixXf(1,1)) );
   }
   CALL_SUBTEST_10( bug987<0>() );
 }

From d428a199ab70bc08db7551457a1e9d8f65d9ebb9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 2 Jul 2018 11:41:09 +0200
Subject: [PATCH 387/680] bug #1562: optimize evaluation of small products of
 the form s*A*B by rewriting them as: s*(A.lazyProduct(B)) to save a costly
 temporary. Measured speedup from 2x to 5x...

---
 Eigen/src/Core/ProductEvaluators.h            | 28 ++++++++++++++++++-
 Eigen/src/Core/products/GeneralMatrixMatrix.h |  8 +++---
 test/product.h                                | 11 ++++++++
 test/product_large.cpp                        |  2 ++
 4 files changed, 44 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 60de771b2..8072a1959 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -396,7 +396,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
     // but easier on the compiler side
     call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<typename Dst::Scalar,Scalar>());
   }
-  
+
   template<typename Dst>
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
@@ -410,6 +410,32 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
     // dst.noalias() -= lhs.lazyProduct(rhs);
     call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>());
   }
+
+  // Catch "dst {,+,-}= (s*A)*B" and evaluate it lazily by moving out the scalar factor:
+  //    dst {,+,-}= s * (A.lazyProduct(B))
+  // This is a huge benefit for heap-allocated matrix types as it save one costly allocation.
+  // For them, this strategy is also faster than simply by-passing the heap allocation through
+  // stack allocation.
+  // For fixed sizes matrices, this is less obvious, it is sometimes x2 faster, but sometimes x3 slower,
+  // and the behavior depends also a lot on the compiler... so let's be conservative and enable them for dynamic-size only,
+  // that is when coming from generic_product_impl<...,GemmProduct> in file GeneralMatrixMatrix.h
+  template<typename Dst, typename Scalar1, typename Scalar2, typename Plain1, typename Xpr2, typename Func>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void eval_dynamic(Dst& dst, const CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                                           const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, Xpr2>& lhs, const Rhs& rhs, const Func &func)
+  {
+    call_assignment_no_alias(dst, lhs.lhs().functor().m_other * lhs.rhs().lazyProduct(rhs), func);
+  }
+
+  // Here, we we always have LhsT==Lhs, but we need to make it a template type to make the above
+  // overload more specialized.
+  template<typename Dst, typename LhsT, typename Func>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void eval_dynamic(Dst& dst, const LhsT& lhs, const Rhs& rhs, const Func &func)
+  {
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), func);
+  }
+  
   
 //   template<typename Dst>
 //   static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index ed4d3182b..dc47cf7cf 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -431,10 +431,10 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
     // to determine the following heuristic.
     // EIGEN_GEMM_TO_COEFFBASED_THRESHOLD is typically defined to 20 in GeneralProduct.h,
     // unless it has been specialized by the user or for a given architecture.
-    // Note that the condition rhs.rows()>0 was required because lazy produc is (was?) not happy with empty inputs.
+    // Note that the condition rhs.rows()>0 was required because lazy product is (was?) not happy with empty inputs.
     // I'm not sure it is still required.
     if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
-      lazyproduct::evalTo(dst, lhs, rhs);
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::assign_op<typename Dst::Scalar,Scalar>());
     else
     {
       dst.setZero();
@@ -446,7 +446,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
-      lazyproduct::addTo(dst, lhs, rhs);
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::add_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst,lhs, rhs, Scalar(1));
   }
@@ -455,7 +455,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
     if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
-      lazyproduct::subTo(dst, lhs, rhs);
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::sub_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
   }
diff --git a/test/product.h b/test/product.h
index 0425a929e..d26e8063d 100644
--- a/test/product.h
+++ b/test/product.h
@@ -111,6 +111,17 @@ template<typename MatrixType> void product(const MatrixType& m)
   vcres.noalias() -= m1.transpose() * v1;
   VERIFY_IS_APPROX(vcres, vc2 - m1.transpose() * v1);
 
+  // test scaled products
+  res = square;
+  res.noalias() = s1 * m1 * m2.transpose();
+  VERIFY_IS_APPROX(res, ((s1*m1).eval() * m2.transpose()));
+  res = square;
+  res.noalias() += s1 * m1 * m2.transpose();
+  VERIFY_IS_APPROX(res, square + ((s1*m1).eval() * m2.transpose()));
+  res = square;
+  res.noalias() -= s1 * m1 * m2.transpose();
+  VERIFY_IS_APPROX(res, square - ((s1*m1).eval() * m2.transpose()));
+
   // test d ?= a+b*c rules
   res.noalias() = square + m1 * m2.transpose();
   VERIFY_IS_APPROX(res, square + m1 * m2.transpose());
diff --git a/test/product_large.cpp b/test/product_large.cpp
index 845cd40ca..14a4f739d 100644
--- a/test/product_large.cpp
+++ b/test/product_large.cpp
@@ -35,6 +35,8 @@ void test_product_large()
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( product(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
+    CALL_SUBTEST_2( product(MatrixXd(internal::random<int>(1,10), internal::random<int>(1,10))) );
+
     CALL_SUBTEST_3( product(MatrixXi(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
     CALL_SUBTEST_4( product(MatrixXcf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2), internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2))) );
     CALL_SUBTEST_5( product(Matrix<float,Dynamic,Dynamic,RowMajor>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );

From d6255649361fde3e937320dfd216a84203adc295 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 2 Jul 2018 11:50:41 +0200
Subject: [PATCH 388/680] Simplify redux_evaluator using inheritance, and
 properly rename parameters in reducers.

---
 Eigen/src/Core/Redux.h | 245 +++++++++++++++++++----------------------
 1 file changed, 115 insertions(+), 130 deletions(-)

diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index 0a99acb21..32574ba60 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -23,22 +23,22 @@ namespace internal {
 * Part 1 : the logic deciding a strategy for vectorization and unrolling
 ***************************************************************************/
 
-template<typename Func, typename Derived>
+template<typename Func, typename Evaluator>
 struct redux_traits
 {
 public:
-    typedef typename find_best_packet<typename Derived::Scalar,Derived::SizeAtCompileTime>::type PacketType;
+    typedef typename find_best_packet<typename Evaluator::Scalar,Evaluator::SizeAtCompileTime>::type PacketType;
   enum {
     PacketSize = unpacket_traits<PacketType>::size,
-    InnerMaxSize = int(Derived::IsRowMajor)
-                 ? Derived::MaxColsAtCompileTime
-                 : Derived::MaxRowsAtCompileTime
+    InnerMaxSize = int(Evaluator::IsRowMajor)
+                 ? Evaluator::MaxColsAtCompileTime
+                 : Evaluator::MaxRowsAtCompileTime
   };
 
   enum {
-    MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit)
+    MightVectorize = (int(Evaluator::Flags)&ActualPacketAccessBit)
                   && (functor_traits<Func>::PacketAccess),
-    MayLinearVectorize = bool(MightVectorize) && (int(Derived::Flags)&LinearAccessBit),
+    MayLinearVectorize = bool(MightVectorize) && (int(Evaluator::Flags)&LinearAccessBit),
     MaySliceVectorize  = bool(MightVectorize) && int(InnerMaxSize)>=3*PacketSize
   };
 
@@ -51,8 +51,8 @@ public:
 
 public:
   enum {
-    Cost = Derived::SizeAtCompileTime == Dynamic ? HugeCost
-         : Derived::SizeAtCompileTime * Derived::CoeffReadCost + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
+    Cost = Evaluator::SizeAtCompileTime == Dynamic ? HugeCost
+         : Evaluator::SizeAtCompileTime * Evaluator::CoeffReadCost + (Evaluator::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
     UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize))
   };
 
@@ -64,9 +64,9 @@ public:
 #ifdef EIGEN_DEBUG_ASSIGN
   static void debug()
   {
-    std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl;
+    std::cerr << "Xpr: " << typeid(typename Evaluator::XprType).name() << std::endl;
     std::cerr.setf(std::ios::hex, std::ios::basefield);
-    EIGEN_DEBUG_VAR(Derived::Flags)
+    EIGEN_DEBUG_VAR(Evaluator::Flags)
     std::cerr.unsetf(std::ios::hex);
     EIGEN_DEBUG_VAR(InnerMaxSize)
     EIGEN_DEBUG_VAR(PacketSize)
@@ -87,88 +87,88 @@ public:
 
 /*** no vectorization ***/
 
-template<typename Func, typename Derived, int Start, int Length>
+template<typename Func, typename Evaluator, int Start, int Length>
 struct redux_novec_unroller
 {
   enum {
     HalfLength = Length/2
   };
 
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Evaluator::Scalar Scalar;
 
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+  static EIGEN_STRONG_INLINE Scalar run(const Evaluator &eval, const Func& func)
   {
-    return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
-                redux_novec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func));
+    return func(redux_novec_unroller<Func, Evaluator, Start, HalfLength>::run(eval,func),
+                redux_novec_unroller<Func, Evaluator, Start+HalfLength, Length-HalfLength>::run(eval,func));
   }
 };
 
-template<typename Func, typename Derived, int Start>
-struct redux_novec_unroller<Func, Derived, Start, 1>
+template<typename Func, typename Evaluator, int Start>
+struct redux_novec_unroller<Func, Evaluator, Start, 1>
 {
   enum {
-    outer = Start / Derived::InnerSizeAtCompileTime,
-    inner = Start % Derived::InnerSizeAtCompileTime
+    outer = Start / Evaluator::InnerSizeAtCompileTime,
+    inner = Start % Evaluator::InnerSizeAtCompileTime
   };
 
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Evaluator::Scalar Scalar;
 
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
+  static EIGEN_STRONG_INLINE Scalar run(const Evaluator &eval, const Func&)
   {
-    return mat.coeffByOuterInner(outer, inner);
+    return eval.coeffByOuterInner(outer, inner);
   }
 };
 
 // This is actually dead code and will never be called. It is required
 // to prevent false warnings regarding failed inlining though
 // for 0 length run() will never be called at all.
-template<typename Func, typename Derived, int Start>
-struct redux_novec_unroller<Func, Derived, Start, 0>
+template<typename Func, typename Evaluator, int Start>
+struct redux_novec_unroller<Func, Evaluator, Start, 0>
 {
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Evaluator::Scalar Scalar;
   EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
+  static EIGEN_STRONG_INLINE Scalar run(const Evaluator&, const Func&) { return Scalar(); }
 };
 
 /*** vectorization ***/
 
-template<typename Func, typename Derived, int Start, int Length>
+template<typename Func, typename Evaluator, int Start, int Length>
 struct redux_vec_unroller
 {
   enum {
-    PacketSize = redux_traits<Func, Derived>::PacketSize,
+    PacketSize = redux_traits<Func, Evaluator>::PacketSize,
     HalfLength = Length/2
   };
 
-  typedef typename Derived::Scalar Scalar;
-  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+  typedef typename Evaluator::Scalar Scalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
 
-  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func)
+  static EIGEN_STRONG_INLINE PacketScalar run(const Evaluator &eval, const Func& func)
   {
     return func.packetOp(
-            redux_vec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
-            redux_vec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func) );
+            redux_vec_unroller<Func, Evaluator, Start, HalfLength>::run(eval,func),
+            redux_vec_unroller<Func, Evaluator, Start+HalfLength, Length-HalfLength>::run(eval,func) );
   }
 };
 
-template<typename Func, typename Derived, int Start>
-struct redux_vec_unroller<Func, Derived, Start, 1>
+template<typename Func, typename Evaluator, int Start>
+struct redux_vec_unroller<Func, Evaluator, Start, 1>
 {
   enum {
-    index = Start * redux_traits<Func, Derived>::PacketSize,
-    outer = index / int(Derived::InnerSizeAtCompileTime),
-    inner = index % int(Derived::InnerSizeAtCompileTime),
-    alignment = Derived::Alignment
+    index = Start * redux_traits<Func, Evaluator>::PacketSize,
+    outer = index / int(Evaluator::InnerSizeAtCompileTime),
+    inner = index % int(Evaluator::InnerSizeAtCompileTime),
+    alignment = Evaluator::Alignment
   };
 
-  typedef typename Derived::Scalar Scalar;
-  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+  typedef typename Evaluator::Scalar Scalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
 
-  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
+  static EIGEN_STRONG_INLINE PacketScalar run(const Evaluator &eval, const Func&)
   {
-    return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
+    return eval.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
   }
 };
 
@@ -176,53 +176,54 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
 * Part 3 : implementation of all cases
 ***************************************************************************/
 
-template<typename Func, typename Derived,
-         int Traversal = redux_traits<Func, Derived>::Traversal,
-         int Unrolling = redux_traits<Func, Derived>::Unrolling
+template<typename Func, typename Evaluator,
+         int Traversal = redux_traits<Func, Evaluator>::Traversal,
+         int Unrolling = redux_traits<Func, Evaluator>::Unrolling
 >
 struct redux_impl;
 
-template<typename Func, typename Derived>
-struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
+template<typename Func, typename Evaluator>
+struct redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>
 {
-  typedef typename Derived::Scalar Scalar;
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+  typedef typename Evaluator::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
+  Scalar run(const Evaluator &eval, const Func& func)
   {
-    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
+    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
     Scalar res;
-    res = mat.coeffByOuterInner(0, 0);
-    for(Index i = 1; i < mat.innerSize(); ++i)
-      res = func(res, mat.coeffByOuterInner(0, i));
-    for(Index i = 1; i < mat.outerSize(); ++i)
-      for(Index j = 0; j < mat.innerSize(); ++j)
-        res = func(res, mat.coeffByOuterInner(i, j));
+    res = eval.coeffByOuterInner(0, 0);
+    for(Index i = 1; i < eval.innerSize(); ++i)
+      res = func(res, eval.coeffByOuterInner(0, i));
+    for(Index i = 1; i < eval.outerSize(); ++i)
+      for(Index j = 0; j < eval.innerSize(); ++j)
+        res = func(res, eval.coeffByOuterInner(i, j));
     return res;
   }
 };
 
-template<typename Func, typename Derived>
-struct redux_impl<Func,Derived, DefaultTraversal, CompleteUnrolling>
-  : public redux_novec_unroller<Func,Derived, 0, Derived::SizeAtCompileTime>
+template<typename Func, typename Evaluator>
+struct redux_impl<Func,Evaluator, DefaultTraversal, CompleteUnrolling>
+  : redux_novec_unroller<Func,Evaluator, 0, Evaluator::SizeAtCompileTime>
 {};
 
-template<typename Func, typename Derived>
-struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
+template<typename Func, typename Evaluator>
+struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, NoUnrolling>
 {
-  typedef typename Derived::Scalar Scalar;
-  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+  typedef typename Evaluator::Scalar Scalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
 
-  static Scalar run(const Derived &mat, const Func& func)
+  static Scalar run(const Evaluator &eval, const Func& func)
   {
-    const Index size = mat.size();
+    const Index size = eval.size();
     
-    const Index packetSize = redux_traits<Func, Derived>::PacketSize;
+    const Index packetSize = redux_traits<Func, Evaluator>::PacketSize;
     const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
     enum {
-      alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
-      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment)
+      alignment0 = (bool(Evaluator::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
+      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Evaluator::Alignment)
     };
-    const Index alignedStart = internal::first_default_aligned(mat.nestedExpression());
+    const Index alignedStart = internal::first_default_aligned(eval.nestedExpression());
     const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
     const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
     const Index alignedEnd2 = alignedStart + alignedSize2;
@@ -230,34 +231,34 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(alignedSize)
     {
-      PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart);
+      PacketScalar packet_res0 = eval.template packet<alignment,PacketScalar>(alignedStart);
       if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
       {
-        PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize);
+        PacketScalar packet_res1 = eval.template packet<alignment,PacketScalar>(alignedStart+packetSize);
         for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
         {
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index));
-          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize));
+          packet_res0 = func.packetOp(packet_res0, eval.template packet<alignment,PacketScalar>(index));
+          packet_res1 = func.packetOp(packet_res1, eval.template packet<alignment,PacketScalar>(index+packetSize));
         }
 
         packet_res0 = func.packetOp(packet_res0,packet_res1);
         if(alignedEnd>alignedEnd2)
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2));
+          packet_res0 = func.packetOp(packet_res0, eval.template packet<alignment,PacketScalar>(alignedEnd2));
       }
       res = func.predux(packet_res0);
 
       for(Index index = 0; index < alignedStart; ++index)
-        res = func(res,mat.coeff(index));
+        res = func(res,eval.coeff(index));
 
       for(Index index = alignedEnd; index < size; ++index)
-        res = func(res,mat.coeff(index));
+        res = func(res,eval.coeff(index));
     }
     else // too small to vectorize anything.
          // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
     {
-      res = mat.coeff(0);
+      res = eval.coeff(0);
       for(Index index = 1; index < size; ++index)
-        res = func(res,mat.coeff(index));
+        res = func(res,eval.coeff(index));
     }
 
     return res;
@@ -265,77 +266,80 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
 };
 
 // NOTE: for SliceVectorizedTraversal we simply bypass unrolling
-template<typename Func, typename Derived, int Unrolling>
-struct redux_impl<Func, Derived, SliceVectorizedTraversal, Unrolling>
+template<typename Func, typename Evaluator, int Unrolling>
+struct redux_impl<Func, Evaluator, SliceVectorizedTraversal, Unrolling>
 {
-  typedef typename Derived::Scalar Scalar;
-  typedef typename redux_traits<Func, Derived>::PacketType PacketType;
+  typedef typename Evaluator::Scalar Scalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketType;
 
-  EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func)
+  EIGEN_DEVICE_FUNC static Scalar run(const Evaluator &eval, const Func& func)
   {
-    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
-    const Index innerSize = mat.innerSize();
-    const Index outerSize = mat.outerSize();
+    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
+    const Index innerSize = eval.innerSize();
+    const Index outerSize = eval.outerSize();
     enum {
-      packetSize = redux_traits<Func, Derived>::PacketSize
+      packetSize = redux_traits<Func, Evaluator>::PacketSize
     };
     const Index packetedInnerSize = ((innerSize)/packetSize)*packetSize;
     Scalar res;
     if(packetedInnerSize)
     {
-      PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0);
+      PacketType packet_res = eval.template packet<Unaligned,PacketType>(0,0);
       for(Index j=0; j<outerSize; ++j)
         for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
-          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i));
+          packet_res = func.packetOp(packet_res, eval.template packetByOuterInner<Unaligned,PacketType>(j,i));
 
       res = func.predux(packet_res);
       for(Index j=0; j<outerSize; ++j)
         for(Index i=packetedInnerSize; i<innerSize; ++i)
-          res = func(res, mat.coeffByOuterInner(j,i));
+          res = func(res, eval.coeffByOuterInner(j,i));
     }
     else // too small to vectorize anything.
          // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
     {
-      res = redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>::run(mat, func);
+      res = redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>::run(eval, func);
     }
 
     return res;
   }
 };
 
-template<typename Func, typename Derived>
-struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
+template<typename Func, typename Evaluator>
+struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, CompleteUnrolling>
 {
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Evaluator::Scalar Scalar;
 
-  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
   enum {
-    PacketSize = redux_traits<Func, Derived>::PacketSize,
-    Size = Derived::SizeAtCompileTime,
+    PacketSize = redux_traits<Func, Evaluator>::PacketSize,
+    Size = Evaluator::SizeAtCompileTime,
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
+  Scalar run(const Evaluator &eval, const Func& func)
   {
-    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
+    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
     if (VectorizedSize > 0) {
-      Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
+      Scalar res = func.predux(redux_vec_unroller<Func, Evaluator, 0, Size / PacketSize>::run(eval,func));
       if (VectorizedSize != Size)
-        res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
+        res = func(res,redux_novec_unroller<Func, Evaluator, VectorizedSize, Size-VectorizedSize>::run(eval,func));
       return res;
     }
     else {
-      return redux_novec_unroller<Func, Derived, 0, Size>::run(mat,func);
+      return redux_novec_unroller<Func, Evaluator, 0, Size>::run(eval,func);
     }
   }
 };
 
 // evaluator adaptor
 template<typename _XprType>
-class redux_evaluator
+class redux_evaluator : public internal::evaluator<_XprType>
 {
+  typedef internal::evaluator<_XprType> Base;
 public:
   typedef _XprType XprType;
-  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : Base(xpr), m_xpr(xpr) {}
   
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
@@ -346,12 +350,10 @@ public:
     MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
     // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
-    Flags = evaluator<XprType>::Flags & ~DirectAccessBit,
+    Flags = Base::Flags & ~DirectAccessBit,
     IsRowMajor = XprType::IsRowMajor,
     SizeAtCompileTime = XprType::SizeAtCompileTime,
-    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
-    CoeffReadCost = evaluator<XprType>::CoeffReadCost,
-    Alignment = evaluator<XprType>::Alignment
+    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime
   };
   
   EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
@@ -359,35 +361,18 @@ public:
   EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
   EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
   EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
-
-  EIGEN_DEVICE_FUNC
-  CoeffReturnType coeff(Index row, Index col) const
-  { return m_evaluator.coeff(row, col); }
-
-  EIGEN_DEVICE_FUNC
-  CoeffReturnType coeff(Index index) const
-  { return m_evaluator.coeff(index); }
-
-  template<int LoadMode, typename PacketType>
-  PacketType packet(Index row, Index col) const
-  { return m_evaluator.template packet<LoadMode,PacketType>(row, col); }
-
-  template<int LoadMode, typename PacketType>
-  PacketType packet(Index index) const
-  { return m_evaluator.template packet<LoadMode,PacketType>(index); }
   
   EIGEN_DEVICE_FUNC
   CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
-  { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  { return Base::coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
   
   template<int LoadMode, typename PacketType>
   PacketType packetByOuterInner(Index outer, Index inner) const
-  { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  { return Base::template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
   
   const XprType & nestedExpression() const { return m_xpr; }
   
 protected:
-  internal::evaluator<XprType> m_evaluator;
   const XprType &m_xpr;
 };
 

From 047677a08d907031bfa596017c15493c982181c5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 2 Jul 2018 12:18:25 +0200
Subject: [PATCH 389/680] Fix regression in changeset
 f05dea6b2326836e5e0243fbaffbece84b833d64 : computeFromHessenberg can take any
 expression for matrixQ, not only an HouseholderSequence.

---
 Eigen/src/Eigenvalues/RealSchur.h | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h
index 4c53344bb..aca8a8279 100644
--- a/Eigen/src/Eigenvalues/RealSchur.h
+++ b/Eigen/src/Eigenvalues/RealSchur.h
@@ -270,8 +270,13 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const EigenBase<InputType>
   // Step 1. Reduce to Hessenberg form
   m_hess.compute(matrix.derived()/scale);
 
-  // Step 2. Reduce to real Schur form  
-  computeFromHessenberg(m_hess.matrixH(), m_hess.matrixQ(), computeU);
+  // Step 2. Reduce to real Schur form
+  // Note: we copy m_hess.matrixQ() into m_matU here and not in computeFromHessenberg
+  //       to be able to pass our working-space buffer for the Householder to Dense evaluation.
+  m_workspaceVector.resize(matrix.cols());
+  if(computeU)
+    m_hess.matrixQ().evalTo(m_matU, m_workspaceVector);
+  computeFromHessenberg(m_hess.matrixH(), m_matU, computeU);
 
   m_matT *= scale;
   
@@ -285,8 +290,8 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa
 
   m_matT = matrixH;
   m_workspaceVector.resize(m_matT.cols());
-  if(computeU)
-    matrixQ.evalTo(m_matU, m_workspaceVector);
+  if(computeU && !internal::is_same_dense(m_matU,matrixQ))
+    m_matU = matrixQ;
   
   Index maxIters = m_maxIters;
   if (maxIters == -1)

From 67ec37f7b0d9ccaaede508109e50528138eaf4d8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 2 Jul 2018 12:54:14 +0200
Subject: [PATCH 390/680] Activate dgmres unit test

---
 unsupported/test/CMakeLists.txt | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index e99eab0e3..fa4e76e42 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -124,6 +124,7 @@ ei_add_test(polynomialsolver)
 ei_add_test(polynomialutils)
 ei_add_test(splines)
 ei_add_test(gmres)
+ei_add_test(dgmres)
 ei_add_test(minres)
 ei_add_test(levenberg_marquardt)
 ei_add_test(kronecker_product)

From 3ae2083e230ec9517cb9b40d1137dec7ba4b2b17 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Jul 2018 13:21:43 +0200
Subject: [PATCH 391/680] Make is_same_dense compatible with different scalar
 types.

---
 Eigen/src/Core/util/XprHelper.h | 9 +++++++--
 test/is_same_dense.cpp          | 8 ++++++++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 16714fa5c..f7a3d9ce7 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -680,13 +680,18 @@ template<typename S1, typename S2> struct glue_shapes;
 template<> struct glue_shapes<DenseShape,TriangularShape> { typedef TriangularShape type;  };
 
 template<typename T1, typename T2>
-bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<has_direct_access<T1>::ret&&has_direct_access<T2>::ret, T1>::type * = 0)
+struct possibly_same_dense {
+  enum { value = has_direct_access<T1>::ret && has_direct_access<T2>::ret && is_same<typename T1::Scalar,typename T2::Scalar>::value };
+};
+
+template<typename T1, typename T2>
+bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<possibly_same_dense<T1,T2>::value>::type * = 0)
 {
   return (mat1.data()==mat2.data()) && (mat1.innerStride()==mat2.innerStride()) && (mat1.outerStride()==mat2.outerStride());
 }
 
 template<typename T1, typename T2>
-bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_access<T1>::ret&&has_direct_access<T2>::ret), T1>::type * = 0)
+bool is_same_dense(const T1 &, const T2 &, typename enable_if<!possibly_same_dense<T1,T2>::value>::type * = 0)
 {
   return false;
 }
diff --git a/test/is_same_dense.cpp b/test/is_same_dense.cpp
index 2c7838ce9..c4e2fbc92 100644
--- a/test/is_same_dense.cpp
+++ b/test/is_same_dense.cpp
@@ -14,9 +14,13 @@ using internal::is_same_dense;
 void test_is_same_dense()
 {
   typedef Matrix<double,Dynamic,Dynamic,ColMajor> ColMatrixXd;
+  typedef Matrix<std::complex<double>,Dynamic,Dynamic,ColMajor> ColMatrixXcd;
   ColMatrixXd m1(10,10);
+  ColMatrixXcd m2(10,10);
   Ref<ColMatrixXd> ref_m1(m1);
+  Ref<ColMatrixXd,0, Stride<Dynamic,Dynamic> >  ref_m2_real(m2.real());
   Ref<const ColMatrixXd> const_ref_m1(m1);
+
   VERIFY(is_same_dense(m1,m1));
   VERIFY(is_same_dense(m1,ref_m1));
   VERIFY(is_same_dense(const_ref_m1,m1));
@@ -30,4 +34,8 @@ void test_is_same_dense()
   
   Ref<const ColMatrixXd> const_ref_m1_col(m1.col(1));
   VERIFY(is_same_dense(m1.col(1),const_ref_m1_col));
+
+
+  VERIFY(!is_same_dense(m1, ref_m2_real));
+  VERIFY(!is_same_dense(m2, ref_m2_real));
 }

From 6a241bd8ee45df0ba112d4d6874888499b51cd34 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 3 Jul 2018 14:02:46 +0200
Subject: [PATCH 392/680] Implement custom inplace triangular product to avoid
 a temporary

---
 Eigen/src/Householder/BlockHouseholder.h | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Householder/BlockHouseholder.h b/Eigen/src/Householder/BlockHouseholder.h
index 01a7ed188..39ce1c2a0 100644
--- a/Eigen/src/Householder/BlockHouseholder.h
+++ b/Eigen/src/Householder/BlockHouseholder.h
@@ -63,8 +63,15 @@ void make_block_householder_triangular_factor(TriangularFactorType& triFactor, c
       triFactor.row(i).tail(rt).noalias() = -hCoeffs(i) * vectors.col(i).tail(rs).adjoint()
                                                         * vectors.bottomRightCorner(rs, rt).template triangularView<UnitLower>();
             
-      // FIXME add .noalias() once the triangular product can work inplace
-      triFactor.row(i).tail(rt) = triFactor.row(i).tail(rt) * triFactor.bottomRightCorner(rt,rt).template triangularView<Upper>();
+      // FIXME use the following line with .noalias() once the triangular product can work inplace
+      // triFactor.row(i).tail(rt) = triFactor.row(i).tail(rt) * triFactor.bottomRightCorner(rt,rt).template triangularView<Upper>();
+      for(Index j=nbVecs-1; j>i; --j)
+      {
+        typename TriangularFactorType::Scalar z = triFactor(i,j);
+        triFactor(i,j) = z * triFactor(j,j);
+        if(nbVecs-j-1>0)
+          triFactor.row(i).tail(nbVecs-j-1) += z * triFactor.row(j).tail(nbVecs-j-1);
+      }
       
     }
     triFactor(i,i) = hCoeffs(i);

From 05371239533012e652de0b88a3e0aa992a48a80f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 5 Jul 2018 09:21:26 +0200
Subject: [PATCH 393/680] bug #1565: help MSVC to generatenot too bad ASM in
 reductions.

---
 Eigen/src/Core/Redux.h | 62 ++++++++++++++++++++++--------------------
 1 file changed, 33 insertions(+), 29 deletions(-)

diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index 32574ba60..ddce65468 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -187,16 +187,17 @@ struct redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>
 {
   typedef typename Evaluator::Scalar Scalar;
 
+  template<typename XprType>
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
-  Scalar run(const Evaluator &eval, const Func& func)
+  Scalar run(const Evaluator &eval, const Func& func, const XprType& xpr)
   {
-    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
+    eigen_assert(xpr.rows()>0 && xpr.cols()>0 && "you are using an empty matrix");
     Scalar res;
     res = eval.coeffByOuterInner(0, 0);
-    for(Index i = 1; i < eval.innerSize(); ++i)
+    for(Index i = 1; i < xpr.innerSize(); ++i)
       res = func(res, eval.coeffByOuterInner(0, i));
-    for(Index i = 1; i < eval.outerSize(); ++i)
-      for(Index j = 0; j < eval.innerSize(); ++j)
+    for(Index i = 1; i < xpr.outerSize(); ++i)
+      for(Index j = 0; j < xpr.innerSize(); ++j)
         res = func(res, eval.coeffByOuterInner(i, j));
     return res;
   }
@@ -205,7 +206,16 @@ struct redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>
 template<typename Func, typename Evaluator>
 struct redux_impl<Func,Evaluator, DefaultTraversal, CompleteUnrolling>
   : redux_novec_unroller<Func,Evaluator, 0, Evaluator::SizeAtCompileTime>
-{};
+{
+  typedef redux_novec_unroller<Func,Evaluator, 0, Evaluator::SizeAtCompileTime> Base;
+  typedef typename Evaluator::Scalar Scalar;
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
+  Scalar run(const Evaluator &eval, const Func& func, const XprType& /*xpr*/)
+  {
+    return Base::run(eval,func);
+  }
+};
 
 template<typename Func, typename Evaluator>
 struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, NoUnrolling>
@@ -213,9 +223,10 @@ struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, NoUnrolling>
   typedef typename Evaluator::Scalar Scalar;
   typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
 
-  static Scalar run(const Evaluator &eval, const Func& func)
+  template<typename XprType>
+  static Scalar run(const Evaluator &eval, const Func& func, const XprType& xpr)
   {
-    const Index size = eval.size();
+    const Index size = xpr.size();
     
     const Index packetSize = redux_traits<Func, Evaluator>::PacketSize;
     const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
@@ -223,7 +234,7 @@ struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, NoUnrolling>
       alignment0 = (bool(Evaluator::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
       alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Evaluator::Alignment)
     };
-    const Index alignedStart = internal::first_default_aligned(eval.nestedExpression());
+    const Index alignedStart = internal::first_default_aligned(xpr);
     const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
     const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
     const Index alignedEnd2 = alignedStart + alignedSize2;
@@ -272,11 +283,12 @@ struct redux_impl<Func, Evaluator, SliceVectorizedTraversal, Unrolling>
   typedef typename Evaluator::Scalar Scalar;
   typedef typename redux_traits<Func, Evaluator>::PacketType PacketType;
 
-  EIGEN_DEVICE_FUNC static Scalar run(const Evaluator &eval, const Func& func)
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC static Scalar run(const Evaluator &eval, const Func& func, const XprType& xpr)
   {
-    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
-    const Index innerSize = eval.innerSize();
-    const Index outerSize = eval.outerSize();
+    eigen_assert(xpr.rows()>0 && xpr.cols()>0 && "you are using an empty matrix");
+    const Index innerSize = xpr.innerSize();
+    const Index outerSize = xpr.outerSize();
     enum {
       packetSize = redux_traits<Func, Evaluator>::PacketSize
     };
@@ -297,7 +309,7 @@ struct redux_impl<Func, Evaluator, SliceVectorizedTraversal, Unrolling>
     else // too small to vectorize anything.
          // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
     {
-      res = redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>::run(eval, func);
+      res = redux_impl<Func, Evaluator, DefaultTraversal, NoUnrolling>::run(eval, func, xpr);
     }
 
     return res;
@@ -316,10 +328,11 @@ struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, CompleteUnrolling>
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
 
+  template<typename XprType>
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
-  Scalar run(const Evaluator &eval, const Func& func)
+  Scalar run(const Evaluator &eval, const Func& func, const XprType &xpr)
   {
-    eigen_assert(eval.rows()>0 && eval.cols()>0 && "you are using an empty matrix");
+    eigen_assert(xpr.rows()>0 && xpr.cols()>0 && "you are using an empty matrix");
     if (VectorizedSize > 0) {
       Scalar res = func.predux(redux_vec_unroller<Func, Evaluator, 0, Size / PacketSize>::run(eval,func));
       if (VectorizedSize != Size)
@@ -339,12 +352,11 @@ class redux_evaluator : public internal::evaluator<_XprType>
   typedef internal::evaluator<_XprType> Base;
 public:
   typedef _XprType XprType;
-  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : Base(xpr), m_xpr(xpr) {}
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : Base(xpr) {}
   
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename XprType::PacketScalar PacketScalar;
-  typedef typename XprType::PacketReturnType PacketReturnType;
   
   enum {
     MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
@@ -356,12 +368,6 @@ public:
     InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime
   };
   
-  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
-  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
-  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
-  EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
-  EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
-  
   EIGEN_DEVICE_FUNC
   CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
   { return Base::coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
@@ -370,10 +376,6 @@ public:
   PacketType packetByOuterInner(Index outer, Index inner) const
   { return Base::template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
   
-  const XprType & nestedExpression() const { return m_xpr; }
-  
-protected:
-  const XprType &m_xpr;
 };
 
 } // end namespace internal
@@ -400,7 +402,9 @@ DenseBase<Derived>::redux(const Func& func) const
   typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
   ThisEvaluator thisEval(derived());
   
-  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
+  // The initial expression is passed to the reducer as an additional argument instead of
+  // passing it as a member of redux_evaluator to help  
+  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func, derived());
 }
 
 /** \returns the minimum of all coefficients of \c *this.

From f7124b3e467363e45c3d906b7003f1520a5f804a Mon Sep 17 00:00:00 2001
From: Andrea Bocci <andrea.bocci@cern.ch>
Date: Mon, 11 Jun 2018 18:33:24 +0200
Subject: [PATCH 394/680] Extend CUDA support to matrix inversion and
 selfadjointeigensolver

---
 Eigen/Core                                    | 39 ++++++++-----
 Eigen/PardisoSupport                          |  0
 Eigen/src/Core/DenseStorage.h                 | 58 +++++++++++++------
 Eigen/src/Core/GeneralProduct.h               |  4 +-
 Eigen/src/Core/MathFunctions.h                | 31 ++++++++--
 Eigen/src/Core/MathFunctionsImpl.h            |  5 +-
 Eigen/src/Core/MatrixBase.h                   |  8 +++
 Eigen/src/Core/NumTraits.h                    |  6 ++
 Eigen/src/Core/PermutationMatrix.h            |  6 +-
 Eigen/src/Core/Product.h                      |  2 +-
 Eigen/src/Core/ProductEvaluators.h            |  3 +-
 Eigen/src/Core/Transpose.h                    |  1 +
 Eigen/src/Core/TriangularMatrix.h             |  2 +
 .../Core/products/SelfadjointMatrixVector.h   | 12 ++--
 .../Core/products/SelfadjointRank2Update.h    |  3 +-
 Eigen/src/Core/util/BlasUtil.h                |  8 +--
 Eigen/src/Core/util/Memory.h                  |  4 +-
 Eigen/src/Core/util/Meta.h                    |  5 +-
 Eigen/src/Core/util/XprHelper.h               |  2 +
 .../src/Eigenvalues/SelfAdjointEigenSolver.h  | 16 +++--
 Eigen/src/Eigenvalues/Tridiagonalization.h    |  9 ++-
 Eigen/src/Householder/Householder.h           |  4 ++
 Eigen/src/Householder/HouseholderSequence.h   | 17 +++++-
 Eigen/src/Jacobi/Jacobi.h                     | 11 +++-
 Eigen/src/LU/Determinant.h                    | 15 +++--
 Eigen/src/LU/InverseImpl.h                    |  2 +
 Eigen/src/plugins/BlockMethods.h              |  1 +
 27 files changed, 200 insertions(+), 74 deletions(-)
 mode change 100755 => 100644 Eigen/PardisoSupport

diff --git a/Eigen/Core b/Eigen/Core
index f6bc18a08..5117461c7 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -45,27 +45,40 @@
   #ifdef EIGEN_EXCEPTIONS
   #undef EIGEN_EXCEPTIONS
   #endif
+#endif
 
-  // All functions callable from CUDA code must be qualified with __device__
-  #ifdef EIGEN_CUDACC
-    // Do not try to vectorize on CUDA and SYCL!
-    #ifndef EIGEN_DONT_VECTORIZE
-    #define EIGEN_DONT_VECTORIZE
+// All functions callable from CUDA code must be qualified with __device__
+#ifdef EIGEN_CUDACC
+  // Do not try to vectorize on CUDA and SYCL!
+  #ifndef EIGEN_DONT_VECTORIZE
+  #define EIGEN_DONT_VECTORIZE
+  #endif
+
+  #define EIGEN_DEVICE_FUNC __host__ __device__
+  // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
+  // works properly on the device side
+  #include <cuda_runtime.h>
+
+  #if EIGEN_HAS_CONSTEXPR
+    // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules
+    #if defined(__NVCC__)
+      // nvcc considers constexpr functions as __host__ __device__ with the option --expt-relaxed-constexpr
+      #ifdef __CUDACC_RELAXED_CONSTEXPR__
+        #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
+      #endif
+    #elif defined(__clang__) && defined(__CUDA__)
+      // clang++ always considers constexpr functions as implicitly __host__ __device__
+      #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
     #endif
-
-    #define EIGEN_DEVICE_FUNC __host__ __device__
-    // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
-    // works properly on the device side
-    #include <cuda_runtime.h>
-  #else
-    #define EIGEN_DEVICE_FUNC
   #endif
 #else
   #define EIGEN_DEVICE_FUNC
 #endif
 
 #ifdef __NVCC__
-#define EIGEN_DONT_VECTORIZE
+  #ifndef EIGEN_DONT_VECTORIZE
+  #define EIGEN_DONT_VECTORIZE
+  #endif
 #endif
 
 // When compiling CUDA device code with NVCC, pull in math functions from the
diff --git a/Eigen/PardisoSupport b/Eigen/PardisoSupport
old mode 100755
new mode 100644
diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h
index 9e58fbf88..3c02a1025 100644
--- a/Eigen/src/Core/DenseStorage.h
+++ b/Eigen/src/Core/DenseStorage.h
@@ -207,7 +207,9 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
       EIGEN_UNUSED_VARIABLE(rows);
       EIGEN_UNUSED_VARIABLE(cols);
     }
-    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) {
+      numext::swap(m_data, other.m_data);
+    }
     EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
     EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
     EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
@@ -267,7 +269,11 @@ template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic
     }
     EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {}
     EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
-    { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
+    {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_rows,other.m_rows);
+      numext::swap(m_cols,other.m_cols);
+    }
     EIGEN_DEVICE_FUNC Index rows() const {return m_rows;}
     EIGEN_DEVICE_FUNC Index cols() const {return m_cols;}
     EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
@@ -296,7 +302,11 @@ template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Si
       return *this; 
     }
     EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {}
-    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
+    {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_rows,other.m_rows);
+    }
     EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
     EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;}
     EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; }
@@ -325,11 +335,14 @@ template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Si
       return *this;
     }
     EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {}
-    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_cols,other.m_cols);
+    }
     EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;}
     EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
-    void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
-    void resize(Index, Index, Index cols) { m_cols = cols; }
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index, Index cols) { m_cols = cols; }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
     EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
@@ -381,16 +394,19 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
     EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
-      using std::swap;
-      swap(m_data, other.m_data);
-      swap(m_rows, other.m_rows);
-      swap(m_cols, other.m_cols);
+      numext::swap(m_data, other.m_data);
+      numext::swap(m_rows, other.m_rows);
+      numext::swap(m_cols, other.m_cols);
       return *this;
     }
 #endif
     EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
     EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
-    { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
+    {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_rows,other.m_rows);
+      numext::swap(m_cols,other.m_cols);
+    }
     EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
     EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
     void conservativeResize(Index size, Index rows, Index cols)
@@ -459,14 +475,16 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
     EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
-      using std::swap;
-      swap(m_data, other.m_data);
-      swap(m_cols, other.m_cols);
+      numext::swap(m_data, other.m_data);
+      numext::swap(m_cols, other.m_cols);
       return *this;
     }
 #endif
     EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
-    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_cols,other.m_cols);
+    }
     EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
     EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
     EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols)
@@ -533,14 +551,16 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
     EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
-      using std::swap;
-      swap(m_data, other.m_data);
-      swap(m_rows, other.m_rows);
+      numext::swap(m_data, other.m_data);
+      numext::swap(m_rows, other.m_rows);
       return *this;
     }
 #endif
     EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
-    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) {
+      numext::swap(m_data,other.m_data);
+      numext::swap(m_rows,other.m_rows);
+    }
     EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
     EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
     void conservativeResize(Index size, Index rows, Index)
diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 694f7cbde..bd2361e9a 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -163,13 +163,13 @@ template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vect
 template<typename Scalar,int Size,int MaxSize>
 struct gemv_static_vector_if<Scalar,Size,MaxSize,false>
 {
-  EIGEN_STRONG_INLINE  Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; }
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; }
 };
 
 template<typename Scalar,int Size>
 struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
 {
-  EIGEN_STRONG_INLINE Scalar* data() { return 0; }
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Scalar* data() { return 0; }
 };
 
 template<typename Scalar,int Size,int MaxSize>
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 954863c39..a5740334a 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -864,7 +864,7 @@ template<typename T> T generic_fast_tanh_float(const T& a_x);
 
 namespace numext {
 
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
+#if (!defined(EIGEN_CUDACC) || defined(EIGEN_CONSTEXPR_ARE_DEVICE_FUNC)) && !defined(__SYCL_DEVICE_ONLY__)
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
@@ -881,19 +881,16 @@ EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
   return max EIGEN_NOT_A_MACRO (x,y);
 }
 
-
 #elif defined(__SYCL_DEVICE_ONLY__)
 template<typename T>
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
 {
-
   return y < x ? y : x;
 }
 
 template<typename T>
 EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
 {
-
   return x < y ? y : x;
 }
 
@@ -937,7 +934,6 @@ EIGEN_ALWAYS_INLINE unsigned long maxi(const unsigned long& x, const unsigned lo
   return cl::sycl::max(x,y);
 }
 
-
 EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
 {
   return cl::sycl::fmin(x,y);
@@ -971,6 +967,19 @@ EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
 {
   return fminf(x, y);
 }
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE double mini(const double& x, const double& y)
+{
+  return fmin(x, y);
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE long double mini(const long double& x, const long double& y)
+{
+  return fminl(x, y);
+}
+
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
@@ -983,6 +992,18 @@ EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
 {
   return fmaxf(x, y);
 }
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE double maxi(const double& x, const double& y)
+{
+  return fmax(x, y);
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE long double maxi(const long double& x, const long double& y)
+{
+  return fmaxl(x, y);
+}
 #endif
 
 
diff --git a/Eigen/src/Core/MathFunctionsImpl.h b/Eigen/src/Core/MathFunctionsImpl.h
index 28737c15e..a23e93ccb 100644
--- a/Eigen/src/Core/MathFunctionsImpl.h
+++ b/Eigen/src/Core/MathFunctionsImpl.h
@@ -67,7 +67,7 @@ T generic_fast_tanh_float(const T& a_x)
 }
 
 template<typename RealScalar>
-EIGEN_STRONG_INLINE
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 RealScalar positive_real_hypot(const RealScalar& x, const RealScalar& y)
 {
   EIGEN_USING_STD_MATH(sqrt);
@@ -82,7 +82,8 @@ template<typename Scalar>
 struct hypot_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  static inline RealScalar run(const Scalar& x, const Scalar& y)
+  static EIGEN_DEVICE_FUNC
+  inline RealScalar run(const Scalar& x, const Scalar& y)
   {
     EIGEN_USING_STD_MATH(abs);
     return positive_real_hypot<RealScalar>(abs(x), abs(y));
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 11435903b..6046c8bae 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -328,6 +328,7 @@ template<typename Derived> class MatrixBase
 
     inline const PartialPivLU<PlainObject> lu() const;
 
+    EIGEN_DEVICE_FUNC
     inline const Inverse<Derived> inverse() const;
 
     template<typename ResultType>
@@ -337,12 +338,15 @@ template<typename Derived> class MatrixBase
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
     ) const;
+
     template<typename ResultType>
     inline void computeInverseWithCheck(
       ResultType& inverse,
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
     ) const;
+
+    EIGEN_DEVICE_FUNC
     Scalar determinant() const;
 
 /////////// Cholesky module ///////////
@@ -414,15 +418,19 @@ template<typename Derived> class MatrixBase
 
 ////////// Householder module ///////////
 
+    EIGEN_DEVICE_FUNC
     void makeHouseholderInPlace(Scalar& tau, RealScalar& beta);
     template<typename EssentialPart>
+    EIGEN_DEVICE_FUNC
     void makeHouseholder(EssentialPart& essential,
                          Scalar& tau, RealScalar& beta) const;
     template<typename EssentialPart>
+    EIGEN_DEVICE_FUNC
     void applyHouseholderOnTheLeft(const EssentialPart& essential,
                                    const Scalar& tau,
                                    Scalar* workspace);
     template<typename EssentialPart>
+    EIGEN_DEVICE_FUNC
     void applyHouseholderOnTheRight(const EssentialPart& essential,
                                     const Scalar& tau,
                                     Scalar* workspace);
diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h
index 5567d4c90..b053cff07 100644
--- a/Eigen/src/Core/NumTraits.h
+++ b/Eigen/src/Core/NumTraits.h
@@ -21,12 +21,14 @@ template< typename T,
           bool is_integer = NumTraits<T>::IsInteger>
 struct default_digits10_impl
 {
+  EIGEN_DEVICE_FUNC
   static int run() { return std::numeric_limits<T>::digits10; }
 };
 
 template<typename T>
 struct default_digits10_impl<T,false,false> // Floating point
 {
+  EIGEN_DEVICE_FUNC
   static int run() {
     using std::log10;
     using std::ceil;
@@ -38,6 +40,7 @@ struct default_digits10_impl<T,false,false> // Floating point
 template<typename T>
 struct default_digits10_impl<T,false,true> // Integer
 {
+  EIGEN_DEVICE_FUNC
   static int run() { return 0; }
 };
 
@@ -49,12 +52,14 @@ template< typename T,
           bool is_integer = NumTraits<T>::IsInteger>
 struct default_digits_impl
 {
+  EIGEN_DEVICE_FUNC
   static int run() { return std::numeric_limits<T>::digits; }
 };
 
 template<typename T>
 struct default_digits_impl<T,false,false> // Floating point
 {
+  EIGEN_DEVICE_FUNC
   static int run() {
     using std::log;
     using std::ceil;
@@ -66,6 +71,7 @@ struct default_digits_impl<T,false,false> // Floating point
 template<typename T>
 struct default_digits_impl<T,false,true> // Integer
 {
+  EIGEN_DEVICE_FUNC
   static int run() { return 0; }
 };
 
diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h
index b1fb455b9..acd085301 100644
--- a/Eigen/src/Core/PermutationMatrix.h
+++ b/Eigen/src/Core/PermutationMatrix.h
@@ -99,13 +99,13 @@ class PermutationBase : public EigenBase<Derived>
     #endif
 
     /** \returns the number of rows */
-    inline Index rows() const { return Index(indices().size()); }
+    inline EIGEN_DEVICE_FUNC Index rows() const { return Index(indices().size()); }
 
     /** \returns the number of columns */
-    inline Index cols() const { return Index(indices().size()); }
+    inline EIGEN_DEVICE_FUNC Index cols() const { return Index(indices().size()); }
 
     /** \returns the size of a side of the respective square matrix, i.e., the number of indices */
-    inline Index size() const { return Index(indices().size()); }
+    inline EIGEN_DEVICE_FUNC Index size() const { return Index(indices().size()); }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename DenseDerived>
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index 3d67d9489..70790dbd4 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -127,7 +127,7 @@ public:
   using Base::derived;
   typedef typename Base::Scalar Scalar;
   
-  EIGEN_STRONG_INLINE operator const Scalar() const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE operator const Scalar() const
   {
     return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
   }
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 8072a1959..76e5083c1 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -767,7 +767,8 @@ struct generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   template<typename Dest>
-  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC
+  void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
     selfadjoint_product_impl<typename Lhs::MatrixType,Lhs::Mode,false,Rhs,0,Rhs::IsVectorAtCompileTime>::run(dst, lhs.nestedExpression(), rhs, alpha);
   }
diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h
index ba7d6e629..d7c204579 100644
--- a/Eigen/src/Core/Transpose.h
+++ b/Eigen/src/Core/Transpose.h
@@ -79,6 +79,7 @@ template<typename MatrixType> class Transpose
     nestedExpression() { return m_matrix; }
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     void resize(Index nrows, Index ncols) {
       m_matrix.resize(ncols,nrows);
     }
diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h
index ab73fcf21..521de6160 100644
--- a/Eigen/src/Core/TriangularMatrix.h
+++ b/Eigen/src/Core/TriangularMatrix.h
@@ -65,6 +65,7 @@ template<typename Derived> class TriangularBase : public EigenBase<Derived>
     inline Index innerStride() const { return derived().innerStride(); }
     
     // dummy resize function
+    EIGEN_DEVICE_FUNC
     void resize(Index rows, Index cols)
     {
       EIGEN_UNUSED_VARIABLE(rows);
@@ -716,6 +717,7 @@ struct unary_evaluator<TriangularView<MatrixType,Mode>, IndexBased>
 {
   typedef TriangularView<MatrixType,Mode> XprType;
   typedef evaluator<typename internal::remove_all<MatrixType>::type> Base;
+  EIGEN_DEVICE_FUNC
   unary_evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {}
 };
 
diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h
index 67390f1d7..d38fd72b2 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixVector.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h
@@ -27,7 +27,8 @@ template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool Conju
 struct selfadjoint_matrix_vector_product
 
 {
-static EIGEN_DONT_INLINE void run(
+static EIGEN_DONT_INLINE EIGEN_DEVICE_FUNC
+void run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
   const Scalar*  rhs,
@@ -36,7 +37,8 @@ static EIGEN_DONT_INLINE void run(
 };
 
 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
-EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
+EIGEN_DONT_INLINE EIGEN_DEVICE_FUNC
+void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
   const Scalar*  rhs,
@@ -62,8 +64,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 
   Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha;
 
-
-  Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
+  Index bound = numext::maxi(Index(0), size-8) & 0xfffffffe;
   if (FirstTriangular)
     bound = size - bound;
 
@@ -175,7 +176,8 @@ struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
   enum { LhsUpLo = LhsMode&(Upper|Lower) };
 
   template<typename Dest>
-  static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC
+  void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
   {
     typedef typename Dest::Scalar ResScalar;
     typedef typename Rhs::Scalar RhsScalar;
diff --git a/Eigen/src/Core/products/SelfadjointRank2Update.h b/Eigen/src/Core/products/SelfadjointRank2Update.h
index d395888e5..09209f733 100644
--- a/Eigen/src/Core/products/SelfadjointRank2Update.h
+++ b/Eigen/src/Core/products/SelfadjointRank2Update.h
@@ -24,7 +24,8 @@ struct selfadjoint_rank2_update_selector;
 template<typename Scalar, typename Index, typename UType, typename VType>
 struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Lower>
 {
-  static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
+  static EIGEN_DEVICE_FUNC
+  void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
   {
     const Index size = u.size();
     for (Index i=0; i<size; ++i)
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index b1791fb3a..705925984 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -289,8 +289,8 @@ template<typename XprType> struct blas_traits
     ExtractType,
     typename _ExtractType::PlainObject
     >::type DirectLinearAccessType;
-  static inline ExtractType extract(const XprType& x) { return x; }
-  static inline const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
+  static inline EIGEN_DEVICE_FUNC ExtractType extract(const XprType& x) { return x; }
+  static inline EIGEN_DEVICE_FUNC const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
 };
 
 // pop conjugate
@@ -318,8 +318,8 @@ struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp
   typedef blas_traits<NestedXpr> Base;
   typedef CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
-  static inline ExtractType extract(const XprType& x) { return Base::extract(x.rhs()); }
-  static inline Scalar extractScalarFactor(const XprType& x)
+  static inline EIGEN_DEVICE_FUNC ExtractType extract(const XprType& x) { return Base::extract(x.rhs()); }
+  static inline EIGEN_DEVICE_FUNC Scalar extractScalarFactor(const XprType& x)
   { return x.lhs().functor().m_other * Base::extractScalarFactor(x.rhs()); }
 };
 template<typename Scalar, typename NestedXpr, typename Plain>
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 53300c388..22d7679c5 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -542,7 +542,7 @@ template<typename T> struct smart_memmove_helper<T,false> {
 
 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
 // to the appropriate stack allocation function
-#ifndef EIGEN_ALLOCA
+#if ! defined EIGEN_ALLOCA && ! defined EIGEN_CUDA_ARCH
   #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
     #define EIGEN_ALLOCA alloca
   #elif EIGEN_COMP_MSVC
@@ -561,12 +561,14 @@ template<typename T> class aligned_stack_memory_handler : noncopyable
      * In this case, the buffer elements will also be destructed when this handler will be destructed.
      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
      **/
+    EIGEN_DEVICE_FUNC
     aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
     {
       if(NumTraits<T>::RequireInitialization && m_ptr)
         Eigen::internal::construct_elements_of_array(m_ptr, size);
     }
+    EIGEN_DEVICE_FUNC
     ~aligned_stack_memory_handler()
     {
       if(NumTraits<T>::RequireInitialization && m_ptr)
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 0d0b8c43a..ef9860c4b 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -544,6 +544,7 @@ using std::numeric_limits;
 // Integer division with rounding up.
 // T is assumed to be an integer type with a>=0, and b>0
 template<typename T>
+EIGEN_DEVICE_FUNC
 T div_ceil(const T &a, const T &b)
 {
   return (a+b-1) / b;
@@ -554,7 +555,7 @@ T div_ceil(const T &a, const T &b)
 template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const X& x,const Y& y) { return x == y; }
 
-#if !defined(EIGEN_CUDA_ARCH)
+#if !defined(EIGEN_CUDA_ARCH) || defined(EIGEN_CONSTEXPR_ARE_DEVICE_FUNC)
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const float& x,const float& y) { return std::equal_to<float>()(x,y); }
 
@@ -565,7 +566,7 @@ bool equal_strict(const double& x,const double& y) { return std::equal_to<double
 template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const X& x,const Y& y) { return x != y; }
 
-#if !defined(EIGEN_CUDA_ARCH)
+#if !defined(EIGEN_CUDA_ARCH) || defined(EIGEN_CONSTEXPR_ARE_DEVICE_FUNC)
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to<float>()(x,y); }
 
diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index f7a3d9ce7..7311adec7 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -685,12 +685,14 @@ struct possibly_same_dense {
 };
 
 template<typename T1, typename T2>
+EIGEN_DEVICE_FUNC
 bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<possibly_same_dense<T1,T2>::value>::type * = 0)
 {
   return (mat1.data()==mat2.data()) && (mat1.innerStride()==mat2.innerStride()) && (mat1.outerStride()==mat2.outerStride());
 }
 
 template<typename T1, typename T2>
+EIGEN_DEVICE_FUNC
 bool is_same_dense(const T1 &, const T2 &, typename enable_if<!possibly_same_dense<T1,T2>::value>::type * = 0)
 {
   return false;
diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
index 040f8d3bb..fbc1ee2f6 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -20,7 +20,9 @@ class GeneralizedSelfAdjointEigenSolver;
 
 namespace internal {
 template<typename SolverType,int Size,bool IsComplex> struct direct_selfadjoint_eigenvalues;
+
 template<typename MatrixType, typename DiagType, typename SubDiagType>
+EIGEN_DEVICE_FUNC
 ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec);
 }
 
@@ -354,7 +356,8 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
     static const int m_maxIterations = 30;
 
   protected:
-    static void check_template_parameters()
+    static EIGEN_DEVICE_FUNC
+    void check_template_parameters()
     {
       EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
     }
@@ -403,7 +406,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
   
   const InputType &matrix(a_matrix.derived());
   
-  using std::abs;
+  EIGEN_USING_STD_MATH(abs);
   eigen_assert(matrix.cols() == matrix.rows());
   eigen_assert((options&~(EigVecMask|GenEigMask))==0
           && (options&EigVecMask)!=EigVecMask
@@ -479,9 +482,10 @@ namespace internal {
   * \returns \c Success or \c NoConvergence
   */
 template<typename MatrixType, typename DiagType, typename SubDiagType>
+EIGEN_DEVICE_FUNC
 ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec)
 {
-  using std::abs;
+  EIGEN_USING_STD_MATH(abs);
 
   ComputationInfo info;
   typedef typename MatrixType::Scalar Scalar;
@@ -535,7 +539,7 @@ ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag
       diag.segment(i,n-i).minCoeff(&k);
       if (k > 0)
       {
-        std::swap(diag[i], diag[k+i]);
+        numext::swap(diag[i], diag[k+i]);
         if(computeEigenvectors)
           eivec.col(i).swap(eivec.col(k+i));
       }
@@ -605,7 +609,7 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
   EIGEN_DEVICE_FUNC
   static inline bool extract_kernel(MatrixType& mat, Ref<VectorType> res, Ref<VectorType> representative)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     Index i0;
     // Find non-zero column i0 (by construction, there must exist a non zero coefficient on the diagonal):
     mat.diagonal().cwiseAbs().maxCoeff(&i0);
@@ -807,7 +811,7 @@ template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
 EIGEN_DEVICE_FUNC
 static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n)
 {
-  using std::abs;
+  EIGEN_USING_STD_MATH(abs);
   RealScalar td = (diag[end-1] - diag[end])*RealScalar(0.5);
   RealScalar e = subdiag[end-1];
   // Note that thanks to scaling, e^2 or td^2 cannot overflow, however they can still
diff --git a/Eigen/src/Eigenvalues/Tridiagonalization.h b/Eigen/src/Eigenvalues/Tridiagonalization.h
index 1d102c17b..c5c1acf46 100644
--- a/Eigen/src/Eigenvalues/Tridiagonalization.h
+++ b/Eigen/src/Eigenvalues/Tridiagonalization.h
@@ -25,6 +25,7 @@ struct traits<TridiagonalizationMatrixTReturnType<MatrixType> >
 };
 
 template<typename MatrixType, typename CoeffVectorType>
+EIGEN_DEVICE_FUNC
 void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs);
 }
 
@@ -344,6 +345,7 @@ namespace internal {
   * \sa Tridiagonalization::packedMatrix()
   */
 template<typename MatrixType, typename CoeffVectorType>
+EIGEN_DEVICE_FUNC
 void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs)
 {
   using numext::conj;
@@ -424,6 +426,7 @@ struct tridiagonalization_inplace_selector;
   * \sa class Tridiagonalization
   */
 template<typename MatrixType, typename DiagonalType, typename SubDiagonalType>
+EIGEN_DEVICE_FUNC
 void tridiagonalization_inplace(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
 {
   eigen_assert(mat.cols()==mat.rows() && diag.size()==mat.rows() && subdiag.size()==mat.rows()-1);
@@ -439,7 +442,8 @@ struct tridiagonalization_inplace_selector
   typedef typename Tridiagonalization<MatrixType>::CoeffVectorType CoeffVectorType;
   typedef typename Tridiagonalization<MatrixType>::HouseholderSequenceType HouseholderSequenceType;
   template<typename DiagonalType, typename SubDiagonalType>
-  static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
+  static EIGEN_DEVICE_FUNC
+  void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
   {
     CoeffVectorType hCoeffs(mat.cols()-1);
     tridiagonalization_inplace(mat,hCoeffs);
@@ -508,7 +512,8 @@ struct tridiagonalization_inplace_selector<MatrixType,1,IsComplex>
   typedef typename MatrixType::Scalar Scalar;
 
   template<typename DiagonalType, typename SubDiagonalType>
-  static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType&, bool extractQ)
+  static EIGEN_DEVICE_FUNC
+  void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType&, bool extractQ)
   {
     diag(0,0) = numext::real(mat(0,0));
     if(extractQ)
diff --git a/Eigen/src/Householder/Householder.h b/Eigen/src/Householder/Householder.h
index a5f336d18..5bc037f00 100644
--- a/Eigen/src/Householder/Householder.h
+++ b/Eigen/src/Householder/Householder.h
@@ -39,6 +39,7 @@ template<int n> struct decrement_size
   *     MatrixBase::applyHouseholderOnTheRight()
   */
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 void MatrixBase<Derived>::makeHouseholderInPlace(Scalar& tau, RealScalar& beta)
 {
   VectorBlock<Derived, internal::decrement_size<Base::SizeAtCompileTime>::ret> essentialPart(derived(), 1, size()-1);
@@ -62,6 +63,7 @@ void MatrixBase<Derived>::makeHouseholderInPlace(Scalar& tau, RealScalar& beta)
   */
 template<typename Derived>
 template<typename EssentialPart>
+EIGEN_DEVICE_FUNC
 void MatrixBase<Derived>::makeHouseholder(
   EssentialPart& essential,
   Scalar& tau,
@@ -110,6 +112,7 @@ void MatrixBase<Derived>::makeHouseholder(
   */
 template<typename Derived>
 template<typename EssentialPart>
+EIGEN_DEVICE_FUNC
 void MatrixBase<Derived>::applyHouseholderOnTheLeft(
   const EssentialPart& essential,
   const Scalar& tau,
@@ -147,6 +150,7 @@ void MatrixBase<Derived>::applyHouseholderOnTheLeft(
   */
 template<typename Derived>
 template<typename EssentialPart>
+EIGEN_DEVICE_FUNC
 void MatrixBase<Derived>::applyHouseholderOnTheRight(
   const EssentialPart& essential,
   const Scalar& tau,
diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h
index fad1d5ab6..a4f40b75c 100644
--- a/Eigen/src/Householder/HouseholderSequence.h
+++ b/Eigen/src/Householder/HouseholderSequence.h
@@ -87,7 +87,7 @@ struct hseq_side_dependent_impl
 {
   typedef Block<const VectorsType, Dynamic, 1> EssentialVectorType;
   typedef HouseholderSequence<VectorsType, CoeffsType, OnTheLeft> HouseholderSequenceType;
-  static inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k)
+  static EIGEN_DEVICE_FUNC inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k)
   {
     Index start = k+1+h.m_shift;
     return Block<const VectorsType,Dynamic,1>(h.m_vectors, start, k, h.rows()-start, 1);
@@ -173,6 +173,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       *
       * \sa setLength(), setShift()
       */
+    EIGEN_DEVICE_FUNC
     HouseholderSequence(const VectorsType& v, const CoeffsType& h)
       : m_vectors(v), m_coeffs(h), m_reverse(false), m_length(v.diagonalSize()),
         m_shift(0)
@@ -180,6 +181,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     }
 
     /** \brief Copy constructor. */
+    EIGEN_DEVICE_FUNC
     HouseholderSequence(const HouseholderSequence& other)
       : m_vectors(other.m_vectors),
         m_coeffs(other.m_coeffs),
@@ -193,12 +195,14 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       * \returns Number of rows 
       * \details This equals the dimension of the space that the transformation acts on.
       */
+    EIGEN_DEVICE_FUNC
     Index rows() const { return Side==OnTheLeft ? m_vectors.rows() : m_vectors.cols(); }
 
     /** \brief Number of columns of transformation viewed as a matrix.
       * \returns Number of columns
       * \details This equals the dimension of the space that the transformation acts on.
       */
+    EIGEN_DEVICE_FUNC
     Index cols() const { return rows(); }
 
     /** \brief Essential part of a Householder vector.
@@ -215,6 +219,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       *
       * \sa setShift(), shift()
       */
+    EIGEN_DEVICE_FUNC
     const EssentialVectorType essentialVector(Index k) const
     {
       eigen_assert(k >= 0 && k < m_length);
@@ -252,7 +257,9 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     AdjointReturnType inverse() const { return adjoint(); }
 
     /** \internal */
-    template<typename DestType> inline void evalTo(DestType& dst) const
+    template<typename DestType>
+    inline EIGEN_DEVICE_FUNC
+    void evalTo(DestType& dst) const
     {
       Matrix<Scalar, DestType::RowsAtCompileTime, 1,
              AutoAlign|ColMajor, DestType::MaxRowsAtCompileTime, 1> workspace(rows());
@@ -261,6 +268,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
 
     /** \internal */
     template<typename Dest, typename Workspace>
+    EIGEN_DEVICE_FUNC
     void evalTo(Dest& dst, Workspace& workspace) const
     {
       workspace.resize(rows());
@@ -394,6 +402,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       *
       * \sa length()
       */
+    EIGEN_DEVICE_FUNC
     HouseholderSequence& setLength(Index length)
     {
       m_length = length;
@@ -411,13 +420,17 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
       *
       * \sa shift()
       */
+    EIGEN_DEVICE_FUNC
     HouseholderSequence& setShift(Index shift)
     {
       m_shift = shift;
       return *this;
     }
 
+    EIGEN_DEVICE_FUNC
     Index length() const { return m_length; }  /**< \brief Returns the length of the Householder sequence. */
+
+    EIGEN_DEVICE_FUNC
     Index shift() const { return m_shift; }    /**< \brief Returns the shift of the Householder sequence. */
 
     /* Necessary for .adjoint() and .conjugate() */
diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index af1228cb8..bba75fc4f 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -90,6 +90,7 @@ template<typename Scalar> class JacobiRotation
   * \sa MatrixBase::makeJacobi(const MatrixBase<Derived>&, Index, Index), MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
   */
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 bool JacobiRotation<Scalar>::makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z)
 {
   using std::sqrt;
@@ -134,6 +135,7 @@ bool JacobiRotation<Scalar>::makeJacobi(const RealScalar& x, const Scalar& y, co
   */
 template<typename Scalar>
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, Index p, Index q)
 {
   return makeJacobi(numext::real(m.coeff(p,p)), m.coeff(p,q), numext::real(m.coeff(q,q)));
@@ -156,6 +158,7 @@ inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, Ind
   * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
   */
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* z)
 {
   makeGivens(p, q, z, typename internal::conditional<NumTraits<Scalar>::IsComplex, internal::true_type, internal::false_type>::type());
@@ -164,6 +167,7 @@ void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar
 
 // specialization for complexes
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::true_type)
 {
   using std::sqrt;
@@ -223,6 +227,7 @@ void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar
 
 // specialization for reals
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::false_type)
 {
   using std::sqrt;
@@ -286,6 +291,7 @@ void apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>&
   */
 template<typename Derived>
 template<typename OtherScalar>
+EIGEN_DEVICE_FUNC
 inline void MatrixBase<Derived>::applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j)
 {
   RowXpr x(this->row(p));
@@ -301,6 +307,7 @@ inline void MatrixBase<Derived>::applyOnTheLeft(Index p, Index q, const JacobiRo
   */
 template<typename Derived>
 template<typename OtherScalar>
+EIGEN_DEVICE_FUNC
 inline void MatrixBase<Derived>::applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j)
 {
   ColXpr x(this->col(p));
@@ -314,7 +321,8 @@ template<typename Scalar, typename OtherScalar,
          int SizeAtCompileTime, int MinAlignment, bool Vectorizable>
 struct apply_rotation_in_the_plane_selector
 {
-  static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
+  static EIGEN_DEVICE_FUNC
+  inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
   {
     for(Index i=0; i<size; ++i)
     {
@@ -441,6 +449,7 @@ struct apply_rotation_in_the_plane_selector<Scalar,OtherScalar,SizeAtCompileTime
 };
 
 template<typename VectorX, typename VectorY, typename OtherScalar>
+EIGEN_DEVICE_FUNC
 void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j)
 {
   typedef typename VectorX::Scalar Scalar;
diff --git a/Eigen/src/LU/Determinant.h b/Eigen/src/LU/Determinant.h
index d6a3c1e5a..6af63a6e7 100644
--- a/Eigen/src/LU/Determinant.h
+++ b/Eigen/src/LU/Determinant.h
@@ -15,6 +15,7 @@ namespace Eigen {
 namespace internal {
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline const typename Derived::Scalar bruteforce_det3_helper
 (const MatrixBase<Derived>& matrix, int a, int b, int c)
 {
@@ -23,6 +24,7 @@ inline const typename Derived::Scalar bruteforce_det3_helper
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 const typename Derived::Scalar bruteforce_det4_helper
 (const MatrixBase<Derived>& matrix, int j, int k, int m, int n)
 {
@@ -44,7 +46,8 @@ template<typename Derived,
 
 template<typename Derived> struct determinant_impl<Derived, 1>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return m.coeff(0,0);
   }
@@ -52,7 +55,8 @@ template<typename Derived> struct determinant_impl<Derived, 1>
 
 template<typename Derived> struct determinant_impl<Derived, 2>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return m.coeff(0,0) * m.coeff(1,1) - m.coeff(1,0) * m.coeff(0,1);
   }
@@ -60,7 +64,8 @@ template<typename Derived> struct determinant_impl<Derived, 2>
 
 template<typename Derived> struct determinant_impl<Derived, 3>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return bruteforce_det3_helper(m,0,1,2)
           - bruteforce_det3_helper(m,1,0,2)
@@ -70,7 +75,8 @@ template<typename Derived> struct determinant_impl<Derived, 3>
 
 template<typename Derived> struct determinant_impl<Derived, 4>
 {
-  static typename traits<Derived>::Scalar run(const Derived& m)
+  static EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     // trick by Martin Costabel to compute 4x4 det with only 30 muls
     return bruteforce_det4_helper(m,0,1,2,3)
@@ -89,6 +95,7 @@ template<typename Derived> struct determinant_impl<Derived, 4>
   * \returns the determinant of this matrix
   */
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline typename internal::traits<Derived>::Scalar MatrixBase<Derived>::determinant() const
 {
   eigen_assert(rows() == cols());
diff --git a/Eigen/src/LU/InverseImpl.h b/Eigen/src/LU/InverseImpl.h
index f49f23360..1bab00c01 100644
--- a/Eigen/src/LU/InverseImpl.h
+++ b/Eigen/src/LU/InverseImpl.h
@@ -290,6 +290,7 @@ template<typename DstXprType, typename XprType>
 struct Assignment<DstXprType, Inverse<XprType>, internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar>, Dense2Dense>
 {
   typedef Inverse<XprType> SrcXprType;
+  EIGEN_DEVICE_FUNC
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar> &)
   {
     Index dstRows = src.rows();
@@ -332,6 +333,7 @@ struct Assignment<DstXprType, Inverse<XprType>, internal::assign_op<typename Dst
   * \sa computeInverseAndDetWithCheck()
   */
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline const Inverse<Derived> MatrixBase<Derived>::inverse() const
 {
   EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsInteger,THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES)
diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h
index a5748525b..67fdebc6f 100644
--- a/Eigen/src/plugins/BlockMethods.h
+++ b/Eigen/src/plugins/BlockMethods.h
@@ -1061,6 +1061,7 @@ EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
 /// \sa block(Index,Index,NRowsType,NColsType), class Block
 ///
 template<int NRows, int NCols>
+EIGEN_DEVICE_FUNC
 inline typename FixedBlockXpr<NRows,NCols>::Type block(Index startRow, Index startCol,
                                                   Index blockRows, Index blockCols)
 {

From ba972fb6b40c1ea4ac991b0fb5fa6908bccfdaa6 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 13 Jun 2018 12:26:18 -0400
Subject: [PATCH 395/680] moving Half headers from CUDA dir to GPU dir,
 removing the HIP versions

---
 Eigen/src/Core/arch/{CUDA => GPU}/Half.h      |    0
 .../Core/arch/{CUDA => GPU}/PacketMathHalf.h  |    0
 .../src/Core/arch/{CUDA => GPU}/TypeCasting.h |    0
 Eigen/src/Core/arch/HIP/hcc/Half.h            |  705 ------------
 Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h  | 1019 -----------------
 Eigen/src/Core/arch/HIP/hcc/TypeCasting.h     |  212 ----
 6 files changed, 1936 deletions(-)
 rename Eigen/src/Core/arch/{CUDA => GPU}/Half.h (100%)
 rename Eigen/src/Core/arch/{CUDA => GPU}/PacketMathHalf.h (100%)
 rename Eigen/src/Core/arch/{CUDA => GPU}/TypeCasting.h (100%)
 delete mode 100644 Eigen/src/Core/arch/HIP/hcc/Half.h
 delete mode 100644 Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
 delete mode 100644 Eigen/src/Core/arch/HIP/hcc/TypeCasting.h

diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/GPU/Half.h
similarity index 100%
rename from Eigen/src/Core/arch/CUDA/Half.h
rename to Eigen/src/Core/arch/GPU/Half.h
diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
similarity index 100%
rename from Eigen/src/Core/arch/CUDA/PacketMathHalf.h
rename to Eigen/src/Core/arch/GPU/PacketMathHalf.h
diff --git a/Eigen/src/Core/arch/CUDA/TypeCasting.h b/Eigen/src/Core/arch/GPU/TypeCasting.h
similarity index 100%
rename from Eigen/src/Core/arch/CUDA/TypeCasting.h
rename to Eigen/src/Core/arch/GPU/TypeCasting.h
diff --git a/Eigen/src/Core/arch/HIP/hcc/Half.h b/Eigen/src/Core/arch/HIP/hcc/Half.h
deleted file mode 100644
index 2ce8a412c..000000000
--- a/Eigen/src/Core/arch/HIP/hcc/Half.h
+++ /dev/null
@@ -1,705 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-//
-// The conversion routines are Copyright (c) Fabian Giesen, 2016.
-// The original license follows:
-//
-// Copyright (c) Fabian Giesen, 2016
-// All rights reserved.
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted.
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-
-// Standard 16-bit float type, mostly useful for GPUs. Defines a new
-// type Eigen::half (inheriting from HIP's __half struct) with
-// operator overloads such that it behaves basically as an arithmetic
-// type. It will be quite slow on CPUs (so it is recommended to stay
-// in fp32 for CPUs, except for simple parameter conversions, I/O
-// to disk and the likes), but fast on GPUs.
-
-
-#ifndef EIGEN_HALF_HIP_H
-#define EIGEN_HALF_HIP_H
-
-#if __cplusplus > 199711L
-#define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
-#else
-#define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type()
-#endif
-
-
-namespace Eigen {
-
-struct half;
-
-namespace half_impl {
-
-#if !defined(EIGEN_HAS_HIP_FP16)
-// Make our own __half_raw definition that is similar to CUDA's.
-struct __half_raw {
-  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
-  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
-  unsigned short x;
-};
-#elif defined(EIGEN_HAS_OLD_HIP_FP16)
-// Make a __half_raw definition that is
-// ++ compatible with that of Eigen and
-// ++ add a implcit conversion to the native __half of the old HIP implementation.
-//
-// Keeping ".x" as "unsigned short" keeps the interface the same between the Eigen and HIP implementation.
-//
-// In the old HIP implementation,
-//   ++ __half is a typedef of __fp16
-//   ++ the "__h*" routines take "__half" arguments
-// so we need to implicitly convert "__half_raw" to "__half" to avoid having to explicitly make 
-// that conversiion in each call to a "__h*" routine...that is why we have "operator __half" routine
-struct __half_raw {
-  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
-  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
-  union {
-    unsigned short x;
-    __half data;
-  };
-  operator __half(void) const { return data; }
-};
-#endif
- 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
-
-struct half_base : public __half_raw {
-  EIGEN_DEVICE_FUNC half_base() {}
-  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
-  EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
-#if defined(EIGEN_HAS_HIP_FP16)
-  #if defined(EIGEN_HAS_OLD_HIP_FP16)
-  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(__half_as_ushort(h)) {}
-  #else
-  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
-  #endif
-#endif
-};
-
-} // namespace half_impl
-
-// Class definition.
-struct half : public half_impl::half_base {
-  #if !defined(EIGEN_HAS_HIP_FP16) || defined(EIGEN_HAS_OLD_HIP_FP16)
-    typedef half_impl::__half_raw __half_raw;
-  #endif
-
-  EIGEN_DEVICE_FUNC half() {}
-
-  EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
-  EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
-#if defined(EIGEN_HAS_HIP_FP16)
-  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
-#endif
-
-  explicit EIGEN_DEVICE_FUNC half(bool b)
-      : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
-  template<class T>
-  explicit EIGEN_DEVICE_FUNC half(const T& val)
-      : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
-  explicit EIGEN_DEVICE_FUNC half(float f)
-      : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
-
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
-    // +0.0 and -0.0 become false, everything else becomes true.
-    return (x & 0x7fff) != 0;
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
-    return static_cast<signed char>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
-    return static_cast<unsigned char>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
-    return static_cast<short>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
-    return static_cast<unsigned short>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
-    return static_cast<int>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
-    return static_cast<unsigned int>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
-    return static_cast<long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
-    return static_cast<unsigned long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
-    return static_cast<long long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
-    return static_cast<unsigned long long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
-    return half_impl::half_to_float(*this);
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
-    return static_cast<double>(half_impl::half_to_float(*this));
-  }
-
-  EIGEN_DEVICE_FUNC half& operator=(const half& other) {
-    x = other.x;
-    return *this;
-  }
-};
-
-} // end namespace Eigen
-
-namespace std {
-template<>
-struct numeric_limits<Eigen::half> {
-  static const bool is_specialized = true;
-  static const bool is_signed = true;
-  static const bool is_integer = false;
-  static const bool is_exact = false;
-  static const bool has_infinity = true;
-  static const bool has_quiet_NaN = true;
-  static const bool has_signaling_NaN = true;
-  static const float_denorm_style has_denorm = denorm_present;
-  static const bool has_denorm_loss = false;
-  static const std::float_round_style round_style = std::round_to_nearest;
-  static const bool is_iec559 = false;
-  static const bool is_bounded = false;
-  static const bool is_modulo = false;
-  static const int digits = 11;
-  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
-  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
-  static const int radix = 2;
-  static const int min_exponent = -13;
-  static const int min_exponent10 = -4;
-  static const int max_exponent = 16;
-  static const int max_exponent10 = 4;
-  static const bool traps = true;
-  static const bool tinyness_before = false;
-
-  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
-  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
-  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
-  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
-  static Eigen::half round_error() { return Eigen::half(0.5); }
-  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
-  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
-  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
-  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
-};
-
-// If std::numeric_limits<T> is specialized, should also specialize
-// std::numeric_limits<const T>, std::numeric_limits<volatile T>, and
-// std::numeric_limits<const volatile T>
-// https://stackoverflow.com/a/16519653/
-template<>
-struct numeric_limits<const Eigen::half> : numeric_limits<Eigen::half> {};
-template<>
-struct numeric_limits<volatile Eigen::half> : numeric_limits<Eigen::half> {};
-template<>
-struct numeric_limits<const volatile Eigen::half> : numeric_limits<Eigen::half> {};
-} // end namespace std
-
-namespace Eigen {
-
-namespace half_impl {
-
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-
-// Intrinsics for native fp16 support. Note that on current hardware,
-// these are no faster than fp32 arithmetic (you need to use the half2
-// versions to get the ALU speed increased), but you do save the
-// conversion steps back and forth.
-
-EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
-  return __hadd(a, b);
-}
-EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
-  return __hmul(a, b);
-}
-EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
-  return __hsub(a, b);
-}
-EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
-  float num = __half2float(a);
-  float denom = __half2float(b);
-  return __float2half(num / denom);
-}
-EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
-  return __hneg(a);
-}
-EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) {
-  a = a + b;
-  return a;
-}
-EIGEN_STRONG_INLINE __device__ half& operator *= (half& a, const half& b) {
-  a = a * b;
-  return a;
-}
-EIGEN_STRONG_INLINE __device__ half& operator -= (half& a, const half& b) {
-  a = a - b;
-  return a;
-}
-EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) {
-  a = a / b;
-  return a;
-}
-EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) {
-  return __heq(a, b);
-}
-EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) {
-  return __hne(a, b);
-}
-EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) {
-  return __hlt(a, b);
-}
-EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) {
-  return __hle(a, b);
-}
-EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) {
-  return __hgt(a, b);
-}
-EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
-  return __hge(a, b);
-}
-
-#else  // Emulate support for half floats
-
-// Definitions for CPUs mostly working through conversion to/from fp32.
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
-  return half(float(a) + float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
-  return half(float(a) * float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
-  return half(float(a) - float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
-  return half(float(a) / float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
-  half result;
-  result.x = a.x ^ 0x8000;
-  return result;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
-  a = half(float(a) + float(b));
-  return a;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
-  a = half(float(a) * float(b));
-  return a;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
-  a = half(float(a) - float(b));
-  return a;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
-  a = half(float(a) / float(b));
-  return a;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
-  return numext::equal_strict(float(a),float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
-  return numext::not_equal_strict(float(a), float(b));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
-  return float(a) < float(b);
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
-  return float(a) <= float(b);
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
-  return float(a) > float(b);
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
-  return float(a) >= float(b);
-}
-
-#endif  // Emulate support for half floats
-
-// Division by an index. Do it in full float precision to avoid accuracy
-// issues in converting the denominator to half.
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
-  return half(static_cast<float>(a) / static_cast<float>(b));
-}
-
-// Conversion routines, including fallbacks for the host or older CUDA.
-// Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
-// these in hardware. If we need more performance on older/other CPUs, they are
-// also possible to vectorize directly.
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x) {
-  __half_raw h;
-  h.x = x;
-  return h;
-}
-
-union FP32 {
-  unsigned int u;
-  float f;
-};
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  __half tmp_ff = __float2half(ff);
-  #if defined(EIGEN_HAS_OLD_HIP_FP16)
-    __half_raw h;
-    h.data = tmp_ff;
-    return h;
-  #else
-    return *(__half_raw*)&tmp_ff;
-  #endif
-
-#elif defined(EIGEN_HAS_FP16_C)
-  __half_raw h;
-  h.x = _cvtss_sh(ff, 0);
-  return h;
-
-#else
-  FP32 f; f.f = ff;
-
-  const FP32 f32infty = { 255 << 23 };
-  const FP32 f16max = { (127 + 16) << 23 };
-  const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
-  unsigned int sign_mask = 0x80000000u;
-  __half_raw o;
-  o.x = static_cast<unsigned short>(0x0u);
-
-  unsigned int sign = f.u & sign_mask;
-  f.u ^= sign;
-
-  // NOTE all the integer compares in this function can be safely
-  // compiled into signed compares since all operands are below
-  // 0x80000000. Important if you want fast straight SSE2 code
-  // (since there's no unsigned PCMPGTD).
-
-  if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
-    o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
-  } else {  // (De)normalized number or zero
-    if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
-      // use a magic value to align our 10 mantissa bits at the bottom of
-      // the float. as long as FP addition is round-to-nearest-even this
-      // just works.
-      f.f += denorm_magic.f;
-
-      // and one integer subtract of the bias later, we have our final float!
-      o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
-    } else {
-      unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
-
-      // update exponent, rounding bias part 1
-      f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
-      // rounding bias part 2
-      f.u += mant_odd;
-      // take the bits!
-      o.x = static_cast<unsigned short>(f.u >> 13);
-    }
-  }
-
-  o.x |= static_cast<unsigned short>(sign >> 16);
-  return o;
-#endif
-}
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return __half2float(h);
-
-#elif defined(EIGEN_HAS_FP16_C)
-  return _cvtsh_ss(h.x);
-
-#else
-  const FP32 magic = { 113 << 23 };
-  const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
-  FP32 o;
-
-  o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
-  unsigned int exp = shifted_exp & o.u;   // just the exponent
-  o.u += (127 - 15) << 23;                // exponent adjust
-
-  // handle exponent special cases
-  if (exp == shifted_exp) {     // Inf/NaN?
-    o.u += (128 - 16) << 23;    // extra exp adjust
-  } else if (exp == 0) {        // Zero/Denormal?
-    o.u += 1 << 23;             // extra exp adjust
-    o.f -= magic.f;             // renormalize
-  }
-
-  o.u |= (h.x & 0x8000) << 16;    // sign bit
-  return o.f;
-#endif
-}
-
-// --- standard functions ---
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
-  return (a.x & 0x7fff) == 0x7c00;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return __hisnan(a);
-#else
-  return (a.x & 0x7fff) > 0x7c00;
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
-  return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
-}
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
-  half result;
-  result.x = a.x & 0x7FFF;
-  return result;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return half(hexp(a));
-#else
-   return half(::expf(float(a)));
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
-  return half(numext::expm1(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return half(hlog(a));
-#else
-  return half(::logf(float(a)));
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
-  return half(numext::log1p(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
-  return half(::log10f(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return half(hsqrt(a));
-#else
-    return half(::sqrtf(float(a)));
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
-  return half(::powf(float(a), float(b)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
-  return half(::sinf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
-  return half(::cosf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
-  return half(::tanf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
-  return half(::tanhf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return half(hfloor(a));
-#else
-  return half(::floorf(float(a)));
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return half(hceil(a));
-#else
-  return half(::ceilf(float(a)));
-#endif
-}
-
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return __hlt(b, a) ? b : a;
-#else
-  const float f1 = static_cast<float>(a);
-  const float f2 = static_cast<float>(b);
-  return f2 < f1 ? b : a;
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-  return __hlt(a, b) ? b : a;
-#else
-  const float f1 = static_cast<float>(a);
-  const float f2 = static_cast<float>(b);
-  return f1 < f2 ? b : a;
-#endif
-}
-
-EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
-  os << static_cast<float>(v);
-  return os;
-}
-
-} // end namespace half_impl
-
-// import Eigen::half_impl::half into Eigen namespace
-// using half_impl::half;
-
-namespace internal {
-
-template<>
-struct random_default_impl<half, false, false>
-{
-  static inline half run(const half& x, const half& y)
-  {
-    return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
-  }
-  static inline half run()
-  {
-    return run(half(-1.f), half(1.f));
-  }
-};
-
-template<> struct is_arithmetic<half> { enum { value = true }; };
-
-} // end namespace internal
-
-template<> struct NumTraits<Eigen::half>
-    : GenericNumTraits<Eigen::half>
-{
-  enum {
-    IsSigned = true,
-    IsInteger = false,
-    IsComplex = false,
-    RequireInitialization = false
-  };
-
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
-    return half_impl::raw_uint16_to_half(0x0800);
-  }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
-    return half_impl::raw_uint16_to_half(0x7bff);
-  }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
-    return half_impl::raw_uint16_to_half(0xfbff);
-  }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
-    return half_impl::raw_uint16_to_half(0x7c00);
-  }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
-    return half_impl::raw_uint16_to_half(0x7c01);
-  }
-};
-
-} // end namespace Eigen
-
-// C-like standard mathematical functions and trancendentals.
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) {
-  Eigen::half result;
-  result.x = a.x & 0x7FFF;
-  return result;
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
-  return Eigen::half(::expf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-    return Eigen::half(hlog(a));
-#else
-  return Eigen::half(::logf(float(a)));
-#endif
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
-  return Eigen::half(::sqrtf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) {
-  return Eigen::half(::powf(float(a), float(b)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) {
-  return Eigen::half(::floorf(float(a)));
-}
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) {
-  return Eigen::half(::ceilf(float(a)));
-}
-
-namespace std {
-
-#if __cplusplus > 199711L
-template <>
-struct hash<Eigen::half> {
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
-    return static_cast<std::size_t>(a.x);
-  }
-};
-#endif
-
-} // end namespace std
-
-
-// Add the missing shfl_xor intrinsic
-#if defined(EIGEN_HAS_HIP_FP16) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
-__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
-  // FIXME
-  //return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
-  return var;
-}
-#endif
-
-// ldg() has an overload for __half, but we also need one for Eigen::half.
-#if defined(EIGEN_HAS_HIP_FP16) && \
-    defined(__HIP_ARCH_HAS_WARP_FUNNEL_SHIFT__) && defined(__HIP_ARCH_HAS_DYNAMIC_PARALLEL__)
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
-  // FIXME
-  //return Eigen::half_impl::raw_uint16_to_half(
-  //    __ldg(reinterpret_cast<const unsigned short*>(ptr)));
-  return *ptr;
-}
-#endif
-
-
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
-namespace Eigen {
-namespace numext {
-
-template<>
-EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-bool (isnan)(const Eigen::half& h) {
-  return (half_impl::isnan)(h);
-}
-
-template<>
-EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-bool (isinf)(const Eigen::half& h) {
-  return (half_impl::isinf)(h);
-}
-
-template<>
-EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-bool (isfinite)(const Eigen::half& h) {
-  return (half_impl::isfinite)(h);
-}
-
-} // namespace Eigen
-}  // namespace numext
-#endif
-
-#endif // EIGEN_HALF_HIP_H
diff --git a/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h b/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
deleted file mode 100644
index 29c3f4671..000000000
--- a/Eigen/src/Core/arch/HIP/hcc/PacketMathHalf.h
+++ /dev/null
@@ -1,1019 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_PACKET_MATH_HALF_HIP_H
-#define EIGEN_PACKET_MATH_HALF_HIP_H
-
-
-namespace Eigen {
-namespace internal {
-
-// Most of the following operations require arch >= 3.0
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-
-template<> struct is_arithmetic<half2> { enum { value = true }; };
-
-template<> struct packet_traits<Eigen::half> : default_packet_traits
-{
-  typedef half2 type;
-  typedef half2 half;
-  enum {
-    Vectorizable = 1,
-    AlignedOnScalar = 1,
-    size=2,
-    HasHalfPacket = 0,
-    HasAdd    = 1,
-    HasMul    = 1,
-    HasDiv    = 1,
-    HasSqrt   = 1,
-    HasRsqrt  = 1,
-    HasExp    = 1,
-    HasExpm1  = 1,
-    HasLog    = 1,
-    HasLog1p  = 1
-  };
-};
-
-template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
-#if defined(EIGEN_HAS_OLD_HIP_FP16)
-  return half2half2(from);
-#else
-  return __half2half2(from);
-#endif
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
-  return *reinterpret_cast<const half2*>(from);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
-  return __halves2half2(from[0], from[1]);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
-  return __halves2half2(from[0], from[0]);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
-  *reinterpret_cast<half2*>(to) = from;
-}
-
-template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
-  to[0] = __low2half(from);
-  to[1] = __high2half(from);
-}
-
-template<>
- __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
-#if defined(EIGEN_HAS_OLD_HIP_FP16)
-  return __halves2half2((*(from+0)), (*(from+1)));
-#else
-  return __ldg((const half2*)from);
-#endif
-}
-
-template<>
-__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
-#if defined(EIGEN_HAS_OLD_HIP_FP16)
-  return __halves2half2((*(from+0)), (*(from+1)));
-#else
-  return __halves2half2(__ldg(from+0), __ldg(from+1));
-#endif
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
-  return __halves2half2(from[0*stride], from[1*stride]);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
-  to[stride*0] = __low2half(from);
-  to[stride*1] = __high2half(from);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
-  return __low2half(a); 
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
-  __half x = __ushort_as_half(__half_as_ushort(__low2half(a)) & 0x7FFF);
-  __half y = __ushort_as_half(__half_as_ushort(__high2half(a)) & 0x7FFF);
-  return __halves2half2(x, y);
-}
-
-
-__device__ EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<half2,2>& kernel) {
-  __half a1 = __low2half(kernel.packet[0]);
-  __half a2 = __high2half(kernel.packet[0]);
-  __half b1 = __low2half(kernel.packet[1]);
-  __half b2 = __high2half(kernel.packet[1]);
-  kernel.packet[0] = __halves2half2(a1, b1);
-  kernel.packet[1] = __halves2half2(a2, b2);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
-  return __halves2half2(a, __hadd(a, __float2half(1.0f)));
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
-  return __hadd2(a, b);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
-  return __hsub2(a, b);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
-  return __hneg2(a);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
-  return __hmul2(a, b);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
-   return __hfma2(a, b, c);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
-#if defined(EIGEN_HAS_OLD_HIP_FP16)
-  return h2div(a, b);
-#else
-  return __h2div(a, b);
-#endif
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
-  float a1 = __low2float(a);
-  float a2 = __high2float(a);
-  float b1 = __low2float(b);
-  float b2 = __high2float(b);
-  __half r1 = a1 < b1 ? __low2half(a) : __low2half(b);
-  __half r2 = a2 < b2 ? __high2half(a) : __high2half(b);
-  return __halves2half2(r1, r2);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
-  float a1 = __low2float(a);
-  float a2 = __high2float(a);
-  float b1 = __low2float(b);
-  float b2 = __high2float(b);
-  __half r1 = a1 > b1 ? __low2half(a) : __low2half(b);
-  __half r2 = a2 > b2 ? __high2half(a) : __high2half(b);
-  return __halves2half2(r1, r2);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
-  return __hadd(__low2half(a), __high2half(a));
-}
-
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
-  __half first = __low2half(a);
-  __half second = __high2half(a);
-  return __hgt(first, second) ? first : second;
-}
-
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
-  __half first = __low2half(a);
-  __half second = __high2half(a);
-  return __hlt(first, second) ? first : second;
-}
-
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
-  return __hmul(__low2half(a), __high2half(a));
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
-  float a1 = __low2float(a);
-  float a2 = __high2float(a);
-  float r1 = log1pf(a1);
-  float r2 = log1pf(a2);
-  return __floats2half2_rn(r1, r2);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
-  float a1 = __low2float(a);
-  float a2 = __high2float(a);
-  float r1 = expm1f(a1);
-  float r2 = expm1f(a2);
-  return __floats2half2_rn(r1, r2);
-}
-
-template<>  __device__ EIGEN_STRONG_INLINE
-half2 plog<half2>(const half2& a) {
-  return h2log(a);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE
-half2 pexp<half2>(const half2& a) {
-  return h2exp(a);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE
-half2 psqrt<half2>(const half2& a) {
-  return h2sqrt(a);
-}
-
-template<> __device__ EIGEN_STRONG_INLINE
-half2 prsqrt<half2>(const half2& a) {
-  return h2rsqrt(a);
-}
-
-#elif defined EIGEN_VECTORIZE_AVX512
-
-typedef struct {
-  __m256i x;
-} Packet16h;
-
-
-template<> struct is_arithmetic<Packet16h> { enum { value = true }; };
-
-template <>
-struct packet_traits<half> : default_packet_traits {
-  typedef Packet16h type;
-  // There is no half-size packet for Packet16h.
-  typedef Packet16h half;
-  enum {
-    Vectorizable = 1,
-    AlignedOnScalar = 1,
-    size = 16,
-    HasHalfPacket = 0,
-    HasAdd    = 0,
-    HasSub    = 0,
-    HasMul    = 0,
-    HasNegate = 0,
-    HasAbs    = 0,
-    HasAbs2   = 0,
-    HasMin    = 0,
-    HasMax    = 0,
-    HasConj   = 0,
-    HasSetLinear = 0,
-    HasDiv = 0,
-    HasSqrt = 0,
-    HasRsqrt = 0,
-    HasExp = 0,
-    HasLog = 0,
-    HasBlend = 0
-  };
-};
-
-
-template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; };
-
-template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) {
-  Packet16h result;
-  result.x = _mm256_set1_epi16(from.x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) {
-  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0)));
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) {
-  Packet16h result;
-  result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) {
-  Packet16h result;
-  result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_store_si256((__m256i*)to, from.x);
-}
-
-template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_storeu_si256((__m256i*)to, from.x);
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h
-ploadquad(const Eigen::half* from) {
-  Packet16h result;
-  unsigned short a = from[0].x;
-  unsigned short b = from[1].x;
-  unsigned short c = from[2].x;
-  unsigned short d = from[3].x;
-  result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a);
-  return result;
-}
-
-EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) {
-#ifdef EIGEN_HAS_FP16_C
-  return _mm512_cvtph_ps(a.x);
-#else
-  EIGEN_ALIGN64 half aux[16];
-  pstore(aux, a);
-  float f0(aux[0]);
-  float f1(aux[1]);
-  float f2(aux[2]);
-  float f3(aux[3]);
-  float f4(aux[4]);
-  float f5(aux[5]);
-  float f6(aux[6]);
-  float f7(aux[7]);
-  float f8(aux[8]);
-  float f9(aux[9]);
-  float fa(aux[10]);
-  float fb(aux[11]);
-  float fc(aux[12]);
-  float fd(aux[13]);
-  float fe(aux[14]);
-  float ff(aux[15]);
-
-  return _mm512_set_ps(
-      ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0);
-#endif
-}
-
-EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
-#ifdef EIGEN_HAS_FP16_C
-  Packet16h result;
-  result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
-  return result;
-#else
-  EIGEN_ALIGN64 float aux[16];
-  pstore(aux, a);
-  half h0(aux[0]);
-  half h1(aux[1]);
-  half h2(aux[2]);
-  half h3(aux[3]);
-  half h4(aux[4]);
-  half h5(aux[5]);
-  half h6(aux[6]);
-  half h7(aux[7]);
-  half h8(aux[8]);
-  half h9(aux[9]);
-  half ha(aux[10]);
-  half hb(aux[11]);
-  half hc(aux[12]);
-  half hd(aux[13]);
-  half he(aux[14]);
-  half hf(aux[15]);
-
-  Packet16h result;
-  result.x = _mm256_set_epi16(
-      hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x,
-      h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
-  return result;
-#endif
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
-  Packet16f af = half2float(a);
-  Packet16f bf = half2float(b);
-  Packet16f rf = padd(af, bf);
-  return float2half(rf);
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
-  Packet16f af = half2float(a);
-  Packet16f bf = half2float(b);
-  Packet16f rf = pmul(af, bf);
-  return float2half(rf);
-}
-
-template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
-  Packet16f from_float = half2float(from);
-  return half(predux(from_float));
-}
-
-template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
-{
-  Packet16h result;
-  result.x = _mm256_set_epi16(
-      from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x,
-      from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x,
-      from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x,
-      from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride)
-{
-  EIGEN_ALIGN64 half aux[16];
-  pstore(aux, from);
-  to[stride*0].x = aux[0].x;
-  to[stride*1].x = aux[1].x;
-  to[stride*2].x = aux[2].x;
-  to[stride*3].x = aux[3].x;
-  to[stride*4].x = aux[4].x;
-  to[stride*5].x = aux[5].x;
-  to[stride*6].x = aux[6].x;
-  to[stride*7].x = aux[7].x;
-  to[stride*8].x = aux[8].x;
-  to[stride*9].x = aux[9].x;
-  to[stride*10].x = aux[10].x;
-  to[stride*11].x = aux[11].x;
-  to[stride*12].x = aux[12].x;
-  to[stride*13].x = aux[13].x;
-  to[stride*14].x = aux[14].x;
-  to[stride*15].x = aux[15].x;
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet16h,16>& kernel) {
-  __m256i a = kernel.packet[0].x;
-  __m256i b = kernel.packet[1].x;
-  __m256i c = kernel.packet[2].x;
-  __m256i d = kernel.packet[3].x;
-  __m256i e = kernel.packet[4].x;
-  __m256i f = kernel.packet[5].x;
-  __m256i g = kernel.packet[6].x;
-  __m256i h = kernel.packet[7].x;
-  __m256i i = kernel.packet[8].x;
-  __m256i j = kernel.packet[9].x;
-  __m256i k = kernel.packet[10].x;
-  __m256i l = kernel.packet[11].x;
-  __m256i m = kernel.packet[12].x;
-  __m256i n = kernel.packet[13].x;
-  __m256i o = kernel.packet[14].x;
-  __m256i p = kernel.packet[15].x;
-
-  __m256i ab_07 = _mm256_unpacklo_epi16(a, b);
-  __m256i cd_07 = _mm256_unpacklo_epi16(c, d);
-  __m256i ef_07 = _mm256_unpacklo_epi16(e, f);
-  __m256i gh_07 = _mm256_unpacklo_epi16(g, h);
-  __m256i ij_07 = _mm256_unpacklo_epi16(i, j);
-  __m256i kl_07 = _mm256_unpacklo_epi16(k, l);
-  __m256i mn_07 = _mm256_unpacklo_epi16(m, n);
-  __m256i op_07 = _mm256_unpacklo_epi16(o, p);
-
-  __m256i ab_8f = _mm256_unpackhi_epi16(a, b);
-  __m256i cd_8f = _mm256_unpackhi_epi16(c, d);
-  __m256i ef_8f = _mm256_unpackhi_epi16(e, f);
-  __m256i gh_8f = _mm256_unpackhi_epi16(g, h);
-  __m256i ij_8f = _mm256_unpackhi_epi16(i, j);
-  __m256i kl_8f = _mm256_unpackhi_epi16(k, l);
-  __m256i mn_8f = _mm256_unpackhi_epi16(m, n);
-  __m256i op_8f = _mm256_unpackhi_epi16(o, p);
-
-  __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07);
-  __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07);
-  __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07);
-  __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07);
-  __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07);
-  __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07);
-  __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07);
-  __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07);
-
-  __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f);
-  __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f);
-  __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f);
-  __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f);
-  __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f);
-  __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f);
-  __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f);
-  __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f);
-
-  __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03);
-  __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03);
-  __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03);
-  __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03);
-  __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47);
-  __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47);
-  __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47);
-  __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47);
-  __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b);
-  __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b);
-  __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b);
-  __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b);
-  __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf);
-  __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf);
-  __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf);
-  __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf);
-
-  // NOTE: no unpacklo/hi instr in this case, so using permute instr.
-  __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
-  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
-  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
-  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
-  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
-  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
-  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
-  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
-  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
-  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
-  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
-  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
-  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
-  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
-  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
-  __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
-
-  kernel.packet[0].x = a_p_0;
-  kernel.packet[1].x = a_p_1;
-  kernel.packet[2].x = a_p_2;
-  kernel.packet[3].x = a_p_3;
-  kernel.packet[4].x = a_p_4;
-  kernel.packet[5].x = a_p_5;
-  kernel.packet[6].x = a_p_6;
-  kernel.packet[7].x = a_p_7;
-  kernel.packet[8].x = a_p_8;
-  kernel.packet[9].x = a_p_9;
-  kernel.packet[10].x = a_p_a;
-  kernel.packet[11].x = a_p_b;
-  kernel.packet[12].x = a_p_c;
-  kernel.packet[13].x = a_p_d;
-  kernel.packet[14].x = a_p_e;
-  kernel.packet[15].x = a_p_f;
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet16h,8>& kernel) {
-  EIGEN_ALIGN64 half in[8][16];
-  pstore<half>(in[0], kernel.packet[0]);
-  pstore<half>(in[1], kernel.packet[1]);
-  pstore<half>(in[2], kernel.packet[2]);
-  pstore<half>(in[3], kernel.packet[3]);
-  pstore<half>(in[4], kernel.packet[4]);
-  pstore<half>(in[5], kernel.packet[5]);
-  pstore<half>(in[6], kernel.packet[6]);
-  pstore<half>(in[7], kernel.packet[7]);
-
-  EIGEN_ALIGN64 half out[8][16];
-
-  for (int i = 0; i < 8; ++i) {
-    for (int j = 0; j < 8; ++j) {
-      out[i][j] = in[j][2*i];
-    }
-    for (int j = 0; j < 8; ++j) {
-      out[i][j+8] = in[j][2*i+1];
-    }
-  }
-
-  kernel.packet[0] = pload<Packet16h>(out[0]);
-  kernel.packet[1] = pload<Packet16h>(out[1]);
-  kernel.packet[2] = pload<Packet16h>(out[2]);
-  kernel.packet[3] = pload<Packet16h>(out[3]);
-  kernel.packet[4] = pload<Packet16h>(out[4]);
-  kernel.packet[5] = pload<Packet16h>(out[5]);
-  kernel.packet[6] = pload<Packet16h>(out[6]);
-  kernel.packet[7] = pload<Packet16h>(out[7]);
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet16h,4>& kernel) {
-  EIGEN_ALIGN64 half in[4][16];
-  pstore<half>(in[0], kernel.packet[0]);
-  pstore<half>(in[1], kernel.packet[1]);
-  pstore<half>(in[2], kernel.packet[2]);
-  pstore<half>(in[3], kernel.packet[3]);
-
-  EIGEN_ALIGN64 half out[4][16];
-
-  for (int i = 0; i < 4; ++i) {
-    for (int j = 0; j < 4; ++j) {
-      out[i][j] = in[j][4*i];
-    }
-    for (int j = 0; j < 4; ++j) {
-      out[i][j+4] = in[j][4*i+1];
-    }
-    for (int j = 0; j < 4; ++j) {
-      out[i][j+8] = in[j][4*i+2];
-    }
-    for (int j = 0; j < 4; ++j) {
-      out[i][j+12] = in[j][4*i+3];
-    }
-  }
-
-  kernel.packet[0] = pload<Packet16h>(out[0]);
-  kernel.packet[1] = pload<Packet16h>(out[1]);
-  kernel.packet[2] = pload<Packet16h>(out[2]);
-  kernel.packet[3] = pload<Packet16h>(out[3]);
-}
-
-
-#elif defined EIGEN_VECTORIZE_AVX
-
-typedef struct {
-  __m128i x;
-} Packet8h;
-
-
-template<> struct is_arithmetic<Packet8h> { enum { value = true }; };
-
-template <>
-struct packet_traits<Eigen::half> : default_packet_traits {
-  typedef Packet8h type;
-  // There is no half-size packet for Packet8h.
-  typedef Packet8h half;
-  enum {
-    Vectorizable = 1,
-    AlignedOnScalar = 1,
-    size = 8,
-    HasHalfPacket = 0,
-    HasAdd    = 0,
-    HasSub    = 0,
-    HasMul    = 0,
-    HasNegate = 0,
-    HasAbs    = 0,
-    HasAbs2   = 0,
-    HasMin    = 0,
-    HasMax    = 0,
-    HasConj   = 0,
-    HasSetLinear = 0,
-    HasDiv = 0,
-    HasSqrt = 0,
-    HasRsqrt = 0,
-    HasExp = 0,
-    HasLog = 0,
-    HasBlend = 0
-  };
-};
-
-
-template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; };
-
-template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) {
-  Packet8h result;
-  result.x = _mm_set1_epi16(from.x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet8h>(const Packet8h& from) {
-  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_extract_epi16(from.x, 0)));
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h pload<Packet8h>(const Eigen::half* from) {
-  Packet8h result;
-  result.x = _mm_load_si128(reinterpret_cast<const __m128i*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h ploadu<Packet8h>(const Eigen::half* from) {
-  Packet8h result;
-  result.x = _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8h& from) {
-  _mm_store_si128(reinterpret_cast<__m128i*>(to), from.x);
-}
-
-template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet8h& from) {
-  _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x);
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h
-ploadquad<Packet8h>(const Eigen::half* from) {
-  Packet8h result;
-  unsigned short a = from[0].x;
-  unsigned short b = from[1].x;
-  result.x = _mm_set_epi16(b, b, b, b, a, a, a, a);
-  return result;
-}
-
-EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) {
-#ifdef EIGEN_HAS_FP16_C
-  return _mm256_cvtph_ps(a.x);
-#else
-  EIGEN_ALIGN32 Eigen::half aux[8];
-  pstore(aux, a);
-  float f0(aux[0]);
-  float f1(aux[1]);
-  float f2(aux[2]);
-  float f3(aux[3]);
-  float f4(aux[4]);
-  float f5(aux[5]);
-  float f6(aux[6]);
-  float f7(aux[7]);
-
-  return _mm256_set_ps(f7, f6, f5, f4, f3, f2, f1, f0);
-#endif
-}
-
-EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
-#ifdef EIGEN_HAS_FP16_C
-  Packet8h result;
-  result.x = _mm256_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
-  return result;
-#else
-  EIGEN_ALIGN32 float aux[8];
-  pstore(aux, a);
-  Eigen::half h0(aux[0]);
-  Eigen::half h1(aux[1]);
-  Eigen::half h2(aux[2]);
-  Eigen::half h3(aux[3]);
-  Eigen::half h4(aux[4]);
-  Eigen::half h5(aux[5]);
-  Eigen::half h6(aux[6]);
-  Eigen::half h7(aux[7]);
-
-  Packet8h result;
-  result.x = _mm_set_epi16(h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
-  return result;
-#endif
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
-
-template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) {
-  Packet8f af = half2float(a);
-  Packet8f bf = half2float(b);
-  Packet8f rf = padd(af, bf);
-  return float2half(rf);
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) {
-  Packet8f af = half2float(a);
-  Packet8f bf = half2float(b);
-  Packet8f rf = pmul(af, bf);
-  return float2half(rf);
-}
-
-template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride)
-{
-  Packet8h result;
-  result.x = _mm_set_epi16(from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half* to, const Packet8h& from, Index stride)
-{
-  EIGEN_ALIGN32 Eigen::half aux[8];
-  pstore(aux, from);
-  to[stride*0].x = aux[0].x;
-  to[stride*1].x = aux[1].x;
-  to[stride*2].x = aux[2].x;
-  to[stride*3].x = aux[3].x;
-  to[stride*4].x = aux[4].x;
-  to[stride*5].x = aux[5].x;
-  to[stride*6].x = aux[6].x;
-  to[stride*7].x = aux[7].x;
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) {
-  Packet8f af = half2float(a);
-  float reduced = predux<Packet8f>(af);
-  return Eigen::half(reduced);
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) {
-  Packet8f af = half2float(a);
-  float reduced = predux_max<Packet8f>(af);
-  return Eigen::half(reduced);
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) {
-  Packet8f af = half2float(a);
-  float reduced = predux_min<Packet8f>(af);
-  return Eigen::half(reduced);
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) {
-  Packet8f af = half2float(a);
-  float reduced = predux_mul<Packet8f>(af);
-  return Eigen::half(reduced);
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet8h,8>& kernel) {
-  __m128i a = kernel.packet[0].x;
-  __m128i b = kernel.packet[1].x;
-  __m128i c = kernel.packet[2].x;
-  __m128i d = kernel.packet[3].x;
-  __m128i e = kernel.packet[4].x;
-  __m128i f = kernel.packet[5].x;
-  __m128i g = kernel.packet[6].x;
-  __m128i h = kernel.packet[7].x;
-
-  __m128i a03b03 = _mm_unpacklo_epi16(a, b);
-  __m128i c03d03 = _mm_unpacklo_epi16(c, d);
-  __m128i e03f03 = _mm_unpacklo_epi16(e, f);
-  __m128i g03h03 = _mm_unpacklo_epi16(g, h);
-  __m128i a47b47 = _mm_unpackhi_epi16(a, b);
-  __m128i c47d47 = _mm_unpackhi_epi16(c, d);
-  __m128i e47f47 = _mm_unpackhi_epi16(e, f);
-  __m128i g47h47 = _mm_unpackhi_epi16(g, h);
-
-  __m128i a01b01c01d01 = _mm_unpacklo_epi32(a03b03, c03d03);
-  __m128i a23b23c23d23 = _mm_unpackhi_epi32(a03b03, c03d03);
-  __m128i e01f01g01h01 = _mm_unpacklo_epi32(e03f03, g03h03);
-  __m128i e23f23g23h23 = _mm_unpackhi_epi32(e03f03, g03h03);
-  __m128i a45b45c45d45 = _mm_unpacklo_epi32(a47b47, c47d47);
-  __m128i a67b67c67d67 = _mm_unpackhi_epi32(a47b47, c47d47);
-  __m128i e45f45g45h45 = _mm_unpacklo_epi32(e47f47, g47h47);
-  __m128i e67f67g67h67 = _mm_unpackhi_epi32(e47f47, g47h47);
-
-  __m128i a0b0c0d0e0f0g0h0 = _mm_unpacklo_epi64(a01b01c01d01, e01f01g01h01);
-  __m128i a1b1c1d1e1f1g1h1 = _mm_unpackhi_epi64(a01b01c01d01, e01f01g01h01);
-  __m128i a2b2c2d2e2f2g2h2 = _mm_unpacklo_epi64(a23b23c23d23, e23f23g23h23);
-  __m128i a3b3c3d3e3f3g3h3 = _mm_unpackhi_epi64(a23b23c23d23, e23f23g23h23);
-  __m128i a4b4c4d4e4f4g4h4 = _mm_unpacklo_epi64(a45b45c45d45, e45f45g45h45);
-  __m128i a5b5c5d5e5f5g5h5 = _mm_unpackhi_epi64(a45b45c45d45, e45f45g45h45);
-  __m128i a6b6c6d6e6f6g6h6 = _mm_unpacklo_epi64(a67b67c67d67, e67f67g67h67);
-  __m128i a7b7c7d7e7f7g7h7 = _mm_unpackhi_epi64(a67b67c67d67, e67f67g67h67);
-
-  kernel.packet[0].x = a0b0c0d0e0f0g0h0;
-  kernel.packet[1].x = a1b1c1d1e1f1g1h1;
-  kernel.packet[2].x = a2b2c2d2e2f2g2h2;
-  kernel.packet[3].x = a3b3c3d3e3f3g3h3;
-  kernel.packet[4].x = a4b4c4d4e4f4g4h4;
-  kernel.packet[5].x = a5b5c5d5e5f5g5h5;
-  kernel.packet[6].x = a6b6c6d6e6f6g6h6;
-  kernel.packet[7].x = a7b7c7d7e7f7g7h7;
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet8h,4>& kernel) {
-  EIGEN_ALIGN32 Eigen::half in[4][8];
-  pstore<Eigen::half>(in[0], kernel.packet[0]);
-  pstore<Eigen::half>(in[1], kernel.packet[1]);
-  pstore<Eigen::half>(in[2], kernel.packet[2]);
-  pstore<Eigen::half>(in[3], kernel.packet[3]);
-
-  EIGEN_ALIGN32 Eigen::half out[4][8];
-
-  for (int i = 0; i < 4; ++i) {
-    for (int j = 0; j < 4; ++j) {
-      out[i][j] = in[j][2*i];
-    }
-    for (int j = 0; j < 4; ++j) {
-      out[i][j+4] = in[j][2*i+1];
-    }
-  }
-
-  kernel.packet[0] = pload<Packet8h>(out[0]);
-  kernel.packet[1] = pload<Packet8h>(out[1]);
-  kernel.packet[2] = pload<Packet8h>(out[2]);
-  kernel.packet[3] = pload<Packet8h>(out[3]);
-}
-
-
-// Disable the following code since it's broken on too many platforms / compilers.
-//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
-#elif 0
-
-typedef struct {
-  __m64 x;
-} Packet4h;
-
-
-template<> struct is_arithmetic<Packet4h> { enum { value = true }; };
-
-template <>
-struct packet_traits<Eigen::half> : default_packet_traits {
-  typedef Packet4h type;
-  // There is no half-size packet for Packet4h.
-  typedef Packet4h half;
-  enum {
-    Vectorizable = 1,
-    AlignedOnScalar = 1,
-    size = 4,
-    HasHalfPacket = 0,
-    HasAdd    = 0,
-    HasSub    = 0,
-    HasMul    = 0,
-    HasNegate = 0,
-    HasAbs    = 0,
-    HasAbs2   = 0,
-    HasMin    = 0,
-    HasMax    = 0,
-    HasConj   = 0,
-    HasSetLinear = 0,
-    HasDiv = 0,
-    HasSqrt = 0,
-    HasRsqrt = 0,
-    HasExp = 0,
-    HasLog = 0,
-    HasBlend = 0
-  };
-};
-
-
-template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; };
-
-template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) {
-  Packet4h result;
-  result.x = _mm_set1_pi16(from.x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) {
-  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; }
-
-template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const Packet4h& b) {
-  __int64_t a64 = _mm_cvtm64_si64(a.x);
-  __int64_t b64 = _mm_cvtm64_si64(b.x);
-
-  Eigen::half h[4];
-
-  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
-  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
-  h[0] = ha + hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
-  h[1] = ha + hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
-  h[2] = ha + hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
-  h[3] = ha + hb;
-  Packet4h result;
-  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const Packet4h& b) {
-  __int64_t a64 = _mm_cvtm64_si64(a.x);
-  __int64_t b64 = _mm_cvtm64_si64(b.x);
-
-  Eigen::half h[4];
-
-  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
-  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
-  h[0] = ha * hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
-  h[1] = ha * hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
-  h[2] = ha * hb;
-  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
-  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
-  h[3] = ha * hb;
-  Packet4h result;
-  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(const Eigen::half* from) {
-  Packet4h result;
-  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(const Eigen::half* from) {
-  Packet4h result;
-  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet4h& from) {
-  __int64_t r = _mm_cvtm64_si64(from.x);
-  *(reinterpret_cast<__int64_t*>(to)) = r;
-}
-
-template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4h& from) {
-  __int64_t r = _mm_cvtm64_si64(from.x);
-  *(reinterpret_cast<__int64_t*>(to)) = r;
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h
-ploadquad<Packet4h>(const Eigen::half* from) {
-  return pset1<Packet4h>(*from);
-}
-
-template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(const Eigen::half* from, Index stride)
-{
-  Packet4h result;
-  result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
-  return result;
-}
-
-template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4h>(Eigen::half* to, const Packet4h& from, Index stride)
-{
-  __int64_t a = _mm_cvtm64_si64(from.x);
-  to[stride*0].x = static_cast<unsigned short>(a);
-  to[stride*1].x = static_cast<unsigned short>(a >> 16);
-  to[stride*2].x = static_cast<unsigned short>(a >> 32);
-  to[stride*3].x = static_cast<unsigned short>(a >> 48);
-}
-
-EIGEN_STRONG_INLINE void
-ptranspose(PacketBlock<Packet4h,4>& kernel) {
-  __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x);
-  __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x);
-  __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x);
-  __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x);
-
-  kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1);
-  kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1);
-  kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3);
-  kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3);
-}
-
-#endif
-
-}
-}
-
-#endif // EIGEN_PACKET_MATH_HALF_HIP_H
diff --git a/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h b/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h
deleted file mode 100644
index 915266a9d..000000000
--- a/Eigen/src/Core/arch/HIP/hcc/TypeCasting.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_TYPE_CASTING_HIP_H
-#define EIGEN_TYPE_CASTING_HIP_H
-
-namespace Eigen {
-
-namespace internal {
-
-template<>
-struct scalar_cast_op<float, Eigen::half> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
-  typedef Eigen::half result_type;
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
-    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-      return __float2half(a);
-    #else
-      return Eigen::half(a);
-    #endif
-  }
-};
-
-template<>
-struct functor_traits<scalar_cast_op<float, Eigen::half> >
-{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
-
-
-template<>
-struct scalar_cast_op<int, Eigen::half> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
-  typedef Eigen::half result_type;
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
-    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-    return __float2half(static_cast<float>(a));
-    #else
-      return Eigen::half(static_cast<float>(a));
-    #endif
-  }
-};
-
-template<>
-struct functor_traits<scalar_cast_op<int, Eigen::half> >
-{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
-
-
-template<>
-struct scalar_cast_op<Eigen::half, float> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
-  typedef float result_type;
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
-    #if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-      return __half2float(a);
-    #else
-      return static_cast<float>(a);
-    #endif
-  }
-};
-
-template<>
-struct functor_traits<scalar_cast_op<Eigen::half, float> >
-{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
-
-
-
-#if defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE)
-
-template <>
-struct type_casting_traits<Eigen::half, float> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 2,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
-  float2 r1 = __half22float2(a);
-  float2 r2 = __half22float2(b);
-  return make_float4(r1.x, r1.y, r2.x, r2.y);
-}
-
-template <>
-struct type_casting_traits<float, Eigen::half> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 2
-  };
-};
-
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
-  // Simply discard the second half of the input
-  return __floats2half2_rn(a.x, a.y);
-}
-
-#elif defined EIGEN_VECTORIZE_AVX512
-template <>
-struct type_casting_traits<half, float> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) {
-  return half2float(a);
-}
-
-template <>
-struct type_casting_traits<float, half> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) {
-  return float2half(a);
-}
-
-#elif defined EIGEN_VECTORIZE_AVX
-
-template <>
-struct type_casting_traits<Eigen::half, float> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
-  return half2float(a);
-}
-
-template <>
-struct type_casting_traits<float, Eigen::half> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
-  return float2half(a);
-}
-
-// Disable the following code since it's broken on too many platforms / compilers.
-//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
-#elif 0
-
-template <>
-struct type_casting_traits<Eigen::half, float> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4h, Packet4f>(const Packet4h& a) {
-  __int64_t a64 = _mm_cvtm64_si64(a.x);
-  Eigen::half h = raw_uint16_to_half(static_cast<unsigned short>(a64));
-  float f1 = static_cast<float>(h);
-  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
-  float f2 = static_cast<float>(h);
-  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
-  float f3 = static_cast<float>(h);
-  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
-  float f4 = static_cast<float>(h);
-  return _mm_set_ps(f4, f3, f2, f1);
-}
-
-template <>
-struct type_casting_traits<float, Eigen::half> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 1
-  };
-};
-
-template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f& a) {
-  EIGEN_ALIGN16 float aux[4];
-  pstore(aux, a);
-  Eigen::half h0(aux[0]);
-  Eigen::half h1(aux[1]);
-  Eigen::half h2(aux[2]);
-  Eigen::half h3(aux[3]);
-
-  Packet4h result;
-  result.x = _mm_set_pi16(h3.x, h2.x, h1.x, h0.x);
-  return result;
-}
-
-#endif
-
-} // end namespace internal
-
-} // end namespace Eigen
-
-#endif // EIGEN_TYPE_CASTING_HIP_H

From b6cc0961b17f6204038158c445eddf411c97a3e2 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Thu, 14 Jun 2018 10:21:54 -0400
Subject: [PATCH 396/680] updates based on PR feedback

There are two major changes (and a few minor ones which are not listed here...see PR discussion for details)

1. Eigen::half implementations for HIP and CUDA have been merged.
This means that
- `CUDA/Half.h` and `HIP/hcc/Half.h` got merged to a new file `GPU/Half.h`
- `CUDA/PacketMathHalf.h` and `HIP/hcc/PacketMathHalf.h` got merged to a new file `GPU/PacketMathHalf.h`
- `CUDA/TypeCasting.h` and `HIP/hcc/TypeCasting.h` got merged to a new file `GPU/TypeCasting.h`

After this change the `HIP/hcc` directory only contains one file `math_constants.h`. That will go away too once that file becomes a part of the HIP install.

2. new macros EIGEN_GPUCC, EIGEN_GPU_COMPILE_PHASE and EIGEN_HAS_GPU_FP16 have been added and the code has been updated to use them where appropriate.
- `EIGEN_GPUCC` is the same as `(EIGEN_CUDACC || EIGEN_HIPCC)`
- `EIGEN_GPU_DEVICE_COMPILE` is the same as `(EIGEN_CUDA_ARCH || EIGEN_HIP_DEVICE_COMPILE)`
- `EIGEN_HAS_GPU_FP16` is the same as `(EIGEN_HAS_CUDA_FP16 or EIGEN_HAS_HIP_FP16)`
---
 Eigen/Core                                    |  71 +++++++--
 Eigen/src/Core/GeneralProduct.h               |   2 +-
 Eigen/src/Core/GenericPacketMath.h            |  12 +-
 Eigen/src/Core/MathFunctions.h                |  56 +++----
 Eigen/src/Core/ProductEvaluators.h            |  18 +--
 Eigen/src/Core/arch/GPU/Half.h                | 131 ++++++++++++----
 Eigen/src/Core/arch/GPU/PacketMathHalf.h      | 145 +++++++++++++++++-
 Eigen/src/Core/arch/GPU/TypeCasting.h         |  18 ++-
 Eigen/src/Core/functors/AssignmentFunctors.h  |   2 +-
 Eigen/src/Core/functors/BinaryFunctors.h      |  10 +-
 Eigen/src/Core/util/BlasUtil.h                |   5 +-
 Eigen/src/Core/util/Memory.h                  |   8 +-
 Eigen/src/Core/util/Meta.h                    |  29 ++--
 Eigen/src/SVD/BDCSVD.h                        |   2 +-
 test/half_float.cpp                           |   2 +-
 test/main.h                                   |  16 +-
 .../CXX11/src/Tensor/TensorDeviceDefault.h    |  14 +-
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |   4 +-
 .../Eigen/CXX11/src/Tensor/TensorIntDiv.h     |  12 +-
 .../Eigen/CXX11/src/Tensor/TensorMacros.h     |   2 +-
 .../Eigen/CXX11/src/Tensor/TensorMeta.h       |   2 +-
 .../Eigen/CXX11/src/Tensor/TensorRandom.h     |   2 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  10 +-
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |   4 +-
 unsupported/Eigen/CXX11/src/util/CXX11Meta.h  |  20 +--
 .../Eigen/CXX11/src/util/EmulateArray.h       |   2 +-
 .../SpecialFunctions/SpecialFunctionsImpl.h   |   4 +-
 .../arch/CUDA/CudaSpecialFunctions.h          |   2 +-
 unsupported/test/cxx11_tensor_hip.cu          |   5 +-
 29 files changed, 425 insertions(+), 185 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index c72d5468a..f67bffd12 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -99,6 +99,61 @@
 #define EIGEN_DONT_VECTORIZE
 #endif
 
+
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// If either EIGEN_CUDACC or EIGEN_HIPCC is defined, then define EIGEN_GPUCC
+//
+#define EIGEN_GPUCC
+//
+// EIGEN_HIPCC implies the HIP compiler and is used to tweak Eigen code for use in HIP kernels
+// EIGEN_CUDACC implies the CUDA compiler and is used to tweak Eigen code for use in CUDA kernels
+//
+// In most cases the same tweaks are required to the Eigen code to enable in both the HIP and CUDA kernels.
+// For those cases, the corresponding code should be guarded with
+//      #if defined(EIGEN_GPUCC)
+// instead of
+//      #if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDACC)
+//
+#endif
+
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// If either EIGEN_CUDA_ARCH or EIGEN_HIP_DEVICE_COMPILE is defined, then define EIGEN_GPU_COMPILE_PHASE
+//
+#define EIGEN_GPU_COMPILE_PHASE
+//
+// GPU compilers (HIPCC, NVCC) typically do two passes over the source code,
+//   + one to compile the source for the "host" (ie CPU)
+//   + another to compile the source for the "device" (ie. GPU)
+// 
+// Code that needs to enabled only during the either the "host" or "device" compilation phase
+// needs to be guarded with a macro that indicates the current compilation phase
+//
+// EIGEN_HIP_DEVICE_COMPILE implies the device compilation phase in HIP
+// EIGEN_CUDA_ARCH implies the device compilation phase in CUDA
+//
+// In most cases, the "host" / "device" specific code is the same for both HIP and CUDA
+// For those cases, the code should be guarded with
+//       #if defined(EIGEN_GPU_COMPILE_PHASE)
+// instead of
+//       #if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDA_ARCH)
+//
+#endif
+
+
 // When compiling CUDA device code with NVCC, or HIP device code with HIPCC
 // pull in math functions from the global namespace.  In host mode, and when
 // device doee with clang, use the std versions.
@@ -312,6 +367,10 @@
   #endif
 #endif
 
+#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
+  #define EIGEN_HAS_GPU_FP16
+#endif
+
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
   #define EIGEN_HAS_OPENMP
 #endif
@@ -475,15 +534,9 @@ using std::ptrdiff_t;
 #endif
 
 // Half float support
-#if defined EIGEN_USE_HIP
-  #include "src/Core/arch/HIP/hcc/Half.h"
-  #include "src/Core/arch/HIP/hcc/PacketMathHalf.h"
-  #include "src/Core/arch/HIP/hcc/TypeCasting.h"
-#else
-  #include "src/Core/arch/CUDA/Half.h"
-  #include "src/Core/arch/CUDA/PacketMathHalf.h"
-  #include "src/Core/arch/CUDA/TypeCasting.h"
-#endif
+#include "src/Core/arch/GPU/Half.h"
+#include "src/Core/arch/GPU/PacketMathHalf.h"
+#include "src/Core/arch/GPU/TypeCasting.h"
 
 #if defined EIGEN_VECTORIZE_CUDA
   #include "src/Core/arch/CUDA/PacketMath.h"
diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 694f7cbde..261f77b99 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -35,7 +35,7 @@ template<int Rows, int Cols, int Depth> struct product_type_selector;
 template<int Size, int MaxSize> struct product_size_category
 {
   enum {
-    #ifndef EIGEN_CUDA_ARCH
+    #ifndef EIGEN_GPU_COMPILE_PHASE
     is_large = MaxSize == Dynamic ||
                Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
                (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 2603bd2f7..b67c41d8a 100644
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -301,13 +301,11 @@ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu
  { pstore(to, from); }
 
 /** \internal tries to do cache prefetching of \a addr */
-template<typename Scalar>
-  #if !defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  inline void prefetch(const Scalar* addr)
+template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
 {
-#ifdef EIGEN_CUDA_ARCH
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+  // do nothing
+#elif defined(EIGEN_CUDA_ARCH)
 #if defined(__LP64__)
   // 64-bit pointer operand constraint for inlined asm
   asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
@@ -534,7 +532,7 @@ inline void palign(PacketType& first, const PacketType& second)
 ***************************************************************************/
 
 // Eigen+CUDA does not support complexes.
-#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
+#if !defined(EIGEN_GPUCC)
 
 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
 { return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index fe6d6585c..6415a7696 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -96,7 +96,7 @@ struct real_default_impl<Scalar,true>
 
 template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
 
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 template<typename T>
 struct real_impl<std::complex<T> >
 {
@@ -144,7 +144,7 @@ struct imag_default_impl<Scalar,true>
 
 template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
 
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 template<typename T>
 struct imag_impl<std::complex<T> >
 {
@@ -260,7 +260,7 @@ struct conj_default_impl<Scalar,true>
 
 template<typename Scalar> struct conj_impl : conj_default_impl<Scalar> {};
 
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 template<typename T>
 struct conj_impl<std::complex<T> >
 {
@@ -773,9 +773,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isfinite_impl(const T& x)
 {
-  #if defined(EIGEN_HIP_DEVICE_COMPILE)
-    return isfinite(x);
-  #elif defined(EIGEN_CUDA_ARCH)
+  #if defined(EIGEN_GPU_COMPILE_PHASE)
     return (::isfinite)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isfinite;
@@ -790,9 +788,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isinf_impl(const T& x)
 {
-  #if defined(EIGEN_HIP_DEVICE_COMPILE)
-    return isinf(x);
-  #elif defined(EIGEN_CUDA_ARCH)
+  #if defined(EIGEN_GPU_COMPILE_PHASE)
     return (::isinf)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isinf;
@@ -807,9 +803,7 @@ EIGEN_DEVICE_FUNC
 typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
 isnan_impl(const T& x)
 {
-  #if defined(EIGEN_HIP_DEVICE_COMPILE)
-    return isnan(x);
-  #elif defined(EIGEN_CUDA_ARCH)
+  #if defined(EIGEN_GPU_COMPILE_PHASE)
     return (::isnan)(x);
   #elif EIGEN_USE_STD_FPCLASSIFY
     using std::isnan;
@@ -875,7 +869,7 @@ template<typename T> T generic_fast_tanh_float(const T& a_x);
 
 namespace numext {
 
-#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE) && !defined(__SYCL_DEVICE_ONLY__)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(__SYCL_DEVICE_ONLY__)
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
@@ -1089,7 +1083,7 @@ EIGEN_ALWAYS_INLINE float   log1p(float x) { return cl::sycl::log1p(x); }
 EIGEN_ALWAYS_INLINE double  log1p(double x) { return cl::sycl::log1p(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log1p(const float &x) { return ::log1pf(x); }
 
@@ -1147,7 +1141,7 @@ EIGEN_ALWAYS_INLINE float   floor(float x) { return cl::sycl::floor(x); }
 EIGEN_ALWAYS_INLINE double  floor(double x) { return cl::sycl::floor(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float floor(const float &x) { return ::floorf(x); }
 
@@ -1168,7 +1162,7 @@ EIGEN_ALWAYS_INLINE float   ceil(float x) { return cl::sycl::ceil(x); }
 EIGEN_ALWAYS_INLINE double  ceil(double x) { return cl::sycl::ceil(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float ceil(const float &x) { return ::ceilf(x); }
 
@@ -1226,7 +1220,7 @@ EIGEN_ALWAYS_INLINE double  log(double x) { return cl::sycl::log(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float log(const float &x) { return ::logf(x); }
 
@@ -1254,7 +1248,7 @@ EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
 EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float abs(const float &x) { return ::fabsf(x); }
 
@@ -1284,7 +1278,7 @@ EIGEN_ALWAYS_INLINE float   exp(float x) { return cl::sycl::exp(x); }
 EIGEN_ALWAYS_INLINE double  exp(double x) { return cl::sycl::exp(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float exp(const float &x) { return ::expf(x); }
 
@@ -1320,7 +1314,7 @@ EIGEN_ALWAYS_INLINE float   expm1(float x) { return cl::sycl::expm1(x); }
 EIGEN_ALWAYS_INLINE double  expm1(double x) { return cl::sycl::expm1(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float expm1(const float &x) { return ::expm1f(x); }
 
@@ -1340,7 +1334,7 @@ EIGEN_ALWAYS_INLINE float   cos(float x) { return cl::sycl::cos(x); }
 EIGEN_ALWAYS_INLINE double  cos(double x) { return cl::sycl::cos(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cos(const float &x) { return ::cosf(x); }
 
@@ -1360,7 +1354,7 @@ EIGEN_ALWAYS_INLINE float   sin(float x) { return cl::sycl::sin(x); }
 EIGEN_ALWAYS_INLINE double  sin(double x) { return cl::sycl::sin(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sin(const float &x) { return ::sinf(x); }
 
@@ -1380,7 +1374,7 @@ EIGEN_ALWAYS_INLINE float   tan(float x) { return cl::sycl::tan(x); }
 EIGEN_ALWAYS_INLINE double  tan(double x) { return cl::sycl::tan(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tan(const float &x) { return ::tanf(x); }
 
@@ -1411,7 +1405,7 @@ EIGEN_ALWAYS_INLINE float   acosh(float x) { return cl::sycl::acosh(x); }
 EIGEN_ALWAYS_INLINE double  acosh(double x) { return cl::sycl::acosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float acos(const float &x) { return ::acosf(x); }
 
@@ -1442,7 +1436,7 @@ EIGEN_ALWAYS_INLINE float   asinh(float x) { return cl::sycl::asinh(x); }
 EIGEN_ALWAYS_INLINE double  asinh(double x) { return cl::sycl::asinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float asin(const float &x) { return ::asinf(x); }
 
@@ -1473,7 +1467,7 @@ EIGEN_ALWAYS_INLINE float   atanh(float x) { return cl::sycl::atanh(x); }
 EIGEN_ALWAYS_INLINE double  atanh(double x) { return cl::sycl::atanh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float atan(const float &x) { return ::atanf(x); }
 
@@ -1494,7 +1488,7 @@ EIGEN_ALWAYS_INLINE float   cosh(float x) { return cl::sycl::cosh(x); }
 EIGEN_ALWAYS_INLINE double  cosh(double x) { return cl::sycl::cosh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float cosh(const float &x) { return ::coshf(x); }
 
@@ -1514,7 +1508,7 @@ EIGEN_ALWAYS_INLINE float   sinh(float x) { return cl::sycl::sinh(x); }
 EIGEN_ALWAYS_INLINE double  sinh(double x) { return cl::sycl::sinh(x); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float sinh(const float &x) { return ::sinhf(x); }
 
@@ -1532,12 +1526,12 @@ T tanh(const T &x) {
 #if defined(__SYCL_DEVICE_ONLY__)
 EIGEN_ALWAYS_INLINE float   tanh(float x) { return cl::sycl::tanh(x); }
 EIGEN_ALWAYS_INLINE double  tanh(double x) { return cl::sycl::tanh(x); }
-#elif (!defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)) && EIGEN_FAST_MATH
+#elif (!defined(EIGEN_GPUCC)) && EIGEN_FAST_MATH
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(float x) { return internal::generic_fast_tanh_float(x); }
 #endif
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(const float &x) { return ::tanhf(x); }
 
@@ -1557,7 +1551,7 @@ EIGEN_ALWAYS_INLINE float   fmod(float x, float y) { return cl::sycl::fmod(x, y)
 EIGEN_ALWAYS_INLINE double  fmod(double x, double y) { return cl::sycl::fmod(x, y); }
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template <>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float fmod(const float& a, const float& b) {
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index e0868daf5..2bb42f74b 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -137,14 +137,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scal
   typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
 {
   typedef Product<Lhs,Rhs,Options> SrcXprType;
-<<<<<<< local
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif  
-  static EIGEN_STRONG_INLINE
-=======
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
->>>>>>> other
   void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
     Index dstRows = src.rows();
@@ -397,14 +390,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   template<typename Dst>
-<<<<<<< local
-  #if defined(EIGEN_HIPCC)  
-  EIGEN_DEVICE_FUNC
-  #endif  
-  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
-=======
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
->>>>>>> other
   {
     // Same as: dst.noalias() = lhs.lazyProduct(rhs);
     // but easier on the compiler side
@@ -865,7 +851,7 @@ struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalSha
     return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col);
   }
   
-#ifndef EIGEN_CUDACC
+#ifndef EIGEN_GPUCC
   template<int LoadMode,typename PacketType>
   EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
   {
@@ -909,7 +895,7 @@ struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape,
     return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col);
   }
   
-#ifndef EIGEN_CUDACC
+#ifndef EIGEN_GPUCC
   template<int LoadMode,typename PacketType>
   EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
   {
diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index c10550050..ab9d27591 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -26,15 +26,15 @@
 
 
 // Standard 16-bit float type, mostly useful for GPUs. Defines a new
-// type Eigen::half (inheriting from CUDA's __half struct) with
+// type Eigen::half (inheriting either from CUDA's or HIP's __half struct) with
 // operator overloads such that it behaves basically as an arithmetic
 // type. It will be quite slow on CPUs (so it is recommended to stay
 // in fp32 for CPUs, except for simple parameter conversions, I/O
 // to disk and the likes), but fast on GPUs.
 
 
-#ifndef EIGEN_HALF_CUDA_H
-#define EIGEN_HALF_CUDA_H
+#ifndef EIGEN_HALF_GPU_H
+#define EIGEN_HALF_GPU_H
 
 #if __cplusplus > 199711L
 #define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
@@ -49,16 +49,41 @@ struct half;
 
 namespace half_impl {
 
-#if !defined(EIGEN_HAS_CUDA_FP16)
+#if !defined(EIGEN_HAS_GPU_FP16)
 // Make our own __half_raw definition that is similar to CUDA's.
 struct __half_raw {
   EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
   explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
   unsigned short x;
 };
-#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+#elif defined(EIGEN_HAS_HIP_FP16)
+ #if defined(EIGEN_HAS_OLD_HIP_FP16)
+// Make a __half_raw definition that is
+// ++ compatible with that of Eigen and
+// ++ add a implcit conversion to the native __half of the old HIP implementation.
+//
+// Keeping ".x" as "unsigned short" keeps the interface the same between the Eigen and HIP implementation.
+//
+// In the old HIP implementation,
+//   ++ __half is a typedef of __fp16
+//   ++ the "__h*" routines take "__half" arguments
+// so we need to implicitly convert "__half_raw" to "__half" to avoid having to explicitly make 
+// that conversiion in each call to a "__h*" routine...that is why we have "operator __half" routine
+struct __half_raw {
+  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
+  union {
+    unsigned short x;
+    __half data;
+  };
+  operator __half(void) const { return data; }
+};
+ #endif
+#elif defined(EIGEN_HAS_CUDA_FP16)
+ #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
 // In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
-typedef __half __half_raw;
+ typedef __half __half_raw;
+ #endif
 #endif
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
@@ -69,8 +94,19 @@ struct half_base : public __half_raw {
   EIGEN_DEVICE_FUNC half_base() {}
   EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
   EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+
+#if defined(EIGEN_HAS_GPU_FP16)
+ #if defined(EIGEN_HAS_HIP_FP16)
+  #if defined(EIGEN_HAS_OLD_HIP_FP16)
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(__half_as_ushort(h)) {}
+  #else
+  EIGEN_DEVICE_FUNC half_base(const __half& h) { x = __half_as_ushort(h); }
+  #endif
+ #elif defined(EIGEN_HAS_CUDA_FP16)
+  #if (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000)
   EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
+  #endif
+ #endif    
 #endif
 };
 
@@ -78,17 +114,38 @@ struct half_base : public __half_raw {
 
 // Class definition.
 struct half : public half_impl::half_base {
-  #if !defined(EIGEN_HAS_CUDA_FP16) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
-    typedef half_impl::__half_raw __half_raw;
-  #endif
+
+  // Writing this out as separate #if-else blocks to make the code easier to follow
+  // The same applies to most #if-else blocks in this file
+#if !defined(EIGEN_HAS_GPU_FP16)
+  typedef half_impl::__half_raw __half_raw;
+#elif defined(EIGEN_HAS_HIP_FP16)
+ #if defined(EIGEN_HAS_OLD_HIP_FP16)
+  typedef half_impl::__half_raw __half_raw;
+ #endif
+#elif defined(EIGEN_HAS_CUDA_FP16)
+  // Note that EIGEN_CUDACC_VER is set to 0 even when compiling with HIP, so (EIGEN_CUDACC_VER < 90000) is true even for HIP!
+  // So keeping this within #if defined(EIGEN_HAS_CUDA_FP16) is needed
+ #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000  
+  typedef half_impl::__half_raw __half_raw;
+ #endif
+#endif
 
   EIGEN_DEVICE_FUNC half() {}
 
   EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
   EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  
+#if defined(EIGEN_HAS_GPU_FP16)
+ #if defined(EIGEN_HAS_HIP_FP16)
   EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+ #elif defined(EIGEN_HAS_CUDA_FP16)
+  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  #endif
+ #endif
 #endif
+  
 
   explicit EIGEN_DEVICE_FUNC half(bool b)
       : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
@@ -201,7 +258,8 @@ namespace Eigen {
 
 namespace half_impl {
 
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(HIP_DEVICE_COMPILE))
 
 // Intrinsics for native fp16 support. Note that on current hardware,
 // these are no faster than fp32 arithmetic (you need to use the half2
@@ -262,7 +320,7 @@ EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
 
 #else  // Emulate support for half floats
 
-// Definitions for CPUs and older CUDA, mostly working through conversion
+// Definitions for CPUs and older HIP+CUDA, mostly working through conversion
 // to/from fp32.
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
@@ -342,7 +400,8 @@ union FP32 {
 };
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   __half tmp_ff = __float2half(ff);
   return *(__half_raw*)&tmp_ff;
 
@@ -398,7 +457,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
 }
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   return __half2float(h);
 
 #elif defined(EIGEN_HAS_FP16_C)
@@ -432,7 +492,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
   return (a.x & 0x7fff) == 0x7c00;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   return __hisnan(a);
 #else
   return (a.x & 0x7fff) > 0x7c00;
@@ -448,7 +509,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
   return result;
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
-#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
   return half(hexp(a));
 #else
    return half(::expf(float(a)));
@@ -458,7 +520,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
   return half(numext::expm1(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
-#if defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   return half(::hlog(a));
 #else
   return half(::logf(float(a)));
@@ -471,7 +534,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
   return half(::log10f(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
-#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
   return half(hsqrt(a));
 #else
     return half(::sqrtf(float(a)));
@@ -493,14 +557,16 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
   return half(::tanhf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
-#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
+#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
   return half(hfloor(a));
 #else
   return half(::floorf(float(a)));
 #endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
-#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
+#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
   return half(hceil(a));
 #else
   return half(::ceilf(float(a)));
@@ -508,7 +574,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
 }
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   return __hlt(b, a) ? b : a;
 #else
   const float f1 = static_cast<float>(a);
@@ -517,7 +584,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
 #endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
   return __hlt(a, b) ? b : a;
 #else
   const float f1 = static_cast<float>(a);
@@ -595,7 +663,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
   return Eigen::half(::expf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
-#if EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
+#if (EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
   return Eigen::half(::hlog(a));
 #else
   return Eigen::half(::logf(float(a)));
@@ -629,9 +698,12 @@ struct hash<Eigen::half> {
 
 
 // Add the missing shfl_xor intrinsic
-#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
+
 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
-  #if EIGEN_CUDACC_VER < 90000
+  #if (EIGEN_CUDACC_VER < 90000) || \
+    defined(EIGEN_HAS_HIP_FP16)
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
   #else
   return static_cast<Eigen::half>(__shfl_xor_sync(0xFFFFFFFF, static_cast<float>(var), laneMask, width));
@@ -640,7 +712,8 @@ __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneM
 #endif
 
 // ldg() has an overload for __half_raw, but we also need one for Eigen::half.
-#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
+#if (defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
   return Eigen::half_impl::raw_uint16_to_half(
       __ldg(reinterpret_cast<const unsigned short*>(ptr)));
@@ -648,7 +721,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr)
 #endif
 
 
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 namespace Eigen {
 namespace numext {
 
@@ -674,4 +747,4 @@ bool (isfinite)(const Eigen::half& h) {
 }  // namespace numext
 #endif
 
-#endif // EIGEN_HALF_CUDA_H
+#endif // EIGEN_HALF_GPU_H
diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index 8a6f209c4..e1ecac1ab 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -7,15 +7,16 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_PACKET_MATH_HALF_CUDA_H
-#define EIGEN_PACKET_MATH_HALF_CUDA_H
+#ifndef EIGEN_PACKET_MATH_HALF_GPU_H
+#define EIGEN_PACKET_MATH_HALF_GPU_H
 
 
 namespace Eigen {
 namespace internal {
 
 // Most of the following operations require arch >= 3.0
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIPCC) && defined(EIGEN_HIP_DEVICE_COMPILE))
 
 template<> struct is_arithmetic<half2> { enum { value = true }; };
 
@@ -43,7 +44,18 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits
 template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return half2half2(from);
+#else  
   return __half2half2(from);
+#endif
+
+#else // EIGEN_CUDA_ARCH
+  return __half2half2(from);
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
@@ -69,20 +81,46 @@ template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half*
 
 template<>
  __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return __halves2half2((*(from+0)), (*(from+1)));
+#else
+  return __ldg((const half2*)from);
+#endif
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 350
    return __ldg((const half2*)from);
 #else
   return __halves2half2(*(from+0), *(from+1));
 #endif
+
+#endif
 }
 
 template<>
 __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return __halves2half2((*(from+0)), (*(from+1)));
+#else
+  return __halves2half2(__ldg(from+0), __ldg(from+1));
+#endif
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 350
    return __halves2half2(__ldg(from+0), __ldg(from+1));
 #else
   return __halves2half2(*(from+0), *(from+1));
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
@@ -117,15 +155,29 @@ ptranspose(PacketBlock<half2,2>& kernel) {
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+  
+  return __halves2half2(a, __hadd(a, __float2half(1.0f)));
+  
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __halves2half2(a, __hadd(a, __float2half(1.0f)));
 #else
   float f = __half2float(a) + 1.0f;
   return __halves2half2(a, __float2half(f));
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hadd2(a, b);
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hadd2(a, b);
 #else
@@ -137,9 +189,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, cons
   float r2 = a2 + b2;
   return __floats2half2_rn(r1, r2);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hsub2(a, b);
+  
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hsub2(a, b);
 #else
@@ -151,9 +211,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, cons
   float r2 = a2 - b2;
   return __floats2half2_rn(r1, r2);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hneg2(a);
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hneg2(a);
 #else
@@ -161,11 +229,19 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
   float a2 = __high2float(a);
   return __floats2half2_rn(-a1, -a2);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hmul2(a, b);
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hmul2(a, b);
 #else
@@ -177,9 +253,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, cons
   float r2 = a2 * b2;
   return __floats2half2_rn(r1, r2);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+   return __hfma2(a, b, c);
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
    return __hfma2(a, b, c);
 #else
@@ -193,9 +277,21 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, con
   float r2 = a2 * b2 + c2;
   return __floats2half2_rn(r1, r2);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+  
+#if defined(EIGEN_HAS_OLD_HIP_FP16)
+  return h2div(a, b);
+#else
+  return __h2div(a, b);
+#endif
+  
+#else // EIGEN_CUDA_ARCH
+  
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -203,6 +299,8 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, cons
   float r1 = a1 / b1;
   float r2 = a2 / b2;
   return __floats2half2_rn(r1, r2);
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
@@ -226,6 +324,12 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, cons
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hadd(__low2half(a), __high2half(a));
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hadd(__low2half(a), __high2half(a));
 #else
@@ -233,9 +337,19 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
   float a2 = __high2float(a);
   return Eigen::half(__float2half(a1 + a2));
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hgt(first, second) ? first : second;
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
@@ -245,9 +359,19 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const ha
   float a2 = __high2float(a);
   return a1 > a2 ? __low2half(a) : __high2half(a);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hlt(first, second) ? first : second;
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
@@ -257,9 +381,17 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const ha
   float a2 = __high2float(a);
   return a1 < a2 ? __low2half(a) : __high2half(a);
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  return __hmul(__low2half(a), __high2half(a));
+
+#else  // EIGEN_CUDA_ARCH
+
 #if EIGEN_CUDA_ARCH >= 530
   return __hmul(__low2half(a), __high2half(a));
 #else
@@ -267,6 +399,8 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const ha
   float a2 = __high2float(a);
   return Eigen::half(__float2half(a1 * a2));
 #endif
+
+#endif
 }
 
 template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
@@ -285,7 +419,8 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
+#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
+  defined(EIGEN_HIP_DEVICE_COMPILE)
 
 template<>  __device__ EIGEN_STRONG_INLINE
 half2 plog<half2>(const half2& a) {
@@ -1130,4 +1265,4 @@ ptranspose(PacketBlock<Packet4h,4>& kernel) {
 }
 }
 
-#endif // EIGEN_PACKET_MATH_HALF_CUDA_H
+#endif // EIGEN_PACKET_MATH_HALF_GPU_H
diff --git a/Eigen/src/Core/arch/GPU/TypeCasting.h b/Eigen/src/Core/arch/GPU/TypeCasting.h
index 30f870c3d..57a55d08b 100644
--- a/Eigen/src/Core/arch/GPU/TypeCasting.h
+++ b/Eigen/src/Core/arch/GPU/TypeCasting.h
@@ -7,8 +7,8 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_TYPE_CASTING_CUDA_H
-#define EIGEN_TYPE_CASTING_CUDA_H
+#ifndef EIGEN_TYPE_CASTING_GPU_H
+#define EIGEN_TYPE_CASTING_GPU_H
 
 namespace Eigen {
 
@@ -19,7 +19,8 @@ struct scalar_cast_op<float, Eigen::half> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef Eigen::half result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
       return __float2half(a);
     #else
       return Eigen::half(a);
@@ -37,7 +38,8 @@ struct scalar_cast_op<int, Eigen::half> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef Eigen::half result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
       return __float2half(static_cast<float>(a));
     #else
       return Eigen::half(static_cast<float>(a));
@@ -55,7 +57,8 @@ struct scalar_cast_op<Eigen::half, float> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
   typedef float result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
-    #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
       return __half2float(a);
     #else
       return static_cast<float>(a);
@@ -69,7 +72,8 @@ struct functor_traits<scalar_cast_op<Eigen::half, float> >
 
 
 
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
 
 template <>
 struct type_casting_traits<Eigen::half, float> {
@@ -209,4 +213,4 @@ template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f
 
 } // end namespace Eigen
 
-#endif // EIGEN_TYPE_CASTING_CUDA_H
+#endif // EIGEN_TYPE_CASTING_GPU_H
diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h
index 1077d8eb0..9765cc763 100644
--- a/Eigen/src/Core/functors/AssignmentFunctors.h
+++ b/Eigen/src/Core/functors/AssignmentFunctors.h
@@ -144,7 +144,7 @@ template<typename Scalar> struct swap_assign_op {
   EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const
   {
-#ifdef EIGEN_CUDACC
+#ifdef EIGEN_GPUCC
     // FIXME is there some kind of cuda::swap?
     Scalar t=b; const_cast<Scalar&>(b)=a; a=t;
 #else
diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h
index e269140bd..401d597d8 100644
--- a/Eigen/src/Core/functors/BinaryFunctors.h
+++ b/Eigen/src/Core/functors/BinaryFunctors.h
@@ -436,10 +436,7 @@ template<typename BinaryOp> struct bind1st_op : BinaryOp {
   typedef typename BinaryOp::second_argument_type second_argument_type;
   typedef typename BinaryOp::result_type          result_type;
 
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC explicit
-  #endif
-  bind1st_op(const first_argument_type &val) : m_value(val) {}
+  EIGEN_DEVICE_FUNC explicit bind1st_op(const first_argument_type &val) : m_value(val) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const second_argument_type& b) const { return BinaryOp::operator()(m_value,b); }
 
@@ -458,10 +455,7 @@ template<typename BinaryOp> struct bind2nd_op : BinaryOp {
   typedef typename BinaryOp::second_argument_type second_argument_type;
   typedef typename BinaryOp::result_type          result_type;
 
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC explicit
-  #endif
-  bind2nd_op(const second_argument_type &val) : m_value(val) {}
+  EIGEN_DEVICE_FUNC explicit bind2nd_op(const second_argument_type &val) : m_value(val) {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const first_argument_type& a) const { return BinaryOp::operator()(a,m_value); }
 
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index a4cde6d95..b1791fb3a 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -163,10 +163,7 @@ class BlasLinearMapper {
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
 
-  #if !defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  EIGEN_ALWAYS_INLINE void prefetch(int i) const {
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
     internal::prefetch(&operator()(i));
   }
 
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 87fcc30f5..059d06874 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -171,7 +171,7 @@ EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
   
     #if defined(EIGEN_HIP_DEVICE_COMPILE)
-    result = aligned_malloc(size);
+    result = ::malloc(size);
     #else
     result = std::malloc(size);
     #endif
@@ -195,7 +195,7 @@ EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
 
     #if defined(EIGEN_HIP_DEVICE_COMPILE)
-    aligned_free(ptr);
+    ::free(ptr);
     #else
     std::free(ptr);
     #endif
@@ -244,7 +244,7 @@ template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std:
   check_that_malloc_is_allowed();
 
   #if defined(EIGEN_HIP_DEVICE_COMPILE)
-  void *result = aligned_malloc(size);
+  void *result = ::malloc(size);
   #else
   void *result = std::malloc(size);
   #endif
@@ -263,7 +263,7 @@ template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void
 template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
 {
   #if defined(EIGEN_HIP_DEVICE_COMPILE)
-  aligned_free(ptr);
+  ::free(ptr);
   #else
   std::free(ptr);
   #endif
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 7f78cc89c..5a358bc12 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -11,16 +11,19 @@
 #ifndef EIGEN_META_H
 #define EIGEN_META_H
 
-#if defined(EIGEN_CUDA_ARCH)
-#include <cfloat>
-#include <math_constants.h>
-#endif
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
-#include <cfloat>
-#include "Eigen/src/Core/arch/HIP/hcc/math_constants.h"
-#endif
+ #include <cfloat>
 
+ #if defined(EIGEN_CUDA_ARCH)
+  #include <math_constants.h>
+ #endif
+
+ #if defined(EIGEN_HIP_DEVICE_COMPILE)
+  #include "Eigen/src/Core/arch/HIP/hcc/math_constants.h"
+  #endif
+
+#endif
 
 #if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
 #include <cstdint>
@@ -181,7 +184,7 @@ template<bool Condition, typename T=void> struct enable_if;
 template<typename T> struct enable_if<true,T>
 { typedef T type; };
 
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 #if !defined(__FLT_EPSILON__)
 #define __FLT_EPSILON__ FLT_EPSILON
 #define __DBL_EPSILON__ DBL_EPSILON
@@ -565,13 +568,13 @@ template<typename T, typename U> struct scalar_product_traits
 
 namespace numext {
   
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
 #else
 template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
 #endif
 
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 using internal::device::numeric_limits;
 #else
 using std::numeric_limits;
@@ -590,7 +593,7 @@ T div_ceil(const T &a, const T &b)
 template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const X& x,const Y& y) { return x == y; }
 
-#if !defined(EIGEN_CUDA_ARCH)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) 
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool equal_strict(const float& x,const float& y) { return std::equal_to<float>()(x,y); }
 
@@ -601,7 +604,7 @@ bool equal_strict(const double& x,const double& y) { return std::equal_to<double
 template<typename X, typename Y> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const X& x,const Y& y) { return x != y; }
 
-#if !defined(EIGEN_CUDA_ARCH)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) 
 template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
 bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to<float>()(x,y); }
 
diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index e977b9623..11df14918 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -1299,7 +1299,7 @@ void BDCSVD<MatrixType>::deflation(Eigen::Index firstCol, Eigen::Index lastCol,
 #endif
 }//end deflation
 
-#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
+#if !defined(EIGEN_GPUCC)
 /** \svd_module
   *
   * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm
diff --git a/test/half_float.cpp b/test/half_float.cpp
index 7734f82cc..1b0ea9482 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -9,7 +9,7 @@
 
 #include "main.h"
 
-#include <Eigen/src/Core/arch/CUDA/Half.h>
+#include <Eigen/src/Core/arch/GPU/Half.h>
 
 // Make sure it's possible to forward declare Eigen::half
 namespace Eigen {
diff --git a/test/main.h b/test/main.h
index 5691af52b..95bbc9eb0 100644
--- a/test/main.h
+++ b/test/main.h
@@ -68,9 +68,19 @@
 // are defined here and any not-parenthesized min/max call will cause a
 // compiler error.
 #if !defined(__HIPCC__)
-  // HIP headers include the <thread> header which contains not-parenthesized
-  // calls to "max", triggering the following check and causing the compile to fail
-  // so disabling the following checks for HIP
+  //
+  // HIP header files include the following files
+  //  <thread>
+  //  <regex>
+  //  <unordered_map>
+  // which seem to contain not-parenthesized calls to "max"/"min", triggering the following check and causing the compile to fail
+  //
+  // Including those header files before the following macro definition for "min" / "max", only partially resolves the issue
+  // This is because other HIP header files also define "isnan" / "isinf" / "isfinite" functions, which are needed in other
+  // headers.
+  //
+  // So instead choosing to simply disable this check for HIP
+  //
   #define min(A,B) please_protect_your_min_with_parentheses
   #define max(A,B) please_protect_your_max_with_parentheses
   #define isnan(X) please_protect_your_isnan_with_parentheses
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
index e94e577fc..5c1c68912 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
@@ -35,7 +35,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE)
     // Running on the host CPU
     return 1;
 #elif defined(EIGEN_HIP_DEVICE_COMPILE)
@@ -48,9 +48,12 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(__SYCL_DEVICE_ONLY__)
     // Running on the host CPU
     return l1CacheSize();
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return 48*1024; // FIXME : update this number for HIP
 #else
     // Running on a CUDA device, return the amount of shared memory available.
     return 48*1024;
@@ -58,9 +61,12 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(__SYCL_DEVICE_ONLY__)
     // Running single threaded on the host CPU
     return l3CacheSize();
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return firstLevelCacheSize(); // FIXME : update this number for HIP
 #else
     // Running on a CUDA device
     return firstLevelCacheSize();
@@ -68,7 +74,7 @@ struct DefaultDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
-#if !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE)
     // Running single threaded on the host CPU
     // Should return an enum that encodes the ISA supported by the CPU
     return 1;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 24a57970a..8bbe449cc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -201,7 +201,7 @@ class TensorExecutor<Expression, GpuDevice, Vectorizable> {
 };
 
 
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+#if defined(EIGEN_GPUCC)
 template <typename Evaluator, typename Index, bool Vectorizable>
 struct EigenMetaKernelEval {
   static __device__ EIGEN_ALWAYS_INLINE
@@ -276,7 +276,7 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
   evaluator.cleanup();
 }
 
-#endif  // EIGEN_CUDACC || EIGEN_HIPCC
+#endif  // EIGEN_GPUCC
 #endif  // EIGEN_USE_GPU
 
 // SYCL Executor policy
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
index 25ba2001e..b6d445c50 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
@@ -35,7 +35,7 @@ namespace {
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   typename internal::enable_if<sizeof(T)==4,int>::type count_leading_zeros(const T val)
   {
-#ifdef EIGEN_CUDA_ARCH
+#ifdef EIGEN_GPU_COMPILE_PHASE
     return __clz(val);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::clz(val);
@@ -53,7 +53,7 @@ namespace {
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   typename internal::enable_if<sizeof(T)==8,int>::type count_leading_zeros(const T val)
   {
-#ifdef EIGEN_CUDA_ARCH
+#ifdef EIGEN_GPU_COMPILE_PHASE
     return __clzll(val);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::clz(val);
@@ -90,7 +90,7 @@ namespace {
 
   template <typename T>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
     return __umulhi(a, b);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
@@ -101,7 +101,7 @@ namespace {
 
   template <typename T>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
-#if defined(EIGEN_CUDA_ARCH)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
     return __umul64hi(a, b);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
@@ -124,7 +124,7 @@ namespace {
   template <typename T>
   struct DividerHelper<64, T> {
     static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
-#if defined(__SIZEOF_INT128__) && !defined(EIGEN_CUDA_ARCH) && !defined(__SYCL_DEVICE_ONLY__)
+#if defined(__SIZEOF_INT128__) && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(__SYCL_DEVICE_ONLY__)
       return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
 #else
       const uint64_t shift = 1ULL << log_div;
@@ -203,7 +203,7 @@ class TensorIntDivisor<int32_t, true> {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
-#ifdef EIGEN_CUDA_ARCH
+#ifdef EIGEN_GPU_COMPILE_PHASE
     return (__umulhi(magic, n) >> shift);
 #elif defined(__SYCL_DEVICE_ONLY__)
     return (cl::sycl::mul_hi(static_cast<uint64_t>(magic), static_cast<uint64_t>(n)) >> shift);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
index 8e1ba486d..c6ca396a3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMacros.h
@@ -27,7 +27,7 @@
  */
 
 // SFINAE requires variadic templates
-#if !defined(EIGEN_CUDACC) && !defined(EIGEN_HIPCC)
+#if !defined(EIGEN_GPUCC)
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   // SFINAE doesn't work for gcc <= 4.7
   #ifdef EIGEN_COMP_GNUC
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index de1075cc1..87be090f9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -52,7 +52,7 @@ struct PacketType : internal::packet_traits<Scalar> {
 };
 
 // For CUDA packet types when using a GpuDevice
-#if defined(EIGEN_USE_GPU) && ((defined(EIGEN_CUDACC) && defined(EIGEN_HAS_CUDA_FP16)) || (defined(EIGEN_HIPCC) && defined(EIGEN_HAS_HIP_FP16)))
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_HAS_GPU_FP16)
 template <>
 struct PacketType<half, GpuDevice> {
   typedef half2 type;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
index 71536a4b9..5a547141a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
@@ -16,7 +16,7 @@ namespace internal {
 namespace {
 
 EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+#if defined(EIGEN_GPU_COMPILE_PHASE)
   // We don't support 3d kernels since we currently only use 1 and
   // 2d kernels.
   assert(threadIdx.z == 0);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index d2fb3fd32..fdd338b96 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -334,12 +334,12 @@ struct OuterReducer {
 };
 
 
-#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
 template <int B, int N, typename S, typename R, typename I>
 __global__ void FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
 
 
-#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
+#if defined(EIGEN_HAS_GPU_FP16)
 template <typename S, typename R, typename I>
 __global__ void ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
 template <int B, int N, typename S, typename R, typename I>
@@ -698,9 +698,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 #ifdef EIGEN_USE_THREADS
   template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
 #endif
-#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
   template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
-#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
+#if defined(EIGEN_HAS_GPU_FP16)
   template <typename S, typename R, typename I> KERNEL_FRIEND void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
   template <int B, int N, typename S, typename R, typename I> KERNEL_FRIEND void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
   template <int NPT, typename S, typename R, typename I> KERNEL_FRIEND void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
@@ -793,7 +793,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   Op m_reducer;
 
   // For full reductions
-#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
   static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
   static const bool RunningOnSycl = false;
 #elif defined(EIGEN_USE_SYCL)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 174a6a064..6d68e256f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -242,7 +242,7 @@ struct ScanLauncher {
   }
 };
 
-#if defined(EIGEN_USE_GPU) && (defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC))
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
 
 // GPU implementation of scan
 // TODO(ibab) This placeholder implementation performs multiple scans in
@@ -286,7 +286,7 @@ struct ScanLauncher<Self, Reducer, GpuDevice> {
 #endif     
   }
 };
-#endif  // EIGEN_USE_GPU && (EIGEN_CUDACC || EIGEN_HIPCC)
+#endif  // EIGEN_USE_GPU && (EIGEN_GPUCC)
 
 }  // end namespace Eigen
 
diff --git a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
index bb584e3f9..8de3bbcab 100644
--- a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
+++ b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
@@ -268,10 +268,7 @@ template<
   typename Reducer
 > struct reduce<Reducer>
 {
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  constexpr static inline int run() { return Reducer::Identity; }
+  EIGEN_DEVICE_FUNC constexpr static inline int run() { return Reducer::Identity; }
 };
 
 template<
@@ -279,10 +276,7 @@ template<
   typename A
 > struct reduce<Reducer, A>
 {
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  constexpr static inline A run(A a) { return a; }
+  EIGEN_DEVICE_FUNC constexpr static inline A run(A a) { return a; }
 };
 
 template<
@@ -291,10 +285,7 @@ template<
   typename... Ts
 > struct reduce<Reducer, A, Ts...>
 {
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  constexpr static inline auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
+  EIGEN_DEVICE_FUNC constexpr static inline auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
     return Reducer::run(a, reduce<Reducer, Ts...>::run(ts...));
   }
 };
@@ -333,10 +324,7 @@ struct greater_equal_zero_op { template<typename A> constexpr static inline auto
 // together in front... (13.0 doesn't work with array_prod/array_reduce/... anyway, but 13.1
 // does...
 template<typename... Ts>
-#if defined(EIGEN_HIPCC)
-EIGEN_DEVICE_FUNC
-#endif
-constexpr inline decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
+EIGEN_DEVICE_FUNC constexpr inline decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
 {
   return reduce<product_op, Ts...>::run(ts...);
 }
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 5b01c5fb7..d91662d96 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -15,7 +15,7 @@
 // The array class is only available starting with cxx11. Emulate our own here
 // if needed. Beware, msvc still doesn't advertise itself as a c++11 compiler!
 // Moreover, CUDA doesn't support the STL containers, so we use our own instead.
-#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC) || defined(EIGEN_AVOID_STL_ARRAY)
+#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_GPUCC) || defined(EIGEN_AVOID_STL_ARRAY)
 
 namespace Eigen {
 template <typename T, size_t n> class array {
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 9a719e3c0..fbbc87661 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -190,7 +190,7 @@ template <>
 struct lgamma_impl<float> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE float run(float x) {
-#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) 
     int dummy;
     return ::lgammaf_r(x, &dummy);
 #else
@@ -203,7 +203,7 @@ template <>
 struct lgamma_impl<double> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE double run(double x) {
-#if !defined(EIGEN_CUDA_ARCH) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) 
     int dummy;
     return ::lgamma_r(x, &dummy);
 #else
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
index 020ac1b62..ad011e305 100644
--- a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
+++ b/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_GPUCC) && defined(EIGEN_USE_GPU)
 
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 plgamma<float4>(const float4& a)
diff --git a/unsupported/test/cxx11_tensor_hip.cu b/unsupported/test/cxx11_tensor_hip.cu
index f1794aa87..b28840267 100644
--- a/unsupported/test/cxx11_tensor_hip.cu
+++ b/unsupported/test/cxx11_tensor_hip.cu
@@ -1222,9 +1222,8 @@ void test_cxx11_tensor_hip()
   CALL_SUBTEST(test_hip_elementwise());
   CALL_SUBTEST(test_hip_props());
   CALL_SUBTEST(test_hip_reduction());
-  // FIXME : uncommenting following tests results in compile failure
-  // CALL_SUBTEST(test_hip_contraction<ColMajor>());
-  // CALL_SUBTEST(test_hip_contraction<RowMajor>());
+  CALL_SUBTEST(test_hip_contraction<ColMajor>());
+  CALL_SUBTEST(test_hip_contraction<RowMajor>());
   CALL_SUBTEST(test_hip_convolution_1d<ColMajor>());
   CALL_SUBTEST(test_hip_convolution_1d<RowMajor>());
   CALL_SUBTEST(test_hip_convolution_inner_dim_col_major_1d());

From b8271bb368d4d2be11f9f493495840481d2e5f2a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 15 Jun 2018 14:21:29 +0200
Subject: [PATCH 397/680] fix md5sum of lapack_addons

---
 lapack/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/lapack/CMakeLists.txt b/lapack/CMakeLists.txt
index 52f18edfc..522ba8a2b 100644
--- a/lapack/CMakeLists.txt
+++ b/lapack/CMakeLists.txt
@@ -49,7 +49,7 @@ if(EIGEN_ENABLE_LAPACK_TESTS)
                   INACTIVITY_TIMEOUT 15
                   TIMEOUT 240
                   STATUS download_status
-                  EXPECTED_MD5 5758ce55afcf79da98de8b9de1615ad5
+                  EXPECTED_MD5 ab5742640617e3221a873aba44bbdc93
                   SHOW_PROGRESS)
                   
     message(STATUS ${download_status})

From a8ab6060df270709b9975102382caa0eb8ef22ec Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 6 Jul 2018 09:58:45 +0200
Subject: [PATCH 398/680] Add unitests for inverse and selfadjoint-eigenvalues
 on CUDA

---
 test/cuda_basic.cu | 39 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 3 deletions(-)

diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu
index ce66c2c78..33e5fd119 100644
--- a/test/cuda_basic.cu
+++ b/test/cuda_basic.cu
@@ -121,7 +121,7 @@ struct diagonal {
 };
 
 template<typename T>
-struct eigenvalues {
+struct eigenvalues_direct {
   EIGEN_DEVICE_FUNC
   void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
   {
@@ -136,6 +136,34 @@ struct eigenvalues {
   }
 };
 
+template<typename T>
+struct eigenvalues {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    typedef Matrix<typename T::Scalar, T::RowsAtCompileTime, 1> Vec;
+    T M(in+i);
+    Map<Vec> res(out+i*Vec::MaxSizeAtCompileTime);
+    T A = M*M.adjoint();
+    SelfAdjointEigenSolver<T> eig;
+    eig.compute(M);
+    res = eig.eigenvalues();
+  }
+};
+
+template<typename T>
+struct matrix_inverse {
+  EIGEN_DEVICE_FUNC
+  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
+  {
+    using namespace Eigen;
+    T M(in+i);
+    Map<T> res(out+i*T::MaxSizeAtCompileTime);
+    res = M.inverse();
+  }
+};
+
 void test_cuda_basic()
 {
   ei_test_init_cuda();
@@ -163,8 +191,13 @@ void test_cuda_basic()
   
   CALL_SUBTEST( run_and_compare_to_cuda(diagonal<Matrix3f,Vector3f>(), nthreads, in, out) );
   CALL_SUBTEST( run_and_compare_to_cuda(diagonal<Matrix4f,Vector4f>(), nthreads, in, out) );
+
+  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix2f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix4f>(), nthreads, in, out) );
   
-  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues<Matrix3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues<Matrix2f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues_direct<Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues_direct<Matrix2f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues<Matrix4f>(), nthreads, in, out) );
 
 }

From f4d623ffa772093169a0ac949c7955bf38bf1bbf Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 6 Jul 2018 17:13:36 +0200
Subject: [PATCH 399/680] Complete Packet8h implementation and test it in
 packetmath unit test

---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 85 +++++++++++++++++++++--
 test/packetmath.cpp                       | 15 ++--
 2 files changed, 90 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index 8a6f209c4..2ee92b4f6 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -493,6 +493,13 @@ template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, con
   return float2half(rf);
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h psub<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = psub(af, bf);
+  return float2half(rf);
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
   Packet16f af = half2float(a);
   Packet16f bf = half2float(b);
@@ -730,10 +737,10 @@ struct packet_traits<Eigen::half> : default_packet_traits {
     AlignedOnScalar = 1,
     size = 8,
     HasHalfPacket = 0,
-    HasAdd    = 0,
-    HasSub    = 0,
-    HasMul    = 0,
-    HasNegate = 0,
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasNegate = 1,
     HasAbs    = 0,
     HasAbs2   = 0,
     HasMin    = 0,
@@ -782,6 +789,17 @@ template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const
   _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x);
 }
 
+template<> EIGEN_STRONG_INLINE Packet8h
+ploaddup<Packet8h>(const Eigen::half*  from) {
+  Packet8h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  result.x = _mm_set_epi16(d, d, c, c, b, b, a, a);
+  return result;
+}
+
 template<> EIGEN_STRONG_INLINE Packet8h
 ploadquad<Packet8h>(const Eigen::half* from) {
   Packet8h result;
@@ -835,6 +853,12 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
 
 template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
 
+template<> EIGEN_STRONG_INLINE Packet8h pnegate(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  Packet8f rf = pnegate(af);
+  return float2half(rf);
+}
+
 template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) {
   Packet8f af = half2float(a);
   Packet8f bf = half2float(b);
@@ -842,6 +866,13 @@ template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const
   return float2half(rf);
 }
 
+template<> EIGEN_STRONG_INLINE Packet8h psub<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = psub(af, bf);
+  return float2half(rf);
+}
+
 template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) {
   Packet8f af = half2float(a);
   Packet8f bf = half2float(b);
@@ -894,6 +925,52 @@ template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h&
   return Eigen::half(reduced);
 }
 
+template<> EIGEN_STRONG_INLINE Packet8h preduxp<Packet8h>(const Packet8h* p) {
+  Packet8f pf[8];
+  pf[0] = half2float(p[0]);
+  pf[1] = half2float(p[1]);
+  pf[2] = half2float(p[2]);
+  pf[3] = half2float(p[3]);
+  pf[4] = half2float(p[4]);
+  pf[5] = half2float(p[5]);
+  pf[6] = half2float(p[6]);
+  pf[7] = half2float(p[7]);
+  Packet8f reduced = preduxp<Packet8f>(pf);
+  return float2half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h preverse(const Packet8h& a)
+{
+  __m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);
+  Packet8h res;
+  res.x = _mm_shuffle_epi8(a.x,m);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pinsertfirst(const Packet8h& a, Eigen::half b)
+{
+  Packet8h res;
+  res.x = _mm_insert_epi16(a.x,int(b.x),0);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pinsertlast(const Packet8h& a, Eigen::half b)
+{
+  Packet8h res;
+  res.x = _mm_insert_epi16(a.x,int(b.x),15);
+  return res;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet8h>
+{
+  static EIGEN_STRONG_INLINE void run(Packet8h& first, const Packet8h& second)
+  {
+    if (Offset!=0)
+      first.x = _mm_alignr_epi8(second.x,first.x, Offset*2);
+  }
+};
+
 EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<Packet8h,8>& kernel) {
   __m128i a = kernel.packet[0].x;
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 4b19be92d..94ee91e3b 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -123,7 +123,7 @@ template<typename Scalar> void packetmath()
   EIGEN_ALIGN_MAX Scalar data2[size];
   EIGEN_ALIGN_MAX Packet packets[PacketSize*2];
   EIGEN_ALIGN_MAX Scalar ref[size];
-  RealScalar refvalue = 0;
+  RealScalar refvalue = RealScalar(0);
   for (int i=0; i<size; ++i)
   {
     data1[i] = internal::random<Scalar>()/RealScalar(PacketSize);
@@ -178,7 +178,8 @@ template<typename Scalar> void packetmath()
   VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasSub);
   VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasMul);
   VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasNegate);
-  VERIFY((internal::is_same<Scalar,int>::value) || (!PacketTraits::Vectorizable) || PacketTraits::HasDiv);
+  // Disabled as it is not clear why it would be mandatory to support division.
+  //VERIFY((internal::is_same<Scalar,int>::value) || (!PacketTraits::Vectorizable) || PacketTraits::HasDiv);
 
   CHECK_CWISE2_IF(PacketTraits::HasAdd, REF_ADD,  internal::padd);
   CHECK_CWISE2_IF(PacketTraits::HasSub, REF_SUB,  internal::psub);
@@ -242,29 +243,30 @@ template<typename Scalar> void packetmath()
     }
   }
 
-  ref[0] = 0;
+  ref[0] = Scalar(0);
   for (int i=0; i<PacketSize; ++i)
     ref[0] += data1[i];
   VERIFY(isApproxAbs(ref[0], internal::predux(internal::pload<Packet>(data1)), refvalue) && "internal::predux");
 
+  if(PacketSize==8 && internal::unpacket_traits<typename internal::unpacket_traits<Packet>::half>::size ==4) // so far, predux_half_dowto4 is only required in such a case
   {
     int HalfPacketSize = PacketSize>4 ? PacketSize/2 : PacketSize;
     for (int i=0; i<HalfPacketSize; ++i)
-      ref[i] = 0;
+      ref[i] = Scalar(0);
     for (int i=0; i<PacketSize; ++i)
       ref[i%HalfPacketSize] += data1[i];
     internal::pstore(data2, internal::predux_half_dowto4(internal::pload<Packet>(data1)));
     VERIFY(areApprox(ref, data2, HalfPacketSize) && "internal::predux_half_dowto4");
   }
 
-  ref[0] = 1;
+  ref[0] = Scalar(1);
   for (int i=0; i<PacketSize; ++i)
     ref[0] *= data1[i];
   VERIFY(internal::isApprox(ref[0], internal::predux_mul(internal::pload<Packet>(data1))) && "internal::predux_mul");
 
   for (int j=0; j<PacketSize; ++j)
   {
-    ref[j] = 0;
+    ref[j] = Scalar(0);
     for (int i=0; i<PacketSize; ++i)
       ref[j] += data1[i+j*PacketSize];
     packets[j] = internal::pload<Packet>(data1+j*PacketSize);
@@ -630,6 +632,7 @@ void test_packetmath()
     CALL_SUBTEST_3( packetmath<int>() );
     CALL_SUBTEST_4( packetmath<std::complex<float> >() );
     CALL_SUBTEST_5( packetmath<std::complex<double> >() );
+    CALL_SUBTEST_6( packetmath<half>() );
 
     CALL_SUBTEST_1( packetmath_notcomplex<float>() );
     CALL_SUBTEST_2( packetmath_notcomplex<double>() );

From 56a33ae57dd536f35be007eb051a636cea4467d0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 6 Jul 2018 17:14:04 +0200
Subject: [PATCH 400/680] test product kernel with half-floats.

---
 test/half_float.cpp | 30 ++++++++++++++++++++++++------
 1 file changed, 24 insertions(+), 6 deletions(-)

diff --git a/test/half_float.cpp b/test/half_float.cpp
index 7734f82cc..487d0d1b2 100644
--- a/test/half_float.cpp
+++ b/test/half_float.cpp
@@ -257,13 +257,31 @@ void test_array()
   ss << a1;
 }
 
+void test_product()
+{
+  typedef Matrix<half,Dynamic,Dynamic> MatrixXh;
+  Index rows  = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
+  Index cols  = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
+  Index depth = internal::random<Index>(1,EIGEN_TEST_MAX_SIZE);
+  MatrixXh Ah = MatrixXh::Random(rows,depth);
+  MatrixXh Bh = MatrixXh::Random(depth,cols);
+  MatrixXh Ch = MatrixXh::Random(rows,cols);
+  MatrixXf Af = Ah.cast<float>();
+  MatrixXf Bf = Bh.cast<float>();
+  MatrixXf Cf = Ch.cast<float>();
+  VERIFY_IS_APPROX(Ch.noalias()+=Ah*Bh, (Cf.noalias()+=Af*Bf).cast<half>());
+}
+
 void test_half_float()
 {
-  CALL_SUBTEST(test_conversion());
   CALL_SUBTEST(test_numtraits());
-  CALL_SUBTEST(test_arithmetic());
-  CALL_SUBTEST(test_comparison());
-  CALL_SUBTEST(test_basic_functions());
-  CALL_SUBTEST(test_trigonometric_functions());
-  CALL_SUBTEST(test_array());
+  for(int i = 0; i < g_repeat; i++) {
+    CALL_SUBTEST(test_conversion());
+    CALL_SUBTEST(test_arithmetic());
+    CALL_SUBTEST(test_comparison());
+    CALL_SUBTEST(test_basic_functions());
+    CALL_SUBTEST(test_trigonometric_functions());
+    CALL_SUBTEST(test_array());
+    CALL_SUBTEST(test_product());
+  }
 }

From a937c5020889f95aadccefc79338ed83c2ff8784 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 6 Jul 2018 17:41:52 +0200
Subject: [PATCH 401/680] palign is not used anymore, so let's relax the unit
 test

---
 test/packetmath.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 94ee91e3b..56e017383 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -171,7 +171,10 @@ template<typename Scalar> void packetmath()
     for (int i=0; i<PacketSize; ++i)
       ref[i] = data1[i+offset];
 
+    // palign is not used anymore, so let's just put a warning if it fails
+    ++g_test_level;
     VERIFY(areApprox(ref, data2, PacketSize) && "internal::palign");
+    --g_test_level;
   }
 
   VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasAdd);

From f2dc048df9dddda48cf8945f3a0a61325ffa6c97 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 6 Jul 2018 17:43:11 +0200
Subject: [PATCH 402/680] complete implementation of Packet16h (AVX512)

---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 55 ++++++++++++++++++++++-
 1 file changed, 54 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index 2ee92b4f6..c068351ce 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -486,6 +486,13 @@ EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h pnegate(const Packet16h& a) {
+  // FIXME we could do that with bit manipulation
+  Packet16f af = half2float(a);
+  Packet16f rf = pnegate(af);
+  return float2half(rf);
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
   Packet16f af = half2float(a);
   Packet16f bf = half2float(b);
@@ -512,6 +519,51 @@ template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
   return half(predux(from_float));
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h preduxp<Packet16h>(const Packet16h* p) {
+  Packet16f pf[16];
+  pf[0] = half2float(p[0]);
+  pf[1] = half2float(p[1]);
+  pf[2] = half2float(p[2]);
+  pf[3] = half2float(p[3]);
+  pf[4] = half2float(p[4]);
+  pf[5] = half2float(p[5]);
+  pf[6] = half2float(p[6]);
+  pf[7] = half2float(p[7]);
+  pf[8] = half2float(p[8]);
+  pf[9] = half2float(p[9]);
+  pf[10] = half2float(p[10]);
+  pf[11] = half2float(p[11]);
+  pf[12] = half2float(p[12]);
+  pf[13] = half2float(p[13]);
+  pf[14] = half2float(p[14]);
+  pf[15] = half2float(p[15]);
+  Packet16f reduced = preduxp<Packet16f>(pf);
+  return float2half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h preverse(const Packet16h& a)
+{
+  __m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);
+  Packet16h res;
+  res.x = _mm256_set_m128i(_mm_shuffle_epi8(_mm256_extractf128_si256(a.x,0),m),
+                           _mm_shuffle_epi8(_mm256_extractf128_si256(a.x,1),m));
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pinsertfirst(const Packet16h& a, Eigen::half b)
+{
+  Packet16h res;
+  res.x = _mm256_insert_epi16(a.x,b.x,0);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pinsertlast(const Packet16h& a, Eigen::half b)
+{
+  Packet16h res;
+  res.x = _mm256_insert_epi16(a.x,b.x,15);
+  return res;
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
 {
   Packet16h result;
@@ -854,6 +906,7 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
 template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
 
 template<> EIGEN_STRONG_INLINE Packet8h pnegate(const Packet8h& a) {
+  // FIXME we could do that with bit manipulation
   Packet8f af = half2float(a);
   Packet8f rf = pnegate(af);
   return float2half(rf);
@@ -957,7 +1010,7 @@ template<> EIGEN_STRONG_INLINE Packet8h pinsertfirst(const Packet8h& a, Eigen::h
 template<> EIGEN_STRONG_INLINE Packet8h pinsertlast(const Packet8h& a, Eigen::half b)
 {
   Packet8h res;
-  res.x = _mm_insert_epi16(a.x,int(b.x),15);
+  res.x = _mm_insert_epi16(a.x,int(b.x),7);
   return res;
 }
 

From 1f54164eca5a8960205b94dd3f3cec87089e4ac6 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 7 Jul 2018 00:15:07 +0200
Subject: [PATCH 403/680] Fix a few issues with Packet16h

---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 33 ++++++++++++++++++-----
 1 file changed, 27 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index c068351ce..ae9193b0d 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -362,10 +362,10 @@ struct packet_traits<half> : default_packet_traits {
     AlignedOnScalar = 1,
     size = 16,
     HasHalfPacket = 0,
-    HasAdd    = 0,
-    HasSub    = 0,
-    HasMul    = 0,
-    HasNegate = 0,
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasNegate = 1,
     HasAbs    = 0,
     HasAbs2   = 0,
     HasMin    = 0,
@@ -414,6 +414,21 @@ template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet1
   _mm256_storeu_si256((__m256i*)to, from.x);
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h
+ploaddup<Packet16h>(const Eigen::half*  from) {
+  Packet16h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  unsigned short e = from[4].x;
+  unsigned short f = from[5].x;
+  unsigned short g = from[6].x;
+  unsigned short h = from[7].x;
+  result.x = _mm256_set_epi16(h, h, g, g, f, f, e, e, d, d, c, c, b, b, a, a);
+  return result;
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h
 ploadquad(const Eigen::half* from) {
   Packet16h result;
@@ -519,6 +534,11 @@ template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
   return half(predux(from_float));
 }
 
+template<> EIGEN_STRONG_INLINE half predux_mul<Packet16h>(const Packet16h& from) {
+  Packet16f from_float = half2float(from);
+  return half(predux_mul(from_float));
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h preduxp<Packet16h>(const Packet16h* p) {
   Packet16f pf[16];
   pf[0] = half2float(p[0]);
@@ -545,8 +565,9 @@ template<> EIGEN_STRONG_INLINE Packet16h preverse(const Packet16h& a)
 {
   __m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);
   Packet16h res;
-  res.x = _mm256_set_m128i(_mm_shuffle_epi8(_mm256_extractf128_si256(a.x,0),m),
-                           _mm_shuffle_epi8(_mm256_extractf128_si256(a.x,1),m));
+  res.x = _mm256_insertf128_si256(
+                    _mm256_castsi128_si256(_mm_shuffle_epi8(_mm256_extractf128_si256(a.x,1),m)),
+                                           _mm_shuffle_epi8(_mm256_extractf128_si256(a.x,0),m), 1);
   return res;
 }
 

From 90a53ca6fd8a79d39f926430ec07f626c8260a20 Mon Sep 17 00:00:00 2001
From: Mark D Ryan <mark.d.ryan@intel.com>
Date: Sat, 16 Jun 2018 15:13:06 -0700
Subject: [PATCH 404/680] Fix the Packet16h version of ptranspose

The AVX512 version of ptranpose for PacketBlock<Packet16h,16> was
reordering the PacketBlock argument incorrectly.  This lead to errors in
the multiplication of matrices composed of 16 bit floats on AVX512
machines, if at least of the matrices was using RowMajor order.  This
error is responsible for one tensorflow unit test failure on AVX512
machines:

//tensorflow/python/kernel_tests:batch_matmul_op_test
---
 Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 28 +++++++++++------------
 1 file changed, 14 insertions(+), 14 deletions(-)

diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
index ae9193b0d..9897bd4e5 100644
--- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -692,20 +692,20 @@ ptranspose(PacketBlock<Packet16h,16>& kernel) {
 
   // NOTE: no unpacklo/hi instr in this case, so using permute instr.
   __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
-  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
-  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
-  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
-  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
-  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
-  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
-  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
-  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
-  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
-  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
-  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
-  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
-  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
-  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
+  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
+  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
+  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
+  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
+  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
+  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
+  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
+  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
+  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
+  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
+  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
+  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
   __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
 
   kernel.packet[0].x = a_p_0;

From 7e41c8f1a98c2a3beed667dca416ea8d20ad373e Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 20 Jun 2018 12:52:30 -0400
Subject: [PATCH 405/680] renaming *Cuda files to *Gpu in the
 unsupported/Eigen/CXX11/src/Tensor and unsupported/test directories

---
 .../Tensor/{TensorContractionCuda.h => TensorContractionGpu.h}    | 0
 .../CXX11/src/Tensor/{TensorDeviceCuda.h => TensorDeviceGpu.h}    | 0
 .../src/Tensor/{TensorReductionCuda.h => TensorReductionGpu.h}    | 0
 .../{cxx11_tensor_argmax_cuda.cu => cxx11_tensor_argmax_gpu.cu}   | 0
 ...nsor_cast_float16_cuda.cu => cxx11_tensor_cast_float16_gpu.cu} | 0
 ...ex_cwise_ops_cuda.cu => cxx11_tensor_complex_cwise_ops_gpu.cu} | 0
 .../{cxx11_tensor_complex_cuda.cu => cxx11_tensor_complex_gpu.cu} | 0
 ...cxx11_tensor_contract_cuda.cu => cxx11_tensor_contract_gpu.cu} | 0
 unsupported/test/{cxx11_tensor_cuda.cu => cxx11_tensor_gpu.cu}    | 0
 ...1_tensor_of_float16_cuda.cu => cxx11_tensor_of_float16_gpu.cu} | 0
 .../{cxx11_tensor_random_cuda.cu => cxx11_tensor_random_gpu.cu}   | 0
 ...x11_tensor_reduction_cuda.cu => cxx11_tensor_reduction_gpu.cu} | 0
 .../test/{cxx11_tensor_scan_cuda.cu => cxx11_tensor_scan_gpu.cu}  | 0
 13 files changed, 0 insertions(+), 0 deletions(-)
 rename unsupported/Eigen/CXX11/src/Tensor/{TensorContractionCuda.h => TensorContractionGpu.h} (100%)
 rename unsupported/Eigen/CXX11/src/Tensor/{TensorDeviceCuda.h => TensorDeviceGpu.h} (100%)
 rename unsupported/Eigen/CXX11/src/Tensor/{TensorReductionCuda.h => TensorReductionGpu.h} (100%)
 rename unsupported/test/{cxx11_tensor_argmax_cuda.cu => cxx11_tensor_argmax_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_cast_float16_cuda.cu => cxx11_tensor_cast_float16_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_complex_cwise_ops_cuda.cu => cxx11_tensor_complex_cwise_ops_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_complex_cuda.cu => cxx11_tensor_complex_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_contract_cuda.cu => cxx11_tensor_contract_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_cuda.cu => cxx11_tensor_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_of_float16_cuda.cu => cxx11_tensor_of_float16_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_random_cuda.cu => cxx11_tensor_random_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_reduction_cuda.cu => cxx11_tensor_reduction_gpu.cu} (100%)
 rename unsupported/test/{cxx11_tensor_scan_cuda.cu => cxx11_tensor_scan_gpu.cu} (100%)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
similarity index 100%
rename from unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
rename to unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
similarity index 100%
rename from unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
rename to unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
similarity index 100%
rename from unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
rename to unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_argmax_cuda.cu
rename to unsupported/test/cxx11_tensor_argmax_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_cast_float16_cuda.cu
rename to unsupported/test/cxx11_tensor_cast_float16_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_complex_cwise_ops_cuda.cu
rename to unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_complex_cuda.cu
rename to unsupported/test/cxx11_tensor_complex_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_contract_cuda.cu
rename to unsupported/test/cxx11_tensor_contract_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_cuda.cu
rename to unsupported/test/cxx11_tensor_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_of_float16_cuda.cu
rename to unsupported/test/cxx11_tensor_of_float16_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_random_cuda.cu
rename to unsupported/test/cxx11_tensor_random_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_reduction_cuda.cu
rename to unsupported/test/cxx11_tensor_reduction_gpu.cu
diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_gpu.cu
similarity index 100%
rename from unsupported/test/cxx11_tensor_scan_cuda.cu
rename to unsupported/test/cxx11_tensor_scan_gpu.cu

From cfdabbcc8f708c06da2bfa4e924edc25619f013a Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 20 Jun 2018 12:57:02 -0400
Subject: [PATCH 406/680] removing the *Hip files from the
 unsupported/Eigen/CXX11/src/Tensor and unsupported/test directories

---
 .../CXX11/src/Tensor/TensorContractionHip.h   | 1521 -----------------
 .../CXX11/src/Tensor/TensorConvolutionHip.h   | 1119 ------------
 .../Eigen/CXX11/src/Tensor/TensorDeviceHip.h  |  352 ----
 .../CXX11/src/Tensor/TensorReductionHip.h     |  815 ---------
 unsupported/test/cxx11_tensor_argmax_hip.cu   |  251 ---
 .../test/cxx11_tensor_cast_float16_hip.cu     |   79 -
 unsupported/test/cxx11_tensor_contract_hip.cu |  215 ---
 unsupported/test/cxx11_tensor_device_hip.cu   |  389 -----
 unsupported/test/cxx11_tensor_hip.cu          | 1295 --------------
 .../test/cxx11_tensor_of_float16_hip.cu       |  498 ------
 unsupported/test/cxx11_tensor_random_hip.cu   |   85 -
 .../test/cxx11_tensor_reduction_hip.cu        |  154 --
 unsupported/test/cxx11_tensor_scan_hip.cu     |   76 -
 13 files changed, 6849 deletions(-)
 delete mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
 delete mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
 delete mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
 delete mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
 delete mode 100644 unsupported/test/cxx11_tensor_argmax_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_cast_float16_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_contract_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_device_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_of_float16_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_random_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_reduction_hip.cu
 delete mode 100644 unsupported/test/cxx11_tensor_scan_hip.cu

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
deleted file mode 100644
index 7561846a3..000000000
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionHip.h
+++ /dev/null
@@ -1,1521 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014-2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
-// Copyright (C) 2015 Navdeep Jaitly <ndjaitly@google.com>
-// Copyright (C) 2014 Eric Martin <eric@ericmart.in>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
-#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
-
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
-
-namespace Eigen {
-
-template<typename Scalar, typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper, bool needs_edge_check>
-__device__ EIGEN_STRONG_INLINE void
-EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
-                               const OutputMapper output, Scalar* lhs_shmem, Scalar* rhs_shmem,
-                       const Index m_size, const Index n_size, const Index k_size) {
-
-  const Index m_block_idx = hipBlockIdx_x;
-  const Index n_block_idx = hipBlockIdx_y;
-
-  const Index base_m = 64 * m_block_idx;
-  const Index base_n = 64 * n_block_idx;
-
-  // declare and initialize 64 registers for output 8x8 block
-
-  // prefetch registers
-  Scalar lhs_pf0;
-  Scalar lhs_pf1;
-  Scalar lhs_pf2;
-  Scalar lhs_pf3;
-  Scalar lhs_pf4;
-  Scalar lhs_pf5;
-  Scalar lhs_pf6;
-  Scalar lhs_pf7;
-
-  Scalar rhs_pf0;
-  Scalar rhs_pf1;
-  Scalar rhs_pf2;
-  Scalar rhs_pf3;
-  Scalar rhs_pf4;
-  Scalar rhs_pf5;
-  Scalar rhs_pf6;
-  Scalar rhs_pf7;
-
-  // shared memory is formatted
-  // (contract idx in block, nocontract idx in block, block idx)
-  // where block idx is column major. This transposition limits the number of
-  // bank conflicts when reading the LHS. The core idea is that since the contracting
-  // index is shared by both sides, then the contracting index should be in hipThreadIdx_x.
-
-  // On the LHS, we pad each row inside of each block with an extra element. This makes
-  // each block 8 rows of 9 elements, which is 72 elements. This gives no bank conflicts
-  // on writes and very few 2-way conflicts on reads. There is an 8x8 grid of these blocks.
-
-  // On the RHS we just add 8 padding elements to the end of each block. This gives no bank
-  // conflicts on writes and also none on reads.
-
-  // storage indices
-  const Index lhs_store_idx_base = hipThreadIdx_y * 72 + hipThreadIdx_x * 9 + hipThreadIdx_z;
-  const Index rhs_store_idx_base = hipThreadIdx_y * 72 + hipThreadIdx_z * 8 + hipThreadIdx_x;
-
-  const Index lhs_store_idx_0 = lhs_store_idx_base + 576 * 0;
-  const Index lhs_store_idx_1 = lhs_store_idx_base + 576 * 1;
-  const Index lhs_store_idx_2 = lhs_store_idx_base + 576 * 2;
-  const Index lhs_store_idx_3 = lhs_store_idx_base + 576 * 3;
-  const Index lhs_store_idx_4 = lhs_store_idx_base + 576 * 4;
-  const Index lhs_store_idx_5 = lhs_store_idx_base + 576 * 5;
-  const Index lhs_store_idx_6 = lhs_store_idx_base + 576 * 6;
-  const Index lhs_store_idx_7 = lhs_store_idx_base + 576 * 7;
-
-  const Index rhs_store_idx_0 = rhs_store_idx_base + 576 * 0;
-  const Index rhs_store_idx_1 = rhs_store_idx_base + 576 * 1;
-  const Index rhs_store_idx_2 = rhs_store_idx_base + 576 * 2;
-  const Index rhs_store_idx_3 = rhs_store_idx_base + 576 * 3;
-  const Index rhs_store_idx_4 = rhs_store_idx_base + 576 * 4;
-  const Index rhs_store_idx_5 = rhs_store_idx_base + 576 * 5;
-  const Index rhs_store_idx_6 = rhs_store_idx_base + 576 * 6;
-  const Index rhs_store_idx_7 = rhs_store_idx_base + 576 * 7;
-
-  // in the loading code, the following variables are important:
-  // hipThreadIdx_x: the vertical position in an 8x8 block
-  // hipThreadIdx_y: the vertical index of the 8x8 block in the grid
-  // hipThreadIdx_z: the horizontal position in an 8x8 block
-  // k: the horizontal index of the 8x8 block in the grid
-  //
-  // The k parameter is implicit (it was the loop counter for a loop that went
-  // from 0 to <8, but now that loop is unrolled in the below code.
-
-  const Index load_idx_vert = hipThreadIdx_x + 8 * hipThreadIdx_y;
-  const Index lhs_vert = base_m + load_idx_vert;
-
-#define prefetchIntoRegisters(base_k)                           \
-  {                                                             \
-    lhs_pf0 = conv(0);                                          \
-    lhs_pf1 = conv(0);                                          \
-    lhs_pf2 = conv(0);                                          \
-    lhs_pf3 = conv(0);                                          \
-    lhs_pf4 = conv(0);                                          \
-    lhs_pf5 = conv(0);                                          \
-    lhs_pf6 = conv(0);                                          \
-    lhs_pf7 = conv(0);                                          \
-                                                                \
-    rhs_pf0 = conv(0);                                          \
-    rhs_pf1 = conv(0);                                          \
-    rhs_pf2 = conv(0);                                          \
-    rhs_pf3 = conv(0);                                          \
-    rhs_pf4 = conv(0);                                          \
-    rhs_pf5 = conv(0);                                          \
-    rhs_pf6 = conv(0);                                          \
-    rhs_pf7 = conv(0);                                          \
-                                                                \
-    if (!needs_edge_check || lhs_vert < m_size) {               \
-      const Index lhs_horiz_0 = base_k + hipThreadIdx_z + 0 * 8;   \
-      const Index lhs_horiz_1 = base_k + hipThreadIdx_z + 1 * 8;   \
-      const Index lhs_horiz_2 = base_k + hipThreadIdx_z + 2 * 8;   \
-      const Index lhs_horiz_3 = base_k + hipThreadIdx_z + 3 * 8;   \
-      const Index lhs_horiz_4 = base_k + hipThreadIdx_z + 4 * 8;   \
-      const Index lhs_horiz_5 = base_k + hipThreadIdx_z + 5 * 8;   \
-      const Index lhs_horiz_6 = base_k + hipThreadIdx_z + 6 * 8;   \
-      const Index lhs_horiz_7 = base_k + hipThreadIdx_z + 7 * 8;   \
-                                                                \
-      if (!needs_edge_check || lhs_horiz_7 < k_size) {          \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
-        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
-        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
-        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
-        lhs_pf7 = lhs(lhs_vert, lhs_horiz_7);                   \
-      } else if (lhs_horiz_6 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
-        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
-        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
-        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
-      } else if (lhs_horiz_5 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
-        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
-        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
-      } else if (lhs_horiz_4 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
-        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
-      } else if (lhs_horiz_3 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
-      } else if (lhs_horiz_2 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
-      } else if (lhs_horiz_1 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
-      } else if (lhs_horiz_0 < k_size) {                        \
-        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
-      }                                                         \
-    }                                                           \
-                                                                \
-    const Index rhs_vert = base_k + load_idx_vert;              \
-    if (!needs_edge_check || rhs_vert < k_size) {               \
-      const Index rhs_horiz_0 = base_n + hipThreadIdx_z + 0 * 8;   \
-      const Index rhs_horiz_1 = base_n + hipThreadIdx_z + 1 * 8;   \
-      const Index rhs_horiz_2 = base_n + hipThreadIdx_z + 2 * 8;   \
-      const Index rhs_horiz_3 = base_n + hipThreadIdx_z + 3 * 8;   \
-      const Index rhs_horiz_4 = base_n + hipThreadIdx_z + 4 * 8;   \
-      const Index rhs_horiz_5 = base_n + hipThreadIdx_z + 5 * 8;   \
-      const Index rhs_horiz_6 = base_n + hipThreadIdx_z + 6 * 8;   \
-      const Index rhs_horiz_7 = base_n + hipThreadIdx_z + 7 * 8;   \
-                                                                \
-      if (rhs_horiz_7 < n_size) {                               \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
-        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
-        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
-        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
-        rhs_pf7 = rhs(rhs_vert, rhs_horiz_7);                   \
-      } else if (rhs_horiz_6 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
-        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
-        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
-        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
-      } else if (rhs_horiz_5 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
-        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
-        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
-      } else if (rhs_horiz_4 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
-        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
-      } else if (rhs_horiz_3 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
-      } else if (rhs_horiz_2 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
-      } else if (rhs_horiz_1 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
-      } else if (rhs_horiz_0 < n_size) {                        \
-        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
-      }                                                         \
-    }                                                           \
-  }                                                             \
-
-#define writeRegToShmem(_)                      \
-  lhs_shmem[lhs_store_idx_0] = lhs_pf0;         \
-  rhs_shmem[rhs_store_idx_0] = rhs_pf0;         \
-                                                \
-  lhs_shmem[lhs_store_idx_1] = lhs_pf1;         \
-  rhs_shmem[rhs_store_idx_1] = rhs_pf1;         \
-                                                \
-  lhs_shmem[lhs_store_idx_2] = lhs_pf2;         \
-  rhs_shmem[rhs_store_idx_2] = rhs_pf2;         \
-                                                \
-  lhs_shmem[lhs_store_idx_3] = lhs_pf3;         \
-  rhs_shmem[rhs_store_idx_3] = rhs_pf3;         \
-                                                \
-  lhs_shmem[lhs_store_idx_4] = lhs_pf4;         \
-  rhs_shmem[rhs_store_idx_4] = rhs_pf4;         \
-                                                \
-  lhs_shmem[lhs_store_idx_5] = lhs_pf5;         \
-  rhs_shmem[rhs_store_idx_5] = rhs_pf5;         \
-                                                \
-  lhs_shmem[lhs_store_idx_6] = lhs_pf6;         \
-  rhs_shmem[rhs_store_idx_6] = rhs_pf6;         \
-                                                \
-  lhs_shmem[lhs_store_idx_7] = lhs_pf7;         \
-  rhs_shmem[rhs_store_idx_7] = rhs_pf7;         \
-
-  // declare and initialize result array
-#define res(i, j) _res_##i##j
-#define initResultRow(i)                        \
-  Scalar res(i, 0) = conv(0);                   \
-  Scalar res(i, 1) = conv(0);                   \
-  Scalar res(i, 2) = conv(0);                   \
-  Scalar res(i, 3) = conv(0);                   \
-  Scalar res(i, 4) = conv(0);                   \
-  Scalar res(i, 5) = conv(0);                   \
-  Scalar res(i, 6) = conv(0);                   \
-  Scalar res(i, 7) = conv(0);                   \
-
-  internal::scalar_cast_op<int, Scalar> conv;
-  initResultRow(0);
-  initResultRow(1);
-  initResultRow(2);
-  initResultRow(3);
-  initResultRow(4);
-  initResultRow(5);
-  initResultRow(6);
-  initResultRow(7);
-#undef initResultRow
-
-  for (Index base_k = 0; base_k < k_size; base_k += 64) {
-    // wait for previous iteration to finish with shmem. Despite common sense,
-    // the code is a bit faster with this here then at bottom of loop
-    __syncthreads();
-
-    prefetchIntoRegisters(base_k);
-    writeRegToShmem();
-
-    #undef prefetchIntoRegisters
-    #undef writeRegToShmem
-
-    // wait for shared mem packing to be done before starting computation
-    __syncthreads();
-
-    // compute 8x8 matrix product by outer product. This involves packing one column
-    // of LHS and one row of RHS into registers (takes 16 registers).
-
-#define lcol(i) _lcol##i
-    Scalar lcol(0);
-    Scalar lcol(1);
-    Scalar lcol(2);
-    Scalar lcol(3);
-    Scalar lcol(4);
-    Scalar lcol(5);
-    Scalar lcol(6);
-    Scalar lcol(7);
-
-#define rrow(j) _rrow##j
-    Scalar rrow(0);
-    Scalar rrow(1);
-    Scalar rrow(2);
-    Scalar rrow(3);
-    Scalar rrow(4);
-    Scalar rrow(5);
-    Scalar rrow(6);
-    Scalar rrow(7);
-
-    // Now x corresponds to k, y to m, and z to n
-    const Scalar* lhs_block = &lhs_shmem[hipThreadIdx_x + 9 * hipThreadIdx_y];
-    const Scalar* rhs_block = &rhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_z];
-
-#define lhs_element(i, j) lhs_block[72 * ((i) + 8 * (j))]
-#define rhs_element(i, j) rhs_block[72 * ((i) + 8 * (j))]
-
-#define loadData(i, j)                          \
-    lcol(0) = lhs_element(0, j);               \
-    rrow(0) = rhs_element(i, 0);               \
-    lcol(1) = lhs_element(1, j);               \
-    rrow(1) = rhs_element(i, 1);               \
-    lcol(2) = lhs_element(2, j);               \
-    rrow(2) = rhs_element(i, 2);               \
-    lcol(3) = lhs_element(3, j);               \
-    rrow(3) = rhs_element(i, 3);               \
-    lcol(4) = lhs_element(4, j);               \
-    rrow(4) = rhs_element(i, 4);               \
-    lcol(5) = lhs_element(5, j);               \
-    rrow(5) = rhs_element(i, 5);               \
-    lcol(6) = lhs_element(6, j);               \
-    rrow(6) = rhs_element(i, 6);               \
-    lcol(7) = lhs_element(7, j);               \
-    rrow(7) = rhs_element(i, 7);               \
-
-#define computeCol(j)                           \
-    res(0, j) += lcol(0) * rrow(j);             \
-    res(1, j) += lcol(1) * rrow(j);             \
-    res(2, j) += lcol(2) * rrow(j);             \
-    res(3, j) += lcol(3) * rrow(j);             \
-    res(4, j) += lcol(4) * rrow(j);             \
-    res(5, j) += lcol(5) * rrow(j);             \
-    res(6, j) += lcol(6) * rrow(j);             \
-    res(7, j) += lcol(7) * rrow(j);             \
-
-#define computePass(i)                          \
-    loadData(i, i);                             \
-                                                \
-    computeCol(0);                              \
-    computeCol(1);                              \
-    computeCol(2);                              \
-    computeCol(3);                              \
-    computeCol(4);                              \
-    computeCol(5);                              \
-    computeCol(6);                              \
-    computeCol(7);                              \
-
-    computePass(0);
-    computePass(1);
-    computePass(2);
-    computePass(3);
-    computePass(4);
-    computePass(5);
-    computePass(6);
-    computePass(7);
-
-#undef lcol
-#undef rrow
-#undef lhs_element
-#undef rhs_element
-#undef loadData
-#undef computeCol
-#undef computePass
-  } // end loop over k
-
-  // we've now iterated over all of the large (ie width 64) k blocks and
-  // accumulated results in registers. At this point thread (x, y, z) contains
-  // the sum across all big k blocks of the product of little k block of index (x, y)
-  // with block of index (y, z). To compute the final output, we need to reduce
-  // the 8 threads over y by summation.
-#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
-
-#define reduceRow(i, mask)                      \
-  shuffleInc(i, 0, mask);                       \
-  shuffleInc(i, 1, mask);                       \
-  shuffleInc(i, 2, mask);                       \
-  shuffleInc(i, 3, mask);                       \
-  shuffleInc(i, 4, mask);                       \
-  shuffleInc(i, 5, mask);                       \
-  shuffleInc(i, 6, mask);                       \
-  shuffleInc(i, 7, mask);                       \
-
-#define reduceMatrix(mask)                      \
-  reduceRow(0, mask);                           \
-  reduceRow(1, mask);                           \
-  reduceRow(2, mask);                           \
-  reduceRow(3, mask);                           \
-  reduceRow(4, mask);                           \
-  reduceRow(5, mask);                           \
-  reduceRow(6, mask);                           \
-  reduceRow(7, mask);                           \
-
-  // actually perform the reduction, now each thread of index (_, y, z)
-  // contains the correct values in its registers that belong in the output
-  // block
-  reduceMatrix(1);
-  reduceMatrix(2);
-  reduceMatrix(4);
-
-#undef shuffleInc
-#undef reduceRow
-#undef reduceMatrix
-
-  // now we need to copy the 64 values into main memory. We can't split work
-  // among threads because all variables are in registers. There's 2 ways
-  // to do this:
-  // (1) have 1 thread do 64 writes from registers into global memory
-  // (2) have 1 thread do 64 writes into shared memory, and then 8 threads
-  //     each do 8 writes into global memory. We can just overwrite the shared
-  //     memory from the problem we just solved.
-  // (2) is slightly faster than (1) due to less branching and more ILP
-
-  // TODO: won't yield much gain, but could just use currently unused shared mem
-  //       and then we won't have to sync
-  // wait for shared mem to be out of use
-  __syncthreads();
-
-#define writeResultShmem(i, j)                                          \
-  lhs_shmem[i + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * j] = res(i, j); \
-
-#define writeRow(i)                             \
-  writeResultShmem(i, 0);                       \
-  writeResultShmem(i, 1);                       \
-  writeResultShmem(i, 2);                       \
-  writeResultShmem(i, 3);                       \
-  writeResultShmem(i, 4);                       \
-  writeResultShmem(i, 5);                       \
-  writeResultShmem(i, 6);                       \
-  writeResultShmem(i, 7);                       \
-
-  if (hipThreadIdx_x == 0) {
-    writeRow(0);
-    writeRow(1);
-    writeRow(2);
-    writeRow(3);
-    writeRow(4);
-    writeRow(5);
-    writeRow(6);
-    writeRow(7);
-  }
-#undef writeResultShmem
-#undef writeRow
-
-  const int max_i_write = numext::mini((int)((m_size - base_m - hipThreadIdx_y + 7) / 8), 8);
-  const int max_j_write = numext::mini((int)((n_size - base_n - hipThreadIdx_z + 7) / 8), 8);
-
-  if (hipThreadIdx_x < max_i_write) {
-    if (max_j_write == 8) {
-      // TODO: can i trade bank conflicts for coalesced writes?
-      Scalar val0 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 0];
-      Scalar val1 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 1];
-      Scalar val2 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 2];
-      Scalar val3 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 3];
-      Scalar val4 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 4];
-      Scalar val5 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 5];
-      Scalar val6 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 6];
-      Scalar val7 = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * 7];
-
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 0) = val0;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 1) = val1;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 2) = val2;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 3) = val3;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 4) = val4;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 5) = val5;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 6) = val6;
-      output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * 7) = val7;
-    } else {
-#pragma unroll 7
-      for (int j = 0; j < max_j_write; j++) {
-        Scalar val = lhs_shmem[hipThreadIdx_x + 8 * hipThreadIdx_y + 64 * hipThreadIdx_z + 512 * j];
-        output(base_m + hipThreadIdx_y + 8 * hipThreadIdx_x, base_n + hipThreadIdx_z + 8 * j) = val;
-      }
-    }
-  }
-#undef res
-}
-
-
-template<typename Scalar, typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper>
-__global__ void
-__launch_bounds__(512, 1)
-EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
-                       const OutputMapper output,
-                       const Index m_size, const Index n_size, const Index k_size) {
-  __shared__ Scalar lhs_shmem[72 * 64];
-  __shared__ Scalar rhs_shmem[72 * 64];
-
-  const Index m_block_idx = hipBlockIdx_x;
-  const Index n_block_idx = hipBlockIdx_y;
-
-  const Index base_m = 64 * m_block_idx;
-  const Index base_n = 64 * n_block_idx;
-
-  if (base_m + 63 < m_size && base_n + 63 < n_size) {
-    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
-  } else {
-    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
-  }
-}
-
-
-template<typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
-         bool CHECK_RHS_BOUNDARY>
-__device__ EIGEN_STRONG_INLINE void
-EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rhs,
-                       const OutputMapper output, float2 lhs_shmem2[][16],
-                       float2 rhs_shmem2[][8], const Index m_size,
-                       const Index n_size, const Index k_size,
-                       const Index base_m, const Index base_n) {
-
-  // prefetch registers
-  float4 lhs_pf0, rhs_pf0;
-
-  float4 results[4];
-  for (int i=0; i < 4; i++) {
-    results[i].x = results[i].y = results[i].z = results[i].w = 0;
-  }
-
-
-#define prefetch_lhs(reg, row, col)                   \
-    if (!CHECK_LHS_BOUNDARY) {                        \
-      if (col < k_size) {                             \
-        /*reg = lhs.template loadPacket<Unaligned>(row, col);*/     \
-        reg.x =lhs(row + 0, col);                     \
-        reg.y =lhs(row + 1, col);                     \
-        reg.z =lhs(row + 2, col);                     \
-        reg.w =lhs(row + 3, col);                     \
-      }                                               \
-    } else {                                          \
-      if (col < k_size) {                             \
-        if (row + 3 < m_size) {                       \
-          /*reg =lhs.template loadPacket<Unaligned>(row, col);*/   \
-          reg.x =lhs(row + 0, col);                   \
-          reg.y =lhs(row + 1, col);                   \
-          reg.z =lhs(row + 2, col);                   \
-          reg.w =lhs(row + 3, col);                   \
-        } else if (row + 2 < m_size) {                \
-          reg.x =lhs(row + 0, col);                   \
-          reg.y =lhs(row + 1, col);                   \
-          reg.z =lhs(row + 2, col);                   \
-        } else if (row + 1 < m_size) {                \
-          reg.x =lhs(row + 0, col);                   \
-          reg.y =lhs(row + 1, col);                   \
-        } else if (row  < m_size) {                   \
-          reg.x =lhs(row + 0, col);                   \
-        }                                             \
-      }                                               \
-    }                                                 \
-
-
-  Index lhs_vert = base_m+hipThreadIdx_x*4;
-
-  for (Index k = 0; k < k_size; k += 16) {
-    //lhs_pf0 = internal::pset1<float4>(0);
-    //rhs_pf0 = internal::pset1<float4>(0);
-    lhs_pf0 = make_float4(0, 0, 0, 0);
-    rhs_pf0 = make_float4(0, 0, 0, 0);
-
-    Index lhs_horiz = hipThreadIdx_y+k;
-    prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
-
-    Index rhs_vert = k+(hipThreadIdx_x%4)*4;
-    Index rhs_horiz0 = (hipThreadIdx_x>>2)+hipThreadIdx_y*4+base_n;
-
-    if (!CHECK_RHS_BOUNDARY) {
-      if ((rhs_vert + 3) < k_size) {
-        // just CHECK_RHS_BOUNDARY
-        //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-        rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
-      } else if (rhs_vert + 2 < k_size) {
-        // just CHECK_RHS_BOUNDARY
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-      } else if (rhs_vert + 1 < k_size) {
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-      } else if (rhs_vert  < k_size) {
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-      }
-    } else {
-      if (rhs_horiz0 < n_size) {
-        if ((rhs_vert + 3) < k_size) {
-          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
-        } else if ((rhs_vert + 2) < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-        } else if ((rhs_vert + 1) < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        } else if (rhs_vert  < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        }
-      }
-    }
-    float x1, x2 ;
-    // the following can be a bitwise operation..... some day.
-    if((hipThreadIdx_x%8) < 4) {
-      x1 = rhs_pf0.y;
-      x2 = rhs_pf0.w;
-    } else {
-      x1 = rhs_pf0.x;
-      x2 = rhs_pf0.z;
-    }
-    x1 = __shfl_xor(x1, 4);
-    x2 = __shfl_xor(x2, 4);
-    if((hipThreadIdx_x%8) < 4) {
-      rhs_pf0.y = x1;
-      rhs_pf0.w = x2;
-    } else {
-      rhs_pf0.x = x1;
-      rhs_pf0.z = x2;
-    }
-
-    // We have 64 features.
-    // Row 0 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 0, 1.
-    // Row 1 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 2, 3.
-    // ...
-    // Row 31 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 62, 63
-    // Row 32 -> times (2, 6, 10, 14, 3, 7, 11, 15) for features 0, 1
-    // ...
-    rhs_shmem2[(hipThreadIdx_x>>3)+ hipThreadIdx_y*2][hipThreadIdx_x%8] = make_float2(rhs_pf0.x, rhs_pf0.y);
-    rhs_shmem2[(hipThreadIdx_x>>3)+ hipThreadIdx_y*2+32][hipThreadIdx_x%8] = make_float2(rhs_pf0.z, rhs_pf0.w);
-
-    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
-    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
-    // ...
-    // Row 15 (time 15) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
-    // Row 16 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63)
-    // ...
-
-    lhs_shmem2[hipThreadIdx_y][hipThreadIdx_x] = make_float2(lhs_pf0.x, lhs_pf0.y);
-    lhs_shmem2[hipThreadIdx_y+16][hipThreadIdx_x] = make_float2(lhs_pf0.z, lhs_pf0.w);
-
-
-#define add_vals(fl1, fl2, fr1, fr2)\
-    results[0].x += fl1.x * fr1.x;\
-    results[0].y += fl1.y * fr1.x;\
-    results[0].z += fl2.x * fr1.x;\
-    results[0].w += fl2.y * fr1.x;\
-\
-    results[1].x += fl1.x * fr1.y;\
-    results[1].y += fl1.y * fr1.y;\
-    results[1].z += fl2.x * fr1.y;\
-    results[1].w += fl2.y * fr1.y;\
-\
-    results[2].x += fl1.x * fr2.x;\
-    results[2].y += fl1.y * fr2.x;\
-    results[2].z += fl2.x * fr2.x;\
-    results[2].w += fl2.y * fr2.x;\
-\
-    results[3].x += fl1.x * fr2.y;\
-    results[3].y += fl1.y * fr2.y;\
-    results[3].z += fl2.x * fr2.y;\
-    results[3].w += fl2.y * fr2.y;\
-
-    __syncthreads();
-
-    // Do the multiplies.
-    #pragma unroll
-    for (int koff = 0; koff < 16; koff ++) {
-      // 32 x threads.
-      float2 fl1 = lhs_shmem2[koff][hipThreadIdx_x];
-      float2 fl2 = lhs_shmem2[koff + 16][hipThreadIdx_x];
-
-      int start_feature = hipThreadIdx_y * 4;
-      float2 fr1 = rhs_shmem2[(start_feature>>1) + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
-      float2 fr2 = rhs_shmem2[(start_feature>>1) + 1 + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
-
-      add_vals(fl1, fl2, fr1, fr2)
-    }
-    __syncthreads();
-  }
-
-#undef prefetch_lhs
-#undef add_vals
-
-  Index horiz_base = hipThreadIdx_y*4+base_n;
-  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
-    for (int i = 0; i < 4; i++) {
-      output(lhs_vert, horiz_base + i) = results[i].x;
-      output(lhs_vert + 1, horiz_base + i) = results[i].y;
-      output(lhs_vert + 2, horiz_base + i) = results[i].z;
-      output(lhs_vert + 3, horiz_base + i) = results[i].w;
-    }
-  } else if (!CHECK_RHS_BOUNDARY) {
-    // CHECK LHS
-    if (lhs_vert + 3 < m_size) {
-      for (int i = 0; i < 4; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        output(lhs_vert + 3, horiz_base + i) = results[i].w;
-      }
-    } else if (lhs_vert + 2 < m_size) {
-      for (int i = 0; i < 4; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-      }
-    } else if (lhs_vert + 1 < m_size) {
-      for (int i = 0; i < 4; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-      }
-    } else if (lhs_vert  < m_size) {
-      for (int i = 0; i < 4; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-      }
-    }
-  } else if (!CHECK_LHS_BOUNDARY) {
-    // CHECK RHS
-    /*
-    int ncols_rem = fminf(n_size- horiz_base, 4);
-    for (int i = 0; i < ncols_rem; i++) {
-      output(lhs_vert, horiz_base + i) = results[i].x;
-      output(lhs_vert + 1, horiz_base + i) = results[i].y;
-      output(lhs_vert + 2, horiz_base + i) = results[i].z;
-      output(lhs_vert + 3, horiz_base + i) = results[i].w;
-    }*/
-    for (int i = 0; i < 4; i++) {
-      if (horiz_base+i < n_size) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        output(lhs_vert + 3, horiz_base + i) = results[i].w;
-       }
-    }
-  } else {
-    // CHECK both boundaries.
-    for (int i = 0; i < 4; i++) {
-      if (horiz_base+i < n_size) {
-        if (lhs_vert < m_size)
-          output(lhs_vert, horiz_base + i) = results[i].x;
-        if (lhs_vert + 1 < m_size)
-          output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        if (lhs_vert + 2 < m_size)
-          output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        if (lhs_vert + 3 < m_size)
-          output(lhs_vert + 3, horiz_base + i) = results[i].w;
-      }
-    }
-  }
-}
-
-
-template<typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
-         bool CHECK_RHS_BOUNDARY>
-__device__ EIGEN_STRONG_INLINE void
-EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
-                       const OutputMapper output, float2 lhs_shmem2[][32],
-                       float2 rhs_shmem2[][8], const Index m_size,
-                       const Index n_size, const Index k_size,
-                       const Index base_m, const Index base_n) {
-
-  // prefetch registers
-  float4 lhs_pf0, lhs_pf1, lhs_pf2, lhs_pf3;
-  float4 rhs_pf0, rhs_pf1;
-
-  float4 results[8];
-  for (int i=0; i < 8; i++) {
-    results[i].x = results[i].y = results[i].z = results[i].w = 0;
-  }
-
-
-  Index lhs_vert = base_m+hipThreadIdx_x*4+(hipThreadIdx_y%4)*32;
-  for (Index k = 0; k < k_size; k += 32) {
-    /*lhs_pf0 = internal::pset1<float4>(0);
-    lhs_pf1 = internal::pset1<float4>(0);
-    lhs_pf2 = internal::pset1<float4>(0);
-    lhs_pf3 = internal::pset1<float4>(0);
-
-    rhs_pf0 = internal::pset1<float4>(0);
-    rhs_pf1 = internal::pset1<float4>(0);*/
-
-
-    lhs_pf0 = make_float4(0, 0, 0, 0);
-    lhs_pf1 = make_float4(0, 0, 0, 0);
-    lhs_pf2 = make_float4(0, 0, 0, 0);
-    lhs_pf3 = make_float4(0, 0, 0, 0);
-
-    rhs_pf0 = make_float4(0, 0, 0, 0);
-    rhs_pf1 = make_float4(0, 0, 0, 0);
-
-     if (!CHECK_LHS_BOUNDARY) {
-      if ((hipThreadIdx_y/4+k+24) < k_size) {
-        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-        //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
-        //lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+24));
-        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-        lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
-        lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
-        lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
-        lhs_pf3.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
-      } else if ((hipThreadIdx_y/4+k+16) < k_size) {
-        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-        //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
-        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-        lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-      } else if ((hipThreadIdx_y/4+k+8) < k_size) {
-        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-        //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-      } else if ((hipThreadIdx_y/4+k) < k_size) {
-        //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-        lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-        lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-        lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-      }
-    } else {
-      // just CHECK_LHS_BOUNDARY
-      if (lhs_vert + 3 < m_size) {
-        if ((hipThreadIdx_y/4+k+24) < k_size) {
-          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-          //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
-          //lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+24));
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
-        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
-          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-          //lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+16));
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
-          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-          //lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k+8));
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        } else if ((hipThreadIdx_y/4+k) < k_size) {
-          //lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (hipThreadIdx_y/4+k));
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf0.w =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        }
-      } else if (lhs_vert + 2 < m_size) {
-        if ((hipThreadIdx_y/4+k+24) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+24));
-        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+16));
-        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k+8));
-        } else if ((hipThreadIdx_y/4+k) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf0.z =lhs(lhs_vert + 2, (hipThreadIdx_y/4+k));
-        }
-      } else if (lhs_vert + 1 < m_size) {
-        if ((hipThreadIdx_y/4+k+24) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
-          lhs_pf3.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+24));
-        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf2.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+16));
-        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf1.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k+8));
-        } else if ((hipThreadIdx_y/4+k) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf0.y =lhs(lhs_vert + 1, (hipThreadIdx_y/4+k));
-        }
-      } else if (lhs_vert < m_size) {
-        if ((hipThreadIdx_y/4+k+24) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-          lhs_pf3.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+24));
-        } else if ((hipThreadIdx_y/4+k+16) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-          lhs_pf2.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+16));
-        } else if ((hipThreadIdx_y/4+k+8) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-          lhs_pf1.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k+8));
-        } else if ((hipThreadIdx_y/4+k) < k_size) {
-          lhs_pf0.x =lhs(lhs_vert + 0, (hipThreadIdx_y/4+k));
-        }
-      }
-    }
-    __syncthreads();
-    Index rhs_vert = k+hipThreadIdx_x*4;
-    Index rhs_horiz0 = hipThreadIdx_y*2+base_n;
-    Index rhs_horiz1 = hipThreadIdx_y*2+1+base_n;
-    if (!CHECK_RHS_BOUNDARY) {
-      if ((rhs_vert + 3) < k_size) {
-        // just CHECK_RHS_BOUNDARY
-        //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-        //rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-        rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
-        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-        rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
-        rhs_pf1.w = rhs(rhs_vert + 3, rhs_horiz1);
-      } else if (rhs_vert + 2 < k_size) {
-        // just CHECK_RHS_BOUNDARY
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-        rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
-      } else if (rhs_vert + 1 < k_size) {
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-      } else if (rhs_vert  < k_size) {
-        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-      }
-    } else {
-      if (rhs_horiz1 < n_size) {
-        if ((rhs_vert + 3) < k_size) {
-          // just CHECK_RHS_BOUNDARY
-          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          //rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
-          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-          rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
-          rhs_pf1.w = rhs(rhs_vert + 3, rhs_horiz1);
-        } else if (rhs_vert + 2 < k_size) {
-          // just CHECK_RHS_BOUNDARY
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-          rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
-        } else if (k+hipThreadIdx_x*4 + 1 < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
-        } else if (k+hipThreadIdx_x*4  < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
-        }
-      } else if (rhs_horiz0 < n_size) {
-        if ((rhs_vert + 3) < k_size) {
-          // just CHECK_RHS_BOUNDARY
-          //rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-          rhs_pf0.w = rhs(rhs_vert + 3, rhs_horiz0);
-        } else if ((rhs_vert + 2) < k_size) {
-          // just CHECK_RHS_BOUNDARY
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
-        } else if ((rhs_vert + 1) < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
-        } else if (rhs_vert  < k_size) {
-          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
-        }
-      }
-    }
-    __syncthreads();
-    // Loaded. Do computation
-    // Row 0 -> times (0, 4, 8, .. 28) for features 0, 1.
-    // Row 1 -> times (0, 4, 8, .. 28) for features 2, 3.
-    // ..
-    // Row 31 -> times (0, 4, 8, .. 28) for features 62, 63
-    rhs_shmem2[hipThreadIdx_y][hipThreadIdx_x] = make_float2(rhs_pf0.x, rhs_pf1.x);
-    // Row 32 -> times (1, 5, 9, .. 29) for features 0, 1.
-    // Row 33 -> times (1, 5, 9, .. 29) for features 2, 3.
-    // ..
-    rhs_shmem2[hipThreadIdx_y+32][hipThreadIdx_x] = make_float2(rhs_pf0.y, rhs_pf1.y);
-    // Row 64 -> times (2, 6, 10, .. 30) for features 0, 1.
-    // Row 65 -> times (2, 6, 10, .. 30) for features 2, 3.
-    rhs_shmem2[hipThreadIdx_y+64][hipThreadIdx_x] = make_float2(rhs_pf0.z, rhs_pf1.z);
-    // Row 96 -> times (3, 7, 11, .. 31) for features 0, 1.
-    // Row 97 -> times (3, 7, 11, .. 31) for features 2, 3.
-    rhs_shmem2[hipThreadIdx_y+96][hipThreadIdx_x] = make_float2(rhs_pf0.w, rhs_pf1.w);
-
-    // LHS.
-    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
-    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
-    // ...
-    // Row 8 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
-    // Row 15 (time 7) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
-
-
-#define add_vals(a_feat1, a_feat2, f1, f2, f3, f4)\
-      results[0].x += a_feat1.x * f1.x;\
-      results[1].x += a_feat1.x * f1.y;\
-      results[2].x += a_feat1.x * f2.x;\
-      results[3].x += a_feat1.x * f2.y;\
-      results[4].x += a_feat1.x * f3.x;\
-      results[5].x += a_feat1.x * f3.y;\
-      results[6].x += a_feat1.x * f4.x;\
-      results[7].x += a_feat1.x * f4.y;\
-\
-      results[0].y += a_feat1.y * f1.x;\
-      results[1].y += a_feat1.y * f1.y;\
-      results[2].y += a_feat1.y * f2.x;\
-      results[3].y += a_feat1.y * f2.y;\
-      results[4].y += a_feat1.y * f3.x;\
-      results[5].y += a_feat1.y * f3.y;\
-      results[6].y += a_feat1.y * f4.x;\
-      results[7].y += a_feat1.y * f4.y;\
-\
-      results[0].z += a_feat2.x * f1.x;\
-      results[1].z += a_feat2.x * f1.y;\
-      results[2].z += a_feat2.x * f2.x;\
-      results[3].z += a_feat2.x * f2.y;\
-      results[4].z += a_feat2.x * f3.x;\
-      results[5].z += a_feat2.x * f3.y;\
-      results[6].z += a_feat2.x * f4.x;\
-      results[7].z += a_feat2.x * f4.y;\
-\
-      results[0].w += a_feat2.y * f1.x;\
-      results[1].w += a_feat2.y * f1.y;\
-      results[2].w += a_feat2.y * f2.x;\
-      results[3].w += a_feat2.y * f2.y;\
-      results[4].w += a_feat2.y * f3.x;\
-      results[5].w += a_feat2.y * f3.y;\
-      results[6].w += a_feat2.y * f4.x;\
-      results[7].w += a_feat2.y * f4.y;\
-
-    lhs_shmem2[hipThreadIdx_y/4][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf0.x, lhs_pf0.y);
-    lhs_shmem2[hipThreadIdx_y/4+8][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf1.x, lhs_pf1.y);
-    lhs_shmem2[hipThreadIdx_y/4+16][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf2.x, lhs_pf2.y);
-    lhs_shmem2[hipThreadIdx_y/4+24][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf3.x, lhs_pf3.y);
-
-    lhs_shmem2[hipThreadIdx_y/4 + 32][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf0.z, lhs_pf0.w);
-    lhs_shmem2[hipThreadIdx_y/4 + 40][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf1.z, lhs_pf1.w);
-    lhs_shmem2[hipThreadIdx_y/4 + 48][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf2.z, lhs_pf2.w);
-    lhs_shmem2[hipThreadIdx_y/4 + 56][hipThreadIdx_x+(hipThreadIdx_y%4)*8] = make_float2(lhs_pf3.z, lhs_pf3.w);
-
-    __syncthreads();
-
-    // Do the multiplies.
-    #pragma unroll
-    for (int koff = 0; koff < 32; koff ++) {
-      float2 a3 = lhs_shmem2[koff][hipThreadIdx_x + (hipThreadIdx_y % 4) * 8];
-      float2 a4 = lhs_shmem2[koff + 32][hipThreadIdx_x + (hipThreadIdx_y % 4) * 8];
-
-      // first feature is at (hipThreadIdx_y/4) * 8 last is at start + 8.
-      int start_feature = (hipThreadIdx_y / 4) * 8;
-
-      float2 br1 = rhs_shmem2[start_feature/2 +     (koff % 4) * 32][koff/4];
-      float2 br2 = rhs_shmem2[start_feature/2 + 1 + (koff % 4) * 32][koff/4];
-      float2 br3 = rhs_shmem2[start_feature/2 + 2 + (koff % 4) * 32][koff/4];
-      float2 br4 = rhs_shmem2[start_feature/2 + 3 + (koff % 4) * 32][koff/4];
-
-      add_vals(a3, a4, br1, br2, br3, br4)
-    }
-    __syncthreads();
-  } // end loop over k
-
-
-  __syncthreads();
-  Index horiz_base = (hipThreadIdx_y/4)*8+base_n;
-  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
-    for (int i = 0; i < 8; i++) {
-      output(lhs_vert, horiz_base + i) = results[i].x;
-      output(lhs_vert + 1, horiz_base + i) = results[i].y;
-      output(lhs_vert + 2, horiz_base + i) = results[i].z;
-      output(lhs_vert + 3, horiz_base + i) = results[i].w;
-    }
-  } else if (!CHECK_RHS_BOUNDARY) {
-    if (lhs_vert + 3 < m_size) {
-      for (int i = 0; i < 8; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        output(lhs_vert + 3, horiz_base + i) = results[i].w;
-      }
-    } else if (lhs_vert + 2 < m_size) {
-      for (int i = 0; i < 8; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-      }
-    } else if (lhs_vert + 1 < m_size) {
-      for (int i = 0; i < 8; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-      }
-    } else if (lhs_vert  < m_size) {
-      for (int i = 0; i < 8; i++) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-      }
-    }
-  } else if (!CHECK_LHS_BOUNDARY) {
-    // CHECK BOUNDARY_B
-    for (int i = 0; i < 8; i++) {
-      if (horiz_base + i < n_size) {
-        output(lhs_vert, horiz_base + i) = results[i].x;
-        output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        output(lhs_vert + 3, horiz_base + i) = results[i].w;
-      }
-    }
-  } else {
-    // CHECK both boundaries.
-    for (int i = 0; i < 8; i++) {
-      if (horiz_base + i < n_size) {
-        if (lhs_vert < m_size)
-          output(lhs_vert, horiz_base + i) = results[i].x;
-        if (lhs_vert + 1 < m_size)
-          output(lhs_vert + 1, horiz_base + i) = results[i].y;
-        if (lhs_vert + 2 < m_size)
-          output(lhs_vert + 2, horiz_base + i) = results[i].z;
-        if (lhs_vert + 3 < m_size)
-          output(lhs_vert + 3, horiz_base + i) = results[i].w;
-      }
-    }
-  }
-}
-
-
-template<typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper>
-__global__ void
-__launch_bounds__(256, 1)
-EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
-                       const OutputMapper output,
-                       const Index m_size, const Index n_size, const Index k_size) {
-  __shared__ float2 lhs_shmem[64*32];
-  __shared__ float2 rhs_shmem[128*8];
-
-  typedef float2 LHS_MEM[64][32];
-  typedef float2 RHS_MEM[128][8];
-
-  typedef float2 LHS_MEM16x16[32][16];
-  typedef float2 RHS_MEM16x16[64][8];
-
-  const Index m_block_idx = hipBlockIdx_x;
-  const Index n_block_idx = hipBlockIdx_y;
-
-  const Index base_m = 128 * m_block_idx;
-  const Index base_n = 64 * n_block_idx;
-
-  bool check_rhs = (base_n + 63) >= n_size;
-  bool check_lhs128 = (base_m + 127) >= m_size;
-
-  if (!check_rhs) {
-    if (!check_lhs128) {
-      // >= 128 rows left
-      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(
-                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
-    } else {
-      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(
-                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
-    }
-  } else {
-    if (!check_lhs128) {
-      // >= 128 rows left
-      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(
-                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
-    } else {
-      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(
-                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
-    }
-  }
-}
-
-template<typename Index, typename LhsMapper,
-         typename RhsMapper, typename OutputMapper>
-__global__ void
-__launch_bounds__(256, 1)
-EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
-                       const OutputMapper output,
-                       const Index m_size, const Index n_size, const Index k_size) {
-  __shared__ float2 lhs_shmem[32][16];
-  __shared__ float2 rhs_shmem[64][8];
-
-  const Index m_block_idx = hipBlockIdx_x;
-  const Index n_block_idx = hipBlockIdx_y;
-
-  const Index base_m = 64 * m_block_idx;
-  const Index base_n = 64 * n_block_idx;
-
-  if (base_m + 63 < m_size) {
-    if (base_n + 63 < n_size) {
-      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
-    } else {
-      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
-    }
-  } else {
-    if (base_n + 63 < n_size) {
-      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
-    } else {
-      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
-    }
-  }
-}
-
-
-template<typename Indices, typename LeftArgType, typename RightArgType>
-struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> :
-    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, GpuDevice> > {
-
-  typedef GpuDevice Device;
-
-  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
-  typedef TensorContractionEvaluatorBase<Self> Base;
-
-  typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
-  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
-  typedef typename XprType::Index Index;
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
-  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
-
-  enum {
-    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
-  };
-
-  // Most of the code is assuming that both input tensors are ColMajor. If the
-  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
-  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
-  // will pretend B is LHS and A is RHS.
-  typedef typename internal::conditional<
-    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
-  typedef typename internal::conditional<
-    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
-
-  static const int LDims =
-      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
-  static const int RDims =
-      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
-  static const int ContractDims = internal::array_size<Indices>::value;
-
-  typedef array<Index, LDims> left_dim_mapper_t;
-  typedef array<Index, RDims> right_dim_mapper_t;
-
-  typedef array<Index, ContractDims> contract_t;
-  typedef array<Index, LDims - ContractDims> left_nocontract_t;
-  typedef array<Index, RDims - ContractDims> right_nocontract_t;
-
-  static const int NumDims = LDims + RDims - 2 * ContractDims;
-
-  typedef DSizes<Index, NumDims> Dimensions;
-
-  // typedefs needed in evalTo
-  typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
-  typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
-
-  typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
-  typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
-
-  typedef typename LeftEvaluator::Dimensions LeftDimensions;
-  typedef typename RightEvaluator::Dimensions RightDimensions;
-
-  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) :
-      Base(op, device) {}
- 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
-
-  // We need to redefine this method to make hipcc happy
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
-    this->m_leftImpl.evalSubExprsIfNeeded(NULL);
-    this->m_rightImpl.evalSubExprsIfNeeded(NULL);
-    if (data) {
-      evalTo(data);
-      return false;
-    } else {
-      this->m_result = static_cast<Scalar *>(this->m_device.allocate(this->dimensions().TotalSize() * sizeof(Scalar)));
-      evalTo(this->m_result);
-      return true;
-    }
-  }
-
-  void evalTo(Scalar* buffer) const {
-    if (this->m_lhs_inner_dim_contiguous) {
-      if (this->m_rhs_inner_dim_contiguous) {
-        if (this->m_rhs_inner_dim_reordered) {
-          evalTyped<true, true, true, Unaligned>(buffer);
-        }
-        else {
-          evalTyped<true, true, false, Unaligned>(buffer);
-        }
-      }
-      else {
-       if (this->m_rhs_inner_dim_reordered) {
-          evalTyped<true, false, true, Unaligned>(buffer);
-        }
-        else {
-          evalTyped<true, false, false, Unaligned>(buffer);
-        }
-      }
-    }
-    else {
-      if (this->m_rhs_inner_dim_contiguous) {
-        if (this->m_rhs_inner_dim_reordered) {
-          evalTyped<false, true, true, Unaligned>(buffer);
-        }
-        else {
-          evalTyped<false, true, false, Unaligned>(buffer);
-        }
-      }
-      else {
-       if (this->m_rhs_inner_dim_reordered) {
-          evalTyped<false, false, true, Unaligned>(buffer);
-        }
-        else {
-          evalTyped<false, false, false, Unaligned>(buffer);
-        }
-      }
-    }
-  }
-
-  template <typename LhsScalar, typename RhsScalar, typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels {
-    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
-    const Index m_blocks = (m + 63) / 64;
-    const Index n_blocks = (n + 63) / 64;
-    const dim3 num_blocks(m_blocks, n_blocks, 1);
-    const dim3 block_size(8, 8, 8);
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>),
-                    dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
-    }
-  };
-
-  template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels<float, float, Index, LhsMapper, RhsMapper, OutputMapper> {
-    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
-      if (m < 768 || n < 768) {
-        const Index m_blocks = (m + 63) / 64;
-        const Index n_blocks = (n + 63) / 64;
-        const dim3 num_blocks(m_blocks, n_blocks, 1);
-        const dim3 block_size(16, 16, 1);
-        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>),
-                        dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
-      } else {
-        const Index m_blocks = (m + 127) / 128;
-        const Index n_blocks = (n + 63) / 64;
-        const dim3 num_blocks(m_blocks, n_blocks, 1);
-        const dim3 block_size(8, 32, 1);
-        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>),
-                        dim3(num_blocks), dim3(block_size), 0, device.stream(), lhs, rhs, output, m, n, k);
-      }
-    }
-  };
-
-  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  void evalTyped(Scalar* buffer) const {
-    // columns in left side, rows in right side
-    const Index k = this->m_k_size;
-    EIGEN_UNUSED_VARIABLE(k)
-
-    // rows in left side
-    const Index m = this->m_i_size;
-
-    // columns in right side
-    const Index n = this->m_j_size;
-
-    // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar)
-    this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
-
-    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
-                                                   LeftEvaluator, left_nocontract_t,
-                                                   contract_t, 4,
-                                                   lhs_inner_dim_contiguous,
-                                                   false, Unaligned> LhsMapper;
-
-    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
-                                                   RightEvaluator, right_nocontract_t,
-                                                   contract_t, 4,
-                                                   rhs_inner_dim_contiguous,
-                                                   rhs_inner_dim_reordered, Unaligned> RhsMapper;
-
-    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
-
-
-    // initialize data mappers
-    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
-                  this->m_left_contracting_strides, this->m_k_strides);
-
-    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
-                  this->m_right_contracting_strides, this->m_k_strides);
-
-    OutputMapper output(buffer, m);
-
-    setHipSharedMemConfig(hipSharedMemBankSizeEightByte);
-    LaunchKernels<LhsScalar, RhsScalar, Index, LhsMapper, RhsMapper, OutputMapper>::Run(lhs, rhs, output,  m, n, k, this->m_device);
-  }
-};
-
-} // end namespace Eigen
-
-#endif // EIGEN_USE_GPU and EIGEN_HIPCC
-#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_HIP_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
deleted file mode 100644
index ba9971050..000000000
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionHip.h
+++ /dev/null
@@ -1,1119 +0,0 @@
-//#include "hip/hip_runtime.h"
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
-#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
-
-namespace Eigen {
-
-/** \class TensorConvolution
-  * \ingroup CXX11_Tensor_Module
-  *
-  * \brief Tensor convolution class.
-  *
-  *
-  */
-namespace internal {
-
-template <typename Index, typename InputDims, int NumKernelDims, int Layout>
-class IndexMapper {
- public:
-  IndexMapper(const InputDims& input_dims, const array<Index, NumKernelDims>& kernel_dims,
-              const array<Index, NumKernelDims>& indices) {
-
-    array<Index, NumDims> dimensions = input_dims;
-    for (int i = 0; i < NumKernelDims; ++i) {
-      const Index index = indices[i];
-      const Index input_dim = input_dims[index];
-      const Index kernel_dim = kernel_dims[i];
-      const Index result_dim = input_dim - kernel_dim + 1;
-      dimensions[index] = result_dim;
-    }
-
-    array<Index, NumDims> inputStrides;
-    array<Index, NumDims> outputStrides;
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      inputStrides[0] = 1;
-      outputStrides[0] = 1;
-      for (int i = 1; i < NumDims; ++i) {
-        inputStrides[i] = inputStrides[i-1] * input_dims[i-1];
-        outputStrides[i] = outputStrides[i-1] * dimensions[i-1];
-      }
-    } else {
-      inputStrides[NumDims - 1] = 1;
-      outputStrides[NumDims - 1] = 1;
-      for (int i = static_cast<int>(NumDims) - 2; i >= 0; --i) {
-        inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
-        outputStrides[i] = outputStrides[i + 1] * dimensions[i + 1];
-      }
-    }
-
-    array<Index, NumDims> hipInputDimensions;
-    array<Index, NumDims> hipOutputDimensions;
-    array<Index, NumDims> tmp = dimensions;
-    array<Index, NumDims> ordering;
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    for (int i = 0; i < NumKernelDims; ++i) {
-      const Index index = i + offset;
-      ordering[index] = indices[i];
-      tmp[indices[i]] = -1;
-      hipInputDimensions[index] = input_dims[indices[i]];
-      hipOutputDimensions[index] = dimensions[indices[i]];
-    }
-
-    int written = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                      ? NumKernelDims
-                      : 0;
-    for (int i = 0; i < NumDims; ++i) {
-      if (tmp[i] >= 0) {
-        ordering[written] = i;
-        hipInputDimensions[written] = input_dims[i];
-        hipOutputDimensions[written] = dimensions[i];
-        ++written;
-      }
-    }
-
-    for (int i = 0; i < NumDims; ++i) {
-      m_inputStrides[i] = inputStrides[ordering[i]];
-      m_outputStrides[i] = outputStrides[ordering[i]];
-    }
-
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int i = 0; i < NumDims; ++i) {
-        if (i > NumKernelDims) {
-          m_hipInputStrides[i] =
-              m_hipInputStrides[i - 1] * hipInputDimensions[i - 1];
-          m_hipOutputStrides[i] =
-              m_hipOutputStrides[i - 1] * hipOutputDimensions[i - 1];
-        } else {
-          m_hipInputStrides[i] = 1;
-          m_hipOutputStrides[i] = 1;
-        }
-      }
-    } else {
-      for (int i = NumDims - 1; i >= 0; --i) {
-        if (i + 1 < offset) {
-          m_hipInputStrides[i] =
-              m_hipInputStrides[i + 1] * hipInputDimensions[i + 1];
-          m_hipOutputStrides[i] =
-              m_hipOutputStrides[i + 1] * hipOutputDimensions[i + 1];
-        } else {
-          m_hipInputStrides[i] = 1;
-          m_hipOutputStrides[i] = 1;
-        }
-      }
-    }
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputPlaneToTensorInputOffset(Index p) const {
-    Index inputIndex = 0;
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int d = NumDims - 1; d > NumKernelDims; --d) {
-        const Index idx = p / m_hipInputStrides[d];
-        inputIndex += idx * m_inputStrides[d];
-        p -= idx * m_hipInputStrides[d];
-      }
-      inputIndex += p * m_inputStrides[NumKernelDims];
-    } else {
-      std::ptrdiff_t limit = 0;
-      if (NumKernelDims < NumDims) {
-        limit = NumDims - NumKernelDims - 1;
-      }
-      for (int d = 0; d < limit; ++d) {
-        const Index idx = p / m_hipInputStrides[d];
-        inputIndex += idx * m_inputStrides[d];
-        p -= idx * m_hipInputStrides[d];
-      }
-      inputIndex += p * m_inputStrides[limit];
-    }
-    return inputIndex;
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputPlaneToTensorOutputOffset(Index p) const {
-    Index outputIndex = 0;
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int d = NumDims - 1; d > NumKernelDims; --d) {
-        const Index idx = p / m_hipOutputStrides[d];
-        outputIndex += idx * m_outputStrides[d];
-        p -= idx * m_hipOutputStrides[d];
-      }
-      outputIndex += p * m_outputStrides[NumKernelDims];
-    } else {
-      std::ptrdiff_t limit = 0;
-      if (NumKernelDims < NumDims) {
-        limit = NumDims - NumKernelDims - 1;
-      }
-      for (int d = 0; d < limit; ++d) {
-        const Index idx = p / m_hipOutputStrides[d];
-        outputIndex += idx * m_outputStrides[d];
-        p -= idx * m_hipOutputStrides[d];
-      }
-      outputIndex += p * m_outputStrides[limit];
-    }
-    return outputIndex;
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_inputStrides[offset];
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_outputStrides[offset];
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i, Index j) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1];
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i, Index j) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1];
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1] +
-           k * m_inputStrides[offset + 2];
-  }
-
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapHipOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
-    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                              ? 0
-                              : NumDims - NumKernelDims;
-    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1] +
-           k * m_outputStrides[offset + 2];
-  }
-
- private:
-  static const int NumDims = internal::array_size<InputDims>::value;
-  array<Index, NumDims> m_inputStrides;
-  array<Index, NumDims> m_outputStrides;
-  array<Index, NumDims> m_hipInputStrides;
-  array<Index, NumDims> m_hipOutputStrides;
-};
-
-
-
-template<typename Dimensions, typename InputXprType, typename KernelXprType>
-struct traits<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >
-{
-  // Type promotion to handle the case where the types of the lhs and the rhs are different.
-  typedef typename promote_storage_type<typename InputXprType::Scalar,
-                                        typename KernelXprType::Scalar>::ret Scalar;
-  typedef typename promote_storage_type<typename traits<InputXprType>::StorageKind,
-                                        typename traits<KernelXprType>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<InputXprType>::Index,
-                                      typename traits<KernelXprType>::Index>::type Index;
-  typedef typename InputXprType::Nested LhsNested;
-  typedef typename KernelXprType::Nested RhsNested;
-  typedef typename remove_reference<LhsNested>::type _LhsNested;
-  typedef typename remove_reference<RhsNested>::type _RhsNested;
-  static const int NumDimensions = traits<InputXprType>::NumDimensions;
-  static const int Layout = traits<InputXprType>::Layout;
-
-  enum {
-    Flags = 0
-  };
-};
-
-template<typename Dimensions, typename InputXprType, typename KernelXprType>
-struct eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, Eigen::Dense>
-{
-  typedef const TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>& type;
-};
-
-template<typename Dimensions, typename InputXprType, typename KernelXprType>
-struct nested<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, 1, typename eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >::type>
-{
-  typedef TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> type;
-};
-
-}  // end namespace internal
-
-
-
-template<typename Indices, typename InputXprType, typename KernelXprType>
-class TensorConvolutionOp : public TensorBase<TensorConvolutionOp<Indices, InputXprType, KernelXprType>, ReadOnlyAccessors>
-{
-  public:
-  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Scalar Scalar;
-  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
-  typedef typename internal::promote_storage_type<typename InputXprType::CoeffReturnType,
-                                                  typename KernelXprType::CoeffReturnType>::ret CoeffReturnType;
-  typedef typename Eigen::internal::nested<TensorConvolutionOp>::type Nested;
-  typedef typename Eigen::internal::traits<TensorConvolutionOp>::StorageKind StorageKind;
-  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Index Index;
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConvolutionOp(const InputXprType& input, const KernelXprType& kernel, const Indices& dims)
-      : m_input_xpr(input), m_kernel_xpr(kernel), m_indices(dims) {}
-
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const Indices& indices() const { return m_indices; }
-
-    /** \returns the nested expressions */
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const typename internal::remove_all<typename InputXprType::Nested>::type&
-    inputExpression() const { return m_input_xpr; }
-
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const typename internal::remove_all<typename KernelXprType::Nested>::type&
-    kernelExpression() const { return m_kernel_xpr; }
-
-  protected:
-    typename InputXprType::Nested m_input_xpr;
-    typename KernelXprType::Nested m_kernel_xpr;
-    const Indices m_indices;
-};
-
-
-template<typename Indices, typename InputArgType, typename KernelArgType, typename Device>
-struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, Device>
-{
-  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
-
-  static const int NumDims = internal::array_size<typename TensorEvaluator<InputArgType, Device>::Dimensions>::value;
-  static const int NumKernelDims = internal::array_size<Indices>::value;
-  typedef typename XprType::Index Index;
-  typedef DSizes<Index, NumDims> Dimensions;
-
-  typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
-  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
-
-  enum {
-    IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<InputArgType, Device>::PacketAccess & TensorEvaluator<KernelArgType, Device>::PacketAccess,
-    Layout = TensorEvaluator<InputArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
-  };
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_kernel(NULL), m_local_kernel(false), m_device(device)
-  {
-    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
-
-    const typename TensorEvaluator<InputArgType, Device>::Dimensions& input_dims = m_inputImpl.dimensions();
-    const typename TensorEvaluator<KernelArgType, Device>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
-
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      m_inputStride[0] = 1;
-      for (int i = 1; i < NumDims; ++i) {
-        m_inputStride[i] = m_inputStride[i - 1] * input_dims[i - 1];
-      }
-    } else {
-      m_inputStride[NumDims - 1] = 1;
-      for (int i = NumDims - 2; i >= 0; --i) {
-        m_inputStride[i] = m_inputStride[i + 1] * input_dims[i + 1];
-      }
-    }
-
-    m_dimensions = m_inputImpl.dimensions();
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int i = 0; i < NumKernelDims; ++i) {
-        const Index index = op.indices()[i];
-        const Index input_dim = input_dims[index];
-        const Index kernel_dim = kernel_dims[i];
-        const Index result_dim = input_dim - kernel_dim + 1;
-        m_dimensions[index] = result_dim;
-        if (i > 0) {
-          m_kernelStride[i] = m_kernelStride[i - 1] * kernel_dims[i - 1];
-        } else {
-          m_kernelStride[0] = 1;
-        }
-        m_indexStride[i] = m_inputStride[index];
-      }
-
-      m_outputStride[0] = 1;
-      for (int i = 1; i < NumDims; ++i) {
-        m_outputStride[i] = m_outputStride[i - 1] * m_dimensions[i - 1];
-      }
-    } else {
-      for (int i = NumKernelDims - 1; i >= 0; --i) {
-        const Index index = op.indices()[i];
-        const Index input_dim = input_dims[index];
-        const Index kernel_dim = kernel_dims[i];
-        const Index result_dim = input_dim - kernel_dim + 1;
-        m_dimensions[index] = result_dim;
-        if (i < NumKernelDims - 1) {
-          m_kernelStride[i] = m_kernelStride[i + 1] * kernel_dims[i + 1];
-        } else {
-          m_kernelStride[NumKernelDims - 1] = 1;
-        }
-        m_indexStride[i] = m_inputStride[index];
-      }
-
-      m_outputStride[NumDims - 1] = 1;
-      for (int i = NumDims - 2; i >= 0; --i) {
-        m_outputStride[i] = m_outputStride[i + 1] * m_dimensions[i + 1];
-      }
-    }
-  }
- 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
-    m_inputImpl.evalSubExprsIfNeeded(NULL);
-    preloadKernel();
-    return true;
-  }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
-    m_inputImpl.cleanup();
-    if (m_local_kernel) {
-      m_device.deallocate((void*)m_kernel);
-      m_local_kernel = false;
-    }
-    m_kernel = NULL;
-  }
-
-  void evalTo(typename XprType::Scalar* buffer) {
-    evalSubExprsIfNeeded(NULL);
-    for (int i = 0; i < dimensions().TotalSize(); ++i) {
-      buffer[i] += coeff(i);
-    }
-    cleanup();
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
-  {
-    CoeffReturnType result = CoeffReturnType(0);
-    convolve(firstInput(index), 0, NumKernelDims-1, result);
-    return result;
-  }
-
-  template<int LoadMode>
-  EIGEN_DEVICE_FUNC PacketReturnType packet(const Index index) const
-  {
-    Index indices[2] = {index, index+PacketSize-1};
-    Index startInputs[2] = {0, 0};
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int i = NumDims - 1; i > 0; --i) {
-        const Index idx0 = indices[0] / m_outputStride[i];
-        const Index idx1 = indices[1] / m_outputStride[i];
-        startInputs[0] += idx0 * m_inputStride[i];
-        startInputs[1] += idx1 * m_inputStride[i];
-        indices[0] -= idx0 * m_outputStride[i];
-        indices[1] -= idx1 * m_outputStride[i];
-      }
-    } else {
-      for (int i = 0; i < NumDims - 1; ++i) {
-        const Index idx0 = indices[0] / m_outputStride[i];
-        const Index idx1 = indices[1] / m_outputStride[i];
-        startInputs[0] += idx0 * m_inputStride[i];
-        startInputs[1] += idx1 * m_inputStride[i];
-        indices[0] -= idx0 * m_outputStride[i];
-        indices[1] -= idx1 * m_outputStride[i];
-      }
-    }
-    startInputs[0] += indices[0];
-    startInputs[1] += indices[1];
-
-    if (startInputs[1]-startInputs[0] == PacketSize-1) {
-      PacketReturnType result = internal::pset1<PacketReturnType>(0);
-      convolvePacket(startInputs[0], 0, NumKernelDims-1, result);
-      return result;
-    } else {
-      EIGEN_ALIGN_MAX Scalar data[PacketSize];
-      data[0] = Scalar(0);
-      convolve(startInputs[0], 0, NumKernelDims-1, data[0]);
-      for (int i = 1; i < PacketSize-1; ++i) {
-        data[i] = Scalar(0);
-        convolve(firstInput(index+i), 0, NumKernelDims-1, data[i]);
-      }
-      data[PacketSize-1] = Scalar(0);
-      convolve(startInputs[1], 0, NumKernelDims-1, data[PacketSize-1]);
-      return internal::pload<PacketReturnType>(data);
-    }
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
-  costPerCoeff(bool vectorized) const {
-    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
-    // We ignore the use of fused multiply-add.
-    const double convolve_compute_cost =
-        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
-    const double firstIndex_compute_cost =
-        NumDims *
-        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
-         TensorOpCost::DivCost<Index>());
-    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
-           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
-                          m_kernelImpl.costPerCoeff(vectorized) +
-                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
-                                       PacketSize));
-  }
-
-  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
-
- private:
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
-    Index startInput = 0;
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      for (int i = NumDims - 1; i > 0; --i) {
-        const Index idx = index / m_outputStride[i];
-        startInput += idx * m_inputStride[i];
-        index -= idx * m_outputStride[i];
-      }
-    } else {
-      for (int i = 0; i < NumDims - 1; ++i) {
-        const Index idx = index / m_outputStride[i];
-        startInput += idx * m_inputStride[i];
-        index -= idx * m_outputStride[i];
-      }
-    }
-    startInput += index;
-    return startInput;
-  }
-
-  EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const {
-    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
-      const Index input = firstIndex + j * m_indexStride[DimIndex];
-      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
-      if (DimIndex > 0) {
-        convolve(input, kernel, DimIndex-1, accum);
-      } else {
-        accum += m_inputImpl.coeff(input) * m_kernel[kernel];
-      }
-    }
-  }
-
-  template <typename Packet>
-  EIGEN_DEVICE_FUNC void convolvePacket(Index firstIndex, Index firstKernel, int DimIndex, Packet& accum) const {
-    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
-      const Index input = firstIndex + j * m_indexStride[DimIndex];
-      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
-      if (DimIndex > 0) {
-        convolvePacket(input, kernel, DimIndex-1, accum);
-      } else {
-        accum = internal::pmadd<Packet>(m_inputImpl.template packet<Unaligned>(input), internal::pset1<Packet>(m_kernel[kernel]), accum);
-      }
-    }
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void preloadKernel() {
-    // Don't make a local copy of the kernel unless we have to (i.e. it's an
-    // expression that needs to be evaluated)
-    const Scalar* in_place = m_kernelImpl.data();
-    if (in_place) {
-      m_kernel = in_place;
-      m_local_kernel = false;
-    } else {
-      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
-      Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
-      typedef TensorEvalToOp<const KernelArgType> EvalTo;
-      EvalTo evalToTmp(local, m_kernelArg);
-      const bool PacketAccess = internal::IsVectorizable<Device, KernelArgType>::value;
-      internal::TensorExecutor<const EvalTo, Device, PacketAccess>::run(evalToTmp, m_device);
-
-      m_kernel = local;
-      m_local_kernel = true;
-    }
-  }
-
-  array<Index, NumDims> m_inputStride;
-  array<Index, NumDims> m_outputStride;
-
-  array<Index, NumKernelDims> m_indexStride;
-  array<Index, NumKernelDims> m_kernelStride;
-  TensorEvaluator<InputArgType, Device> m_inputImpl;
-  TensorEvaluator<KernelArgType, Device> m_kernelImpl;
-  Dimensions m_dimensions;
-
-  KernelArgType m_kernelArg;
-  const Scalar* m_kernel;
-  bool m_local_kernel;
-  const Device& m_device;
-};
-
-
-
-
-// Use an optimized implementation of the evaluation code for GPUs whenever possible.
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
-
-template <int StaticKernelSize>
-struct GetKernelSize {
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int /*kernelSize*/) const {
-    return StaticKernelSize;
-  }
-};
-template <>
-struct GetKernelSize<Dynamic> {
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int kernelSize) const {
-    return kernelSize;
-  }
-};
-
-template <typename InputEvaluator, typename Index, typename InputDims,
-          int StaticKernelSize>
-__global__ void EigenConvolutionKernel1D(
-    InputEvaluator eval,
-    const internal::IndexMapper<Index, InputDims, 1, InputEvaluator::Layout>
-        indexMapper,
-    const float* __restrict kernel, const int numPlanes, const int numX,
-    const int maxX, const int kernelSize, float* buffer) {
-  HIP_DYNAMIC_SHARED( float, s)
-
-  const int first_x = hipBlockIdx_x * maxX;
-  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
-  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSize>()(kernelSize);
-  const int num_x_output = last_x - first_x + 1;
-
-  const int first_plane = hipBlockIdx_y * hipBlockDim_y;
-  const int plane_stride = hipBlockDim_y * hipGridDim_y;
-
-  for (int p = first_plane + hipThreadIdx_y; p < numPlanes; p += plane_stride) {
-    // Load inputs to shared memory
-    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
-    const int plane_kernel_offset = hipThreadIdx_y * num_x_input;
-    #pragma unroll
-    for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
-      const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x);
-      s[i + plane_kernel_offset] = eval.coeff(tensor_index);
-    }
-
-    __syncthreads();
-
-    // Compute the convolution
-    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
-
-    #pragma unroll
-    for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
-      const int kernel_offset = plane_kernel_offset + i;
-      float result = 0.0f;
-      #pragma unroll
-      for (int k = 0; k < GetKernelSize<StaticKernelSize>()(kernelSize); ++k) {
-        result += s[k + kernel_offset] * kernel[k];
-      }
-      const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x);
-      buffer[tensor_index] = result;
-    }
-    __syncthreads();
-  }
-};
-
-template <typename InputEvaluator, typename Index, typename InputDims,
-          int StaticKernelSizeX, int StaticKernelSizeY>
-__global__ void EigenConvolutionKernel2D(
-    InputEvaluator eval,
-    const internal::IndexMapper<Index, InputDims, 2, InputEvaluator::Layout>
-        indexMapper,
-    const float* __restrict kernel, const int numPlanes, const int numX,
-    const int maxX, const int numY, const int maxY, const int kernelSizeX,
-    const int kernelSizeY, float* buffer) {
-  HIP_DYNAMIC_SHARED( float, s)
-
-  const int first_x = hipBlockIdx_x * maxX;
-  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
-  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSizeX>()(kernelSizeX);
-  const int num_x_output = last_x - first_x + 1;
-
-  const int first_y = hipBlockIdx_y * maxY;
-  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
-  const int num_y_input = last_y - first_y + GetKernelSize<StaticKernelSizeY>()(kernelSizeY);
-  const int num_y_output = last_y - first_y + 1;
-
-  const int first_plane = hipBlockIdx_z * hipBlockDim_z;
-  const int plane_stride = hipBlockDim_z * hipGridDim_z;
-
-  for (int p = first_plane + hipThreadIdx_z; p < numPlanes; p += plane_stride) {
-
-    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
-    const int plane_kernel_offset = hipThreadIdx_z * num_y_input;
-
-    // Load inputs to shared memory
-    #pragma unroll
-    for (int j = hipThreadIdx_y; j < num_y_input; j += hipBlockDim_y) {
-      const int input_offset = num_x_input * (j + plane_kernel_offset);
-      #pragma unroll
-      for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
-        const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x, j+first_y);
-        s[i + input_offset] = eval.coeff(tensor_index);
-      }
-    }
-
-    __syncthreads();
-
-    // Convolution
-    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
-
-    #pragma unroll
-    for (int j = hipThreadIdx_y; j < num_y_output; j += hipBlockDim_y) {
-      #pragma unroll
-      for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
-        float result = 0.0f;
-        #pragma unroll
-        for (int l = 0; l < GetKernelSize<StaticKernelSizeY>()(kernelSizeY); ++l) {
-          const int kernel_offset = kernelSizeX * l;
-          const int input_offset = i + num_x_input * (j + l + plane_kernel_offset);
-          #pragma unroll
-          for (int k = 0; k < GetKernelSize<StaticKernelSizeX>()(kernelSizeX); ++k) {
-            result += s[k + input_offset] * kernel[k + kernel_offset];
-          }
-        }
-        const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x, j+first_y);
-        buffer[tensor_index] = result;
-      }
-    }
-
-    __syncthreads();
-  }
-};
-
-template <typename InputEvaluator, typename Index, typename InputDims>
-__global__ void EigenConvolutionKernel3D( 
-    InputEvaluator eval,
-    const internal::IndexMapper<Index, InputDims, 3, InputEvaluator::Layout>
-        indexMapper,
-    const float* __restrict kernel, const size_t numPlanes, const size_t numX,
-    const size_t maxX, const size_t numY, const size_t maxY, const size_t numZ,
-    const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY,
-    const size_t kernelSizeZ, float* buffer) {
-  HIP_DYNAMIC_SHARED( float, s)
-
-  // Load inputs to shared memory
-  const int first_x = hipBlockIdx_x * maxX;
-  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
-  const int num_x_input = last_x - first_x + kernelSizeX;
-
-  const int first_y = hipBlockIdx_y * maxY;
-  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
-  const int num_y_input = last_y - first_y + kernelSizeY;
-
-  const int first_z = hipBlockIdx_z * maxZ;
-  const int last_z = (first_z + maxZ < numZ ? first_z + maxZ : numZ) - 1;
-  const int num_z_input = last_z - first_z + kernelSizeZ;
-
-  for (int p = 0; p < numPlanes; ++p) {
-
-    const int plane_input_offset = indexMapper.mapHipInputPlaneToTensorInputOffset(p);
-    const int plane_kernel_offset = 0;
-
-    for (int k = hipThreadIdx_z; k < num_z_input; k += hipBlockDim_z) {
-      for (int j = hipThreadIdx_y; j < num_y_input; j += hipBlockDim_y) {
-        for (int i = hipThreadIdx_x; i < num_x_input; i += hipBlockDim_x) {
-          const int tensor_index = plane_input_offset + indexMapper.mapHipInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z);
-          s[i + num_x_input * (j + num_y_input * (k + plane_kernel_offset))] = eval.coeff(tensor_index);
-        }
-      }
-    }
-
-    __syncthreads();
-
-    // Convolution
-    const int num_z_output = last_z - first_z + 1;
-    const int num_y_output = last_y - first_y + 1;
-    const int num_x_output = last_x - first_x + 1;
-    const int plane_output_offset = indexMapper.mapHipOutputPlaneToTensorOutputOffset(p);
-
-    for (int k = hipThreadIdx_z; k < num_z_output; k += hipBlockDim_z) {
-      for (int j = hipThreadIdx_y; j < num_y_output; j += hipBlockDim_y) {
-        for (int i = hipThreadIdx_x; i < num_x_output; i += hipBlockDim_x) {
-          float result = 0.0f;
-          for (int n = 0; n < kernelSizeZ; ++n) {
-            for (int m = 0; m < kernelSizeY; ++m) {
-              for (int l = 0; l < kernelSizeX; ++l) {
-                result += s[i + l + num_x_input * (j + m + num_y_input * (k + n + plane_kernel_offset))] * kernel[l + kernelSizeX * (m + kernelSizeY * n)];
-              }
-            }
-          }
-          const int tensor_index = plane_output_offset + indexMapper.mapHipOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z);
-          buffer[tensor_index] = result;
-        }
-      }
-    }
-    __syncthreads();
-  }
-};
-
-
-
-template<typename Indices, typename InputArgType, typename KernelArgType>
-struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, GpuDevice>
-{
-  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
-
-  static const int NumDims =  internal::array_size<typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions>::value;
-  static const int NumKernelDims = internal::array_size<Indices>::value;
-  typedef typename XprType::Index Index;
-  typedef DSizes<Index, NumDims> Dimensions;
-  typedef typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions KernelDimensions;
-
-  enum {
-    IsAligned = TensorEvaluator<InputArgType, GpuDevice>::IsAligned & TensorEvaluator<KernelArgType, GpuDevice>::IsAligned,
-    PacketAccess = false,
-    Layout = TensorEvaluator<InputArgType, GpuDevice>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
-  };
-
-  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const GpuDevice& device)
-      : m_inputImpl(op.inputExpression(), device), m_kernelArg(op.kernelExpression()), m_kernelImpl(op.kernelExpression(), device), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device)
-  {
-    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, GpuDevice>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, GpuDevice>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
-
-    const typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions& input_dims = m_inputImpl.dimensions();
-    const typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
-
-    m_dimensions = m_inputImpl.dimensions();
-    for (int i = 0; i < NumKernelDims; ++i) {
-      const Index index = op.indices()[i];
-      const Index input_dim = input_dims[index];
-      const Index kernel_dim = kernel_dims[i];
-      const Index result_dim = input_dim - kernel_dim + 1;
-      m_dimensions[index] = result_dim;
-    }
-  }
- 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {}
-
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
-  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
-  typedef typename InputArgType::Scalar Scalar;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
-
-  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dimensions; }
-
-  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
-    preloadKernel();
-    m_inputImpl.evalSubExprsIfNeeded(NULL);
-    if (data) {
-      executeEval(data);
-      return false;
-    } else {
-      m_buf = (Scalar*)m_device.allocate(dimensions().TotalSize() * sizeof(Scalar));
-      executeEval(m_buf);
-      return true;
-    }
-  }
-
-  EIGEN_STRONG_INLINE void cleanup() {
-    m_inputImpl.cleanup();
-    if (m_buf) {
-      m_device.deallocate(m_buf);
-      m_buf = NULL;
-    }
-    if (m_local_kernel) {
-      m_device.deallocate((void*)m_kernel);
-      m_local_kernel = false;
-    }
-    m_kernel = NULL;
-  }
-
-  EIGEN_STRONG_INLINE void preloadKernel() {
-    // Don't make a local copy of the kernel unless we have to (i.e. it's an
-    // expression that needs to be evaluated)
-    const Scalar* in_place = m_kernelImpl.data();
-    if (in_place) {
-      m_kernel = in_place;
-      m_local_kernel = false;
-    } else {
-      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
-      Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
-      typedef TensorEvalToOp<const KernelArgType> EvalTo;
-      EvalTo evalToTmp(local, m_kernelArg);
-      const bool PacketAccess = internal::IsVectorizable<GpuDevice, KernelArgType>::value;
-      internal::TensorExecutor<const EvalTo, GpuDevice, PacketAccess>::run(evalToTmp, m_device);
-
-      m_kernel = local;
-      m_local_kernel = true;
-    }
-  }
-
-  static unsigned int ceil(unsigned int num, unsigned int denom) {
-    const unsigned int rounded_toward_zero = num / denom;
-    if (num > rounded_toward_zero * denom) {
-      return rounded_toward_zero + 1;
-    }
-    return rounded_toward_zero;
-  }
-
-  void executeEval(Scalar* data) const {
-    typedef typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions InputDims;
-
-    const int maxSharedMem = m_device.sharedMemPerBlock();
-    const int maxThreadsPerBlock = m_device.maxHipThreadsPerBlock();
-    const int maxBlocksPerProcessor = m_device.maxHipThreadsPerMultiProcessor() / maxThreadsPerBlock;
-    const int numMultiProcessors = m_device.getNumHipMultiProcessors();
-    const int hipWarpSize = 32;
-
-    switch (NumKernelDims) {
-      case 1: {
-        const int kernel_size = m_kernelImpl.dimensions().TotalSize();
-
-        const int numX = dimensions()[m_indices[0]];
-        const int numP = dimensions().TotalSize() / numX;
-        int maxX;
-        dim3 block_size;
-
-        const int single_stride_dim =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                ? 0
-                : m_inputImpl.dimensions().rank() - 1;
-        if (m_indices[0] == single_stride_dim) {
-          // Maximum the reuse
-          const int inner_dim = ((maxSharedMem / (sizeof(Scalar)) - kernel_size + 1 + 31) / 32) * 32;
-          maxX = numext::mini<int>(inner_dim, numX);
-          const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size - 1 + maxX) * sizeof(Scalar)), numP);
-          block_size.x = numext::mini(maxThreadsPerBlock, maxX);
-          block_size.y = numext::mini<int>(maxThreadsPerBlock / block_size.x, maxP);
-        }
-        else {
-          // Read as much as possible alongside the inner most dimension, that is the plane
-          const int inner_dim = maxSharedMem / ((hipWarpSize + kernel_size) * sizeof(Scalar));
-          const int maxP = numext::mini<int>(inner_dim, numP);
-          maxX = numext::mini<int>(maxSharedMem / (inner_dim * sizeof(Scalar)) - kernel_size + 1, numX);
-
-          block_size.x = numext::mini(hipWarpSize, maxX);
-          block_size.y = numext::mini<int>(maxThreadsPerBlock/block_size.x, maxP);
-        }
-
-        const int shared_mem = block_size.y * (maxX + kernel_size - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
-
-        const int num_x_blocks = ceil(numX, maxX);
-        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
-        const int num_y_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks);
-
-        dim3 num_blocks(num_x_blocks, numext::mini<int>(num_y_blocks, ceil(numP, block_size.y)));
-
-
-        //cout << "launching 1D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " maxX: " << maxX << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
-
-        const array<Index, 1> indices(m_indices[0]);
-        const array<Index, 1> kernel_dims(m_kernelImpl.dimensions()[0]);
-        internal::IndexMapper<Index, InputDims, 1, Layout> indexMapper(
-            m_inputImpl.dimensions(), kernel_dims, indices);
-        switch(kernel_size) {
-          case 4: {
-            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4>),
-                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data);
-            break;
-          }
-          case 7: {
-            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7>),
-                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data);
-            break;
-          }
-          default: {
-            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic>),
-                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data);
-          }
-        }
-        break;
-      }
-
-      case 2: {
-        const int idxX =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 1;
-        const int idxY =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 0;
-        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
-        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
-
-        const int numX = dimensions()[m_indices[idxX]];
-        const int numY = dimensions()[m_indices[idxY]];
-        const int numP = dimensions().TotalSize() / (numX*numY);
-
-        const float scaling_factor = sqrtf(static_cast<float>(maxSharedMem) / (sizeof(Scalar) * kernel_size_y * kernel_size_x));
-
-        // Snap maxX to warp size
-        int inner_dim = ((static_cast<int>(scaling_factor * kernel_size_x) - kernel_size_x + 1 + 32) / 32) * 32;
-        const int maxX = numext::mini<int>(inner_dim, numX);
-        const int maxY = numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1)) - kernel_size_y + 1, numY);
-        const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size_x - 1 + maxX) * (kernel_size_y - 1 + maxY) * sizeof(Scalar)), numP);
-
-        dim3 block_size;
-        block_size.x = numext::mini(1024, maxX);
-        block_size.y = numext::mini<int>(1024/block_size.x, maxY);
-        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxP);
-
-        const int shared_mem = block_size.z * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
-
-        const int num_x_blocks = ceil(numX, maxX);
-        const int num_y_blocks = ceil(numY, maxY);
-        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
-        const int num_z_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks * num_y_blocks);
-
-        dim3 num_blocks(num_x_blocks, num_y_blocks, numext::mini<int>(num_z_blocks, ceil(numP, block_size.z)));
-
-
-        //cout << "launching 2D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z << " maxX: " << maxX << " maxY: " << maxY << " maxP: " << maxP << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
-
-        const array<Index, 2> indices(m_indices[idxX], m_indices[idxY]);
-        const array<Index, 2> kernel_dims(m_kernelImpl.dimensions()[idxX],
-                                          m_kernelImpl.dimensions()[idxY]);
-        internal::IndexMapper<Index, InputDims, 2, Layout> indexMapper(
-            m_inputImpl.dimensions(), kernel_dims, indices);
-        switch (kernel_size_x) {
-          case 4: {
-            switch (kernel_size_y) {
-              case 7: {
-                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, 7>),
-                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data);
-                break;
-              }
-              default: {
-                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, Dynamic>),
-                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data);
-                break;
-              }
-            }
-            break;
-          }
-          case 7: {
-            switch (kernel_size_y) {
-              case 4: {
-                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, 4>),
-                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data);
-                break;
-              }
-              default: {
-                hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, Dynamic>),
-                                dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data);
-                break;
-              }
-            }
-            break;
-          }
-          default: {
-            hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic, Dynamic>),
-                            dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data);
-            break;
-          }
-        }
-        break;
-      }
-
-      case 3: {
-        const int idxX =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 2;
-        const int idxY =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 1;
-        const int idxZ =
-            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 2 : 0;
-
-        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
-        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
-        const int kernel_size_z = m_kernelImpl.dimensions()[idxZ];
-
-        const int numX = dimensions()[m_indices[idxX]];
-        const int numY = dimensions()[m_indices[idxY]];
-        const int numZ = dimensions()[m_indices[idxZ]];
-        const int numP = dimensions().TotalSize() / (numX*numY*numZ);
-
-        const int maxX = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * kernel_size_y * kernel_size_z) - kernel_size_x + 1, numX));
-        const int maxY = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * kernel_size_z) - kernel_size_y + 1, numY));
-        const int maxZ = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1)) - kernel_size_z + 1, numZ));
-
-        dim3 block_size;
-        block_size.x = numext::mini(32, maxX);
-        block_size.y = numext::mini(32, maxY);
-        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxZ);
-        dim3 num_blocks(ceil(numX, maxX), ceil(numY, maxY), ceil(numZ, maxZ));
-
-        const int shared_mem = (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * (maxZ + kernel_size_z - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
-
-        //cout << "launching 3D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z  << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
-        const array<Index, 3> indices(m_indices[idxX], m_indices[idxY],
-                                      m_indices[idxZ]);
-        const array<Index, 3> kernel_dims(m_kernelImpl.dimensions()[idxX],
-                                          m_kernelImpl.dimensions()[idxY],
-                                          m_kernelImpl.dimensions()[idxZ]);
-        internal::IndexMapper<Index, InputDims, 3, Layout> indexMapper(
-            m_inputImpl.dimensions(), kernel_dims, indices);
-
-        hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenConvolutionKernel3D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims>),
-                        dim3(num_blocks), dim3(block_size), shared_mem, m_device.stream(), m_inputImpl, indexMapper, m_kernel,
-                        numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data);
-        break;
-      }
-
-      default: {
-        EIGEN_STATIC_ASSERT((NumKernelDims >= 1 && NumKernelDims <= 3), THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE);
-      }
-    }
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
-  {
-    eigen_assert(m_buf);
-    eigen_assert(index < m_dimensions.TotalSize());
-    return m_buf[index];
-  }
-
-  template<int LoadMode>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(const Index index) const
-  {
-    eigen_assert(m_buf);
-    eigen_assert(index < m_dimensions.TotalSize());
-    return internal::ploadt<PacketReturnType, LoadMode>(m_buf+index);
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
-  costPerCoeff(bool vectorized) const {
-    // TODO(rmlarsen): FIXME: For now, this is just a copy of the CPU cost
-    // model.
-    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
-    // We ignore the use of fused multiply-add.
-    const double convolve_compute_cost =
-        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
-    const double firstIndex_compute_cost =
-        NumDims *
-        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
-         TensorOpCost::DivCost<Index>());
-    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
-           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
-                          m_kernelImpl.costPerCoeff(vectorized) +
-                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
-                                       PacketSize));
-  }
-
- private:
-  // No assignment (copies are needed by the kernels)
-  TensorEvaluator& operator = (const TensorEvaluator&);
-
-  TensorEvaluator<InputArgType, GpuDevice> m_inputImpl;
-  TensorEvaluator<KernelArgType, GpuDevice> m_kernelImpl;
-  KernelArgType m_kernelArg;
-  Indices m_indices;
-  Dimensions m_dimensions;
-  Scalar* m_buf;
-  const Scalar* m_kernel;
-  bool m_local_kernel;
-
-  const GpuDevice& m_device;
-};
-#endif
-
-
-} // end namespace Eigen
-
-#endif // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
deleted file mode 100644
index c0e110987..000000000
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceHip.h
+++ /dev/null
@@ -1,352 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H)
-#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H
-
-#if defined(EIGEN_HIPCC)
-#include "hip/hip_runtime.h"
-#include "hip/hip_runtime_api.h"
-#endif
-#include <unistd.h> //for sleep function
-
-namespace Eigen {
-
-static const int kHipScratchSize = 1024;
-
-// This defines an interface that GPUDevice can take to use
-// HIP streams underneath.
-class StreamInterface {
- public:
-  virtual ~StreamInterface() {}
-
-  virtual const hipStream_t& stream() const = 0;
-  virtual const hipDeviceProp_t& deviceProperties() const = 0;
-
-  // Allocate memory on the actual device where the computation will run
-  virtual void* allocate(size_t num_bytes) const = 0;
-  virtual void deallocate(void* buffer) const = 0;
-
-  // Return a scratchpad buffer of size 1k
-  virtual void* scratchpad() const = 0;
-
-  // Return a semaphore. The semaphore is initially initialized to 0, and
-  // each kernel using it is responsible for resetting to 0 upon completion
-  // to maintain the invariant that the semaphore is always equal to 0 upon
-  // each kernel start.
-  virtual unsigned int* semaphore() const = 0;
-};
-
-static hipDeviceProp_t* m_deviceProperties;
-static bool m_devicePropInitialized = false;
-
-static void initializeDeviceProp() {
-  if (!m_devicePropInitialized) {
-    // Attempts to ensure proper behavior in the case of multiple threads
-    // calling this function simultaneously. This would be trivial to
-    // implement if we could use std::mutex, but unfortunately mutex don't
-    // compile with nvcc, so we resort to atomics and thread fences instead.
-    // Note that if the caller uses a compiler that doesn't support c++11 we
-    // can't ensure that the initialization is thread safe.
-#if 0 && __cplusplus >= 201103L
-    static std::atomic<bool> first(true);
-    if (first.exchange(false)) {
-#else
-    static bool first = true;
-    if (first) {
-      first = false;
-#endif
-      // We're the first thread to reach this point.
-      int num_devices;
-      hipError_t status = hipGetDeviceCount(&num_devices);
-      if (status != hipSuccess) {
-        std::cerr << "Failed to get the number of HIP devices: "
-                  << hipGetErrorString(status)
-                  << std::endl;
-        assert(status == hipSuccess);
-      }
-      m_deviceProperties = new hipDeviceProp_t[num_devices];
-      for (int i = 0; i < num_devices; ++i) {
-        status = hipGetDeviceProperties(&m_deviceProperties[i], i);
-        if (status != hipSuccess) {
-          std::cerr << "Failed to initialize HIP device #"
-                    << i
-                    << ": "
-                    << hipGetErrorString(status)
-                    << std::endl;
-          assert(status == hipSuccess);
-        }
-      }
-
-#if 0 && __cplusplus >= 201103L
-      std::atomic_thread_fence(std::memory_order_release);
-#endif
-      m_devicePropInitialized = true;
-    } else {
-      // Wait for the other thread to inititialize the properties.
-      while (!m_devicePropInitialized) {
-#if 0 && __cplusplus >= 201103L
-        std::atomic_thread_fence(std::memory_order_acquire);
-#endif
-        sleep(1);
-      }
-    }
-  }
-}
-
-static const hipStream_t default_stream = 0x00;//TODO: Use hipStreamDefault instead of 0x00;
-
-class HipStreamDevice : public StreamInterface {
- public:
-  // Use the default stream on the current device
-  HipStreamDevice() : stream_(&default_stream), scratch_(NULL), semaphore_(NULL) {
-    hipGetDevice(&device_);
-    initializeDeviceProp();
-  }
-  // Use the default stream on the specified device
-  HipStreamDevice(int device) : stream_(&default_stream), device_(device), scratch_(NULL), semaphore_(NULL) {
-    initializeDeviceProp();
-  }
-  // Use the specified stream. Note that it's the
-  // caller responsibility to ensure that the stream can run on
-  // the specified device. If no device is specified the code
-  // assumes that the stream is associated to the current gpu device.
-  HipStreamDevice(const hipStream_t* stream, int device = -1)
-      : stream_(stream), device_(device), scratch_(NULL), semaphore_(NULL) {
-    if (device < 0) {
-      hipGetDevice(&device_);
-    } else {
-      int num_devices;
-      hipError_t err = hipGetDeviceCount(&num_devices);
-      EIGEN_UNUSED_VARIABLE(err)
-      assert(err == hipSuccess);
-      assert(device < num_devices);
-      device_ = device;
-    }
-    initializeDeviceProp();
-  }
-
-  virtual ~HipStreamDevice() {
-    if (scratch_) {
-      deallocate(scratch_);
-    }
-  }
-
-  const hipStream_t& stream() const { return *stream_; }
-  const hipDeviceProp_t& deviceProperties() const {
-    return m_deviceProperties[device_];
-  }
-  virtual void* allocate(size_t num_bytes) const {
-    hipError_t err = hipSetDevice(device_);
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-    void* result;
-    err = hipMalloc(&result, num_bytes);
-    assert(err == hipSuccess);
-    assert(result != NULL);
-    return result;
-  }
-  virtual void deallocate(void* buffer) const {
-    hipError_t err = hipSetDevice(device_);
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-    assert(buffer != NULL);
-    err = hipFree(buffer);
-    assert(err == hipSuccess);
-  }
-
-  virtual void* scratchpad() const {
-    if (scratch_ == NULL) {
-      scratch_ = allocate(kHipScratchSize + sizeof(unsigned int));
-    }
-    return scratch_;
-  }
-
-  virtual unsigned int* semaphore() const {
-    if (semaphore_ == NULL) {
-      char* scratch = static_cast<char*>(scratchpad()) + kHipScratchSize;
-      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
-      //hipError_t err = hipMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
-      hipError_t err = hipMemset(semaphore_, 0, sizeof(unsigned int));
-      EIGEN_UNUSED_VARIABLE(err)
-      assert(err == hipSuccess);
-    }
-    return semaphore_;
-  }
-
- private:
-  const hipStream_t* stream_;
-  int device_;
-  mutable void* scratch_;
-  mutable unsigned int* semaphore_;
-};
-
-struct GpuDevice {
-  // The StreamInterface is not owned: the caller is
-  // responsible for its initialization and eventual destruction.
-  explicit GpuDevice(const StreamInterface* stream) : stream_(stream), max_blocks_(INT_MAX) {
-    eigen_assert(stream);
-  }
-  explicit GpuDevice(const StreamInterface* stream, int num_blocks) : stream_(stream), max_blocks_(num_blocks) {
-    eigen_assert(stream);
-  }
-  // TODO(bsteiner): This is an internal API, we should not expose it.
-  EIGEN_STRONG_INLINE const hipStream_t& stream() const {
-    return stream_->stream();
-  }
-
-  EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
-    return stream_->allocate(num_bytes);
-  }
-
-  EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
-    stream_->deallocate(buffer);
-  }
-
-  EIGEN_STRONG_INLINE void* scratchpad() const {
-    return stream_->scratchpad();
-  }
-
-  EIGEN_STRONG_INLINE unsigned int* semaphore() const {
-    return stream_->semaphore();
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
-#if !defined(EIGEN_HIP_DEVICE_COMPILE)
-    hipError_t err = hipMemcpyAsync(dst, src, n, hipMemcpyDeviceToDevice,
-                                      stream_->stream());
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-#else
-  eigen_assert(false && "The default device should be used instead to generate kernel code");
-#endif
-  }
-
-  EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
-    hipError_t err =
-        hipMemcpyAsync(dst, src, n, hipMemcpyHostToDevice, stream_->stream());
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-  }
-
-   EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
-    hipError_t err =
-        hipMemcpyAsync(dst, src, n, hipMemcpyDeviceToHost, stream_->stream());
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
-#if !defined(EIGEN_HIP_DEVICE_COMPILE)
-    //TODO:hipError_t err = hipMemsetAsync(buffer, c, n, stream_->stream());
-    hipError_t err = hipMemset(buffer, c, n);
-    EIGEN_UNUSED_VARIABLE(err)
-    assert(err == hipSuccess);
-#else
-  eigen_assert(false && "The default device should be used instead to generate kernel code");
-#endif
-  }
-
-  EIGEN_STRONG_INLINE size_t numThreads() const {
-    // FIXME
-    return 32;
-  }
-
-  EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
-    // FIXME
-    return 48*1024;
-  }
-
-  EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
-    // We won't try to take advantage of the l2 cache for the time being, and
-    // there is no l3 cache on hip devices.
-    return firstLevelCacheSize();
-  }
-
-// FIXME - this will move into HIP
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
-#undef assert
-#define assert(COND)
-#endif
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void synchronize() const {
-#if defined(EIGEN_HIPCC) && \
-    !defined(EIGEN_HIP_DEVICE_COMPILE)
-    hipError_t err = hipStreamSynchronize(stream_->stream());
-    if (err != hipSuccess) {
-      std::cerr << "Error detected in HIP stream: "
-                << hipGetErrorString(err)
-                << std::endl;
-      assert(err == hipSuccess);
-    }
-#else
-    assert(false && "The default device should be used instead to generate kernel code");
-#endif
-  }
-
-  EIGEN_STRONG_INLINE int getNumHipMultiProcessors() const {
-    return stream_->deviceProperties().multiProcessorCount;
-  }
-  EIGEN_STRONG_INLINE int maxHipThreadsPerBlock() const {
-    return stream_->deviceProperties().maxThreadsPerBlock;
-  }
-  EIGEN_STRONG_INLINE int maxHipThreadsPerMultiProcessor() const {
-    return stream_->deviceProperties().maxThreadsPerMultiProcessor;
-  }
-  EIGEN_STRONG_INLINE int sharedMemPerBlock() const {
-    return stream_->deviceProperties().sharedMemPerBlock;
-  }
-   EIGEN_STRONG_INLINE int majorDeviceVersion() const {
-    return stream_->deviceProperties().major;
-  }
-  EIGEN_STRONG_INLINE int minorDeviceVersion() const {
-    return stream_->deviceProperties().minor;
-  }
-
-  EIGEN_STRONG_INLINE int maxBlocks() const {
-    return max_blocks_;
-  }
-
-  // This function checks if the HIP runtime recorded an error for the
-  // underlying stream device.
-  inline bool ok() const {
-#if defined(EIGEN_HIPCC)
-    hipError_t error = hipStreamQuery(stream_->stream());
-    return (error == hipSuccess) || (error == hipErrorNotReady);
-#else
-    return false;
-#endif
-  }
-
- private:
-  const StreamInterface* stream_;
-  int max_blocks_;
-};
-
-#define LAUNCH_HIP_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
-  hipLaunchKernelGGL(HIP_KERNEL_NAME(kernel), dim3(gridsize), dim3(blocksize), (sharedmem), (device).stream(), (__VA_ARGS__)); \
-  assert(hipGetLastError() == hipSuccess);
-
-
-// FIXME: Should be device and kernel specific.
-#if defined(EIGEN_HIPCC)
-static EIGEN_DEVICE_FUNC inline void setHipSharedMemConfig(hipSharedMemConfig config) {
-#if !defined(EIGEN_HIP_DEVICE_COMPILE)
-  hipError_t status = hipDeviceSetSharedMemConfig(config);
-  EIGEN_UNUSED_VARIABLE(status)
-  assert(status == hipSuccess);
-#else
-  EIGEN_UNUSED_VARIABLE(config)
-#endif
-}
-#endif
-
-}  // end namespace Eigen
-
-#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_HIP_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
deleted file mode 100644
index 5304a22c5..000000000
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionHip.h
+++ /dev/null
@@ -1,815 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
-#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
-
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
-#include "Eigen/src/Core/arch/HIP/hcc/math_constants.h"
-#endif
-
-#if defined(EIGEN_HIPCC)
-#define HIP_WARP_SIZE 64
-#endif
-
-namespace Eigen {
-namespace internal {
-
-
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_HIPCC)
-// Full reducers for GPU, don't vectorize for now
-
-// Reducer function that enables multiple hip thread to safely accumulate at the same
-// output address. It basically reads the current value of the output variable, and
-// attempts to update it with the new value. If in the meantime another hip thread
-// updated the content of the output address it will try again.
-template <typename T, typename R>
-__device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) {
-#if defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
-  if (sizeof(T) == 4)
-  {
-    unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
-    unsigned int newval = oldval;
-    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
-    if (newval == oldval) {
-      return;
-    }
-    unsigned int readback;
-    while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
-      oldval = readback;
-      newval = oldval;
-      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
-      if (newval == oldval) {
-        return;
-      }
-    }
-  }
-  else if (sizeof(T) == 8) {
-    unsigned long long oldval = *reinterpret_cast<unsigned long long*>(output);
-    unsigned long long newval = oldval;
-    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
-    if (newval == oldval) {
-      return;
-    }
-    unsigned long long readback;
-    while ((readback = atomicCAS((unsigned long long*)output, oldval, newval)) != oldval) {
-      oldval = readback;
-      newval = oldval;
-      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
-      if (newval == oldval) {
-        return;
-      }
-    }
-  }
-  else {
-    assert(0 && "Wordsize not supported");
-  }
-#else
-  assert(0 && "Shouldn't be called on unsupported device");
-#endif
-}
-
-// We extend atomicExch to support extra data types
-template <typename Type>
-__device__ inline Type atomicExchCustom(Type* address, Type val) {
-  return atomicExch(address, val);
-}
-
-template <>
-__device__ inline double atomicExchCustom(double* address, double val) {
-  unsigned long long int* address_as_ull = reinterpret_cast<unsigned long long int*>(address);
-  return __longlong_as_double(atomicExch(address_as_ull, __double_as_longlong(val)));
-}
-
-#if defined(EIGEN_HAS_HIP_FP16)
-template <template <typename T> class R>
-__device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer) {
-  unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
-  unsigned int newval = oldval;
-  reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
-  if (newval == oldval) {
-    return;
-  }
-  unsigned int readback;
-  while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
-    oldval = readback;
-    newval = oldval;
-    reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
-    if (newval == oldval) {
-      return;
-    }
-  }
-}
-#endif
-
-template <>
-__device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
-#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
-    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
-  atomicAdd(output, accum);
-#else
-  assert(0 && "Shouldn't be called on unsupported device");
-#endif
-}
-
-
-template <typename CoeffType, typename Index>
-__global__ void ReductionInitKernel(const CoeffType val, Index num_preserved_coeffs, CoeffType* output) {
-  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
-  const Index num_threads = hipBlockDim_x * hipGridDim_x;
-  for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
-    output[i] = val;
-  }
-}
-
-
-template <int BlockSize, int NumPerThread, typename Self,
-          typename Reducer, typename Index>
-__global__ void FullReductionKernel(const Self input, Index num_coeffs,
-                                    typename Self::CoeffReturnType* output, unsigned int* semaphore, Reducer reducer) {
-#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
-    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
-  // Initialize the output value
-  const Index first_index = hipBlockIdx_x * BlockSize * NumPerThread + hipThreadIdx_x;
-  if (hipGridDim_x == 1) {
-    if (first_index == 0) {
-      *output = reducer.initialize();
-    }
-  }
-  else {
-    if (hipThreadIdx_x == 0) {
-      unsigned int block = atomicCAS(semaphore, 0u, 1u);
-      if (block == 0) {
-        // We're the first block to run, initialize the output value
-        atomicExchCustom(output, reducer.initialize());
-        __threadfence();
-        atomicExch(semaphore, 2u);
-      }
-      else {
-        // Wait for the first block to initialize the output value.
-        // Use atomicCAS here to ensure that the reads aren't cached
-        unsigned int val;
-        do {
-          val = atomicCAS(semaphore, 2u, 2u);
-        }
-        while (val < 2u);
-      }
-    }
-  }
-
-  __syncthreads();
-
-  eigen_assert(hipGridDim_x == 1 || *semaphore >= 2u);
-
-  typename Self::CoeffReturnType accum = reducer.initialize();
-  Index max_iter = numext::mini<Index>(num_coeffs - first_index, NumPerThread*BlockSize);
-  for (Index i = 0; i < max_iter; i+=BlockSize) {
-    const Index index = first_index + i;
-    eigen_assert(index < num_coeffs);
-    typename Self::CoeffReturnType val = input.m_impl.coeff(index);
-    reducer.reduce(val, &accum);
-  }
-
-#pragma unroll
-  for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
-    // XXX use std::is_floating_point to determine the type of accum
-    if (std::is_floating_point<typename Self::CoeffReturnType>::value) {
-      reducer.reduce(__shfl_down(static_cast<float>(accum), offset, HIP_WARP_SIZE), &accum);
-    } else {
-      reducer.reduce(__shfl_down(static_cast<int>(accum), offset, HIP_WARP_SIZE), &accum);
-    }
-  }
-
-  if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
-    atomicReduce(output, accum, reducer);
-  }
-
-  if (hipGridDim_x > 1 && hipThreadIdx_x == 0) {
-    // Let the last block reset the semaphore
-    atomicInc(semaphore, hipGridDim_x + 1);
-    __threadfence_system();
-  }
-
-#else
-  assert(0 && "Shouldn't be called on unsupported device");
-#endif
-}
-
-
-#if defined(EIGEN_HAS_HIP_FP16)
-template <typename Self,
-          typename Reducer, typename Index>
-__global__ void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half2* scratch) {
-  eigen_assert(hipBlockDim_x == 1);
-  eigen_assert(hipGridDim_x == 1);
-  if (num_coeffs % 2 != 0) {
-    half last = input.m_impl.coeff(num_coeffs-1);
-    *scratch = __halves2half2(last, reducer.initialize());
-  } else {
-    *scratch = reducer.template initializePacket<half2>();
-  }
-}
-
-template <typename Self,
-          typename Reducer, typename Index>
-__global__ void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output) {
-  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
-  const Index num_threads = hipBlockDim_x * hipGridDim_x;
-  const Index num_packets = num_coeffs / 2;
-  for (Index i = thread_id; i < num_packets; i += num_threads) {
-    ((half2*)output)[i] = reducer.template initializePacket<half2>();
-  }
-
-  if (thread_id == 0 && num_coeffs % 2 != 0) {
-    output[num_coeffs-1] = reducer.initialize();
-  }
-}
-
-template <int BlockSize, int NumPerThread, typename Self,
-          typename Reducer, typename Index>
-__global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs,
-                                    half* output, half2* scratch) {
-  eigen_assert(NumPerThread % 2 == 0);
-
-  const Index first_index = hipBlockIdx_x * BlockSize * NumPerThread + 2*hipThreadIdx_x;
-
-  // Initialize the output value if it wasn't initialized by the ReductionInitKernel
-  if (hipGridDim_x == 1 && first_index == 0) {
-    if (num_coeffs % 2 != 0) {
-      half last = input.m_impl.coeff(num_coeffs-1);
-      *scratch = __halves2half2(last, reducer.initialize());
-    } else {
-      *scratch = reducer.template initializePacket<half2>();
-    }
-    __syncthreads();
-  }
-
-  half2 accum = reducer.template initializePacket<half2>();
-  const Index max_iter = numext::mini<Index>((num_coeffs - first_index) / 2, NumPerThread*BlockSize / 2);
-  for (Index i = 0; i < max_iter; i += BlockSize) {
-    const Index index = first_index + 2*i;
-    eigen_assert(index + 1 < num_coeffs);
-    half2 val = input.m_impl.template packet<Unaligned>(index);
-    reducer.reducePacket(val, &accum);
-  }
-
-#pragma unroll
-  for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
-    // FIXME : remove this workaround once we have native half/half2 support for __shfl_down
-    union { int i; half2 h; } wka_in, wka_out;
-    wka_in.h = accum;
-    wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
-    reducer.reducePacket(wka_out.h, &accum);
-  }
-
-  if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
-    atomicReduce(scratch, accum, reducer);
-  }
-
-  __syncthreads();
-
-  if (hipGridDim_x == 1 && first_index == 0) {
-    half tmp = __low2half(*scratch);
-    reducer.reduce(__high2half(*scratch), &tmp);
-    *output = tmp;
-  }
-}
-
-template <typename Op>
-__global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2* scratch) {
-  eigen_assert(hipThreadIdx_x == 1);
-  half tmp = __low2half(*scratch);
-  reducer.reduce(__high2half(*scratch), &tmp);
-  *output = tmp;
-}
-
-#endif
-
-template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
-struct FullReductionLauncher {
-  static void run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index) {
-    assert(false && "Should only be called on doubles, floats and half floats");
-  }
-};
-
-namespace {
-  std::mutex __eigen_reduction_hip_mutex;
-}
-
-// Specialization for float and double
-template <typename Self, typename Op, typename OutputType, bool PacketAccess>
-struct FullReductionLauncher<
-    Self, Op, OutputType, PacketAccess,
-    typename internal::enable_if<
-      internal::is_same<float, OutputType>::value ||
-      internal::is_same<double, OutputType>::value,
-    void>::type> {
-  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) {
-    // guard FullReductionLauncher with a mutex so only 1 FullReductionKernel
-    // is dispatched at a time
-    std::lock_guard<std::mutex> lock(__eigen_reduction_hip_mutex);
-
-    typedef typename Self::Index Index;
-    typedef typename Self::CoeffReturnType Scalar;
-    const int block_size = 256;
-    const int num_per_thread = 128;
-    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-
-    unsigned int* semaphore = NULL;
-    if (num_blocks > 1) {
-      semaphore = device.semaphore();
-
-      unsigned int semaphore_host = 0xFF;
-      hipMemcpy(&semaphore_host, semaphore, sizeof(unsigned int), hipMemcpyDeviceToHost);
-      if (semaphore_host != 0) {
-        std::cerr << "[WARN][EIGEN][FullReductionLauncher] incorrect semaphore value: "
-                  << semaphore_host << "\n";
-        // wait for all commands on the device to complete so semaphore value
-        // is reset to 0
-        hipDeviceSynchronize();
-
-        // read again
-        hipMemcpy(&semaphore_host, semaphore, sizeof(unsigned int), hipMemcpyDeviceToHost);
-        if (semaphore_host != 0) {
-          std::cerr << "[ERROR][EIGEN][FullReductionLauncher] CRITICAL incorrect semaphore value: "
-                    << semaphore_host << ", apply manual override to 0\n";
-
-          // force set semaphore value to be 0
-          semaphore_host = 0;
-          hipMemcpy(semaphore, &semaphore_host, sizeof(unsigned int), hipMemcpyHostToDevice);
-        }
-      }
-    }
-
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
-                       dim3(num_blocks), dim3(block_size), 0, device.stream(), self, num_coeffs, output, semaphore, reducer);
-  }
-};
-
-#if defined(EIGEN_HAS_HIP_FP16)
-template <typename Self, typename Op>
-struct FullReductionLauncher<Self, Op, Eigen::half, false> {
-  static void run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index) {
-    assert(false && "Should not be called since there is no packet accessor");
-  }
-};
-
-template <typename Self, typename Op>
-struct FullReductionLauncher<Self, Op, Eigen::half, true> {
-  static void run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs) {
-    typedef typename Self::Index Index;
-
-    const int block_size = 256;
-    const int num_per_thread = 128;
-    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    half2* scratch = static_cast<half2*>(device.scratchpad());
-
-    if (num_blocks > 1) {
-      // We initialize the output and the scrathpad outside the reduction kernel when we can't be sure that there
-      // won't be a race conditions between multiple thread blocks.
-      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitFullReduxKernelHalfFloat<Self, Op, Index>),
-                         dim3(1), dim3(1), 0, device.stream(), reducer, self, num_coeffs, scratch);
-    }
-
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(FullReductionKernelHalfFloat<block_size, num_per_thread, Self, Op, Index>),
-                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs, output, scratch);
-
-    if (num_blocks > 1) {
-      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionCleanupKernelHalfFloat<Op>),
-                         dim3(1), dim3(1), 0, device.stream(), reducer, output, scratch);
-    }
-  }
-};
-#endif
-
-
-template <typename Self, typename Op, bool Vectorizable>
-struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
-  // Unfortunately nvidia doesn't support well exotic types such as complex,
-  // so reduce the scope of the optimized version of the code to the simple cases
-  // of doubles, floats and half floats
-#if defined(EIGEN_HAS_HIP_FP16)
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
-      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-       internal::is_same<typename Self::CoeffReturnType, double>::value ||
-       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
-                                                (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-                                                 internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif
-
-  template <typename OutputType>
-  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
-    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
-    const Index num_coeffs = array_prod(self.m_impl.dimensions());
-    // Don't crash when we're called with an input tensor of size 0.
-    if (num_coeffs == 0) {
-      return;
-    }
-
-    FullReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs);
-  }
-};
-
-
-template <int NumPerThread, typename Self,
-          typename Reducer, typename Index>
-__global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
-                                         typename Self::CoeffReturnType* output) {
-#if defined(EIGEN_HIP_DEVICE_COMPILE) && (__HIP_DEVICE_COMPILE__ == 1) &&\
-    defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)
-  typedef typename Self::CoeffReturnType Type;
-  eigen_assert(hipBlockDim_y == 1);
-  eigen_assert(hipBlockDim_z == 1);
-  eigen_assert(hipGridDim_y == 1);
-  eigen_assert(hipGridDim_z == 1);
-
-  const int unroll_times = 16;
-  eigen_assert(NumPerThread % unroll_times == 0);
-
-  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, hipBlockDim_x * NumPerThread);
-  const Index num_input_blocks = input_col_blocks * num_preserved_coeffs;
-
-  const Index num_threads = hipBlockDim_x * hipGridDim_x;
-  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
-
-  // Initialize the output values if they weren't initialized by the ReductionInitKernel
-  if (hipGridDim_x == 1) {
-    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
-      output[i] = reducer.initialize();
-    }
-    __syncthreads();
-  }
-
-  for (Index i = hipBlockIdx_x; i < num_input_blocks; i += hipGridDim_x) {
-    const Index row = i / input_col_blocks;
-
-    if (row < num_preserved_coeffs) {
-      const Index col_block = i % input_col_blocks;
-      const Index col_begin = col_block * hipBlockDim_x * NumPerThread + hipThreadIdx_x;
-
-      Type reduced_val = reducer.initialize();
-
-      for (Index j = 0; j < NumPerThread; j += unroll_times) {
-        const Index last_col = col_begin + hipBlockDim_x * (j + unroll_times - 1);
-        if (last_col >= num_coeffs_to_reduce) {
-          for (Index col = col_begin + hipBlockDim_x * j; col < num_coeffs_to_reduce; col += hipBlockDim_x) {
-            const Type val = input.m_impl.coeff(row * num_coeffs_to_reduce + col);
-            reducer.reduce(val, &reduced_val);
-          }
-          break;
-        } else {
-          // Faster version of the loop with no branches after unrolling.
-#pragma unroll
-          for (int k = 0; k < unroll_times; ++k) {
-            const Index col = col_begin + hipBlockDim_x * (j + k);
-            reducer.reduce(input.m_impl.coeff(row * num_coeffs_to_reduce + col), &reduced_val);
-          }
-        }
-      }
-
-#pragma unroll
-      for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
-        // XXX use std::is_floating_point to determine the type of reduced_val
-        if (std::is_floating_point<Type>::value) {
-          reducer.reduce(__shfl_down(static_cast<float>(reduced_val), offset), &reduced_val);
-        } else {
-          reducer.reduce(__shfl_down(static_cast<int>(reduced_val), offset), &reduced_val);
-        }
-      }
-
-      if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
-        atomicReduce(&(output[row]), reduced_val, reducer);
-      }
-    }
-  }
-#else
-  assert(0 && "Shouldn't be called on unsupported device");
-#endif
-}
-
-#if defined(EIGEN_HAS_HIP_FP16)
-
-template <int NumPerThread, typename Self,
-          typename Reducer, typename Index>
-__global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
-                                              half* output) {
-  eigen_assert(hipBlockDim_y == 1);
-  eigen_assert(hipBlockDim_z == 1);
-  eigen_assert(hipGridDim_y == 1);
-  eigen_assert(hipGridDim_z == 1);
-
-  const int unroll_times = 16;
-  eigen_assert(NumPerThread % unroll_times == 0);
-  eigen_assert(unroll_times % 2 == 0);
-
-  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, hipBlockDim_x * NumPerThread * 2);
-  const Index num_input_blocks = divup<Index>(input_col_blocks * num_preserved_coeffs, 2);
-
-  const Index num_threads = hipBlockDim_x * hipGridDim_x;
-  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
-
-  // Initialize the output values if they weren't initialized by the ReductionInitKernel
-  if (hipGridDim_x == 1) {
-    Index i = 2*thread_id;
-    for (; i + 1 < num_preserved_coeffs; i += 2*num_threads) {
-      half* loc = output + i;
-      *((half2*)loc) = reducer.template initializePacket<half2>();
-    }
-    if (i < num_preserved_coeffs) {
-      output[i] = reducer.initialize();
-    }
-    __syncthreads();
-  }
-
-  for (Index i = hipBlockIdx_x; i < num_input_blocks; i += hipGridDim_x) {
-    const Index row = 2 * (i / input_col_blocks);
-
-    if (row + 1 < num_preserved_coeffs) {
-      const Index col_block = i % input_col_blocks;
-      const Index col_begin = 2 * (col_block * hipBlockDim_x * NumPerThread + hipThreadIdx_x);
-
-      half2 reduced_val1 = reducer.template initializePacket<half2>();
-      half2 reduced_val2 = reducer.template initializePacket<half2>();
-
-      for (Index j = 0; j < NumPerThread; j += unroll_times) {
-        const Index last_col = col_begin + hipBlockDim_x * (j + unroll_times - 1) * 2;
-        if (last_col >= num_coeffs_to_reduce) {
-          Index col = col_begin + hipBlockDim_x * j;
-          for (; col + 1 < num_coeffs_to_reduce; col += hipBlockDim_x) {
-            const half2 val1 = input.m_impl.template packet<Unaligned>(row * num_coeffs_to_reduce + col);
-            reducer.reducePacket(val1, &reduced_val1);
-            const half2 val2 = input.m_impl.template packet<Unaligned>((row+1) * num_coeffs_to_reduce + col);
-            reducer.reducePacket(val2, &reduced_val2);
-          }
-          if (col < num_coeffs_to_reduce) {
-            // Peel;
-            const half last1 = input.m_impl.coeff(row * num_coeffs_to_reduce + col);
-            const half2 val1 = __halves2half2(last1, reducer.initialize());
-            reducer.reducePacket(val1, &reduced_val1);
-            const half last2 = input.m_impl.coeff((row+1) * num_coeffs_to_reduce + col);
-            const half2 val2 = __halves2half2(last2, reducer.initialize());
-            reducer.reducePacket(val2, &reduced_val2);
-          }
-          break;
-        } else {
-          // Faster version of the loop with no branches after unrolling.
-#pragma unroll
-          for (int k = 0; k < unroll_times; ++k) {
-            const Index col = col_begin + hipBlockDim_x * (j + k) * 2;
-            reducer.reducePacket(input.m_impl.template packet<Unaligned>(row * num_coeffs_to_reduce + col), &reduced_val1);
-            reducer.reducePacket(input.m_impl.template packet<Unaligned>((row + 1)* num_coeffs_to_reduce + col), &reduced_val2);
-          }
-        }
-      }
-
-#pragma unroll
-      for (int offset = HIP_WARP_SIZE/2; offset > 0; offset /= 2) {
-	// FIXME : remove this workaround once we have native half/half2 support for __shfl_down
-	union { int i; half2 h; } wka_in, wka_out;
-
-	wka_in.h = reduced_val1;
-	wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
-        reducer.reducePacket(wka_out.h, &reduced_val1);
-	
-	wka_in.h = reduced_val2;
-	wka_out.i = __shfl_down(wka_in.i, offset, HIP_WARP_SIZE);
-        reducer.reducePacket(wka_out.h, &reduced_val2);
-      }
-
-      half val1 =  __low2half(reduced_val1);
-      reducer.reduce(__high2half(reduced_val1), &val1);
-      half val2 =  __low2half(reduced_val2);
-      reducer.reduce(__high2half(reduced_val2), &val2);
-      half2 val = __halves2half2(val1, val2);
-
-      if ((hipThreadIdx_x & (HIP_WARP_SIZE - 1)) == 0) {
-        half* loc = output + row;
-        atomicReduce((half2*)loc, val, reducer);
-      }
-    }
-  }
-}
-
-#endif
-
-template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
-struct InnerReductionLauncher {
-  static bool run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device");
-    return true;
-  }
-};
-
-// Specialization for float and double
-template <typename Self, typename Op, typename OutputType, bool PacketAccess>
-struct InnerReductionLauncher<
-  Self, Op, OutputType, PacketAccess,
-  typename internal::enable_if<
-    internal::is_same<float, OutputType>::value ||
-    internal::is_same<double, OutputType>::value,
-  void>::type> {
-  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
-    typedef typename Self::Index Index;
-
-    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
-    const int block_size = 256;
-    const int num_per_thread = 128;
-    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / block_size;
-    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-
-    if (num_blocks > 1) {
-      // We initialize the outputs outside the reduction kernel when we can't be sure that there
-      // won't be a race conditions between multiple thread blocks.
-      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / 1024;
-      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernel<OutputType, Index>),
-                         dim3(num_blocks), dim3(1024), 0, device.stream(),
-                         reducer.initialize(), num_preserved_vals, output);
-    }
-
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(InnerReductionKernel<num_per_thread, Self, Op, Index>),
-                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self,
-                       num_coeffs_to_reduce, num_preserved_vals, output);
-
-    return false;
-  }
-};
-
-#if defined(EIGEN_HAS_HIP_FP16)
-template <typename Self, typename Op>
-struct InnerReductionLauncher<Self, Op, Eigen::half, false> {
-  static bool run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should not be called since there is no packet accessor");
-    return true;
-  }
-};
-
-template <typename Self, typename Op>
-struct InnerReductionLauncher<Self, Op, Eigen::half, true> {
-  static bool run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
-    typedef typename Self::Index Index;
-
-    if (num_preserved_vals % 2 != 0) {
-      // Not supported yet, revert to the slower code path
-      return true;
-    }
-
-    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
-    const int block_size = /*256*/128;
-    const int num_per_thread = /*128*/64;
-    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / block_size;
-    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-
-    if (num_blocks > 1) {
-      // We initialize the outputs outside the reduction kernel when we can't be sure that there
-      // won't be a race conditions between multiple thread blocks.
-      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / 1024;
-      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernelHalfFloat<Self, Op, Index>),
-                         dim3(1), dim3(1), 0, device.stream(), reducer, self, num_preserved_vals, output);
-    }
-
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(InnerReductionKernelHalfFloat<num_per_thread, Self, Op, Index>),
-                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
-
-    return false;
-  }
-};
-#endif
-
-
-template <typename Self, typename Op>
-struct InnerReducer<Self, Op, GpuDevice> {
-  // Unfortunately nvidia doesn't support well exotic types such as complex,
-  // so reduce the scope of the optimized version of the code to the simple case
-  // of floats and half floats.
-#if defined(EIGEN_HAS_HIP_FP16)
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
-      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-       internal::is_same<typename Self::CoeffReturnType, double>::value ||
-       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
-                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif
-
-  template <typename OutputType>
-  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
-    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
-    const Index num_coeffs = array_prod(self.m_impl.dimensions());
-    // Don't crash when we're called with an input tensor of size 0.
-    if (num_coeffs == 0) {
-      return true;
-    }
-    // It's faster to use the usual code.
-    if (num_coeffs_to_reduce <= 128) {
-      return true;
-    }
-
-    return InnerReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals);
-  }
-};
-
-template <int NumPerThread, typename Self,
-          typename Reducer, typename Index>
-__global__ void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
-                                     typename Self::CoeffReturnType* output) {
-  const Index num_threads = hipBlockDim_x * hipGridDim_x;
-  const Index thread_id = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
-  // Initialize the output values if they weren't initialized by the ReductionInitKernel
-  if (hipGridDim_x == 1) {
-    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
-      output[i] = reducer.initialize();
-    }
-    __syncthreads();
-  }
-
-  // Do the reduction.
-  const Index max_iter = num_preserved_coeffs * divup<Index>(num_coeffs_to_reduce, NumPerThread);
-  for (Index i = thread_id; i < max_iter; i += num_threads) {
-    const Index input_col = i % num_preserved_coeffs;
-    const Index input_row = (i / num_preserved_coeffs) * NumPerThread;
-    typename Self::CoeffReturnType reduced_val = reducer.initialize();
-    const Index max_row = numext::mini(input_row + NumPerThread, num_coeffs_to_reduce);
-    for (Index j = input_row; j < max_row; j++) {
-      typename Self::CoeffReturnType val = input.m_impl.coeff(j * num_preserved_coeffs + input_col);
-      reducer.reduce(val, &reduced_val);
-    }
-    atomicReduce(&(output[input_col]), reduced_val, reducer);
-  }
-}
-
-
-template <typename Self, typename Op>
-struct OuterReducer<Self, Op, GpuDevice> {
-  // Unfortunately nvidia doesn't support well exotic types such as complex,
-  // so reduce the scope of the optimized version of the code to the simple case
-  // of floats.
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
-                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
-  template <typename Device, typename OutputType>
-  static bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should only be called to reduce doubles or floats on a gpu device");
-    return true;
-  }
-
-  static bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
-    typedef typename Self::Index Index;
-
-    // It's faster to use the usual code.
-    if (num_coeffs_to_reduce <= 32) {
-      return true;
-    }
-
-    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
-    const int block_size = 256;
-    const int num_per_thread = 16;
-    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / block_size;
-    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-
-    if (num_blocks > 1) {
-      // We initialize the outputs in the reduction kernel itself when we don't have to worry
-      // about race conditions between multiple thread blocks.
-      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumHipMultiProcessors() *
-                             device.maxHipThreadsPerMultiProcessor() / 1024;
-      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      hipLaunchKernelGGL(HIP_KERNEL_NAME(ReductionInitKernel<float, Index>),
-                         dim3(num_blocks), dim3(1024), 0, device.stream(),
-                         reducer.initialize(), num_preserved_vals, output);
-    }
-
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(OuterReductionKernel<num_per_thread, Self, Op, Index>),
-                       dim3(num_blocks), dim3(block_size), 0, device.stream(), reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
-
-    return false;
-  }
-};
-
-#endif
-
-
-} // end namespace internal
-} // end namespace Eigen
-
-#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_HIP_H
diff --git a/unsupported/test/cxx11_tensor_argmax_hip.cu b/unsupported/test/cxx11_tensor_argmax_hip.cu
deleted file mode 100644
index 57d6ca000..000000000
--- a/unsupported/test/cxx11_tensor_argmax_hip.cu
+++ /dev/null
@@ -1,251 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_FUNC cxx11_tensor_hip
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-using Eigen::Tensor;
-
-template <int Layout>
-void test_hip_simple_argmax()
-{
-  Tensor<double, 3, Layout> in(Eigen::array<DenseIndex, 3>(72,53,97));
-  Tensor<DenseIndex, 1, Layout> out_max(Eigen::array<DenseIndex, 1>(1));
-  Tensor<DenseIndex, 1, Layout> out_min(Eigen::array<DenseIndex, 1>(1));
-  in.setRandom();
-  in *= in.constant(100.0);
-  in(0, 0, 0) = -1000.0;
-  in(71, 52, 96) = 1000.0;
-
-  std::size_t in_bytes = in.size() * sizeof(double);
-  std::size_t out_bytes = out_max.size() * sizeof(DenseIndex);
-
-  double* d_in;
-  DenseIndex* d_out_max;
-  DenseIndex* d_out_min;
-  hipMalloc((void**)(&d_in), in_bytes);
-  hipMalloc((void**)(&d_out_max), out_bytes);
-  hipMalloc((void**)(&d_out_min), out_bytes);
-
-  hipMemcpy(d_in, in.data(), in_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<double, 3, Layout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 3>(72,53,97));
-  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_max(d_out_max, Eigen::array<DenseIndex, 1>(1));
-  Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_min(d_out_min, Eigen::array<DenseIndex, 1>(1));
-
-  gpu_out_max.device(gpu_device) = gpu_in.argmax();
-  gpu_out_min.device(gpu_device) = gpu_in.argmin();
-
-  assert(hipMemcpyAsync(out_max.data(), d_out_max, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipMemcpyAsync(out_min.data(), d_out_min, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  VERIFY_IS_EQUAL(out_max(Eigen::array<DenseIndex, 1>(0)), 72*53*97 - 1);
-  VERIFY_IS_EQUAL(out_min(Eigen::array<DenseIndex, 1>(0)), 0);
-
-  hipFree(d_in);
-  hipFree(d_out_max);
-  hipFree(d_out_min);
-}
-
-template <int DataLayout>
-void test_hip_argmax_dim()
-{
-  Tensor<float, 4, DataLayout> tensor(2,3,5,7);
-  std::vector<int> dims;
-  dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
-
-  for (int dim = 0; dim < 4; ++dim) {
-    tensor.setRandom();
-    tensor = (tensor + tensor.constant(0.5)).log();
-
-    array<DenseIndex, 3> out_shape;
-    for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
-
-    Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
-
-    array<DenseIndex, 4> ix;
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 3; ++j) {
-        for (int k = 0; k < 5; ++k) {
-          for (int l = 0; l < 7; ++l) {
-            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
-            if (ix[dim] != 0) continue;
-            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
-            tensor(ix) = 10.0;
-          }
-        }
-      }
-    }
-
-    std::size_t in_bytes = tensor.size() * sizeof(float);
-    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
-
-    float* d_in;
-    DenseIndex* d_out;
-    hipMalloc((void**)(&d_in), in_bytes);
-    hipMalloc((void**)(&d_out), out_bytes);
-
-    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
-
-    Eigen::HipStreamDevice stream;
-    Eigen::GpuDevice gpu_device(&stream);
-
-    Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
-    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
-
-    gpu_out.device(gpu_device) = gpu_in.argmax(dim);
-
-    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-    VERIFY_IS_EQUAL(tensor_arg.size(),
-                    size_t(2*3*5*7 / tensor.dimension(dim)));
-
-    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
-      // Expect max to be in the first index of the reduced dimension
-      VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
-    }
-
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 3; ++j) {
-        for (int k = 0; k < 5; ++k) {
-          for (int l = 0; l < 7; ++l) {
-            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
-            if (ix[dim] != tensor.dimension(dim) - 1) continue;
-            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
-            tensor(ix) = 20.0;
-          }
-        }
-      }
-    }
-
-    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
-
-    gpu_out.device(gpu_device) = gpu_in.argmax(dim);
-
-    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
-      // Expect max to be in the last index of the reduced dimension
-      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
-    }
-
-    hipFree(d_in);
-    hipFree(d_out);
-  }
-}
-
-template <int DataLayout>
-void test_hip_argmin_dim()
-{
-  Tensor<float, 4, DataLayout> tensor(2,3,5,7);
-  std::vector<int> dims;
-  dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
-
-  for (int dim = 0; dim < 4; ++dim) {
-    tensor.setRandom();
-    tensor = (tensor + tensor.constant(0.5)).log();
-
-    array<DenseIndex, 3> out_shape;
-    for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
-
-    Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
-
-    array<DenseIndex, 4> ix;
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 3; ++j) {
-        for (int k = 0; k < 5; ++k) {
-          for (int l = 0; l < 7; ++l) {
-            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
-            if (ix[dim] != 0) continue;
-            // suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
-            tensor(ix) = -10.0;
-          }
-        }
-      }
-    }
-
-    std::size_t in_bytes = tensor.size() * sizeof(float);
-    std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
-
-    float* d_in;
-    DenseIndex* d_out;
-    hipMalloc((void**)(&d_in), in_bytes);
-    hipMalloc((void**)(&d_out), out_bytes);
-
-    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
-
-    Eigen::HipStreamDevice stream;
-    Eigen::GpuDevice gpu_device(&stream);
-
-    Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
-    Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
-
-    gpu_out.device(gpu_device) = gpu_in.argmin(dim);
-
-    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-    VERIFY_IS_EQUAL(tensor_arg.size(),
-                    2*3*5*7 / tensor.dimension(dim));
-
-    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
-      // Expect min to be in the first index of the reduced dimension
-      VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
-    }
-
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 3; ++j) {
-        for (int k = 0; k < 5; ++k) {
-          for (int l = 0; l < 7; ++l) {
-            ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
-            if (ix[dim] != tensor.dimension(dim) - 1) continue;
-            // suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
-            tensor(ix) = -20.0;
-          }
-        }
-      }
-    }
-
-    hipMemcpy(d_in, tensor.data(), in_bytes, hipMemcpyHostToDevice);
-
-    gpu_out.device(gpu_device) = gpu_in.argmin(dim);
-
-    assert(hipMemcpyAsync(tensor_arg.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-    assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-    for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
-      // Expect max to be in the last index of the reduced dimension
-      VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
-    }
-
-    hipFree(d_in);
-    hipFree(d_out);
-  }
-}
-
-void test_cxx11_tensor_hip()
-{
-  CALL_SUBTEST(test_hip_simple_argmax<RowMajor>());
-  CALL_SUBTEST(test_hip_simple_argmax<ColMajor>());
-  CALL_SUBTEST(test_hip_argmax_dim<RowMajor>());
-  CALL_SUBTEST(test_hip_argmax_dim<ColMajor>());
-  CALL_SUBTEST(test_hip_argmin_dim<RowMajor>());
-  CALL_SUBTEST(test_hip_argmin_dim<ColMajor>());
-}
diff --git a/unsupported/test/cxx11_tensor_cast_float16_hip.cu b/unsupported/test/cxx11_tensor_cast_float16_hip.cu
deleted file mode 100644
index bf6a49df4..000000000
--- a/unsupported/test/cxx11_tensor_cast_float16_hip.cu
+++ /dev/null
@@ -1,79 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_hip
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-using Eigen::Tensor;
-
-void test_hip_conversion() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  Tensor<float, 1> floats(num_elem);
-  floats.setRandom();
-
-  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
-      d_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
-      d_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
-      d_conv, num_elem);
-
-  gpu_device.memcpyHostToDevice(d_float, floats.data(), num_elem*sizeof(float));
-
-  gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
-  gpu_conv.device(gpu_device) = gpu_half.cast<float>();
-
-  Tensor<float, 1> initial(num_elem);
-  Tensor<float, 1> final(num_elem);
-  gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    VERIFY_IS_APPROX(initial(i), final(i));
-  }
-
-  gpu_device.deallocate(d_float);
-  gpu_device.deallocate(d_half);
-  gpu_device.deallocate(d_conv);
-}
-
-
-void test_fallback_conversion() {
-  int num_elem = 101;
-  Tensor<float, 1> floats(num_elem);
-  floats.setRandom();
-
-  Eigen::Tensor<Eigen::half, 1> halfs = floats.cast<Eigen::half>();
-  Eigen::Tensor<float, 1> conv = halfs.cast<float>();
-
-  for (int i = 0; i < num_elem; ++i) {
-    VERIFY_IS_APPROX(floats(i), conv(i));
-  }
-}
-
-
-void test_cxx11_tensor_cast_float16_hip()
-{
-  CALL_SUBTEST(test_hip_conversion());
-  CALL_SUBTEST(test_fallback_conversion());
-}
diff --git a/unsupported/test/cxx11_tensor_contract_hip.cu b/unsupported/test/cxx11_tensor_contract_hip.cu
deleted file mode 100644
index 652af0ab0..000000000
--- a/unsupported/test/cxx11_tensor_contract_hip.cu
+++ /dev/null
@@ -1,215 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-// Copyright (C) 2014 Navdeep Jaitly <ndjaitly@google.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_hip
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-
-using Eigen::Tensor;
-typedef Tensor<float, 1>::DimensionPair DimPair;
-
-template<int DataLayout>
-void test_hip_contraction(int m_size, int k_size, int n_size)
-{
-  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
-  // with these dimensions, the output has 300 * 140 elements, which is
-  // more than 30 * 1024, which is the number of threads in blocks on
-  // a 15 SM GK110 GPU
-  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
-  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
-  Tensor<float, 2, DataLayout> t_result(m_size, n_size);
-  Tensor<float, 2, DataLayout> t_result_gpu(m_size, n_size);
-  Eigen::array<DimPair, 1> dims(DimPair(1, 0));
-
-  t_left.setRandom();
-  t_right.setRandom();
-
-  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
-  std::size_t t_right_bytes = t_right.size() * sizeof(float);
-  std::size_t t_result_bytes = t_result.size() * sizeof(float);
-
-  float* d_t_left;
-  float* d_t_right;
-  float* d_t_result;
-
-  hipMalloc((void**)(&d_t_left), t_left_bytes);
-  hipMalloc((void**)(&d_t_right), t_right_bytes);
-  hipMalloc((void**)(&d_t_result), t_result_bytes);
-
-  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
-      gpu_t_left(d_t_left, Eigen::array<int, 2>(m_size, k_size));
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
-      gpu_t_right(d_t_right, Eigen::array<int, 2>(k_size, n_size));
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
-      gpu_t_result(d_t_result, Eigen::array<int, 2>(m_size, n_size));
-
-
-  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
-  t_result = t_left.contract(t_right, dims);
-
-  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
-  for (DenseIndex i = 0; i < t_result.size(); i++) {
-    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
-      continue;
-    }
-    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
-      continue;
-    }
-    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
-              << " vs " <<  t_result_gpu(i) << std::endl;
-    assert(false);
-  }
-
-  hipFree((void*)d_t_left);
-  hipFree((void*)d_t_right);
-  hipFree((void*)d_t_result);
-}
-
-
-template<int DataLayout>
-void test_scalar(int m_size, int k_size, int n_size)
-{
-  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
-  // with these dimensions, the output has 300 * 140 elements, which is
-  // more than 30 * 1024, which is the number of threads in blocks on
-  // a 15 SM GK110 GPU
-  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
-  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
-  Tensor<float, 0, DataLayout> t_result;
-  Tensor<float, 0, DataLayout> t_result_gpu;
-  Eigen::array<DimPair, 2> dims(DimPair(0, 0), DimPair(1, 1));
-
-  t_left.setRandom();
-  t_right.setRandom();
-
-  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
-  std::size_t t_right_bytes = t_right.size() * sizeof(float);
-  std::size_t t_result_bytes = sizeof(float);
-
-  float* d_t_left;
-  float* d_t_right;
-  float* d_t_result;
-
-  hipMalloc((void**)(&d_t_left), t_left_bytes);
-  hipMalloc((void**)(&d_t_right), t_right_bytes);
-  hipMalloc((void**)(&d_t_result), t_result_bytes);
-
-  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
-      gpu_t_left(d_t_left, m_size, k_size);
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
-      gpu_t_right(d_t_right, k_size, n_size);
-  Eigen::TensorMap<Eigen::Tensor<float, 0, DataLayout> >
-      gpu_t_result(d_t_result);
-
-  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
-  t_result = t_left.contract(t_right, dims);
-
-  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
-  if (fabs(t_result() - t_result_gpu()) > 1e-4f &&
-      !Eigen::internal::isApprox(t_result(), t_result_gpu(), 1e-4f)) {
-    std::cout << "mismatch detected: " << t_result()
-              << " vs " <<  t_result_gpu() << std::endl;
-    assert(false);
-  }
-
-  hipFree((void*)d_t_left);
-  hipFree((void*)d_t_right);
-  hipFree((void*)d_t_result);
-}
-
-
-template<int DataLayout>
-void test_hip_contraction_m() {
-  for (int k = 32; k < 256; k++) {
-    test_hip_contraction<ColMajor>(k, 128, 128);
-    test_hip_contraction<RowMajor>(k, 128, 128);
-  }
-}
-
-template<int DataLayout>
-void test_hip_contraction_k() {
-  for (int k = 32; k < 256; k++) {
-    test_hip_contraction<ColMajor>(128, k, 128);
-    test_hip_contraction<RowMajor>(128, k, 128);
-  }
-}
-
-template<int DataLayout>
-void test_hip_contraction_n() {
-  for (int k = 32; k < 256; k++) {
-    test_hip_contraction<ColMajor>(128, 128, k);
-    test_hip_contraction<RowMajor>(128, 128, k);
-  }
-}
-
-
-template<int DataLayout>
-void test_hip_contraction_sizes() {
-  int m_sizes[] = { 31,  39,   63,   64,   65,
-                   127, 129,  255,  257 , 511,
-                   512, 513, 1023, 1024, 1025};
-
-  int n_sizes[] = { 31,  39,   63,   64,   65,
-                   127, 129,  255,  257,  511,
-                   512, 513, 1023, 1024, 1025};
-
-  int k_sizes[] = {  31,   39,  63,  64,   65,
-                     95,   96, 127, 129,  255,
-                    257,  511, 512, 513, 1023,
-                   1024, 1025};
-
-  for (int i = 0; i < 15; i++) {
-    for (int j = 0; j < 15; j++) {
-      for (int k = 0; k < 17; k++) {
-        test_hip_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
-      }
-    }
-  }
-}
-
-void test_cxx11_tensor_hip()
-{
-  CALL_SUBTEST(test_hip_contraction<ColMajor>(128, 128, 128));
-  CALL_SUBTEST(test_hip_contraction<RowMajor>(128, 128, 128));
-
-  CALL_SUBTEST(test_scalar<ColMajor>(128, 128, 128));
-  CALL_SUBTEST(test_scalar<RowMajor>(128, 128, 128));
-
-  CALL_SUBTEST(test_hip_contraction_m<ColMajor>());
-  CALL_SUBTEST(test_hip_contraction_m<RowMajor>());
-
-  CALL_SUBTEST(test_hip_contraction_k<ColMajor>());
-  CALL_SUBTEST(test_hip_contraction_k<RowMajor>());
-
-  CALL_SUBTEST(test_hip_contraction_n<ColMajor>());
-  CALL_SUBTEST(test_hip_contraction_n<RowMajor>());
-
-  // Commenting out these tests due to long runtimes
-  // CALL_SUBTEST(test_hip_contraction_sizes<ColMajor>());
-  // CALL_SUBTEST(test_hip_contraction_sizes<RowMajor>());
-}
diff --git a/unsupported/test/cxx11_tensor_device_hip.cu b/unsupported/test/cxx11_tensor_device_hip.cu
deleted file mode 100644
index b98c481ff..000000000
--- a/unsupported/test/cxx11_tensor_device_hip.cu
+++ /dev/null
@@ -1,389 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_device
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-using Eigen::Tensor;
-using Eigen::RowMajor;
-
-// Context for evaluation on cpu
-struct CPUContext {
-  CPUContext(const Eigen::Tensor<float, 3>& in1, Eigen::Tensor<float, 3>& in2, Eigen::Tensor<float, 3>& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(2,2), kernel_3d_(2,2,2) {
-    kernel_1d_(0) = 3.14f;
-    kernel_1d_(1) = 2.7f;
-
-    kernel_2d_(0,0) = 3.14f;
-    kernel_2d_(1,0) = 2.7f;
-    kernel_2d_(0,1) = 0.2f;
-    kernel_2d_(1,1) = 7.0f;
-
-    kernel_3d_(0,0,0) = 3.14f;
-    kernel_3d_(0,1,0) = 2.7f;
-    kernel_3d_(0,0,1) = 0.2f;
-    kernel_3d_(0,1,1) = 7.0f;
-    kernel_3d_(1,0,0) = -1.0f;
-    kernel_3d_(1,1,0) = -0.3f;
-    kernel_3d_(1,0,1) = -0.7f;
-    kernel_3d_(1,1,1) = -0.5f;
-  }
-
-  const Eigen::DefaultDevice& device() const { return cpu_device_; }
-
-  const Eigen::Tensor<float, 3>& in1() const { return in1_; }
-  const Eigen::Tensor<float, 3>& in2() const { return in2_; }
-  Eigen::Tensor<float, 3>& out() { return out_; }
-  const Eigen::Tensor<float, 1>& kernel1d() const { return kernel_1d_; }
-  const Eigen::Tensor<float, 2>& kernel2d() const { return kernel_2d_; }
-  const Eigen::Tensor<float, 3>& kernel3d() const { return kernel_3d_; }
-
- private:
-  const Eigen::Tensor<float, 3>& in1_;
-  const Eigen::Tensor<float, 3>& in2_;
-  Eigen::Tensor<float, 3>& out_;
-
-  Eigen::Tensor<float, 1> kernel_1d_;
-  Eigen::Tensor<float, 2> kernel_2d_;
-  Eigen::Tensor<float, 3> kernel_3d_;
-
-  Eigen::DefaultDevice cpu_device_;
-};
-
-
-// Context for evaluation on GPU
-struct GPUContext {
-  GPUContext(const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1, Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2, Eigen::TensorMap<Eigen::Tensor<float, 3> >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) {
-    assert(hipMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == hipSuccess);
-    float kernel_1d_val[] = {3.14f, 2.7f};
-    assert(hipMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
-
-    assert(hipMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == hipSuccess);
-    float kernel_2d_val[] = {3.14f, 2.7f, 0.2f, 7.0f};
-    assert(hipMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
-
-    assert(hipMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == hipSuccess);
-    float kernel_3d_val[] = {3.14f, -1.0f, 2.7f, -0.3f, 0.2f, -0.7f, 7.0f, -0.5f};
-    assert(hipMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), hipMemcpyHostToDevice) == hipSuccess);
-  }
-  ~GPUContext() {
-    assert(hipFree(kernel_1d_) == hipSuccess);
-    assert(hipFree(kernel_2d_) == hipSuccess);
-    assert(hipFree(kernel_3d_) == hipSuccess);
-  }
-
-  const Eigen::GpuDevice& device() const { return gpu_device_; }
-
-  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1() const { return in1_; }
-  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2() const { return in2_; }
-  Eigen::TensorMap<Eigen::Tensor<float, 3> >& out() { return out_; }
-  Eigen::TensorMap<Eigen::Tensor<float, 1> > kernel1d() const { return Eigen::TensorMap<Eigen::Tensor<float, 1> >(kernel_1d_, 2); }
-  Eigen::TensorMap<Eigen::Tensor<float, 2> > kernel2d() const { return Eigen::TensorMap<Eigen::Tensor<float, 2> >(kernel_2d_, 2, 2); }
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > kernel3d() const { return Eigen::TensorMap<Eigen::Tensor<float, 3> >(kernel_3d_, 2, 2, 2); }
-
- private:
-  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1_;
-  const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2_;
-  Eigen::TensorMap<Eigen::Tensor<float, 3> >& out_;
-
-  float* kernel_1d_;
-  float* kernel_2d_;
-  float* kernel_3d_;
-
-  Eigen::HipStreamDevice stream_;
-  Eigen::GpuDevice gpu_device_;
-};
-
-
-// The actual expression to evaluate
-template <typename Context>
-void test_contextual_eval(Context* context)
-{
-  context->out().device(context->device()) = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f);
-}
-
-template <typename Context>
-void test_forced_contextual_eval(Context* context)
-{
-  context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f);
-}
-
-template <typename Context>
-void test_compound_assignment(Context* context)
-{
-  context->out().device(context->device()) = context->in1().constant(2.718f);
-  context->out().device(context->device()) += context->in1() + context->in2() * 3.14f;
-}
-
-
-template <typename Context>
-void test_contraction(Context* context)
-{
-  Eigen::array<std::pair<int, int>, 2> dims;
-  dims[0] = std::make_pair(1, 1);
-  dims[1] = std::make_pair(2, 2);
-
-  Eigen::array<int, 2> shape(40, 50*70);
-
-  Eigen::DSizes<int, 2> indices(0,0);
-  Eigen::DSizes<int, 2> sizes(40,40);
-
-  context->out().reshape(shape).slice(indices, sizes).device(context->device()) = context->in1().contract(context->in2(), dims);
-}
-
-
-template <typename Context>
-void test_1d_convolution(Context* context)
-{
-  Eigen::DSizes<int, 3> indices(0,0,0);
-  Eigen::DSizes<int, 3> sizes(40,49,70);
-
-  Eigen::array<int, 1> dims(1);
-  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel1d(), dims);
-}
-
-template <typename Context>
-void test_2d_convolution(Context* context)
-{
-  Eigen::DSizes<int, 3> indices(0,0,0);
-  Eigen::DSizes<int, 3> sizes(40,49,69);
-
-  Eigen::array<int, 2> dims(1,2);
-  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel2d(), dims);
-}
-
-template <typename Context>
-void test_3d_convolution(Context* context)
-{
-  Eigen::DSizes<int, 3> indices(0,0,0);
-  Eigen::DSizes<int, 3> sizes(39,49,69);
-
-  Eigen::array<int, 3> dims(0,1,2);
-  context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel3d(), dims);
-}
-
-
-void test_cpu() {
-  Eigen::Tensor<float, 3> in1(40,50,70);
-  Eigen::Tensor<float, 3> in2(40,50,70);
-  Eigen::Tensor<float, 3> out(40,50,70);
-
-  in1 = in1.random() + in1.constant(10.0f);
-  in2 = in2.random() + in2.constant(10.0f);
-
-  CPUContext context(in1, in2, out);
-  test_contextual_eval(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_forced_contextual_eval(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_compound_assignment(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_contraction(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 40; ++j) {
-      const float result = out(i,j,0);
-      float expected = 0;
-      for (int k = 0; k < 50; ++k) {
-        for (int l = 0; l < 70; ++l) {
-          expected += in1(i, k, l) * in2(j, k, l);
-        }
-      }
-      VERIFY_IS_APPROX(expected, result);
-    }
-  }
-
-  test_1d_convolution(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
-      }
-    }
-  }
-
-  test_2d_convolution(&context);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 69; ++k) {
-        const float result = out(i,j,k);
-        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f) +
-                               (in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
-        if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
-          continue;
-        }
-        VERIFY_IS_APPROX(expected, result);
-      }
-    }
-  }
-
-  test_3d_convolution(&context);
-  for (int i = 0; i < 39; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 69; ++k) {
-        const float result = out(i,j,k);
-        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
-                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f) +
-                               (in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
-                                in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
-        if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
-          continue;
-        }
-        VERIFY_IS_APPROX(expected, result);
-      }
-    }
-  }
-}
-
-void test_gpu() {
-  Eigen::Tensor<float, 3> in1(40,50,70);
-  Eigen::Tensor<float, 3> in2(40,50,70);
-  Eigen::Tensor<float, 3> out(40,50,70);
-  in1 = in1.random() + in1.constant(10.0f);
-  in2 = in2.random() + in2.constant(10.0f);
-
-  std::size_t in1_bytes = in1.size() * sizeof(float);
-  std::size_t in2_bytes = in2.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_in1;
-  float* d_in2;
-  float* d_out;
-  hipMalloc((void**)(&d_in1), in1_bytes);
-  hipMalloc((void**)(&d_in2), in2_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, 40,50,70);
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, 40,50,70);
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, 40,50,70);
-
-  GPUContext context(gpu_in1, gpu_in2, gpu_out);
-  test_contextual_eval(&context);
-  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_forced_contextual_eval(&context);
-  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_compound_assignment(&context);
-  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 50; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
-      }
-    }
-  }
-
-  test_contraction(&context);
-  assert(hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 40; ++j) {
-      const float result = out(i,j,0);
-      float expected = 0;
-      for (int k = 0; k < 50; ++k) {
-        for (int l = 0; l < 70; ++l) {
-          expected += in1(i, k, l) * in2(j, k, l);
-        }
-      }
-      VERIFY_IS_APPROX(expected, result);
-    }
-  }
-
-  test_1d_convolution(&context);
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
-  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 70; ++k) {
-        VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
-      }
-    }
-  }
-
-  test_2d_convolution(&context);
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
-  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
-  for (int i = 0; i < 40; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 69; ++k) {
-        const float result = out(i,j,k);
-        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
-                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
-        VERIFY_IS_APPROX(expected, result);
-      }
-    }
-  }
-
-  /*
-  test_3d_convolution(&context);
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, context.device().stream()) == hipSuccess);
-  assert(hipStreamSynchronize(context.device().stream()) == hipSuccess);
-  for (int i = 0; i < 39; ++i) {
-    for (int j = 0; j < 49; ++j) {
-      for (int k = 0; k < 69; ++k) {
-       const float result = out(i,j,k);
-        const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
-                                in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f +
-                                in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
-                                in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
-        VERIFY_IS_APPROX(expected, result);
-      }
-    }
-  }
-  */
-}
-
-
-void test_cxx11_tensor_device()
-{
-  CALL_SUBTEST(test_cpu());
-  CALL_SUBTEST(test_gpu());
-}
diff --git a/unsupported/test/cxx11_tensor_hip.cu b/unsupported/test/cxx11_tensor_hip.cu
deleted file mode 100644
index b28840267..000000000
--- a/unsupported/test/cxx11_tensor_hip.cu
+++ /dev/null
@@ -1,1295 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_hip
-#define EIGEN_USE_GPU
-
-#ifdef __NVCC__
-#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
-#include <cuda_fp16.h>
-#endif
-#endif
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-using Eigen::Tensor;
-
-void test_hip_nullary() {
-  Tensor<float, 1, 0, int> in1(2);
-  Tensor<float, 1, 0, int> in2(2);
-  in1.setRandom();
-  in2.setRandom();
-
-  std::size_t tensor_bytes = in1.size() * sizeof(float);
-
-  float* d_in1;
-  float* d_in2;
-  hipMalloc((void**)(&d_in1), tensor_bytes);
-  hipMalloc((void**)(&d_in2), tensor_bytes);
-  hipMemcpy(d_in1, in1.data(), tensor_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in2, in2.data(), tensor_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in1(
-      d_in1, 2);
-  Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in2(
-      d_in2, 2);
-
-  gpu_in1.device(gpu_device) = gpu_in1.constant(3.14f);
-  gpu_in2.device(gpu_device) = gpu_in2.random();
-
-  Tensor<float, 1, 0, int> new1(2);
-  Tensor<float, 1, 0, int> new2(2);
-
-  assert(hipMemcpyAsync(new1.data(), d_in1, tensor_bytes, hipMemcpyDeviceToHost,
-                         gpu_device.stream()) == hipSuccess);
-  assert(hipMemcpyAsync(new2.data(), d_in2, tensor_bytes, hipMemcpyDeviceToHost,
-                         gpu_device.stream()) == hipSuccess);
-
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 2; ++i) {
-    VERIFY_IS_APPROX(new1(i), 3.14f);
-    VERIFY_IS_NOT_EQUAL(new2(i), in2(i));
-  }
-
-  hipFree(d_in1);
-  hipFree(d_in2);
-}
-
-void test_hip_elementwise_small() {
-  Tensor<float, 1> in1(Eigen::array<Eigen::DenseIndex, 1>{2});
-  Tensor<float, 1> in2(Eigen::array<Eigen::DenseIndex, 1>{2});
-  Tensor<float, 1> out(Eigen::array<Eigen::DenseIndex, 1>{2});
-  in1.setRandom();
-  in2.setRandom();
-
-  std::size_t in1_bytes = in1.size() * sizeof(float);
-  std::size_t in2_bytes = in2.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_in1;
-  float* d_in2;
-  float* d_out;
-  hipMalloc((void**)(&d_in1), in1_bytes);
-  hipMalloc((void**)(&d_in2), in2_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
-      d_in1, Eigen::array<Eigen::DenseIndex, 1>{2});
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in2(
-      d_in2, Eigen::array<Eigen::DenseIndex, 1>{2});
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_out(
-      d_out, Eigen::array<Eigen::DenseIndex, 1>{2});
-
-  gpu_out.device(gpu_device) = gpu_in1 + gpu_in2;
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost,
-                         gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 2; ++i) {
-    VERIFY_IS_APPROX(
-        out(Eigen::array<Eigen::DenseIndex, 1>{i}),
-        in1(Eigen::array<Eigen::DenseIndex, 1>{i}) + in2(Eigen::array<Eigen::DenseIndex, 1>{i}));
-  }
-
-  hipFree(d_in1);
-  hipFree(d_in2);
-  hipFree(d_out);
-}
-
-void test_hip_elementwise()
-{
-  Tensor<float, 3> in1(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Tensor<float, 3> in2(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Tensor<float, 3> in3(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Tensor<float, 3> out(Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  in1.setRandom();
-  in2.setRandom();
-  in3.setRandom();
-
-  std::size_t in1_bytes = in1.size() * sizeof(float);
-  std::size_t in2_bytes = in2.size() * sizeof(float);
-  std::size_t in3_bytes = in3.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_in1;
-  float* d_in2;
-  float* d_in3;
-  float* d_out;
-  hipMalloc((void**)(&d_in1), in1_bytes);
-  hipMalloc((void**)(&d_in2), in2_bytes);
-  hipMalloc((void**)(&d_in3), in3_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in2, in2.data(), in2_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in3, in3.data(), in3_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in3(d_in3, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-  Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, Eigen::array<Eigen::DenseIndex, 3>{72,53,97});
-
-  gpu_out.device(gpu_device) = gpu_in1 + gpu_in2 * gpu_in3;
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 72; ++i) {
-    for (int j = 0; j < 53; ++j) {
-      for (int k = 0; k < 97; ++k) {
-        VERIFY_IS_APPROX(out(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}), in1(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}) + in2(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}) * in3(Eigen::array<Eigen::DenseIndex, 3>{i,j,k}));
-      }
-    }
-  }
-
-  hipFree(d_in1);
-  hipFree(d_in2);
-  hipFree(d_in3);
-  hipFree(d_out);
-}
-
-void test_hip_props() {
-  Tensor<float, 1> in1(200);
-  Tensor<bool, 1> out(200);
-  in1.setRandom();
-
-  std::size_t in1_bytes = in1.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(bool);
-
-  float* d_in1;
-  bool* d_out;
-  hipMalloc((void**)(&d_in1), in1_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
-      d_in1, 200);
-  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_out(
-      d_out, 200);
-
-  gpu_out.device(gpu_device) = (gpu_in1.isnan)();
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost,
-                         gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 200; ++i) {
-    VERIFY_IS_EQUAL(out(i), (std::isnan)(in1(i)));
-  }
-
-  hipFree(d_in1);
-  hipFree(d_out);
-}
-
-void test_hip_reduction()
-{
-  Tensor<float, 4> in1(72,53,97,113);
-  Tensor<float, 2> out(72,97);
-  in1.setRandom();
-
-  std::size_t in1_bytes = in1.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_in1;
-  float* d_out;
-  hipMalloc((void**)(&d_in1), in1_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_in1, in1.data(), in1_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4> > gpu_in1(d_in1, 72,53,97,113);
-  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
-
-  array<Eigen::DenseIndex, 2> reduction_axis;
-  reduction_axis[0] = 1;
-  reduction_axis[1] = 3;
-
-  gpu_out.device(gpu_device) = gpu_in1.maximum(reduction_axis);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 72; ++i) {
-    for (int j = 0; j < 97; ++j) {
-      float expected = 0;
-      for (int k = 0; k < 53; ++k) {
-        for (int l = 0; l < 113; ++l) {
-          expected =
-              std::max<float>(expected, in1(i, k, j, l));
-        }
-      }
-      VERIFY_IS_APPROX(out(i,j), expected);
-    }
-  }
-
-  hipFree(d_in1);
-  hipFree(d_out);
-}
-
-template<int DataLayout>
-void test_hip_contraction()
-{
-  // with these dimensions, the output has 300 * 140 elements, which is
-  // more than 30 * 1024, which is the number of threads in blocks on
-  // a 15 SM GK110 GPU
-  Tensor<float, 4, DataLayout> t_left(6, 50, 3, 31);
-  Tensor<float, 5, DataLayout> t_right(Eigen::array<Eigen::DenseIndex, 5>{3, 31, 7, 20, 1});
-  Tensor<float, 5, DataLayout> t_result(Eigen::array<Eigen::DenseIndex, 5>{6, 50, 7, 20, 1});
-
-  t_left.setRandom();
-  t_right.setRandom();
-
-  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
-  std::size_t t_right_bytes = t_right.size() * sizeof(float);
-  std::size_t t_result_bytes = t_result.size() * sizeof(float);
-
-  float* d_t_left;
-  float* d_t_right;
-  float* d_t_result;
-
-  hipMalloc((void**)(&d_t_left), t_left_bytes);
-  hipMalloc((void**)(&d_t_right), t_right_bytes);
-  hipMalloc((void**)(&d_t_result), t_result_bytes);
-
-  hipMemcpy(d_t_left, t_left.data(), t_left_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_t_right, t_right.data(), t_right_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_t_left(d_t_left, 6, 50, 3, 31);
-  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_t_right(d_t_right, 3, 31, 7, 20, 1);
-  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_t_result(d_t_result, 6, 50, 7, 20, 1);
-
-  typedef Eigen::Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> > MapXf;
-  MapXf m_left(t_left.data(), 300, 93);
-  MapXf m_right(t_right.data(), 93, 140);
-  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(300, 140);
-
-  typedef Tensor<float, 1>::DimensionPair DimPair;
-  Eigen::array<DimPair, 2> dims;
-  dims[0] = DimPair(2, 0);
-  dims[1] = DimPair(3, 1);
-
-  m_result = m_left * m_right;
-  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
-
-  hipMemcpy(t_result.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
-
-  for (DenseIndex i = 0; i < t_result.size(); i++) {
-    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4f) {
-      std::cout << "mismatch detected at index " << i << ": " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
-      assert(false);
-    }
-  }
-
-  hipFree(d_t_left);
-  hipFree(d_t_right);
-  hipFree(d_t_result);
-}
-
-
-template<int DataLayout>
-void test_hip_convolution_1d()
-{
-  Tensor<float, 4, DataLayout> input(74,37,11,137);
-  Tensor<float, 1, DataLayout> kernel(4);
-  Tensor<float, 4, DataLayout> out(74,34,11,137);
-  input = input.constant(10.0f) + input.random();
-  kernel = kernel.constant(7.0f) + kernel.random();
-
-  std::size_t input_bytes = input.size() * sizeof(float);
-  std::size_t kernel_bytes = kernel.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_input;
-  float* d_kernel;
-  float* d_out;
-  hipMalloc((void**)(&d_input), input_bytes);
-  hipMalloc((void**)(&d_kernel), kernel_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input, 74,37,11,137);
-  Eigen::TensorMap<Eigen::Tensor<float, 1, DataLayout> > gpu_kernel(d_kernel, 4);
-  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out, 74,34,11,137);
-
-  Eigen::array<Eigen::DenseIndex, 1> dims{1};
-  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 74; ++i) {
-    for (int j = 0; j < 34; ++j) {
-      for (int k = 0; k < 11; ++k) {
-        for (int l = 0; l < 137; ++l) {
-          const float result = out(i,j,k,l);
-          const float expected = input(i,j+0,k,l) * kernel(0) + input(i,j+1,k,l) * kernel(1) +
-                                 input(i,j+2,k,l) * kernel(2) + input(i,j+3,k,l) * kernel(3);
-          VERIFY_IS_APPROX(result, expected);
-        }
-      }
-    }
-  }
-
-  hipFree(d_input);
-  hipFree(d_kernel);
-  hipFree(d_out);
-}
-
-void test_hip_convolution_inner_dim_col_major_1d()
-{
-  Tensor<float, 4, ColMajor> input(74,9,11,7);
-  Tensor<float, 1, ColMajor> kernel(4);
-  Tensor<float, 4, ColMajor> out(71,9,11,7);
-  input = input.constant(10.0f) + input.random();
-  kernel = kernel.constant(7.0f) + kernel.random();
-
-  std::size_t input_bytes = input.size() * sizeof(float);
-  std::size_t kernel_bytes = kernel.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_input;
-  float* d_kernel;
-  float* d_out;
-  hipMalloc((void**)(&d_input), input_bytes);
-  hipMalloc((void**)(&d_kernel), kernel_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_input(d_input,74,9,11,7);
-  Eigen::TensorMap<Eigen::Tensor<float, 1, ColMajor> > gpu_kernel(d_kernel,4);
-  Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_out(d_out,71,9,11,7);
-
-  Eigen::array<Eigen::DenseIndex, 1> dims{0};
-  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 71; ++i) {
-    for (int j = 0; j < 9; ++j) {
-      for (int k = 0; k < 11; ++k) {
-        for (int l = 0; l < 7; ++l) {
-          const float result = out(i,j,k,l);
-          const float expected = input(i+0,j,k,l) * kernel(0) + input(i+1,j,k,l) * kernel(1) +
-                                 input(i+2,j,k,l) * kernel(2) + input(i+3,j,k,l) * kernel(3);
-          VERIFY_IS_APPROX(result, expected);
-        }
-      }
-    }
-  }
-
-  hipFree(d_input);
-  hipFree(d_kernel);
-  hipFree(d_out);
-}
-
-void test_hip_convolution_inner_dim_row_major_1d()
-{
-  Tensor<float, 4, RowMajor> input(7,9,11,74);
-  Tensor<float, 1, RowMajor> kernel(4);
-  Tensor<float, 4, RowMajor> out(7,9,11,71);
-  input = input.constant(10.0f) + input.random();
-  kernel = kernel.constant(7.0f) + kernel.random();
-
-  std::size_t input_bytes = input.size() * sizeof(float);
-  std::size_t kernel_bytes = kernel.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_input;
-  float* d_kernel;
-  float* d_out;
-  hipMalloc((void**)(&d_input), input_bytes);
-  hipMalloc((void**)(&d_kernel), kernel_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_input(d_input, 7,9,11,74);
-  Eigen::TensorMap<Eigen::Tensor<float, 1, RowMajor> > gpu_kernel(d_kernel, 4);
-  Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_out(d_out, 7,9,11,71);
-
-  Eigen::array<Eigen::DenseIndex, 1> dims{3};
-  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 7; ++i) {
-    for (int j = 0; j < 9; ++j) {
-      for (int k = 0; k < 11; ++k) {
-        for (int l = 0; l < 71; ++l) {
-          const float result = out(i,j,k,l);
-          const float expected = input(i,j,k,l+0) * kernel(0) + input(i,j,k,l+1) * kernel(1) +
-                                 input(i,j,k,l+2) * kernel(2) + input(i,j,k,l+3) * kernel(3);
-          VERIFY_IS_APPROX(result, expected);
-        }
-      }
-    }
-  }
-
-  hipFree(d_input);
-  hipFree(d_kernel);
-  hipFree(d_out);
-}
-
-template<int DataLayout>
-void test_hip_convolution_2d()
-{
-  Tensor<float, 4, DataLayout> input(74,37,11,137);
-  Tensor<float, 2, DataLayout> kernel(3,4);
-  Tensor<float, 4, DataLayout> out(74,35,8,137);
-  input = input.constant(10.0f) + input.random();
-  kernel = kernel.constant(7.0f) + kernel.random();
-
-  std::size_t input_bytes = input.size() * sizeof(float);
-  std::size_t kernel_bytes = kernel.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_input;
-  float* d_kernel;
-  float* d_out;
-  hipMalloc((void**)(&d_input), input_bytes);
-  hipMalloc((void**)(&d_kernel), kernel_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input,74,37,11,137);
-  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> > gpu_kernel(d_kernel,3,4);
-  Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out,74,35,8,137);
-
-  Eigen::array<Eigen::DenseIndex, 2> dims{1,2};
-  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 74; ++i) {
-    for (int j = 0; j < 35; ++j) {
-      for (int k = 0; k < 8; ++k) {
-        for (int l = 0; l < 137; ++l) {
-          const float result = out(i,j,k,l);
-          const float expected = input(i,j+0,k+0,l) * kernel(0,0) +
-                                 input(i,j+1,k+0,l) * kernel(1,0) +
-                                 input(i,j+2,k+0,l) * kernel(2,0) +
-                                 input(i,j+0,k+1,l) * kernel(0,1) +
-                                 input(i,j+1,k+1,l) * kernel(1,1) +
-                                 input(i,j+2,k+1,l) * kernel(2,1) +
-                                 input(i,j+0,k+2,l) * kernel(0,2) +
-                                 input(i,j+1,k+2,l) * kernel(1,2) +
-                                 input(i,j+2,k+2,l) * kernel(2,2) +
-                                 input(i,j+0,k+3,l) * kernel(0,3) +
-                                 input(i,j+1,k+3,l) * kernel(1,3) +
-                                 input(i,j+2,k+3,l) * kernel(2,3);
-          VERIFY_IS_APPROX(result, expected);
-        }
-      }
-    }
-  }
-
-  hipFree(d_input);
-  hipFree(d_kernel);
-  hipFree(d_out);
-}
-
-template<int DataLayout>
-void test_hip_convolution_3d()
-{
-  Tensor<float, 5, DataLayout> input(Eigen::array<Eigen::DenseIndex, 5>{74,37,11,137,17});
-  Tensor<float, 3, DataLayout> kernel(3,4,2);
-  Tensor<float, 5, DataLayout> out(Eigen::array<Eigen::DenseIndex, 5>{74,35,8,136,17});
-  input = input.constant(10.0f) + input.random();
-  kernel = kernel.constant(7.0f) + kernel.random();
-
-  std::size_t input_bytes = input.size() * sizeof(float);
-  std::size_t kernel_bytes = kernel.size() * sizeof(float);
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_input;
-  float* d_kernel;
-  float* d_out;
-  hipMalloc((void**)(&d_input), input_bytes);
-  hipMalloc((void**)(&d_kernel), kernel_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  hipMemcpy(d_input, input.data(), input_bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_kernel, kernel.data(), kernel_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;    
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_input(d_input,74,37,11,137,17);
-  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> > gpu_kernel(d_kernel,3,4,2);
-  Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_out(d_out,74,35,8,136,17);
-
-  Eigen::array<Eigen::DenseIndex, 3> dims{1,2,3};
-  gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 74; ++i) {
-    for (int j = 0; j < 35; ++j) {
-      for (int k = 0; k < 8; ++k) {
-        for (int l = 0; l < 136; ++l) {
-          for (int m = 0; m < 17; ++m) {
-            const float result = out(i,j,k,l,m);
-            const float expected = input(i,j+0,k+0,l+0,m) * kernel(0,0,0) +
-                                   input(i,j+1,k+0,l+0,m) * kernel(1,0,0) +
-                                   input(i,j+2,k+0,l+0,m) * kernel(2,0,0) +
-                                   input(i,j+0,k+1,l+0,m) * kernel(0,1,0) +
-                                   input(i,j+1,k+1,l+0,m) * kernel(1,1,0) +
-                                   input(i,j+2,k+1,l+0,m) * kernel(2,1,0) +
-                                   input(i,j+0,k+2,l+0,m) * kernel(0,2,0) +
-                                   input(i,j+1,k+2,l+0,m) * kernel(1,2,0) +
-                                   input(i,j+2,k+2,l+0,m) * kernel(2,2,0) +
-                                   input(i,j+0,k+3,l+0,m) * kernel(0,3,0) +
-                                   input(i,j+1,k+3,l+0,m) * kernel(1,3,0) +
-                                   input(i,j+2,k+3,l+0,m) * kernel(2,3,0) +
-                                   input(i,j+0,k+0,l+1,m) * kernel(0,0,1) +
-                                   input(i,j+1,k+0,l+1,m) * kernel(1,0,1) +
-                                   input(i,j+2,k+0,l+1,m) * kernel(2,0,1) +
-                                   input(i,j+0,k+1,l+1,m) * kernel(0,1,1) +
-                                   input(i,j+1,k+1,l+1,m) * kernel(1,1,1) +
-                                   input(i,j+2,k+1,l+1,m) * kernel(2,1,1) +
-                                   input(i,j+0,k+2,l+1,m) * kernel(0,2,1) +
-                                   input(i,j+1,k+2,l+1,m) * kernel(1,2,1) +
-                                   input(i,j+2,k+2,l+1,m) * kernel(2,2,1) +
-                                   input(i,j+0,k+3,l+1,m) * kernel(0,3,1) +
-                                   input(i,j+1,k+3,l+1,m) * kernel(1,3,1) +
-                                   input(i,j+2,k+3,l+1,m) * kernel(2,3,1);
-            VERIFY_IS_APPROX(result, expected);
-          }
-        }
-      }
-    }
-  }
-
-  hipFree(d_input);
-  hipFree(d_kernel);
-  hipFree(d_out);
-}
-
-
-template <typename Scalar>
-void test_hip_lgamma(const Scalar stddev)
-{
-  Tensor<Scalar, 2> in(72,97);
-  in.setRandom();
-  in *= in.constant(stddev);
-  Tensor<Scalar, 2> out(72,97);
-  out.setZero();
-
-  std::size_t bytes = in.size() * sizeof(Scalar);
-
-  Scalar* d_in;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
-
-  gpu_out.device(gpu_device) = gpu_in.lgamma();
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 72; ++i) {
-    for (int j = 0; j < 97; ++j) {
-      VERIFY_IS_APPROX(out(i,j), (std::lgamma)(in(i,j)));
-    }
-  }
-
-  hipFree(d_in);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_digamma()
-{
-  Tensor<Scalar, 1> in(7);
-  Tensor<Scalar, 1> out(7);
-  Tensor<Scalar, 1> expected_out(7);
-  out.setZero();
-
-  in(0) = Scalar(1);
-  in(1) = Scalar(1.5);
-  in(2) = Scalar(4);
-  in(3) = Scalar(-10.5);
-  in(4) = Scalar(10000.5);
-  in(5) = Scalar(0);
-  in(6) = Scalar(-1);
-
-  expected_out(0) = Scalar(-0.5772156649015329);
-  expected_out(1) = Scalar(0.03648997397857645);
-  expected_out(2) = Scalar(1.2561176684318);
-  expected_out(3) = Scalar(2.398239129535781);
-  expected_out(4) = Scalar(9.210340372392849);
-  expected_out(5) = std::numeric_limits<Scalar>::infinity();
-  expected_out(6) = std::numeric_limits<Scalar>::infinity();
-
-  std::size_t bytes = in.size() * sizeof(Scalar);
-
-  Scalar* d_in;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 7);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 7);
-
-  gpu_out.device(gpu_device) = gpu_in.digamma();
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 5; ++i) {
-    VERIFY_IS_APPROX(out(i), expected_out(i));
-  }
-  for (int i = 5; i < 7; ++i) {
-    VERIFY_IS_EQUAL(out(i), expected_out(i));
-  }
-
-  hipFree(d_in);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_zeta()
-{
-  Tensor<Scalar, 1> in_x(6);
-  Tensor<Scalar, 1> in_q(6);
-  Tensor<Scalar, 1> out(6);
-  Tensor<Scalar, 1> expected_out(6);
-  out.setZero();
-
-  in_x(0) = Scalar(1);
-  in_x(1) = Scalar(1.5);
-  in_x(2) = Scalar(4);
-  in_x(3) = Scalar(-10.5);
-  in_x(4) = Scalar(10000.5);
-  in_x(5) = Scalar(3);
-  
-  in_q(0) = Scalar(1.2345);
-  in_q(1) = Scalar(2);
-  in_q(2) = Scalar(1.5);
-  in_q(3) = Scalar(3);
-  in_q(4) = Scalar(1.0001);
-  in_q(5) = Scalar(-2.5);
-
-  expected_out(0) = std::numeric_limits<Scalar>::infinity();
-  expected_out(1) = Scalar(1.61237534869);
-  expected_out(2) = Scalar(0.234848505667);
-  expected_out(3) = Scalar(1.03086757337e-5);
-  expected_out(4) = Scalar(0.367879440865);
-  expected_out(5) = Scalar(0.054102025820864097);
-
-  std::size_t bytes = in_x.size() * sizeof(Scalar);
-
-  Scalar* d_in_x;
-  Scalar* d_in_q;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in_x), bytes);
-  hipMalloc((void**)(&d_in_q), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in_q, in_q.data(), bytes, hipMemcpyHostToDevice);
-  
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_q(d_in_q, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 6);
-
-  gpu_out.device(gpu_device) = gpu_in_x.zeta(gpu_in_q);
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  VERIFY_IS_EQUAL(out(0), expected_out(0));
-  VERIFY((std::isnan)(out(3)));
-
-  for (int i = 1; i < 6; ++i) {
-    if (i != 3) {
-      VERIFY_IS_APPROX(out(i), expected_out(i));
-    }
-  }
-
-  hipFree(d_in_x);
-  hipFree(d_in_q);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_polygamma()
-{
-  Tensor<Scalar, 1> in_x(7);
-  Tensor<Scalar, 1> in_n(7);
-  Tensor<Scalar, 1> out(7);
-  Tensor<Scalar, 1> expected_out(7);
-  out.setZero();
-
-  in_n(0) = Scalar(1);
-  in_n(1) = Scalar(1);
-  in_n(2) = Scalar(1);
-  in_n(3) = Scalar(17);
-  in_n(4) = Scalar(31);
-  in_n(5) = Scalar(28);
-  in_n(6) = Scalar(8);
-  
-  in_x(0) = Scalar(2);
-  in_x(1) = Scalar(3);
-  in_x(2) = Scalar(25.5);
-  in_x(3) = Scalar(4.7);
-  in_x(4) = Scalar(11.8);
-  in_x(5) = Scalar(17.7);
-  in_x(6) = Scalar(30.2);
-
-  expected_out(0) = Scalar(0.644934066848);
-  expected_out(1) = Scalar(0.394934066848);
-  expected_out(2) = Scalar(0.0399946696496);
-  expected_out(3) = Scalar(293.334565435);
-  expected_out(4) = Scalar(0.445487887616);
-  expected_out(5) = Scalar(-2.47810300902e-07);
-  expected_out(6) = Scalar(-8.29668781082e-09);
-
-  std::size_t bytes = in_x.size() * sizeof(Scalar);
-
-  Scalar* d_in_x;
-  Scalar* d_in_n;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in_x), bytes);
-  hipMalloc((void**)(&d_in_n), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in_n, in_n.data(), bytes, hipMemcpyHostToDevice);
-  
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 7);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_n(d_in_n, 7);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 7);
-
-  gpu_out.device(gpu_device) = gpu_in_n.polygamma(gpu_in_x);
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 7; ++i) {
-    VERIFY_IS_APPROX(out(i), expected_out(i));
-  }
-
-  hipFree(d_in_x);
-  hipFree(d_in_n);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_igamma()
-{
-  Tensor<Scalar, 2> a(6, 6);
-  Tensor<Scalar, 2> x(6, 6);
-  Tensor<Scalar, 2> out(6, 6);
-  out.setZero();
-
-  Scalar a_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
-  Scalar x_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
-
-  for (int i = 0; i < 6; ++i) {
-    for (int j = 0; j < 6; ++j) {
-      a(i, j) = a_s[i];
-      x(i, j) = x_s[j];
-    }
-  }
-
-  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
-  Scalar igamma_s[][6] = {{0.0, nan, nan, nan, nan, nan},
-                          {0.0, 0.6321205588285578, 0.7768698398515702,
-                           0.9816843611112658, 9.999500016666262e-05, 1.0},
-                          {0.0, 0.4275932955291202, 0.608374823728911,
-                           0.9539882943107686, 7.522076445089201e-07, 1.0},
-                          {0.0, 0.01898815687615381, 0.06564245437845008,
-                           0.5665298796332909, 4.166333347221828e-18, 1.0},
-                          {0.0, 0.9999780593618628, 0.9999899967080838,
-                           0.9999996219837988, 0.9991370418689945, 1.0},
-                          {0.0, 0.0, 0.0, 0.0, 0.0, 0.5042041932513908}};
-
-
-
-  std::size_t bytes = a.size() * sizeof(Scalar);
-
-  Scalar* d_a;
-  Scalar* d_x;
-  Scalar* d_out;
-  assert(hipMalloc((void**)(&d_a), bytes) == hipSuccess);
-  assert(hipMalloc((void**)(&d_x), bytes) == hipSuccess);
-  assert(hipMalloc((void**)(&d_out), bytes) == hipSuccess);
-
-  hipMemcpy(d_a, a.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_x, x.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_x(d_x, 6, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 6, 6);
-
-  gpu_out.device(gpu_device) = gpu_a.igamma(gpu_x);
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 6; ++i) {
-    for (int j = 0; j < 6; ++j) {
-      if ((std::isnan)(igamma_s[i][j])) {
-        VERIFY((std::isnan)(out(i, j)));
-      } else {
-        VERIFY_IS_APPROX(out(i, j), igamma_s[i][j]);
-      }
-    }
-  }
-
-  hipFree(d_a);
-  hipFree(d_x);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_igammac()
-{
-  Tensor<Scalar, 2> a(6, 6);
-  Tensor<Scalar, 2> x(6, 6);
-  Tensor<Scalar, 2> out(6, 6);
-  out.setZero();
-
-  Scalar a_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
-  Scalar x_s[] = {Scalar(0), Scalar(1), Scalar(1.5), Scalar(4), Scalar(0.0001), Scalar(1000.5)};
-
-  for (int i = 0; i < 6; ++i) {
-    for (int j = 0; j < 6; ++j) {
-      a(i, j) = a_s[i];
-      x(i, j) = x_s[j];
-    }
-  }
-
-  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
-  Scalar igammac_s[][6] = {{nan, nan, nan, nan, nan, nan},
-                           {1.0, 0.36787944117144233, 0.22313016014842982,
-                            0.018315638888734182, 0.9999000049998333, 0.0},
-                           {1.0, 0.5724067044708798, 0.3916251762710878,
-                            0.04601170568923136, 0.9999992477923555, 0.0},
-                           {1.0, 0.9810118431238462, 0.9343575456215499,
-                            0.4334701203667089, 1.0, 0.0},
-                           {1.0, 2.1940638138146658e-05, 1.0003291916285e-05,
-                            3.7801620118431334e-07, 0.0008629581310054535,
-                            0.0},
-                           {1.0, 1.0, 1.0, 1.0, 1.0, 0.49579580674813944}};
-
-  std::size_t bytes = a.size() * sizeof(Scalar);
-
-  Scalar* d_a;
-  Scalar* d_x;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_a), bytes);
-  hipMalloc((void**)(&d_x), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_a, a.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_x, x.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_x(d_x, 6, 6);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 6, 6);
-
-  gpu_out.device(gpu_device) = gpu_a.igammac(gpu_x);
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 6; ++i) {
-    for (int j = 0; j < 6; ++j) {
-      if ((std::isnan)(igammac_s[i][j])) {
-        VERIFY((std::isnan)(out(i, j)));
-      } else {
-        VERIFY_IS_APPROX(out(i, j), igammac_s[i][j]);
-      }
-    }
-  }
-
-  hipFree(d_a);
-  hipFree(d_x);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_erf(const Scalar stddev)
-{
-  Tensor<Scalar, 2> in(72,97);
-  in.setRandom();
-  in *= in.constant(stddev);
-  Tensor<Scalar, 2> out(72,97);
-  out.setZero();
-
-  std::size_t bytes = in.size() * sizeof(Scalar);
-
-  Scalar* d_in;
-  Scalar* d_out;
-  assert(hipMalloc((void**)(&d_in), bytes) == hipSuccess);
-  assert(hipMalloc((void**)(&d_out), bytes) == hipSuccess);
-
-  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
-
-  gpu_out.device(gpu_device) = gpu_in.erf();
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 72; ++i) {
-    for (int j = 0; j < 97; ++j) {
-      VERIFY_IS_APPROX(out(i,j), (std::erf)(in(i,j)));
-    }
-  }
-
-  hipFree(d_in);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_erfc(const Scalar stddev)
-{
-  Tensor<Scalar, 2> in(72,97);
-  in.setRandom();
-  in *= in.constant(stddev);
-  Tensor<Scalar, 2> out(72,97);
-  out.setZero();
-
-  std::size_t bytes = in.size() * sizeof(Scalar);
-
-  Scalar* d_in;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in, in.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_out(d_out, 72, 97);
-
-  gpu_out.device(gpu_device) = gpu_in.erfc();
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 0; i < 72; ++i) {
-    for (int j = 0; j < 97; ++j) {
-      VERIFY_IS_APPROX(out(i,j), (std::erfc)(in(i,j)));
-    }
-  }
-
-  hipFree(d_in);
-  hipFree(d_out);
-}
-
-template <typename Scalar>
-void test_hip_betainc()
-{
-  Tensor<Scalar, 1> in_x(125);
-  Tensor<Scalar, 1> in_a(125);
-  Tensor<Scalar, 1> in_b(125);
-  Tensor<Scalar, 1> out(125);
-  Tensor<Scalar, 1> expected_out(125);
-  out.setZero();
-
-  Scalar nan = std::numeric_limits<Scalar>::quiet_NaN();
-
-  Array<Scalar, 1, Dynamic> x(125);
-  Array<Scalar, 1, Dynamic> a(125);
-  Array<Scalar, 1, Dynamic> b(125);
-  Array<Scalar, 1, Dynamic> v(125);
-
-  a << 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
-      0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999,
-      0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999,
-      0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
-      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999,
-      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999,
-      999.999, 999.999, 999.999, 999.999, 999.999, 999.999, 999.999;
-
-  b << 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
-      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
-      999.999, 999.999, 999.999, 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 999.999, 999.999, 999.999, 999.999, 999.999, 0.0, 0.0,
-      0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
-      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
-      999.999, 999.999, 999.999, 0.0, 0.0, 0.0, 0.0, 0.0, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.999, 0.999, 0.999, 0.999, 0.999, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 999.999, 999.999, 999.999, 999.999, 999.999, 0.0, 0.0,
-      0.0, 0.0, 0.0, 0.03062277660168379, 0.03062277660168379,
-      0.03062277660168379, 0.03062277660168379, 0.03062277660168379, 0.999,
-      0.999, 0.999, 0.999, 0.999, 31.62177660168379, 31.62177660168379,
-      31.62177660168379, 31.62177660168379, 31.62177660168379, 999.999, 999.999,
-      999.999, 999.999, 999.999;
-
-  x << -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8,
-      1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5,
-      0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2,
-      0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1,
-      0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1,
-      -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8,
-      1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5,
-      0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2,
-      0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1, -0.1, 0.2, 0.5, 0.8, 1.1;
-
-  v << nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
-      nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
-      nan, nan, 0.47972119876364683, 0.5, 0.5202788012363533, nan, nan,
-      0.9518683957740043, 0.9789663010413743, 0.9931729188073435, nan, nan,
-      0.999995949033062, 0.9999999999993698, 0.9999999999999999, nan, nan,
-      0.9999999999999999, 0.9999999999999999, 0.9999999999999999, nan, nan, nan,
-      nan, nan, nan, nan, 0.006827081192655869, 0.0210336989586256,
-      0.04813160422599567, nan, nan, 0.20014344256217678, 0.5000000000000001,
-      0.7998565574378232, nan, nan, 0.9991401428435834, 0.999999999698403,
-      0.9999999999999999, nan, nan, 0.9999999999999999, 0.9999999999999999,
-      0.9999999999999999, nan, nan, nan, nan, nan, nan, nan,
-      1.0646600232370887e-25, 6.301722877826246e-13, 4.050966937974938e-06, nan,
-      nan, 7.864342668429763e-23, 3.015969667594166e-10, 0.0008598571564165444,
-      nan, nan, 6.031987710123844e-08, 0.5000000000000007, 0.9999999396801229,
-      nan, nan, 0.9999999999999999, 0.9999999999999999, 0.9999999999999999, nan,
-      nan, nan, nan, nan, nan, nan, 0.0, 7.029920380986636e-306,
-      2.2450728208591345e-101, nan, nan, 0.0, 9.275871147869727e-302,
-      1.2232913026152827e-97, nan, nan, 0.0, 3.0891393081932924e-252,
-      2.9303043666183996e-60, nan, nan, 2.248913486879199e-196,
-      0.5000000000004947, 0.9999999999999999, nan;
-
-  for (int i = 0; i < 125; ++i) {
-    in_x(i) = x(i);
-    in_a(i) = a(i);
-    in_b(i) = b(i);
-    expected_out(i) = v(i);
-  }
-
-  std::size_t bytes = in_x.size() * sizeof(Scalar);
-
-  Scalar* d_in_x;
-  Scalar* d_in_a;
-  Scalar* d_in_b;
-  Scalar* d_out;
-  hipMalloc((void**)(&d_in_x), bytes);
-  hipMalloc((void**)(&d_in_a), bytes);
-  hipMalloc((void**)(&d_in_b), bytes);
-  hipMalloc((void**)(&d_out), bytes);
-
-  hipMemcpy(d_in_x, in_x.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in_a, in_a.data(), bytes, hipMemcpyHostToDevice);
-  hipMemcpy(d_in_b, in_b.data(), bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 125);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_a(d_in_a, 125);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_b(d_in_b, 125);
-  Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 125);
-
-  gpu_out.device(gpu_device) = betainc(gpu_in_a, gpu_in_b, gpu_in_x);
-
-  assert(hipMemcpyAsync(out.data(), d_out, bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  for (int i = 1; i < 125; ++i) {
-    if ((std::isnan)(expected_out(i))) {
-      VERIFY((std::isnan)(out(i)));
-    } else {
-      VERIFY_IS_APPROX(out(i), expected_out(i));
-    }
-  }
-
-  hipFree(d_in_x);
-  hipFree(d_in_a);
-  hipFree(d_in_b);
-  hipFree(d_out);
-}
-
-
-void test_cxx11_tensor_hip()
-{
-  CALL_SUBTEST(test_hip_nullary());
-  CALL_SUBTEST(test_hip_elementwise_small());
-  CALL_SUBTEST(test_hip_elementwise());
-  CALL_SUBTEST(test_hip_props());
-  CALL_SUBTEST(test_hip_reduction());
-  CALL_SUBTEST(test_hip_contraction<ColMajor>());
-  CALL_SUBTEST(test_hip_contraction<RowMajor>());
-  CALL_SUBTEST(test_hip_convolution_1d<ColMajor>());
-  CALL_SUBTEST(test_hip_convolution_1d<RowMajor>());
-  CALL_SUBTEST(test_hip_convolution_inner_dim_col_major_1d());
-  CALL_SUBTEST(test_hip_convolution_inner_dim_row_major_1d());
-  CALL_SUBTEST(test_hip_convolution_2d<ColMajor>());
-  CALL_SUBTEST(test_hip_convolution_2d<RowMajor>());
-  // The following two tests commented out due to long runtime
-  // CALL_SUBTEST(test_hip_convolution_3d<ColMajor>());
-  // CALL_SUBTEST(test_hip_convolution_3d<RowMajor>());
-
-#if __cplusplus > 199711L
-  // std::erf, std::erfc, and so on where only added in c++11. We use them
-  // as a golden reference to validate the results produced by Eigen. Therefore
-  // we can only run these tests if we use a c++11 compiler.
-
-  CALL_SUBTEST(test_hip_lgamma<float>(1.0f));
-  CALL_SUBTEST(test_hip_lgamma<float>(100.0f));
-  CALL_SUBTEST(test_hip_lgamma<float>(0.01f));
-  CALL_SUBTEST(test_hip_lgamma<float>(0.001f));
-
-  CALL_SUBTEST(test_hip_lgamma<double>(1.0));
-  CALL_SUBTEST(test_hip_lgamma<double>(100.0));
-  CALL_SUBTEST(test_hip_lgamma<double>(0.01));
-  CALL_SUBTEST(test_hip_lgamma<double>(0.001));
-
-  CALL_SUBTEST(test_hip_erf<float>(1.0f));
-  CALL_SUBTEST(test_hip_erf<float>(100.0f));
-  CALL_SUBTEST(test_hip_erf<float>(0.01f));
-  CALL_SUBTEST(test_hip_erf<float>(0.001f));
-
-  CALL_SUBTEST(test_hip_erfc<float>(1.0f));
-  // CALL_SUBTEST(test_hip_erfc<float>(100.0f)); // HIP erfc lacks precision for large inputs
-  CALL_SUBTEST(test_hip_erfc<float>(5.0f)); 
-  CALL_SUBTEST(test_hip_erfc<float>(0.01f));
-  CALL_SUBTEST(test_hip_erfc<float>(0.001f));
-
-  CALL_SUBTEST(test_hip_erf<double>(1.0));
-  CALL_SUBTEST(test_hip_erf<double>(100.0));
-  CALL_SUBTEST(test_hip_erf<double>(0.01));
-  CALL_SUBTEST(test_hip_erf<double>(0.001));
-
-  CALL_SUBTEST(test_hip_erfc<double>(1.0));
-  // CALL_SUBTEST(test_hip_erfc<double>(100.0)); // HIP erfc lacks precision for large inputs
-  CALL_SUBTEST(test_hip_erfc<double>(5.0)); 
-  CALL_SUBTEST(test_hip_erfc<double>(0.01));
-  CALL_SUBTEST(test_hip_erfc<double>(0.001));
-
-  // Following tests have functional failures on some seeds
-  // CALL_SUBTEST(test_hip_digamma<float>());
-  // CALL_SUBTEST(test_hip_digamma<double>());
-
-  // Following tests have functional failures on some seeds
-  // CALL_SUBTEST(test_hip_polygamma<float>());
-  // CALL_SUBTEST(test_hip_polygamma<double>());
-
-  // Following tests have functional failures on some seeds
-  // CALL_SUBTEST(test_hip_zeta<float>());
-  // CALL_SUBTEST(test_hip_zeta<double>());
-
-  CALL_SUBTEST(test_hip_igamma<float>());
-  CALL_SUBTEST(test_hip_igammac<float>());
-
-  CALL_SUBTEST(test_hip_igamma<double>());
-  CALL_SUBTEST(test_hip_igammac<double>());
-
-  CALL_SUBTEST(test_hip_betainc<float>());
-  CALL_SUBTEST(test_hip_betainc<double>());
-#endif
-}
diff --git a/unsupported/test/cxx11_tensor_of_float16_hip.cu b/unsupported/test/cxx11_tensor_of_float16_hip.cu
deleted file mode 100644
index 6c1a401bf..000000000
--- a/unsupported/test/cxx11_tensor_of_float16_hip.cu
+++ /dev/null
@@ -1,498 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_hip
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-
-using Eigen::Tensor;
-
-template<typename>
-void test_hip_numext() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  bool* d_res_half = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
-  bool* d_res_float = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
-      d_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_half(
-      d_res_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_float(
-      d_res_float, num_elem);
-
-  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
-  gpu_res_float.device(gpu_device) = gpu_float.unaryExpr(Eigen::internal::scalar_isnan_op<float>());
-  gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().unaryExpr(Eigen::internal::scalar_isnan_op<Eigen::half>());
-
-  Tensor<bool, 1> half_prec(num_elem);
-  Tensor<bool, 1> full_prec(num_elem);
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(bool));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(bool));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking numext " << i << std::endl;
-    VERIFY_IS_EQUAL(full_prec(i), half_prec(i));
-  }
-
-  gpu_device.deallocate(d_float);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-
-#ifdef EIGEN_HAS_HIP_FP16
-
-template<typename>
-void test_hip_conversion() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-  
-  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
-      d_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
-      d_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
-      d_conv, num_elem);
-
-  gpu_float.device(gpu_device) = gpu_float.random();
-  gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
-  gpu_conv.device(gpu_device) = gpu_half.cast<float>();
-
-  Tensor<float, 1> initial(num_elem);
-  Tensor<float, 1> final(num_elem);
-  gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
-
-  for (int i = 0; i < num_elem; ++i) {
-    VERIFY_IS_APPROX(initial(i), final(i));
-  }
-
-  gpu_device.deallocate(d_float);
-  gpu_device.deallocate(d_half);
-  gpu_device.deallocate(d_conv);
-}
-
-template<typename>
-void test_hip_unary() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
-      d_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
-      d_res_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
-      d_res_float, num_elem);
-
-  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
-  gpu_res_float.device(gpu_device) = gpu_float.abs();
-  gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().cast<float>();
-
-  Tensor<float, 1> half_prec(num_elem);
-  Tensor<float, 1> full_prec(num_elem);
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking unary " << i << std::endl;
-    VERIFY_IS_APPROX(full_prec(i), half_prec(i));
-  }
-
-  gpu_device.deallocate(d_float);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-template<typename>
-void test_hip_elementwise() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(
-      d_float1, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(
-      d_float2, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
-      d_res_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
-      d_res_float, num_elem);
-
-  gpu_float1.device(gpu_device) = gpu_float1.random();
-  gpu_float2.device(gpu_device) = gpu_float2.random();
-  gpu_res_float.device(gpu_device) = (gpu_float1 + gpu_float2) * gpu_float1;
-  gpu_res_half.device(gpu_device) = ((gpu_float1.cast<Eigen::half>() + gpu_float2.cast<Eigen::half>()) * gpu_float1.cast<Eigen::half>()).cast<float>();
-
-  Tensor<float, 1> half_prec(num_elem);
-  Tensor<float, 1> full_prec(num_elem);
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking elemwise " << i << ": full prec = " << full_prec(i) << " vs half prec = " << half_prec(i) << std::endl;
-    VERIFY_IS_APPROX(static_cast<Eigen::half>(full_prec(i)), static_cast<Eigen::half>(half_prec(i)));
-  }
-
-  gpu_device.deallocate(d_float1);
-  gpu_device.deallocate(d_float2);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-template<typename>
-void test_hip_trancendental() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float3 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_res1_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res1_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res2_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res2_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res3_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res3_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(d_float1, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(d_float2, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float3(d_float3, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half(d_res1_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float(d_res1_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half(d_res2_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float(d_res2_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_half(d_res3_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_float(d_res3_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res4_half(d_res3_half, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res4_float(d_res3_float, num_elem);
-
-  gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
-  gpu_float2.device(gpu_device) = gpu_float2.random() + gpu_float1.constant(0.5f);
-  gpu_float3.device(gpu_device) = gpu_float3.random();
-  gpu_res1_float.device(gpu_device) = gpu_float1.exp().cast<Eigen::half>();
-  gpu_res2_float.device(gpu_device) = gpu_float2.log().cast<Eigen::half>();
-  gpu_res3_float.device(gpu_device) = gpu_float3.log1p().cast<Eigen::half>();
-  gpu_res4_float.device(gpu_device) = gpu_float3.expm1().cast<Eigen::half>();
-
-  gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>();
-  gpu_res1_half.device(gpu_device) = gpu_res1_half.exp();
-
-  gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>();
-  gpu_res2_half.device(gpu_device) = gpu_res2_half.log();
-
-  gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>();
-  gpu_res3_half.device(gpu_device) = gpu_res3_half.log1p();
-
-  gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>();
-  gpu_res3_half.device(gpu_device) = gpu_res3_half.expm1();
-
-  Tensor<float, 1> input1(num_elem);
-  Tensor<Eigen::half, 1> half_prec1(num_elem);
-  Tensor<Eigen::half, 1> full_prec1(num_elem);
-  Tensor<float, 1> input2(num_elem);
-  Tensor<Eigen::half, 1> half_prec2(num_elem);
-  Tensor<Eigen::half, 1> full_prec2(num_elem);
-  Tensor<float, 1> input3(num_elem);
-  Tensor<Eigen::half, 1> half_prec3(num_elem);
-  Tensor<Eigen::half, 1> full_prec3(num_elem);
-  gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem*sizeof(Eigen::half));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i) << " half = " << half_prec1(i) << std::endl;
-    VERIFY_IS_APPROX(full_prec1(i), half_prec1(i));
-  }
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
-    if(std::abs(input2(i)-1.f)<0.05f) // log lacks accurary nearby 1
-      VERIFY_IS_APPROX(full_prec2(i)+Eigen::half(0.1f), half_prec2(i)+Eigen::half(0.1f));
-    else
-      VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
-  }
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i) << " half = " << half_prec3(i) << std::endl;
-    VERIFY_IS_APPROX(full_prec3(i), half_prec3(i));
-  }
-  gpu_device.deallocate(d_float1);
-  gpu_device.deallocate(d_float2);
-  gpu_device.deallocate(d_float3);
-  gpu_device.deallocate(d_res1_half);
-  gpu_device.deallocate(d_res1_float);
-  gpu_device.deallocate(d_res2_half);
-  gpu_device.deallocate(d_res2_float);
-  gpu_device.deallocate(d_res3_float);
-  gpu_device.deallocate(d_res3_half);
-}
-
-template<typename>
-void test_hip_contractions() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int rows = 23;
-  int cols = 23;
-  int num_elem = rows*cols;
-
-  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
-      d_float1, rows, cols);
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
-      d_float2, rows, cols);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(
-      d_res_half, rows, cols);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(
-      d_res_float, rows, cols);
-
-  gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
-  gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float2.constant(0.5f);
-
-  typedef Tensor<float, 2>::DimensionPair DimPair;
-  Eigen::array<DimPair, 1> dims(DimPair(1, 0));
-  gpu_res_float.device(gpu_device) = gpu_float1.contract(gpu_float2, dims).cast<Eigen::half>();
-  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().contract(gpu_float2.cast<Eigen::half>(), dims);
-
-  Tensor<Eigen::half, 2> half_prec(rows, cols);
-  Tensor<Eigen::half, 2> full_prec(rows, cols);
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(Eigen::half));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < rows; ++i) {
-    for (int j = 0; j < cols; ++j) {
-      std::cout << "Checking contract " << i << " " << j << full_prec(i, j) << " " << half_prec(i, j) << std::endl;
-      if (numext::abs(full_prec(i, j) - half_prec(i, j)) > Eigen::half(1e-2f)) {
-        VERIFY_IS_APPROX(full_prec(i, j), half_prec(i, j));
-      }
-    }
-  }
-
-  gpu_device.deallocate(d_float1);
-  gpu_device.deallocate(d_float2);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-template<typename>
-void test_hip_reductions(int size1, int size2, int redux) {
-
-   std::cout << "Reducing " << size1 << " by " << size2
-             << " tensor along dim " << redux << std::endl; 
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = size1*size2;
-  int result_size = (redux == 1 ? size1 : size2);
-
-  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
-  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
-      d_float1, size1, size2);
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
-      d_float2, size1, size2);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(
-      d_res_half, result_size);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(
-      d_res_float, result_size);
-
-  gpu_float1.device(gpu_device) = gpu_float1.random() * 2.0f;
-  gpu_float2.device(gpu_device) = gpu_float2.random() * 2.0f;
-
-  Eigen::array<int, 1> redux_dim(redux);
-  gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim).cast<Eigen::half>();
-  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim);
-
-  Tensor<Eigen::half, 1> half_prec(result_size);
-  Tensor<Eigen::half, 1> full_prec(result_size);
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, result_size*sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, result_size*sizeof(Eigen::half));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < result_size; ++i) {
-    std::cout << "EXPECTED " << full_prec(i) << " GOT " << half_prec(i) << std::endl;
-    VERIFY_IS_APPROX(full_prec(i), half_prec(i));
-  }
-
-  gpu_device.deallocate(d_float1);
-  gpu_device.deallocate(d_float2);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-template<typename>
-void test_hip_reductions() {
-  test_hip_reductions<void>(13, 13, 0);
-  test_hip_reductions<void>(13, 13, 1);
-
-  test_hip_reductions<void>(35, 36, 0);
-  test_hip_reductions<void>(35, 36, 1);
-
-  test_hip_reductions<void>(36, 35, 0);
-  test_hip_reductions<void>(36, 35, 1);
-}
-
-template<typename>
-void test_hip_full_reductions() {
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int size = 13;
-  int num_elem = size*size;
-
-  float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
-  Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
-      d_float1, size, size);
-  Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
-      d_float2, size, size);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(
-      d_res_half);
-  Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(
-      d_res_float);
-
-  gpu_float1.device(gpu_device) = gpu_float1.random();
-  gpu_float2.device(gpu_device) = gpu_float2.random();
-
-  gpu_res_float.device(gpu_device) = gpu_float1.sum().cast<Eigen::half>();
-  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum();
-
-  Tensor<Eigen::half, 0> half_prec;
-  Tensor<Eigen::half, 0> full_prec;
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
-  gpu_device.synchronize();
-
-  VERIFY_IS_APPROX(full_prec(), half_prec());
-
-  gpu_res_float.device(gpu_device) = gpu_float1.maximum().cast<Eigen::half>();
-  gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().maximum();
-  gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
-  gpu_device.synchronize();
-
-  VERIFY_IS_APPROX(full_prec(), half_prec());
-
-  gpu_device.deallocate(d_float1);
-  gpu_device.deallocate(d_float2);
-  gpu_device.deallocate(d_res_half);
-  gpu_device.deallocate(d_res_float);
-}
-
-template<typename>
-void test_hip_forced_evals() {
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-  int num_elem = 101;
-
-  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_half1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_half2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
-  float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
-
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
-      d_float, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half1(
-      d_res_half1, num_elem);
- Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Unaligned> gpu_res_half2(
-      d_res_half2, num_elem);
-  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
-      d_res_float, num_elem);
-
-  Eigen::array<int, 1> no_bcast;
-  no_bcast[0] = 1;
-
-  gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
-  gpu_res_float.device(gpu_device) = gpu_float.abs();
-  gpu_res_half1.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().eval().cast<float>();
-  gpu_res_half2.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().broadcast(no_bcast).eval().cast<float>();
-
-  Tensor<float, 1> half_prec1(num_elem);
-  Tensor<float, 1> half_prec2(num_elem);
-  Tensor<float, 1> full_prec(num_elem);
-  gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res_half1, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res_half1, num_elem*sizeof(float));
-  gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
-  gpu_device.synchronize();
-
-  for (int i = 0; i < num_elem; ++i) {
-    std::cout << "Checking forced eval " << i << full_prec(i) << " vs " << half_prec1(i) << " vs " << half_prec2(i) << std::endl;
-    VERIFY_IS_APPROX(full_prec(i), half_prec1(i));
-    VERIFY_IS_APPROX(full_prec(i), half_prec2(i));
-  }
-
-  gpu_device.deallocate(d_float);
-  gpu_device.deallocate(d_res_half1);
-  gpu_device.deallocate(d_res_half2);
-  gpu_device.deallocate(d_res_float);
-}
-#endif
-
-
-void test_cxx11_tensor_of_float16_hip()
-{
-  CALL_SUBTEST(test_hip_numext<void>());
-
-#ifdef EIGEN_HAS_HIP_FP16
-  CALL_SUBTEST(test_hip_conversion<void>());
-  CALL_SUBTEST(test_hip_unary<void>());
-  CALL_SUBTEST(test_hip_elementwise<void>());
-  CALL_SUBTEST(test_hip_trancendental<void>());
-  CALL_SUBTEST(test_hip_contractions<void>());
-  CALL_SUBTEST(test_hip_reductions<void>());
-  CALL_SUBTEST(test_hip_full_reductions<void>());
-  CALL_SUBTEST(test_hip_forced_evals<void>());
-#else
-  std::cout << "Half floats are not supported by this version of hip: skipping the test" << std::endl;
-#endif
-}
diff --git a/unsupported/test/cxx11_tensor_random_hip.cu b/unsupported/test/cxx11_tensor_random_hip.cu
deleted file mode 100644
index 7d7e72ca2..000000000
--- a/unsupported/test/cxx11_tensor_random_hip.cu
+++ /dev/null
@@ -1,85 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_random_hip
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <Eigen/CXX11/Tensor>
-
-
-void test_hip_random_uniform()
-{
-  Tensor<float, 2> out(72,97);
-  out.setZero();
-
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_out;
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
-
-  gpu_out.device(gpu_device) = gpu_out.random();
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-
-  // For now we just check thes code doesn't crash.
-  // TODO: come up with a valid test of randomness
-}
-
-
-void test_hip_random_normal()
-{
-  Tensor<float, 2> out(72,97);
-  out.setZero();
-
-  std::size_t out_bytes = out.size() * sizeof(float);
-
-  float* d_out;
-  hipMalloc((void**)(&d_out), out_bytes);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
-
-  Eigen::internal::NormalRandomGenerator<float> gen(true);
-  gpu_out.device(gpu_device) = gpu_out.random(gen);
-
-  assert(hipMemcpyAsync(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost, gpu_device.stream()) == hipSuccess);
-  assert(hipStreamSynchronize(gpu_device.stream()) == hipSuccess);
-}
-
-static void test_complex()
-{
-  Tensor<std::complex<float>, 1> vec(6);
-  vec.setRandom();
-
-  // Fixme: we should check that the generated numbers follow a uniform
-  // distribution instead.
-  for (int i = 1; i < 6; ++i) {
-    VERIFY_IS_NOT_EQUAL(vec(i), vec(i-1));
-  }
-}
-
-
-void test_cxx11_tensor_random_hip()
-{
-  CALL_SUBTEST(test_hip_random_uniform());
-  CALL_SUBTEST(test_hip_random_normal());
-  CALL_SUBTEST(test_complex());
-}
diff --git a/unsupported/test/cxx11_tensor_reduction_hip.cu b/unsupported/test/cxx11_tensor_reduction_hip.cu
deleted file mode 100644
index c5aad05be..000000000
--- a/unsupported/test/cxx11_tensor_reduction_hip.cu
+++ /dev/null
@@ -1,154 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_reduction_hip
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-
-template<typename Type, int DataLayout>
-static void test_full_reductions() {
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  const int num_rows = internal::random<int>(1024, 5*1024);
-  const int num_cols = internal::random<int>(1024, 5*1024);
-
-  Tensor<Type, 2, DataLayout> in(num_rows, num_cols);
-  in.setRandom();
-
-  Tensor<Type, 0, DataLayout> full_redux;
-  full_redux = in.sum();
-
-  std::size_t in_bytes = in.size() * sizeof(Type);
-  std::size_t out_bytes = full_redux.size() * sizeof(Type);
-  Type* gpu_in_ptr = static_cast<Type*>(gpu_device.allocate(in_bytes));
-  Type* gpu_out_ptr = static_cast<Type*>(gpu_device.allocate(out_bytes));
-  gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
-
-  TensorMap<Tensor<Type, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols);
-  TensorMap<Tensor<Type, 0, DataLayout> > out_gpu(gpu_out_ptr);
-
-  out_gpu.device(gpu_device) = in_gpu.sum();
-
-  Tensor<Type, 0, DataLayout> full_redux_gpu;
-  gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
-  gpu_device.synchronize();
-
-  // Check that the CPU and GPU reductions return the same result.
-  VERIFY_IS_APPROX(full_redux(), full_redux_gpu());
-
-  gpu_device.deallocate(gpu_in_ptr);
-  gpu_device.deallocate(gpu_out_ptr);
-}
-
-template<typename Type, int DataLayout>
-static void test_first_dim_reductions() {
-  int dim_x = 33;
-  int dim_y = 1;
-  int dim_z = 128;
-
-  Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z);
-  in.setRandom();
-
-  Eigen::array<int, 1> red_axis;
-  red_axis[0] = 0;
-  Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
-
-  // Create device
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice dev(&stream);
-  
-  // Create data(T)
-  Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type));
-  Type* out_data = (Type*)dev.allocate(dim_z*dim_y*sizeof(Type));
-  Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z);
-  Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_y, dim_z);
-  
-  // Perform operation
-  dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type));
-  gpu_out.device(dev) = gpu_in.sum(red_axis);
-  gpu_out.device(dev) += gpu_in.sum(red_axis);
-  Tensor<Type, 2, DataLayout> redux_gpu(dim_y, dim_z);
-  dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type));
-  dev.synchronize();
-
-  // Check that the CPU and GPU reductions return the same result.
-  for (int i = 0; i < gpu_out.size(); ++i) {
-    VERIFY_IS_APPROX(2*redux(i), redux_gpu(i));
-  }
-
-  dev.deallocate(in_data);
-  dev.deallocate(out_data);
-}
-
-template<typename Type, int DataLayout>
-static void test_last_dim_reductions() {
-  int dim_x = 128;
-  int dim_y = 1;
-  int dim_z = 33;
-
-  Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z);
-  in.setRandom();
-
-  Eigen::array<int, 1> red_axis;
-  red_axis[0] = 2;
-  Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
-
-  // Create device
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice dev(&stream);
-  
-  // Create data
-  Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type));
-  Type* out_data = (Type*)dev.allocate(dim_x*dim_y*sizeof(Type));
-  Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z);
-  Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_x, dim_y);
-  
-  // Perform operation
-  dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type));
-  gpu_out.device(dev) = gpu_in.sum(red_axis);
-  gpu_out.device(dev) += gpu_in.sum(red_axis);
-  Tensor<Type, 2, DataLayout> redux_gpu(dim_x, dim_y);
-  dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type));
-  dev.synchronize();
-
-  // Check that the CPU and GPU reductions return the same result.
-  for (int i = 0; i < gpu_out.size(); ++i) {
-    VERIFY_IS_APPROX(2*redux(i), redux_gpu(i));
-  }
-
-  dev.deallocate(in_data);
-  dev.deallocate(out_data);
-}
-
-
-void test_cxx11_tensor_reduction_hip() {
-  CALL_SUBTEST((test_full_reductions<float, ColMajor>()));
-  CALL_SUBTEST((test_full_reductions<double, ColMajor>()));
-  CALL_SUBTEST((test_full_reductions<float, RowMajor>()));
-  CALL_SUBTEST((test_full_reductions<double, RowMajor>()));
-  
-  CALL_SUBTEST((test_first_dim_reductions<float, ColMajor>()));
-  CALL_SUBTEST((test_first_dim_reductions<double, ColMajor>()));
-  CALL_SUBTEST((test_first_dim_reductions<float, RowMajor>()));
-// Outer reductions of doubles aren't supported just yet.  					      
-//  CALL_SUBTEST((test_first_dim_reductions<double, RowMajor>()))
-
-  CALL_SUBTEST((test_last_dim_reductions<float, ColMajor>()));
-// Outer reductions of doubles aren't supported just yet.  					      
-//  CALL_SUBTEST((test_last_dim_reductions<double, ColMajor>()));
-  CALL_SUBTEST((test_last_dim_reductions<float, RowMajor>()));
-  CALL_SUBTEST((test_last_dim_reductions<double, RowMajor>()));
-}
diff --git a/unsupported/test/cxx11_tensor_scan_hip.cu b/unsupported/test/cxx11_tensor_scan_hip.cu
deleted file mode 100644
index f4c4f59b9..000000000
--- a/unsupported/test/cxx11_tensor_scan_hip.cu
+++ /dev/null
@@ -1,76 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_scan_hip
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-#define EIGEN_USE_GPU
-
-#include "main.h"
-#include <unsupported/Eigen/CXX11/Tensor>
-
-using Eigen::Tensor;
-typedef Tensor<float, 1>::DimensionPair DimPair;
-
-template<int DataLayout>
-void test_hip_cumsum(int m_size, int k_size, int n_size)
-{
-  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
-  Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size);
-  Tensor<float, 3, DataLayout> t_result(m_size, k_size, n_size);
-  Tensor<float, 3, DataLayout> t_result_gpu(m_size, k_size, n_size);
-
-  t_input.setRandom();
-
-  std::size_t t_input_bytes = t_input.size()  * sizeof(float);
-  std::size_t t_result_bytes = t_result.size() * sizeof(float);
-
-  float* d_t_input;
-  float* d_t_result;
-
-  hipMalloc((void**)(&d_t_input), t_input_bytes);
-  hipMalloc((void**)(&d_t_result), t_result_bytes);
-
-  hipMemcpy(d_t_input, t_input.data(), t_input_bytes, hipMemcpyHostToDevice);
-
-  Eigen::HipStreamDevice stream;
-  Eigen::GpuDevice gpu_device(&stream);
-
-  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
-      gpu_t_input(d_t_input, Eigen::array<int, 3>(m_size, k_size, n_size));
-  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
-      gpu_t_result(d_t_result, Eigen::array<int, 3>(m_size, k_size, n_size));
-
-  gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1);
-  t_result = t_input.cumsum(1);
-
-  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
-  for (DenseIndex i = 0; i < t_result.size(); i++) {
-    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
-      continue;
-    }
-    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
-      continue;
-    }
-    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
-              << " vs " <<  t_result_gpu(i) << std::endl;
-    assert(false);
-  }
-
-  hipFree((void*)d_t_input);
-  hipFree((void*)d_t_result);
-}
-
-
-void test_cxx11_tensor_scan_hip()
-{
-  CALL_SUBTEST(test_hip_cumsum<ColMajor>(128, 128, 128));
-  CALL_SUBTEST(test_hip_cumsum<RowMajor>(128, 128, 128));
-}

From 1bb6fa99a31d2dcf5431087d3f238e2dcca03084 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 20 Jun 2018 16:44:58 -0400
Subject: [PATCH 407/680] merging the CUDA and HIP implementation for the
 Tensor directory and the unit tests

---
 cmake/EigenTesting.cmake                      |   3 +-
 unsupported/Eigen/CXX11/Tensor                |  24 +-
 .../CXX11/src/Tensor/TensorContractionGpu.h   | 189 ++++-
 .../CXX11/src/Tensor/TensorConvolution.h      | 136 ++--
 .../Eigen/CXX11/src/Tensor/TensorDeviceGpu.h  | 173 +++--
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |  21 +-
 .../src/Tensor/TensorGpuHipCudaDefines.h      |  86 +++
 .../src/Tensor/TensorGpuHipCudaUndefines.h    |  39 +
 .../CXX11/src/Tensor/TensorReductionGpu.h     | 190 +++--
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |   8 +-
 unsupported/test/CMakeLists.txt               |  47 +-
 unsupported/test/cxx11_tensor_argmax_gpu.cu   |  90 +--
 .../test/cxx11_tensor_cast_float16_gpu.cu     |  10 +-
 .../cxx11_tensor_complex_cwise_ops_gpu.cu     |   2 +-
 unsupported/test/cxx11_tensor_complex_gpu.cu  |   8 +-
 unsupported/test/cxx11_tensor_contract_gpu.cu |  92 +--
 unsupported/test/cxx11_tensor_device.cu       |  58 +-
 unsupported/test/cxx11_tensor_gpu.cu          | 694 +++++++++---------
 .../test/cxx11_tensor_of_float16_gpu.cu       |  78 +-
 unsupported/test/cxx11_tensor_random_gpu.cu   |  29 +-
 .../test/cxx11_tensor_reduction_gpu.cu        |  10 +-
 unsupported/test/cxx11_tensor_scan_gpu.cu     |  25 +-
 22 files changed, 1208 insertions(+), 804 deletions(-)
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index f818ae840..1d4486c05 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -20,7 +20,8 @@ macro(ei_add_test_internal testname testname_with_suffix)
 
   if(EIGEN_ADD_TEST_FILENAME_EXTENSION STREQUAL cu)
     if(EIGEN_TEST_HIP)
-      hip_add_executable(${targetname} ${filename} HIPCC_OPTIONS "-DEIGEN_USE_HIP")
+      hip_reset_flags()
+      hip_add_executable(${targetname} ${filename} HIPCC_OPTIONS "-DEIGEN_USE_HIP ${ARGV2}")
     elseif(EIGEN_TEST_CUDA_CLANG)
       set_source_files_properties(${filename} PROPERTIES LANGUAGE CXX)
       if(CUDA_64_BIT_DEVICE_CODE)
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 4b7c7d724..ddbbcfba2 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -99,11 +99,7 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorCostModel.h"
 #include "src/Tensor/TensorDeviceDefault.h"
 #include "src/Tensor/TensorDeviceThreadPool.h"
-#if defined(EIGEN_USE_HIP)
-  #include "src/Tensor/TensorDeviceHip.h"
-#else
-  #include "src/Tensor/TensorDeviceCuda.h"
-#endif
+#include "src/Tensor/TensorDeviceGpu.h"
 #include "src/Tensor/TensorDeviceSycl.h"
 #include "src/Tensor/TensorIndexList.h"
 #include "src/Tensor/TensorDimensionList.h"
@@ -120,28 +116,16 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorEvaluator.h"
 #include "src/Tensor/TensorExpr.h"
 #include "src/Tensor/TensorReduction.h"
-#if defined(EIGEN_USE_HIP)
-  #include "src/Tensor/TensorReductionHip.h"
-#else
-  #include "src/Tensor/TensorReductionCuda.h"
-#endif
+#include "src/Tensor/TensorReductionGpu.h"
 #include "src/Tensor/TensorArgMax.h"
 #include "src/Tensor/TensorConcatenation.h"
 #include "src/Tensor/TensorContractionMapper.h"
 #include "src/Tensor/TensorContractionBlocking.h"
 #include "src/Tensor/TensorContraction.h"
 #include "src/Tensor/TensorContractionThreadPool.h"
-#if defined(EIGEN_USE_HIP)
-  #include "src/Tensor/TensorContractionHip.h"
-#else
-  #include "src/Tensor/TensorContractionCuda.h"
-#endif
+#include "src/Tensor/TensorContractionGpu.h"
 #include "src/Tensor/TensorConversion.h"
-#if defined(EIGEN_USE_HIP)
-  #include "src/Tensor/TensorConvolutionHip.h"
-#else
-  #include "src/Tensor/TensorConvolution.h"
-#endif   
+#include "src/Tensor/TensorConvolution.h"
 #include "src/Tensor/TensorFFT.h"
 #include "src/Tensor/TensorPatch.h"
 #include "src/Tensor/TensorImagePatch.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index 903bc51cc..238754424 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -9,10 +9,10 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
-#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
 
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
 
 namespace Eigen {
 
@@ -388,7 +388,7 @@ EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
   // the sum across all big k blocks of the product of little k block of index (x, y)
   // with block of index (y, z). To compute the final output, we need to reduce
   // the 8 threads over y by summation.
-#if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
 #define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
 #else
 #define shuffleInc(i, j, mask) res(i, j) += __shfl_xor_sync(0xFFFFFFFF, res(i, j), mask)
@@ -503,7 +503,11 @@ EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 template<typename Scalar, typename Index, typename LhsMapper,
          typename RhsMapper, typename OutputMapper>
 __global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(512, 1)
+#else  
 __launch_bounds__(512)
+#endif  
 EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
                        const OutputMapper output,
                        const Index m_size, const Index n_size, const Index k_size) {
@@ -542,7 +546,45 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     results[i].x = results[i].y = results[i].z = results[i].w = 0;
   }
 
+#if defined(EIGEN_HIPCC)
 
+#define prefetch_lhs(reg, row, col)                            \
+    if (!CHECK_LHS_BOUNDARY) {                                 \
+      if (col < k_size) {                                      \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+          reg.z =lhs(row + 2, col);                            \
+          reg.w =lhs(row + 3, col);                            \
+      }                                                        \
+    } else {                                                   \
+      if (col < k_size) {                                      \
+        if (row + 3 < m_size) {                                \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+          reg.z =lhs(row + 2, col);                            \
+          reg.w =lhs(row + 3, col);                            \
+        } else if (row + 2 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+          reg.z =lhs(row + 2, col);                            \
+        } else if (row + 1 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+        } else if (row  < m_size) {                            \
+          reg.x =lhs(row + 0, col);                            \
+        }                                                      \
+      }                                                        \
+    }							       \
+
+#define prefetch_rhs_hipcc(reg, row, col)                  \
+      reg.x =rhs(row + 0, col);				   \
+      reg.y =rhs(row + 1, col);				   \
+      reg.z =rhs(row + 2, col);                            \
+      reg.w =rhs(row + 3, col);                            \
+
+
+#else
+  
 #define prefetch_lhs(reg, row, col)                            \
     if (!CHECK_LHS_BOUNDARY) {                                 \
       if (col < k_size) {                                      \
@@ -563,14 +605,21 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
           reg.x =lhs(row + 0, col);                            \
         }                                                      \
       }                                                        \
-    }                                                          \
+    }							       \
 
+#endif
 
   Index lhs_vert = base_m+threadIdx.x*4;
 
   for (Index k = 0; k < k_size; k += 16) {
+
+#if defined(EIGEN_HIPCC)
+    lhs_pf0 = make_float4(0, 0, 0, 0);
+    rhs_pf0 = make_float4(0, 0, 0, 0);
+#else
     lhs_pf0 = internal::pset1<float4>(0);
     rhs_pf0 = internal::pset1<float4>(0);
+#endif
 
     Index lhs_horiz = threadIdx.y+k;
     prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
@@ -581,7 +630,11 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
+#if defined(EIGEN_HIPCC)
+	prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
+#else
         rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+#endif
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -596,7 +649,11 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     } else {
       if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
+#if defined(EIGEN_HIPCC)
+	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
+#else
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+#endif
         } else if ((rhs_vert + 2) < k_size) {
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
           rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
@@ -618,7 +675,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
       x1 = rhs_pf0.x;
       x2 = rhs_pf0.z;
     }
-    #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+    #if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
     x1 = __shfl_xor(x1, 4);
     x2 = __shfl_xor(x2, 4);
     #else
@@ -695,7 +752,11 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
 
 #undef prefetch_lhs
 #undef add_vals
-
+  
+#if defined(EIGEN_HIPCC)  
+#undef prefetch_rhs_hipcc
+#endif
+  
   Index horiz_base = threadIdx.y*4+base_n;
   if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
     for (int i = 0; i < 4; i++) {
@@ -784,9 +845,33 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     results[i].x = results[i].y = results[i].z = results[i].w = 0;
   }
 
+#if defined(EIGEN_HIPCC)
+
+#define prefetch_lhs_hipcc(reg, row, col)                  \
+      reg.x =lhs(row + 0, col);                            \
+      reg.y =lhs(row + 1, col);                            \
+      reg.z =lhs(row + 2, col);                            \
+      reg.w =lhs(row + 3, col);                            
+
+#define prefetch_rhs_hipcc(reg, row, col)                  \
+      reg.x =rhs(row + 0, col);                            \
+      reg.y =rhs(row + 1, col);                            \
+      reg.z =rhs(row + 2, col);                            \
+      reg.w =rhs(row + 3, col);                            
+
+#endif
 
   Index lhs_vert = base_m+threadIdx.x*4+(threadIdx.y%4)*32;
   for (Index k = 0; k < k_size; k += 32) {
+#if defined(EIGEN_HIPCC)
+    lhs_pf0 = make_float4(0, 0, 0, 0);
+    lhs_pf1 = make_float4(0, 0, 0, 0);
+    lhs_pf2 = make_float4(0, 0, 0, 0);
+    lhs_pf3 = make_float4(0, 0, 0, 0);
+
+    rhs_pf0 = make_float4(0, 0, 0, 0);
+    rhs_pf1 = make_float4(0, 0, 0, 0);
+#else  
     lhs_pf0 = internal::pset1<float4>(0);
     lhs_pf1 = internal::pset1<float4>(0);
     lhs_pf2 = internal::pset1<float4>(0);
@@ -794,40 +879,85 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 
     rhs_pf0 = internal::pset1<float4>(0);
     rhs_pf1 = internal::pset1<float4>(0);
+#endif
 
      if (!CHECK_LHS_BOUNDARY) {
       if ((threadIdx.y/4+k+24) < k_size) {
+#if defined(EIGEN_HIPCC)
+	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+	prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
+	prefetch_lhs_hipcc(lhs_pf3, lhs_vert, (threadIdx.y/4+k+24))
+#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
         lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+#endif
       } else if ((threadIdx.y/4+k+16) < k_size) {
+#if defined(EIGEN_HIPCC)
+	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+	prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
+#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+#endif
       } else if ((threadIdx.y/4+k+8) < k_size) {
+#if defined(EIGEN_HIPCC)
+	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+#endif
       } else if ((threadIdx.y/4+k) < k_size) {
+#if defined(EIGEN_HIPCC)
+	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+#endif
       }
     } else {
       // just CHECK_LHS_BOUNDARY
       if (lhs_vert + 3 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
+#if defined(EIGEN_HIPCC)
+	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+	  prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
+	  prefetch_lhs_hipcc(lhs_pf3, lhs_vert, (threadIdx.y/4+k+24))
+#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
           lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
           lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+#endif
         } else if ((threadIdx.y/4+k+16) < k_size) {
+#if defined(EIGEN_HIPCC)
+	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+	  prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
+#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
           lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+#endif
         } else if ((threadIdx.y/4+k+8) < k_size) {
+#if defined(EIGEN_HIPCC)
+	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
+#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+#endif
         } else if ((threadIdx.y/4+k) < k_size) {
+#if defined(EIGEN_HIPCC)
+	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
+#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+#endif
         }
       } else if (lhs_vert + 2 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
@@ -916,8 +1046,13 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
+#if defined(EIGEN_HIPCC)
+	prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
+	prefetch_rhs_hipcc(rhs_pf1, rhs_vert, rhs_horiz1)
+#else	
         rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
         rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+#endif
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -939,8 +1074,13 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       if (rhs_horiz1 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
+#if defined(EIGEN_HIPCC)
+  	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
+	  prefetch_rhs_hipcc(rhs_pf1, rhs_vert, rhs_horiz1)
+#else	
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
           rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+#endif
         } else if (rhs_vert + 2 < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -961,7 +1101,11 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       } else if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
+#if defined(EIGEN_HIPCC)
+  	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
+#else	
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+#endif
         } else if ((rhs_vert + 2) < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -1069,7 +1213,11 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     __syncthreads();
   } // end loop over k
 
-
+#if defined(EIGEN_HIPCC)
+#undef prefetch_lhs_hipcc
+#undef prefetch_rhs_hipcc
+#endif
+  
   __syncthreads();
   Index horiz_base = (threadIdx.y/4)*8+base_n;
   if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
@@ -1134,7 +1282,11 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 template<typename Index, typename LhsMapper,
          typename RhsMapper, typename OutputMapper>
 __global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(256, 1)
+#else  
 __launch_bounds__(256)
+#endif
 EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
                        const OutputMapper output,
                        const Index m_size, const Index n_size, const Index k_size) {
@@ -1177,7 +1329,11 @@ EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
 template<typename Index, typename LhsMapper,
          typename RhsMapper, typename OutputMapper>
 __global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(256, 1)
+#else  
 __launch_bounds__(256)
+#endif
 EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
                        const OutputMapper output,
                        const Index m_size, const Index n_size, const Index k_size) {
@@ -1323,7 +1479,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     const Index n_blocks = (n + 63) / 64;
     const dim3 num_blocks(m_blocks, n_blocks, 1);
     const dim3 block_size(8, 8, 8);
-    LAUNCH_CUDA_KERNEL((EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+    LAUNCH_GPU_KERNEL((EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
     }
   };
 
@@ -1334,13 +1490,13 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         const Index n_blocks = (n + 63) / 64;
         const dim3 num_blocks(m_blocks, n_blocks, 1);
         const dim3 block_size(16, 16, 1);
-        LAUNCH_CUDA_KERNEL((EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+        LAUNCH_GPU_KERNEL((EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
       } else {
         const Index m_blocks = (m + 127) / 128;
         const Index n_blocks = (n + 63) / 64;
         const dim3 num_blocks(m_blocks, n_blocks, 1);
         const dim3 block_size(8, 32, 1);
-        LAUNCH_CUDA_KERNEL((EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+        LAUNCH_GPU_KERNEL((EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
       }
     }
   };
@@ -1384,12 +1540,17 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
     OutputMapper output(buffer, m);
 
-    setCudaSharedMemConfig(cudaSharedMemBankSizeEightByte);
+#if defined(EIGEN_USE_HIP)
+    setGpuSharedMemConfig(hipSharedMemBankSizeEightByte);
+#else
+    setGpuSharedMemConfig(cudaSharedMemBankSizeEightByte);
+#endif
+
     LaunchKernels<LhsScalar, RhsScalar, Index, LhsMapper, RhsMapper, OutputMapper>::Run(lhs, rhs, output,  m, n, k, this->m_device);
   }
 };
 
 } // end namespace Eigen
 
-#endif // EIGEN_USE_GPU and EIGEN_CUDACC
-#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H
+#endif // EIGEN_USE_GPU and EIGEN_GPUCC
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 84d5be173..3110887e1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -54,8 +54,8 @@ class IndexMapper {
       }
     }
 
-    array<Index, NumDims> cudaInputDimensions;
-    array<Index, NumDims> cudaOutputDimensions;
+    array<Index, NumDims> gpuInputDimensions;
+    array<Index, NumDims> gpuOutputDimensions;
     array<Index, NumDims> tmp = dimensions;
     array<Index, NumDims> ordering;
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
@@ -65,8 +65,8 @@ class IndexMapper {
       const Index index = i + offset;
       ordering[index] = indices[i];
       tmp[indices[i]] = -1;
-      cudaInputDimensions[index] = input_dims[indices[i]];
-      cudaOutputDimensions[index] = dimensions[indices[i]];
+      gpuInputDimensions[index] = input_dims[indices[i]];
+      gpuOutputDimensions[index] = dimensions[indices[i]];
     }
 
     int written = static_cast<int>(Layout) == static_cast<int>(ColMajor)
@@ -75,8 +75,8 @@ class IndexMapper {
     for (int i = 0; i < NumDims; ++i) {
       if (tmp[i] >= 0) {
         ordering[written] = i;
-        cudaInputDimensions[written] = input_dims[i];
-        cudaOutputDimensions[written] = dimensions[i];
+        gpuInputDimensions[written] = input_dims[i];
+        gpuOutputDimensions[written] = dimensions[i];
         ++written;
       }
     }
@@ -89,37 +89,37 @@ class IndexMapper {
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int i = 0; i < NumDims; ++i) {
         if (i > NumKernelDims) {
-          m_cudaInputStrides[i] =
-              m_cudaInputStrides[i - 1] * cudaInputDimensions[i - 1];
-          m_cudaOutputStrides[i] =
-              m_cudaOutputStrides[i - 1] * cudaOutputDimensions[i - 1];
+          m_gpuInputStrides[i] =
+              m_gpuInputStrides[i - 1] * gpuInputDimensions[i - 1];
+          m_gpuOutputStrides[i] =
+              m_gpuOutputStrides[i - 1] * gpuOutputDimensions[i - 1];
         } else {
-          m_cudaInputStrides[i] = 1;
-          m_cudaOutputStrides[i] = 1;
+          m_gpuInputStrides[i] = 1;
+          m_gpuOutputStrides[i] = 1;
         }
       }
     } else {
       for (int i = NumDims - 1; i >= 0; --i) {
         if (static_cast<size_t>(i + 1) < offset) {
-          m_cudaInputStrides[i] =
-              m_cudaInputStrides[i + 1] * cudaInputDimensions[i + 1];
-          m_cudaOutputStrides[i] =
-              m_cudaOutputStrides[i + 1] * cudaOutputDimensions[i + 1];
+          m_gpuInputStrides[i] =
+              m_gpuInputStrides[i + 1] * gpuInputDimensions[i + 1];
+          m_gpuOutputStrides[i] =
+              m_gpuOutputStrides[i + 1] * gpuOutputDimensions[i + 1];
         } else {
-          m_cudaInputStrides[i] = 1;
-          m_cudaOutputStrides[i] = 1;
+          m_gpuInputStrides[i] = 1;
+          m_gpuOutputStrides[i] = 1;
         }
       }
     }
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputPlaneToTensorInputOffset(Index p) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputPlaneToTensorInputOffset(Index p) const {
     Index inputIndex = 0;
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int d = NumDims - 1; d > NumKernelDims; --d) {
-        const Index idx = p / m_cudaInputStrides[d];
+        const Index idx = p / m_gpuInputStrides[d];
         inputIndex += idx * m_inputStrides[d];
-        p -= idx * m_cudaInputStrides[d];
+        p -= idx * m_gpuInputStrides[d];
       }
       inputIndex += p * m_inputStrides[NumKernelDims];
     } else {
@@ -128,22 +128,22 @@ class IndexMapper {
         limit = NumDims - NumKernelDims - 1;
       }
       for (int d = 0; d < limit; ++d) {
-        const Index idx = p / m_cudaInputStrides[d];
+        const Index idx = p / m_gpuInputStrides[d];
         inputIndex += idx * m_inputStrides[d];
-        p -= idx * m_cudaInputStrides[d];
+        p -= idx * m_gpuInputStrides[d];
       }
       inputIndex += p * m_inputStrides[limit];
     }
     return inputIndex;
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputPlaneToTensorOutputOffset(Index p) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputPlaneToTensorOutputOffset(Index p) const {
     Index outputIndex = 0;
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int d = NumDims - 1; d > NumKernelDims; --d) {
-        const Index idx = p / m_cudaOutputStrides[d];
+        const Index idx = p / m_gpuOutputStrides[d];
         outputIndex += idx * m_outputStrides[d];
-        p -= idx * m_cudaOutputStrides[d];
+        p -= idx * m_gpuOutputStrides[d];
       }
       outputIndex += p * m_outputStrides[NumKernelDims];
     } else {
@@ -152,44 +152,44 @@ class IndexMapper {
         limit = NumDims - NumKernelDims - 1;
       }
       for (int d = 0; d < limit; ++d) {
-        const Index idx = p / m_cudaOutputStrides[d];
+        const Index idx = p / m_gpuOutputStrides[d];
         outputIndex += idx * m_outputStrides[d];
-        p -= idx * m_cudaOutputStrides[d];
+        p -= idx * m_gpuOutputStrides[d];
       }
       outputIndex += p * m_outputStrides[limit];
     }
     return outputIndex;
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
     return i * m_inputStrides[offset];
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
     return i * m_outputStrides[offset];
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i, Index j) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
     return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1];
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i, Index j) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
     return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1];
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
@@ -197,7 +197,7 @@ class IndexMapper {
            k * m_inputStrides[offset + 2];
   }
 
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
     const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
                               ? 0
                               : NumDims - NumKernelDims;
@@ -209,8 +209,8 @@ class IndexMapper {
   static const int NumDims = internal::array_size<InputDims>::value;
   array<Index, NumDims> m_inputStrides;
   array<Index, NumDims> m_outputStrides;
-  array<Index, NumDims> m_cudaInputStrides;
-  array<Index, NumDims> m_cudaOutputStrides;
+  array<Index, NumDims> m_gpuInputStrides;
+  array<Index, NumDims> m_gpuOutputStrides;
 };
 
 
@@ -553,7 +553,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
 
 
 // Use an optimized implementation of the evaluation code for GPUs whenever possible.
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
 
 template <int StaticKernelSize>
 struct GetKernelSize {
@@ -576,8 +576,12 @@ __global__ void EigenConvolutionKernel1D(
         indexMapper,
     const float* __restrict kernel, const int numPlanes, const int numX,
     const int maxX, const int kernelSize, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)	    
+#else
   extern __shared__ float s[];
-
+#endif
+  
   const int first_x = blockIdx.x * maxX;
   const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
   const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSize>()(kernelSize);
@@ -588,18 +592,18 @@ __global__ void EigenConvolutionKernel1D(
 
   for (int p = first_plane + threadIdx.y; p < numPlanes; p += plane_stride) {
     // Load inputs to shared memory
-    const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p);
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
     const int plane_kernel_offset = threadIdx.y * num_x_input;
     #pragma unroll
     for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
-      const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x);
+      const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x);
       s[i + plane_kernel_offset] = eval.coeff(tensor_index);
     }
 
     __syncthreads();
 
     // Compute the convolution
-    const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p);
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
 
     #pragma unroll
     for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
@@ -609,7 +613,7 @@ __global__ void EigenConvolutionKernel1D(
       for (int k = 0; k < GetKernelSize<StaticKernelSize>()(kernelSize); ++k) {
         result += s[k + kernel_offset] * kernel[k];
       }
-      const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x);
+      const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x);
       buffer[tensor_index] = result;
     }
     __syncthreads();
@@ -625,7 +629,11 @@ __global__ void EigenConvolutionKernel2D(
     const float* __restrict kernel, const int numPlanes, const int numX,
     const int maxX, const int numY, const int maxY, const int kernelSizeX,
     const int kernelSizeY, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)	    
+#else
   extern __shared__ float s[];
+#endif
 
   const int first_x = blockIdx.x * maxX;
   const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
@@ -642,7 +650,7 @@ __global__ void EigenConvolutionKernel2D(
 
   for (int p = first_plane + threadIdx.z; p < numPlanes; p += plane_stride) {
 
-    const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p);
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
     const int plane_kernel_offset = threadIdx.z * num_y_input;
 
     // Load inputs to shared memory
@@ -651,7 +659,7 @@ __global__ void EigenConvolutionKernel2D(
       const int input_offset = num_x_input * (j + plane_kernel_offset);
       #pragma unroll
       for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
-        const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x, j+first_y);
+        const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x, j+first_y);
         s[i + input_offset] = eval.coeff(tensor_index);
       }
     }
@@ -659,7 +667,7 @@ __global__ void EigenConvolutionKernel2D(
     __syncthreads();
 
     // Convolution
-    const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p);
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
 
     #pragma unroll
     for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) {
@@ -675,7 +683,7 @@ __global__ void EigenConvolutionKernel2D(
             result += s[k + input_offset] * kernel[k + kernel_offset];
           }
         }
-        const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x, j+first_y);
+        const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x, j+first_y);
         buffer[tensor_index] = result;
       }
     }
@@ -693,7 +701,11 @@ __global__ void EigenConvolutionKernel3D(
     const size_t maxX, const size_t numY, const size_t maxY, const size_t numZ,
     const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY,
     const size_t kernelSizeZ, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)	    
+#else
   extern __shared__ float s[];
+#endif
 
   // Load inputs to shared memory
   const int first_x = blockIdx.x * maxX;
@@ -710,13 +722,13 @@ __global__ void EigenConvolutionKernel3D(
 
   for (int p = 0; p < numPlanes; ++p) {
 
-    const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p);
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
     const int plane_kernel_offset = 0;
 
     for (int k = threadIdx.z; k < num_z_input; k += blockDim.z) {
       for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) {
         for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
-          const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z);
+          const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z);
           s[i + num_x_input * (j + num_y_input * (k + plane_kernel_offset))] = eval.coeff(tensor_index);
         }
       }
@@ -728,7 +740,7 @@ __global__ void EigenConvolutionKernel3D(
     const int num_z_output = last_z - first_z + 1;
     const int num_y_output = last_y - first_y + 1;
     const int num_x_output = last_x - first_x + 1;
-    const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p);
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
 
     for (int k = threadIdx.z; k < num_z_output; k += blockDim.z) {
       for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) {
@@ -741,7 +753,7 @@ __global__ void EigenConvolutionKernel3D(
               }
             }
           }
-          const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z);
+          const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z);
           buffer[tensor_index] = result;
         }
       }
@@ -854,9 +866,9 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
     typedef typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions InputDims;
 
     const int maxSharedMem = m_device.sharedMemPerBlock();
-    const int maxThreadsPerBlock = m_device.maxCudaThreadsPerBlock();
-    const int maxBlocksPerProcessor = m_device.maxCudaThreadsPerMultiProcessor() / maxThreadsPerBlock;
-    const int numMultiProcessors = m_device.getNumCudaMultiProcessors();
+    const int maxThreadsPerBlock = m_device.maxGpuThreadsPerBlock();
+    const int maxBlocksPerProcessor = m_device.maxGpuThreadsPerMultiProcessor() / maxThreadsPerBlock;
+    const int numMultiProcessors = m_device.getNumGpuMultiProcessors();
     const int warpSize = 32;
 
     switch (NumKernelDims) {
@@ -908,15 +920,15 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
             m_inputImpl.dimensions(), kernel_dims, indices);
         switch(kernel_size) {
           case 4: {
-            LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data);
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data);
             break;
           }
           case 7: {
-            LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data);
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data);
             break;
           }
           default: {
-            LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data);
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data);
           }
         }
         break;
@@ -969,11 +981,11 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           case 4: {
             switch (kernel_size_y) {
               case 7: {
-                LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data);
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data);
                 break;
               }
               default: {
-                LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data);
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data);
                 break;
               }
             }
@@ -982,18 +994,18 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           case 7: {
             switch (kernel_size_y) {
               case 4: {
-                LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data);
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data);
                 break;
               }
               default: {
-                LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data);
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data);
                 break;
               }
             }
             break;
           }
           default: {
-            LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data);
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data);
             break;
           }
         }
@@ -1039,7 +1051,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
         internal::IndexMapper<Index, InputDims, 3, Layout> indexMapper(
             m_inputImpl.dimensions(), kernel_dims, indices);
 
-        LAUNCH_CUDA_KERNEL((EigenConvolutionKernel3D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data);
+        LAUNCH_GPU_KERNEL((EigenConvolutionKernel3D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data);
         break;
       }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
index ded7129da..64ef32793 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
@@ -7,21 +7,26 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_CUDA_H)
-#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_CUDA_H
+#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H)
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H
+
+// This header file container defines fo gpu* macros which will resolve to
+// their equivalent hip* or cuda* versions depending on the compiler in use
+// A separte header (included at the end of this file) will undefine all 
+#include "TensorGpuHipCudaDefines.h"
 
 namespace Eigen {
 
-static const int kCudaScratchSize = 1024;
+static const int kGpuScratchSize = 1024;
 
 // This defines an interface that GPUDevice can take to use
-// CUDA streams underneath.
+// HIP / CUDA streams underneath.
 class StreamInterface {
  public:
   virtual ~StreamInterface() {}
 
-  virtual const cudaStream_t& stream() const = 0;
-  virtual const cudaDeviceProp& deviceProperties() const = 0;
+  virtual const gpuStream_t& stream() const = 0;
+  virtual const gpuDeviceProp_t& deviceProperties() const = 0;
 
   // Allocate memory on the actual device where the computation will run
   virtual void* allocate(size_t num_bytes) const = 0;
@@ -37,7 +42,7 @@ class StreamInterface {
   virtual unsigned int* semaphore() const = 0;
 };
 
-static cudaDeviceProp* m_deviceProperties;
+static gpuDeviceProp_t* m_deviceProperties;
 static bool m_devicePropInitialized = false;
 
 static void initializeDeviceProp() {
@@ -58,23 +63,23 @@ static void initializeDeviceProp() {
 #endif
       // We're the first thread to reach this point.
       int num_devices;
-      cudaError_t status = cudaGetDeviceCount(&num_devices);
-      if (status != cudaSuccess) {
-        std::cerr << "Failed to get the number of CUDA devices: "
-                  << cudaGetErrorString(status)
+      gpuError_t status = gpuGetDeviceCount(&num_devices);
+      if (status != gpuSuccess) {
+        std::cerr << "Failed to get the number of GPU devices: "
+                  << gpuGetErrorString(status)
                   << std::endl;
-        assert(status == cudaSuccess);
+        assert(status == gpuSuccess);
       }
-      m_deviceProperties = new cudaDeviceProp[num_devices];
+      m_deviceProperties = new gpuDeviceProp_t[num_devices];
       for (int i = 0; i < num_devices; ++i) {
-        status = cudaGetDeviceProperties(&m_deviceProperties[i], i);
-        if (status != cudaSuccess) {
-          std::cerr << "Failed to initialize CUDA device #"
+        status = gpuGetDeviceProperties(&m_deviceProperties[i], i);
+        if (status != gpuSuccess) {
+          std::cerr << "Failed to initialize GPU device #"
                     << i
                     << ": "
-                    << cudaGetErrorString(status)
+                    << gpuGetErrorString(status)
                     << std::endl;
-          assert(status == cudaSuccess);
+          assert(status == gpuSuccess);
         }
       }
 
@@ -94,87 +99,87 @@ static void initializeDeviceProp() {
   }
 }
 
-static const cudaStream_t default_stream = cudaStreamDefault;
+static const gpuStream_t default_stream = gpuStreamDefault;
 
-class CudaStreamDevice : public StreamInterface {
+class GpuStreamDevice : public StreamInterface {
  public:
   // Use the default stream on the current device
-  CudaStreamDevice() : stream_(&default_stream), scratch_(NULL), semaphore_(NULL) {
-    cudaGetDevice(&device_);
+  GpuStreamDevice() : stream_(&default_stream), scratch_(NULL), semaphore_(NULL) {
+    gpuGetDevice(&device_);
     initializeDeviceProp();
   }
   // Use the default stream on the specified device
-  CudaStreamDevice(int device) : stream_(&default_stream), device_(device), scratch_(NULL), semaphore_(NULL) {
+  GpuStreamDevice(int device) : stream_(&default_stream), device_(device), scratch_(NULL), semaphore_(NULL) {
     initializeDeviceProp();
   }
   // Use the specified stream. Note that it's the
   // caller responsibility to ensure that the stream can run on
   // the specified device. If no device is specified the code
   // assumes that the stream is associated to the current gpu device.
-  CudaStreamDevice(const cudaStream_t* stream, int device = -1)
+  GpuStreamDevice(const gpuStream_t* stream, int device = -1)
       : stream_(stream), device_(device), scratch_(NULL), semaphore_(NULL) {
     if (device < 0) {
-      cudaGetDevice(&device_);
+      gpuGetDevice(&device_);
     } else {
       int num_devices;
-      cudaError_t err = cudaGetDeviceCount(&num_devices);
+      gpuError_t err = gpuGetDeviceCount(&num_devices);
       EIGEN_UNUSED_VARIABLE(err)
-      assert(err == cudaSuccess);
+      assert(err == gpuSuccess);
       assert(device < num_devices);
       device_ = device;
     }
     initializeDeviceProp();
   }
 
-  virtual ~CudaStreamDevice() {
+  virtual ~GpuStreamDevice() {
     if (scratch_) {
       deallocate(scratch_);
     }
   }
 
-  const cudaStream_t& stream() const { return *stream_; }
-  const cudaDeviceProp& deviceProperties() const {
+  const gpuStream_t& stream() const { return *stream_; }
+  const gpuDeviceProp_t& deviceProperties() const {
     return m_deviceProperties[device_];
   }
   virtual void* allocate(size_t num_bytes) const {
-    cudaError_t err = cudaSetDevice(device_);
+    gpuError_t err = gpuSetDevice(device_);
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
     void* result;
-    err = cudaMalloc(&result, num_bytes);
-    assert(err == cudaSuccess);
+    err = gpuMalloc(&result, num_bytes);
+    assert(err == gpuSuccess);
     assert(result != NULL);
     return result;
   }
   virtual void deallocate(void* buffer) const {
-    cudaError_t err = cudaSetDevice(device_);
+    gpuError_t err = gpuSetDevice(device_);
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
     assert(buffer != NULL);
-    err = cudaFree(buffer);
-    assert(err == cudaSuccess);
+    err = gpuFree(buffer);
+    assert(err == gpuSuccess);
   }
 
   virtual void* scratchpad() const {
     if (scratch_ == NULL) {
-      scratch_ = allocate(kCudaScratchSize + sizeof(unsigned int));
+      scratch_ = allocate(kGpuScratchSize + sizeof(unsigned int));
     }
     return scratch_;
   }
 
   virtual unsigned int* semaphore() const {
     if (semaphore_ == NULL) {
-      char* scratch = static_cast<char*>(scratchpad()) + kCudaScratchSize;
+      char* scratch = static_cast<char*>(scratchpad()) + kGpuScratchSize;
       semaphore_ = reinterpret_cast<unsigned int*>(scratch);
-      cudaError_t err = cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
+      gpuError_t err = gpuMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
       EIGEN_UNUSED_VARIABLE(err)
-      assert(err == cudaSuccess);
+      assert(err == gpuSuccess);
     }
     return semaphore_;
   }
 
  private:
-  const cudaStream_t* stream_;
+  const gpuStream_t* stream_;
   int device_;
   mutable void* scratch_;
   mutable unsigned int* semaphore_;
@@ -190,7 +195,7 @@ struct GpuDevice {
     eigen_assert(stream);
   }
   // TODO(bsteiner): This is an internal API, we should not expose it.
-  EIGEN_STRONG_INLINE const cudaStream_t& stream() const {
+  EIGEN_STRONG_INLINE const gpuStream_t& stream() const {
     return stream_->stream();
   }
 
@@ -211,11 +216,11 @@ struct GpuDevice {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
-#ifndef EIGEN_CUDA_ARCH
-    cudaError_t err = cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice,
+#ifndef EIGEN_GPU_COMPILE_PHASE
+    gpuError_t err = gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToDevice,
                                       stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
 #else
     EIGEN_UNUSED_VARIABLE(dst);
     EIGEN_UNUSED_VARIABLE(src);
@@ -225,24 +230,24 @@ struct GpuDevice {
   }
 
   EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
-    cudaError_t err =
-        cudaMemcpyAsync(dst, src, n, cudaMemcpyHostToDevice, stream_->stream());
+    gpuError_t err =
+        gpuMemcpyAsync(dst, src, n, gpuMemcpyHostToDevice, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
   }
 
   EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
-    cudaError_t err =
-        cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToHost, stream_->stream());
+    gpuError_t err =
+        gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToHost, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
-#ifndef EIGEN_CUDA_ARCH
-    cudaError_t err = cudaMemsetAsync(buffer, c, n, stream_->stream());
+#ifndef EIGEN_GPU_COMPILE_PHASE
+    gpuError_t err = gpuMemsetAsync(buffer, c, n, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == cudaSuccess);
+    assert(err == gpuSuccess);
 #else
   eigen_assert(false && "The default device should be used instead to generate kernel code");
 #endif
@@ -260,31 +265,31 @@ struct GpuDevice {
 
   EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
     // We won't try to take advantage of the l2 cache for the time being, and
-    // there is no l3 cache on cuda devices.
+    // there is no l3 cache on hip/cuda devices.
     return firstLevelCacheSize();
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void synchronize() const {
-#if defined(EIGEN_CUDACC) && !defined(EIGEN_CUDA_ARCH)
-    cudaError_t err = cudaStreamSynchronize(stream_->stream());
-    if (err != cudaSuccess) {
-      std::cerr << "Error detected in CUDA stream: "
-                << cudaGetErrorString(err)
+#if defined(EIGEN_GPUCC) && !defined(EIGEN_GPU_COMPILE_PHASE)
+    gpuError_t err = gpuStreamSynchronize(stream_->stream());
+    if (err != gpuSuccess) {
+      std::cerr << "Error detected in GPU stream: "
+                << gpuGetErrorString(err)
                 << std::endl;
-      assert(err == cudaSuccess);
+      assert(err == gpuSuccess);
     }
 #else
     assert(false && "The default device should be used instead to generate kernel code");
 #endif
   }
 
-  EIGEN_STRONG_INLINE int getNumCudaMultiProcessors() const {
+  EIGEN_STRONG_INLINE int getNumGpuMultiProcessors() const {
     return stream_->deviceProperties().multiProcessorCount;
   }
-  EIGEN_STRONG_INLINE int maxCudaThreadsPerBlock() const {
+  EIGEN_STRONG_INLINE int maxGpuThreadsPerBlock() const {
     return stream_->deviceProperties().maxThreadsPerBlock;
   }
-  EIGEN_STRONG_INLINE int maxCudaThreadsPerMultiProcessor() const {
+  EIGEN_STRONG_INLINE int maxGpuThreadsPerMultiProcessor() const {
     return stream_->deviceProperties().maxThreadsPerMultiProcessor;
   }
   EIGEN_STRONG_INLINE int sharedMemPerBlock() const {
@@ -301,12 +306,12 @@ struct GpuDevice {
     return max_blocks_;
   }
 
-  // This function checks if the CUDA runtime recorded an error for the
+  // This function checks if the GPU runtime recorded an error for the
   // underlying stream device.
   inline bool ok() const {
-#ifdef EIGEN_CUDACC
-    cudaError_t error = cudaStreamQuery(stream_->stream());
-    return (error == cudaSuccess) || (error == cudaErrorNotReady);
+#ifdef EIGEN_GPUCC
+    gpuError_t error = gpuStreamQuery(stream_->stream());
+    return (error == gpuSuccess) || (error == gpuErrorNotReady);
 #else
     return false;
 #endif
@@ -317,18 +322,27 @@ struct GpuDevice {
   int max_blocks_;
 };
 
-#define LAUNCH_CUDA_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
+#if defined(EIGEN_HIPCC)
+
+#define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
+  hipLaunchKernelGGL(kernel, dim3(gridsize), dim3(blocksize), (sharedmem), (device).stream(), __VA_ARGS__); \
+  assert(hipGetLastError() == hipSuccess);
+
+#else
+ 
+#define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
   (kernel) <<< (gridsize), (blocksize), (sharedmem), (device).stream() >>> (__VA_ARGS__);   \
   assert(cudaGetLastError() == cudaSuccess);
 
-
+#endif
+ 
 // FIXME: Should be device and kernel specific.
-#ifdef EIGEN_CUDACC
-static EIGEN_DEVICE_FUNC inline void setCudaSharedMemConfig(cudaSharedMemConfig config) {
-#ifndef EIGEN_CUDA_ARCH
-  cudaError_t status = cudaDeviceSetSharedMemConfig(config);
+#ifdef EIGEN_GPUCC
+static EIGEN_DEVICE_FUNC inline void setGpuSharedMemConfig(gpuSharedMemConfig config) {
+#ifndef EIGEN_GPU_COMPILE_PHASE
+  gpuError_t status = gpuDeviceSetSharedMemConfig(config);
   EIGEN_UNUSED_VARIABLE(status)
-  assert(status == cudaSuccess);
+  assert(status == gpuSuccess);
 #else
   EIGEN_UNUSED_VARIABLE(config)
 #endif
@@ -337,4 +351,7 @@ static EIGEN_DEVICE_FUNC inline void setCudaSharedMemConfig(cudaSharedMemConfig
 
 }  // end namespace Eigen
 
-#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_CUDA_H
+// undefine all the gpu* macros we defined at the beginning of the file
+#include "TensorGpuHipCudaUndefines.h"
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 8bbe449cc..1181c2753 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -250,28 +250,17 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
   TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
   const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
   if (needs_assign) {
-#if defined(EIGEN_HIPCC)
-    const int block_size = device.maxHipThreadsPerBlock();
-    const int max_blocks = device.getNumHipMultiProcessors() *
-                           device.maxHipThreadsPerMultiProcessor() / block_size;
-    const Index size = array_prod(evaluator.dimensions());
-    // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
-    const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
- 
-    hipLaunchKernelGGL(HIP_KERNEL_NAME(EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),
-        dim3(num_blocks), dim3(block_size), 0, device.stream(), evaluator, size);
-#else
-    const int block_size = device.maxCudaThreadsPerBlock();
-    const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / block_size;
+
+    const int block_size = device.maxGpuThreadsPerBlock();
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
     const Index size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
     const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
 
-    LAUNCH_CUDA_KERNEL(
+    LAUNCH_GPU_KERNEL(
         (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),
         num_blocks, block_size, 0, device, evaluator, size);
-#endif
   }
   evaluator.cleanup();
 }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
new file mode 100644
index 000000000..f009ae855
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
@@ -0,0 +1,86 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Deven Desai <deven.desai.amd@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H)
+#define EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
+
+// Note that we are using EIGEN_USE_HIP here instead of EIGEN_HIPCC...this is by design
+// There is code in the Tensorflow codebase that will define EIGEN_USE_GPU,  but
+// for some reason gets sent to the gcc/host compiler instead of the gpu/nvcc/hipcc compiler
+// When compiling such files, gcc will end up trying to pick up the CUDA headers by 
+// default (see the code within "unsupported/Eigen/CXX11/Tensor" that is guarded by EIGEN_USE_GPU)
+// This will obsviously not work when trying to compile tensorflow on a sytem with no CUDA
+// To work around this issue for HIP systems (and leave the default behaviour intact), the
+// HIP tensorflow build defines EIGEN_USE_HIP when compiling all source files, and 
+// "unsupported/Eigen/CXX11/Tensor" has been updated to use HIP header when EIGEN_USE_HIP is
+// defined. In continuation of that requirement, the guard here needs to be EIGEN_USE_HIP as well
+
+#if defined(EIGEN_USE_HIP)
+
+#define gpuStream_t hipStream_t
+#define gpuDeviceProp_t hipDeviceProp_t
+#define gpuError_t hipError_t
+#define gpuSuccess hipSuccess
+#define gpuErrorNotReady hipErrorNotReady
+#define gpuGetDeviceCount hipGetDeviceCount
+#define gpuGetErrorString hipGetErrorString
+#define gpuGetDeviceProperties hipGetDeviceProperties
+// FIXME : use hipStreamDefault instead of 0x00
+#define gpuStreamDefault 0x00
+#define gpuGetDevice hipGetDevice
+#define gpuSetDevice hipSetDevice
+#define gpuMalloc hipMalloc
+#define gpuFree hipFree
+#define gpuMemsetAsync hipMemsetAsync
+#define gpuMemcpyAsync hipMemcpyAsync
+#define gpuMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define gpuMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define gpuMemcpyHostToDevice hipMemcpyHostToDevice
+#define gpuStreamQuery hipStreamQuery
+#define gpuSharedMemConfig hipSharedMemConfig
+#define gpuDeviceSetSharedMemConfig hipDeviceSetSharedMemConfig
+#define gpuStreamSynchronize hipStreamSynchronize
+#define gpuMemcpy hipMemcpy
+
+#else
+
+#define gpuStream_t cudaStream_t
+#define gpuDeviceProp_t cudaDeviceProp
+#define gpuError_t cudaError_t
+#define gpuSuccess cudaSuccess
+#define gpuErrorNotReady cudaErrorNotReady
+#define gpuGetDeviceCount cudaGetDeviceCount
+#define gpuGetErrorString cudaGetErrorString
+#define gpuGetDeviceProperties cudaGetDeviceProperties
+#define gpuStreamDefault cudaStreamDefault
+#define gpuGetDevice cudaGetDevice
+#define gpuSetDevice cudaSetDevice
+#define gpuMalloc cudaMalloc
+#define gpuFree cudaFree
+#define gpuMemsetAsync cudaMemsetAsync
+#define gpuMemcpyAsync cudaMemcpyAsync
+#define gpuMemcpyDeviceToDevice cudaMemcpyDeviceToDevice
+#define gpuMemcpyDeviceToHost cudaMemcpyDeviceToHost
+#define gpuMemcpyHostToDevice cudaMemcpyHostToDevice
+#define gpuStreamQuery cudaStreamQuery
+#define gpuSharedMemConfig cudaSharedMemConfig
+#define gpuDeviceSetSharedMemConfig cudaDeviceSetSharedMemConfig
+#define gpuStreamSynchronize cudaStreamSynchronize
+#define gpuMemcpy cudaMemcpy
+
+#endif
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+// HIPCC does not support the use of assert on the GPU side.
+#undef assert
+#define assert(COND)
+#endif
+
+#endif  // EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
new file mode 100644
index 000000000..9bc0708ed
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
@@ -0,0 +1,39 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Deven Desai <deven.desai.amd@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H)
+
+#undef gpuStream_t
+#undef gpuDeviceProp_t 
+#undef gpuError_t
+#undef gpuSuccess
+#undef gpuErrorNotReady
+#undef gpuGetDeviceCount
+#undef gpuGetErrorString
+#undef gpuGetDeviceProperties
+#undef gpuStreamDefault
+#undef gpuGetDevice
+#undef gpuSetDevice
+#undef gpuMalloc
+#undef gpuFree
+#undef gpuMemsetAsync
+#undef gpuMemcpyAsync
+#undef gpuMemcpyDeviceToDevice
+#undef gpuMemcpyDeviceToHost
+#undef gpuMemcpyHostToDevice
+#undef gpuStreamQuery
+#undef gpuSharedMemConfig
+#undef gpuDeviceSetSharedMemConfig
+#undef gpuStreamSynchronize
+#undef gpuMemcpy
+
+#undef EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
+
+#endif // EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
index ebcbd6f41..ca854d670 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
@@ -7,23 +7,23 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_CUDA_H
-#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_CUDA_H
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
 
 namespace Eigen {
 namespace internal {
 
 
-#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
 // Full reducers for GPU, don't vectorize for now
 
-// Reducer function that enables multiple cuda thread to safely accumulate at the same
+// Reducer function that enables multiple gpu thread to safely accumulate at the same
 // output address. It basically reads the current value of the output variable, and
-// attempts to update it with the new value. If in the meantime another cuda thread
+// attempts to update it with the new value. If in the meantime another gpu thread
 // updated the content of the output address it will try again.
 template <typename T, typename R>
 __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) {
-#if EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
   if (sizeof(T) == 4)
   {
     unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
@@ -79,7 +79,7 @@ __device__ inline double atomicExchCustom(double* address, double val) {
   return __longlong_as_double(atomicExch(address_as_ull, __double_as_longlong(val)));
 }
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 template <template <typename T> class R>
 __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer) {
   unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
@@ -98,11 +98,11 @@ __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer
     }
   }
 }
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
 template <>
 __device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
-#if EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
   atomicAdd(output, accum);
 #else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
@@ -124,7 +124,7 @@ template <int BlockSize, int NumPerThread, typename Self,
           typename Reducer, typename Index>
 __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num_coeffs,
                                     typename Self::CoeffReturnType* output, unsigned int* semaphore) {
-#if EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
   // Initialize the output value
   const Index first_index = blockIdx.x * BlockSize * NumPerThread + threadIdx.x;
   if (gridDim.x == 1) {
@@ -168,7 +168,14 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
 
 #pragma unroll
   for (int offset = warpSize/2; offset > 0; offset /= 2) {
-  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+  #if defined(EIGEN_HIPCC)
+    // XXX use std::is_floating_point to determine the type of accum
+    if (std::is_floating_point<typename Self::CoeffReturnType>::value) {
+      reducer.reduce(__shfl_down(static_cast<float>(accum), offset, warpSize), &accum);
+    } else {
+      reducer.reduce(__shfl_down(static_cast<int>(accum), offset, warpSize), &accum);
+    }
+  #elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
     reducer.reduce(__shfl_down(accum, offset, warpSize), &accum);
   #else
     reducer.reduce(__shfl_down_sync(0xFFFFFFFF, accum, offset, warpSize), &accum);
@@ -182,6 +189,9 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
   if (gridDim.x > 1 && threadIdx.x == 0) {
     // Let the last block reset the semaphore
     atomicInc(semaphore, gridDim.x + 1);
+#if defined(EIGEN_HIPCC)
+    __threadfence_system();
+#endif
   }
 #else // EIGEN_CUDA_ARCH >= 300
   assert(0 && "Shouldn't be called on unsupported device");
@@ -189,7 +199,7 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
 }
 
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 template <typename Self,
           typename Reducer, typename Index>
 __global__ void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half2* scratch) {
@@ -227,6 +237,21 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
   const Index first_index = blockIdx.x * BlockSize * NumPerThread + 2*threadIdx.x;
 
   // Initialize the output value if it wasn't initialized by the ReductionInitKernel
+
+#if defined(EIGEN_HIPCC)
+  
+  if (gridDim.x == 1 && first_index == 0) {
+    if (num_coeffs % 2 != 0) {
+      half last = input.m_impl.coeff(num_coeffs-1);
+      *scratch = __halves2half2(last, reducer.initialize());
+    } else {
+      *scratch = reducer.template initializePacket<half2>();
+    }
+    __syncthreads();
+  }
+  
+#else
+  
   if (gridDim.x == 1) {
     if (first_index == 0) {
       if (num_coeffs % 2 != 0) {
@@ -238,6 +263,8 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
     }
     __syncthreads();
   }
+  
+#endif
 
   half2 accum = reducer.template initializePacket<half2>();
   const Index max_iter = numext::mini<Index>((num_coeffs - first_index) / 2, NumPerThread*BlockSize / 2);
@@ -250,7 +277,13 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
 
 #pragma unroll
   for (int offset = warpSize/2; offset > 0; offset /= 2) {
-  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+  #if defined(EIGEN_HIPCC)
+    // FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+    union { int i; half2 h; } wka_in, wka_out;
+    wka_in.h = accum;
+    wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
+    reducer.reducePacket(wka_out.h, &accum);
+  #elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
     reducer.reducePacket(__shfl_down(accum, offset, warpSize), &accum);
   #else
     int temp = __shfl_down_sync(0xFFFFFFFF, *(int*)(&accum), (unsigned)offset, warpSize);
@@ -262,6 +295,17 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
     atomicReduce(scratch, accum, reducer);
   }
 
+#if defined(EIGEN_HIPCC)
+  __syncthreads();
+
+  if (gridDim.x == 1 && first_index == 0) {
+    half tmp = __low2half(*scratch);
+    reducer.reduce(__high2half(*scratch), &tmp);
+    *output = tmp;
+  }
+
+#else
+    
   if (gridDim.x == 1) {
     __syncthreads();
     if (first_index == 0) {
@@ -270,6 +314,8 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
       *output = tmp;
     }
   }
+
+#endif
 }
 
 template <typename Op>
@@ -280,7 +326,7 @@ __global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2
   *output = tmp;
 }
 
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct FullReductionLauncher {
@@ -298,6 +344,7 @@ struct FullReductionLauncher<
       internal::is_same<double, OutputType>::value,
     void>::type> {
   static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) {
+
     typedef typename Self::Index Index;
     const int block_size = 256;
     const int num_per_thread = 128;
@@ -308,12 +355,12 @@ struct FullReductionLauncher<
       semaphore = device.semaphore();
     }
 
-    LAUNCH_CUDA_KERNEL((FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
+    LAUNCH_GPU_KERNEL((FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs, output, semaphore);
   }
 };
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 template <typename Self, typename Op>
 struct FullReductionLauncher<Self, Op, Eigen::half, false> {
   static void run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index) {
@@ -334,20 +381,20 @@ struct FullReductionLauncher<Self, Op, Eigen::half, true> {
     if (num_blocks > 1) {
       // We initialize the output and the scrathpad outside the reduction kernel when we can't be sure that there
       // won't be a race conditions between multiple thread blocks.
-      LAUNCH_CUDA_KERNEL((ReductionInitFullReduxKernelHalfFloat<Self, Op, Index>),
+      LAUNCH_GPU_KERNEL((ReductionInitFullReduxKernelHalfFloat<Self, Op, Index>),
                          1, 1, 0, device, reducer, self, num_coeffs, scratch);
     }
 
-    LAUNCH_CUDA_KERNEL((FullReductionKernelHalfFloat<block_size, num_per_thread, Self, Op, Index>),
+    LAUNCH_GPU_KERNEL((FullReductionKernelHalfFloat<block_size, num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs, output, scratch);
 
     if (num_blocks > 1) {
-      LAUNCH_CUDA_KERNEL((ReductionCleanupKernelHalfFloat<Op>),
+      LAUNCH_GPU_KERNEL((ReductionCleanupKernelHalfFloat<Op>),
                          1, 1, 0, device, reducer, output, scratch);
     }
   }
 };
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
 
 template <typename Self, typename Op, bool Vectorizable>
@@ -355,16 +402,16 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
   // Unfortunately nvidia doesn't support well exotic types such as complex,
   // so reduce the scope of the optimized version of the code to the simple cases
   // of doubles, floats and half floats
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else // EIGEN_HAS_CUDA_FP16
+#else // EIGEN_HAS_GPU_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
   template <typename OutputType>
   static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
@@ -384,7 +431,7 @@ template <int NumPerThread, typename Self,
           typename Reducer, typename Index>
 __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
                                          typename Self::CoeffReturnType* output) {
-#if EIGEN_CUDA_ARCH >= 300
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
   typedef typename Self::CoeffReturnType Type;
   eigen_assert(blockDim.y == 1);
   eigen_assert(blockDim.z == 1);
@@ -437,7 +484,14 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
 
 #pragma unroll
       for (int offset = warpSize/2; offset > 0; offset /= 2) {
-      #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+      #if defined(EIGEN_HIPCC)
+        // XXX use std::is_floating_point to determine the type of reduced_val
+        if (std::is_floating_point<Type>::value) {
+          reducer.reduce(__shfl_down(static_cast<float>(reduced_val), offset), &reduced_val);
+        } else {
+          reducer.reduce(__shfl_down(static_cast<int>(reduced_val), offset), &reduced_val);
+        }
+      #elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
         reducer.reduce(__shfl_down(reduced_val, offset), &reduced_val);
       #else
         reducer.reduce(__shfl_down_sync(0xFFFFFFFF, reduced_val, offset), &reduced_val);
@@ -454,7 +508,7 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
 #endif // EIGEN_CUDA_ARCH >= 300
 }
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 
 template <int NumPerThread, typename Self,
           typename Reducer, typename Index>
@@ -531,7 +585,18 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
 
 #pragma unroll
       for (int offset = warpSize/2; offset > 0; offset /= 2) {
-      #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+      #if defined(EIGEN_HIPCC)
+	// FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+	union { int i; half2 h; } wka_in, wka_out;
+
+	wka_in.h = reduced_val1;
+	wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
+        reducer.reducePacket(wka_out.h, &reduced_val1);
+	
+	wka_in.h = reduced_val2;
+	wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
+        reducer.reducePacket(wka_out.h, &reduced_val2);
+      #elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
         reducer.reducePacket(__shfl_down(reduced_val1, offset, warpSize), &reduced_val1);
         reducer.reducePacket(__shfl_down(reduced_val2, offset, warpSize), &reduced_val2);
       #else
@@ -556,7 +621,7 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
   }
 }
 
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct InnerReductionLauncher {
@@ -581,30 +646,30 @@ struct InnerReductionLauncher<
     const int block_size = 256;
     const int num_per_thread = 128;
     const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / block_size;
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
     const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
 
     if (num_blocks > 1) {
       // We initialize the outputs outside the reduction kernel when we can't be sure that there
       // won't be a race conditions between multiple thread blocks.
       const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / 1024;
+      const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / 1024;
       const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      LAUNCH_CUDA_KERNEL((ReductionInitKernel<OutputType, Index>),
+      LAUNCH_GPU_KERNEL((ReductionInitKernel<OutputType, Index>),
                          num_blocks, 1024, 0, device, reducer.initialize(),
                          num_preserved_vals, output);
     }
 
-    LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
+    LAUNCH_GPU_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
 
     return false;
   }
 };
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 template <typename Self, typename Op>
 struct InnerReductionLauncher<Self, Op, Eigen::half, false> {
   static bool run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index, typename Self::Index) {
@@ -627,28 +692,28 @@ struct InnerReductionLauncher<Self, Op, Eigen::half, true> {
     const int block_size = /*256*/128;
     const int num_per_thread = /*128*/64;
     const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / block_size;
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
     const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
 
     if (num_blocks > 1) {
       // We initialize the outputs outside the reduction kernel when we can't be sure that there
       // won't be a race conditions between multiple thread blocks.
       const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / 1024;
+      const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / 1024;
       const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      LAUNCH_CUDA_KERNEL((ReductionInitKernelHalfFloat<Self, Op, Index>),
+      LAUNCH_GPU_KERNEL((ReductionInitKernelHalfFloat<Self, Op, Index>),
                          1, 1, 0, device, reducer, self, num_preserved_vals, output);
     }
 
-    LAUNCH_CUDA_KERNEL((InnerReductionKernelHalfFloat<num_per_thread, Self, Op, Index>),
+    LAUNCH_GPU_KERNEL((InnerReductionKernelHalfFloat<num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
 
     return false;
   }
 };
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
 
 template <typename Self, typename Op>
@@ -656,16 +721,16 @@ struct InnerReducer<Self, Op, GpuDevice> {
   // Unfortunately nvidia doesn't support well exotic types such as complex,
   // so reduce the scope of the optimized version of the code to the simple case
   // of floats and half floats.
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
-#else // EIGEN_HAS_CUDA_FP16
+#else // EIGEN_HAS_GPU_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                  (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                   internal::is_same<typename Self::CoeffReturnType, double>::value);
-#endif // EIGEN_HAS_CUDA_FP16
+#endif // EIGEN_HAS_GPU_FP16
 
   template <typename OutputType>
   static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
@@ -723,7 +788,20 @@ struct OuterReducer<Self, Op, GpuDevice> {
                                                  (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                   internal::is_same<typename Self::CoeffReturnType, double>::value);
   template <typename Device, typename OutputType>
-  static EIGEN_DEVICE_FUNC bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
+  static
+    #if !defined(EIGEN_HIPCC)
+    // FIXME :  leaving this EIGEN_DEVICE_FUNC in, results in the following runtime error
+    //          (in the cxx11_tensor_reduction_gpu test)
+    //
+    // terminate called after throwing an instance of 'std::runtime_error'
+    //   what():  No device code available for function: _ZN5Eigen8internal20OuterReductionKernelIL...
+    //
+    // dont know why this happens (and why is it a runtime error instead of a compile time errror)
+    //
+    // this will be fixed by HIP PR#457
+    EIGEN_DEVICE_FUNC
+    #endif
+    bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
     assert(false && "Should only be called to reduce doubles or floats on a gpu device");
     return true;
   }
@@ -740,33 +818,37 @@ struct OuterReducer<Self, Op, GpuDevice> {
     const int block_size = 256;
     const int num_per_thread = 16;
     const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
-    const int max_blocks = device.getNumCudaMultiProcessors() *
-                           device.maxCudaThreadsPerMultiProcessor() / block_size;
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
     const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
 
     if (num_blocks > 1) {
       // We initialize the outputs in the reduction kernel itself when we don't have to worry
       // about race conditions between multiple thread blocks.
       const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
-      const int max_blocks = device.getNumCudaMultiProcessors() *
-                             device.maxCudaThreadsPerMultiProcessor() / 1024;
+      const int max_blocks = device.getNumGpuMultiProcessors() *
+                             device.maxGpuThreadsPerMultiProcessor() / 1024;
       const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
-      LAUNCH_CUDA_KERNEL((ReductionInitKernel<float, Index>),
+      LAUNCH_GPU_KERNEL((ReductionInitKernel<float, Index>),
                          num_blocks, 1024, 0, device, reducer.initialize(),
                          num_preserved_vals, output);
     }
 
-    LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread, Self, Op, Index>),
+    LAUNCH_GPU_KERNEL((OuterReductionKernel<num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
 
     return false;
   }
 };
 
-#endif // defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#endif // defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
 
 
 } // end namespace internal
 } // end namespace Eigen
 
-#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_CUDA_H
+#if defined(EIGEN_HIPCC)
+#undef warpSize
+#endif
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 6d68e256f..39717efaa 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -278,12 +278,8 @@ struct ScanLauncher<Self, Reducer, GpuDevice> {
      Index total_size = internal::array_prod(self.dimensions());
      Index num_blocks = (total_size / self.size() + 63) / 64;
      Index block_size = 64;
-#if defined(EIGEN_HIPCC)     
-     hipLaunchKernelGGL(HIP_KERNEL_NAME(ScanKernel<Self, Reducer>), dim3(num_blocks),
-			dim3(block_size), 0, self.device().stream(), self, total_size, data);
-#else
-     LAUNCH_CUDA_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
-#endif     
+
+     LAUNCH_GPU_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
   }
 };
 #endif  // EIGEN_USE_GPU && (EIGEN_GPUCC)
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 05b141e21..97e0669a6 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -274,24 +274,24 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
   cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include")
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu")
 
-  ei_add_test(cxx11_tensor_complex_cuda)
-  ei_add_test(cxx11_tensor_complex_cwise_ops_cuda)
-  ei_add_test(cxx11_tensor_reduction_cuda)
-  ei_add_test(cxx11_tensor_argmax_cuda)
-  ei_add_test(cxx11_tensor_cast_float16_cuda)
-  ei_add_test(cxx11_tensor_scan_cuda)
+  ei_add_test(cxx11_tensor_complex_gpu)
+  ei_add_test(cxx11_tensor_complex_cwise_ops_gpu)
+  ei_add_test(cxx11_tensor_reduction_gpu)
+  ei_add_test(cxx11_tensor_argmax_gpu)
+  ei_add_test(cxx11_tensor_cast_float16_gpu)
+  ei_add_test(cxx11_tensor_scan_gpu)
 
   # Contractions require arch 3.0 or higher
   if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 29)
     ei_add_test(cxx11_tensor_device)
-    ei_add_test(cxx11_tensor_cuda)
-    ei_add_test(cxx11_tensor_contract_cuda)
-    ei_add_test(cxx11_tensor_of_float16_cuda)
+    ei_add_test(cxx11_tensor_gpu)
+    ei_add_test(cxx11_tensor_contract_gpu)
+    ei_add_test(cxx11_tensor_of_float16_gpu)
   endif()
 
   # The random number generation code requires arch 3.5 or greater.
   if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 34)
-    ei_add_test(cxx11_tensor_random_cuda)
+    ei_add_test(cxx11_tensor_random_gpu)
   endif()
 
 
@@ -318,18 +318,23 @@ if (EIGEN_TEST_HIP)
 	include_directories(${HIP_PATH}/include)
 
 	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
+	#
+	# complex datatype is not yet supported by HIP
+	# so leaving out those tests for now
+	#
+	# ei_add_test(cxx11_tensor_complex_gpu)
+	# ei_add_test(cxx11_tensor_complex_cwise_ops_gpu)
+	#
+	ei_add_test(cxx11_tensor_reduction_gpu)
+	ei_add_test(cxx11_tensor_argmax_gpu)
+	ei_add_test(cxx11_tensor_cast_float16_gpu)
+	ei_add_test(cxx11_tensor_scan_gpu)
+	ei_add_test(cxx11_tensor_device)
 
-	# ei_add_test(cxx11_tensor_complex_hip)
-	# ei_add_test(cxx11_tensor_complex_cwise_ops_hip)
-	ei_add_test(cxx11_tensor_reduction_hip)
-	ei_add_test(cxx11_tensor_argmax_hip)
-	ei_add_test(cxx11_tensor_cast_float16_hip)
-	ei_add_test(cxx11_tensor_scan_hip)
-	ei_add_test(cxx11_tensor_device_hip)
-	ei_add_test(cxx11_tensor_hip)
-	ei_add_test(cxx11_tensor_contract_hip)
-	ei_add_test(cxx11_tensor_of_float16_hip)
-	ei_add_test(cxx11_tensor_random_hip)
+	ei_add_test(cxx11_tensor_gpu)
+	ei_add_test(cxx11_tensor_contract_gpu)
+	ei_add_test(cxx11_tensor_of_float16_gpu)
+	ei_add_test(cxx11_tensor_random_gpu)
 	
 	unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 	
diff --git a/unsupported/test/cxx11_tensor_argmax_gpu.cu b/unsupported/test/cxx11_tensor_argmax_gpu.cu
index 3d73d491a..541a27865 100644
--- a/unsupported/test/cxx11_tensor_argmax_gpu.cu
+++ b/unsupported/test/cxx11_tensor_argmax_gpu.cu
@@ -9,16 +9,18 @@
 
 
 #define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_FUNC cxx11_tensor_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_gpu
 #define EIGEN_USE_GPU
 
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
+
 using Eigen::Tensor;
 
 template <int Layout>
-void test_cuda_simple_argmax()
+void test_gpu_simple_argmax()
 {
   Tensor<double, 3, Layout> in(Eigen::array<DenseIndex, 3>(72,53,97));
   Tensor<DenseIndex, 1, Layout> out_max(Eigen::array<DenseIndex, 1>(1));
@@ -34,13 +36,13 @@ void test_cuda_simple_argmax()
   double* d_in;
   DenseIndex* d_out_max;
   DenseIndex* d_out_min;
-  cudaMalloc((void**)(&d_in), in_bytes);
-  cudaMalloc((void**)(&d_out_max), out_bytes);
-  cudaMalloc((void**)(&d_out_min), out_bytes);
+  gpuMalloc((void**)(&d_in), in_bytes);
+  gpuMalloc((void**)(&d_out_max), out_bytes);
+  gpuMalloc((void**)(&d_out_min), out_bytes);
 
-  cudaMemcpy(d_in, in.data(), in_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in.data(), in_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<double, 3, Layout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 3>(72,53,97));
@@ -50,20 +52,20 @@ void test_cuda_simple_argmax()
   gpu_out_max.device(gpu_device) = gpu_in.argmax();
   gpu_out_min.device(gpu_device) = gpu_in.argmin();
 
-  assert(cudaMemcpyAsync(out_max.data(), d_out_max, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaMemcpyAsync(out_min.data(), d_out_min, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out_max.data(), d_out_max, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuMemcpyAsync(out_min.data(), d_out_min, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   VERIFY_IS_EQUAL(out_max(Eigen::array<DenseIndex, 1>(0)), 72*53*97 - 1);
   VERIFY_IS_EQUAL(out_min(Eigen::array<DenseIndex, 1>(0)), 0);
 
-  cudaFree(d_in);
-  cudaFree(d_out_max);
-  cudaFree(d_out_min);
+  gpuFree(d_in);
+  gpuFree(d_out_max);
+  gpuFree(d_out_min);
 }
 
 template <int DataLayout>
-void test_cuda_argmax_dim()
+void test_gpu_argmax_dim()
 {
   Tensor<float, 4, DataLayout> tensor(2,3,5,7);
   std::vector<int> dims;
@@ -97,12 +99,12 @@ void test_cuda_argmax_dim()
 
     float* d_in;
     DenseIndex* d_out;
-    cudaMalloc((void**)(&d_in), in_bytes);
-    cudaMalloc((void**)(&d_out), out_bytes);
+    gpuMalloc((void**)(&d_in), in_bytes);
+    gpuMalloc((void**)(&d_out), out_bytes);
 
-    cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
+    gpuMemcpy(d_in, tensor.data(), in_bytes, gpuMemcpyHostToDevice);
 
-    Eigen::CudaStreamDevice stream;
+    Eigen::GpuStreamDevice stream;
     Eigen::GpuDevice gpu_device(&stream);
 
     Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
@@ -110,8 +112,8 @@ void test_cuda_argmax_dim()
 
     gpu_out.device(gpu_device) = gpu_in.argmax(dim);
 
-    assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-    assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+    assert(gpuMemcpyAsync(tensor_arg.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+    assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
     VERIFY_IS_EQUAL(tensor_arg.size(),
                     size_t(2*3*5*7 / tensor.dimension(dim)));
@@ -134,25 +136,25 @@ void test_cuda_argmax_dim()
       }
     }
 
-    cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
+    gpuMemcpy(d_in, tensor.data(), in_bytes, gpuMemcpyHostToDevice);
 
     gpu_out.device(gpu_device) = gpu_in.argmax(dim);
 
-    assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-    assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+    assert(gpuMemcpyAsync(tensor_arg.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+    assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
     for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
       // Expect max to be in the last index of the reduced dimension
       VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
     }
 
-    cudaFree(d_in);
-    cudaFree(d_out);
+    gpuFree(d_in);
+    gpuFree(d_out);
   }
 }
 
 template <int DataLayout>
-void test_cuda_argmin_dim()
+void test_gpu_argmin_dim()
 {
   Tensor<float, 4, DataLayout> tensor(2,3,5,7);
   std::vector<int> dims;
@@ -186,12 +188,12 @@ void test_cuda_argmin_dim()
 
     float* d_in;
     DenseIndex* d_out;
-    cudaMalloc((void**)(&d_in), in_bytes);
-    cudaMalloc((void**)(&d_out), out_bytes);
+    gpuMalloc((void**)(&d_in), in_bytes);
+    gpuMalloc((void**)(&d_out), out_bytes);
 
-    cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
+    gpuMemcpy(d_in, tensor.data(), in_bytes, gpuMemcpyHostToDevice);
 
-    Eigen::CudaStreamDevice stream;
+    Eigen::GpuStreamDevice stream;
     Eigen::GpuDevice gpu_device(&stream);
 
     Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
@@ -199,8 +201,8 @@ void test_cuda_argmin_dim()
 
     gpu_out.device(gpu_device) = gpu_in.argmin(dim);
 
-    assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-    assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+    assert(gpuMemcpyAsync(tensor_arg.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+    assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
     VERIFY_IS_EQUAL(tensor_arg.size(),
                     2*3*5*7 / tensor.dimension(dim));
@@ -223,29 +225,29 @@ void test_cuda_argmin_dim()
       }
     }
 
-    cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
+    gpuMemcpy(d_in, tensor.data(), in_bytes, gpuMemcpyHostToDevice);
 
     gpu_out.device(gpu_device) = gpu_in.argmin(dim);
 
-    assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-    assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+    assert(gpuMemcpyAsync(tensor_arg.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+    assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
     for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
       // Expect max to be in the last index of the reduced dimension
       VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
     }
 
-    cudaFree(d_in);
-    cudaFree(d_out);
+    gpuFree(d_in);
+    gpuFree(d_out);
   }
 }
 
-void test_cxx11_tensor_cuda()
+void test_cxx11_tensor_gpu()
 {
-  CALL_SUBTEST_1(test_cuda_simple_argmax<RowMajor>());
-  CALL_SUBTEST_1(test_cuda_simple_argmax<ColMajor>());
-  CALL_SUBTEST_2(test_cuda_argmax_dim<RowMajor>());
-  CALL_SUBTEST_2(test_cuda_argmax_dim<ColMajor>());
-  CALL_SUBTEST_3(test_cuda_argmin_dim<RowMajor>());
-  CALL_SUBTEST_3(test_cuda_argmin_dim<ColMajor>());
+  CALL_SUBTEST_1(test_gpu_simple_argmax<RowMajor>());
+  CALL_SUBTEST_1(test_gpu_simple_argmax<ColMajor>());
+  CALL_SUBTEST_2(test_gpu_argmax_dim<RowMajor>());
+  CALL_SUBTEST_2(test_gpu_argmax_dim<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_argmin_dim<RowMajor>());
+  CALL_SUBTEST_3(test_gpu_argmin_dim<ColMajor>());
 }
diff --git a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
index 816e03220..a2928b0b3 100644
--- a/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_cast_float16_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_gpu
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
@@ -18,8 +18,8 @@
 
 using Eigen::Tensor;
 
-void test_cuda_conversion() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_conversion() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -72,8 +72,8 @@ void test_fallback_conversion() {
 }
 
 
-void test_cxx11_tensor_cast_float16_cuda()
+void test_cxx11_tensor_cast_float16_gpu()
 {
-  CALL_SUBTEST(test_cuda_conversion());
+  CALL_SUBTEST(test_gpu_conversion());
   CALL_SUBTEST(test_fallback_conversion());
 }
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
index aac780905..af67348aa 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
@@ -28,7 +28,7 @@ void test_cuda_complex_cwise_ops() {
   cudaMalloc((void**)(&d_in2), complex_bytes);
   cudaMalloc((void**)(&d_out), complex_bytes);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<std::complex<T>, 1, 0, int>, Eigen::Aligned> gpu_in1(
diff --git a/unsupported/test/cxx11_tensor_complex_gpu.cu b/unsupported/test/cxx11_tensor_complex_gpu.cu
index a52350f85..45b49d266 100644
--- a/unsupported/test/cxx11_tensor_complex_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_gpu.cu
@@ -34,7 +34,7 @@ void test_cuda_nullary() {
   cudaMemcpy(d_in1, in1.data(), complex_bytes, cudaMemcpyHostToDevice);
   cudaMemcpy(d_in2, in2.data(), complex_bytes, cudaMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<std::complex<float>, 1, 0, int>, Eigen::Aligned> gpu_in1(
@@ -70,7 +70,7 @@ void test_cuda_nullary() {
 
 static void test_cuda_sum_reductions() {
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   const int num_rows = internal::random<int>(1024, 5*1024);
@@ -106,7 +106,7 @@ static void test_cuda_sum_reductions() {
 
 static void test_cuda_mean_reductions() {
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   const int num_rows = internal::random<int>(1024, 5*1024);
@@ -142,7 +142,7 @@ static void test_cuda_mean_reductions() {
 
 static void test_cuda_product_reductions() {
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   const int num_rows = internal::random<int>(1024, 5*1024);
diff --git a/unsupported/test/cxx11_tensor_contract_gpu.cu b/unsupported/test/cxx11_tensor_contract_gpu.cu
index 3621e2aa6..061d0464e 100644
--- a/unsupported/test/cxx11_tensor_contract_gpu.cu
+++ b/unsupported/test/cxx11_tensor_contract_gpu.cu
@@ -10,19 +10,20 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_gpu
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
 
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;
 
 template<int DataLayout>
-void test_cuda_contraction(int m_size, int k_size, int n_size)
+void test_gpu_contraction(int m_size, int k_size, int n_size)
 {
   std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
   // with these dimensions, the output has 300 * 140 elements, which is
@@ -45,14 +46,14 @@ void test_cuda_contraction(int m_size, int k_size, int n_size)
   float* d_t_right;
   float* d_t_result;
 
-  cudaMalloc((void**)(&d_t_left), t_left_bytes);
-  cudaMalloc((void**)(&d_t_right), t_right_bytes);
-  cudaMalloc((void**)(&d_t_result), t_result_bytes);
+  gpuMalloc((void**)(&d_t_left), t_left_bytes);
+  gpuMalloc((void**)(&d_t_right), t_right_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
 
-  cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_t_left, t_left.data(), t_left_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_t_right, t_right.data(), t_right_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
@@ -66,7 +67,7 @@ void test_cuda_contraction(int m_size, int k_size, int n_size)
   gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
   t_result = t_left.contract(t_right, dims);
 
-  cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
+  gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
   for (DenseIndex i = 0; i < t_result.size(); i++) {
     if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
       continue;
@@ -79,9 +80,9 @@ void test_cuda_contraction(int m_size, int k_size, int n_size)
     assert(false);
   }
 
-  cudaFree((void*)d_t_left);
-  cudaFree((void*)d_t_right);
-  cudaFree((void*)d_t_result);
+  gpuFree((void*)d_t_left);
+  gpuFree((void*)d_t_right);
+  gpuFree((void*)d_t_result);
 }
 
 
@@ -109,14 +110,14 @@ void test_scalar(int m_size, int k_size, int n_size)
   float* d_t_right;
   float* d_t_result;
 
-  cudaMalloc((void**)(&d_t_left), t_left_bytes);
-  cudaMalloc((void**)(&d_t_right), t_right_bytes);
-  cudaMalloc((void**)(&d_t_result), t_result_bytes);
+  gpuMalloc((void**)(&d_t_left), t_left_bytes);
+  gpuMalloc((void**)(&d_t_right), t_right_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
 
-  cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_t_left, t_left.data(), t_left_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_t_right, t_right.data(), t_right_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
@@ -129,7 +130,7 @@ void test_scalar(int m_size, int k_size, int n_size)
   gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
   t_result = t_left.contract(t_right, dims);
 
-  cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
+  gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
   if (fabs(t_result() - t_result_gpu()) > 1e-4f &&
       !Eigen::internal::isApprox(t_result(), t_result_gpu(), 1e-4f)) {
     std::cout << "mismatch detected: " << t_result()
@@ -137,39 +138,39 @@ void test_scalar(int m_size, int k_size, int n_size)
     assert(false);
   }
 
-  cudaFree((void*)d_t_left);
-  cudaFree((void*)d_t_right);
-  cudaFree((void*)d_t_result);
+  gpuFree((void*)d_t_left);
+  gpuFree((void*)d_t_right);
+  gpuFree((void*)d_t_result);
 }
 
 
 template<int DataLayout>
-void test_cuda_contraction_m() {
+void test_gpu_contraction_m() {
   for (int k = 32; k < 256; k++) {
-    test_cuda_contraction<ColMajor>(k, 128, 128);
-    test_cuda_contraction<RowMajor>(k, 128, 128);
+    test_gpu_contraction<ColMajor>(k, 128, 128);
+    test_gpu_contraction<RowMajor>(k, 128, 128);
   }
 }
 
 template<int DataLayout>
-void test_cuda_contraction_k() {
+void test_gpu_contraction_k() {
   for (int k = 32; k < 256; k++) {
-    test_cuda_contraction<ColMajor>(128, k, 128);
-    test_cuda_contraction<RowMajor>(128, k, 128);
+    test_gpu_contraction<ColMajor>(128, k, 128);
+    test_gpu_contraction<RowMajor>(128, k, 128);
   }
 }
 
 template<int DataLayout>
-void test_cuda_contraction_n() {
+void test_gpu_contraction_n() {
   for (int k = 32; k < 256; k++) {
-    test_cuda_contraction<ColMajor>(128, 128, k);
-    test_cuda_contraction<RowMajor>(128, 128, k);
+    test_gpu_contraction<ColMajor>(128, 128, k);
+    test_gpu_contraction<RowMajor>(128, 128, k);
   }
 }
 
 
 template<int DataLayout>
-void test_cuda_contraction_sizes() {
+void test_gpu_contraction_sizes() {
   int m_sizes[] = { 31,  39,   63,   64,   65,
                    127, 129,  255,  257 , 511,
                    512, 513, 1023, 1024, 1025};
@@ -186,29 +187,32 @@ void test_cuda_contraction_sizes() {
   for (int i = 0; i < 15; i++) {
     for (int j = 0; j < 15; j++) {
       for (int k = 0; k < 17; k++) {
-        test_cuda_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
+        test_gpu_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
       }
     }
   }
 }
 
-void test_cxx11_tensor_cuda()
+void test_cxx11_tensor_gpu()
 {
-  CALL_SUBTEST_1(test_cuda_contraction<ColMajor>(128, 128, 128));
-  CALL_SUBTEST_1(test_cuda_contraction<RowMajor>(128, 128, 128));
+  CALL_SUBTEST_1(test_gpu_contraction<ColMajor>(128, 128, 128));
+  CALL_SUBTEST_1(test_gpu_contraction<RowMajor>(128, 128, 128));
 
   CALL_SUBTEST_1(test_scalar<ColMajor>(128, 128, 128));
   CALL_SUBTEST_1(test_scalar<RowMajor>(128, 128, 128));
 
-  CALL_SUBTEST_2(test_cuda_contraction_m<ColMajor>());
-  CALL_SUBTEST_3(test_cuda_contraction_m<RowMajor>());
+  CALL_SUBTEST_2(test_gpu_contraction_m<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_contraction_m<RowMajor>());
 
-  CALL_SUBTEST_4(test_cuda_contraction_k<ColMajor>());
-  CALL_SUBTEST_5(test_cuda_contraction_k<RowMajor>());
+  CALL_SUBTEST_4(test_gpu_contraction_k<ColMajor>());
+  CALL_SUBTEST_5(test_gpu_contraction_k<RowMajor>());
 
-  CALL_SUBTEST_6(test_cuda_contraction_n<ColMajor>());
-  CALL_SUBTEST_7(test_cuda_contraction_n<RowMajor>());
+  CALL_SUBTEST_6(test_gpu_contraction_n<ColMajor>());
+  CALL_SUBTEST_7(test_gpu_contraction_n<RowMajor>());
 
-  CALL_SUBTEST_8(test_cuda_contraction_sizes<ColMajor>());
-  CALL_SUBTEST_9(test_cuda_contraction_sizes<RowMajor>());
+#if !defined(EIGEN_USE_HIP)
+// disable these subtests for HIP
+  CALL_SUBTEST_8(test_gpu_contraction_sizes<ColMajor>());
+  CALL_SUBTEST_9(test_gpu_contraction_sizes<RowMajor>());
+#endif	
 }
diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu
index 7c14bc187..52215fc39 100644
--- a/unsupported/test/cxx11_tensor_device.cu
+++ b/unsupported/test/cxx11_tensor_device.cu
@@ -16,6 +16,7 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
 
 using Eigen::Tensor;
 using Eigen::RowMajor;
@@ -66,22 +67,22 @@ struct CPUContext {
 // Context for evaluation on GPU
 struct GPUContext {
   GPUContext(const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1, Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2, Eigen::TensorMap<Eigen::Tensor<float, 3> >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) {
-    assert(cudaMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == cudaSuccess);
+    assert(gpuMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == gpuSuccess);
     float kernel_1d_val[] = {3.14f, 2.7f};
-    assert(cudaMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
+    assert(gpuMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), gpuMemcpyHostToDevice) == gpuSuccess);
 
-    assert(cudaMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == cudaSuccess);
+    assert(gpuMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == gpuSuccess);
     float kernel_2d_val[] = {3.14f, 2.7f, 0.2f, 7.0f};
-    assert(cudaMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
+    assert(gpuMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), gpuMemcpyHostToDevice) == gpuSuccess);
 
-    assert(cudaMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == cudaSuccess);
+    assert(gpuMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == gpuSuccess);
     float kernel_3d_val[] = {3.14f, -1.0f, 2.7f, -0.3f, 0.2f, -0.7f, 7.0f, -0.5f};
-    assert(cudaMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
+    assert(gpuMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), gpuMemcpyHostToDevice) == gpuSuccess);
   }
   ~GPUContext() {
-    assert(cudaFree(kernel_1d_) == cudaSuccess);
-    assert(cudaFree(kernel_2d_) == cudaSuccess);
-    assert(cudaFree(kernel_3d_) == cudaSuccess);
+    assert(gpuFree(kernel_1d_) == gpuSuccess);
+    assert(gpuFree(kernel_2d_) == gpuSuccess);
+    assert(gpuFree(kernel_3d_) == gpuSuccess);
   }
 
   const Eigen::GpuDevice& device() const { return gpu_device_; }
@@ -102,7 +103,7 @@ struct GPUContext {
   float* kernel_2d_;
   float* kernel_3d_;
 
-  Eigen::CudaStreamDevice stream_;
+  Eigen::GpuStreamDevice stream_;
   Eigen::GpuDevice gpu_device_;
 };
 
@@ -281,12 +282,12 @@ void test_gpu() {
   float* d_in1;
   float* d_in2;
   float* d_out;
-  cudaMalloc((void**)(&d_in1), in1_bytes);
-  cudaMalloc((void**)(&d_in2), in2_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in1), in1_bytes);
+  gpuMalloc((void**)(&d_in2), in2_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in1, in1.data(), in1_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in2, in2.data(), in2_bytes, gpuMemcpyHostToDevice);
 
   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, 40,50,70);
   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, 40,50,70);
@@ -294,7 +295,7 @@ void test_gpu() {
 
   GPUContext context(gpu_in1, gpu_in2, gpu_out);
   test_contextual_eval(&context);
-  assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
+  assert(gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 50; ++j) {
       for (int k = 0; k < 70; ++k) {
@@ -304,7 +305,7 @@ void test_gpu() {
   }
 
   test_forced_contextual_eval(&context);
-  assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
+  assert(gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 50; ++j) {
       for (int k = 0; k < 70; ++k) {
@@ -314,7 +315,7 @@ void test_gpu() {
   }
 
   test_compound_assignment(&context);
-  assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
+  assert(gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 50; ++j) {
       for (int k = 0; k < 70; ++k) {
@@ -324,7 +325,7 @@ void test_gpu() {
   }
 
   test_contraction(&context);
-  assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
+  assert(gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 40; ++j) {
       const float result = out(i,j,0);
@@ -339,8 +340,8 @@ void test_gpu() {
   }
 
   test_1d_convolution(&context);
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, context.device().stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(context.device().stream()) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 49; ++j) {
       for (int k = 0; k < 70; ++k) {
@@ -350,8 +351,8 @@ void test_gpu() {
   }
 
   test_2d_convolution(&context);
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, context.device().stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(context.device().stream()) == gpuSuccess);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 49; ++j) {
       for (int k = 0; k < 69; ++k) {
@@ -363,9 +364,13 @@ void test_gpu() {
     }
   }
 
+#if !defined(EIGEN_USE_HIP)
+// disable this test on the HIP platform
+// 3D tensor convolutions seem to hang on the HIP platform
+
   test_3d_convolution(&context);
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, context.device().stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(context.device().stream()) == gpuSuccess);
   for (int i = 0; i < 39; ++i) {
     for (int j = 0; j < 49; ++j) {
       for (int k = 0; k < 69; ++k) {
@@ -378,6 +383,9 @@ void test_gpu() {
       }
     }
   }
+
+#endif
+ 
 }
 
 
diff --git a/unsupported/test/cxx11_tensor_gpu.cu b/unsupported/test/cxx11_tensor_gpu.cu
index f238ed5be..285441182 100644
--- a/unsupported/test/cxx11_tensor_gpu.cu
+++ b/unsupported/test/cxx11_tensor_gpu.cu
@@ -9,15 +9,17 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_gpu
 #define EIGEN_USE_GPU
 
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
+
 using Eigen::Tensor;
 
-void test_cuda_nullary() {
+void test_gpu_nullary() {
   Tensor<float, 1, 0, int> in1(2);
   Tensor<float, 1, 0, int> in2(2);
   in1.setRandom();
@@ -27,12 +29,12 @@ void test_cuda_nullary() {
 
   float* d_in1;
   float* d_in2;
-  cudaMalloc((void**)(&d_in1), tensor_bytes);
-  cudaMalloc((void**)(&d_in2), tensor_bytes);
-  cudaMemcpy(d_in1, in1.data(), tensor_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in2, in2.data(), tensor_bytes, cudaMemcpyHostToDevice);
+  gpuMalloc((void**)(&d_in1), tensor_bytes);
+  gpuMalloc((void**)(&d_in2), tensor_bytes);
+  gpuMemcpy(d_in1, in1.data(), tensor_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in2, in2.data(), tensor_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in1(
@@ -46,23 +48,23 @@ void test_cuda_nullary() {
   Tensor<float, 1, 0, int> new1(2);
   Tensor<float, 1, 0, int> new2(2);
 
-  assert(cudaMemcpyAsync(new1.data(), d_in1, tensor_bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaMemcpyAsync(new2.data(), d_in2, tensor_bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(new1.data(), d_in1, tensor_bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuMemcpyAsync(new2.data(), d_in2, tensor_bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
 
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 2; ++i) {
     VERIFY_IS_APPROX(new1(i), 3.14f);
     VERIFY_IS_NOT_EQUAL(new2(i), in2(i));
   }
 
-  cudaFree(d_in1);
-  cudaFree(d_in2);
+  gpuFree(d_in1);
+  gpuFree(d_in2);
 }
 
-void test_cuda_elementwise_small() {
+void test_gpu_elementwise_small() {
   Tensor<float, 1> in1(Eigen::array<Eigen::DenseIndex, 1>(2));
   Tensor<float, 1> in2(Eigen::array<Eigen::DenseIndex, 1>(2));
   Tensor<float, 1> out(Eigen::array<Eigen::DenseIndex, 1>(2));
@@ -76,14 +78,14 @@ void test_cuda_elementwise_small() {
   float* d_in1;
   float* d_in2;
   float* d_out;
-  cudaMalloc((void**)(&d_in1), in1_bytes);
-  cudaMalloc((void**)(&d_in2), in2_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in1), in1_bytes);
+  gpuMalloc((void**)(&d_in2), in2_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in1, in1.data(), in1_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in2, in2.data(), in2_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
@@ -95,9 +97,9 @@ void test_cuda_elementwise_small() {
 
   gpu_out.device(gpu_device) = gpu_in1 + gpu_in2;
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 2; ++i) {
     VERIFY_IS_APPROX(
@@ -105,12 +107,12 @@ void test_cuda_elementwise_small() {
         in1(Eigen::array<Eigen::DenseIndex, 1>(i)) + in2(Eigen::array<Eigen::DenseIndex, 1>(i)));
   }
 
-  cudaFree(d_in1);
-  cudaFree(d_in2);
-  cudaFree(d_out);
+  gpuFree(d_in1);
+  gpuFree(d_in2);
+  gpuFree(d_out);
 }
 
-void test_cuda_elementwise()
+void test_gpu_elementwise()
 {
   Tensor<float, 3> in1(Eigen::array<Eigen::DenseIndex, 3>(72,53,97));
   Tensor<float, 3> in2(Eigen::array<Eigen::DenseIndex, 3>(72,53,97));
@@ -129,16 +131,16 @@ void test_cuda_elementwise()
   float* d_in2;
   float* d_in3;
   float* d_out;
-  cudaMalloc((void**)(&d_in1), in1_bytes);
-  cudaMalloc((void**)(&d_in2), in2_bytes);
-  cudaMalloc((void**)(&d_in3), in3_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in1), in1_bytes);
+  gpuMalloc((void**)(&d_in2), in2_bytes);
+  gpuMalloc((void**)(&d_in3), in3_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in3, in3.data(), in3_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in1, in1.data(), in1_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in2, in2.data(), in2_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in3, in3.data(), in3_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, Eigen::array<Eigen::DenseIndex, 3>(72,53,97));
@@ -148,8 +150,8 @@ void test_cuda_elementwise()
 
   gpu_out.device(gpu_device) = gpu_in1 + gpu_in2 * gpu_in3;
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 72; ++i) {
     for (int j = 0; j < 53; ++j) {
@@ -159,13 +161,13 @@ void test_cuda_elementwise()
     }
   }
 
-  cudaFree(d_in1);
-  cudaFree(d_in2);
-  cudaFree(d_in3);
-  cudaFree(d_out);
+  gpuFree(d_in1);
+  gpuFree(d_in2);
+  gpuFree(d_in3);
+  gpuFree(d_out);
 }
 
-void test_cuda_props() {
+void test_gpu_props() {
   Tensor<float, 1> in1(200);
   Tensor<bool, 1> out(200);
   in1.setRandom();
@@ -175,12 +177,12 @@ void test_cuda_props() {
 
   float* d_in1;
   bool* d_out;
-  cudaMalloc((void**)(&d_in1), in1_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in1), in1_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in1, in1.data(), in1_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1(
@@ -190,19 +192,19 @@ void test_cuda_props() {
 
   gpu_out.device(gpu_device) = (gpu_in1.isnan)();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 200; ++i) {
     VERIFY_IS_EQUAL(out(i), (std::isnan)(in1(i)));
   }
 
-  cudaFree(d_in1);
-  cudaFree(d_out);
+  gpuFree(d_in1);
+  gpuFree(d_out);
 }
 
-void test_cuda_reduction()
+void test_gpu_reduction()
 {
   Tensor<float, 4> in1(72,53,97,113);
   Tensor<float, 2> out(72,97);
@@ -213,12 +215,12 @@ void test_cuda_reduction()
 
   float* d_in1;
   float* d_out;
-  cudaMalloc((void**)(&d_in1), in1_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in1), in1_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in1, in1.data(), in1_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4> > gpu_in1(d_in1, 72,53,97,113);
@@ -230,8 +232,8 @@ void test_cuda_reduction()
 
   gpu_out.device(gpu_device) = gpu_in1.maximum(reduction_axis);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 72; ++i) {
     for (int j = 0; j < 97; ++j) {
@@ -246,12 +248,12 @@ void test_cuda_reduction()
     }
   }
 
-  cudaFree(d_in1);
-  cudaFree(d_out);
+  gpuFree(d_in1);
+  gpuFree(d_out);
 }
 
 template<int DataLayout>
-void test_cuda_contraction()
+void test_gpu_contraction()
 {
   // with these dimensions, the output has 300 * 140 elements, which is
   // more than 30 * 1024, which is the number of threads in blocks on
@@ -271,14 +273,14 @@ void test_cuda_contraction()
   float* d_t_right;
   float* d_t_result;
 
-  cudaMalloc((void**)(&d_t_left), t_left_bytes);
-  cudaMalloc((void**)(&d_t_right), t_right_bytes);
-  cudaMalloc((void**)(&d_t_result), t_result_bytes);
+  gpuMalloc((void**)(&d_t_left), t_left_bytes);
+  gpuMalloc((void**)(&d_t_right), t_right_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
 
-  cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_t_left, t_left.data(), t_left_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_t_right, t_right.data(), t_right_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_t_left(d_t_left, 6, 50, 3, 31);
@@ -298,7 +300,7 @@ void test_cuda_contraction()
   m_result = m_left * m_right;
   gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
 
-  cudaMemcpy(t_result.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
+  gpuMemcpy(t_result.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
 
   for (DenseIndex i = 0; i < t_result.size(); i++) {
     if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4f) {
@@ -307,13 +309,13 @@ void test_cuda_contraction()
     }
   }
 
-  cudaFree(d_t_left);
-  cudaFree(d_t_right);
-  cudaFree(d_t_result);
+  gpuFree(d_t_left);
+  gpuFree(d_t_right);
+  gpuFree(d_t_result);
 }
 
 template<int DataLayout>
-void test_cuda_convolution_1d()
+void test_gpu_convolution_1d()
 {
   Tensor<float, 4, DataLayout> input(74,37,11,137);
   Tensor<float, 1, DataLayout> kernel(4);
@@ -328,14 +330,14 @@ void test_cuda_convolution_1d()
   float* d_input;
   float* d_kernel;
   float* d_out;
-  cudaMalloc((void**)(&d_input), input_bytes);
-  cudaMalloc((void**)(&d_kernel), kernel_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_input), input_bytes);
+  gpuMalloc((void**)(&d_kernel), kernel_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_input, input.data(), input_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_kernel, kernel.data(), kernel_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input, 74,37,11,137);
@@ -345,8 +347,8 @@ void test_cuda_convolution_1d()
   Eigen::array<Eigen::DenseIndex, 1> dims(1);
   gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 74; ++i) {
     for (int j = 0; j < 34; ++j) {
@@ -361,12 +363,12 @@ void test_cuda_convolution_1d()
     }
   }
 
-  cudaFree(d_input);
-  cudaFree(d_kernel);
-  cudaFree(d_out);
+  gpuFree(d_input);
+  gpuFree(d_kernel);
+  gpuFree(d_out);
 }
 
-void test_cuda_convolution_inner_dim_col_major_1d()
+void test_gpu_convolution_inner_dim_col_major_1d()
 {
   Tensor<float, 4, ColMajor> input(74,9,11,7);
   Tensor<float, 1, ColMajor> kernel(4);
@@ -381,14 +383,14 @@ void test_cuda_convolution_inner_dim_col_major_1d()
   float* d_input;
   float* d_kernel;
   float* d_out;
-  cudaMalloc((void**)(&d_input), input_bytes);
-  cudaMalloc((void**)(&d_kernel), kernel_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_input), input_bytes);
+  gpuMalloc((void**)(&d_kernel), kernel_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_input, input.data(), input_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_kernel, kernel.data(), kernel_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_input(d_input,74,9,11,7);
@@ -398,8 +400,8 @@ void test_cuda_convolution_inner_dim_col_major_1d()
   Eigen::array<Eigen::DenseIndex, 1> dims(0);
   gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 71; ++i) {
     for (int j = 0; j < 9; ++j) {
@@ -414,12 +416,12 @@ void test_cuda_convolution_inner_dim_col_major_1d()
     }
   }
 
-  cudaFree(d_input);
-  cudaFree(d_kernel);
-  cudaFree(d_out);
+  gpuFree(d_input);
+  gpuFree(d_kernel);
+  gpuFree(d_out);
 }
 
-void test_cuda_convolution_inner_dim_row_major_1d()
+void test_gpu_convolution_inner_dim_row_major_1d()
 {
   Tensor<float, 4, RowMajor> input(7,9,11,74);
   Tensor<float, 1, RowMajor> kernel(4);
@@ -434,14 +436,14 @@ void test_cuda_convolution_inner_dim_row_major_1d()
   float* d_input;
   float* d_kernel;
   float* d_out;
-  cudaMalloc((void**)(&d_input), input_bytes);
-  cudaMalloc((void**)(&d_kernel), kernel_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_input), input_bytes);
+  gpuMalloc((void**)(&d_kernel), kernel_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_input, input.data(), input_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_kernel, kernel.data(), kernel_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_input(d_input, 7,9,11,74);
@@ -451,8 +453,8 @@ void test_cuda_convolution_inner_dim_row_major_1d()
   Eigen::array<Eigen::DenseIndex, 1> dims(3);
   gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 7; ++i) {
     for (int j = 0; j < 9; ++j) {
@@ -467,13 +469,13 @@ void test_cuda_convolution_inner_dim_row_major_1d()
     }
   }
 
-  cudaFree(d_input);
-  cudaFree(d_kernel);
-  cudaFree(d_out);
+  gpuFree(d_input);
+  gpuFree(d_kernel);
+  gpuFree(d_out);
 }
 
 template<int DataLayout>
-void test_cuda_convolution_2d()
+void test_gpu_convolution_2d()
 {
   Tensor<float, 4, DataLayout> input(74,37,11,137);
   Tensor<float, 2, DataLayout> kernel(3,4);
@@ -488,14 +490,14 @@ void test_cuda_convolution_2d()
   float* d_input;
   float* d_kernel;
   float* d_out;
-  cudaMalloc((void**)(&d_input), input_bytes);
-  cudaMalloc((void**)(&d_kernel), kernel_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_input), input_bytes);
+  gpuMalloc((void**)(&d_kernel), kernel_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_input, input.data(), input_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_kernel, kernel.data(), kernel_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_input(d_input,74,37,11,137);
@@ -505,8 +507,8 @@ void test_cuda_convolution_2d()
   Eigen::array<Eigen::DenseIndex, 2> dims(1,2);
   gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 74; ++i) {
     for (int j = 0; j < 35; ++j) {
@@ -531,13 +533,13 @@ void test_cuda_convolution_2d()
     }
   }
 
-  cudaFree(d_input);
-  cudaFree(d_kernel);
-  cudaFree(d_out);
+  gpuFree(d_input);
+  gpuFree(d_kernel);
+  gpuFree(d_out);
 }
 
 template<int DataLayout>
-void test_cuda_convolution_3d()
+void test_gpu_convolution_3d()
 {
   Tensor<float, 5, DataLayout> input(Eigen::array<Eigen::DenseIndex, 5>(74,37,11,137,17));
   Tensor<float, 3, DataLayout> kernel(3,4,2);
@@ -552,14 +554,14 @@ void test_cuda_convolution_3d()
   float* d_input;
   float* d_kernel;
   float* d_out;
-  cudaMalloc((void**)(&d_input), input_bytes);
-  cudaMalloc((void**)(&d_kernel), kernel_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_input), input_bytes);
+  gpuMalloc((void**)(&d_kernel), kernel_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_input, input.data(), input_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_kernel, kernel.data(), kernel_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;    
+  Eigen::GpuStreamDevice stream;    
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_input(d_input,74,37,11,137,17);
@@ -569,8 +571,8 @@ void test_cuda_convolution_3d()
   Eigen::array<Eigen::DenseIndex, 3> dims(1,2,3);
   gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 74; ++i) {
     for (int j = 0; j < 35; ++j) {
@@ -609,14 +611,14 @@ void test_cuda_convolution_3d()
     }
   }
 
-  cudaFree(d_input);
-  cudaFree(d_kernel);
-  cudaFree(d_out);
+  gpuFree(d_input);
+  gpuFree(d_kernel);
+  gpuFree(d_out);
 }
 
 
 template <typename Scalar>
-void test_cuda_lgamma(const Scalar stddev)
+void test_gpu_lgamma(const Scalar stddev)
 {
   Tensor<Scalar, 2> in(72,97);
   in.setRandom();
@@ -628,12 +630,12 @@ void test_cuda_lgamma(const Scalar stddev)
 
   Scalar* d_in;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in, in.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
@@ -641,8 +643,8 @@ void test_cuda_lgamma(const Scalar stddev)
 
   gpu_out.device(gpu_device) = gpu_in.lgamma();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 72; ++i) {
     for (int j = 0; j < 97; ++j) {
@@ -650,12 +652,12 @@ void test_cuda_lgamma(const Scalar stddev)
     }
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_digamma()
+void test_gpu_digamma()
 {
   Tensor<Scalar, 1> in(7);
   Tensor<Scalar, 1> out(7);
@@ -682,12 +684,12 @@ void test_cuda_digamma()
 
   Scalar* d_in;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in, in.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 7);
@@ -695,8 +697,8 @@ void test_cuda_digamma()
 
   gpu_out.device(gpu_device) = gpu_in.digamma();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 5; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
@@ -705,12 +707,12 @@ void test_cuda_digamma()
     VERIFY_IS_EQUAL(out(i), expected_out(i));
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_zeta()
+void test_gpu_zeta()
 {
   Tensor<Scalar, 1> in_x(6);
   Tensor<Scalar, 1> in_q(6);
@@ -744,14 +746,14 @@ void test_cuda_zeta()
   Scalar* d_in_x;
   Scalar* d_in_q;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in_x), bytes);
-  cudaMalloc((void**)(&d_in_q), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in_x), bytes);
+  gpuMalloc((void**)(&d_in_q), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in_x, in_x.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in_q, in_q.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in_x, in_x.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in_q, in_q.data(), bytes, gpuMemcpyHostToDevice);
   
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 6);
@@ -760,8 +762,8 @@ void test_cuda_zeta()
 
   gpu_out.device(gpu_device) = gpu_in_x.zeta(gpu_in_q);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   VERIFY_IS_EQUAL(out(0), expected_out(0));
   VERIFY((std::isnan)(out(3)));
@@ -772,13 +774,13 @@ void test_cuda_zeta()
     }
   }
 
-  cudaFree(d_in_x);
-  cudaFree(d_in_q);
-  cudaFree(d_out);
+  gpuFree(d_in_x);
+  gpuFree(d_in_q);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_polygamma()
+void test_gpu_polygamma()
 {
   Tensor<Scalar, 1> in_x(7);
   Tensor<Scalar, 1> in_n(7);
@@ -815,14 +817,14 @@ void test_cuda_polygamma()
   Scalar* d_in_x;
   Scalar* d_in_n;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in_x), bytes);
-  cudaMalloc((void**)(&d_in_n), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in_x), bytes);
+  gpuMalloc((void**)(&d_in_n), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in_x, in_x.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in_n, in_n.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in_x, in_x.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in_n, in_n.data(), bytes, gpuMemcpyHostToDevice);
   
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 7);
@@ -831,20 +833,20 @@ void test_cuda_polygamma()
 
   gpu_out.device(gpu_device) = gpu_in_n.polygamma(gpu_in_x);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 7; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
   }
 
-  cudaFree(d_in_x);
-  cudaFree(d_in_n);
-  cudaFree(d_out);
+  gpuFree(d_in_x);
+  gpuFree(d_in_n);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_igamma()
+void test_gpu_igamma()
 {
   Tensor<Scalar, 2> a(6, 6);
   Tensor<Scalar, 2> x(6, 6);
@@ -880,14 +882,14 @@ void test_cuda_igamma()
   Scalar* d_a;
   Scalar* d_x;
   Scalar* d_out;
-  assert(cudaMalloc((void**)(&d_a), bytes) == cudaSuccess);
-  assert(cudaMalloc((void**)(&d_x), bytes) == cudaSuccess);
-  assert(cudaMalloc((void**)(&d_out), bytes) == cudaSuccess);
+  assert(gpuMalloc((void**)(&d_a), bytes) == gpuSuccess);
+  assert(gpuMalloc((void**)(&d_x), bytes) == gpuSuccess);
+  assert(gpuMalloc((void**)(&d_out), bytes) == gpuSuccess);
 
-  cudaMemcpy(d_a, a.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_x, x.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_a, a.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_x, x.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
@@ -896,8 +898,8 @@ void test_cuda_igamma()
 
   gpu_out.device(gpu_device) = gpu_a.igamma(gpu_x);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 6; ++i) {
     for (int j = 0; j < 6; ++j) {
@@ -909,13 +911,13 @@ void test_cuda_igamma()
     }
   }
 
-  cudaFree(d_a);
-  cudaFree(d_x);
-  cudaFree(d_out);
+  gpuFree(d_a);
+  gpuFree(d_x);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_igammac()
+void test_gpu_igammac()
 {
   Tensor<Scalar, 2> a(6, 6);
   Tensor<Scalar, 2> x(6, 6);
@@ -950,14 +952,14 @@ void test_cuda_igammac()
   Scalar* d_a;
   Scalar* d_x;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_a), bytes);
-  cudaMalloc((void**)(&d_x), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_a), bytes);
+  gpuMalloc((void**)(&d_x), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_a, a.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_x, x.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_a, a.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_x, x.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_a(d_a, 6, 6);
@@ -966,8 +968,8 @@ void test_cuda_igammac()
 
   gpu_out.device(gpu_device) = gpu_a.igammac(gpu_x);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 6; ++i) {
     for (int j = 0; j < 6; ++j) {
@@ -979,13 +981,13 @@ void test_cuda_igammac()
     }
   }
 
-  cudaFree(d_a);
-  cudaFree(d_x);
-  cudaFree(d_out);
+  gpuFree(d_a);
+  gpuFree(d_x);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_erf(const Scalar stddev)
+void test_gpu_erf(const Scalar stddev)
 {
   Tensor<Scalar, 2> in(72,97);
   in.setRandom();
@@ -997,12 +999,12 @@ void test_cuda_erf(const Scalar stddev)
 
   Scalar* d_in;
   Scalar* d_out;
-  assert(cudaMalloc((void**)(&d_in), bytes) == cudaSuccess);
-  assert(cudaMalloc((void**)(&d_out), bytes) == cudaSuccess);
+  assert(gpuMalloc((void**)(&d_in), bytes) == gpuSuccess);
+  assert(gpuMalloc((void**)(&d_out), bytes) == gpuSuccess);
 
-  cudaMemcpy(d_in, in.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
@@ -1010,8 +1012,8 @@ void test_cuda_erf(const Scalar stddev)
 
   gpu_out.device(gpu_device) = gpu_in.erf();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 72; ++i) {
     for (int j = 0; j < 97; ++j) {
@@ -1019,12 +1021,12 @@ void test_cuda_erf(const Scalar stddev)
     }
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_erfc(const Scalar stddev)
+void test_gpu_erfc(const Scalar stddev)
 {
   Tensor<Scalar, 2> in(72,97);
   in.setRandom();
@@ -1036,12 +1038,12 @@ void test_cuda_erfc(const Scalar stddev)
 
   Scalar* d_in;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in, in.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 2> > gpu_in(d_in, 72, 97);
@@ -1049,8 +1051,8 @@ void test_cuda_erfc(const Scalar stddev)
 
   gpu_out.device(gpu_device) = gpu_in.erfc();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 72; ++i) {
     for (int j = 0; j < 97; ++j) {
@@ -1058,12 +1060,12 @@ void test_cuda_erfc(const Scalar stddev)
     }
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_betainc()
+void test_gpu_betainc()
 {
   Tensor<Scalar, 1> in_x(125);
   Tensor<Scalar, 1> in_a(125);
@@ -1172,16 +1174,16 @@ void test_cuda_betainc()
   Scalar* d_in_a;
   Scalar* d_in_b;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in_x), bytes);
-  cudaMalloc((void**)(&d_in_a), bytes);
-  cudaMalloc((void**)(&d_in_b), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in_x), bytes);
+  gpuMalloc((void**)(&d_in_a), bytes);
+  gpuMalloc((void**)(&d_in_b), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in_x, in_x.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in_a, in_a.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_in_b, in_b.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in_x, in_x.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in_a, in_a.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_in_b, in_b.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 125);
@@ -1191,8 +1193,8 @@ void test_cuda_betainc()
 
   gpu_out.device(gpu_device) = betainc(gpu_in_a, gpu_in_b, gpu_in_x);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 1; i < 125; ++i) {
     if ((std::isnan)(expected_out(i))) {
@@ -1202,14 +1204,14 @@ void test_cuda_betainc()
     }
   }
 
-  cudaFree(d_in_x);
-  cudaFree(d_in_a);
-  cudaFree(d_in_b);
-  cudaFree(d_out);
+  gpuFree(d_in_x);
+  gpuFree(d_in_a);
+  gpuFree(d_in_b);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_i0e()
+void test_gpu_i0e()
 {
   Tensor<Scalar, 1> in_x(21);
   Tensor<Scalar, 1> out(21);
@@ -1238,12 +1240,12 @@ void test_cuda_i0e()
 
   Scalar* d_in;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in, in_x.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in_x.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 21);
@@ -1251,20 +1253,20 @@ void test_cuda_i0e()
 
   gpu_out.device(gpu_device) = gpu_in.i0e();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 21; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_i1e()
+void test_gpu_i1e()
 {
   Tensor<Scalar, 1> in_x(21);
   Tensor<Scalar, 1> out(21);
@@ -1293,12 +1295,12 @@ void test_cuda_i1e()
 
   Scalar* d_in;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_in), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_in), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_in, in_x.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in, in_x.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in(d_in, 21);
@@ -1306,20 +1308,20 @@ void test_cuda_i1e()
 
   gpu_out.device(gpu_device) = gpu_in.i1e();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 21; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
   }
 
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_igamma_der_a()
+void test_gpu_igamma_der_a()
 {
   Tensor<Scalar, 1> in_x(30);
   Tensor<Scalar, 1> in_a(30);
@@ -1365,14 +1367,14 @@ void test_cuda_igamma_der_a()
   Scalar* d_a;
   Scalar* d_x;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_a), bytes);
-  cudaMalloc((void**)(&d_x), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_a), bytes);
+  gpuMalloc((void**)(&d_x), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_a, in_a.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_x, in_x.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_a, in_a.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_x, in_x.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_a(d_a, 30);
@@ -1381,21 +1383,21 @@ void test_cuda_igamma_der_a()
 
   gpu_out.device(gpu_device) = gpu_a.igamma_der_a(gpu_x);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 30; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
   }
 
-  cudaFree(d_a);
-  cudaFree(d_x);
-  cudaFree(d_out);
+  gpuFree(d_a);
+  gpuFree(d_x);
+  gpuFree(d_out);
 }
 
 template <typename Scalar>
-void test_cuda_gamma_sample_der_alpha()
+void test_gpu_gamma_sample_der_alpha()
 {
   Tensor<Scalar, 1> in_alpha(30);
   Tensor<Scalar, 1> in_sample(30);
@@ -1441,14 +1443,14 @@ void test_cuda_gamma_sample_der_alpha()
   Scalar* d_alpha;
   Scalar* d_sample;
   Scalar* d_out;
-  cudaMalloc((void**)(&d_alpha), bytes);
-  cudaMalloc((void**)(&d_sample), bytes);
-  cudaMalloc((void**)(&d_out), bytes);
+  gpuMalloc((void**)(&d_alpha), bytes);
+  gpuMalloc((void**)(&d_sample), bytes);
+  gpuMalloc((void**)(&d_out), bytes);
 
-  cudaMemcpy(d_alpha, in_alpha.data(), bytes, cudaMemcpyHostToDevice);
-  cudaMemcpy(d_sample, in_sample.data(), bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_alpha, in_alpha.data(), bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_sample, in_sample.data(), bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_alpha(d_alpha, 30);
@@ -1457,101 +1459,115 @@ void test_cuda_gamma_sample_der_alpha()
 
   gpu_out.device(gpu_device) = gpu_alpha.gamma_sample_der_alpha(gpu_sample);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost,
-                         gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, bytes, gpuMemcpyDeviceToHost,
+                         gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   for (int i = 0; i < 30; ++i) {
     VERIFY_IS_APPROX(out(i), expected_out(i));
   }
 
-  cudaFree(d_alpha);
-  cudaFree(d_sample);
-  cudaFree(d_out);
+  gpuFree(d_alpha);
+  gpuFree(d_sample);
+  gpuFree(d_out);
 }
 
-void test_cxx11_tensor_cuda()
+void test_cxx11_tensor_gpu()
 {
-  CALL_SUBTEST_1(test_cuda_nullary());
-  CALL_SUBTEST_1(test_cuda_elementwise_small());
-  CALL_SUBTEST_1(test_cuda_elementwise());
-  CALL_SUBTEST_1(test_cuda_props());
-  CALL_SUBTEST_1(test_cuda_reduction());
-  CALL_SUBTEST_2(test_cuda_contraction<ColMajor>());
-  CALL_SUBTEST_2(test_cuda_contraction<RowMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_1d<ColMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_1d<RowMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_inner_dim_col_major_1d());
-  CALL_SUBTEST_3(test_cuda_convolution_inner_dim_row_major_1d());
-  CALL_SUBTEST_3(test_cuda_convolution_2d<ColMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_2d<RowMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_3d<ColMajor>());
-  CALL_SUBTEST_3(test_cuda_convolution_3d<RowMajor>());
+  CALL_SUBTEST_1(test_gpu_nullary());
+  CALL_SUBTEST_1(test_gpu_elementwise_small());
+  CALL_SUBTEST_1(test_gpu_elementwise());
+  CALL_SUBTEST_1(test_gpu_props());
+  CALL_SUBTEST_1(test_gpu_reduction());
+  CALL_SUBTEST_2(test_gpu_contraction<ColMajor>());
+  CALL_SUBTEST_2(test_gpu_contraction<RowMajor>());
+  CALL_SUBTEST_3(test_gpu_convolution_1d<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_convolution_1d<RowMajor>());
+  CALL_SUBTEST_3(test_gpu_convolution_inner_dim_col_major_1d());
+  CALL_SUBTEST_3(test_gpu_convolution_inner_dim_row_major_1d());
+  CALL_SUBTEST_3(test_gpu_convolution_2d<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_convolution_2d<RowMajor>());
+#if !defined(EIGEN_USE_HIP)
+// disable these tests on HIP for now.
+// they hang..need to investigate and fix
+  CALL_SUBTEST_3(test_gpu_convolution_3d<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_convolution_3d<RowMajor>());
+#endif
 
 #if __cplusplus > 199711L
   // std::erf, std::erfc, and so on where only added in c++11. We use them
   // as a golden reference to validate the results produced by Eigen. Therefore
   // we can only run these tests if we use a c++11 compiler.
-  CALL_SUBTEST_4(test_cuda_lgamma<float>(1.0f));
-  CALL_SUBTEST_4(test_cuda_lgamma<float>(100.0f));
-  CALL_SUBTEST_4(test_cuda_lgamma<float>(0.01f));
-  CALL_SUBTEST_4(test_cuda_lgamma<float>(0.001f));
+  CALL_SUBTEST_4(test_gpu_lgamma<float>(1.0f));
+  CALL_SUBTEST_4(test_gpu_lgamma<float>(100.0f));
+  CALL_SUBTEST_4(test_gpu_lgamma<float>(0.01f));
+  CALL_SUBTEST_4(test_gpu_lgamma<float>(0.001f));
 
-  CALL_SUBTEST_4(test_cuda_lgamma<double>(1.0));
-  CALL_SUBTEST_4(test_cuda_lgamma<double>(100.0));
-  CALL_SUBTEST_4(test_cuda_lgamma<double>(0.01));
-  CALL_SUBTEST_4(test_cuda_lgamma<double>(0.001));
+  CALL_SUBTEST_4(test_gpu_lgamma<double>(1.0));
+  CALL_SUBTEST_4(test_gpu_lgamma<double>(100.0));
+  CALL_SUBTEST_4(test_gpu_lgamma<double>(0.01));
+  CALL_SUBTEST_4(test_gpu_lgamma<double>(0.001));
 
-  CALL_SUBTEST_4(test_cuda_erf<float>(1.0f));
-  CALL_SUBTEST_4(test_cuda_erf<float>(100.0f));
-  CALL_SUBTEST_4(test_cuda_erf<float>(0.01f));
-  CALL_SUBTEST_4(test_cuda_erf<float>(0.001f));
+  CALL_SUBTEST_4(test_gpu_erf<float>(1.0f));
+  CALL_SUBTEST_4(test_gpu_erf<float>(100.0f));
+  CALL_SUBTEST_4(test_gpu_erf<float>(0.01f));
+  CALL_SUBTEST_4(test_gpu_erf<float>(0.001f));
 
-  CALL_SUBTEST_4(test_cuda_erfc<float>(1.0f));
-  // CALL_SUBTEST(test_cuda_erfc<float>(100.0f));
-  CALL_SUBTEST_4(test_cuda_erfc<float>(5.0f)); // CUDA erfc lacks precision for large inputs
-  CALL_SUBTEST_4(test_cuda_erfc<float>(0.01f));
-  CALL_SUBTEST_4(test_cuda_erfc<float>(0.001f));
+  CALL_SUBTEST_4(test_gpu_erfc<float>(1.0f));
+  // CALL_SUBTEST(test_gpu_erfc<float>(100.0f));
+  CALL_SUBTEST_4(test_gpu_erfc<float>(5.0f)); // GPU erfc lacks precision for large inputs
+  CALL_SUBTEST_4(test_gpu_erfc<float>(0.01f));
+  CALL_SUBTEST_4(test_gpu_erfc<float>(0.001f));
 
-  CALL_SUBTEST_4(test_cuda_erf<double>(1.0));
-  CALL_SUBTEST_4(test_cuda_erf<double>(100.0));
-  CALL_SUBTEST_4(test_cuda_erf<double>(0.01));
-  CALL_SUBTEST_4(test_cuda_erf<double>(0.001));
+  CALL_SUBTEST_4(test_gpu_erf<double>(1.0));
+  CALL_SUBTEST_4(test_gpu_erf<double>(100.0));
+  CALL_SUBTEST_4(test_gpu_erf<double>(0.01));
+  CALL_SUBTEST_4(test_gpu_erf<double>(0.001));
 
-  CALL_SUBTEST_4(test_cuda_erfc<double>(1.0));
-  // CALL_SUBTEST(test_cuda_erfc<double>(100.0));
-  CALL_SUBTEST_4(test_cuda_erfc<double>(5.0)); // CUDA erfc lacks precision for large inputs
-  CALL_SUBTEST_4(test_cuda_erfc<double>(0.01));
-  CALL_SUBTEST_4(test_cuda_erfc<double>(0.001));
+  CALL_SUBTEST_4(test_gpu_erfc<double>(1.0));
+  // CALL_SUBTEST(test_gpu_erfc<double>(100.0));
+  CALL_SUBTEST_4(test_gpu_erfc<double>(5.0)); // GPU erfc lacks precision for large inputs
+  CALL_SUBTEST_4(test_gpu_erfc<double>(0.01));
+  CALL_SUBTEST_4(test_gpu_erfc<double>(0.001));
 
-  CALL_SUBTEST_5(test_cuda_digamma<float>());
-  CALL_SUBTEST_5(test_cuda_digamma<double>());
+#if !defined(EIGEN_USE_HIP)
+// disable these tests on HIP for now.
+  CALL_SUBTEST_5(test_gpu_digamma<float>());
+  CALL_SUBTEST_5(test_gpu_digamma<double>());
 
-  CALL_SUBTEST_5(test_cuda_polygamma<float>());
-  CALL_SUBTEST_5(test_cuda_polygamma<double>());
+  CALL_SUBTEST_5(test_gpu_polygamma<float>());
+  CALL_SUBTEST_5(test_gpu_polygamma<double>());
 
-  CALL_SUBTEST_5(test_cuda_zeta<float>());
-  CALL_SUBTEST_5(test_cuda_zeta<double>());
+  CALL_SUBTEST_5(test_gpu_zeta<float>());
+  CALL_SUBTEST_5(test_gpu_zeta<double>());
+#endif
 
-  CALL_SUBTEST_5(test_cuda_igamma<float>());
-  CALL_SUBTEST_5(test_cuda_igammac<float>());
+  CALL_SUBTEST_5(test_gpu_igamma<float>());
+  CALL_SUBTEST_5(test_gpu_igammac<float>());
 
-  CALL_SUBTEST_5(test_cuda_igamma<double>());
-  CALL_SUBTEST_5(test_cuda_igammac<double>());
+  CALL_SUBTEST_5(test_gpu_igamma<double>());
+  CALL_SUBTEST_5(test_gpu_igammac<double>());
 
-  CALL_SUBTEST_6(test_cuda_betainc<float>());
-  CALL_SUBTEST_6(test_cuda_betainc<double>());
+#if !defined(EIGEN_USE_HIP)
+// disable these tests on HIP for now.
+  CALL_SUBTEST_6(test_gpu_betainc<float>());
+  CALL_SUBTEST_6(test_gpu_betainc<double>());
 
-  CALL_SUBTEST_6(test_cuda_i0e<float>());
-  CALL_SUBTEST_6(test_cuda_i0e<double>());
+  CALL_SUBTEST_6(test_gpu_i0e<float>());
+  CALL_SUBTEST_6(test_gpu_i0e<double>());
 
-  CALL_SUBTEST_6(test_cuda_i1e<float>());
-  CALL_SUBTEST_6(test_cuda_i1e<double>());
+  CALL_SUBTEST_6(test_gpu_i1e<float>());
+  CALL_SUBTEST_6(test_gpu_i1e<double>());
 
-  CALL_SUBTEST_6(test_cuda_igamma_der_a<float>());
-  CALL_SUBTEST_6(test_cuda_igamma_der_a<double>());
+  CALL_SUBTEST_6(test_gpu_i1e<float>());
+  CALL_SUBTEST_6(test_gpu_i1e<double>());
+
+  CALL_SUBTEST_6(test_gpu_igamma_der_a<float>());
+  CALL_SUBTEST_6(test_gpu_igamma_der_a<double>());
+
+  CALL_SUBTEST_6(test_gpu_gamma_sample_der_alpha<float>());
+  CALL_SUBTEST_6(test_gpu_gamma_sample_der_alpha<double>());
+#endif
 
-  CALL_SUBTEST_6(test_cuda_gamma_sample_der_alpha<float>());
-  CALL_SUBTEST_6(test_cuda_gamma_sample_der_alpha<double>());
 #endif
 }
diff --git a/unsupported/test/cxx11_tensor_of_float16_gpu.cu b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
index 7a751ff02..150fde8bf 100644
--- a/unsupported/test/cxx11_tensor_of_float16_gpu.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_gpu
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
@@ -20,8 +20,8 @@
 using Eigen::Tensor;
 
 template<typename>
-void test_cuda_numext() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_numext() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -57,11 +57,11 @@ void test_cuda_numext() {
 }
 
 
-#ifdef EIGEN_HAS_CUDA_FP16
+#ifdef EIGEN_HAS_GPU_FP16
 
 template<typename>
-void test_cuda_conversion() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_conversion() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
   
@@ -95,8 +95,8 @@ void test_cuda_conversion() {
 }
 
 template<typename>
-void test_cuda_unary() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_unary() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -132,8 +132,8 @@ void test_cuda_unary() {
 }
 
 template<typename>
-void test_cuda_elementwise() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_elementwise() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -174,8 +174,8 @@ void test_cuda_elementwise() {
 }
 
 template<typename>
-void test_cuda_trancendental() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_trancendental() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -268,8 +268,8 @@ void test_cuda_trancendental() {
 }
 
 template<typename>
-void test_cuda_contractions() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_contractions() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int rows = 23;
   int cols = 23;
@@ -319,12 +319,12 @@ void test_cuda_contractions() {
 }
 
 template<typename>
-void test_cuda_reductions(int size1, int size2, int redux) {
+void test_gpu_reductions(int size1, int size2, int redux) {
 
    std::cout << "Reducing " << size1 << " by " << size2
              << " tensor along dim " << redux << std::endl; 
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = size1*size2;
   int result_size = (redux == 1 ? size1 : size2);
@@ -368,20 +368,20 @@ void test_cuda_reductions(int size1, int size2, int redux) {
 }
 
 template<typename>
-void test_cuda_reductions() {
-  test_cuda_reductions<void>(13, 13, 0);
-  test_cuda_reductions<void>(13, 13, 1);
+void test_gpu_reductions() {
+  test_gpu_reductions<void>(13, 13, 0);
+  test_gpu_reductions<void>(13, 13, 1);
 
-  test_cuda_reductions<void>(35, 36, 0);
-  test_cuda_reductions<void>(35, 36, 1);
+  test_gpu_reductions<void>(35, 36, 0);
+  test_gpu_reductions<void>(35, 36, 1);
 
-  test_cuda_reductions<void>(36, 35, 0);
-  test_cuda_reductions<void>(36, 35, 1);
+  test_gpu_reductions<void>(36, 35, 0);
+  test_gpu_reductions<void>(36, 35, 1);
 }
 
 template<typename>
-void test_cuda_full_reductions() {
-  Eigen::CudaStreamDevice stream;
+void test_gpu_full_reductions() {
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int size = 13;
   int num_elem = size*size;
@@ -429,9 +429,9 @@ void test_cuda_full_reductions() {
 }
 
 template<typename>
-void test_cuda_forced_evals() {
+void test_gpu_forced_evals() {
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
   int num_elem = 101;
 
@@ -479,20 +479,20 @@ void test_cuda_forced_evals() {
 #endif
 
 
-void test_cxx11_tensor_of_float16_cuda()
+void test_cxx11_tensor_of_float16_gpu()
 {
-  CALL_SUBTEST_1(test_cuda_numext<void>());
+  CALL_SUBTEST_1(test_gpu_numext<void>());
 
-#ifdef EIGEN_HAS_CUDA_FP16
-  CALL_SUBTEST_1(test_cuda_conversion<void>());
-  CALL_SUBTEST_1(test_cuda_unary<void>());
-  CALL_SUBTEST_1(test_cuda_elementwise<void>());
-  CALL_SUBTEST_1(test_cuda_trancendental<void>());
-  CALL_SUBTEST_2(test_cuda_contractions<void>());
-  CALL_SUBTEST_3(test_cuda_reductions<void>());
-  CALL_SUBTEST_4(test_cuda_full_reductions<void>());
-  CALL_SUBTEST_5(test_cuda_forced_evals<void>());
+#ifdef EIGEN_HAS_GPU_FP16
+  CALL_SUBTEST_1(test_gpu_conversion<void>());
+  CALL_SUBTEST_1(test_gpu_unary<void>());
+  CALL_SUBTEST_1(test_gpu_elementwise<void>());
+  CALL_SUBTEST_1(test_gpu_trancendental<void>());
+  CALL_SUBTEST_2(test_gpu_contractions<void>());
+  CALL_SUBTEST_3(test_gpu_reductions<void>());
+  CALL_SUBTEST_4(test_gpu_full_reductions<void>());
+  CALL_SUBTEST_5(test_gpu_forced_evals<void>());
 #else
-  std::cout << "Half floats are not supported by this version of cuda: skipping the test" << std::endl;
+  std::cout << "Half floats are not supported by this version of gpu: skipping the test" << std::endl;
 #endif
 }
diff --git a/unsupported/test/cxx11_tensor_random_gpu.cu b/unsupported/test/cxx11_tensor_random_gpu.cu
index 389c0a8c2..da5977f09 100644
--- a/unsupported/test/cxx11_tensor_random_gpu.cu
+++ b/unsupported/test/cxx11_tensor_random_gpu.cu
@@ -9,15 +9,16 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_random_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_random_gpu
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
 #include "main.h"
 #include <Eigen/CXX11/Tensor>
 
+#include <Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
 
-void test_cuda_random_uniform()
+void test_gpu_random_uniform()
 {
   Tensor<float, 2> out(72,97);
   out.setZero();
@@ -25,24 +26,24 @@ void test_cuda_random_uniform()
   std::size_t out_bytes = out.size() * sizeof(float);
 
   float* d_out;
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
 
   gpu_out.device(gpu_device) = gpu_out.random();
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 
   // For now we just check this code doesn't crash.
   // TODO: come up with a valid test of randomness
 }
 
 
-void test_cuda_random_normal()
+void test_gpu_random_normal()
 {
   Tensor<float, 2> out(72,97);
   out.setZero();
@@ -50,9 +51,9 @@ void test_cuda_random_normal()
   std::size_t out_bytes = out.size() * sizeof(float);
 
   float* d_out;
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
@@ -60,8 +61,8 @@ void test_cuda_random_normal()
   Eigen::internal::NormalRandomGenerator<float> gen(true);
   gpu_out.device(gpu_device) = gpu_out.random(gen);
 
-  assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
-  assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
+  assert(gpuMemcpyAsync(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost, gpu_device.stream()) == gpuSuccess);
+  assert(gpuStreamSynchronize(gpu_device.stream()) == gpuSuccess);
 }
 
 static void test_complex()
@@ -77,9 +78,9 @@ static void test_complex()
 }
 
 
-void test_cxx11_tensor_random_cuda()
+void test_cxx11_tensor_random_gpu()
 {
-  CALL_SUBTEST(test_cuda_random_uniform());
-  CALL_SUBTEST(test_cuda_random_normal());
+  CALL_SUBTEST(test_gpu_random_uniform());
+  CALL_SUBTEST(test_gpu_random_normal());
   CALL_SUBTEST(test_complex());
 }
diff --git a/unsupported/test/cxx11_tensor_reduction_gpu.cu b/unsupported/test/cxx11_tensor_reduction_gpu.cu
index ec0669704..a36759303 100644
--- a/unsupported/test/cxx11_tensor_reduction_gpu.cu
+++ b/unsupported/test/cxx11_tensor_reduction_gpu.cu
@@ -9,7 +9,7 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_reduction_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_reduction_gpu
 #define EIGEN_USE_GPU
 
 #include "main.h"
@@ -19,7 +19,7 @@
 template<typename Type, int DataLayout>
 static void test_full_reductions() {
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   const int num_rows = internal::random<int>(1024, 5*1024);
@@ -67,7 +67,7 @@ static void test_first_dim_reductions() {
   Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
 
   // Create device
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice dev(&stream);
   
   // Create data(T)
@@ -107,7 +107,7 @@ static void test_last_dim_reductions() {
   Tensor<Type, 2, DataLayout> redux = in.sum(red_axis);
 
   // Create device
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice dev(&stream);
   
   // Create data
@@ -134,7 +134,7 @@ static void test_last_dim_reductions() {
 }
 
 
-void test_cxx11_tensor_reduction_cuda() {
+void test_cxx11_tensor_reduction_gpu() {
   CALL_SUBTEST_1((test_full_reductions<float, ColMajor>()));
   CALL_SUBTEST_1((test_full_reductions<double, ColMajor>()));
   CALL_SUBTEST_2((test_full_reductions<float, RowMajor>()));
diff --git a/unsupported/test/cxx11_tensor_scan_gpu.cu b/unsupported/test/cxx11_tensor_scan_gpu.cu
index 1d4edef11..51cd3a3cf 100644
--- a/unsupported/test/cxx11_tensor_scan_gpu.cu
+++ b/unsupported/test/cxx11_tensor_scan_gpu.cu
@@ -9,19 +9,20 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cxx11_tensor_scan_cuda
+#define EIGEN_TEST_FUNC cxx11_tensor_scan_gpu
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 #define EIGEN_USE_GPU
 
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#include <Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
 
 using Eigen::Tensor;
 typedef Tensor<float, 1>::DimensionPair DimPair;
 
 template<int DataLayout>
-void test_cuda_cumsum(int m_size, int k_size, int n_size)
+void test_gpu_cumsum(int m_size, int k_size, int n_size)
 {
   std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
   Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size);
@@ -36,12 +37,12 @@ void test_cuda_cumsum(int m_size, int k_size, int n_size)
   float* d_t_input;
   float* d_t_result;
 
-  cudaMalloc((void**)(&d_t_input), t_input_bytes);
-  cudaMalloc((void**)(&d_t_result), t_result_bytes);
+  gpuMalloc((void**)(&d_t_input), t_input_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
 
-  cudaMemcpy(d_t_input, t_input.data(), t_input_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_t_input, t_input.data(), t_input_bytes, gpuMemcpyHostToDevice);
 
-  Eigen::CudaStreamDevice stream;
+  Eigen::GpuStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
 
   Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
@@ -52,7 +53,7 @@ void test_cuda_cumsum(int m_size, int k_size, int n_size)
   gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1);
   t_result = t_input.cumsum(1);
 
-  cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
+  gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
   for (DenseIndex i = 0; i < t_result.size(); i++) {
     if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
       continue;
@@ -65,13 +66,13 @@ void test_cuda_cumsum(int m_size, int k_size, int n_size)
     assert(false);
   }
 
-  cudaFree((void*)d_t_input);
-  cudaFree((void*)d_t_result);
+  gpuFree((void*)d_t_input);
+  gpuFree((void*)d_t_result);
 }
 
 
-void test_cxx11_tensor_scan_cuda()
+void test_cxx11_tensor_scan_gpu()
 {
-  CALL_SUBTEST_1(test_cuda_cumsum<ColMajor>(128, 128, 128));
-  CALL_SUBTEST_2(test_cuda_cumsum<RowMajor>(128, 128, 128));
+  CALL_SUBTEST_1(test_gpu_cumsum<ColMajor>(128, 128, 128));
+  CALL_SUBTEST_2(test_gpu_cumsum<RowMajor>(128, 128, 128));
 }

From 359dd77ec3583fa44a927e3ebcee902bd37f6623 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Jul 2018 11:03:39 +0200
Subject: [PATCH 408/680] Fix legitimate "declaration shadows a typedef"
 warning

---
 Eigen/src/Core/products/GeneralBlockPanelKernel.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index dd4f366ee..1890efd4d 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -1526,10 +1526,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           // The following piece of code won't work for 512 bit registers
           // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size
           // as nr (which is currently 4) for the return type.
-          typedef typename unpacket_traits<SResPacket>::half SResPacketHalf;
+          const int SResPacketHalfSize = unpacket_traits<typename unpacket_traits<SResPacket>::half>::size;
           if ((SwappedTraits::LhsProgress % 4) == 0 &&
               (SwappedTraits::LhsProgress <= 8) &&
-              (SwappedTraits::LhsProgress!=8 || unpacket_traits<SResPacketHalf>::size==nr))
+              (SwappedTraits::LhsProgress!=8 || SResPacketHalfSize==nr))
           {
             SAccPacket C0, C1, C2, C3;
             straits.initAcc(C0);

From ec323b7e6643d9da3d9cda86570cc7724e0a8d3c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Jul 2018 11:13:19 +0200
Subject: [PATCH 409/680] Skip null numerators in triangular-vector-solve (as
 in BLAS TRSV).

---
 .../Core/products/TriangularSolverVector.h    | 21 +++++++++++--------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/Eigen/src/Core/products/TriangularSolverVector.h b/Eigen/src/Core/products/TriangularSolverVector.h
index b994759b2..647317016 100644
--- a/Eigen/src/Core/products/TriangularSolverVector.h
+++ b/Eigen/src/Core/products/TriangularSolverVector.h
@@ -58,7 +58,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
       {
         // let's directly call the low level product function because:
         // 1 - it is faster to compile
-        // 2 - it is slighlty faster at runtime
+        // 2 - it is slightly faster at runtime
         Index startRow = IsLower ? pi : pi-actualPanelWidth;
         Index startCol = IsLower ? 0 : pi;
 
@@ -77,7 +77,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
         if (k>0)
           rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum();
 
-        if(!(Mode & UnitDiag))
+        if((!(Mode & UnitDiag)) && numext::not_equal_strict(rhs[i],RhsScalar(0)))
           rhs[i] /= cjLhs(i,i);
       }
     }
@@ -114,20 +114,23 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
       for(Index k=0; k<actualPanelWidth; ++k)
       {
         Index i = IsLower ? pi+k : pi-k-1;
-        if(!(Mode & UnitDiag))
-          rhs[i] /= cjLhs.coeff(i,i);
+        if(numext::not_equal_strict(rhs[i],RhsScalar(0)))
+        {
+          if(!(Mode & UnitDiag))
+            rhs[i] /= cjLhs.coeff(i,i);
 
-        Index r = actualPanelWidth - k - 1; // remaining size
-        Index s = IsLower ? i+1 : i-r;
-        if (r>0)
-          Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+          Index r = actualPanelWidth - k - 1; // remaining size
+          Index s = IsLower ? i+1 : i-r;
+          if (r>0)
+            Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+        }
       }
       Index r = IsLower ? size - endBlock : startBlock; // remaining size
       if (r > 0)
       {
         // let's directly call the low level product function because:
         // 1 - it is faster to compile
-        // 2 - it is slighlty faster at runtime
+        // 2 - it is slightly faster at runtime
         general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
             r, actualPanelWidth,
             LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),

From 6190aa5632fb698fa66d2dad2949275089f15738 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Jul 2018 11:23:16 +0200
Subject: [PATCH 410/680] bug #1567: add optimized path for tensor broadcasting
 and 'Channel First' shape

---
 .../CXX11/src/Tensor/TensorBroadcasting.h     | 70 +++++++++++++++++--
 .../test/cxx11_tensor_broadcasting.cpp        | 57 +++++++++++++++
 2 files changed, 123 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 9ab6b3565..b35b36475 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -161,6 +161,22 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
         }
       }
     }
+
+    // Handle special format like NCHW, its input shape is '[1, N..., 1]' and
+    // broadcast shape is '[N, 1..., N]'
+    if (!oneByN && !nByOne) {
+      if (input_dims[0] == 1 && input_dims[NumDims-1] == 1 && NumDims > 2) {
+        nByOne = true;
+        oneByN = true;
+        for (int i = 1; i < NumDims-1; ++i) {
+          if (broadcast[i] != 1) {
+            nByOne = false;
+            oneByN = false;
+            break;
+          }
+        }
+      }
+    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -256,24 +272,70 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     }
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      if (oneByN) {
+      if (oneByN && !nByOne) {
         return packetNByOne<LoadMode>(index);
-      } else if (nByOne) {
+      } else if (!oneByN && nByOne) {
         return packetOneByN<LoadMode>(index);
+      } else if (oneByN && nByOne) {
+        return packetOneByNByOne<LoadMode>(index);
       } else {
         return packetColMajor<LoadMode>(index);
       }
     } else {
-      if (oneByN) {
+      if (oneByN && !nByOne) {
         return packetOneByN<LoadMode>(index);
-      } else if (nByOne) {
+      } else if (!oneByN && nByOne) {
         return packetNByOne<LoadMode>(index);
+      } else if (oneByN && nByOne) {
+        return packetOneByNByOne<LoadMode>(index);
       } else {
         return packetRowMajor<LoadMode>(index);
       }
     }
   }
 
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByNByOne
+  (Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    Index startDim, endDim;
+    Index inputIndex, outputOffset, batchedIndex;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      startDim = NumDims - 1;
+      endDim = 1;
+    } else {
+      startDim = 0;
+      endDim = NumDims - 2;
+    }
+
+    batchedIndex = index % m_outputStrides[startDim];
+    inputIndex   = batchedIndex / m_outputStrides[endDim];
+    outputOffset = batchedIndex % m_outputStrides[endDim];
+
+    if (outputOffset + PacketSize <= m_outputStrides[endDim]) {
+      values[0] = m_impl.coeff(inputIndex);
+      return internal::pload1<PacketReturnType>(values);
+    } else {
+      for (int i = 0, cur = 0; i < PacketSize; ++i, ++cur) {
+        if (outputOffset + cur < m_outputStrides[endDim]) {
+          values[i] = m_impl.coeff(inputIndex);
+        } else {
+          ++inputIndex;
+          inputIndex = (inputIndex == m_inputStrides[startDim] ? 0 : inputIndex);
+          values[i] = m_impl.coeff(inputIndex);
+          outputOffset = 0;
+          cur = 0;
+        }
+      }
+      return internal::pload<PacketReturnType>(values);
+    }
+  }
+
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByN(Index index) const
   {
diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp
index a9d268ea6..f0ff03184 100644
--- a/unsupported/test/cxx11_tensor_broadcasting.cpp
+++ b/unsupported/test/cxx11_tensor_broadcasting.cpp
@@ -238,6 +238,59 @@ static void test_simple_broadcasting_n_by_one()
   }
 }
 
+template <int DataLayout>
+static void test_simple_broadcasting_one_by_n_by_one_1d()
+{
+  Tensor<float, 3, DataLayout> tensor(1,7,1);
+  tensor.setRandom();
+  array<ptrdiff_t, 3> broadcasts;
+  broadcasts[0] = 5;
+  broadcasts[1] = 1;
+  broadcasts[2] = 13;
+  Tensor<float, 3, DataLayout> broadcasted;
+  broadcasted = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcasted.dimension(0), 5);
+  VERIFY_IS_EQUAL(broadcasted.dimension(1), 7);
+  VERIFY_IS_EQUAL(broadcasted.dimension(2), 13);
+
+  for (int i = 0; i < 5; ++i) {
+    for (int j = 0; j < 7; ++j) {
+      for (int k = 0; k < 13; ++k) {
+        VERIFY_IS_EQUAL(tensor(0,j%7,0), broadcasted(i,j,k));
+      }
+    }
+  }
+}
+
+template <int DataLayout>
+static void test_simple_broadcasting_one_by_n_by_one_2d()
+{
+  Tensor<float, 4, DataLayout> tensor(1,7,13,1);
+  tensor.setRandom();
+  array<ptrdiff_t, 4> broadcasts;
+  broadcasts[0] = 5;
+  broadcasts[1] = 1;
+  broadcasts[2] = 1;
+  broadcasts[3] = 19;
+  Tensor<float, 4, DataLayout> broadcast;
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 5);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 7);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 13);
+  VERIFY_IS_EQUAL(broadcast.dimension(3), 19);
+
+  for (int i = 0; i < 5; ++i) {
+    for (int j = 0; j < 7; ++j) {
+      for (int k = 0; k < 13; ++k) {
+        for (int l = 0; l < 19; ++l) {
+          VERIFY_IS_EQUAL(tensor(0,j%7,k%13,0), broadcast(i,j,k,l));
+        }
+      }
+    }
+  }
+}
 
 void test_cxx11_tensor_broadcasting()
 {
@@ -253,4 +306,8 @@ void test_cxx11_tensor_broadcasting()
   CALL_SUBTEST(test_simple_broadcasting_n_by_one<RowMajor>());
   CALL_SUBTEST(test_simple_broadcasting_one_by_n<ColMajor>());
   CALL_SUBTEST(test_simple_broadcasting_n_by_one<ColMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n_by_one_1d<ColMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n_by_one_2d<ColMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n_by_one_1d<RowMajor>());
+  CALL_SUBTEST(test_simple_broadcasting_one_by_n_by_one_2d<RowMajor>());
 }

From de9e31a06d0324862d9200d08eb3cc4d3d07e660 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 9 Jul 2018 15:41:14 +0200
Subject: [PATCH 411/680] Introduce the macro ei_declare_local_nested_eval to
 help allocating on the stack local temporaries via alloca, and let
 outer-products makes a good use of it. If successful, we should use it
 everywhere nested_eval is used to declare local dense temporaries.

---
 Eigen/src/Core/ProductEvaluators.h |  4 +-
 Eigen/src/Core/util/Memory.h       | 80 ++++++++++++++++++++++++++++--
 Eigen/src/Core/util/XprHelper.h    |  2 +-
 test/product_notemporary.cpp       | 10 +++-
 4 files changed, 89 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 76e5083c1..22ad32ae3 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -272,7 +272,7 @@ template<typename Dst, typename Lhs, typename Rhs, typename Func>
 void EIGEN_DEVICE_FUNC outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
 {
   evaluator<Rhs> rhsEval(rhs);
-  typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs);
+  ei_declare_local_nested_eval(Lhs,lhs,Rhs::SizeAtCompileTime,actual_lhs);
   // FIXME if cols is large enough, then it might be useful to make sure that lhs is sequentially stored
   // FIXME not very good if rhs is real and lhs complex while alpha is real too
   const Index cols = dst.cols();
@@ -285,7 +285,7 @@ template<typename Dst, typename Lhs, typename Rhs, typename Func>
 void EIGEN_DEVICE_FUNC outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
 {
   evaluator<Lhs> lhsEval(lhs);
-  typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs);
+  ei_declare_local_nested_eval(Rhs,rhs,Lhs::SizeAtCompileTime,actual_rhs);
   // FIXME if rows is large enough, then it might be useful to make sure that rhs is sequentially stored
   // FIXME not very good if lhs is real and rhs complex while alpha is real too
   const Index rows = dst.rows();
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 22d7679c5..43920bdc8 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -582,6 +582,60 @@ template<typename T> class aligned_stack_memory_handler : noncopyable
     bool m_deallocate;
 };
 
+#ifdef EIGEN_ALLOCA
+
+template<typename Xpr, int NbEvaluations,
+         bool MapExternalBuffer = nested_eval<Xpr,NbEvaluations>::Evaluate && Xpr::MaxSizeAtCompileTime==Dynamic
+         >
+struct local_nested_eval_wrapper
+{
+  static const bool NeedExternalBuffer = false;
+  typedef typename Xpr::Scalar Scalar;
+  typedef typename nested_eval<Xpr,NbEvaluations>::type ObjectType;
+  ObjectType object;
+
+  EIGEN_DEVICE_FUNC
+  local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr) : object(xpr)
+  {
+    EIGEN_UNUSED_VARIABLE(ptr);
+    eigen_internal_assert(ptr==0);
+  }
+};
+
+template<typename Xpr, int NbEvaluations>
+struct local_nested_eval_wrapper<Xpr,NbEvaluations,true>
+{
+  static const bool NeedExternalBuffer = true;
+  typedef typename Xpr::Scalar Scalar;
+  typedef typename plain_object_eval<Xpr>::type PlainObject;
+  typedef Map<PlainObject,EIGEN_DEFAULT_ALIGN_BYTES> ObjectType;
+  ObjectType object;
+
+  EIGEN_DEVICE_FUNC
+  local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr)
+    : object(ptr==0 ? reinterpret_cast<Scalar*>(Eigen::internal::aligned_malloc(sizeof(Scalar)*xpr.size())) : ptr, xpr.rows(), xpr.cols()),
+      m_deallocate(ptr==0)
+  {
+    if(NumTraits<Scalar>::RequireInitialization && object.data())
+      Eigen::internal::construct_elements_of_array(object.data(), object.size());
+    object = xpr;
+  }
+
+  EIGEN_DEVICE_FUNC
+  ~local_nested_eval_wrapper()
+  {
+    if(NumTraits<Scalar>::RequireInitialization && object.data())
+      Eigen::internal::destruct_elements_of_array(object.data(), object.size());
+    if(m_deallocate)
+      Eigen::internal::aligned_free(object.data());
+  }
+
+private:
+  bool m_deallocate;
+};
+
+#endif // EIGEN_ALLOCA
+
 template<typename T> class scoped_array : noncopyable
 {
   T* m_ptr;
@@ -609,9 +663,11 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 } // end namespace internal
 
 /** \internal
-  * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
-  * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
-  * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
+  * 
+  * The macro ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) declares, allocates,
+  * and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
+  * if the size in bytes is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
+  * (currently, this is Linux, OSX and Visual Studio only). Otherwise the memory is allocated on the heap.
   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
   * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
   * Here is an example:
@@ -622,6 +678,14 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
   * }
   * \endcode
   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
+  * 
+  * The macro ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) is analogue to
+  * \code
+  *   typename internal::nested_eval<XPRT_T,N>::type NAME(XPR);
+  * \endcode
+  * with the advantage of using aligned stack allocation even if the maximal size of XPR at compile time is unknown.
+  * This is accomplished through alloca if this later is supported and if the required number of bytes
+  * is below EIGEN_STACK_ALLOCATION_LIMIT.
   */
 #ifdef EIGEN_ALLOCA
 
@@ -641,6 +705,13 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,sizeof(TYPE)*SIZE>EIGEN_STACK_ALLOCATION_LIMIT)
 
+
+  #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) \
+    Eigen::internal::local_nested_eval_wrapper<XPR_T,N> EIGEN_CAT(NAME,_wrapper)(XPR, reinterpret_cast<typename XPR_T::Scalar*>( \
+      ( (Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::NeedExternalBuffer) && ((sizeof(typename XPR_T::Scalar)*XPR.size())<=EIGEN_STACK_ALLOCATION_LIMIT) ) \
+        ? EIGEN_ALIGNED_ALLOCA( sizeof(typename XPR_T::Scalar)*XPR.size() ) : 0 ) ) ; \
+    typename Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::ObjectType NAME(EIGEN_CAT(NAME,_wrapper).object)
+
 #else
 
   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
@@ -648,6 +719,9 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
 
+
+#define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR);
+
 #endif
 
 
diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 7311adec7..3926d3a6c 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -460,7 +460,7 @@ template<typename T, int n, typename PlainObject = typename plain_object_eval<T>
 {
   enum {
     ScalarReadCost = NumTraits<typename traits<T>::Scalar>::ReadCost,
-    CoeffReadCost = evaluator<T>::CoeffReadCost,  // NOTE What if an evaluator evaluate itself into a tempory?
+    CoeffReadCost = evaluator<T>::CoeffReadCost,  // NOTE What if an evaluator evaluate itself into a temporary?
                                                   //      Then CoeffReadCost will be small (e.g., 1) but we still have to evaluate, especially if n>1.
                                                   //      This situation is already taken care by the EvalBeforeNestingBit flag, which is turned ON
                                                   //      for all evaluator creating a temporary. This flag is then propagated by the parent evaluators.
diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp
index 30592b79e..062180f42 100644
--- a/test/product_notemporary.cpp
+++ b/test/product_notemporary.cpp
@@ -128,11 +128,19 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m)
   VERIFY_EVALUATION_COUNT( cvres.noalias() = (rm3+rm3) * (m1*cv1), 1 );
 
   // Check outer products
+  #ifdef EIGEN_ALLOCA
+  bool temp_via_alloca = m3.rows()*sizeof(Scalar) <= EIGEN_STACK_ALLOCATION_LIMIT;
+  #else
+  bool temp_via_alloca = false;
+  #endif
   m3 = cv1 * rv1;
   VERIFY_EVALUATION_COUNT( m3.noalias() = cv1 * rv1, 0 );
-  VERIFY_EVALUATION_COUNT( m3.noalias() = (cv1+cv1) * (rv1+rv1), 1 );
+  VERIFY_EVALUATION_COUNT( m3.noalias() = (cv1+cv1) * (rv1+rv1), temp_via_alloca ? 0 : 1 );
   VERIFY_EVALUATION_COUNT( m3.noalias() = (m1*cv1) * (rv1), 1 );
   VERIFY_EVALUATION_COUNT( m3.noalias() += (m1*cv1) * (rv1), 1 );
+  rm3 = cv1 * rv1;
+  VERIFY_EVALUATION_COUNT( rm3.noalias() = cv1 * rv1, 0 );
+  VERIFY_EVALUATION_COUNT( rm3.noalias() = (cv1+cv1) * (rv1+rv1), temp_via_alloca ? 0 : 1 );
   VERIFY_EVALUATION_COUNT( rm3.noalias() = (cv1) * (rv1 * m1), 1 );
   VERIFY_EVALUATION_COUNT( rm3.noalias() -= (cv1) * (rv1 * m1), 1 );
   VERIFY_EVALUATION_COUNT( rm3.noalias() = (m1*cv1) * (rv1 * m1), 2 );

From 9357838f94d2907996adadc7e5200376f3561ed4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 10 Jul 2018 09:10:15 +0200
Subject: [PATCH 412/680] bug #1543: improve linear indexing for general block
 expressions

---
 Eigen/src/Core/CoreEvaluators.h | 6 +++---
 test/block.cpp                  | 2 ++
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 6231f2edd..b65ec4ffb 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1080,7 +1080,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
     : m_argImpl(block.nestedExpression()), 
       m_startRow(block.startRow()), 
       m_startCol(block.startCol()),
-      m_linear_offset(InnerPanel?(XprType::IsRowMajor ? block.startRow()*block.cols() : block.startCol()*block.rows()):0)
+      m_linear_offset((InnerPanel|| XprType::IsVectorAtCompileTime)?(XprType::IsRowMajor ? block.startRow()*block.cols() + block.startCol() : block.startCol()*block.rows() + block.startRow()):0)
   { }
  
   typedef typename XprType::Scalar Scalar;
@@ -1088,7 +1088,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
 
   enum {
     RowsAtCompileTime = XprType::RowsAtCompileTime,
-    ForwardLinearAccess = InnerPanel && bool(evaluator<ArgType>::Flags&LinearAccessBit)
+    ForwardLinearAccess = (InnerPanel || XprType::IsVectorAtCompileTime) && bool(evaluator<ArgType>::Flags&LinearAccessBit)
   };
  
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -1162,7 +1162,7 @@ protected:
   evaluator<ArgType> m_argImpl;
   const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
   const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
-  const variable_if_dynamic<Index, InnerPanel ? Dynamic : 0> m_linear_offset;
+  const variable_if_dynamic<Index, (InnerPanel || XprType::IsVectorAtCompileTime) ? Dynamic : 0> m_linear_offset;
 };
 
 // TODO: This evaluator does not actually use the child evaluator; 
diff --git a/test/block.cpp b/test/block.cpp
index 8c4dd87be..9c2424662 100644
--- a/test/block.cpp
+++ b/test/block.cpp
@@ -162,9 +162,11 @@ template<typename MatrixType> void block(const MatrixType& m)
   // expressions without direct access
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,rows-r1,cols-c1).block(r2-r1,c2-c1,rows-r2,cols-c2)) , ((m1+m2).block(r2,c2,rows-r2,cols-c2)) );
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).row(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)) );
+  VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).row(0)) , ((m1+m2).eval().row(r1).segment(c1,c2-c1+1)) );
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).col(0)) , ((m1+m2).col(c1).segment(r1,r2-r1+1)) );
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).transpose().col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
   VERIFY_IS_APPROX( ((m1+m2).transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
+  VERIFY_IS_APPROX( ((m1+m2).template block<1,Dynamic>(r1,c1,1,c2-c1+1)) , ((m1+m2).eval().row(r1).segment(c1,c2-c1+1)) );
 
   VERIFY_IS_APPROX( (m1*1).topRows(r1),  m1.topRows(r1) );
   VERIFY_IS_APPROX( (m1*1).leftCols(c1), m1.leftCols(c1) );

From fe723d6129f2fde60e91e7cef52a7907a0cb479f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 10 Jul 2018 09:10:32 +0200
Subject: [PATCH 413/680] Fix conversion warning

---
 Eigen/src/Core/util/IntegralConstant.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/IntegralConstant.h b/Eigen/src/Core/util/IntegralConstant.h
index 78a4705cd..bf99cd8ab 100644
--- a/Eigen/src/Core/util/IntegralConstant.h
+++ b/Eigen/src/Core/util/IntegralConstant.h
@@ -192,7 +192,7 @@ inline internal::FixedInt<N> fix() { return internal::FixedInt<N>(); }
 // The generic typename T is mandatory. Otherwise, a code like fix<N> could refer to either the function above or this next overload.
 // This way a code like fix<N> can only refer to the previous function.
 template<int N,typename T>
-inline internal::VariableAndFixedInt<N> fix(T val) { return internal::VariableAndFixedInt<N>(val); }
+inline internal::VariableAndFixedInt<N> fix(T val) { return internal::VariableAndFixedInt<N>(internal::convert_index<int>(val)); }
 #endif
 
 #else // EIGEN_PARSED_BY_DOXYGEN

From 1625476091c2c4576fe9f2eef91df4d0444dc9e0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Jul 2018 14:00:24 +0200
Subject: [PATCH 414/680] Add internall::is_identity compile-time helper

---
 Eigen/src/Core/util/XprHelper.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index 3926d3a6c..e3231c712 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -676,6 +676,14 @@ template<typename T> struct is_diagonal<DiagonalWrapper<T> >
 template<typename T, int S> struct is_diagonal<DiagonalMatrix<T,S> >
 { enum { ret = true }; };
 
+
+template<typename T> struct is_identity
+{ enum { value = false }; };
+
+template<typename T> struct is_identity<CwiseNullaryOp<internal::scalar_identity_op<typename T::Scalar>, T> >
+{ enum { value = true }; };
+
+
 template<typename S1, typename S2> struct glue_shapes;
 template<> struct glue_shapes<DenseShape,TriangularShape> { typedef TriangularShape type;  };
 

From f00d08cc0a987fa624209b920608b56638404f13 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Jul 2018 14:01:47 +0200
Subject: [PATCH 415/680] Optimize extraction of Q in SparseQR by exploiting
 the structure of the identity matrix.

---
 Eigen/src/SparseQR/SparseQR.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h
index 7409fcae9..1a28389e8 100644
--- a/Eigen/src/SparseQR/SparseQR.h
+++ b/Eigen/src/SparseQR/SparseQR.h
@@ -640,7 +640,8 @@ struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived
       // Compute res = Q * other column by column
       for(Index j = 0; j < res.cols(); j++)
       {
-        for (Index k = diagSize-1; k >=0; k--)
+        Index start_k = internal::is_identity<Derived>::value ? numext::mini(j,diagSize-1) : diagSize-1;
+        for (Index k = start_k; k >=0; k--)
         {
           Scalar tau = Scalar(0);
           tau = m_qr.m_Q.col(k).dot(res.col(j));

From 471cfe5ff7577b51adc37de61ca39ad63a2d0b40 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 11 Jul 2018 09:22:04 -0400
Subject: [PATCH 416/680] renaming CUDA* to GPU* for some header files

---
 Eigen/src/Core/arch/{CUDA => GPU}/MathFunctions.h                 | 0
 Eigen/src/Core/arch/{CUDA => GPU}/PacketMath.h                    | 0
 .../{CUDA/CudaSpecialFunctions.h => GPU/GpuSpecialFunctions.h}    | 0
 3 files changed, 0 insertions(+), 0 deletions(-)
 rename Eigen/src/Core/arch/{CUDA => GPU}/MathFunctions.h (100%)
 rename Eigen/src/Core/arch/{CUDA => GPU}/PacketMath.h (100%)
 rename unsupported/Eigen/src/SpecialFunctions/arch/{CUDA/CudaSpecialFunctions.h => GPU/GpuSpecialFunctions.h} (100%)

diff --git a/Eigen/src/Core/arch/CUDA/MathFunctions.h b/Eigen/src/Core/arch/GPU/MathFunctions.h
similarity index 100%
rename from Eigen/src/Core/arch/CUDA/MathFunctions.h
rename to Eigen/src/Core/arch/GPU/MathFunctions.h
diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h
similarity index 100%
rename from Eigen/src/Core/arch/CUDA/PacketMath.h
rename to Eigen/src/Core/arch/GPU/PacketMath.h
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h
similarity index 100%
rename from unsupported/Eigen/src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h
rename to unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h

From dec47a64930b2a5b68c55706cbdba7f945781d0e Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 11 Jul 2018 09:26:54 -0400
Subject: [PATCH 417/680] renaming CUDA* to GPU* for some header files

---
 test/{cuda_basic.cu => gpu_basic.cu} | 0
 test/{cuda_common.h => gpu_common.h} | 0
 2 files changed, 0 insertions(+), 0 deletions(-)
 rename test/{cuda_basic.cu => gpu_basic.cu} (100%)
 rename test/{cuda_common.h => gpu_common.h} (100%)

diff --git a/test/cuda_basic.cu b/test/gpu_basic.cu
similarity index 100%
rename from test/cuda_basic.cu
rename to test/gpu_basic.cu
diff --git a/test/cuda_common.h b/test/gpu_common.h
similarity index 100%
rename from test/cuda_common.h
rename to test/gpu_common.h

From 1fe0b749042320501c59378f2860d9322b0c6e19 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 11 Jul 2018 09:28:44 -0400
Subject: [PATCH 418/680] deleting hip specific files that are no longer
 required

---
 test/hip_basic.cu | 172 ----------------------------------------------
 test/hip_common.h | 103 ---------------------------
 2 files changed, 275 deletions(-)
 delete mode 100644 test/hip_basic.cu
 delete mode 100644 test/hip_common.h

diff --git a/test/hip_basic.cu b/test/hip_basic.cu
deleted file mode 100644
index 2e1bf94a4..000000000
--- a/test/hip_basic.cu
+++ /dev/null
@@ -1,172 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2015-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-// workaround issue between gcc >= 4.7 and cuda 5.5
-#if (defined __GNUC__) && (__GNUC__>4 || __GNUC_MINOR__>=7)
-  #undef _GLIBCXX_ATOMIC_BUILTINS
-  #undef _GLIBCXX_USE_INT128
-#endif
-
-#define EIGEN_TEST_NO_LONGDOUBLE
-#define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC hip_basic
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
-
-#include <hip/hip_runtime.h>
-
-#include "main.h"
-#include "hip_common.h"
-
-// Check that dense modules can be properly parsed by hipcc
-#include <Eigen/Dense>
-
-// struct Foo{
-//   EIGEN_DEVICE_FUNC
-//   void operator()(int i, const float* mats, float* vecs) const {
-//     using namespace Eigen;
-//   //   Matrix3f M(data);
-//   //   Vector3f x(data+9);
-//   //   Map<Vector3f>(data+9) = M.inverse() * x;
-//     Matrix3f M(mats+i/16);
-//     Vector3f x(vecs+i*3);
-//   //   using std::min;
-//   //   using std::sqrt;
-//     Map<Vector3f>(vecs+i*3) << x.minCoeff(), 1, 2;// / x.dot(x);//(M.inverse() *  x) / x.x();
-//     //x = x*2 + x.y() * x + x * x.maxCoeff() - x / x.sum();
-//   }
-// };
-
-template<typename T>
-struct coeff_wise {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
-  {
-    using namespace Eigen;
-    T x1(in+i);
-    T x2(in+i+1);
-    T x3(in+i+2);
-    Map<T> res(out+i*T::MaxSizeAtCompileTime);
-    
-    res.array() += (in[0] * x1 + x2).array() * x3.array();
-  }
-};
-
-template<typename T>
-struct replicate {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
-  {
-    using namespace Eigen;
-    T x1(in+i);
-    int step   = x1.size() * 4;
-    int stride = 3 * step;
-    
-    typedef Map<Array<typename T::Scalar,Dynamic,Dynamic> > MapType;
-    MapType(out+i*stride+0*step, x1.rows()*2, x1.cols()*2) = x1.replicate(2,2);
-    MapType(out+i*stride+1*step, x1.rows()*3, x1.cols()) = in[i] * x1.colwise().replicate(3);
-    MapType(out+i*stride+2*step, x1.rows(), x1.cols()*3) = in[i] * x1.rowwise().replicate(3);
-  }
-};
-
-template<typename T>
-struct redux {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
-  {
-    using namespace Eigen;
-    int N = 10;
-    T x1(in+i);
-    out[i*N+0] = x1.minCoeff();
-    out[i*N+1] = x1.maxCoeff();
-    out[i*N+2] = x1.sum();
-    out[i*N+3] = x1.prod();
-    out[i*N+4] = x1.matrix().squaredNorm();
-    out[i*N+5] = x1.matrix().norm();
-    out[i*N+6] = x1.colwise().sum().maxCoeff();
-    out[i*N+7] = x1.rowwise().maxCoeff().sum();
-    out[i*N+8] = x1.matrix().colwise().squaredNorm().sum();
-  }
-};
-
-template<typename T1, typename T2>
-struct prod_test {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const
-  {
-    using namespace Eigen;
-    typedef Matrix<typename T1::Scalar, T1::RowsAtCompileTime, T2::ColsAtCompileTime> T3;
-    T1 x1(in+i);
-    T2 x2(in+i+1);
-    Map<T3> res(out+i*T3::MaxSizeAtCompileTime);
-    res += in[i] * x1 * x2;
-  }
-};
-
-template<typename T1, typename T2>
-struct diagonal {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const
-  {
-    using namespace Eigen;
-    T1 x1(in+i);
-    Map<T2> res(out+i*T2::MaxSizeAtCompileTime);
-    res += x1.diagonal();
-  }
-};
-
-template<typename T>
-struct eigenvalues {
-  EIGEN_DEVICE_FUNC
-  void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
-  {
-    using namespace Eigen;
-    typedef Matrix<typename T::Scalar, T::RowsAtCompileTime, 1> Vec;
-    T M(in+i);
-    Map<Vec> res(out+i*Vec::MaxSizeAtCompileTime);
-    T A = M*M.adjoint();
-    SelfAdjointEigenSolver<T> eig;
-    eig.computeDirect(M);
-    res = eig.eigenvalues();
-  }
-};
-
-void test_hip_basic()
-{
-  ei_test_init_hip();
-  
-  int nthreads = 100;
-  Eigen::VectorXf in, out;
-  
-  #ifndef __HIP_DEVICE_COMPILE__ 
-  int data_size = nthreads * 512;
-  in.setRandom(data_size);
-  out.setRandom(data_size);
-  #endif
-  
-  CALL_SUBTEST( run_and_compare_to_hip(coeff_wise<Vector3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_hip(coeff_wise<Array44f>(), nthreads, in, out) );
-
-  // FIXME compile fails when we uncomment the followig two tests
-  // CALL_SUBTEST( run_and_compare_to_hip(replicate<Array4f>(), nthreads, in, out) );
-  // CALL_SUBTEST( run_and_compare_to_hip(replicate<Array33f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_hip(redux<Array4f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_hip(redux<Matrix3f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_hip(prod_test<Matrix3f,Matrix3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_hip(prod_test<Matrix4f,Vector4f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_hip(diagonal<Matrix3f,Vector3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_hip(diagonal<Matrix4f,Vector4f>(), nthreads, in, out) );
-
-  // FIXME : Runtime failure occurs when we uncomment the following two tests
-  // CALL_SUBTEST( run_and_compare_to_hip(eigenvalues<Matrix3f>(), nthreads, in, out) );
-  // CALL_SUBTEST( run_and_compare_to_hip(eigenvalues<Matrix2f>(), nthreads, in, out) );
-
-}
diff --git a/test/hip_common.h b/test/hip_common.h
deleted file mode 100644
index 251585c52..000000000
--- a/test/hip_common.h
+++ /dev/null
@@ -1,103 +0,0 @@
-
-#ifndef EIGEN_TEST_HIP_COMMON_H
-#define EIGEN_TEST_HIP_COMMON_H
-
-#include "hip/hip_runtime.h"
-#include "hip/hip_runtime_api.h"
-#include <iostream>
-
-#ifndef __HIPCC__
-dim3 threadIdx, blockDim, blockIdx;
-#endif
-
-template<typename Kernel, typename Input, typename Output>
-void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out)
-{
-  for(int i=0; i<n; i++)
-    ker(i, in.data(), out.data());
-}
-
-
-template<typename Kernel, typename Input, typename Output>
-__global__ __attribute__((used))
-void run_on_hip_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
-{
-  int i = hipThreadIdx_x + hipBlockIdx_x*hipBlockDim_x;
-  if(i<n) {
-    ker(i, in, out);
-  }
-}
-
-
-template<typename Kernel, typename Input, typename Output>
-void run_on_hip(const Kernel& ker, int n, const Input& in, Output& out)
-{
-  typename Input::Scalar*  d_in;
-  typename Output::Scalar* d_out;
-  std::ptrdiff_t in_bytes  = in.size()  * sizeof(typename Input::Scalar);
-  std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
-  
-  hipMalloc((void**)(&d_in),  in_bytes);
-  hipMalloc((void**)(&d_out), out_bytes);
-  
-  hipMemcpy(d_in,  in.data(),  in_bytes,  hipMemcpyHostToDevice);
-  hipMemcpy(d_out, out.data(), out_bytes, hipMemcpyHostToDevice);
-  
-  // Simple and non-optimal 1D mapping assuming n is not too large
-  // That's only for unit testing!
-  dim3 Blocks(128);
-  dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
-
-  hipDeviceSynchronize();
-  hipLaunchKernelGGL(HIP_KERNEL_NAME(run_on_hip_meta_kernel<Kernel,
-                                                         typename std::decay<decltype(*d_in)>::type,
-                                                         typename std::decay<decltype(*d_out)>::type>), 
-                  dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
-  hipDeviceSynchronize();
-  
-  // check inputs have not been modified
-  hipMemcpy(const_cast<typename Input::Scalar*>(in.data()),  d_in,  in_bytes,  hipMemcpyDeviceToHost);
-  hipMemcpy(out.data(), d_out, out_bytes, hipMemcpyDeviceToHost);
-  
-  hipFree(d_in);
-  hipFree(d_out);
-}
-
-
-template<typename Kernel, typename Input, typename Output>
-void run_and_compare_to_hip(const Kernel& ker, int n, const Input& in, Output& out)
-{
-  Input  in_ref,  in_hip;
-  Output out_ref, out_hip;
-  #ifndef __HIP_DEVICE_COMPILE__
-  in_ref = in_hip = in;
-  out_ref = out_hip = out;
-  #endif
-  run_on_cpu (ker, n, in_ref,  out_ref);
-  run_on_hip(ker, n, in_hip, out_hip);
-  #ifndef __HIP_DEVICE_COMPILE__
-  VERIFY_IS_APPROX(in_ref, in_hip);
-  VERIFY_IS_APPROX(out_ref, out_hip);
-  #endif
-}
-
-
-void ei_test_init_hip()
-{
-  int device = 0;
-  hipDeviceProp_t deviceProp;
-  hipGetDeviceProperties(&deviceProp, device);
-  std::cout << "HIP device info:\n";
-  std::cout << "  name:                        " << deviceProp.name << "\n";
-  std::cout << "  capability:                  " << deviceProp.major << "." << deviceProp.minor << "\n";
-  std::cout << "  multiProcessorCount:         " << deviceProp.multiProcessorCount << "\n";
-  std::cout << "  maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n";
-  std::cout << "  warpSize:                    " << deviceProp.warpSize << "\n";
-  std::cout << "  regsPerBlock:                " << deviceProp.regsPerBlock << "\n";
-  std::cout << "  concurrentKernels:           " << deviceProp.concurrentKernels << "\n";
-  std::cout << "  clockRate:                   " << deviceProp.clockRate << "\n";
-  std::cout << "  canMapHostMemory:            " << deviceProp.canMapHostMemory << "\n";
-  std::cout << "  computeMode:                 " << deviceProp.computeMode << "\n";
-}
-
-#endif // EIGEN_TEST_HIP_COMMON_H

From 876f392c396318f33454168db36ed54308e54e0d Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Wed, 11 Jul 2018 10:39:54 -0400
Subject: [PATCH 419/680] Updates corresponding to the latest round of PR
 feedback

The major changes are

1. Moving CUDA/PacketMath.h to GPU/PacketMath.h
2. Moving CUDA/MathFunctions.h to GPU/MathFunction.h
3. Moving CUDA/CudaSpecialFunctions.h to GPU/GpuSpecialFunctions.h
    The above three changes effectively enable the Eigen "Packet" layer for the HIP platform

4. Merging the "hip_basic" and "cuda_basic" unit tests into one ("gpu_basic")
5. Updating the "EIGEN_DEVICE_FUNC" marking in some places

The change has been tested on the HIP and CUDA platforms.
---
 Eigen/Core                                    |  17 +-
 Eigen/src/Core/MathFunctions.h                |  10 ++
 Eigen/src/Core/arch/GPU/MathFunctions.h       |   8 +-
 Eigen/src/Core/arch/GPU/PacketMath.h          |   8 +-
 Eigen/src/Core/products/GeneralMatrixVector.h |   8 +-
 test/CMakeLists.txt                           |   4 +-
 test/gpu_basic.cu                             |  63 ++++----
 test/gpu_common.h                             |  86 ++++++----
 .../src/Tensor/TensorContractionBlocking.h    |  14 ++
 .../CXX11/src/Tensor/TensorContractionGpu.h   | 147 ------------------
 .../src/Tensor/TensorGpuHipCudaDefines.h      |   5 +-
 .../src/Tensor/TensorGpuHipCudaUndefines.h    |   1 +
 .../Eigen/CXX11/src/Tensor/TensorIndexList.h  |   5 +-
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |   5 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  11 +-
 .../CXX11/src/Tensor/TensorReductionGpu.h     |  41 +----
 unsupported/Eigen/SpecialFunctions            |   4 +-
 .../arch/GPU/GpuSpecialFunctions.h            |   6 +-
 18 files changed, 158 insertions(+), 285 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index 4336de91d..647a10831 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -81,6 +81,7 @@
       // clang++ always considers constexpr functions as implicitly __host__ __device__
       #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
     #endif
+  #endif
 
 #elif defined(EIGEN_HIPCC)
   // Do not try to vectorize on HIP
@@ -92,7 +93,7 @@
   // We need hip_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
   // works properly on the device side
   #include <hip/hip_runtime.h>
-    
+  
   #if defined(__HIP_DEVICE_COMPILE__) && !defined(EIGEN_NO_HIP)
     // analogous to EIGEN_CUDA_ARCH, but for HIP
     #define EIGEN_HIP_DEVICE_COMPILE __HIP_DEVICE_COMPILE__
@@ -356,7 +357,7 @@
 #endif
 
 #if defined EIGEN_CUDACC
-  #define EIGEN_VECTORIZE_CUDA
+  #define EIGEN_VECTORIZE_GPU
   #include <vector_types.h>
   #if EIGEN_CUDACC_VER >= 70500
     #define EIGEN_HAS_CUDA_FP16
@@ -369,14 +370,20 @@
 #endif
 
 #if defined(EIGEN_HIPCC) && defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  #define EIGEN_VECTORIZE_GPU
+  #include <hip/hip_vector_types.h>
+
   #define EIGEN_HAS_HIP_FP16
   #include <hip/hip_fp16.h>
+
   #define HIP_PATCH_WITH_NEW_FP16 18215
   #if (HIP_VERSION_PATCH < HIP_PATCH_WITH_NEW_FP16)
     #define EIGEN_HAS_OLD_HIP_FP16
     // Old HIP implementation does not have a explicit typedef for "half2"
     typedef __half2 half2;
   #endif
+  
 #endif
 
 #if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
@@ -550,9 +557,9 @@ using std::ptrdiff_t;
 #include "src/Core/arch/GPU/PacketMathHalf.h"
 #include "src/Core/arch/GPU/TypeCasting.h"
 
-#if defined EIGEN_VECTORIZE_CUDA
-  #include "src/Core/arch/CUDA/PacketMath.h"
-  #include "src/Core/arch/CUDA/MathFunctions.h"
+#if defined EIGEN_VECTORIZE_GPU
+  #include "src/Core/arch/GPU/PacketMath.h"
+  #include "src/Core/arch/GPU/MathFunctions.h"
 #endif
 
 #include "src/Core/arch/Default/Settings.h"
diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 72aa68d45..f16476a92 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -982,7 +982,12 @@ template<>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE long double mini(const long double& x, const long double& y)
 {
+#if defined(EIGEN_HIPCC)
+  // no "fminl" on HIP yet
+  return (x < y) ? x : y;
+#else
   return fminl(x, y);
+#endif
 }
 
 template<typename T>
@@ -1007,7 +1012,12 @@ template<>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE long double maxi(const long double& x, const long double& y)
 {
+#if defined(EIGEN_HIPCC)
+  // no "fmaxl" on HIP yet
+  return (x > y) ? x : y;
+#else
   return fmaxl(x, y);
+#endif
 }
 #endif
 
diff --git a/Eigen/src/Core/arch/GPU/MathFunctions.h b/Eigen/src/Core/arch/GPU/MathFunctions.h
index ff6256ce0..d2b3a2568 100644
--- a/Eigen/src/Core/arch/GPU/MathFunctions.h
+++ b/Eigen/src/Core/arch/GPU/MathFunctions.h
@@ -7,8 +7,8 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_MATH_FUNCTIONS_CUDA_H
-#define EIGEN_MATH_FUNCTIONS_CUDA_H
+#ifndef EIGEN_MATH_FUNCTIONS_GPU_H
+#define EIGEN_MATH_FUNCTIONS_GPU_H
 
 namespace Eigen {
 
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_GPUCC) && defined(EIGEN_USE_GPU)
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 plog<float4>(const float4& a)
 {
@@ -100,4 +100,4 @@ double2 prsqrt<double2>(const double2& a)
 
 } // end namespace Eigen
 
-#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
+#endif // EIGEN_MATH_FUNCTIONS_GPU_H
diff --git a/Eigen/src/Core/arch/GPU/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h
index ab8e477f4..ddf37b9c1 100644
--- a/Eigen/src/Core/arch/GPU/PacketMath.h
+++ b/Eigen/src/Core/arch/GPU/PacketMath.h
@@ -7,8 +7,8 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_PACKET_MATH_CUDA_H
-#define EIGEN_PACKET_MATH_CUDA_H
+#ifndef EIGEN_PACKET_MATH_GPU_H
+#define EIGEN_PACKET_MATH_GPU_H
 
 namespace Eigen {
 
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_GPUCC) && defined(EIGEN_USE_GPU)
 template<> struct is_arithmetic<float4>  { enum { value = true }; };
 template<> struct is_arithmetic<double2> { enum { value = true }; };
 
@@ -338,4 +338,4 @@ ptranspose(PacketBlock<double2,2>& kernel) {
 } // end namespace Eigen
 
 
-#endif // EIGEN_PACKET_MATH_CUDA_H
+#endif // EIGEN_PACKET_MATH_GPU_H
diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h
index b2a71bc6f..767feb99d 100644
--- a/Eigen/src/Core/products/GeneralMatrixVector.h
+++ b/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -48,7 +48,7 @@ typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
 typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
 typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
 
-EIGEN_DONT_INLINE static void run(
+EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE static void run(
   Index rows, Index cols,
   const LhsMapper& lhs,
   const RhsMapper& rhs,
@@ -57,7 +57,7 @@ EIGEN_DONT_INLINE static void run(
 };
 
 template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
-EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
+EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
   Index rows, Index cols,
   const LhsMapper& alhs,
   const RhsMapper& rhs,
@@ -231,7 +231,7 @@ typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
 typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
 typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
 
-EIGEN_DONT_INLINE static void run(
+EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE static void run(
   Index rows, Index cols,
   const LhsMapper& lhs,
   const RhsMapper& rhs,
@@ -240,7 +240,7 @@ EIGEN_DONT_INLINE static void run(
 };
 
 template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
-EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
+EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
   Index rows, Index cols,
   const LhsMapper& alhs,
   const RhsMapper& rhs,
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index f20df119a..d312d16e4 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -399,7 +399,7 @@ if(CUDA_FOUND)
   cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR})
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
   
-  ei_add_test(cuda_basic)
+  ei_add_test(gpu_basic)
   
   unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 
@@ -429,7 +429,7 @@ if (EIGEN_TEST_HIP)
 	include_directories(${HIP_PATH}/include)
 
 	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
-	ei_add_test(hip_basic)
+	ei_add_test(gpu_basic)
 	unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 	
       elseif (${HIP_PLATFORM} STREQUAL "nvcc")
diff --git a/test/gpu_basic.cu b/test/gpu_basic.cu
index 33e5fd119..897834dff 100644
--- a/test/gpu_basic.cu
+++ b/test/gpu_basic.cu
@@ -15,13 +15,11 @@
 
 #define EIGEN_TEST_NO_LONGDOUBLE
 #define EIGEN_TEST_NO_COMPLEX
-#define EIGEN_TEST_FUNC cuda_basic
+#define EIGEN_TEST_FUNC gpu_basic
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
 
-#include <math_constants.h>
-#include <cuda.h>
 #include "main.h"
-#include "cuda_common.h"
+#include "gpu_common.h"
 
 // Check that dense modules can be properly parsed by nvcc
 #include <Eigen/Dense>
@@ -164,40 +162,51 @@ struct matrix_inverse {
   }
 };
 
-void test_cuda_basic()
+void test_gpu_basic()
 {
-  ei_test_init_cuda();
+  ei_test_init_gpu();
   
   int nthreads = 100;
   Eigen::VectorXf in, out;
   
-  #ifndef __CUDA_ARCH__
+  #if !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
   int data_size = nthreads * 512;
   in.setRandom(data_size);
   out.setRandom(data_size);
   #endif
   
-  CALL_SUBTEST( run_and_compare_to_cuda(coeff_wise<Vector3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(coeff_wise<Array44f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_cuda(replicate<Array4f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(replicate<Array33f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_cuda(redux<Array4f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(redux<Matrix3f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_cuda(prod_test<Matrix3f,Matrix3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(prod_test<Matrix4f,Vector4f>(), nthreads, in, out) );
-  
-  CALL_SUBTEST( run_and_compare_to_cuda(diagonal<Matrix3f,Vector3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(diagonal<Matrix4f,Vector4f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(coeff_wise<Vector3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(coeff_wise<Array44f>(), nthreads, in, out) );
 
-  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix2f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(matrix_inverse<Matrix4f>(), nthreads, in, out) );
+#if !defined(EIGEN_USE_HIP)
+  // FIXME
+  // These subtests result in a compile failure on the HIP platform
+  //
+  //  eigen-upstream/Eigen/src/Core/Replicate.h:61:65: error:
+  //           base class 'internal::dense_xpr_base<Replicate<Array<float, 4, 1, 0, 4, 1>, -1, -1> >::type'
+  //           (aka 'ArrayBase<Eigen::Replicate<Eigen::Array<float, 4, 1, 0, 4, 1>, -1, -1> >') has protected default constructor
+  CALL_SUBTEST( run_and_compare_to_gpu(replicate<Array4f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(replicate<Array33f>(), nthreads, in, out) );
+#endif
   
-  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues_direct<Matrix3f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues_direct<Matrix2f>(), nthreads, in, out) );
-  CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues<Matrix4f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(redux<Array4f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(redux<Matrix3f>(), nthreads, in, out) );
+  
+  CALL_SUBTEST( run_and_compare_to_gpu(prod_test<Matrix3f,Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(prod_test<Matrix4f,Vector4f>(), nthreads, in, out) );
+  
+  CALL_SUBTEST( run_and_compare_to_gpu(diagonal<Matrix3f,Vector3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(diagonal<Matrix4f,Vector4f>(), nthreads, in, out) );
 
+  CALL_SUBTEST( run_and_compare_to_gpu(matrix_inverse<Matrix2f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(matrix_inverse<Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(matrix_inverse<Matrix4f>(), nthreads, in, out) );
+  
+#if !defined(EIGEN_USE_HIP)
+  // FIXME
+  // These subtests result in a linking error on the HIP platform
+  CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues_direct<Matrix3f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues_direct<Matrix2f>(), nthreads, in, out) );
+  CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues<Matrix4f>(), nthreads, in, out) );
+#endif
 }
diff --git a/test/gpu_common.h b/test/gpu_common.h
index 9737693ac..3030af6dc 100644
--- a/test/gpu_common.h
+++ b/test/gpu_common.h
@@ -1,13 +1,22 @@
 
-#ifndef EIGEN_TEST_CUDA_COMMON_H
-#define EIGEN_TEST_CUDA_COMMON_H
+#ifndef EIGEN_TEST_GPU_COMMON_H
+#define EIGEN_TEST_GPU_COMMON_H
+
+#ifdef EIGEN_USE_HIP
+  #include <hip/hip_runtime.h>
+  #include <hip/hip_runtime_api.h>
+#else
+  #include <cuda.h>
+  #include <cuda_runtime.h>
+  #include <cuda_runtime_api.h>
+#endif
 
-#include <cuda.h>
-#include <cuda_runtime.h>
-#include <cuda_runtime_api.h>
 #include <iostream>
 
-#ifndef __CUDACC__
+#define EIGEN_USE_GPU
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
+
+#if !defined(__CUDACC__) && !defined(__HIPCC__)
 dim3 threadIdx, blockDim, blockIdx;
 #endif
 
@@ -21,7 +30,7 @@ void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out)
 
 template<typename Kernel, typename Input, typename Output>
 __global__
-void run_on_cuda_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
+void run_on_gpu_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
 {
   int i = threadIdx.x + blockIdx.x*blockDim.x;
   if(i<n) {
@@ -31,61 +40,70 @@ void run_on_cuda_meta_kernel(const Kernel ker, int n, const Input* in, Output* o
 
 
 template<typename Kernel, typename Input, typename Output>
-void run_on_cuda(const Kernel& ker, int n, const Input& in, Output& out)
+void run_on_gpu(const Kernel& ker, int n, const Input& in, Output& out)
 {
   typename Input::Scalar*  d_in;
   typename Output::Scalar* d_out;
   std::ptrdiff_t in_bytes  = in.size()  * sizeof(typename Input::Scalar);
   std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
   
-  cudaMalloc((void**)(&d_in),  in_bytes);
-  cudaMalloc((void**)(&d_out), out_bytes);
+  gpuMalloc((void**)(&d_in),  in_bytes);
+  gpuMalloc((void**)(&d_out), out_bytes);
   
-  cudaMemcpy(d_in,  in.data(),  in_bytes,  cudaMemcpyHostToDevice);
-  cudaMemcpy(d_out, out.data(), out_bytes, cudaMemcpyHostToDevice);
+  gpuMemcpy(d_in,  in.data(),  in_bytes,  gpuMemcpyHostToDevice);
+  gpuMemcpy(d_out, out.data(), out_bytes, gpuMemcpyHostToDevice);
   
   // Simple and non-optimal 1D mapping assuming n is not too large
   // That's only for unit testing!
   dim3 Blocks(128);
   dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
 
-  cudaThreadSynchronize();
-  run_on_cuda_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
-  cudaThreadSynchronize();
+  gpuDeviceSynchronize();
+  
+#ifdef EIGEN_USE_HIP
+  hipLaunchKernelGGL(run_on_gpu_meta_kernel<Kernel,
+		     typename std::decay<decltype(*d_in)>::type,
+		     typename std::decay<decltype(*d_out)>::type>, 
+		     dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
+#else
+  run_on_gpu_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
+#endif
+  
+  gpuDeviceSynchronize();
   
   // check inputs have not been modified
-  cudaMemcpy(const_cast<typename Input::Scalar*>(in.data()),  d_in,  in_bytes,  cudaMemcpyDeviceToHost);
-  cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost);
+  gpuMemcpy(const_cast<typename Input::Scalar*>(in.data()),  d_in,  in_bytes,  gpuMemcpyDeviceToHost);
+  gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost);
   
-  cudaFree(d_in);
-  cudaFree(d_out);
+  gpuFree(d_in);
+  gpuFree(d_out);
 }
 
 
 template<typename Kernel, typename Input, typename Output>
-void run_and_compare_to_cuda(const Kernel& ker, int n, const Input& in, Output& out)
+void run_and_compare_to_gpu(const Kernel& ker, int n, const Input& in, Output& out)
 {
-  Input  in_ref,  in_cuda;
-  Output out_ref, out_cuda;
-  #ifndef __CUDA_ARCH__
-  in_ref = in_cuda = in;
-  out_ref = out_cuda = out;
+  Input  in_ref,  in_gpu;
+  Output out_ref, out_gpu;
+  #if !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
+  in_ref = in_gpu = in;
+  out_ref = out_gpu = out;
   #endif
   run_on_cpu (ker, n, in_ref,  out_ref);
-  run_on_cuda(ker, n, in_cuda, out_cuda);
-  #ifndef __CUDA_ARCH__
-  VERIFY_IS_APPROX(in_ref, in_cuda);
-  VERIFY_IS_APPROX(out_ref, out_cuda);
+  run_on_gpu(ker, n, in_gpu, out_gpu);
+  #if !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
+  VERIFY_IS_APPROX(in_ref, in_gpu);
+  VERIFY_IS_APPROX(out_ref, out_gpu);
   #endif
 }
 
 
-void ei_test_init_cuda()
+void ei_test_init_gpu()
 {
   int device = 0;
-  cudaDeviceProp deviceProp;
-  cudaGetDeviceProperties(&deviceProp, device);
-  std::cout << "CUDA device info:\n";
+  gpuDeviceProp_t deviceProp;
+  gpuGetDeviceProperties(&deviceProp, device);
+  std::cout << "GPU device info:\n";
   std::cout << "  name:                        " << deviceProp.name << "\n";
   std::cout << "  capability:                  " << deviceProp.major << "." << deviceProp.minor << "\n";
   std::cout << "  multiProcessorCount:         " << deviceProp.multiProcessorCount << "\n";
@@ -98,4 +116,4 @@ void ei_test_init_cuda()
   std::cout << "  computeMode:                 " << deviceProp.computeMode << "\n";
 }
 
-#endif // EIGEN_TEST_CUDA_COMMON_H
+#endif // EIGEN_TEST_GPU_COMMON_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
index a1e292108..8c1af1da8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -28,6 +28,20 @@ class TensorContractionBlocking {
   typedef typename LhsMapper::Scalar LhsScalar;
   typedef typename RhsMapper::Scalar RhsScalar;
 
+ /*
+   adding EIGEN_DEVICE_FUNC unconditionally to 'TensorContractionBlocking' constructor in `TensorContractionBlocking.h`
+     requires adding EIGEN_DEVICE_FUNC to `computeProductBlockingSizes` in `GeneralBlockPanelKernel.h`
+     which in turn, requires adding EIGEN_DEVICE_FUNC to `evaluateProductBlockingSizesHeuristic` in `GeneralBlockPanelKernel.h`
+     which in turn, requires adding EIGEN_DEVICE_FUNC to `manage_caching_sizes` in `GeneralBlockPanelKernel.h`
+     (else HIPCC will error out)
+
+   However adding EIGEN_DEVICE_FUNC to `manage_caching_sizes` in `GeneralBlockPanelKernel.h`
+   results in NVCC erroring out with the following error
+
+   ../Eigen/src/Core/products/GeneralBlockPanelKernel.h(57): error #2901:
+      dynamic initialization is not supported for function-scope static variables within a __device__/__global__ function
+ */
+ 
   #if !defined(EIGEN_HIPCC)
   EIGEN_DEVICE_FUNC
   #endif
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index 238754424..a4f92ee44 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -546,45 +546,6 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     results[i].x = results[i].y = results[i].z = results[i].w = 0;
   }
 
-#if defined(EIGEN_HIPCC)
-
-#define prefetch_lhs(reg, row, col)                            \
-    if (!CHECK_LHS_BOUNDARY) {                                 \
-      if (col < k_size) {                                      \
-          reg.x =lhs(row + 0, col);                            \
-          reg.y =lhs(row + 1, col);                            \
-          reg.z =lhs(row + 2, col);                            \
-          reg.w =lhs(row + 3, col);                            \
-      }                                                        \
-    } else {                                                   \
-      if (col < k_size) {                                      \
-        if (row + 3 < m_size) {                                \
-          reg.x =lhs(row + 0, col);                            \
-          reg.y =lhs(row + 1, col);                            \
-          reg.z =lhs(row + 2, col);                            \
-          reg.w =lhs(row + 3, col);                            \
-        } else if (row + 2 < m_size) {                         \
-          reg.x =lhs(row + 0, col);                            \
-          reg.y =lhs(row + 1, col);                            \
-          reg.z =lhs(row + 2, col);                            \
-        } else if (row + 1 < m_size) {                         \
-          reg.x =lhs(row + 0, col);                            \
-          reg.y =lhs(row + 1, col);                            \
-        } else if (row  < m_size) {                            \
-          reg.x =lhs(row + 0, col);                            \
-        }                                                      \
-      }                                                        \
-    }							       \
-
-#define prefetch_rhs_hipcc(reg, row, col)                  \
-      reg.x =rhs(row + 0, col);				   \
-      reg.y =rhs(row + 1, col);				   \
-      reg.z =rhs(row + 2, col);                            \
-      reg.w =rhs(row + 3, col);                            \
-
-
-#else
-  
 #define prefetch_lhs(reg, row, col)                            \
     if (!CHECK_LHS_BOUNDARY) {                                 \
       if (col < k_size) {                                      \
@@ -607,19 +568,12 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
       }                                                        \
     }							       \
 
-#endif
-
   Index lhs_vert = base_m+threadIdx.x*4;
 
   for (Index k = 0; k < k_size; k += 16) {
 
-#if defined(EIGEN_HIPCC)
-    lhs_pf0 = make_float4(0, 0, 0, 0);
-    rhs_pf0 = make_float4(0, 0, 0, 0);
-#else
     lhs_pf0 = internal::pset1<float4>(0);
     rhs_pf0 = internal::pset1<float4>(0);
-#endif
 
     Index lhs_horiz = threadIdx.y+k;
     prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
@@ -630,11 +584,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-#if defined(EIGEN_HIPCC)
-	prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
-#else
         rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-#endif
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -649,11 +599,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     } else {
       if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
-#if defined(EIGEN_HIPCC)
-	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
-#else
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-#endif
         } else if ((rhs_vert + 2) < k_size) {
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
           rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
@@ -753,10 +699,6 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
 #undef prefetch_lhs
 #undef add_vals
   
-#if defined(EIGEN_HIPCC)  
-#undef prefetch_rhs_hipcc
-#endif
-  
   Index horiz_base = threadIdx.y*4+base_n;
   if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
     for (int i = 0; i < 4; i++) {
@@ -845,33 +787,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     results[i].x = results[i].y = results[i].z = results[i].w = 0;
   }
 
-#if defined(EIGEN_HIPCC)
-
-#define prefetch_lhs_hipcc(reg, row, col)                  \
-      reg.x =lhs(row + 0, col);                            \
-      reg.y =lhs(row + 1, col);                            \
-      reg.z =lhs(row + 2, col);                            \
-      reg.w =lhs(row + 3, col);                            
-
-#define prefetch_rhs_hipcc(reg, row, col)                  \
-      reg.x =rhs(row + 0, col);                            \
-      reg.y =rhs(row + 1, col);                            \
-      reg.z =rhs(row + 2, col);                            \
-      reg.w =rhs(row + 3, col);                            
-
-#endif
-
   Index lhs_vert = base_m+threadIdx.x*4+(threadIdx.y%4)*32;
   for (Index k = 0; k < k_size; k += 32) {
-#if defined(EIGEN_HIPCC)
-    lhs_pf0 = make_float4(0, 0, 0, 0);
-    lhs_pf1 = make_float4(0, 0, 0, 0);
-    lhs_pf2 = make_float4(0, 0, 0, 0);
-    lhs_pf3 = make_float4(0, 0, 0, 0);
-
-    rhs_pf0 = make_float4(0, 0, 0, 0);
-    rhs_pf1 = make_float4(0, 0, 0, 0);
-#else  
     lhs_pf0 = internal::pset1<float4>(0);
     lhs_pf1 = internal::pset1<float4>(0);
     lhs_pf2 = internal::pset1<float4>(0);
@@ -879,85 +796,40 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 
     rhs_pf0 = internal::pset1<float4>(0);
     rhs_pf1 = internal::pset1<float4>(0);
-#endif
 
      if (!CHECK_LHS_BOUNDARY) {
       if ((threadIdx.y/4+k+24) < k_size) {
-#if defined(EIGEN_HIPCC)
-	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-	prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
-	prefetch_lhs_hipcc(lhs_pf3, lhs_vert, (threadIdx.y/4+k+24))
-#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
         lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
-#endif
       } else if ((threadIdx.y/4+k+16) < k_size) {
-#if defined(EIGEN_HIPCC)
-	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-	prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
-#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-#endif
       } else if ((threadIdx.y/4+k+8) < k_size) {
-#if defined(EIGEN_HIPCC)
-	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
         lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-#endif
       } else if ((threadIdx.y/4+k) < k_size) {
-#if defined(EIGEN_HIPCC)
-	prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-#else
         lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-#endif
       }
     } else {
       // just CHECK_LHS_BOUNDARY
       if (lhs_vert + 3 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
-#if defined(EIGEN_HIPCC)
-	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-	  prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
-	  prefetch_lhs_hipcc(lhs_pf3, lhs_vert, (threadIdx.y/4+k+24))
-#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
           lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
           lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
-#endif
         } else if ((threadIdx.y/4+k+16) < k_size) {
-#if defined(EIGEN_HIPCC)
-	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-	  prefetch_lhs_hipcc(lhs_pf2, lhs_vert, (threadIdx.y/4+k+16))
-#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
           lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-#endif
         } else if ((threadIdx.y/4+k+8) < k_size) {
-#if defined(EIGEN_HIPCC)
-	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-	  prefetch_lhs_hipcc(lhs_pf1, lhs_vert, (threadIdx.y/4+k+8))
-#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
           lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-#endif
         } else if ((threadIdx.y/4+k) < k_size) {
-#if defined(EIGEN_HIPCC)
-	  prefetch_lhs_hipcc(lhs_pf0, lhs_vert, (threadIdx.y/4+k))
-#else
           lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-#endif
         }
       } else if (lhs_vert + 2 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
@@ -1046,13 +918,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-#if defined(EIGEN_HIPCC)
-	prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
-	prefetch_rhs_hipcc(rhs_pf1, rhs_vert, rhs_horiz1)
-#else	
         rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
         rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
-#endif
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -1074,13 +941,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       if (rhs_horiz1 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-#if defined(EIGEN_HIPCC)
-  	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
-	  prefetch_rhs_hipcc(rhs_pf1, rhs_vert, rhs_horiz1)
-#else	
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
           rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
-#endif
         } else if (rhs_vert + 2 < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -1101,11 +963,7 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       } else if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-#if defined(EIGEN_HIPCC)
-  	  prefetch_rhs_hipcc(rhs_pf0, rhs_vert, rhs_horiz0)
-#else	
           rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-#endif
         } else if ((rhs_vert + 2) < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -1213,11 +1071,6 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     __syncthreads();
   } // end loop over k
 
-#if defined(EIGEN_HIPCC)
-#undef prefetch_lhs_hipcc
-#undef prefetch_rhs_hipcc
-#endif
-  
   __syncthreads();
   Index horiz_base = (threadIdx.y/4)*8+base_n;
   if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
index f009ae855..9966955f7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
@@ -32,8 +32,7 @@
 #define gpuGetDeviceCount hipGetDeviceCount
 #define gpuGetErrorString hipGetErrorString
 #define gpuGetDeviceProperties hipGetDeviceProperties
-// FIXME : use hipStreamDefault instead of 0x00
-#define gpuStreamDefault 0x00
+#define gpuStreamDefault hipStreamDefault
 #define gpuGetDevice hipGetDevice
 #define gpuSetDevice hipSetDevice
 #define gpuMalloc hipMalloc
@@ -47,6 +46,7 @@
 #define gpuSharedMemConfig hipSharedMemConfig
 #define gpuDeviceSetSharedMemConfig hipDeviceSetSharedMemConfig
 #define gpuStreamSynchronize hipStreamSynchronize
+#define gpuDeviceSynchronize hipDeviceSynchronize
 #define gpuMemcpy hipMemcpy
 
 #else
@@ -73,6 +73,7 @@
 #define gpuSharedMemConfig cudaSharedMemConfig
 #define gpuDeviceSetSharedMemConfig cudaDeviceSetSharedMemConfig
 #define gpuStreamSynchronize cudaStreamSynchronize
+#define gpuDeviceSynchronize cudaDeviceSynchronize
 #define gpuMemcpy cudaMemcpy
 
 #endif
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
index 9bc0708ed..db394bcbb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
@@ -32,6 +32,7 @@
 #undef gpuSharedMemConfig
 #undef gpuDeviceSetSharedMemConfig
 #undef gpuStreamSynchronize
+#undef gpuDeviceSynchronize
 #undef gpuMemcpy
 
 #undef EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index 835efbf72..8810d78cf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -351,10 +351,7 @@ struct IndexPairList : internal::IndexTuple<FirstType, OtherTypes...> {
 namespace internal {
 
 template<typename FirstType, typename... OtherTypes>
-  #if defined(EIGEN_HIPCC)
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
-  #endif
-  size_t array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
   size_t result = 1;
   for (int i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
     result *= sizes[i];
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 2a979845b..cda49f8fe 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -859,10 +859,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
     return inputIndex;
   }
   
-  #if defined(EIGEN_HIPCC)
-  EIGEN_DEVICE_FUNC
-  #endif
-  static EIGEN_STRONG_INLINE Index clamp(Index value, Index min, Index max) {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index clamp(Index value, Index min, Index max) {
 #ifndef __SYCL_DEVICE_ONLY__
     return numext::maxi(min, numext::mini(max,value));
 #else
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index fdd338b96..ce573d730 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -497,6 +497,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   EIGEN_STRONG_INLINE
     #if !defined(EIGEN_HIPCC)
+    // Marking this as EIGEN_DEVICE_FUNC for HIPCC requires also doing the same for all the functions
+    // being called within here, which then leads to proliferation of EIGEN_DEVICE_FUNC markings, one
+    // of which will eventually result in an NVCC error
     EIGEN_DEVICE_FUNC
     #endif
     bool evalSubExprsIfNeeded(typename MakePointer_<CoeffReturnType>::Type data) {
@@ -778,17 +781,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   // Indexed by reduced dimensions.
   array<Index, NumReducedDims> m_reducedDims;
 
-#if defined(EIGEN_HIPCC)
- public:
-#endif
-  
   // Evaluator for the input expression.
   TensorEvaluator<ArgType, Device> m_impl;
 
-#if defined(EIGEN_HIPCC)
- private:
-#endif
-  
   // Operation to apply for computing the reduction.
   Op m_reducer;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
index ca854d670..a691e530a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
@@ -169,7 +169,9 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
 #pragma unroll
   for (int offset = warpSize/2; offset > 0; offset /= 2) {
   #if defined(EIGEN_HIPCC)
-    // XXX use std::is_floating_point to determine the type of accum
+    // use std::is_floating_point to determine the type of reduced_val 
+    // This is needed because when Type == double, hipcc will give a "call to __shfl_down is ambguous" error 
+    // and list the float and int versions of __shfl_down as the candidate functions. 
     if (std::is_floating_point<typename Self::CoeffReturnType>::value) {
       reducer.reduce(__shfl_down(static_cast<float>(accum), offset, warpSize), &accum);
     } else {
@@ -238,20 +240,6 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
 
   // Initialize the output value if it wasn't initialized by the ReductionInitKernel
 
-#if defined(EIGEN_HIPCC)
-  
-  if (gridDim.x == 1 && first_index == 0) {
-    if (num_coeffs % 2 != 0) {
-      half last = input.m_impl.coeff(num_coeffs-1);
-      *scratch = __halves2half2(last, reducer.initialize());
-    } else {
-      *scratch = reducer.template initializePacket<half2>();
-    }
-    __syncthreads();
-  }
-  
-#else
-  
   if (gridDim.x == 1) {
     if (first_index == 0) {
       if (num_coeffs % 2 != 0) {
@@ -264,8 +252,6 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
     __syncthreads();
   }
   
-#endif
-
   half2 accum = reducer.template initializePacket<half2>();
   const Index max_iter = numext::mini<Index>((num_coeffs - first_index) / 2, NumPerThread*BlockSize / 2);
   for (Index i = 0; i < max_iter; i += BlockSize) {
@@ -295,17 +281,6 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
     atomicReduce(scratch, accum, reducer);
   }
 
-#if defined(EIGEN_HIPCC)
-  __syncthreads();
-
-  if (gridDim.x == 1 && first_index == 0) {
-    half tmp = __low2half(*scratch);
-    reducer.reduce(__high2half(*scratch), &tmp);
-    *output = tmp;
-  }
-
-#else
-    
   if (gridDim.x == 1) {
     __syncthreads();
     if (first_index == 0) {
@@ -314,8 +289,6 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
       *output = tmp;
     }
   }
-
-#endif
 }
 
 template <typename Op>
@@ -485,7 +458,9 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
 #pragma unroll
       for (int offset = warpSize/2; offset > 0; offset /= 2) {
       #if defined(EIGEN_HIPCC)
-        // XXX use std::is_floating_point to determine the type of reduced_val
+        // use std::is_floating_point to determine the type of reduced_val 
+	// This is needed because when Type == double, hipcc will give a "call to __shfl_down is ambguous" error 
+	// and list the float and int versions of __shfl_down as the candidate functions. 
         if (std::is_floating_point<Type>::value) {
           reducer.reduce(__shfl_down(static_cast<float>(reduced_val), offset), &reduced_val);
         } else {
@@ -847,8 +822,4 @@ struct OuterReducer<Self, Op, GpuDevice> {
 } // end namespace internal
 } // end namespace Eigen
 
-#if defined(EIGEN_HIPCC)
-#undef warpSize
-#endif
-
 #endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
diff --git a/unsupported/Eigen/SpecialFunctions b/unsupported/Eigen/SpecialFunctions
index 9441ba8f5..44fd99b43 100644
--- a/unsupported/Eigen/SpecialFunctions
+++ b/unsupported/Eigen/SpecialFunctions
@@ -53,8 +53,8 @@ namespace Eigen {
 #include "src/SpecialFunctions/SpecialFunctionsFunctors.h"
 #include "src/SpecialFunctions/SpecialFunctionsArrayAPI.h"
 
-#if defined EIGEN_VECTORIZE_CUDA
-  #include "src/SpecialFunctions/arch/CUDA/CudaSpecialFunctions.h"
+#if defined EIGEN_VECTORIZE_GPU
+  #include "src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h"
 #endif
 
 namespace Eigen {
diff --git a/unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h b/unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h
index ad011e305..40abcee3a 100644
--- a/unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h
+++ b/unsupported/Eigen/src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h
@@ -7,8 +7,8 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_CUDA_SPECIALFUNCTIONS_H
-#define EIGEN_CUDA_SPECIALFUNCTIONS_H
+#ifndef EIGEN_GPU_SPECIALFUNCTIONS_H
+#define EIGEN_GPU_SPECIALFUNCTIONS_H
 
 namespace Eigen {
 
@@ -223,4 +223,4 @@ pi1e<double2>(const double2& x) {
 
 } // end namespace Eigen
 
-#endif // EIGEN_CUDA_SPECIALFUNCTIONS_H
+#endif // EIGEN_GPU_SPECIALFUNCTIONS_H

From fb336877364eedbfbe6f865a213bd81ddcc0b15a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Jul 2018 17:08:30 +0200
Subject: [PATCH 420/680] Fix double ;;

---
 Eigen/src/Core/util/Memory.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 43920bdc8..aaa05a19c 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -720,7 +720,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
 
 
-#define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR);
+#define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR)
 
 #endif
 

From d193cc87f4ff26a3ea3187235015c06f005e3960 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Jul 2018 17:09:23 +0200
Subject: [PATCH 421/680] Fix regression in
 9357838f94d2907996adadc7e5200376f3561ed4

---
 Eigen/src/Core/CoreEvaluators.h | 6 +++---
 test/block.cpp                  | 9 ++++++++-
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index b65ec4ffb..4029b8da7 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1080,7 +1080,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
     : m_argImpl(block.nestedExpression()), 
       m_startRow(block.startRow()), 
       m_startCol(block.startCol()),
-      m_linear_offset((InnerPanel|| XprType::IsVectorAtCompileTime)?(XprType::IsRowMajor ? block.startRow()*block.cols() + block.startCol() : block.startCol()*block.rows() + block.startRow()):0)
+      m_linear_offset(ForwardLinearAccess?(ArgType::IsRowMajor ? block.startRow()*block.nestedExpression().cols() + block.startCol() : block.startCol()*block.nestedExpression().rows() + block.startRow()):0)
   { }
  
   typedef typename XprType::Scalar Scalar;
@@ -1088,7 +1088,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
 
   enum {
     RowsAtCompileTime = XprType::RowsAtCompileTime,
-    ForwardLinearAccess = (InnerPanel || XprType::IsVectorAtCompileTime) && bool(evaluator<ArgType>::Flags&LinearAccessBit)
+    ForwardLinearAccess = (InnerPanel || XprType::IsRowMajor==ArgType::IsRowMajor) && bool(evaluator<ArgType>::Flags&LinearAccessBit)
   };
  
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -1162,7 +1162,7 @@ protected:
   evaluator<ArgType> m_argImpl;
   const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
   const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
-  const variable_if_dynamic<Index, (InnerPanel || XprType::IsVectorAtCompileTime) ? Dynamic : 0> m_linear_offset;
+  const variable_if_dynamic<Index, ForwardLinearAccess ? Dynamic : 0> m_linear_offset;
 };
 
 // TODO: This evaluator does not actually use the child evaluator; 
diff --git a/test/block.cpp b/test/block.cpp
index 9c2424662..0c1d2b7b8 100644
--- a/test/block.cpp
+++ b/test/block.cpp
@@ -166,7 +166,14 @@ template<typename MatrixType> void block(const MatrixType& m)
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).col(0)) , ((m1+m2).col(c1).segment(r1,r2-r1+1)) );
   VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).transpose().col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
   VERIFY_IS_APPROX( ((m1+m2).transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() );
-  VERIFY_IS_APPROX( ((m1+m2).template block<1,Dynamic>(r1,c1,1,c2-c1+1)) , ((m1+m2).eval().row(r1).segment(c1,c2-c1+1)) );
+  VERIFY_IS_APPROX( ((m1+m2).template block<Dynamic,1>(r1,c1,r2-r1+1,1)) , ((m1+m2).eval().col(c1).eval().segment(r1,r2-r1+1)) );
+  VERIFY_IS_APPROX( ((m1+m2).template block<1,Dynamic>(r1,c1,1,c2-c1+1)) , ((m1+m2).eval().row(r1).eval().segment(c1,c2-c1+1)) );
+  VERIFY_IS_APPROX( ((m1+m2).transpose().template block<1,Dynamic>(c1,r1,1,r2-r1+1)) , ((m1+m2).eval().col(c1).eval().segment(r1,r2-r1+1)).transpose() );
+  VERIFY_IS_APPROX( (m1+m2).row(r1).eval(), (m1+m2).eval().row(r1) );
+  VERIFY_IS_APPROX( (m1+m2).adjoint().col(r1).eval(), (m1+m2).adjoint().eval().col(r1) );
+  VERIFY_IS_APPROX( (m1+m2).adjoint().row(c1).eval(), (m1+m2).adjoint().eval().row(c1) );
+  VERIFY_IS_APPROX( (m1*1).row(r1).segment(c1,c2-c1+1).eval(), m1.row(r1).eval().segment(c1,c2-c1+1).eval() );
+  VERIFY_IS_APPROX( m1.col(c1).reverse().segment(r1,r2-r1+1).eval(),m1.col(c1).reverse().eval().segment(r1,r2-r1+1).eval() );
 
   VERIFY_IS_APPROX( (m1*1).topRows(r1),  m1.topRows(r1) );
   VERIFY_IS_APPROX( (m1*1).leftCols(c1), m1.leftCols(c1) );

From 8a5955a052f64fc32099b052a221e0b5a8e92cbd Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 11 Jul 2018 17:16:50 +0200
Subject: [PATCH 422/680] Optimize the product of a householder-sequence with
 the identity, and optimize the evaluation of a HouseholderSequence to a dense
 matrix using faster blocked product.

---
 CTestConfig.cmake                           |  8 ++---
 Eigen/src/Householder/HouseholderSequence.h | 40 +++++++++++++++------
 bench/bench_gemm.cpp                        |  2 +-
 3 files changed, 35 insertions(+), 15 deletions(-)

diff --git a/CTestConfig.cmake b/CTestConfig.cmake
index 4c0027824..8b4cd798e 100644
--- a/CTestConfig.cmake
+++ b/CTestConfig.cmake
@@ -11,7 +11,7 @@ set(CTEST_DROP_METHOD "http")
 set(CTEST_DROP_SITE "manao.inria.fr")
 set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen")
 set(CTEST_DROP_SITE_CDASH TRUE)
-set(CTEST_PROJECT_SUBPROJECTS
-Official
-Unsupported
-)
+#set(CTEST_PROJECT_SUBPROJECTS
+#Official
+#Unsupported
+#)
diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h
index a4f40b75c..af19701a4 100644
--- a/Eigen/src/Householder/HouseholderSequence.h
+++ b/Eigen/src/Householder/HouseholderSequence.h
@@ -295,6 +295,14 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
         for(Index k = 0; k<cols()-vecs ; ++k)
           dst.col(k).tail(rows()-k-1).setZero();
       }
+      else if(m_length>BlockSize)
+      {
+        dst.setIdentity(rows(), rows());
+        if(m_reverse)
+          applyThisOnTheLeft(dst,workspace,true);
+        else
+          applyThisOnTheLeft(dst,workspace,true);
+      }
       else
       {
         dst.setIdentity(rows(), rows());
@@ -332,24 +340,27 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     }
 
     /** \internal */
-    template<typename Dest> inline void applyThisOnTheLeft(Dest& dst) const
+    template<typename Dest> inline void applyThisOnTheLeft(Dest& dst, bool inputIsIdentity = false) const
     {
       Matrix<Scalar,1,Dest::ColsAtCompileTime,RowMajor,1,Dest::MaxColsAtCompileTime> workspace;
-      applyThisOnTheLeft(dst, workspace);
+      applyThisOnTheLeft(dst, workspace, inputIsIdentity);
     }
 
     /** \internal */
     template<typename Dest, typename Workspace>
-    inline void applyThisOnTheLeft(Dest& dst, Workspace& workspace) const
+    inline void applyThisOnTheLeft(Dest& dst, Workspace& workspace, bool inputIsIdentity = false) const
     {
-      const Index BlockSize = 48;
+      if(inputIsIdentity && m_reverse)
+        inputIsIdentity = false;
       // if the entries are large enough, then apply the reflectors by block
       if(m_length>=BlockSize && dst.cols()>1)
       {
-        for(Index i = 0; i < m_length; i+=BlockSize)
+        // Make sure we have at least 2 useful blocks, otherwise it is point-less:
+        Index blockSize = m_length<2*BlockSize ? (m_length+1)/2 : BlockSize;
+        for(Index i = 0; i < m_length; i+=blockSize)
         {
-          Index end = m_reverse ? (std::min)(m_length,i+BlockSize) : m_length-i;
-          Index k = m_reverse ? i : (std::max)(Index(0),end-BlockSize);
+          Index end = m_reverse ? (std::min)(m_length,i+blockSize) : m_length-i;
+          Index k = m_reverse ? i : (std::max)(Index(0),end-blockSize);
           Index bs = end-k;
           Index start = k + m_shift;
           
@@ -359,7 +370,14 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
                                                                    Side==OnTheRight ? bs : m_vectors.rows()-start,
                                                                    Side==OnTheRight ? m_vectors.cols()-start : bs);
           typename internal::conditional<Side==OnTheRight, Transpose<SubVectorsType>, SubVectorsType&>::type sub_vecs(sub_vecs1);
-          Block<Dest,Dynamic,Dynamic> sub_dst(dst,dst.rows()-rows()+m_shift+k,0, rows()-m_shift-k,dst.cols());
+
+          Index dstStart = dst.rows()-rows()+m_shift+k;
+          Index dstRows  = rows()-m_shift-k;
+          Block<Dest,Dynamic,Dynamic> sub_dst(dst,
+                                              dstStart,
+                                              inputIsIdentity ? dstStart : 0,
+                                              dstRows,
+                                              inputIsIdentity ? dstRows : dst.cols());
           apply_block_householder_on_the_left(sub_dst, sub_vecs, m_coeffs.segment(k, bs), !m_reverse);
         }
       }
@@ -369,7 +387,8 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
         for(Index k = 0; k < m_length; ++k)
         {
           Index actual_k = m_reverse ? k : m_length-k-1;
-          dst.bottomRows(rows()-m_shift-actual_k)
+          Index dstStart = rows()-m_shift-actual_k;
+          dst.bottomRightCorner(dstStart, inputIsIdentity ? dstStart : dst.cols())
             .applyHouseholderOnTheLeft(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data());
         }
       }
@@ -387,7 +406,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     {
       typename internal::matrix_type_times_scalar_type<Scalar, OtherDerived>::Type
         res(other.template cast<typename internal::matrix_type_times_scalar_type<Scalar,OtherDerived>::ResultScalar>());
-      applyThisOnTheLeft(res);
+      applyThisOnTheLeft(res, internal::is_identity<OtherDerived>::value && res.rows()==res.cols());
       return res;
     }
 
@@ -461,6 +480,7 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     bool m_reverse;
     Index m_length;
     Index m_shift;
+    enum { BlockSize = 48 };
 };
 
 /** \brief Computes the product of a matrix with a Householder sequence.
diff --git a/bench/bench_gemm.cpp b/bench/bench_gemm.cpp
index 8528c5587..688d99c4a 100644
--- a/bench/bench_gemm.cpp
+++ b/bench/bench_gemm.cpp
@@ -129,7 +129,7 @@ void matlab_cplx_real(const M& ar, const M& ai, const M& b, M& cr, M& ci)
 template<typename A, typename B, typename C>
 EIGEN_DONT_INLINE void gemm(const A& a, const B& b, C& c)
 {
- c.noalias() += a * b;
+  c.noalias() += a * b;
 }
 
 int main(int argc, char ** argv)

From 21cf4a1a8b0868b6dcb74aa4ebe6317d5316b2e1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 09:57:19 +0200
Subject: [PATCH 423/680] Make is_convertible more robust and conformant to
 std::is_convertible

---
 Eigen/src/Core/util/Meta.h | 12 +++++++-----
 test/meta.cpp              | 36 +++++++++++++++++++++++++++++-------
 2 files changed, 36 insertions(+), 12 deletions(-)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index ef9860c4b..3d2bdd12e 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -145,16 +145,19 @@ private:
   struct yes {int a[1];};
   struct no  {int a[2];};
 
-  static yes test(const To&, int);
+  template<typename T>
+  static yes test(T, int);
+
+  template<typename T>
   static no  test(any_conversion, ...);
 
 public:
-  static From ms_from;
+  static From* ms_from;
 #ifdef __INTEL_COMPILER
   #pragma warning push
   #pragma warning ( disable : 2259 )
 #endif
-  enum { value = sizeof(test(ms_from, 0))==sizeof(yes) };
+  enum { value = sizeof(test<To>(*ms_from, 0))==sizeof(yes) };
 #ifdef __INTEL_COMPILER
   #pragma warning pop
 #endif
@@ -163,8 +166,7 @@ public:
 template<typename From, typename To>
 struct is_convertible
 {
-  enum { value = is_convertible_impl<typename remove_all<From>::type,
-                                     typename remove_all<To  >::type>::value };
+  enum { value = is_convertible_impl<From,To>::value };
 };
 
 /** \internal Allows to enable/disable an overload
diff --git a/test/meta.cpp b/test/meta.cpp
index bd505762e..4904f3ba4 100644
--- a/test/meta.cpp
+++ b/test/meta.cpp
@@ -19,6 +19,14 @@ struct FooReturnType {
   typedef int ReturnType;
 };
 
+struct MyInterface {
+  virtual void func() = 0;
+  virtual ~MyInterface() {}
+};
+struct MyImpl : public MyInterface {
+  void func() {}
+};
+
 void test_meta()
 {
   VERIFY((internal::conditional<(3<4),internal::true_type, internal::false_type>::type::value));
@@ -62,14 +70,19 @@ void test_meta()
   VERIFY(( internal::is_same<const float,internal::remove_pointer<const float*>::type >::value));
   VERIFY(( internal::is_same<float,internal::remove_pointer<float* const >::type >::value));
   
-  VERIFY(( internal::is_convertible<float,double>::value ));
-  VERIFY(( internal::is_convertible<int,double>::value ));
-  VERIFY(( internal::is_convertible<double,int>::value ));
-  VERIFY((!internal::is_convertible<std::complex<double>,double>::value ));
-  VERIFY(( internal::is_convertible<Array33f,Matrix3f>::value ));
+
+  // is_convertible
+  STATIC_CHECK(( internal::is_convertible<float,double>::value ));
+  STATIC_CHECK(( internal::is_convertible<int,double>::value ));
+  STATIC_CHECK(( internal::is_convertible<int, short>::value ));
+  STATIC_CHECK(( internal::is_convertible<short, int>::value ));
+  STATIC_CHECK(( internal::is_convertible<double,int>::value ));
+  STATIC_CHECK(( internal::is_convertible<double,std::complex<double> >::value ));
+  STATIC_CHECK((!internal::is_convertible<std::complex<double>,double>::value ));
+  STATIC_CHECK(( internal::is_convertible<Array33f,Matrix3f>::value ));
 //   VERIFY((!internal::is_convertible<Matrix3f,Matrix3d>::value )); //does not work because the conversion is prevented by a static assertion
-  VERIFY((!internal::is_convertible<Array33f,int>::value ));
-  VERIFY((!internal::is_convertible<MatrixXf,float>::value ));
+  STATIC_CHECK((!internal::is_convertible<Array33f,int>::value ));
+  STATIC_CHECK((!internal::is_convertible<MatrixXf,float>::value ));
   {
     float f;
     MatrixXf A, B;
@@ -80,6 +93,15 @@ void test_meta()
     VERIFY(( check_is_convertible(A*B, A) ));
   }
 
+  STATIC_CHECK(( !internal::is_convertible<MyInterface, MyImpl>::value ));
+  STATIC_CHECK(( !internal::is_convertible<MyImpl, MyInterface>::value ));
+  STATIC_CHECK((  internal::is_convertible<MyImpl, const MyInterface&>::value ));
+  {
+    int i;
+    VERIFY(( check_is_convertible(fix<3>(), i) ));
+    VERIFY((!check_is_convertible(i, fix<DynamicIndex>()) ));
+  }
+
   VERIFY((  internal::has_ReturnType<FooReturnType>::value ));
   VERIFY((  internal::has_ReturnType<ScalarBinaryOpTraits<int,int> >::value ));
   VERIFY(( !internal::has_ReturnType<MatrixXf>::value ));

From 8ef267ccbd3ef04af5b725a22599ab11b9da2fbf Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 09:58:29 +0200
Subject: [PATCH 424/680] spellcheck

---
 Eigen/src/Core/util/IndexedViewHelper.h | 2 +-
 Eigen/src/Core/util/SymbolicIndex.h     | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/util/IndexedViewHelper.h b/Eigen/src/Core/util/IndexedViewHelper.h
index ab01c857f..dcba731f0 100644
--- a/Eigen/src/Core/util/IndexedViewHelper.h
+++ b/Eigen/src/Core/util/IndexedViewHelper.h
@@ -117,7 +117,7 @@ template<> struct get_compile_time_incr<SingleRange> {
   enum { value = 1 }; // 1 or 0 ??
 };
 
-// Turn a single index into something that looks like an array (i.e., that exposes a .size(), and operatro[](int) methods)
+// Turn a single index into something that looks like an array (i.e., that exposes a .size(), and operator[](int) methods)
 template<typename T, int XprSize>
 struct IndexedViewCompatibleType<T,XprSize,typename internal::enable_if<internal::is_integral<T>::value>::type> {
   // Here we could simply use Array, but maybe it's less work for the compiler to use
diff --git a/Eigen/src/Core/util/SymbolicIndex.h b/Eigen/src/Core/util/SymbolicIndex.h
index bb6349eb9..1ade6d67b 100644
--- a/Eigen/src/Core/util/SymbolicIndex.h
+++ b/Eigen/src/Core/util/SymbolicIndex.h
@@ -35,7 +35,7 @@ namespace Eigen {
   * std::cout << expr98.eval(x=6) << "\n";
   * \endcode
   *
-  * It is currently only used internally to define and minipulate the placeholders::last and placeholders::end symbols in Eigen::seq and Eigen::seqN.
+  * It is currently only used internally to define and manipulate the placeholders::last and placeholders::end symbols in Eigen::seq and Eigen::seqN.
   *
   */
 namespace Symbolic {
@@ -187,7 +187,7 @@ public:
 
 template<typename T>
 struct is_symbolic {
-  // BaseExpr has no conversion ctor, so we only have to check whether T can be staticaly cast to its base class BaseExpr<T>.
+  // BaseExpr has no conversion ctor, so we only have to check whether T can be statically cast to its base class BaseExpr<T>.
   enum { value = internal::is_convertible<T,BaseExpr<T> >::value };
 };
 

From 8a40dda5a6c7847b466ca4d10d177639d1646e6d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 09:59:00 +0200
Subject: [PATCH 425/680] Add some basic unit-tests

---
 test/symbolic_index.cpp | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/test/symbolic_index.cpp b/test/symbolic_index.cpp
index 1db85144b..4c079857b 100644
--- a/test/symbolic_index.cpp
+++ b/test/symbolic_index.cpp
@@ -58,6 +58,15 @@ bool is_same_symb(const T1& a, const T2& b, Index size)
   return a.eval(last=size-1) == b.eval(last=size-1);
 }
 
+template<typename T>
+void check_is_symbolic(const T&) {
+  STATIC_CHECK(( Symbolic::is_symbolic<T>::value ))
+}
+
+template<typename T>
+void check_isnot_symbolic(const T&) {
+  STATIC_CHECK(( !Symbolic::is_symbolic<T>::value ))
+}
 
 #define VERIFY_EQ_INT(A,B) VERIFY_IS_APPROX(int(A),int(B))
 
@@ -66,6 +75,13 @@ void check_symbolic_index()
   using Eigen::placeholders::last;
   using Eigen::placeholders::end;
 
+  check_is_symbolic(last);
+  check_is_symbolic(end);
+  check_is_symbolic(last+1);
+  check_is_symbolic(last-end);
+  check_is_symbolic(2*last-end/2);
+  check_isnot_symbolic(fix<3>());
+
   Index size=100;
 
   // First, let's check FixedInt arithmetic:

From a4ea611ca7fa887f7b129a81942c8fa98ae4b6dd Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 09:59:44 +0200
Subject: [PATCH 426/680] Remove useless specialization thanks to
 is_convertible being more robust.

---
 Eigen/src/Core/util/SymbolicIndex.h | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/Eigen/src/Core/util/SymbolicIndex.h b/Eigen/src/Core/util/SymbolicIndex.h
index 1ade6d67b..b9fe733b7 100644
--- a/Eigen/src/Core/util/SymbolicIndex.h
+++ b/Eigen/src/Core/util/SymbolicIndex.h
@@ -191,13 +191,6 @@ struct is_symbolic {
   enum { value = internal::is_convertible<T,BaseExpr<T> >::value };
 };
 
-// Specialization for functions, because is_convertible fails in this case.
-// Useful in c++98/11 mode when testing is_symbolic<decltype(fix<N>)>
-template<typename T>
-struct is_symbolic<T (*)()> {
-  enum { value = false };
-};
-
 /** Represents the actual value of a symbol identified by its tag
   *
   * It is the return type of SymbolValue::operator=, and most of the time this is only way it is used.

From 6cd6551b264f720e2b44f0d14e1f78becb002b5e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 10:50:53 +0200
Subject: [PATCH 427/680] Add deprecated header files for TensorFlow

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h | 6 ++++++
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h      | 6 ++++++
 unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h   | 6 ++++++
 3 files changed, 18 insertions(+)
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
new file mode 100644
index 000000000..3f315fedc
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorContractionGpu.h file"
+#endif
+
+#include "TensorContractionGpu.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
new file mode 100644
index 000000000..f77923933
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorDeviceGpu.h file"
+#endif
+
+#include "TensorDeviceGpu.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
new file mode 100644
index 000000000..68780cd3c
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorReductionGpu.h file"
+#endif
+
+#include "TensorReductionGpu.h"

From a9060378d32411043f3e6995092d10070fa0ce1b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 11:07:09 +0200
Subject: [PATCH 428/680] bug #1570: fix warning

---
 Eigen/src/Core/CoreEvaluators.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 4029b8da7..264446f65 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -1088,7 +1088,7 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
 
   enum {
     RowsAtCompileTime = XprType::RowsAtCompileTime,
-    ForwardLinearAccess = (InnerPanel || XprType::IsRowMajor==ArgType::IsRowMajor) && bool(evaluator<ArgType>::Flags&LinearAccessBit)
+    ForwardLinearAccess = (InnerPanel || int(XprType::IsRowMajor)==int(ArgType::IsRowMajor)) && bool(evaluator<ArgType>::Flags&LinearAccessBit)
   };
  
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE

From 8bdb214fd0513e954dd2da86989338144b9ac5b0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 11:17:53 +0200
Subject: [PATCH 429/680] remove double ;;

---
 Eigen/src/Core/arch/SSE/PacketMath.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index c944d2c0e..2e815c0c5 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -663,7 +663,7 @@ template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
   // TODO try to call _mm_mul_epu32 directly
   EIGEN_ALIGN16 int aux[4];
   pstore(aux, a);
-  return  (aux[0] * aux[1]) * (aux[2] * aux[3]);;
+  return  (aux[0] * aux[1]) * (aux[2] * aux[3]);
 }
 
 // min

From a12b8a8c75ebef312509da643424951725519348 Mon Sep 17 00:00:00 2001
From: "Daniele E. Domenichelli" <daniele.domenichelli@iit.it>
Date: Wed, 11 Jul 2018 16:31:50 +0200
Subject: [PATCH 430/680] FindEigen3: Set Eigen3_FOUND variable

---
 cmake/FindEigen3.cmake | 1 +
 1 file changed, 1 insertion(+)

diff --git a/cmake/FindEigen3.cmake b/cmake/FindEigen3.cmake
index 52efb4e15..4d9bc1a52 100644
--- a/cmake/FindEigen3.cmake
+++ b/cmake/FindEigen3.cmake
@@ -68,6 +68,7 @@ if (EIGEN3_INCLUDE_DIR)
   # in cache already
   _eigen3_check_version()
   set(EIGEN3_FOUND ${EIGEN3_VERSION_OK})
+  set(Eigen3_FOUND ${EIGEN3_VERSION_OK})
 
 else (EIGEN3_INCLUDE_DIR)
   

From 6d451cf2b6a6122beb5c12259eaec28d891ee083 Mon Sep 17 00:00:00 2001
From: Julian Kent <anonymous@invalid.net>
Date: Sat, 10 Feb 2018 13:44:05 +0100
Subject: [PATCH 431/680] Add missing consts for rows and cols functions in
 SparseLU

---
 Eigen/src/SparseLU/SparseLU.h                  | 8 ++++----
 Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h | 4 ++--
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h
index 3bc06f4e3..5d2419353 100644
--- a/Eigen/src/SparseLU/SparseLU.h
+++ b/Eigen/src/SparseLU/SparseLU.h
@@ -706,8 +706,8 @@ struct SparseLUMatrixLReturnType : internal::no_assignment_operator
   typedef typename MappedSupernodalType::Scalar Scalar;
   explicit SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL)
   { }
-  Index rows() { return m_mapL.rows(); }
-  Index cols() { return m_mapL.cols(); }
+  Index rows() const { return m_mapL.rows(); }
+  Index cols() const { return m_mapL.cols(); }
   template<typename Dest>
   void solveInPlace( MatrixBase<Dest> &X) const
   {
@@ -723,8 +723,8 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator
   SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU)
   : m_mapL(mapL),m_mapU(mapU)
   { }
-  Index rows() { return m_mapL.rows(); }
-  Index cols() { return m_mapL.cols(); }
+  Index rows() const { return m_mapL.rows(); }
+  Index cols() const { return m_mapL.cols(); }
 
   template<typename Dest>   void solveInPlace(MatrixBase<Dest> &X) const
   {
diff --git a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h
index 721e1883b..8583b1b69 100644
--- a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h
+++ b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h
@@ -75,12 +75,12 @@ class MappedSuperNodalMatrix
     /**
      * Number of rows
      */
-    Index rows() { return m_row; }
+    Index rows() const { return m_row; }
     
     /**
      * Number of columns
      */
-    Index cols() { return m_col; }
+    Index cols() const { return m_col; }
     
     /**
      * Return the array of nonzero values packed by column

From 937ad18221e940cef96617ade46e6f84e81d0b9e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 16 Feb 2018 22:25:11 +0400
Subject: [PATCH 432/680] add unit test for SimplicialCholesky and Boost
 multiprec.

---
 test/boostmultiprec.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/test/boostmultiprec.cpp b/test/boostmultiprec.cpp
index e06e9bdaf..bafbd3b96 100644
--- a/test/boostmultiprec.cpp
+++ b/test/boostmultiprec.cpp
@@ -55,6 +55,10 @@
 #include "bdcsvd.cpp"
 #endif
 
+#ifdef EIGEN_TEST_PART_11
+#include "simplicial_cholesky.cpp"
+#endif
+
 #include <Eigen/Dense>
 
 #undef min
@@ -197,5 +201,7 @@ void test_boostmultiprec()
 
   CALL_SUBTEST_9(( jacobisvd(Mat(internal::random<int>(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE), internal::random<int>(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) ));
   CALL_SUBTEST_10(( bdcsvd(Mat(internal::random<int>(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE), internal::random<int>(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) ));
+
+  CALL_SUBTEST_11(( test_simplicial_cholesky_T<Real,int>() ));
 }
 

From adb134d47e70906ed15d877d7a0fe03f3f665c02 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 16 Feb 2018 22:26:01 +0400
Subject: [PATCH 433/680] Fix implicit conversion from 0.0 to scalar

---
 Eigen/src/OrderingMethods/Ordering.h               | 2 +-
 Eigen/src/SparseCholesky/SimplicialCholesky_impl.h | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/OrderingMethods/Ordering.h b/Eigen/src/OrderingMethods/Ordering.h
index 7ea9b14d7..34dbef487 100644
--- a/Eigen/src/OrderingMethods/Ordering.h
+++ b/Eigen/src/OrderingMethods/Ordering.h
@@ -31,7 +31,7 @@ void ordering_helper_at_plus_a(const MatrixType& A, MatrixType& symmat)
   for (int i = 0; i < C.rows(); i++) 
   {
       for (typename MatrixType::InnerIterator it(C, i); it; ++it)
-        it.valueRef() = 0.0;
+        it.valueRef() = typename MatrixType::Scalar(0);
   }
   symmat = C + A;
 }
diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
index 31e06995b..84a1bf2bd 100644
--- a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
+++ b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
@@ -122,7 +122,7 @@ void SimplicialCholeskyBase<Derived>::factorize_preordered(const CholMatrixType&
   for(StorageIndex k = 0; k < size; ++k)
   {
     // compute nonzero pattern of kth row of L, in topological order
-    y[k] = 0.0;                     // Y(0:k) is now all zero
+    y[k] = Scalar(0);                     // Y(0:k) is now all zero
     StorageIndex top = size;               // stack for pattern is empty
     tags[k] = k;                    // mark node k as visited
     m_nonZerosPerCol[k] = 0;        // count of nonzeros in column k of L
@@ -146,12 +146,12 @@ void SimplicialCholeskyBase<Derived>::factorize_preordered(const CholMatrixType&
     /* compute numerical values kth row of L (a sparse triangular solve) */
 
     RealScalar d = numext::real(y[k]) * m_shiftScale + m_shiftOffset;    // get D(k,k), apply the shift function, and clear Y(k)
-    y[k] = 0.0;
+    y[k] = Scalar(0);
     for(; top < size; ++top)
     {
       Index i = pattern[top];       /* pattern[top:n-1] is pattern of L(:,k) */
       Scalar yi = y[i];             /* get and clear Y(i) */
-      y[i] = 0.0;
+      y[i] = Scalar(0);
 
       /* the nonzero entry L(k,i) */
       Scalar l_ki;

From 1eff6cf8a77f1b8699671d31f8f307a6fd9170ea Mon Sep 17 00:00:00 2001
From: Yuefeng Zhou <yuefengz@google.com>
Date: Tue, 20 Feb 2018 16:50:05 -0800
Subject: [PATCH 434/680] Use device's allocate function instead of
 internal::aligned_malloc. This would make it easier to track memory usage in
 device instances.

---
 .../Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h      | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index d30cc96ab..6fb69910e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -377,7 +377,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
           divup<size_t>(bm_ * bk_ * sizeof(LhsScalar), align) * align;
       size_t rhs_size =
           divup<size_t>(bn_ * bk_ * sizeof(RhsScalar), align) * align;
-      packed_mem_ = static_cast<char*>(internal::aligned_malloc(
+      packed_mem_ = static_cast<char*>(device_.allocate(
           (nm0_ * lhs_size + nn0_ * rhs_size) * std::min<size_t>(nk_, P - 1)));
       char* mem = static_cast<char*>(packed_mem_);
       for (Index x = 0; x < numext::mini<Index>(nk_, P - 1); x++) {
@@ -399,7 +399,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         for (Index m = 0; m < nm_; m++) delete[] state_kernel_[x][m];
         delete[] state_kernel_[x];
       }
-      internal::aligned_free(packed_mem_);
+      device_.deallocate(packed_mem_);
     }
 
     void run() {

From e79c5149bf049994b86ab19aac24b0dde2bcc387 Mon Sep 17 00:00:00 2001
From: Mark D Ryan <mark.d.ryan@intel.com>
Date: Mon, 25 Jun 2018 05:05:02 -0700
Subject: [PATCH 435/680] Fix AVX512 implementations of psqrt

This commit fixes the AVX512 implementations of psqrt in the same
way that 3ed67cb0bb4af65fbf243df598604a8c7630bf7d
 fixed the AVX2 version of this function.  The
AVX512 versions of psqrt incorrectly return -0.0 for negative
values, instead of NaN.  Fixing the issues requires adding
some additional instructions that slow down the algorithms.  A
similar test to the one used in 3ed67cb0bb4af65fbf243df598604a8c7630bf7d
 shows that the
corrected Packet16f code runs at 73% of the speed of the existing code,
while the corrected Packed8d function runs at 68% of the original.
---
 Eigen/src/Core/arch/AVX512/MathFunctions.h | 47 +++++++++-------------
 1 file changed, 19 insertions(+), 28 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 9c1717f76..81a3b4f62 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -258,48 +258,39 @@ pexp<Packet8d>(const Packet8d& _x) {
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
 psqrt<Packet16f>(const Packet16f& _x) {
-  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
-  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
-  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+  Packet16f half = pmul(_x, pset1<Packet16f>(.5f));
+  __mmask16 denormal_mask = _mm512_kand(
+      _mm512_cmp_ps_mask(_x, pset1<Packet16f>((std::numeric_limits<float>::min)()),
+                        _CMP_LT_OQ),
+      _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_GE_OQ));
 
-  Packet16f neg_half = pmul(_x, p16f_minus_half);
-
-  // select only the inverse sqrt of positive normal inputs (denormals are
-  // flushed to zero and cause infs as well).
-  __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ);
-  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_setzero_ps(), _mm512_rsqrt14_ps(_x));
+  Packet16f x = _mm512_rsqrt14_ps(_x);
 
   // Do a single step of Newton's iteration.
-  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+  x = pmul(x, psub(pset1<Packet16f>(1.5f), pmul(half, pmul(x,x))));
 
-  // Multiply the original _x by it's reciprocal square root to extract the
-  // square root.
-  return pmul(_x, x);
+  // Flush results for denormals to zero.
+  return _mm512_mask_blend_ps(denormal_mask, pmul(_x,x), _mm512_setzero_ps());
 }
 
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
 psqrt<Packet8d>(const Packet8d& _x) {
-  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
-  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
-  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+  Packet8d half = pmul(_x, pset1<Packet8d>(.5f));
+  __mmask16 denormal_mask = _mm512_kand(
+      _mm512_cmp_pd_mask(_x, pset1<Packet8d>((std::numeric_limits<double>::min)()),
+                        _CMP_LT_OQ),
+      _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_GE_OQ));
 
-  Packet8d neg_half = pmul(_x, p8d_minus_half);
+  Packet8d x = _mm512_rsqrt14_pd(_x);
 
-  // select only the inverse sqrt of positive normal inputs (denormals are
-  // flushed to zero and cause infs as well).
-  __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ);
-  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_setzero_pd(), _mm512_rsqrt14_pd(_x));
-
-  // Do a first step of Newton's iteration.
-  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+  // Do a single step of Newton's iteration.
+  x = pmul(x, psub(pset1<Packet8d>(1.5f), pmul(half, pmul(x,x))));
 
   // Do a second step of Newton's iteration.
-  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+  x = pmul(x, psub(pset1<Packet8d>(1.5f), pmul(half, pmul(x,x))));
 
-  // Multiply the original _x by it's reciprocal square root to extract the
-  // square root.
-  return pmul(_x, x);
+  return _mm512_mask_blend_pd(denormal_mask, pmul(_x,x), _mm512_setzero_pd());
 }
 #else
 template <>

From b347eb0b1c6fd0c9ee09d3493e72d68a2d514db5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 11:56:18 +0200
Subject: [PATCH 436/680] Fix doc

---
 unsupported/Eigen/NonLinearOptimization | 32 +++++++++++++++----------
 1 file changed, 19 insertions(+), 13 deletions(-)

diff --git a/unsupported/Eigen/NonLinearOptimization b/unsupported/Eigen/NonLinearOptimization
index c22b89054..28eaa5f29 100644
--- a/unsupported/Eigen/NonLinearOptimization
+++ b/unsupported/Eigen/NonLinearOptimization
@@ -30,8 +30,8 @@
   * actually linear. But if this is so, you should probably better use other
   * methods more fitted to this special case.
   *
-  * One algorithm allows to find an extremum of such a system (Levenberg
-  * Marquardt algorithm) and the second one is used to find 
+  * One algorithm allows to find a least-squares solution of such a system
+  * (Levenberg-Marquardt algorithm) and the second one is used to find 
   * a zero for the system (Powell hybrid "dogleg" method).
   *
   * This code is a port of minpack (http://en.wikipedia.org/wiki/MINPACK).
@@ -58,35 +58,41 @@
   * There are two kinds of tests : those that come from examples bundled with cminpack.
   * They guaranty we get the same results as the original algorithms (value for 'x',
   * for the number of evaluations of the function, and for the number of evaluations
-  * of the jacobian if ever).
+  * of the Jacobian if ever).
   * 
   * Other tests were added by myself at the very beginning of the 
-  * process and check the results for levenberg-marquardt using the reference data 
+  * process and check the results for Levenberg-Marquardt using the reference data 
   * on http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml. Since then i've 
   * carefully checked that the same results were obtained when modifying the
   * code. Please note that we do not always get the exact same decimals as they do,
   * but this is ok : they use 128bits float, and we do the tests using the C type 'double',
   * which is 64 bits on most platforms (x86 and amd64, at least).
-  * I've performed those tests on several other implementations of levenberg-marquardt, and
+  * I've performed those tests on several other implementations of Levenberg-Marquardt, and
   * (c)minpack performs VERY well compared to those, both in accuracy and speed.
   * 
   * The documentation for running the tests is on the wiki
   * http://eigen.tuxfamily.org/index.php?title=Tests
   * 
-  * \section API API : overview of methods
+  * \section API API: overview of methods
   * 
-  * Both algorithms can use either the jacobian (provided by the user) or compute 
-  * an approximation by themselves (actually using Eigen \ref NumericalDiff_Module).
-  * The part of API referring to the latter use 'NumericalDiff' in the method names
-  * (exemple: LevenbergMarquardt.minimizeNumericalDiff() ) 
+  * Both algorithms needs a functor computing the Jacobian. It can be computed by
+  * hand, using auto-differentiation (see \ref AutoDiff_Module), or using numerical
+  * differences (see \ref NumericalDiff_Module). For instance:
+  *\code
+  * MyFunc func;
+  * NumericalDiff<MyFunc> func_with_num_diff(func);
+  * LevenbergMarquardt<NumericalDiff<MyFunc> > lm(func_with_num_diff);
+  * \endcode
+  * For HybridNonLinearSolver, the method solveNumericalDiff() does the above wrapping for
+  * you.
   * 
   * The methods LevenbergMarquardt.lmder1()/lmdif1()/lmstr1() and 
   * HybridNonLinearSolver.hybrj1()/hybrd1() are specific methods from the original 
   * minpack package that you probably should NOT use until you are porting a code that
-  *  was previously using minpack. They just define a 'simple' API with default values 
+  * was previously using minpack. They just define a 'simple' API with default values 
   * for some parameters.
   * 
-  * All algorithms are provided using Two APIs :
+  * All algorithms are provided using two APIs :
   *     - one where the user inits the algorithm, and uses '*OneStep()' as much as he wants : 
   * this way the caller have control over the steps
   *     - one where the user just calls a method (optimize() or solve()) which will 
@@ -94,7 +100,7 @@
   *  convenience.
   * 
   * As an example, the method LevenbergMarquardt::minimize() is 
-  * implemented as follow : 
+  * implemented as follow: 
   * \code
   * Status LevenbergMarquardt<FunctorType,Scalar>::minimize(FVectorType  &x, const int mode)
   * {

From 9a6a43319f31c03cda67c4ff772de339d0f19b8f Mon Sep 17 00:00:00 2001
From: Thales Sabino <thales@codeplay.com>
Date: Thu, 12 Jul 2018 11:20:59 +0100
Subject: [PATCH 437/680] Fix cxx11_tensor_fft not building on Windows.

The type used in Eigen::DSizes needs to be at least 8 bytes long. Internally Tensor tries to convert this to an __int64 on Windows and this fails to build. On Linux, long and long long are both 8 byte integer types.
* * *
Changing from "long long" to "std::int64_t".
---
 unsupported/test/cxx11_tensor_fft.cpp | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_fft.cpp b/unsupported/test/cxx11_tensor_fft.cpp
index a55369477..ef5f1a312 100644
--- a/unsupported/test/cxx11_tensor_fft.cpp
+++ b/unsupported/test/cxx11_tensor_fft.cpp
@@ -228,10 +228,13 @@ template <typename RealScalar>
 static void test_fft_non_power_of_2_round_trip(int exponent) {
   int n = (1 << exponent) + 1;
 
-  Eigen::DSizes<long, 1> dimensions;
+  // The dimension type needs to be at least 8 bytes long for the
+  // Tensor constructor to work. On Windows, long is only 4 bytes long,
+  // so use long long here to force the usage of a 8 bytes integer type.
+  Eigen::DSizes<std::int64_t, 1> dimensions;
   dimensions[0] = n;
-  const DSizes<long, 1> arr = dimensions;
-  Tensor<RealScalar, 1, ColMajor, long> input;
+  const DSizes<std::int64_t, 1> arr = dimensions;
+  Tensor<RealScalar, 1, ColMajor, std::int64_t> input;
 
   input.resize(arr);
   input.setRandom();
@@ -242,7 +245,7 @@ static void test_fft_non_power_of_2_round_trip(int exponent) {
   Tensor<std::complex<RealScalar>, 1, ColMajor> forward =
       input.template fft<BothParts, FFT_FORWARD>(fft);
 
-  Tensor<RealScalar, 1, ColMajor, long> output =
+  Tensor<RealScalar, 1, ColMajor, std::int64_t> output =
       forward.template fft<RealPart, FFT_REVERSE>(fft);
 
   for (int i = 0; i < n; ++i) {

From 006e18e52bfef7bac5db144dff241f685b383b39 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 16:57:41 +0200
Subject: [PATCH 438/680] Cleanup the mess in Eigen/Core by moving CUDA/HIP
 stuff at more appropriate places (Macros.h), and alignment/vectorization
 logic is now in util/ConfigureVectorization.h

---
 Eigen/Core                                   | 387 +----------------
 Eigen/src/Core/util/ConfigureVectorization.h | 415 ++++++++++++++++++
 Eigen/src/Core/util/Macros.h                 | 432 ++++++++++---------
 3 files changed, 647 insertions(+), 587 deletions(-)
 create mode 100644 Eigen/src/Core/util/ConfigureVectorization.h

diff --git a/Eigen/Core b/Eigen/Core
index 647a10831..fd6edc018 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -14,181 +14,26 @@
 // first thing Eigen does: stop the compiler from committing suicide
 #include "src/Core/util/DisableStupidWarnings.h"
 
-#if defined(__CUDACC__) && !defined(EIGEN_NO_CUDA)
-  #define EIGEN_CUDACC __CUDACC__
-#endif
+// then include this file where all our macros are defined. It's really important to do it first because
+// it's where we do all the compiler/OS/arch detections and define most defaults.
+#include "src/Core/util/Macros.h"
 
-#if defined(__CUDA_ARCH__) && !defined(EIGEN_NO_CUDA)
-  #define EIGEN_CUDA_ARCH __CUDA_ARCH__
-#endif
+// This detects SSE/AVX/NEON/etc. and configure alignment settings
+#include "src/Core/util/ConfigureVectorization.h"
 
-#if defined(__HIPCC__) && !defined(EIGEN_NO_HIP)
-  // analogous to EIGEN_CUDACC, but for HIP
-  #define EIGEN_HIPCC __HIPCC__
-#endif
-
-// NVCC is not supported as the target platform for HIPCC
-// Note that this also makes EIGEN_CUDACC and EIGEN_HIPCC mutually exclusive
-#if defined(__NVCC__) && defined(__HIPCC__)
-  #error "NVCC as the target platform for HIPCC is currently not supported."
-#endif
-
-// Starting with CUDA 9 the composite __CUDACC_VER__ is not available.
-#if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
-#define EIGEN_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
-#elif defined(__CUDACC_VER__)
-#define EIGEN_CUDACC_VER __CUDACC_VER__
-#else
-#define EIGEN_CUDACC_VER 0
-#endif
-
-// Handle NVCC/CUDA/SYCL
-#if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__) || defined(EIGEN_HIPCC)
-  // Do not try asserts on CUDA, HIP and SYCL!
-  #ifndef EIGEN_NO_DEBUG
-  #define EIGEN_NO_DEBUG
-  #endif
-
-  #ifdef EIGEN_INTERNAL_DEBUGGING
-  #undef EIGEN_INTERNAL_DEBUGGING
-  #endif
-
-  #ifdef EIGEN_EXCEPTIONS
-  #undef EIGEN_EXCEPTIONS
-  #endif
-#endif
-
-// All functions callable from CUDA code must be qualified with __device__
-#ifdef EIGEN_CUDACC
-  // Do not try to vectorize on CUDA and SYCL!
-  #ifndef EIGEN_DONT_VECTORIZE
-  #define EIGEN_DONT_VECTORIZE
-  #endif
-
-  #define EIGEN_DEVICE_FUNC __host__ __device__
-  // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
-  // works properly on the device side
+// We need cuda_runtime.h/hip_runtime.h to ensure that
+// the EIGEN_USING_STD_MATH macro works properly on the device side
+#if defined(EIGEN_CUDACC)
   #include <cuda_runtime.h>
-
-  #if EIGEN_HAS_CONSTEXPR
-    // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules
-    #if defined(__NVCC__)
-      // nvcc considers constexpr functions as __host__ __device__ with the option --expt-relaxed-constexpr
-      #ifdef __CUDACC_RELAXED_CONSTEXPR__
-        #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
-      #endif
-    #elif defined(__clang__) && defined(__CUDA__)
-      // clang++ always considers constexpr functions as implicitly __host__ __device__
-      #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
-    #endif
-  #endif
-
 #elif defined(EIGEN_HIPCC)
-  // Do not try to vectorize on HIP
-  #ifndef EIGEN_DONT_VECTORIZE
-  #define EIGEN_DONT_VECTORIZE
-  #endif
-
-  #define EIGEN_DEVICE_FUNC __host__ __device__
-  // We need hip_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
-  // works properly on the device side
   #include <hip/hip_runtime.h>
-  
-  #if defined(__HIP_DEVICE_COMPILE__) && !defined(EIGEN_NO_HIP)
-    // analogous to EIGEN_CUDA_ARCH, but for HIP
-    #define EIGEN_HIP_DEVICE_COMPILE __HIP_DEVICE_COMPILE__
-    // Note this check needs to come after we include hip_runtime.h since
-    // hip_runtime.h includes hip_common.h which in turn has the define
-    // for __HIP_DEVICE_COMPILE__
-  #endif
-
-#else
-  #define EIGEN_DEVICE_FUNC
 #endif
 
-#ifdef __NVCC__
-  #ifndef EIGEN_DONT_VECTORIZE
-  #define EIGEN_DONT_VECTORIZE
-  #endif
-#endif
-
-
-#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
-//
-// If either EIGEN_CUDACC or EIGEN_HIPCC is defined, then define EIGEN_GPUCC
-//
-#define EIGEN_GPUCC
-//
-// EIGEN_HIPCC implies the HIP compiler and is used to tweak Eigen code for use in HIP kernels
-// EIGEN_CUDACC implies the CUDA compiler and is used to tweak Eigen code for use in CUDA kernels
-//
-// In most cases the same tweaks are required to the Eigen code to enable in both the HIP and CUDA kernels.
-// For those cases, the corresponding code should be guarded with
-//      #if defined(EIGEN_GPUCC)
-// instead of
-//      #if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
-//
-// For cases where the tweak is specific to HIP, the code should be guarded with
-//      #if defined(EIGEN_HIPCC)
-//
-// For cases where the tweak is specific to CUDA, the code should be guarded with
-//      #if defined(EIGEN_CUDACC)
-//
-#endif
-
-#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
-//
-// If either EIGEN_CUDA_ARCH or EIGEN_HIP_DEVICE_COMPILE is defined, then define EIGEN_GPU_COMPILE_PHASE
-//
-#define EIGEN_GPU_COMPILE_PHASE
-//
-// GPU compilers (HIPCC, NVCC) typically do two passes over the source code,
-//   + one to compile the source for the "host" (ie CPU)
-//   + another to compile the source for the "device" (ie. GPU)
-// 
-// Code that needs to enabled only during the either the "host" or "device" compilation phase
-// needs to be guarded with a macro that indicates the current compilation phase
-//
-// EIGEN_HIP_DEVICE_COMPILE implies the device compilation phase in HIP
-// EIGEN_CUDA_ARCH implies the device compilation phase in CUDA
-//
-// In most cases, the "host" / "device" specific code is the same for both HIP and CUDA
-// For those cases, the code should be guarded with
-//       #if defined(EIGEN_GPU_COMPILE_PHASE)
-// instead of
-//       #if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
-//
-// For cases where the tweak is specific to HIP, the code should be guarded with
-//      #if defined(EIGEN_HIP_DEVICE_COMPILE)
-//
-// For cases where the tweak is specific to CUDA, the code should be guarded with
-//      #if defined(EIGEN_CUDA_ARCH)
-//
-#endif
-
-
-// When compiling CUDA device code with NVCC, or HIP device code with HIPCC
-// pull in math functions from the global namespace.  In host mode, and when
-// device doee with clang, use the std versions.
-#if (defined(EIGEN_CUDA_ARCH) && defined(__NVCC__)) || (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIPCC__))
-  #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
-#else
-  #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
-#endif
-
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL) && !defined(EIGEN_HIP_DEVICE_COMPILE)
-  #define EIGEN_EXCEPTIONS
-#endif
 
 #ifdef EIGEN_EXCEPTIONS
   #include <new>
 #endif
 
-// then include this file where all our macros are defined. It's really important to do it first because
-// it's where we do all the alignment settings (platform detection and honoring the user's will if he
-// defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization.
-#include "src/Core/util/Macros.h"
-
 // Disable the ipa-cp-clone optimization flag with MinGW 6.x or newer (enabled by default with -O3)
 // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=556 for details.
 #if EIGEN_COMP_MINGW && EIGEN_GNUC_AT_LEAST(4,6)
@@ -201,190 +46,6 @@
 // and inclusion of their respective header files
 #include "src/Core/util/MKL_support.h"
 
-// if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
-// account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
-#if EIGEN_MAX_ALIGN_BYTES==0
-  #ifndef EIGEN_DONT_VECTORIZE
-    #define EIGEN_DONT_VECTORIZE
-  #endif
-#endif
-
-#if EIGEN_COMP_MSVC
-  #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
-  #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
-    // Remember that usage of defined() in a #define is undefined by the standard.
-    // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
-    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
-      #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
-    #endif
-  #endif
-#else
-  // Remember that usage of defined() in a #define is undefined by the standard
-  #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
-    #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
-  #endif
-#endif
-
-#ifndef EIGEN_DONT_VECTORIZE
-
-  #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER)
-
-    // Defines symbols for compile-time detection of which instructions are
-    // used.
-    // EIGEN_VECTORIZE_YY is defined if and only if the instruction set YY is used
-    #define EIGEN_VECTORIZE
-    #define EIGEN_VECTORIZE_SSE
-    #define EIGEN_VECTORIZE_SSE2
-
-    // Detect sse3/ssse3/sse4:
-    // gcc and icc defines __SSE3__, ...
-    // there is no way to know about this on msvc. You can define EIGEN_VECTORIZE_SSE* if you
-    // want to force the use of those instructions with msvc.
-    #ifdef __SSE3__
-      #define EIGEN_VECTORIZE_SSE3
-    #endif
-    #ifdef __SSSE3__
-      #define EIGEN_VECTORIZE_SSSE3
-    #endif
-    #ifdef __SSE4_1__
-      #define EIGEN_VECTORIZE_SSE4_1
-    #endif
-    #ifdef __SSE4_2__
-      #define EIGEN_VECTORIZE_SSE4_2
-    #endif
-    #ifdef __AVX__
-      #define EIGEN_VECTORIZE_AVX
-      #define EIGEN_VECTORIZE_SSE3
-      #define EIGEN_VECTORIZE_SSSE3
-      #define EIGEN_VECTORIZE_SSE4_1
-      #define EIGEN_VECTORIZE_SSE4_2
-    #endif
-    #ifdef __AVX2__
-      #define EIGEN_VECTORIZE_AVX2
-      #define EIGEN_VECTORIZE_AVX
-      #define EIGEN_VECTORIZE_SSE3
-      #define EIGEN_VECTORIZE_SSSE3
-      #define EIGEN_VECTORIZE_SSE4_1
-      #define EIGEN_VECTORIZE_SSE4_2
-    #endif
-    #ifdef __FMA__
-      #define EIGEN_VECTORIZE_FMA
-    #endif
-    #if defined(__AVX512F__)
-      #define EIGEN_VECTORIZE_AVX512
-      #define EIGEN_VECTORIZE_AVX2
-      #define EIGEN_VECTORIZE_AVX
-      #define EIGEN_VECTORIZE_FMA
-      #define EIGEN_VECTORIZE_SSE3
-      #define EIGEN_VECTORIZE_SSSE3
-      #define EIGEN_VECTORIZE_SSE4_1
-      #define EIGEN_VECTORIZE_SSE4_2
-      #ifdef __AVX512DQ__
-        #define EIGEN_VECTORIZE_AVX512DQ
-      #endif
-      #ifdef __AVX512ER__
-        #define EIGEN_VECTORIZE_AVX512ER
-      #endif
-    #endif
-
-    // include files
-
-    // This extern "C" works around a MINGW-w64 compilation issue
-    // https://sourceforge.net/tracker/index.php?func=detail&aid=3018394&group_id=202880&atid=983354
-    // In essence, intrin.h is included by windows.h and also declares intrinsics (just as emmintrin.h etc. below do).
-    // However, intrin.h uses an extern "C" declaration, and g++ thus complains of duplicate declarations
-    // with conflicting linkage.  The linkage for intrinsics doesn't matter, but at that stage the compiler doesn't know;
-    // so, to avoid compile errors when windows.h is included after Eigen/Core, ensure intrinsics are extern "C" here too.
-    // notice that since these are C headers, the extern "C" is theoretically needed anyways.
-    extern "C" {
-      // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
-      // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
-      #if EIGEN_COMP_ICC >= 1110
-        #include <immintrin.h>
-      #else
-        #include <mmintrin.h>
-        #include <emmintrin.h>
-        #include <xmmintrin.h>
-        #ifdef  EIGEN_VECTORIZE_SSE3
-        #include <pmmintrin.h>
-        #endif
-        #ifdef EIGEN_VECTORIZE_SSSE3
-        #include <tmmintrin.h>
-        #endif
-        #ifdef EIGEN_VECTORIZE_SSE4_1
-        #include <smmintrin.h>
-        #endif
-        #ifdef EIGEN_VECTORIZE_SSE4_2
-        #include <nmmintrin.h>
-        #endif
-        #if defined(EIGEN_VECTORIZE_AVX) || defined(EIGEN_VECTORIZE_AVX512)
-        #include <immintrin.h>
-        #endif
-      #endif
-    } // end extern "C"
-  #elif defined __VSX__
-    #define EIGEN_VECTORIZE
-    #define EIGEN_VECTORIZE_VSX
-    #include <altivec.h>
-    // We need to #undef all these ugly tokens defined in <altivec.h>
-    // => use __vector instead of vector
-    #undef bool
-    #undef vector
-    #undef pixel
-  #elif defined __ALTIVEC__
-    #define EIGEN_VECTORIZE
-    #define EIGEN_VECTORIZE_ALTIVEC
-    #include <altivec.h>
-    // We need to #undef all these ugly tokens defined in <altivec.h>
-    // => use __vector instead of vector
-    #undef bool
-    #undef vector
-    #undef pixel
-  #elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
-    #define EIGEN_VECTORIZE
-    #define EIGEN_VECTORIZE_NEON
-    #include <arm_neon.h>
-  #elif (defined __s390x__ && defined __VEC__)
-    #define EIGEN_VECTORIZE
-    #define EIGEN_VECTORIZE_ZVECTOR
-    #include <vecintrin.h>
-  #endif
-#endif
-
-#if defined(__F16C__) && !defined(EIGEN_COMP_CLANG)
-  // We can use the optimized fp16 to float and float to fp16 conversion routines
-  #define EIGEN_HAS_FP16_C
-#endif
-
-#if defined EIGEN_CUDACC
-  #define EIGEN_VECTORIZE_GPU
-  #include <vector_types.h>
-  #if EIGEN_CUDACC_VER >= 70500
-    #define EIGEN_HAS_CUDA_FP16
-  #endif
-#endif
-
-#if defined EIGEN_HAS_CUDA_FP16
-  #include <host_defines.h>
-  #include <cuda_fp16.h>
-#endif
-
-#if defined(EIGEN_HIPCC) && defined(EIGEN_HIP_DEVICE_COMPILE)
-
-  #define EIGEN_VECTORIZE_GPU
-  #include <hip/hip_vector_types.h>
-
-  #define EIGEN_HAS_HIP_FP16
-  #include <hip/hip_fp16.h>
-
-  #define HIP_PATCH_WITH_NEW_FP16 18215
-  #if (HIP_VERSION_PATCH < HIP_PATCH_WITH_NEW_FP16)
-    #define EIGEN_HAS_OLD_HIP_FP16
-    // Old HIP implementation does not have a explicit typedef for "half2"
-    typedef __half2 half2;
-  #endif
-  
-#endif
 
 #if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
   #define EIGEN_HAS_GPU_FP16
@@ -443,38 +104,6 @@
   #include <SYCL/sycl.hpp>
 #endif
 
-/** \brief Namespace containing all symbols from the %Eigen library. */
-namespace Eigen {
-
-inline static const char *SimdInstructionSetsInUse(void) {
-#if defined(EIGEN_VECTORIZE_AVX512)
-  return "AVX512, FMA, AVX2, AVX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
-#elif defined(EIGEN_VECTORIZE_AVX)
-  return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
-#elif defined(EIGEN_VECTORIZE_SSE4_2)
-  return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
-#elif defined(EIGEN_VECTORIZE_SSE4_1)
-  return "SSE, SSE2, SSE3, SSSE3, SSE4.1";
-#elif defined(EIGEN_VECTORIZE_SSSE3)
-  return "SSE, SSE2, SSE3, SSSE3";
-#elif defined(EIGEN_VECTORIZE_SSE3)
-  return "SSE, SSE2, SSE3";
-#elif defined(EIGEN_VECTORIZE_SSE2)
-  return "SSE, SSE2";
-#elif defined(EIGEN_VECTORIZE_ALTIVEC)
-  return "AltiVec";
-#elif defined(EIGEN_VECTORIZE_VSX)
-  return "VSX";
-#elif defined(EIGEN_VECTORIZE_NEON)
-  return "ARM NEON";
-#elif defined(EIGEN_VECTORIZE_ZVECTOR)
-  return "S390X ZVECTOR";
-#else
-  return "None";
-#endif
-}
-
-} // end namespace Eigen
 
 #if defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS || defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API || defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS || defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API || defined EIGEN2_SUPPORT
 // This will generate an error message:
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
new file mode 100644
index 000000000..27f65f672
--- /dev/null
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -0,0 +1,415 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2018 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONFIGURE_VECTORIZATION_H
+#define EIGEN_CONFIGURE_VECTORIZATION_H
+
+// FIXME: not sure why this is needed, perhaps it is not needed anymore.
+#ifdef __NVCC__
+  #ifndef EIGEN_DONT_VECTORIZE
+  #define EIGEN_DONT_VECTORIZE
+  #endif
+#endif
+
+//------------------------------------------------------------------------------------------
+// Static and dynamic alignment control
+//
+// The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES
+// as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively.
+// The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not,
+// a default value is automatically computed based on architecture, compiler, and OS.
+//
+// This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX}
+// to be used to declare statically aligned buffers.
+//------------------------------------------------------------------------------------------
+
+
+/* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
+ * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
+ * so that vectorization doesn't affect binary compatibility.
+ *
+ * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
+ * vectorized and non-vectorized code.
+ */
+#if (defined EIGEN_CUDACC)
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
+#elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#elif EIGEN_COMP_MSVC
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
+#elif EIGEN_COMP_SUNCC
+  // FIXME not sure about this one:
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#else
+  #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
+#endif
+
+// If the user explicitly disable vectorization, then we also disable alignment
+#if defined(EIGEN_DONT_VECTORIZE)
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
+#elif defined(__AVX512F__)
+  // 64 bytes static alignment is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
+#elif defined(__AVX__)
+  // 32 bytes static alignment is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
+#else
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
+#endif
+
+
+// EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense
+#define EIGEN_MIN_ALIGN_BYTES 16
+
+// Defined the boundary (in bytes) on which the data needs to be aligned. Note
+// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
+// aligned at all regardless of the value of this #define.
+
+#if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN))  && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
+#endif
+
+// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprecated
+// They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
+#if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
+  #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
+    #undef EIGEN_MAX_STATIC_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+#endif
+
+#ifndef EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
+  // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
+  // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
+  // certain common platform (compiler+architecture combinations) to avoid these problems.
+  // Only static alignment is really problematic (relies on nonstandard compiler extensions),
+  // try to keep heap alignment even when we have to disable static alignment.
+  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
+  // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
+  // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
+  // 4.8 and newer seem definitely unaffected.
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #else
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
+  #endif
+
+  // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
+  #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
+  && !EIGEN_GCC3_OR_OLDER \
+  && !EIGEN_COMP_SUNCC \
+  && !EIGEN_OS_QNX
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
+  #else
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
+  #endif
+
+  #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+  #else
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+  #endif
+
+#endif
+
+// If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_ALIGN_BYTES
+#if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES
+#undef EIGEN_MAX_STATIC_ALIGN_BYTES
+#define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
+  #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
+#endif
+
+// At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
+// It takes into account both the user choice to explicitly enable/disable alignment (by setting EIGEN_MAX_STATIC_ALIGN_BYTES)
+// and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
+// Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
+
+
+// Shortcuts to EIGEN_ALIGN_TO_BOUNDARY
+#define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
+#define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
+#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32)
+#define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64)
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES)
+#else
+#define EIGEN_ALIGN_MAX
+#endif
+
+
+// Dynamic alignment control
+
+#if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0
+#error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN.
+#endif
+
+#ifdef EIGEN_DONT_ALIGN
+  #ifdef EIGEN_MAX_ALIGN_BYTES
+    #undef EIGEN_MAX_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_ALIGN_BYTES 0
+#elif !defined(EIGEN_MAX_ALIGN_BYTES)
+  #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+#else
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
+#endif
+
+
+#ifndef EIGEN_UNALIGNED_VECTORIZE
+#define EIGEN_UNALIGNED_VECTORIZE 1
+#endif
+
+//----------------------------------------------------------------------
+
+
+
+// if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
+// account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
+#if EIGEN_MAX_ALIGN_BYTES==0
+  #ifndef EIGEN_DONT_VECTORIZE
+    #define EIGEN_DONT_VECTORIZE
+  #endif
+#endif
+
+
+// The following (except #include <malloc.h> and _M_IX86_FP ??) can likely be
+// removed as gcc 4.1 and msvc 2008 are not supported anyways.
+#if EIGEN_COMP_MSVC
+  #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
+  #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
+    // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
+    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
+      #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
+    #endif
+  #endif
+#else
+  #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
+    #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
+  #endif
+#endif
+
+
+#ifndef EIGEN_DONT_VECTORIZE
+
+  #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER)
+
+    // Defines symbols for compile-time detection of which instructions are
+    // used.
+    // EIGEN_VECTORIZE_YY is defined if and only if the instruction set YY is used
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_SSE
+    #define EIGEN_VECTORIZE_SSE2
+
+    // Detect sse3/ssse3/sse4:
+    // gcc and icc defines __SSE3__, ...
+    // there is no way to know about this on msvc. You can define EIGEN_VECTORIZE_SSE* if you
+    // want to force the use of those instructions with msvc.
+    #ifdef __SSE3__
+      #define EIGEN_VECTORIZE_SSE3
+    #endif
+    #ifdef __SSSE3__
+      #define EIGEN_VECTORIZE_SSSE3
+    #endif
+    #ifdef __SSE4_1__
+      #define EIGEN_VECTORIZE_SSE4_1
+    #endif
+    #ifdef __SSE4_2__
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __AVX__
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_SSE3
+      #define EIGEN_VECTORIZE_SSSE3
+      #define EIGEN_VECTORIZE_SSE4_1
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __AVX2__
+      #define EIGEN_VECTORIZE_AVX2
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_SSE3
+      #define EIGEN_VECTORIZE_SSSE3
+      #define EIGEN_VECTORIZE_SSE4_1
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __FMA__
+      #define EIGEN_VECTORIZE_FMA
+    #endif
+    #if defined(__AVX512F__)
+      #define EIGEN_VECTORIZE_AVX512
+      #define EIGEN_VECTORIZE_AVX2
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_FMA
+      #define EIGEN_VECTORIZE_SSE3
+      #define EIGEN_VECTORIZE_SSSE3
+      #define EIGEN_VECTORIZE_SSE4_1
+      #define EIGEN_VECTORIZE_SSE4_2
+      #ifdef __AVX512DQ__
+        #define EIGEN_VECTORIZE_AVX512DQ
+      #endif
+      #ifdef __AVX512ER__
+        #define EIGEN_VECTORIZE_AVX512ER
+      #endif
+    #endif
+
+    // include files
+
+    // This extern "C" works around a MINGW-w64 compilation issue
+    // https://sourceforge.net/tracker/index.php?func=detail&aid=3018394&group_id=202880&atid=983354
+    // In essence, intrin.h is included by windows.h and also declares intrinsics (just as emmintrin.h etc. below do).
+    // However, intrin.h uses an extern "C" declaration, and g++ thus complains of duplicate declarations
+    // with conflicting linkage.  The linkage for intrinsics doesn't matter, but at that stage the compiler doesn't know;
+    // so, to avoid compile errors when windows.h is included after Eigen/Core, ensure intrinsics are extern "C" here too.
+    // notice that since these are C headers, the extern "C" is theoretically needed anyways.
+    extern "C" {
+      // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
+      // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
+      #if EIGEN_COMP_ICC >= 1110
+        #include <immintrin.h>
+      #else
+        #include <mmintrin.h>
+        #include <emmintrin.h>
+        #include <xmmintrin.h>
+        #ifdef  EIGEN_VECTORIZE_SSE3
+        #include <pmmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSSE3
+        #include <tmmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSE4_1
+        #include <smmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSE4_2
+        #include <nmmintrin.h>
+        #endif
+        #if defined(EIGEN_VECTORIZE_AVX) || defined(EIGEN_VECTORIZE_AVX512)
+        #include <immintrin.h>
+        #endif
+      #endif
+    } // end extern "C"
+
+  #elif defined __VSX__
+
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_VSX
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
+
+  #elif defined __ALTIVEC__
+
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_ALTIVEC
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
+
+  #elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
+
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_NEON
+    #include <arm_neon.h>
+
+  #elif (defined __s390x__ && defined __VEC__)
+
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_ZVECTOR
+    #include <vecintrin.h>
+
+  #endif
+#endif
+
+#if defined(__F16C__) && !defined(EIGEN_COMP_CLANG)
+  // We can use the optimized fp16 to float and float to fp16 conversion routines
+  #define EIGEN_HAS_FP16_C
+#endif
+
+#if defined EIGEN_CUDACC
+  #define EIGEN_VECTORIZE_GPU
+  #include <vector_types.h>
+  #if EIGEN_CUDACC_VER >= 70500
+    #define EIGEN_HAS_CUDA_FP16
+  #endif
+#endif
+
+#if defined(EIGEN_HAS_CUDA_FP16)
+  #include <host_defines.h>
+  #include <cuda_fp16.h>
+#endif
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+
+  #define EIGEN_VECTORIZE_GPU
+  #include <hip/hip_vector_types.h>
+
+  #define EIGEN_HAS_HIP_FP16
+  #include <hip/hip_fp16.h>
+
+  #define HIP_PATCH_WITH_NEW_FP16 18215
+  #if (HIP_VERSION_PATCH < HIP_PATCH_WITH_NEW_FP16)
+    #define EIGEN_HAS_OLD_HIP_FP16
+    // Old HIP implementation does not have a explicit typedef for "half2"
+    typedef __half2 half2;
+  #endif
+
+#endif
+
+
+/** \brief Namespace containing all symbols from the %Eigen library. */
+namespace Eigen {
+
+inline static const char *SimdInstructionSetsInUse(void) {
+#if defined(EIGEN_VECTORIZE_AVX512)
+  return "AVX512, FMA, AVX2, AVX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_AVX)
+  return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_SSE4_2)
+  return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_SSE4_1)
+  return "SSE, SSE2, SSE3, SSSE3, SSE4.1";
+#elif defined(EIGEN_VECTORIZE_SSSE3)
+  return "SSE, SSE2, SSE3, SSSE3";
+#elif defined(EIGEN_VECTORIZE_SSE3)
+  return "SSE, SSE2, SSE3";
+#elif defined(EIGEN_VECTORIZE_SSE2)
+  return "SSE, SSE2";
+#elif defined(EIGEN_VECTORIZE_ALTIVEC)
+  return "AltiVec";
+#elif defined(EIGEN_VECTORIZE_VSX)
+  return "VSX";
+#elif defined(EIGEN_VECTORIZE_NEON)
+  return "ARM NEON";
+#elif defined(EIGEN_VECTORIZE_ZVECTOR)
+  return "S390X ZVECTOR";
+#else
+  return "None";
+#endif
+}
+
+} // end namespace Eigen
+
+
+#endif // EIGEN_CONFIGURE_VECTORIZATION_H
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 64b7be423..46ca0193a 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -11,6 +11,10 @@
 #ifndef EIGEN_MACROS_H
 #define EIGEN_MACROS_H
 
+//------------------------------------------------------------------------------------------
+// Eigen version and basic defaults
+//------------------------------------------------------------------------------------------
+
 #define EIGEN_WORLD_VERSION 3
 #define EIGEN_MAJOR_VERSION 3
 #define EIGEN_MINOR_VERSION 90
@@ -19,7 +23,40 @@
                                       (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
                                                                  EIGEN_MINOR_VERSION>=z))))
 
+#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor
+#else
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor
+#endif
+
+#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
+#endif
+
+// Upperbound on the C++ version to use.
+// Expected values are 03, 11, 14, 17, etc.
+// By default, let's use an arbitrarily large C++ version.
+#ifndef EIGEN_MAX_CPP_VER
+#define EIGEN_MAX_CPP_VER 99
+#endif
+
+/** Allows to disable some optimizations which might affect the accuracy of the result.
+  * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
+  * They currently include:
+  *   - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization.
+  */
+#ifndef EIGEN_FAST_MATH
+#define EIGEN_FAST_MATH 1
+#endif
+
+#ifndef EIGEN_STACK_ALLOCATION_LIMIT
+// 131072 == 128 KB
+#define EIGEN_STACK_ALLOCATION_LIMIT 131072
+#endif
+
+//------------------------------------------------------------------------------------------
 // Compiler identification, EIGEN_COMP_*
+//------------------------------------------------------------------------------------------
 
 /// \internal EIGEN_COMP_GNUC set to 1 for all compilers compatible with GCC
 #ifdef __GNUC__
@@ -147,7 +184,11 @@
 #endif
 
 
+
+//------------------------------------------------------------------------------------------
 // Architecture identification, EIGEN_ARCH_*
+//------------------------------------------------------------------------------------------
+
 
 #if defined(__x86_64__) || defined(_M_X64) || defined(__amd64)
   #define EIGEN_ARCH_x86_64 1
@@ -217,7 +258,9 @@
 
 
 
+//------------------------------------------------------------------------------------------
 // Operating system identification, EIGEN_OS_*
+//------------------------------------------------------------------------------------------
 
 /// \internal EIGEN_OS_UNIX set to 1 if the OS is a unix variant
 #if defined(__unix__) || defined(__unix)
@@ -319,6 +362,106 @@
 #endif
 
 
+//------------------------------------------------------------------------------------------
+// Detect GPU compilers and architectures
+//------------------------------------------------------------------------------------------
+
+// NVCC is not supported as the target platform for HIPCC
+// Note that this also makes EIGEN_CUDACC and EIGEN_HIPCC mutually exclusive
+#if defined(__NVCC__) && defined(__HIPCC__)
+  #error "NVCC as the target platform for HIPCC is currently not supported."
+#endif
+
+#if defined(__CUDACC__) && !defined(EIGEN_NO_CUDA)
+  // Means the compiler is either nvcc or clang with CUDA enabled
+  #define EIGEN_CUDACC __CUDACC__
+#endif
+
+#if defined(__CUDA_ARCH__) && !defined(EIGEN_NO_CUDA)
+  // Means we are generating code for the device
+  #define EIGEN_CUDA_ARCH __CUDA_ARCH__
+#endif
+
+// Starting with CUDA 9 the composite __CUDACC_VER__ is not available.
+#if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9)
+  #define EIGEN_CUDACC_VER  ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100))
+#elif defined(__CUDACC_VER__)
+  #define EIGEN_CUDACC_VER __CUDACC_VER__
+#else
+  #define EIGEN_CUDACC_VER 0
+#endif
+
+#if defined(__HIPCC__) && !defined(EIGEN_NO_HIP)
+  // Means the compiler is HIPCC (analogous to EIGEN_CUDACC, but for HIP)
+  #define EIGEN_HIPCC __HIPCC__
+
+  // We need hip_common.h here because __HIP_DEVICE_COMPILE__ is defined in this header.
+  #include <hip/hip_common.h>
+
+  #if defined(__HIP_DEVICE_COMPILE__)
+    // analogous to EIGEN_CUDA_ARCH, but for HIP
+    #define EIGEN_HIP_DEVICE_COMPILE __HIP_DEVICE_COMPILE__
+  #endif
+#endif
+
+// Unify CUDA/HIPCC
+
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// If either EIGEN_CUDACC or EIGEN_HIPCC is defined, then define EIGEN_GPUCC
+//
+#define EIGEN_GPUCC
+//
+// EIGEN_HIPCC implies the HIP compiler and is used to tweak Eigen code for use in HIP kernels
+// EIGEN_CUDACC implies the CUDA compiler and is used to tweak Eigen code for use in CUDA kernels
+//
+// In most cases the same tweaks are required to the Eigen code to enable in both the HIP and CUDA kernels.
+// For those cases, the corresponding code should be guarded with
+//      #if defined(EIGEN_GPUCC)
+// instead of
+//      #if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDACC)
+//
+#endif
+
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// If either EIGEN_CUDA_ARCH or EIGEN_HIP_DEVICE_COMPILE is defined, then define EIGEN_GPU_COMPILE_PHASE
+//
+#define EIGEN_GPU_COMPILE_PHASE
+//
+// GPU compilers (HIPCC, NVCC) typically do two passes over the source code,
+//   + one to compile the source for the "host" (ie CPU)
+//   + another to compile the source for the "device" (ie. GPU)
+//
+// Code that needs to enabled only during the either the "host" or "device" compilation phase
+// needs to be guarded with a macro that indicates the current compilation phase
+//
+// EIGEN_HIP_DEVICE_COMPILE implies the device compilation phase in HIP
+// EIGEN_CUDA_ARCH implies the device compilation phase in CUDA
+//
+// In most cases, the "host" / "device" specific code is the same for both HIP and CUDA
+// For those cases, the code should be guarded with
+//       #if defined(EIGEN_GPU_COMPILE_PHASE)
+// instead of
+//       #if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDA_ARCH)
+//
+#endif
+
+//------------------------------------------------------------------------------------------
+// Detect Compiler/Architecture/OS specific features
+//------------------------------------------------------------------------------------------
 
 #if EIGEN_GNUC_AT_MOST(4,3) && !EIGEN_COMP_CLANG
   // see bug 89
@@ -327,20 +470,6 @@
   #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1
 #endif
 
-// This macro can be used to prevent from macro expansion, e.g.:
-//   std::max EIGEN_NOT_A_MACRO(a,b)
-#define EIGEN_NOT_A_MACRO
-
-#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
-#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor
-#else
-#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor
-#endif
-
-#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
-#endif
-
 // Cross compiler wrapper around LLVM's __has_builtin
 #ifdef __has_builtin
 #  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
@@ -362,13 +491,6 @@
 #define EIGEN_HAS_STATIC_ARRAY_TEMPLATE 0
 #endif
 
-// Upperbound on the C++ version to use.
-// Expected values are 03, 11, 14, 17, etc.
-// By default, let's use an arbitrarily large C++ version.
-#ifndef EIGEN_MAX_CPP_VER
-#define EIGEN_MAX_CPP_VER 99
-#endif
-
 #if EIGEN_MAX_CPP_VER>=11 && (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900)
 #define EIGEN_HAS_CXX11 1
 #else
@@ -442,22 +564,22 @@
 // Does the compiler fully support const expressions? (as in c++14)
 #ifndef EIGEN_HAS_CONSTEXPR
 
-#if defined(EIGEN_CUDACC)
-// Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
-#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500))
+  #if defined(EIGEN_CUDACC)
+  // Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
+  #if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500))
+    #define EIGEN_HAS_CONSTEXPR 1
+  #endif
+  #elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
+    (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L)) || \
+    (EIGEN_COMP_CLANG >= 306 && (__cplusplus > 199711L)))
   #define EIGEN_HAS_CONSTEXPR 1
-#endif
-#elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
-  (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L)) || \
-  (EIGEN_COMP_CLANG >= 306 && (__cplusplus > 199711L)))
-#define EIGEN_HAS_CONSTEXPR 1
-#endif
+  #endif
 
-#ifndef EIGEN_HAS_CONSTEXPR
-#define EIGEN_HAS_CONSTEXPR 0
-#endif
+  #ifndef EIGEN_HAS_CONSTEXPR
+  #define EIGEN_HAS_CONSTEXPR 0
+  #endif
 
-#endif
+#endif // EIGEN_HAS_CONSTEXPR
 
 // Does the compiler support C++11 math?
 // Let's be conservative and enable the default C++11 implementation only if we are sure it exists
@@ -495,15 +617,29 @@
   #endif
 #endif
 
-/** Allows to disable some optimizations which might affect the accuracy of the result.
-  * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
-  * They currently include:
-  *   - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization.
-  */
-#ifndef EIGEN_FAST_MATH
-#define EIGEN_FAST_MATH 1
+
+#if defined(EIGEN_CUDACC) && EIGEN_HAS_CONSTEXPR
+  // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules
+  #if defined(__NVCC__)
+    // nvcc considers constexpr functions as __host__ __device__ with the option --expt-relaxed-constexpr
+    #ifdef __CUDACC_RELAXED_CONSTEXPR__
+      #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
+    #endif
+  #elif defined(__clang__) && defined(__CUDA__)
+    // clang++ always considers constexpr functions as implicitly __host__ __device__
+    #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
+  #endif
 #endif
 
+
+//------------------------------------------------------------------------------------------
+// Preprocessor programming helpers
+//------------------------------------------------------------------------------------------
+
+// This macro can be used to prevent from macro expansion, e.g.:
+//   std::max EIGEN_NOT_A_MACRO(a,b)
+#define EIGEN_NOT_A_MACRO
+
 #define EIGEN_DEBUG_VAR(x) std::cerr << #x << " = " << x << std::endl;
 
 // concatenate two tokens
@@ -555,6 +691,36 @@
 #define EIGEN_PERMISSIVE_EXPR
 #endif
 
+// GPU stuff
+
+// Disable some features when compiling with GPU compilers (NVCC/clang-cuda/SYCL/HIPCC)
+#if defined(EIGEN_CUDACC) || defined(__SYCL_DEVICE_ONLY__) || defined(EIGEN_HIPCC)
+  // Do not try asserts on device code
+  #ifndef EIGEN_NO_DEBUG
+  #define EIGEN_NO_DEBUG
+  #endif
+
+  #ifdef EIGEN_INTERNAL_DEBUGGING
+  #undef EIGEN_INTERNAL_DEBUGGING
+  #endif
+
+  #ifdef EIGEN_EXCEPTIONS
+  #undef EIGEN_EXCEPTIONS
+  #endif
+#endif
+
+// All functions callable from CUDA/HIP code must be qualified with __device__
+#ifdef EIGEN_GPUCC
+  #ifndef EIGEN_DONT_VECTORIZE
+  #define EIGEN_DONT_VECTORIZE
+  #endif
+
+  #define EIGEN_DEVICE_FUNC __host__ __device__
+#else
+  #define EIGEN_DEVICE_FUNC
+#endif
+
+
 // this macro allows to get rid of linking errors about multiply defined functions.
 //  - static is not very good because it prevents definitions from different object files to be merged.
 //           So static causes the resulting linked executable to be bloated with multiple copies of the same function.
@@ -666,169 +832,6 @@ namespace Eigen {
 #  define EIGEN_CONST_CONDITIONAL(cond)  cond
 #endif
 
-//------------------------------------------------------------------------------------------
-// Static and dynamic alignment control
-//
-// The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES
-// as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively.
-// The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not,
-// a default value is automatically computed based on architecture, compiler, and OS.
-//
-// This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX}
-// to be used to declare statically aligned buffers.
-//------------------------------------------------------------------------------------------
-
-
-/* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
- * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
- * so that vectorization doesn't affect binary compatibility.
- *
- * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
- * vectorized and non-vectorized code.
- */
-#if (defined EIGEN_CUDACC)
-  #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
-#elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
-  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
-#elif EIGEN_COMP_MSVC
-  #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
-#elif EIGEN_COMP_SUNCC
-  // FIXME not sure about this one:
-  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
-#else
-  #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
-#endif
-
-// If the user explicitly disable vectorization, then we also disable alignment
-#if defined(EIGEN_DONT_VECTORIZE)
-  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
-#elif defined(__AVX512F__)
-  // 64 bytes static alignment is preferred only if really required
-  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
-#elif defined(__AVX__)
-  // 32 bytes static alignment is preferred only if really required
-  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
-#else
-  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
-#endif
-
-
-// EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense
-#define EIGEN_MIN_ALIGN_BYTES 16
-
-// Defined the boundary (in bytes) on which the data needs to be aligned. Note
-// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
-// aligned at all regardless of the value of this #define.
-
-#if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN))  && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0
-#error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
-#endif
-
-// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprecated
-// They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
-#if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
-  #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
-    #undef EIGEN_MAX_STATIC_ALIGN_BYTES
-  #endif
-  #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
-#endif
-
-#ifndef EIGEN_MAX_STATIC_ALIGN_BYTES
-
-  // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES
-
-  // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
-  // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
-  // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
-  // certain common platform (compiler+architecture combinations) to avoid these problems.
-  // Only static alignment is really problematic (relies on nonstandard compiler extensions),
-  // try to keep heap alignment even when we have to disable static alignment.
-  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
-  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
-  #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
-  // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
-  // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
-  // 4.8 and newer seem definitely unaffected.
-  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
-  #else
-  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
-  #endif
-
-  // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
-  #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
-  && !EIGEN_GCC3_OR_OLDER \
-  && !EIGEN_COMP_SUNCC \
-  && !EIGEN_OS_QNX
-    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
-  #else
-    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
-  #endif
-
-  #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT
-    #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
-  #else
-    #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
-  #endif
-
-#endif
-
-// If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_ALIGN_BYTES
-#if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES
-#undef EIGEN_MAX_STATIC_ALIGN_BYTES
-#define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
-#endif
-
-#if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
-  #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
-#endif
-
-// At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
-// It takes into account both the user choice to explicitly enable/disable alignment (by setting EIGEN_MAX_STATIC_ALIGN_BYTES)
-// and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
-// Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
-
-
-// Shortcuts to EIGEN_ALIGN_TO_BOUNDARY
-#define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
-#define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
-#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32)
-#define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64)
-#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
-#define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES)
-#else
-#define EIGEN_ALIGN_MAX
-#endif
-
-
-// Dynamic alignment control
-
-#if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0
-#error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN.
-#endif
-
-#ifdef EIGEN_DONT_ALIGN
-  #ifdef EIGEN_MAX_ALIGN_BYTES
-    #undef EIGEN_MAX_ALIGN_BYTES
-  #endif
-  #define EIGEN_MAX_ALIGN_BYTES 0
-#elif !defined(EIGEN_MAX_ALIGN_BYTES)
-  #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
-#endif
-
-#if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES
-#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
-#else
-#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
-#endif
-
-
-#ifndef EIGEN_UNALIGNED_VECTORIZE
-#define EIGEN_UNALIGNED_VECTORIZE 1
-#endif
-
-//----------------------------------------------------------------------
-
-
 #ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD
   #define EIGEN_RESTRICT
 #endif
@@ -836,10 +839,6 @@ namespace Eigen {
   #define EIGEN_RESTRICT __restrict
 #endif
 
-#ifndef EIGEN_STACK_ALLOCATION_LIMIT
-// 131072 == 128 KB
-#define EIGEN_STACK_ALLOCATION_LIMIT 131072
-#endif
 
 #ifndef EIGEN_DEFAULT_IO_FORMAT
 #ifdef EIGEN_MAKING_DOCS
@@ -854,6 +853,18 @@ namespace Eigen {
 // just an empty macro !
 #define EIGEN_EMPTY
 
+
+// When compiling CUDA/HIP device code with NVCC or HIPCC
+// pull in math functions from the global namespace.
+// In host mode, and when device code is compiled with clang,
+// use the std versions.
+#if (defined(EIGEN_CUDA_ARCH) && defined(__NVCC__)) || defined(EIGEN_HIP_DEVICE_COMPILE)
+  #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
+#else
+  #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
+#endif
+
+
 #if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 || EIGEN_CUDACC_VER>0)
   // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324)
   #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
@@ -1002,6 +1013,11 @@ namespace Eigen {
   EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME)
 
 
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(EIGEN_CUDA_ARCH) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL) && !defined(EIGEN_HIP_DEVICE_COMPILE)
+  #define EIGEN_EXCEPTIONS
+#endif
+
+
 #ifdef EIGEN_EXCEPTIONS
 #  define EIGEN_THROW_X(X) throw X
 #  define EIGEN_THROW throw

From eb3d8f68bb3103210b6ed3ea13cde2787d2613ab Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 16:59:47 +0200
Subject: [PATCH 439/680] fix unused warning

---
 Eigen/src/Core/Redux.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index ddce65468..e449ef3ac 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -332,6 +332,7 @@ struct redux_impl<Func, Evaluator, LinearVectorizedTraversal, CompleteUnrolling>
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
   Scalar run(const Evaluator &eval, const Func& func, const XprType &xpr)
   {
+    EIGEN_ONLY_USED_FOR_DEBUG(xpr)
     eigen_assert(xpr.rows()>0 && xpr.cols()>0 && "you are using an empty matrix");
     if (VectorizedSize > 0) {
       Scalar res = func.predux(redux_vec_unroller<Func, Evaluator, 0, Size / PacketSize>::run(eval,func));

From 98728312c846613729158c0d8d00eb65b264573e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:00:37 +0200
Subject: [PATCH 440/680] Fix compilation regarding std::array

---
 Eigen/src/Core/util/Meta.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 748f24b1e..33dc0034c 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -360,6 +360,13 @@ template<typename T, int N> struct array_size<T (&)[N]> {
 };
 
 #if EIGEN_HAS_CXX11
+}}
+namespace std {
+  template<class T, std::size_t N> struct array;
+}
+namespace Eigen {
+namespace internal {
+
 template<typename T, std::size_t N> struct array_size<const std::array<T,N> > {
   enum { value = N };
 };

From 5c73c9223ac5911c2cf31cfd55d7987bd2b08281 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:01:07 +0200
Subject: [PATCH 441/680] Fix shadowing typedefs

---
 Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h | 2 --
 Eigen/src/Jacobi/Jacobi.h                     | 2 +-
 2 files changed, 1 insertion(+), 3 deletions(-)

diff --git a/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h b/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
index dbbd4806a..66e5a3dbb 100644
--- a/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
+++ b/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
@@ -66,7 +66,6 @@ template<typename Derived>
 inline typename MatrixBase<Derived>::EigenvaluesReturnType
 MatrixBase<Derived>::eigenvalues() const
 {
-  typedef typename internal::traits<Derived>::Scalar Scalar;
   return internal::eigenvalues_selector<Derived, NumTraits<Scalar>::IsComplex>::run(derived());
 }
 
@@ -88,7 +87,6 @@ template<typename MatrixType, unsigned int UpLo>
 EIGEN_DEVICE_FUNC inline typename SelfAdjointView<MatrixType, UpLo>::EigenvaluesReturnType
 SelfAdjointView<MatrixType, UpLo>::eigenvalues() const
 {
-  typedef typename SelfAdjointView<MatrixType, UpLo>::PlainObject PlainObject;
   PlainObject thisAsMatrix(*this);
   return SelfAdjointEigenSolver<PlainObject>(thisAsMatrix, false).eigenvalues();
 }
diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index bba75fc4f..176ff9d07 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -95,7 +95,7 @@ bool JacobiRotation<Scalar>::makeJacobi(const RealScalar& x, const Scalar& y, co
 {
   using std::sqrt;
   using std::abs;
-  typedef typename NumTraits<Scalar>::Real RealScalar;
+
   RealScalar deno = RealScalar(2)*abs(y);
   if(deno < (std::numeric_limits<RealScalar>::min)())
   {

From bec013b2c965880a6dc50a0d3a12b52d07b30352 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:02:18 +0200
Subject: [PATCH 442/680] fix unused warning

---
 test/gpu_common.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/test/gpu_common.h b/test/gpu_common.h
index 3030af6dc..60a78ccd7 100644
--- a/test/gpu_common.h
+++ b/test/gpu_common.h
@@ -88,6 +88,9 @@ void run_and_compare_to_gpu(const Kernel& ker, int n, const Input& in, Output& o
   #if !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
   in_ref = in_gpu = in;
   out_ref = out_gpu = out;
+  #else
+  EIGEN_UNUSED_VARIABLE(in);
+  EIGEN_UNUSED_VARIABLE(out);
   #endif
   run_on_cpu (ker, n, in_ref,  out_ref);
   run_on_gpu(ker, n, in_gpu, out_gpu);

From 63185be8b240ee1a9ef476af2fc06676d3d89fe0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:03:14 +0200
Subject: [PATCH 443/680] Disable eigenvalues test for clang-cuda

---
 test/gpu_basic.cu | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/test/gpu_basic.cu b/test/gpu_basic.cu
index 897834dff..285b87910 100644
--- a/test/gpu_basic.cu
+++ b/test/gpu_basic.cu
@@ -129,7 +129,7 @@ struct eigenvalues_direct {
     Map<Vec> res(out+i*Vec::MaxSizeAtCompileTime);
     T A = M*M.adjoint();
     SelfAdjointEigenSolver<T> eig;
-    eig.computeDirect(M);
+    eig.computeDirect(A);
     res = eig.eigenvalues();
   }
 };
@@ -145,7 +145,7 @@ struct eigenvalues {
     Map<Vec> res(out+i*Vec::MaxSizeAtCompileTime);
     T A = M*M.adjoint();
     SelfAdjointEigenSolver<T> eig;
-    eig.compute(M);
+    eig.compute(A);
     res = eig.eigenvalues();
   }
 };
@@ -202,11 +202,14 @@ void test_gpu_basic()
   CALL_SUBTEST( run_and_compare_to_gpu(matrix_inverse<Matrix3f>(), nthreads, in, out) );
   CALL_SUBTEST( run_and_compare_to_gpu(matrix_inverse<Matrix4f>(), nthreads, in, out) );
   
-#if !defined(EIGEN_USE_HIP)
-  // FIXME
-  // These subtests result in a linking error on the HIP platform
   CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues_direct<Matrix3f>(), nthreads, in, out) );
   CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues_direct<Matrix2f>(), nthreads, in, out) );
+
+#if defined(__NVCC__)
+  // FIXME
+  // These subtests compiles only with nvcc and fail with HIPCC and clang-cuda
   CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues<Matrix4f>(), nthreads, in, out) );
+  typedef Matrix<float,6,6> Matrix6f;
+  CALL_SUBTEST( run_and_compare_to_gpu(eigenvalues<Matrix6f>(), nthreads, in, out) );
 #endif
 }

From 12e1ebb68b93b218a3553af67fee8535d79e857e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:16:40 +0200
Subject: [PATCH 444/680] Remove local Index typedef from unit-tests

---
 test/adjoint.cpp                      |  1 -
 test/array.cpp                        |  7 +------
 test/array_for_matrix.cpp             |  6 +-----
 test/array_replicate.cpp              |  1 -
 test/array_reverse.cpp                |  1 -
 test/basicstuff.cpp                   | 10 ++++------
 test/block.cpp                        |  3 ---
 test/cholesky.cpp                     |  3 ---
 test/conservative_resize.cpp          |  1 -
 test/corners.cpp                      |  1 -
 test/determinant.cpp                  |  1 -
 test/diagonal.cpp                     |  1 -
 test/diagonalmatrices.cpp             |  1 -
 test/dontalign.cpp                    |  1 -
 test/eigen2support.cpp                |  1 -
 test/eigensolver_complex.cpp          |  1 -
 test/eigensolver_generalized_real.cpp |  1 -
 test/eigensolver_generic.cpp          |  1 -
 test/eigensolver_selfadjoint.cpp      |  1 -
 test/geo_alignedbox.cpp               |  2 --
 test/geo_hyperplane.cpp               |  1 -
 test/geo_parametrizedline.cpp         |  1 -
 test/householder.cpp                  |  1 -
 test/integer_types.cpp                |  2 --
 test/inverse.cpp                      |  1 -
 test/jacobi.cpp                       |  1 -
 test/jacobisvd.cpp                    |  1 -
 test/linearstructure.cpp              |  1 -
 test/lu.cpp                           |  2 --
 test/mapped_matrix.cpp                |  4 ----
 test/mapstaticmethods.cpp             |  2 --
 test/mapstride.cpp                    |  2 --
 test/miscmatrices.cpp                 |  1 -
 test/nomalloc.cpp                     |  3 +--
 test/permutationmatrices.cpp          |  1 -
 test/product_extra.cpp                |  1 -
 test/product_notemporary.cpp          |  1 -
 test/product_selfadjoint.cpp          |  1 -
 test/product_symm.cpp                 |  1 -
 test/product_syrk.cpp                 |  1 -
 test/product_trmv.cpp                 |  1 -
 test/qr.cpp                           |  2 --
 test/qr_colpivoting.cpp               |  4 ----
 test/qr_fullpivoting.cpp              |  2 --
 test/qtvector.cpp                     |  2 --
 test/real_qz.cpp                      |  1 -
 test/redux.cpp                        |  2 --
 test/ref.cpp                          |  2 --
 test/schur_real.cpp                   |  2 --
 test/selfadjoint.cpp                  |  1 -
 test/sizeof.cpp                       |  2 +-
 test/stable_norm.cpp                  |  1 -
 test/stddeque.cpp                     |  2 --
 test/stddeque_overload.cpp            |  4 ++--
 test/stdlist.cpp                      |  2 --
 test/stdlist_overload.cpp             |  4 ++--
 test/stdvector.cpp                    |  4 ++--
 test/stdvector_overload.cpp           |  4 ++--
 test/svd_common.h                     |  5 -----
 test/svd_fill.h                       |  1 -
 test/swap.cpp                         |  4 ++--
 test/triangular.cpp                   |  5 ++---
 test/upperbidiagonalization.cpp       |  4 ++--
 test/vectorwiseop.cpp                 |  2 --
 test/visitor.cpp                      |  2 --
 65 files changed, 22 insertions(+), 117 deletions(-)

diff --git a/test/adjoint.cpp b/test/adjoint.cpp
index bdea51c10..37032d220 100644
--- a/test/adjoint.cpp
+++ b/test/adjoint.cpp
@@ -70,7 +70,6 @@ template<typename MatrixType> void adjoint(const MatrixType& m)
      Transpose.h Conjugate.h Dot.h
   */
   using std::abs;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
diff --git a/test/array.cpp b/test/array.cpp
index f7f3ba780..7e228ed54 100644
--- a/test/array.cpp
+++ b/test/array.cpp
@@ -11,7 +11,6 @@
 
 template<typename ArrayType> void array(const ArrayType& m)
 {
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
   typedef typename ArrayType::RealScalar RealScalar;
   typedef Array<Scalar, ArrayType::RowsAtCompileTime, 1> ColVectorType;
@@ -130,7 +129,6 @@ template<typename ArrayType> void array(const ArrayType& m)
 template<typename ArrayType> void comparisons(const ArrayType& m)
 {
   using std::abs;
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -197,7 +195,7 @@ template<typename ArrayType> void comparisons(const ArrayType& m)
   RealScalar a = m1.abs().mean();
   VERIFY( (m1<-a || m1>a).count() == (m1.abs()>a).count());
 
-  typedef Array<typename ArrayType::Index, Dynamic, 1> ArrayOfIndices;
+  typedef Array<Index, Dynamic, 1> ArrayOfIndices;
 
   // TODO allows colwise/rowwise for array
   VERIFY_IS_APPROX(((m1.abs()+1)>RealScalar(0.1)).colwise().count(), ArrayOfIndices::Constant(cols,rows).transpose());
@@ -208,7 +206,6 @@ template<typename ArrayType> void array_real(const ArrayType& m)
 {
   using std::abs;
   using std::sqrt;
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -322,7 +319,6 @@ template<typename ArrayType> void array_real(const ArrayType& m)
 
 template<typename ArrayType> void array_complex(const ArrayType& m)
 {
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -427,7 +423,6 @@ template<typename ArrayType> void array_complex(const ArrayType& m)
 
 template<typename ArrayType> void min_max(const ArrayType& m)
 {
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
 
   Index rows = m.rows();
diff --git a/test/array_for_matrix.cpp b/test/array_for_matrix.cpp
index b8721391f..3dc5dc487 100644
--- a/test/array_for_matrix.cpp
+++ b/test/array_for_matrix.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void array_for_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> ColVectorType;
   typedef Matrix<Scalar, 1, MatrixType::ColsAtCompileTime> RowVectorType; 
@@ -83,7 +82,6 @@ template<typename MatrixType> void array_for_matrix(const MatrixType& m)
 template<typename MatrixType> void comparisons(const MatrixType& m)
 {
   using std::abs;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -140,7 +138,7 @@ template<typename MatrixType> void comparisons(const MatrixType& m)
   RealScalar a = m1.cwiseAbs().mean();
   VERIFY( ((m1.array()<-a).matrix() || (m1.array()>a).matrix()).count() == (m1.cwiseAbs().array()>a).count());
 
-  typedef Matrix<typename MatrixType::Index, Dynamic, 1> VectorOfIndices;
+  typedef Matrix<Index, Dynamic, 1> VectorOfIndices;
 
   // TODO allows colwise/rowwise for array
   VERIFY_IS_APPROX(((m1.array().abs()+1)>RealScalar(0.1)).matrix().colwise().count(), VectorOfIndices::Constant(cols,rows).transpose());
@@ -172,7 +170,6 @@ template<typename VectorType> void lpNorm(const VectorType& v)
 
 template<typename MatrixType> void cwise_min_max(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows();
@@ -211,7 +208,6 @@ template<typename MatrixType> void cwise_min_max(const MatrixType& m)
 
 template<typename MatrixTraits> void resize(const MatrixTraits& t)
 {
-  typedef typename MatrixTraits::Index Index;
   typedef typename MatrixTraits::Scalar Scalar;
   typedef Matrix<Scalar,Dynamic,Dynamic> MatrixType;
   typedef Array<Scalar,Dynamic,Dynamic> Array2DType;
diff --git a/test/array_replicate.cpp b/test/array_replicate.cpp
index 779c8fc2f..0dad5bace 100644
--- a/test/array_replicate.cpp
+++ b/test/array_replicate.cpp
@@ -14,7 +14,6 @@ template<typename MatrixType> void replicate(const MatrixType& m)
   /* this test covers the following files:
      Replicate.cpp
   */
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   typedef Matrix<Scalar, Dynamic, Dynamic> MatrixX;
diff --git a/test/array_reverse.cpp b/test/array_reverse.cpp
index c9d9f90c3..9d5b9a66d 100644
--- a/test/array_reverse.cpp
+++ b/test/array_reverse.cpp
@@ -15,7 +15,6 @@ using namespace std;
 
 template<typename MatrixType> void reverse(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
 
diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp
index c346ce6cb..0fbeb42eb 100644
--- a/test/basicstuff.cpp
+++ b/test/basicstuff.cpp
@@ -13,7 +13,6 @@
 
 template<typename MatrixType> void basicStuff(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, MatrixType::RowsAtCompileTime> SquareMatrixType;
@@ -124,22 +123,22 @@ template<typename MatrixType> void basicStuff(const MatrixType& m)
   
   // check automatic transposition
   sm2.setZero();
-  for(typename MatrixType::Index i=0;i<rows;++i)
+  for(Index i=0;i<rows;++i)
     sm2.col(i) = sm1.row(i);
   VERIFY_IS_APPROX(sm2,sm1.transpose());
   
   sm2.setZero();
-  for(typename MatrixType::Index i=0;i<rows;++i)
+  for(Index i=0;i<rows;++i)
     sm2.col(i).noalias() = sm1.row(i);
   VERIFY_IS_APPROX(sm2,sm1.transpose());
   
   sm2.setZero();
-  for(typename MatrixType::Index i=0;i<rows;++i)
+  for(Index i=0;i<rows;++i)
     sm2.col(i).noalias() += sm1.row(i);
   VERIFY_IS_APPROX(sm2,sm1.transpose());
   
   sm2.setZero();
-  for(typename MatrixType::Index i=0;i<rows;++i)
+  for(Index i=0;i<rows;++i)
     sm2.col(i).noalias() -= sm1.row(i);
   VERIFY_IS_APPROX(sm2,-sm1.transpose());
   
@@ -160,7 +159,6 @@ template<typename MatrixType> void basicStuff(const MatrixType& m)
 
 template<typename MatrixType> void basicStuffComplex(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef Matrix<RealScalar, MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime> RealMatrixType;
diff --git a/test/block.cpp b/test/block.cpp
index 0c1d2b7b8..a45dd501d 100644
--- a/test/block.cpp
+++ b/test/block.cpp
@@ -39,7 +39,6 @@ is_same_block(const T1& a, const T2& b)
 
 template<typename MatrixType> void block(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
@@ -227,7 +226,6 @@ template<typename MatrixType> void block(const MatrixType& m)
 template<typename MatrixType>
 void compare_using_data_and_stride(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
   Index size = m.size();
@@ -261,7 +259,6 @@ void compare_using_data_and_stride(const MatrixType& m)
 template<typename MatrixType>
 void data_and_stride(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/cholesky.cpp b/test/cholesky.cpp
index b4b6bda7d..62c391de4 100644
--- a/test/cholesky.cpp
+++ b/test/cholesky.cpp
@@ -57,7 +57,6 @@ template<typename MatrixType,template <typename,int> class CholType> void test_c
 
 template<typename MatrixType> void cholesky(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      LLT.h LDLT.h
   */
@@ -289,8 +288,6 @@ template<typename MatrixType> void cholesky_cplx(const MatrixType& m)
 
   // test mixing real/scalar types
 
-  typedef typename MatrixType::Index Index;
-
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/conservative_resize.cpp b/test/conservative_resize.cpp
index 498421b4c..21a1db4ac 100644
--- a/test/conservative_resize.cpp
+++ b/test/conservative_resize.cpp
@@ -17,7 +17,6 @@ template <typename Scalar, int Storage>
 void run_matrix_tests()
 {
   typedef Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic, Storage> MatrixType;
-  typedef typename MatrixType::Index Index;
 
   MatrixType m, n;
 
diff --git a/test/corners.cpp b/test/corners.cpp
index 3c64c32a1..32edadb25 100644
--- a/test/corners.cpp
+++ b/test/corners.cpp
@@ -15,7 +15,6 @@
 
 template<typename MatrixType> void corners(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/determinant.cpp b/test/determinant.cpp
index 758f3afbb..b8c9babb3 100644
--- a/test/determinant.cpp
+++ b/test/determinant.cpp
@@ -16,7 +16,6 @@ template<typename MatrixType> void determinant(const MatrixType& m)
   /* this test covers the following files:
      Determinant.h
   */
-  typedef typename MatrixType::Index Index;
   Index size = m.rows();
 
   MatrixType m1(size, size), m2(size, size);
diff --git a/test/diagonal.cpp b/test/diagonal.cpp
index 0b5ae8280..8ed9b4682 100644
--- a/test/diagonal.cpp
+++ b/test/diagonal.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void diagonal(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows();
diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp
index a4ff10239..c55733df8 100644
--- a/test/diagonalmatrices.cpp
+++ b/test/diagonalmatrices.cpp
@@ -11,7 +11,6 @@
 using namespace std;
 template<typename MatrixType> void diagonalmatrices(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   enum { Rows = MatrixType::RowsAtCompileTime, Cols = MatrixType::ColsAtCompileTime };
   typedef Matrix<Scalar, Rows, 1> VectorType;
diff --git a/test/dontalign.cpp b/test/dontalign.cpp
index 4643cfed6..ac00112ed 100644
--- a/test/dontalign.cpp
+++ b/test/dontalign.cpp
@@ -19,7 +19,6 @@
 template<typename MatrixType>
 void dontalign(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, MatrixType::RowsAtCompileTime> SquareMatrixType;
diff --git a/test/eigen2support.cpp b/test/eigen2support.cpp
index ad1d98091..ac6931a0e 100644
--- a/test/eigen2support.cpp
+++ b/test/eigen2support.cpp
@@ -13,7 +13,6 @@
 
 template<typename MatrixType> void eigen2support(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows();
diff --git a/test/eigensolver_complex.cpp b/test/eigensolver_complex.cpp
index 03d5774ef..8f279dc00 100644
--- a/test/eigensolver_complex.cpp
+++ b/test/eigensolver_complex.cpp
@@ -71,7 +71,6 @@ void verify_is_approx_upto_permutation(const VectorType& vec1, const VectorType&
 
 template<typename MatrixType> void eigensolver(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      ComplexEigenSolver.h, and indirectly ComplexSchur.h
   */
diff --git a/test/eigensolver_generalized_real.cpp b/test/eigensolver_generalized_real.cpp
index 9c0838ba4..12a15ab81 100644
--- a/test/eigensolver_generalized_real.cpp
+++ b/test/eigensolver_generalized_real.cpp
@@ -15,7 +15,6 @@
 
 template<typename MatrixType> void generalized_eigensolver_real(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      GeneralizedEigenSolver.h
   */
diff --git a/test/eigensolver_generic.cpp b/test/eigensolver_generic.cpp
index d0e644d4b..07bf65e03 100644
--- a/test/eigensolver_generic.cpp
+++ b/test/eigensolver_generic.cpp
@@ -14,7 +14,6 @@
 
 template<typename MatrixType> void eigensolver(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      EigenSolver.h
   */
diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp
index 39ad4130e..0e39b5364 100644
--- a/test/eigensolver_selfadjoint.cpp
+++ b/test/eigensolver_selfadjoint.cpp
@@ -68,7 +68,6 @@ template<typename MatrixType> void selfadjointeigensolver_essential_check(const
 
 template<typename MatrixType> void selfadjointeigensolver(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      EigenSolver.h, SelfAdjointEigenSolver.h (and indirectly: Tridiagonalization.h)
   */
diff --git a/test/geo_alignedbox.cpp b/test/geo_alignedbox.cpp
index 223ff5eea..87bf1140f 100644
--- a/test/geo_alignedbox.cpp
+++ b/test/geo_alignedbox.cpp
@@ -33,7 +33,6 @@ template<typename BoxType> void alignedbox(const BoxType& _box)
   /* this test covers the following files:
      AlignedBox.h
   */
-  typedef typename BoxType::Index Index;  
   typedef typename BoxType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef Matrix<Scalar, BoxType::AmbientDimAtCompileTime, 1> VectorType;
@@ -95,7 +94,6 @@ template<typename BoxType>
 void alignedboxCastTests(const BoxType& _box)
 {
   // casting  
-  typedef typename BoxType::Index Index;
   typedef typename BoxType::Scalar Scalar;
   typedef Matrix<Scalar, BoxType::AmbientDimAtCompileTime, 1> VectorType;
 
diff --git a/test/geo_hyperplane.cpp b/test/geo_hyperplane.cpp
index 27892850d..b3a48c585 100644
--- a/test/geo_hyperplane.cpp
+++ b/test/geo_hyperplane.cpp
@@ -19,7 +19,6 @@ template<typename HyperplaneType> void hyperplane(const HyperplaneType& _plane)
      Hyperplane.h
   */
   using std::abs;
-  typedef typename HyperplaneType::Index Index;
   const Index dim = _plane.dim();
   enum { Options = HyperplaneType::Options };
   typedef typename HyperplaneType::Scalar Scalar;
diff --git a/test/geo_parametrizedline.cpp b/test/geo_parametrizedline.cpp
index 29c1b105c..14cf4fd48 100644
--- a/test/geo_parametrizedline.cpp
+++ b/test/geo_parametrizedline.cpp
@@ -19,7 +19,6 @@ template<typename LineType> void parametrizedline(const LineType& _line)
      ParametrizedLine.h
   */
   using std::abs;
-  typedef typename LineType::Index Index;
   const Index dim = _line.dim();
   typedef typename LineType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
diff --git a/test/householder.cpp b/test/householder.cpp
index 78d44d48a..ccda46811 100644
--- a/test/householder.cpp
+++ b/test/householder.cpp
@@ -12,7 +12,6 @@
 
 template<typename MatrixType> void householder(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   static bool even = true;
   even = !even;
   /* this test covers the following files:
diff --git a/test/integer_types.cpp b/test/integer_types.cpp
index 25126315a..36295598f 100644
--- a/test/integer_types.cpp
+++ b/test/integer_types.cpp
@@ -18,7 +18,6 @@
 
 template<typename MatrixType> void signed_integer_type_tests(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   enum { is_signed = (Scalar(-1) > Scalar(0)) ? 0 : 1 };
@@ -49,7 +48,6 @@ template<typename MatrixType> void signed_integer_type_tests(const MatrixType& m
 
 template<typename MatrixType> void integer_type_tests(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   VERIFY(NumTraits<Scalar>::IsInteger);
diff --git a/test/inverse.cpp b/test/inverse.cpp
index 97fe6ff92..be607cc8b 100644
--- a/test/inverse.cpp
+++ b/test/inverse.cpp
@@ -14,7 +14,6 @@
 template<typename MatrixType> void inverse(const MatrixType& m)
 {
   using std::abs;
-  typedef typename MatrixType::Index Index;
   /* this test covers the following files:
      Inverse.h
   */
diff --git a/test/jacobi.cpp b/test/jacobi.cpp
index 7ccd4124b..319e4767a 100644
--- a/test/jacobi.cpp
+++ b/test/jacobi.cpp
@@ -14,7 +14,6 @@
 template<typename MatrixType, typename JacobiScalar>
 void jacobi(const MatrixType& m = MatrixType())
 {
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp
index 7f5f71562..72007bff5 100644
--- a/test/jacobisvd.cpp
+++ b/test/jacobisvd.cpp
@@ -36,7 +36,6 @@ void jacobisvd(const MatrixType& a = MatrixType(), bool pickrandom = true)
 template<typename MatrixType> void jacobisvd_verify_assert(const MatrixType& m)
 {
   svd_verify_assert<JacobiSVD<MatrixType> >(m);
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp
index 17474af10..b6559b2a0 100644
--- a/test/linearstructure.cpp
+++ b/test/linearstructure.cpp
@@ -19,7 +19,6 @@ template<typename MatrixType> void linearStructure(const MatrixType& m)
   /* this test covers the following files:
      CwiseUnaryOp.h, CwiseBinaryOp.h, SelfCwiseBinaryOp.h 
   */
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
 
diff --git a/test/lu.cpp b/test/lu.cpp
index 9787f4d86..cad703dd0 100644
--- a/test/lu.cpp
+++ b/test/lu.cpp
@@ -18,7 +18,6 @@ typename MatrixType::RealScalar matrix_l1_norm(const MatrixType& m) {
 
 template<typename MatrixType> void lu_non_invertible()
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::RealScalar RealScalar;
   /* this test covers the following files:
      LU.h
@@ -181,7 +180,6 @@ template<typename MatrixType> void lu_partial_piv()
   /* this test covers the following files:
      PartialPivLU.h
   */
-  typedef typename MatrixType::Index Index;
   typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
   Index size = internal::random<Index>(1,4);
 
diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp
index 7b9e1fe52..a530fab41 100644
--- a/test/mapped_matrix.cpp
+++ b/test/mapped_matrix.cpp
@@ -17,7 +17,6 @@
 
 template<typename VectorType> void map_class_vector(const VectorType& m)
 {
-  typedef typename VectorType::Index Index;
   typedef typename VectorType::Scalar Scalar;
 
   Index size = m.size();
@@ -51,7 +50,6 @@ template<typename VectorType> void map_class_vector(const VectorType& m)
 
 template<typename MatrixType> void map_class_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows(), cols = m.cols(), size = rows*cols;
@@ -121,7 +119,6 @@ template<typename MatrixType> void map_class_matrix(const MatrixType& m)
 
 template<typename VectorType> void map_static_methods(const VectorType& m)
 {
-  typedef typename VectorType::Index Index;
   typedef typename VectorType::Scalar Scalar;
 
   Index size = m.size();
@@ -163,7 +160,6 @@ template<typename Scalar>
 void map_not_aligned_on_scalar()
 {
   typedef Matrix<Scalar,Dynamic,Dynamic> MatrixType;
-  typedef typename MatrixType::Index Index;
   Index size = 11;
   Scalar* array1 = internal::aligned_new<Scalar>((size+1)*(size+1)+1);
   Scalar* array2 = reinterpret_cast<Scalar*>(sizeof(Scalar)/2+std::size_t(array1));
diff --git a/test/mapstaticmethods.cpp b/test/mapstaticmethods.cpp
index 06272d106..8156ca939 100644
--- a/test/mapstaticmethods.cpp
+++ b/test/mapstaticmethods.cpp
@@ -69,7 +69,6 @@ struct mapstaticmethods_impl<PlainObjectType, true, false>
 {
   static void run(const PlainObjectType& m)
   {
-    typedef typename PlainObjectType::Index Index;
     Index rows = m.rows(), cols = m.cols();
 
     int i = internal::random<int>(2,5), j = internal::random<int>(2,5);
@@ -116,7 +115,6 @@ struct mapstaticmethods_impl<PlainObjectType, true, true>
 {
   static void run(const PlainObjectType& v)
   {
-    typedef typename PlainObjectType::Index Index;
     Index size = v.size();
 
     int i = internal::random<int>(2,5);
diff --git a/test/mapstride.cpp b/test/mapstride.cpp
index de77dc5de..d785148cf 100644
--- a/test/mapstride.cpp
+++ b/test/mapstride.cpp
@@ -11,7 +11,6 @@
 
 template<int Alignment,typename VectorType> void map_class_vector(const VectorType& m)
 {
-  typedef typename VectorType::Index Index;
   typedef typename VectorType::Scalar Scalar;
 
   Index size = m.size();
@@ -50,7 +49,6 @@ template<int Alignment,typename VectorType> void map_class_vector(const VectorTy
 
 template<int Alignment,typename MatrixType> void map_class_matrix(const MatrixType& _m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = _m.rows(), cols = _m.cols();
diff --git a/test/miscmatrices.cpp b/test/miscmatrices.cpp
index ef20dc749..f17291c40 100644
--- a/test/miscmatrices.cpp
+++ b/test/miscmatrices.cpp
@@ -14,7 +14,6 @@ template<typename MatrixType> void miscMatrices(const MatrixType& m)
   /* this test covers the following files:
      DiagonalMatrix.h Ones.h
   */
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   Index rows = m.rows();
diff --git a/test/nomalloc.cpp b/test/nomalloc.cpp
index 50756c2fb..d1ca79d68 100644
--- a/test/nomalloc.cpp
+++ b/test/nomalloc.cpp
@@ -24,7 +24,6 @@ template<typename MatrixType> void nomalloc(const MatrixType& m)
 {
   /* this test check no dynamic memory allocation are issued with fixed-size matrices
   */
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows();
@@ -173,7 +172,7 @@ template<typename MatrixType> void test_reference(const MatrixType& m) {
   typedef typename MatrixType::Scalar Scalar;
   enum { Flag          =  MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor};
   enum { TransposeFlag = !MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor};
-  typename MatrixType::Index rows = m.rows(), cols=m.cols();
+  Index rows = m.rows(), cols=m.cols();
   typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic, Flag         > MatrixX;
   typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic, TransposeFlag> MatrixXT;
   // Dynamic reference:
diff --git a/test/permutationmatrices.cpp b/test/permutationmatrices.cpp
index b2229cfe9..e885f0e04 100644
--- a/test/permutationmatrices.cpp
+++ b/test/permutationmatrices.cpp
@@ -14,7 +14,6 @@
 using namespace std;
 template<typename MatrixType> void permutationmatrices(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   enum { Rows = MatrixType::RowsAtCompileTime, Cols = MatrixType::ColsAtCompileTime,
          Options = MatrixType::Options };
diff --git a/test/product_extra.cpp b/test/product_extra.cpp
index e2b855bff..de2709d8b 100644
--- a/test/product_extra.cpp
+++ b/test/product_extra.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void product_extra(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, 1, Dynamic> RowVectorType;
   typedef Matrix<Scalar, Dynamic, 1> ColVectorType;
diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp
index 062180f42..b90165b2f 100644
--- a/test/product_notemporary.cpp
+++ b/test/product_notemporary.cpp
@@ -15,7 +15,6 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m)
 {
   /* This test checks the number of temporaries created
    * during the evaluation of a complex expression */
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
   typedef Matrix<Scalar, 1, Dynamic> RowVectorType;
diff --git a/test/product_selfadjoint.cpp b/test/product_selfadjoint.cpp
index 3d768aa7e..88d68391b 100644
--- a/test/product_selfadjoint.cpp
+++ b/test/product_selfadjoint.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void product_selfadjoint(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
   typedef Matrix<Scalar, 1, MatrixType::RowsAtCompileTime> RowVectorType;
diff --git a/test/product_symm.cpp b/test/product_symm.cpp
index 8c44383f9..7d1042a4f 100644
--- a/test/product_symm.cpp
+++ b/test/product_symm.cpp
@@ -16,7 +16,6 @@ template<typename Scalar, int Size, int OtherSize> void symm(int size = Size, in
   typedef Matrix<Scalar, OtherSize, Size> Rhs2;
   enum { order = OtherSize==1 ? 0 : RowMajor };
   typedef Matrix<Scalar, Size, OtherSize,order> Rhs3;
-  typedef typename MatrixType::Index Index;
 
   Index rows = size;
   Index cols = size;
diff --git a/test/product_syrk.cpp b/test/product_syrk.cpp
index e10f0f2f2..3ebbe14ca 100644
--- a/test/product_syrk.cpp
+++ b/test/product_syrk.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void syrk(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime, RowMajor> RMatrixType;
   typedef Matrix<Scalar, MatrixType::ColsAtCompileTime, Dynamic> Rhs1;
diff --git a/test/product_trmv.cpp b/test/product_trmv.cpp
index 57a202afc..65d66e57b 100644
--- a/test/product_trmv.cpp
+++ b/test/product_trmv.cpp
@@ -11,7 +11,6 @@
 
 template<typename MatrixType> void trmv(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> VectorType;
diff --git a/test/qr.cpp b/test/qr.cpp
index 00ef99ef9..02e3ed74c 100644
--- a/test/qr.cpp
+++ b/test/qr.cpp
@@ -12,8 +12,6 @@
 
 template<typename MatrixType> void qr(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
-
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/qr_colpivoting.cpp b/test/qr_colpivoting.cpp
index 26ed27f5c..96c0badb7 100644
--- a/test/qr_colpivoting.cpp
+++ b/test/qr_colpivoting.cpp
@@ -14,8 +14,6 @@
 
 template <typename MatrixType>
 void cod() {
-  typedef typename MatrixType::Index Index;
-
   Index rows = internal::random<Index>(2, EIGEN_TEST_MAX_SIZE);
   Index cols = internal::random<Index>(2, EIGEN_TEST_MAX_SIZE);
   Index cols2 = internal::random<Index>(2, EIGEN_TEST_MAX_SIZE);
@@ -94,7 +92,6 @@ void cod_fixedsize() {
 template<typename MatrixType> void qr()
 {
   using std::sqrt;
-  typedef typename MatrixType::Index Index;
 
   Index rows = internal::random<Index>(2,EIGEN_TEST_MAX_SIZE), cols = internal::random<Index>(2,EIGEN_TEST_MAX_SIZE), cols2 = internal::random<Index>(2,EIGEN_TEST_MAX_SIZE);
   Index rank = internal::random<Index>(1, (std::min)(rows, cols)-1);
@@ -211,7 +208,6 @@ template<typename MatrixType> void qr_kahan_matrix()
 {
   using std::sqrt;
   using std::abs;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
 
diff --git a/test/qr_fullpivoting.cpp b/test/qr_fullpivoting.cpp
index 3d9accea6..c270edf1a 100644
--- a/test/qr_fullpivoting.cpp
+++ b/test/qr_fullpivoting.cpp
@@ -13,8 +13,6 @@
 
 template<typename MatrixType> void qr()
 {
-  typedef typename MatrixType::Index Index;
-
   static const int Rows = MatrixType::RowsAtCompileTime, Cols = MatrixType::ColsAtCompileTime;
   Index max_size = EIGEN_TEST_MAX_SIZE;
   Index min_size = numext::maxi(1,EIGEN_TEST_MAX_SIZE/10);
diff --git a/test/qtvector.cpp b/test/qtvector.cpp
index 2be885e48..22df0d515 100644
--- a/test/qtvector.cpp
+++ b/test/qtvector.cpp
@@ -18,8 +18,6 @@
 template<typename MatrixType>
 void check_qtvector_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
-
   Index rows = m.rows();
   Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
diff --git a/test/real_qz.cpp b/test/real_qz.cpp
index 99ac31235..3c1492e4b 100644
--- a/test/real_qz.cpp
+++ b/test/real_qz.cpp
@@ -18,7 +18,6 @@ template<typename MatrixType> void real_qz(const MatrixType& m)
      RealQZ.h
   */
   using std::abs;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   
   Index dim = m.cols();
diff --git a/test/redux.cpp b/test/redux.cpp
index 2bade3735..213f080aa 100644
--- a/test/redux.cpp
+++ b/test/redux.cpp
@@ -16,7 +16,6 @@
 
 template<typename MatrixType> void matrixRedux(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
 
@@ -81,7 +80,6 @@ template<typename MatrixType> void matrixRedux(const MatrixType& m)
 template<typename VectorType> void vectorRedux(const VectorType& w)
 {
   using std::abs;
-  typedef typename VectorType::Index Index;
   typedef typename VectorType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   Index size = w.size();
diff --git a/test/ref.cpp b/test/ref.cpp
index 9dd2c0449..704495aff 100644
--- a/test/ref.cpp
+++ b/test/ref.cpp
@@ -32,7 +32,6 @@
 
 template<typename MatrixType> void ref_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
   typedef Matrix<Scalar,Dynamic,Dynamic,MatrixType::Options> DynMatrixType;
@@ -80,7 +79,6 @@ template<typename MatrixType> void ref_matrix(const MatrixType& m)
 
 template<typename VectorType> void ref_vector(const VectorType& m)
 {
-  typedef typename VectorType::Index Index;
   typedef typename VectorType::Scalar Scalar;
   typedef typename VectorType::RealScalar RealScalar;
   typedef Matrix<Scalar,Dynamic,1,VectorType::Options> DynMatrixType;
diff --git a/test/schur_real.cpp b/test/schur_real.cpp
index 4aede87df..e5229e6e8 100644
--- a/test/schur_real.cpp
+++ b/test/schur_real.cpp
@@ -13,8 +13,6 @@
 
 template<typename MatrixType> void verifyIsQuasiTriangular(const MatrixType& T)
 {
-  typedef typename MatrixType::Index Index;
-
   const Index size = T.cols();
   typedef typename MatrixType::Scalar Scalar;
 
diff --git a/test/selfadjoint.cpp b/test/selfadjoint.cpp
index aaa4888bd..bb11cc351 100644
--- a/test/selfadjoint.cpp
+++ b/test/selfadjoint.cpp
@@ -15,7 +15,6 @@
 
 template<typename MatrixType> void selfadjoint(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index rows = m.rows();
diff --git a/test/sizeof.cpp b/test/sizeof.cpp
index 03ad20453..e552bae7a 100644
--- a/test/sizeof.cpp
+++ b/test/sizeof.cpp
@@ -15,7 +15,7 @@ template<typename MatrixType> void verifySizeOf(const MatrixType&)
   if (MatrixType::RowsAtCompileTime!=Dynamic && MatrixType::ColsAtCompileTime!=Dynamic)
     VERIFY_IS_EQUAL(std::ptrdiff_t(sizeof(MatrixType)),std::ptrdiff_t(sizeof(Scalar))*std::ptrdiff_t(MatrixType::SizeAtCompileTime));
   else
-    VERIFY_IS_EQUAL(sizeof(MatrixType),sizeof(Scalar*) + 2 * sizeof(typename MatrixType::Index));
+    VERIFY_IS_EQUAL(sizeof(MatrixType),sizeof(Scalar*) + 2 * sizeof(Index));
 }
 
 void test_sizeof()
diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp
index 0dcf072fe..255cae9b3 100644
--- a/test/stable_norm.cpp
+++ b/test/stable_norm.cpp
@@ -21,7 +21,6 @@ template<typename MatrixType> void stable_norm(const MatrixType& m)
   */
   using std::sqrt;
   using std::abs;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   
diff --git a/test/stddeque.cpp b/test/stddeque.cpp
index bb4b476f3..b511c4e61 100644
--- a/test/stddeque.cpp
+++ b/test/stddeque.cpp
@@ -15,8 +15,6 @@
 template<typename MatrixType>
 void check_stddeque_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
-  
   Index rows = m.rows();
   Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
diff --git a/test/stddeque_overload.cpp b/test/stddeque_overload.cpp
index 4da618bbf..0541bc1bd 100644
--- a/test/stddeque_overload.cpp
+++ b/test/stddeque_overload.cpp
@@ -28,8 +28,8 @@ EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(Quaterniond)
 template<typename MatrixType>
 void check_stddeque_matrix(const MatrixType& m)
 {
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
   std::deque<MatrixType> v(10, MatrixType(rows,cols)), w(20, y);
   v[5] = x;
diff --git a/test/stdlist.cpp b/test/stdlist.cpp
index 17cce779a..23cbe9039 100644
--- a/test/stdlist.cpp
+++ b/test/stdlist.cpp
@@ -15,8 +15,6 @@
 template<typename MatrixType>
 void check_stdlist_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
-  
   Index rows = m.rows();
   Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
diff --git a/test/stdlist_overload.cpp b/test/stdlist_overload.cpp
index bb910bd43..386124ad4 100644
--- a/test/stdlist_overload.cpp
+++ b/test/stdlist_overload.cpp
@@ -44,8 +44,8 @@ void set(Container & c, Position position, const Value & value)
 template<typename MatrixType>
 void check_stdlist_matrix(const MatrixType& m)
 {
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
   std::list<MatrixType> v(10, MatrixType(rows,cols)), w(20, y);
   typename std::list<MatrixType>::iterator itv = get(v, 5);
diff --git a/test/stdvector.cpp b/test/stdvector.cpp
index 50cb3341d..cf3646870 100644
--- a/test/stdvector.cpp
+++ b/test/stdvector.cpp
@@ -14,8 +14,8 @@
 template<typename MatrixType>
 void check_stdvector_matrix(const MatrixType& m)
 {
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
   std::vector<MatrixType,Eigen::aligned_allocator<MatrixType> > v(10, MatrixType(rows,cols)), w(20, y);
   v[5] = x;
diff --git a/test/stdvector_overload.cpp b/test/stdvector_overload.cpp
index 959665954..0c86ce8a4 100644
--- a/test/stdvector_overload.cpp
+++ b/test/stdvector_overload.cpp
@@ -28,8 +28,8 @@ EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(Quaterniond)
 template<typename MatrixType>
 void check_stdvector_matrix(const MatrixType& m)
 {
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
   MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols);
   std::vector<MatrixType> v(10, MatrixType(rows,cols)), w(20, y);
   v[5] = x;
diff --git a/test/svd_common.h b/test/svd_common.h
index 605d5dfef..cba066593 100644
--- a/test/svd_common.h
+++ b/test/svd_common.h
@@ -23,7 +23,6 @@
 template<typename SvdType, typename MatrixType>
 void svd_check_full(const MatrixType& m, const SvdType& svd)
 {
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
@@ -101,7 +100,6 @@ void svd_least_square(const MatrixType& m, unsigned int computationOptions)
 {
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
@@ -168,7 +166,6 @@ template<typename MatrixType>
 void svd_min_norm(const MatrixType& m, unsigned int computationOptions)
 {
   typedef typename MatrixType::Scalar Scalar;
-  typedef typename MatrixType::Index Index;
   Index cols = m.cols();
 
   enum {
@@ -261,7 +258,6 @@ void svd_test_all_computation_options(const MatrixType& m, bool full_only)
     CALL_SUBTEST(( svd_min_norm(m, ComputeThinU | ComputeThinV) ));
 
     // test reconstruction
-    typedef typename MatrixType::Index Index;
     Index diagSize = (std::min)(m.rows(), m.cols());
     SvdType svd(m, ComputeThinU | ComputeThinV);
     VERIFY_IS_APPROX(m, svd.matrixU().leftCols(diagSize) * svd.singularValues().asDiagonal() * svd.matrixV().leftCols(diagSize).adjoint());
@@ -437,7 +433,6 @@ template<typename SvdType,typename MatrixType>
 void svd_verify_assert(const MatrixType& m)
 {
   typedef typename MatrixType::Scalar Scalar;
-  typedef typename MatrixType::Index Index;
   Index rows = m.rows();
   Index cols = m.cols();
 
diff --git a/test/svd_fill.h b/test/svd_fill.h
index 3877c0c7e..d68647e99 100644
--- a/test/svd_fill.h
+++ b/test/svd_fill.h
@@ -23,7 +23,6 @@ void svd_fill_random(MatrixType &m, int Option = 0)
   using std::pow;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
-  typedef typename MatrixType::Index Index;
   Index diagSize = (std::min)(m.rows(), m.cols());
   RealScalar s = std::numeric_limits<RealScalar>::max_exponent10/4;
   s = internal::random<RealScalar>(1,s);
diff --git a/test/swap.cpp b/test/swap.cpp
index f76e3624d..a294da719 100644
--- a/test/swap.cpp
+++ b/test/swap.cpp
@@ -28,8 +28,8 @@ template<typename MatrixType> void swap(const MatrixType& m)
   typedef typename MatrixType::Scalar Scalar;
 
   eigen_assert((!internal::is_same<MatrixType,OtherMatrixType>::value));
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
   
   // construct 3 matrix guaranteed to be distinct
   MatrixType m1 = MatrixType::Random(rows,cols);
diff --git a/test/triangular.cpp b/test/triangular.cpp
index b96856486..d50342ccc 100644
--- a/test/triangular.cpp
+++ b/test/triangular.cpp
@@ -19,8 +19,8 @@ template<typename MatrixType> void triangular_square(const MatrixType& m)
 
   RealScalar largerEps = 10*test_precision<RealScalar>();
 
-  typename MatrixType::Index rows = m.rows();
-  typename MatrixType::Index cols = m.cols();
+  Index rows = m.rows();
+  Index cols = m.cols();
 
   MatrixType m1 = MatrixType::Random(rows, cols),
              m2 = MatrixType::Random(rows, cols),
@@ -134,7 +134,6 @@ template<typename MatrixType> void triangular_square(const MatrixType& m)
 
 template<typename MatrixType> void triangular_rect(const MatrixType& m)
 {
-  typedef const typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   enum { Rows =  MatrixType::RowsAtCompileTime, Cols =  MatrixType::ColsAtCompileTime };
diff --git a/test/upperbidiagonalization.cpp b/test/upperbidiagonalization.cpp
index 847b34b55..27a747875 100644
--- a/test/upperbidiagonalization.cpp
+++ b/test/upperbidiagonalization.cpp
@@ -12,8 +12,8 @@
 
 template<typename MatrixType> void upperbidiag(const MatrixType& m)
 {
-  const typename MatrixType::Index rows = m.rows();
-  const typename MatrixType::Index cols = m.cols();
+  const Index rows = m.rows();
+  const Index cols = m.cols();
 
   typedef Matrix<typename MatrixType::RealScalar, MatrixType::RowsAtCompileTime,  MatrixType::ColsAtCompileTime> RealMatrixType;
   typedef Matrix<typename MatrixType::Scalar, MatrixType::ColsAtCompileTime,  MatrixType::RowsAtCompileTime> TransposeMatrixType;
diff --git a/test/vectorwiseop.cpp b/test/vectorwiseop.cpp
index f3ab561ee..a099d17c8 100644
--- a/test/vectorwiseop.cpp
+++ b/test/vectorwiseop.cpp
@@ -15,7 +15,6 @@
 
 template<typename ArrayType> void vectorwiseop_array(const ArrayType& m)
 {
-  typedef typename ArrayType::Index Index;
   typedef typename ArrayType::Scalar Scalar;
   typedef Array<Scalar, ArrayType::RowsAtCompileTime, 1> ColVectorType;
   typedef Array<Scalar, 1, ArrayType::ColsAtCompileTime> RowVectorType;
@@ -129,7 +128,6 @@ template<typename ArrayType> void vectorwiseop_array(const ArrayType& m)
 
 template<typename MatrixType> void vectorwiseop_matrix(const MatrixType& m)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, 1> ColVectorType;
diff --git a/test/visitor.cpp b/test/visitor.cpp
index 844170ec6..7f4efab97 100644
--- a/test/visitor.cpp
+++ b/test/visitor.cpp
@@ -12,7 +12,6 @@
 template<typename MatrixType> void matrixVisitor(const MatrixType& p)
 {
   typedef typename MatrixType::Scalar Scalar;
-  typedef typename MatrixType::Index Index;
 
   Index rows = p.rows();
   Index cols = p.cols();
@@ -65,7 +64,6 @@ template<typename MatrixType> void matrixVisitor(const MatrixType& p)
 template<typename VectorType> void vectorVisitor(const VectorType& w)
 {
   typedef typename VectorType::Scalar Scalar;
-  typedef typename VectorType::Index Index;
 
   Index size = w.size();
 

From 44ea5f7623d1d5211de753b4ea719b63d55b228a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 12 Jul 2018 17:19:38 +0200
Subject: [PATCH 445/680] Add unit test for -Tensor<complex> on GPU

---
 unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
index af67348aa..3b078d1d1 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
@@ -48,11 +48,13 @@ void test_cuda_complex_cwise_ops() {
     Add = 0,
     Sub,
     Mul,
-    Div
+    Div,
+    Neg,
+    NbOps
   };
 
   Tensor<std::complex<T>, 1, 0, int> actual(kNumItems);
-  for (int op = Add; op <= Div; op++) {
+  for (int op = Add; op < NbOps; op++) {
     std::complex<T> expected;
     switch (static_cast<CwiseOp>(op)) {
       case Add:
@@ -71,6 +73,10 @@ void test_cuda_complex_cwise_ops() {
         gpu_out.device(gpu_device) = gpu_in1 / gpu_in2;
         expected = a / b;
         break;
+      case Neg:
+        gpu_out.device(gpu_device) = -gpu_in1;
+        expected = -a;
+        break;
     }
     assert(cudaMemcpyAsync(actual.data(), d_out, complex_bytes, cudaMemcpyDeviceToHost,
                            gpu_device.stream()) == cudaSuccess);

From 3875fb05aaa019c700d68b074e61471a96c85b36 Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Fri, 6 Jul 2018 16:04:30 -0700
Subject: [PATCH 446/680] Add support for MIPS SIMD (MSA)

---
 CMakeLists.txt                               |    6 +
 Eigen/src/Core/arch/MSA/Complex.h            |  759 ++++++++++
 Eigen/src/Core/arch/MSA/MathFunctions.h      |  387 +++++
 Eigen/src/Core/arch/MSA/PacketMath.h         | 1317 ++++++++++++++++++
 Eigen/src/Core/util/ConfigureVectorization.h |    2 +-
 Eigen/src/Core/util/Constants.h              |    3 +
 cmake/EigenTesting.cmake                     |    8 +
 7 files changed, 2481 insertions(+), 1 deletion(-)
 create mode 100644 Eigen/src/Core/arch/MSA/Complex.h
 create mode 100644 Eigen/src/Core/arch/MSA/MathFunctions.h
 create mode 100644 Eigen/src/Core/arch/MSA/PacketMath.h

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 91545bdb0..0547ee681 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -256,6 +256,12 @@ if(NOT MSVC)
     message(STATUS "Enabling VSX in tests/examples")
   endif()
 
+  option(EIGEN_TEST_MSA "Enable/Disable MSA in tests/examples" OFF)
+  if(EIGEN_TEST_MSA)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mmsa")
+    message(STATUS "Enabling MSA in tests/examples")
+  endif()
+
   option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF)
   if(EIGEN_TEST_NEON)
     if(EIGEN_TEST_FMA)
diff --git a/Eigen/src/Core/arch/MSA/Complex.h b/Eigen/src/Core/arch/MSA/Complex.h
new file mode 100644
index 000000000..9a45cf51e
--- /dev/null
+++ b/Eigen/src/Core/arch/MSA/Complex.h
@@ -0,0 +1,759 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Wave Computing, Inc.
+// Written by:
+//   Chris Larsen
+//   Alexey Frunze (afrunze@wavecomp.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_MSA_H
+#define EIGEN_COMPLEX_MSA_H
+
+#include <iostream>
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- float ----------
+struct Packet2cf {
+  EIGEN_STRONG_INLINE Packet2cf() {
+  }
+  EIGEN_STRONG_INLINE explicit Packet2cf(const std::complex<float>& a,
+                                         const std::complex<float>& b) {
+    Packet4f t = { std::real(a), std::imag(a), std::real(b), std::imag(b) };
+    v = t;
+  }
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {
+  }
+  EIGEN_STRONG_INLINE Packet2cf(const Packet2cf& a) : v(a.v) {
+  }
+  EIGEN_STRONG_INLINE Packet2cf& operator=(const Packet2cf& b) {
+    v = b.v;
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet2cf conjugate(void) const {
+    return Packet2cf((Packet4f)__builtin_msa_bnegi_d((v2u64)v, 63));
+  }
+  EIGEN_STRONG_INLINE Packet2cf& operator*=(const Packet2cf& b) {
+    Packet4f v1, v2;
+
+    // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
+    v1 = (Packet4f)__builtin_msa_ilvev_w((v4i32)v, (v4i32)v);
+    // Get the imag values of a | a1_im | a1_im | a2_im | a2_im |
+    v2 = (Packet4f)__builtin_msa_ilvod_w((v4i32)v, (v4i32)v);
+    // Multiply the real a with b
+    v1 = pmul(v1, b.v);
+    // Multiply the imag a with b
+    v2 = pmul(v2, b.v);
+    // Conjugate v2
+    v2 = Packet2cf(v2).conjugate().v;
+    // Swap real/imag elements in v2.
+    v2 = (Packet4f)__builtin_msa_shf_w((v4i32)v2, EIGEN_MSA_SHF_I8(1, 0, 3, 2));
+    // Add and return the result
+    v = padd(v1, v2);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet2cf operator*(const Packet2cf& b) const {
+    return Packet2cf(*this) *= b;
+  }
+  EIGEN_STRONG_INLINE Packet2cf& operator+=(const Packet2cf& b) {
+    v = padd(v, b.v);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet2cf operator+(const Packet2cf& b) const {
+    return Packet2cf(*this) += b;
+  }
+  EIGEN_STRONG_INLINE Packet2cf& operator-=(const Packet2cf& b) {
+    v = psub(v, b.v);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet2cf operator-(const Packet2cf& b) const {
+    return Packet2cf(*this) -= b;
+  }
+  EIGEN_STRONG_INLINE Packet2cf& operator/=(const Packet2cf& b) {
+    *this *= b.conjugate();
+    Packet4f s = pmul<Packet4f>(b.v, b.v);
+    s = padd(s, (Packet4f)__builtin_msa_shf_w((v4i32)s, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+    v = pdiv(v, s);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet2cf operator/(const Packet2cf& b) const {
+    return Packet2cf(*this) /= b;
+  }
+  EIGEN_STRONG_INLINE Packet2cf operator-(void) const {
+    return Packet2cf(pnegate(v));
+  }
+
+  Packet4f v;
+};
+
+inline std::ostream& operator<<(std::ostream& os, const Packet2cf& value) {
+  os << "[ (" << value.v[0] << ", " << value.v[1]
+     << "i),"
+        "  ("
+     << value.v[2] << ", " << value.v[3] << "i) ]";
+  return os;
+}
+
+template <>
+struct packet_traits<std::complex<float> > : default_packet_traits {
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd = 1,
+    HasSub = 1,
+    HasMul = 1,
+    HasDiv = 1,
+    HasNegate = 1,
+    HasAbs = 0,
+    HasAbs2 = 0,
+    HasMin = 0,
+    HasMax = 0,
+    HasSetLinear = 0,
+    HasBlend = 1
+  };
+};
+
+template <>
+struct unpacket_traits<Packet2cf> {
+  typedef std::complex<float> type;
+  enum { size = 2, alignment = Aligned16 };
+  typedef Packet2cf half;
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) {
+  EIGEN_MSA_DEBUG;
+
+  float f0 = from.real(), f1 = from.imag();
+  Packet4f v0 = { f0, f0, f0, f0 };
+  Packet4f v1 = { f1, f1, f1, f1 };
+  return Packet2cf((Packet4f)__builtin_msa_ilvr_w((Packet4i)v1, (Packet4i)v0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a + b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a - b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return -a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a.conjugate();
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a * b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf(pand(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf(por(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf(pxor(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf(pandnot(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) {
+  EIGEN_MSA_DEBUG;
+
+  return pset1<Packet2cf>(*from);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<std::complex<float> >(std::complex<float>* to,
+                                                      const Packet2cf& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_STORE pstore<float>((float*)to, from.v);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to,
+                                                       const Packet2cf& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_STORE pstoreu<float>((float*)to, from.v);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(
+    const std::complex<float>* from, Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf(from[0 * stride], from[1 * stride]);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to,
+                                                                       const Packet2cf& from,
+                                                                       Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  *to = std::complex<float>(from.v[0], from.v[1]);
+  to += stride;
+  *to = std::complex<float>(from.v[2], from.v[3]);
+}
+
+template <>
+EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float>* addr) {
+  EIGEN_MSA_DEBUG;
+
+  prefetch(reinterpret_cast<const float*>(addr));
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return std::complex<float>(a.v[0], a.v[1]);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf((Packet4f)__builtin_msa_shf_w((v4i32)a.v, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet2cf((Packet4f)__builtin_msa_shf_w((v4i32)a.v, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f value = (Packet4f)preverse((Packet2d)a.v);
+  value += a.v;
+  return std::complex<float>(value[0], value[1]);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f sum1, sum2, sum;
+
+  // Add the first two 64-bit float32x2_t of vecs[0]
+  sum1 = (Packet4f)__builtin_msa_ilvr_d((v2i64)vecs[1].v, (v2i64)vecs[0].v);
+  sum2 = (Packet4f)__builtin_msa_ilvl_d((v2i64)vecs[1].v, (v2i64)vecs[0].v);
+  sum = padd(sum1, sum2);
+
+  return Packet2cf(sum);
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) {
+  EIGEN_MSA_DEBUG;
+
+  return std::complex<float>((a.v[0] * a.v[2]) - (a.v[1] * a.v[3]),
+                             (a.v[0] * a.v[3]) + (a.v[1] * a.v[2]));
+}
+
+template <int Offset>
+struct palign_impl<Offset, Packet2cf> {
+  EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second) {
+    if (Offset == 1) {
+      first.v = (Packet4f)__builtin_msa_sldi_b((v16i8)second.v, (v16i8)first.v, Offset * 8);
+    }
+  }
+};
+
+template <>
+struct conj_helper<Packet2cf, Packet2cf, false, true> {
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y,
+                                      const Packet2cf& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template <>
+struct conj_helper<Packet2cf, Packet2cf, true, false> {
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y,
+                                      const Packet2cf& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template <>
+struct conj_helper<Packet2cf, Packet2cf, true, true> {
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y,
+                                      const Packet2cf& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a / b;
+}
+
+inline std::ostream& operator<<(std::ostream& os, const PacketBlock<Packet2cf, 2>& value) {
+  os << "[ " << value.packet[0] << ", " << std::endl << "  " << value.packet[1] << " ]";
+  return os;
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f tmp =
+      (Packet4f)__builtin_msa_ilvl_d((v2i64)kernel.packet[1].v, (v2i64)kernel.packet[0].v);
+  kernel.packet[0].v =
+      (Packet4f)__builtin_msa_ilvr_d((v2i64)kernel.packet[1].v, (v2i64)kernel.packet[0].v);
+  kernel.packet[1].v = tmp;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket,
+                                     const Packet2cf& elsePacket) {
+  return (Packet2cf)(Packet4f)pblend<Packet2d>(ifPacket, (Packet2d)thenPacket.v,
+                                               (Packet2d)elsePacket.v);
+}
+
+//---------- double ----------
+
+struct Packet1cd {
+  EIGEN_STRONG_INLINE Packet1cd() {
+  }
+  EIGEN_STRONG_INLINE explicit Packet1cd(const std::complex<double>& a) {
+    v[0] = std::real(a);
+    v[1] = std::imag(a);
+  }
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {
+  }
+  EIGEN_STRONG_INLINE Packet1cd(const Packet1cd& a) : v(a.v) {
+  }
+  EIGEN_STRONG_INLINE Packet1cd& operator=(const Packet1cd& b) {
+    v = b.v;
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet1cd conjugate(void) const {
+    static const v2u64 p2ul_CONJ_XOR = { 0x0, 0x8000000000000000 };
+    return (Packet1cd)pxor(v, (Packet2d)p2ul_CONJ_XOR);
+  }
+  EIGEN_STRONG_INLINE Packet1cd& operator*=(const Packet1cd& b) {
+    Packet2d v1, v2;
+
+    // Get the real values of a | a1_re | a1_re
+    v1 = (Packet2d)__builtin_msa_ilvev_d((v2i64)v, (v2i64)v);
+    // Get the imag values of a | a1_im | a1_im
+    v2 = (Packet2d)__builtin_msa_ilvod_d((v2i64)v, (v2i64)v);
+    // Multiply the real a with b
+    v1 = pmul(v1, b.v);
+    // Multiply the imag a with b
+    v2 = pmul(v2, b.v);
+    // Conjugate v2
+    v2 = Packet1cd(v2).conjugate().v;
+    // Swap real/imag elements in v2.
+    v2 = (Packet2d)__builtin_msa_shf_w((v4i32)v2, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+    // Add and return the result
+    v = padd(v1, v2);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet1cd operator*(const Packet1cd& b) const {
+    return Packet1cd(*this) *= b;
+  }
+  EIGEN_STRONG_INLINE Packet1cd& operator+=(const Packet1cd& b) {
+    v = padd(v, b.v);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet1cd operator+(const Packet1cd& b) const {
+    return Packet1cd(*this) += b;
+  }
+  EIGEN_STRONG_INLINE Packet1cd& operator-=(const Packet1cd& b) {
+    v = psub(v, b.v);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet1cd operator-(const Packet1cd& b) const {
+    return Packet1cd(*this) -= b;
+  }
+  EIGEN_STRONG_INLINE Packet1cd& operator/=(const Packet1cd& b) {
+    *this *= b.conjugate();
+    Packet2d s = pmul<Packet2d>(b.v, b.v);
+    s = padd(s, preverse<Packet2d>(s));
+    v = pdiv(v, s);
+    return *this;
+  }
+  EIGEN_STRONG_INLINE Packet1cd operator/(const Packet1cd& b) const {
+    return Packet1cd(*this) /= b;
+  }
+  EIGEN_STRONG_INLINE Packet1cd operator-(void) const {
+    return Packet1cd(pnegate(v));
+  }
+
+  Packet2d v;
+};
+
+inline std::ostream& operator<<(std::ostream& os, const Packet1cd& value) {
+  os << "[ (" << value.v[0] << ", " << value.v[1] << "i) ]";
+  return os;
+}
+
+template <>
+struct packet_traits<std::complex<double> > : default_packet_traits {
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd = 1,
+    HasSub = 1,
+    HasMul = 1,
+    HasDiv = 1,
+    HasNegate = 1,
+    HasAbs = 0,
+    HasAbs2 = 0,
+    HasMin = 0,
+    HasMax = 0,
+    HasSetLinear = 0
+  };
+};
+
+template <>
+struct unpacket_traits<Packet1cd> {
+  typedef std::complex<double> type;
+  enum { size = 1, alignment = Aligned16 };
+  typedef Packet1cd half;
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(from);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a + b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a - b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return -a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a.conjugate();
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a * b;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(pand(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(por(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(pxor(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(pandnot(a.v, b.v));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) {
+  EIGEN_MSA_DEBUG;
+
+  return pset1<Packet1cd>(*from);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<std::complex<double> >(std::complex<double>* to,
+                                                       const Packet1cd& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_STORE pstore<double>((double*)to, from.v);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double>* to,
+                                                        const Packet1cd& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_STORE pstoreu<double>((double*)to, from.v);
+}
+
+template <>
+EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double>* addr) {
+  EIGEN_MSA_DEBUG;
+
+  prefetch(reinterpret_cast<const double*>(addr));
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(
+    const std::complex<double>* from, Index stride __attribute__((unused))) {
+  EIGEN_MSA_DEBUG;
+
+  Packet1cd res;
+  res.v[0] = std::real(from[0]);
+  res.v[1] = std::imag(from[0]);
+  return res;
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to,
+                                                                        const Packet1cd& from,
+                                                                        Index stride
+                                                                        __attribute__((unused))) {
+  EIGEN_MSA_DEBUG;
+
+  pstore(to, from);
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return std::complex<double>(a.v[0], a.v[1]);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return pfirst(a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) {
+  EIGEN_MSA_DEBUG;
+
+  return vecs[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) {
+  EIGEN_MSA_DEBUG;
+
+  return pfirst(a);
+}
+
+template <int Offset>
+struct palign_impl<Offset, Packet1cd> {
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes
+    // boundary...
+  }
+};
+
+template <>
+struct conj_helper<Packet1cd, Packet1cd, false, true> {
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y,
+                                      const Packet1cd& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template <>
+struct conj_helper<Packet1cd, Packet1cd, true, false> {
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y,
+                                      const Packet1cd& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template <>
+struct conj_helper<Packet1cd, Packet1cd, true, true> {
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y,
+                                      const Packet1cd& c) const {
+    return padd(pmul(x, y), c);
+  }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  EIGEN_MSA_DEBUG;
+
+  return a / b;
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip /*<Packet1cd>*/ (const Packet1cd& x) {
+  EIGEN_MSA_DEBUG;
+
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+inline std::ostream& operator<<(std::ostream& os, const PacketBlock<Packet1cd, 2>& value) {
+  os << "[ " << value.packet[0] << ", " << std::endl << "  " << value.packet[1] << " ]";
+  return os;
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd, 2>& kernel) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d v1, v2;
+
+  v1 = (Packet2d)__builtin_msa_ilvev_d((v2i64)kernel.packet[0].v, (v2i64)kernel.packet[1].v);
+  // Get the imag values of a
+  v2 = (Packet2d)__builtin_msa_ilvod_d((v2i64)kernel.packet[0].v, (v2i64)kernel.packet[1].v);
+
+  kernel.packet[0].v = v1;
+  kernel.packet[1].v = v2;
+}
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_COMPLEX_MSA_H
diff --git a/Eigen/src/Core/arch/MSA/MathFunctions.h b/Eigen/src/Core/arch/MSA/MathFunctions.h
new file mode 100644
index 000000000..98e23e36f
--- /dev/null
+++ b/Eigen/src/Core/arch/MSA/MathFunctions.h
@@ -0,0 +1,387 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// Copyright (C) 2018 Wave Computing, Inc.
+// Written by:
+//   Chris Larsen
+//   Alexey Frunze (afrunze@wavecomp.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+/* The tanh function of this file is an adaptation of
+ * template<typename T> T generic_fast_tanh_float(const T&)
+ * from MathFunctionsImpl.h.
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_MSA_H
+#define EIGEN_MATH_FUNCTIONS_MSA_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+plog<Packet4f>(const Packet4f& _x) {
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292e-2f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+  static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+
+  // Convert negative argument into NAN (quiet negative, to be specific).
+  Packet4f zero = (Packet4f)__builtin_msa_ldi_w(0);
+  Packet4i neg_mask = __builtin_msa_fclt_w(_x, zero);
+  Packet4i zero_mask = __builtin_msa_fceq_w(_x, zero);
+  Packet4f non_neg_x_or_nan = padd(_x, (Packet4f)neg_mask);  // Add 0.0 or NAN.
+  Packet4f x = non_neg_x_or_nan;
+
+  // Extract exponent from x = mantissa * 2**exponent, where 1.0 <= mantissa < 2.0.
+  // N.B. the exponent is one less of what frexpf() would return.
+  Packet4i e_int = __builtin_msa_ftint_s_w(__builtin_msa_flog2_w(x));
+  // Multiply x by 2**(-exponent-1) to get 0.5 <= x < 1.0 as from frexpf().
+  x = __builtin_msa_fexp2_w(x, (Packet4i)__builtin_msa_nori_b((v16u8)e_int, 0));
+
+  /*
+     if (x < SQRTHF) {
+       x = x + x - 1.0;
+     } else {
+       e += 1;
+       x = x - 1.0;
+     }
+  */
+  Packet4f xx = padd(x, x);
+  Packet4i ge_mask = __builtin_msa_fcle_w(p4f_cephes_SQRTHF, x);
+  e_int = psub(e_int, ge_mask);
+  x = (Packet4f)__builtin_msa_bsel_v((v16u8)ge_mask, (v16u8)xx, (v16u8)x);
+  x = psub(x, p4f_1);
+  Packet4f e = __builtin_msa_ffint_s_w(e_int);
+
+  Packet4f x2 = pmul(x, x);
+  Packet4f x3 = pmul(x2, x);
+
+  Packet4f y, y1, y2;
+  y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
+  y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
+  y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
+  y = pmadd(y, x, p4f_cephes_log_p2);
+  y1 = pmadd(y1, x, p4f_cephes_log_p5);
+  y2 = pmadd(y2, x, p4f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  y = pmadd(e, p4f_cephes_log_q1, y);
+  x = __builtin_msa_fmsub_w(x, x2, p4f_half);
+  x = padd(x, y);
+  x = pmadd(e, p4f_cephes_log_q2, x);
+
+  // x is now the logarithm result candidate. We still need to handle the
+  // extreme arguments of zero and positive infinity, though.
+  // N.B. if the argument is +INFINITY, x is NAN because the polynomial terms
+  // contain infinities of both signs (see the coefficients and code above).
+  // INFINITY - INFINITY is NAN.
+
+  // If the argument is +INFINITY, make it the new result candidate.
+  // To achieve that we choose the smaller of the result candidate and the
+  // argument.
+  // This is correct for all finite pairs of values (the logarithm is smaller
+  // than the argument).
+  // This is also correct in the special case when the argument is +INFINITY
+  // and the result candidate is NAN. This is because the fmin.df instruction
+  // prefers non-NANs to NANs.
+  x = __builtin_msa_fmin_w(x, non_neg_x_or_nan);
+
+  // If the argument is zero (including -0.0), the result becomes -INFINITY.
+  Packet4i neg_infs = __builtin_msa_slli_w(zero_mask, 23);
+  x = (Packet4f)__builtin_msa_bsel_v((v16u8)zero_mask, (v16u8)x, (v16u8)neg_infs);
+
+  return x;
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+pexp<Packet4f>(const Packet4f& _x) {
+  // Limiting single-precision pexp's argument to [-128, +128] lets pexp
+  // reach 0 and INFINITY naturally.
+  static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -128.0f);
+  static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, +128.0f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894e-2f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+
+  Packet4f x = _x;
+
+  // Clamp x.
+  x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(x, p4f_exp_lo), (v16u8)x,
+                                     (v16u8)p4f_exp_lo);
+  x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_exp_hi, x), (v16u8)x,
+                                     (v16u8)p4f_exp_hi);
+
+  // Round to nearest integer by adding 0.5 (with x's sign) and truncating.
+  Packet4f x2_add = (Packet4f)__builtin_msa_binsli_w((v4u32)p4f_half, (v4u32)x, 0);
+  Packet4f x2 = pmadd(x, p4f_cephes_LOG2EF, x2_add);
+  Packet4i x2_int = __builtin_msa_ftrunc_s_w(x2);
+  Packet4f x2_int_f = __builtin_msa_ffint_s_w(x2_int);
+
+  x = __builtin_msa_fmsub_w(x, x2_int_f, p4f_cephes_exp_C1);
+  x = __builtin_msa_fmsub_w(x, x2_int_f, p4f_cephes_exp_C2);
+
+  Packet4f z = pmul(x, x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // y *= 2**exponent.
+  y = __builtin_msa_fexp2_w(y, x2_int);
+
+  return y;
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+ptanh<Packet4f>(const Packet4f& _x) {
+  static _EIGEN_DECLARE_CONST_Packet4f(tanh_tiny, 1e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(tanh_hi, 9.0f);
+  // The monomial coefficients of the numerator polynomial (odd).
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-5f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-8f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
+  static _EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
+  // The monomial coefficients of the denominator polynomial (even).
+  static _EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-6f);
+
+  Packet4f x = pabs(_x);
+  Packet4i tiny_mask = __builtin_msa_fclt_w(x, p4f_tanh_tiny);
+
+  // Clamp the inputs to the range [-9, 9] since anything outside
+  // this range is -/+1.0f in single-precision.
+  x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_tanh_hi, x), (v16u8)x,
+                                     (v16u8)p4f_tanh_hi);
+
+  // Since the polynomials are odd/even, we need x**2.
+  Packet4f x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial p.
+  Packet4f p = pmadd(x2, p4f_alpha_13, p4f_alpha_11);
+  p = pmadd(x2, p, p4f_alpha_9);
+  p = pmadd(x2, p, p4f_alpha_7);
+  p = pmadd(x2, p, p4f_alpha_5);
+  p = pmadd(x2, p, p4f_alpha_3);
+  p = pmadd(x2, p, p4f_alpha_1);
+  p = pmul(x, p);
+
+  // Evaluate the denominator polynomial q.
+  Packet4f q = pmadd(x2, p4f_beta_6, p4f_beta_4);
+  q = pmadd(x2, q, p4f_beta_2);
+  q = pmadd(x2, q, p4f_beta_0);
+
+  // Divide the numerator by the denominator.
+  p = pdiv(p, q);
+
+  // Reinstate the sign.
+  p = (Packet4f)__builtin_msa_binsli_w((v4u32)p, (v4u32)_x, 0);
+
+  // When the argument is very small in magnitude it's more accurate to just return it.
+  p = (Packet4f)__builtin_msa_bsel_v((v16u8)tiny_mask, (v16u8)p, (v16u8)_x);
+
+  return p;
+}
+
+template <bool sine>
+Packet4f psincos_inner_msa_float(const Packet4f& _x) {
+  static _EIGEN_DECLARE_CONST_Packet4f(sincos_max_arg, 13176795.0f);  // Approx. (2**24) / (4/Pi).
+  static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1, -0.78515625f);
+  static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
+  static _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891e-4f);
+  static _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611e-1f);
+  static _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948e-5f);
+  static _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765e-3f);
+  static _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827e-2f);
+  static _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f);  // 4/Pi.
+  static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+
+  Packet4f x = pabs(_x);
+
+  // Translate infinite arguments into NANs.
+  Packet4f zero_or_nan_if_inf = psub(_x, _x);
+  x = padd(x, zero_or_nan_if_inf);
+  // Prevent sin/cos from generating values larger than 1.0 in magnitude
+  // for very large arguments by setting x to 0.0.
+  Packet4i small_or_nan_mask = __builtin_msa_fcult_w(x, p4f_sincos_max_arg);
+  x = pand(x, (Packet4f)small_or_nan_mask);
+
+  // Scale x by 4/Pi to find x's octant.
+  Packet4f y = pmul(x, p4f_cephes_FOPI);
+  // Get the octant. We'll reduce x by this number of octants or by one more than it.
+  Packet4i y_int = __builtin_msa_ftrunc_s_w(y);
+  // x's from even-numbered octants will translate to octant 0: [0, +Pi/4].
+  // x's from odd-numbered octants will translate to octant -1: [-Pi/4, 0].
+  // Adjustment for odd-numbered octants: octant = (octant + 1) & (~1).
+  Packet4i y_int1 = __builtin_msa_addvi_w(y_int, 1);
+  Packet4i y_int2 = (Packet4i)__builtin_msa_bclri_w((Packet4ui)y_int1, 0);
+  y = __builtin_msa_ffint_s_w(y_int2);
+
+  // Compute the sign to apply to the polynomial.
+  Packet4i sign_mask = sine ? pxor(__builtin_msa_slli_w(y_int1, 29), (Packet4i)_x)
+                            : __builtin_msa_slli_w(__builtin_msa_addvi_w(y_int, 3), 29);
+
+  // Get the polynomial selection mask.
+  // We'll calculate both (sin and cos) polynomials and then select from the two.
+  Packet4i poly_mask = __builtin_msa_ceqi_w(__builtin_msa_slli_w(y_int2, 30), 0);
+
+  // Reduce x by y octants to get: -Pi/4 <= x <= +Pi/4.
+  // The magic pass: "Extended precision modular arithmetic"
+  // x = ((x - y * DP1) - y * DP2) - y * DP3
+  Packet4f tmp1 = pmul(y, p4f_minus_cephes_DP1);
+  Packet4f tmp2 = pmul(y, p4f_minus_cephes_DP2);
+  Packet4f tmp3 = pmul(y, p4f_minus_cephes_DP3);
+  x = padd(x, tmp1);
+  x = padd(x, tmp2);
+  x = padd(x, tmp3);
+
+  // Evaluate the cos(x) polynomial.
+  y = p4f_coscof_p0;
+  Packet4f z = pmul(x, x);
+  y = pmadd(y, z, p4f_coscof_p1);
+  y = pmadd(y, z, p4f_coscof_p2);
+  y = pmul(y, z);
+  y = pmul(y, z);
+  y = __builtin_msa_fmsub_w(y, z, p4f_half);
+  y = padd(y, p4f_1);
+
+  // Evaluate the sin(x) polynomial.
+  Packet4f y2 = p4f_sincof_p0;
+  y2 = pmadd(y2, z, p4f_sincof_p1);
+  y2 = pmadd(y2, z, p4f_sincof_p2);
+  y2 = pmul(y2, z);
+  y2 = pmadd(y2, x, x);
+
+  // Select the correct result from the two polynomials.
+  y = sine ? (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y, (v16u8)y2)
+           : (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y2, (v16u8)y);
+
+  // Update the sign.
+  sign_mask = pxor(sign_mask, (Packet4i)y);
+  y = (Packet4f)__builtin_msa_binsli_w((v4u32)y, (v4u32)sign_mask, 0);
+  return y;
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+psin<Packet4f>(const Packet4f& x) {
+  return psincos_inner_msa_float</* sine */ true>(x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+pcos<Packet4f>(const Packet4f& x) {
+  return psincos_inner_msa_float</* sine */ false>(x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d
+pexp<Packet2d>(const Packet2d& _x) {
+  // Limiting double-precision pexp's argument to [-1024, +1024] lets pexp
+  // reach 0 and INFINITY naturally.
+  static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -1024.0);
+  static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, +1024.0);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+  static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+  static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+  static _EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
+  static _EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
+
+  Packet2d x = _x;
+
+  // Clamp x.
+  x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(x, p2d_exp_lo), (v16u8)x,
+                                     (v16u8)p2d_exp_lo);
+  x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(p2d_exp_hi, x), (v16u8)x,
+                                     (v16u8)p2d_exp_hi);
+
+  // Round to nearest integer by adding 0.5 (with x's sign) and truncating.
+  Packet2d x2_add = (Packet2d)__builtin_msa_binsli_d((v2u64)p2d_half, (v2u64)x, 0);
+  Packet2d x2 = pmadd(x, p2d_cephes_LOG2EF, x2_add);
+  Packet2l x2_long = __builtin_msa_ftrunc_s_d(x2);
+  Packet2d x2_long_d = __builtin_msa_ffint_s_d(x2_long);
+
+  x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C1);
+  x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C2);
+
+  x2 = pmul(x, x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px, psub(qx, px));
+  x = pmadd(p2d_2, x, p2d_1);
+
+  // x *= 2**exponent.
+  x = __builtin_msa_fexp2_d(x, x2_long);
+
+  return x;
+}
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_MSA_H
diff --git a/Eigen/src/Core/arch/MSA/PacketMath.h b/Eigen/src/Core/arch/MSA/PacketMath.h
new file mode 100644
index 000000000..94c15d132
--- /dev/null
+++ b/Eigen/src/Core/arch/MSA/PacketMath.h
@@ -0,0 +1,1317 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Wave Computing, Inc.
+// Written by:
+//   Chris Larsen
+//   Alexey Frunze (afrunze@wavecomp.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_MSA_H
+#define EIGEN_PACKET_MATH_MSA_H
+
+#include <iostream>
+#include <string>
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32
+#endif
+
+#if 0
+#define EIGEN_MSA_DEBUG                                                             \
+  static bool firstTime = true;                                                     \
+  do {                                                                              \
+    if (firstTime) {                                                                \
+      std::cout << __FILE__ << ':' << __LINE__ << ':' << __FUNCTION__ << std::endl; \
+      firstTime = false;                                                            \
+    }                                                                               \
+  } while (0)
+#else
+#define EIGEN_MSA_DEBUG
+#endif
+
+#define EIGEN_MSA_SHF_I8(a, b, c, d) (((d) << 6) | ((c) << 4) | ((b) << 2) | (a))
+
+typedef v4f32 Packet4f;
+typedef v4i32 Packet4i;
+typedef v4u32 Packet4ui;
+
+#define _EIGEN_DECLARE_CONST_Packet4f(NAME, X) const Packet4f p4f_##NAME = { X, X, X, X }
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME, X) const Packet4i p4i_##NAME = { X, X, X, X }
+#define _EIGEN_DECLARE_CONST_Packet4ui(NAME, X) const Packet4ui p4ui_##NAME = { X, X, X, X }
+
+inline std::ostream& operator<<(std::ostream& os, const Packet4f& value) {
+  os << "[ " << value[0] << ", " << value[1] << ", " << value[2] << ", " << value[3] << " ]";
+  return os;
+}
+
+inline std::ostream& operator<<(std::ostream& os, const Packet4i& value) {
+  os << "[ " << value[0] << ", " << value[1] << ", " << value[2] << ", " << value[3] << " ]";
+  return os;
+}
+
+inline std::ostream& operator<<(std::ostream& os, const Packet4ui& value) {
+  os << "[ " << value[0] << ", " << value[1] << ", " << value[2] << ", " << value[3] << " ]";
+  return os;
+}
+
+template <>
+struct packet_traits<float> : default_packet_traits {
+  typedef Packet4f type;
+  typedef Packet4f half;  // Packet2f intrinsics not implemented yet
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,  // Packet2f intrinsics not implemented yet
+    // FIXME check the Has*
+    HasDiv = 1,
+    HasSin = EIGEN_FAST_MATH,
+    HasCos = EIGEN_FAST_MATH,
+    HasTanh = EIGEN_FAST_MATH,
+    HasLog = 1,
+    HasExp = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasBlend = 1
+  };
+};
+
+template <>
+struct packet_traits<int32_t> : default_packet_traits {
+  typedef Packet4i type;
+  typedef Packet4i half;  // Packet2i intrinsics not implemented yet
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,  // Packet2i intrinsics not implemented yet
+    // FIXME check the Has*
+    HasDiv = 1,
+    HasBlend = 1
+  };
+};
+
+template <>
+struct unpacket_traits<Packet4f> {
+  typedef float type;
+  enum { size = 4, alignment = Aligned16 };
+  typedef Packet4f half;
+};
+
+template <>
+struct unpacket_traits<Packet4i> {
+  typedef int32_t type;
+  enum { size = 4, alignment = Aligned16 };
+  typedef Packet4i half;
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f v = { from, from, from, from };
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t& from) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fill_w(from);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float* from) {
+  EIGEN_MSA_DEBUG;
+
+  float f = *from;
+  Packet4f v = { f, f, f, f };
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pload1<Packet4i>(const int32_t* from) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fill_w(*from);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fadd_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_addv_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) {
+  EIGEN_MSA_DEBUG;
+
+  static const Packet4f countdown = { 0.0f, 1.0f, 2.0f, 3.0f };
+  return padd(pset1<Packet4f>(a), countdown);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int32_t& a) {
+  EIGEN_MSA_DEBUG;
+
+  static const Packet4i countdown = { 0, 1, 2, 3 };
+  return padd(pset1<Packet4i>(a), countdown);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fsub_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_subv_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_bnegi_w((v4u32)a, 31);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_addvi_w((v4i32)__builtin_msa_nori_b((v16u8)a, 0), 1);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fmul_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_mulv_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fdiv_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_div_s_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fmadd_w(c, a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) {
+  EIGEN_MSA_DEBUG;
+
+  // Use "asm" construct to avoid __builtin_msa_maddv_w GNU C bug.
+  Packet4i value = c;
+  __asm__("maddv.w %w[value], %w[a], %w[b]\n"
+          // Outputs
+          : [value] "+f"(value)
+          // Inputs
+          : [a] "f"(a), [b] "f"(b));
+  return value;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_and_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4i)__builtin_msa_and_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_or_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4i)__builtin_msa_or_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_xor_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4i)__builtin_msa_xor_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+  return pand(a, (Packet4f)__builtin_msa_xori_b((v16u8)b, 255));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return pand(a, (Packet4i)__builtin_msa_xori_b((v16u8)b, 255));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  // This prefers numbers to NaNs.
+  return __builtin_msa_fmin_w(a, b);
+#else
+  // This prefers NaNs to numbers.
+  Packet4i aNaN = __builtin_msa_fcun_w(a, a);
+  Packet4i aMinOrNaN = por(__builtin_msa_fclt_w(a, b), aNaN);
+  return (Packet4f)__builtin_msa_bsel_v((v16u8)aMinOrNaN, (v16u8)b, (v16u8)a);
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_min_s_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  // This prefers numbers to NaNs.
+  return __builtin_msa_fmax_w(a, b);
+#else
+  // This prefers NaNs to numbers.
+  Packet4i aNaN = __builtin_msa_fcun_w(a, a);
+  Packet4i aMaxOrNaN = por(__builtin_msa_fclt_w(b, a), aNaN);
+  return (Packet4f)__builtin_msa_bsel_v((v16u8)aMaxOrNaN, (v16u8)b, (v16u8)a);
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_max_s_w(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_LOAD return (Packet4f)__builtin_msa_ld_w(const_cast<float*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int32_t* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_LOAD return __builtin_msa_ld_w(const_cast<int32_t*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return (Packet4f)__builtin_msa_ld_w(const_cast<float*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int32_t* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return (Packet4i)__builtin_msa_ld_w(const_cast<int32_t*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) {
+  EIGEN_MSA_DEBUG;
+
+  float f0 = from[0], f1 = from[1];
+  Packet4f v0 = { f0, f0, f0, f0 };
+  Packet4f v1 = { f1, f1, f1, f1 };
+  return (Packet4f)__builtin_msa_ilvr_d((v2i64)v1, (v2i64)v0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int32_t* from) {
+  EIGEN_MSA_DEBUG;
+
+  int32_t i0 = from[0], i1 = from[1];
+  Packet4i v0 = { i0, i0, i0, i0 };
+  Packet4i v1 = { i1, i1, i1, i1 };
+  return (Packet4i)__builtin_msa_ilvr_d((v2i64)v1, (v2i64)v0);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_STORE __builtin_msa_st_w((Packet4i)from, to, 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<int32_t>(int32_t* to, const Packet4i& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_STORE __builtin_msa_st_w(from, to, 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_STORE __builtin_msa_st_w((Packet4i)from, to, 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<int32_t>(int32_t* to, const Packet4i& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_STORE __builtin_msa_st_w(from, to, 0);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  float f = *from;
+  Packet4f v = { f, f, f, f };
+  v[1] = from[stride];
+  v[2] = from[2 * stride];
+  v[3] = from[3 * stride];
+  return v;
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  int32_t i = *from;
+  Packet4i v = { i, i, i, i };
+  v[1] = from[stride];
+  v[2] = from[2 * stride];
+  v[3] = from[3 * stride];
+  return v;
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from,
+                                                        Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  *to = from[0];
+  to += stride;
+  *to = from[1];
+  to += stride;
+  *to = from[2];
+  to += stride;
+  *to = from[3];
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from,
+                                                          Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  *to = from[0];
+  to += stride;
+  *to = from[1];
+  to += stride;
+  *to = from[2];
+  to += stride;
+  *to = from[3];
+}
+
+template <>
+EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) {
+  EIGEN_MSA_DEBUG;
+
+  __builtin_prefetch(addr);
+}
+
+template <>
+EIGEN_STRONG_INLINE void prefetch<int32_t>(const int32_t* addr) {
+  EIGEN_MSA_DEBUG;
+
+  __builtin_prefetch(addr);
+}
+
+template <>
+EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_shf_w((v4i32)a, EIGEN_MSA_SHF_I8(3, 2, 1, 0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_shf_w(a, EIGEN_MSA_SHF_I8(3, 2, 1, 0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet4f)__builtin_msa_bclri_w((v4u32)a, 31);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4i zero = __builtin_msa_ldi_w(0);
+  return __builtin_msa_add_a_w(zero, a);
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f s = padd(a, (Packet4f)__builtin_msa_shf_w((v4i32)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  s = padd(s, (Packet4f)__builtin_msa_shf_w((v4i32)s, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return s[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) {
+  EIGEN_MSA_DEBUG;
+
+  v4i32 tmp1, tmp2, tmp3, tmp4;
+  Packet4f sum;
+
+  tmp1 = __builtin_msa_ilvr_w((v4i32)vecs[1], (v4i32)vecs[0]);
+  tmp2 = __builtin_msa_ilvr_w((v4i32)vecs[3], (v4i32)vecs[2]);
+  tmp3 = __builtin_msa_ilvl_w((v4i32)vecs[1], (v4i32)vecs[0]);
+  tmp4 = __builtin_msa_ilvl_w((v4i32)vecs[3], (v4i32)vecs[2]);
+
+  sum = (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
+  sum = padd(sum, (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1));
+  sum = padd(sum, (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3));
+  sum = padd(sum, (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3));
+
+  return sum;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) {
+  EIGEN_MSA_DEBUG;
+
+  v4i32 tmp1, tmp2, tmp3, tmp4;
+  Packet4i sum;
+
+  tmp1 = __builtin_msa_ilvr_w((v4i32)vecs[1], (v4i32)vecs[0]);
+  tmp2 = __builtin_msa_ilvr_w((v4i32)vecs[3], (v4i32)vecs[2]);
+  tmp3 = __builtin_msa_ilvl_w((v4i32)vecs[1], (v4i32)vecs[0]);
+  tmp4 = __builtin_msa_ilvl_w((v4i32)vecs[3], (v4i32)vecs[2]);
+
+  sum = (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
+  sum = padd(sum, (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1));
+  sum = padd(sum, (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3));
+  sum = padd(sum, (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3));
+
+  return sum;
+}
+
+template <>
+EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4i s = padd(a, __builtin_msa_shf_w(a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  s = padd(s, __builtin_msa_shf_w(s, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return s[0];
+}
+
+// Other reduction functions:
+// mul
+template <>
+EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4f p = pmul(a, (Packet4f)__builtin_msa_shf_w((v4i32)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  p = pmul(p, (Packet4f)__builtin_msa_shf_w((v4i32)p, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return p[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE int32_t predux_mul<Packet4i>(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4i p = pmul(a, __builtin_msa_shf_w(a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  p = pmul(p, __builtin_msa_shf_w(p, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return p[0];
+}
+
+// min
+template <>
+EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  // Swap 64-bit halves of a.
+  Packet4f swapped = (Packet4f)__builtin_msa_shf_w((Packet4i)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+#if !EIGEN_FAST_MATH
+  // Detect presence of NaNs from pairs a[0]-a[2] and a[1]-a[3] as two 32-bit
+  // masks of all zeroes/ones in low 64 bits.
+  v16u8 unord = (v16u8)__builtin_msa_fcun_w(a, swapped);
+  // Combine the two masks into one: 64 ones if no NaNs, otherwise 64 zeroes.
+  unord = (v16u8)__builtin_msa_ceqi_d((v2i64)unord, 0);
+#endif
+  // Continue with min computation.
+  Packet4f v = __builtin_msa_fmin_w(a, swapped);
+  v = __builtin_msa_fmin_w(
+      v, (Packet4f)__builtin_msa_shf_w((Packet4i)v, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+#if !EIGEN_FAST_MATH
+  // Based on the mask select between v and 4 qNaNs.
+  v16u8 qnans = (v16u8)__builtin_msa_fill_w(0x7FC00000);
+  v = (Packet4f)__builtin_msa_bsel_v(unord, qnans, (v16u8)v);
+#endif
+  return v[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE int32_t predux_min<Packet4i>(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4i m = pmin(a, __builtin_msa_shf_w(a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  m = pmin(m, __builtin_msa_shf_w(m, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return m[0];
+}
+
+// max
+template <>
+EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  // Swap 64-bit halves of a.
+  Packet4f swapped = (Packet4f)__builtin_msa_shf_w((Packet4i)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+#if !EIGEN_FAST_MATH
+  // Detect presence of NaNs from pairs a[0]-a[2] and a[1]-a[3] as two 32-bit
+  // masks of all zeroes/ones in low 64 bits.
+  v16u8 unord = (v16u8)__builtin_msa_fcun_w(a, swapped);
+  // Combine the two masks into one: 64 ones if no NaNs, otherwise 64 zeroes.
+  unord = (v16u8)__builtin_msa_ceqi_d((v2i64)unord, 0);
+#endif
+  // Continue with max computation.
+  Packet4f v = __builtin_msa_fmax_w(a, swapped);
+  v = __builtin_msa_fmax_w(
+      v, (Packet4f)__builtin_msa_shf_w((Packet4i)v, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+#if !EIGEN_FAST_MATH
+  // Based on the mask select between v and 4 qNaNs.
+  v16u8 qnans = (v16u8)__builtin_msa_fill_w(0x7FC00000);
+  v = (Packet4f)__builtin_msa_bsel_v(unord, qnans, (v16u8)v);
+#endif
+  return v[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet4i m = pmax(a, __builtin_msa_shf_w(a, EIGEN_MSA_SHF_I8(2, 3, 0, 1)));
+  m = pmax(m, __builtin_msa_shf_w(m, EIGEN_MSA_SHF_I8(1, 0, 3, 2)));
+  return m[0];
+}
+
+#define PALIGN_MSA(Offset, Type, Command)                                                \
+  template <>                                                                            \
+  struct palign_impl<Offset, Type> {                                                     \
+    EIGEN_STRONG_INLINE static void run(Type& first, const Type& second) {               \
+      if (Offset != 0) first = (Type)(Command((v16i8)second, (v16i8)first, Offset * 4)); \
+    }                                                                                    \
+  };
+
+PALIGN_MSA(0, Packet4f, __builtin_msa_sldi_b)
+PALIGN_MSA(1, Packet4f, __builtin_msa_sldi_b)
+PALIGN_MSA(2, Packet4f, __builtin_msa_sldi_b)
+PALIGN_MSA(3, Packet4f, __builtin_msa_sldi_b)
+PALIGN_MSA(0, Packet4i, __builtin_msa_sldi_b)
+PALIGN_MSA(1, Packet4i, __builtin_msa_sldi_b)
+PALIGN_MSA(2, Packet4i, __builtin_msa_sldi_b)
+PALIGN_MSA(3, Packet4i, __builtin_msa_sldi_b)
+
+#undef PALIGN_MSA
+
+inline std::ostream& operator<<(std::ostream& os, const PacketBlock<Packet4f, 4>& value) {
+  os << "[ " << value.packet[0] << "," << std::endl
+     << "  " << value.packet[1] << "," << std::endl
+     << "  " << value.packet[2] << "," << std::endl
+     << "  " << value.packet[3] << " ]";
+  return os;
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel) {
+  EIGEN_MSA_DEBUG;
+
+  v4i32 tmp1, tmp2, tmp3, tmp4;
+
+  tmp1 = __builtin_msa_ilvr_w((v4i32)kernel.packet[1], (v4i32)kernel.packet[0]);
+  tmp2 = __builtin_msa_ilvr_w((v4i32)kernel.packet[3], (v4i32)kernel.packet[2]);
+  tmp3 = __builtin_msa_ilvl_w((v4i32)kernel.packet[1], (v4i32)kernel.packet[0]);
+  tmp4 = __builtin_msa_ilvl_w((v4i32)kernel.packet[3], (v4i32)kernel.packet[2]);
+
+  kernel.packet[0] = (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
+  kernel.packet[1] = (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1);
+  kernel.packet[2] = (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3);
+  kernel.packet[3] = (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3);
+}
+
+inline std::ostream& operator<<(std::ostream& os, const PacketBlock<Packet4i, 4>& value) {
+  os << "[ " << value.packet[0] << "," << std::endl
+     << "  " << value.packet[1] << "," << std::endl
+     << "  " << value.packet[2] << "," << std::endl
+     << "  " << value.packet[3] << " ]";
+  return os;
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4i, 4>& kernel) {
+  EIGEN_MSA_DEBUG;
+
+  v4i32 tmp1, tmp2, tmp3, tmp4;
+
+  tmp1 = __builtin_msa_ilvr_w(kernel.packet[1], kernel.packet[0]);
+  tmp2 = __builtin_msa_ilvr_w(kernel.packet[3], kernel.packet[2]);
+  tmp3 = __builtin_msa_ilvl_w(kernel.packet[1], kernel.packet[0]);
+  tmp4 = __builtin_msa_ilvl_w(kernel.packet[3], kernel.packet[2]);
+
+  kernel.packet[0] = (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
+  kernel.packet[1] = (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1);
+  kernel.packet[2] = (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3);
+  kernel.packet[3] = (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fsqrt_w(a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f prsqrt(const Packet4f& a) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  return __builtin_msa_frsqrt_w(a);
+#else
+  Packet4f ones = __builtin_msa_ffint_s_w(__builtin_msa_ldi_w(1));
+  return pdiv(ones, psqrt(a));
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) {
+  Packet4f v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"  // 3 = round towards -INFINITY.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.w %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) {
+  Packet4f v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"
+      "xori    %[new_mode], %[new_mode], 1\n"  // 2 = round towards +INFINITY.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.w %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) {
+  Packet4f v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"
+      "xori    %[new_mode], %[new_mode], 3\n"  // 0 = round to nearest, ties to even.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.w %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket,
+                                    const Packet4f& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2],
+                       ifPacket.select[3] };
+  Packet4i mask = __builtin_msa_ceqi_w((Packet4i)select, 0);
+  return (Packet4f)__builtin_msa_bsel_v((v16u8)mask, (v16u8)thenPacket, (v16u8)elsePacket);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket,
+                                    const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2],
+                       ifPacket.select[3] };
+  Packet4i mask = __builtin_msa_ceqi_w((Packet4i)select, 0);
+  return (Packet4i)__builtin_msa_bsel_v((v16u8)mask, (v16u8)thenPacket, (v16u8)elsePacket);
+}
+
+//---------- double ----------
+
+typedef v2f64 Packet2d;
+typedef v2i64 Packet2l;
+typedef v2u64 Packet2ul;
+
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME, X) const Packet2d p2d_##NAME = { X, X }
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME, X) const Packet2l p2l_##NAME = { X, X }
+#define _EIGEN_DECLARE_CONST_Packet2ul(NAME, X) const Packet2ul p2ul_##NAME = { X, X }
+
+inline std::ostream& operator<<(std::ostream& os, const Packet2d& value) {
+  os << "[ " << value[0] << ", " << value[1] << " ]";
+  return os;
+}
+
+inline std::ostream& operator<<(std::ostream& os, const Packet2l& value) {
+  os << "[ " << value[0] << ", " << value[1] << " ]";
+  return os;
+}
+
+inline std::ostream& operator<<(std::ostream& os, const Packet2ul& value) {
+  os << "[ " << value[0] << ", " << value[1] << " ]";
+  return os;
+}
+
+template <>
+struct packet_traits<double> : default_packet_traits {
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+    // FIXME check the Has*
+    HasDiv = 1,
+    HasExp = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasBlend = 1
+  };
+};
+
+template <>
+struct unpacket_traits<Packet2d> {
+  typedef double type;
+  enum { size = 2, alignment = Aligned16 };
+  typedef Packet2d half;
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d value = { from, from };
+  return value;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fadd_d(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) {
+  EIGEN_MSA_DEBUG;
+
+  static const Packet2d countdown = { 0.0, 1.0 };
+  return padd(pset1<Packet2d>(a), countdown);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fsub_d(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_bnegi_d((v2u64)a, 63);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fmul_d(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fdiv_d(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fmadd_d(c, a, b);
+}
+
+// Logical Operations are not supported for float, so we have to reinterpret casts using MSA
+// intrinsics
+template <>
+EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_and_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_or_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_xor_v((v16u8)a, (v16u8)b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+  return pand(a, (Packet2d)__builtin_msa_xori_b((v16u8)b, 255));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return (Packet2d)__builtin_msa_ld_d(const_cast<double*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  // This prefers numbers to NaNs.
+  return __builtin_msa_fmin_d(a, b);
+#else
+  // This prefers NaNs to numbers.
+  v2i64 aNaN = __builtin_msa_fcun_d(a, a);
+  v2i64 aMinOrNaN = por(__builtin_msa_fclt_d(a, b), aNaN);
+  return (Packet2d)__builtin_msa_bsel_v((v16u8)aMinOrNaN, (v16u8)b, (v16u8)a);
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  // This prefers numbers to NaNs.
+  return __builtin_msa_fmax_d(a, b);
+#else
+  // This prefers NaNs to numbers.
+  v2i64 aNaN = __builtin_msa_fcun_d(a, a);
+  v2i64 aMaxOrNaN = por(__builtin_msa_fclt_d(b, a), aNaN);
+  return (Packet2d)__builtin_msa_bsel_v((v16u8)aMaxOrNaN, (v16u8)b, (v16u8)a);
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_LOAD return (Packet2d)__builtin_msa_ld_d(const_cast<double*>(from), 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d value = { *from, *from };
+  return value;
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_ALIGNED_STORE __builtin_msa_st_d((v2i64)from, to, 0);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) {
+  EIGEN_MSA_DEBUG;
+
+  EIGEN_DEBUG_UNALIGNED_STORE __builtin_msa_st_d((v2i64)from, to, 0);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d value;
+  value[0] = *from;
+  from += stride;
+  value[1] = *from;
+  return value;
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from,
+                                                         Index stride) {
+  EIGEN_MSA_DEBUG;
+
+  *to = from[0];
+  to += stride;
+  *to = from[1];
+}
+
+template <>
+EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) {
+  EIGEN_MSA_DEBUG;
+
+  __builtin_prefetch(addr);
+}
+
+template <>
+EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return a[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_shf_w((v4i32)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return (Packet2d)__builtin_msa_bclri_d((v2u64)a, 63);
+}
+
+template <>
+EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d s = padd(a, preverse(a));
+  return s[0];
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d v0 = (Packet2d)__builtin_msa_ilvev_d((v2i64)vecs[1], (v2i64)vecs[0]);
+  Packet2d v1 = (Packet2d)__builtin_msa_ilvod_d((v2i64)vecs[1], (v2i64)vecs[0]);
+
+  return padd(v0, v1);
+}
+
+// Other reduction functions:
+// mul
+template <>
+EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d p = pmul(a, preverse(a));
+  return p[0];
+}
+
+// min
+template <>
+EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  Packet2d swapped = (Packet2d)__builtin_msa_shf_w((Packet4i)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+  Packet2d v = __builtin_msa_fmin_d(a, swapped);
+  return v[0];
+#else
+  double a0 = a[0], a1 = a[1];
+  return ((std::isnan)(a0) || a0 < a1) ? a0 : a1;
+#endif
+}
+
+// max
+template <>
+EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  Packet2d swapped = (Packet2d)__builtin_msa_shf_w((Packet4i)a, EIGEN_MSA_SHF_I8(2, 3, 0, 1));
+  Packet2d v = __builtin_msa_fmax_d(a, swapped);
+  return v[0];
+#else
+  double a0 = a[0], a1 = a[1];
+  return ((std::isnan)(a0) || a0 > a1) ? a0 : a1;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+  return __builtin_msa_fsqrt_d(a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d prsqrt(const Packet2d& a) {
+  EIGEN_MSA_DEBUG;
+
+#if EIGEN_FAST_MATH
+  return __builtin_msa_frsqrt_d(a);
+#else
+  Packet2d ones = __builtin_msa_ffint_s_d(__builtin_msa_ldi_d(1));
+  return pdiv(ones, psqrt(a));
+#endif
+}
+
+#define PALIGN_MSA(Offset, Type, Command)                                                \
+  template <>                                                                            \
+  struct palign_impl<Offset, Type> {                                                     \
+    EIGEN_STRONG_INLINE static void run(Type& first, const Type& second) {               \
+      if (Offset != 0) first = (Type)(Command((v16i8)second, (v16i8)first, Offset * 8)); \
+    }                                                                                    \
+  };
+
+PALIGN_MSA(0, Packet2d, __builtin_msa_sldi_b)
+PALIGN_MSA(1, Packet2d, __builtin_msa_sldi_b)
+
+#undef PALIGN_MSA
+
+inline std::ostream& operator<<(std::ostream& os, const PacketBlock<Packet2d, 2>& value) {
+  os << "[ " << value.packet[0] << "," << std::endl << "  " << value.packet[1] << " ]";
+  return os;
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2d, 2>& kernel) {
+  EIGEN_MSA_DEBUG;
+
+  Packet2d trn1 = (Packet2d)__builtin_msa_ilvev_d((v2i64)kernel.packet[1], (v2i64)kernel.packet[0]);
+  Packet2d trn2 = (Packet2d)__builtin_msa_ilvod_d((v2i64)kernel.packet[1], (v2i64)kernel.packet[0]);
+  kernel.packet[0] = trn1;
+  kernel.packet[1] = trn2;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) {
+  Packet2d v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"  // 3 = round towards -INFINITY.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.d %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) {
+  Packet2d v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"
+      "xori    %[new_mode], %[new_mode], 1\n"  // 2 = round towards +INFINITY.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.d %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) {
+  Packet2d v = a;
+  int32_t old_mode, new_mode;
+  asm volatile(
+      "cfcmsa  %[old_mode], $1\n"
+      "ori     %[new_mode], %[old_mode], 3\n"
+      "xori    %[new_mode], %[new_mode], 3\n"  // 0 = round to nearest, ties to even.
+      "ctcmsa  $1, %[new_mode]\n"
+      "frint.d %w[v], %w[v]\n"
+      "ctcmsa  $1, %[old_mode]\n"
+      :  // outputs
+      [old_mode] "=r"(old_mode), [new_mode] "=r"(new_mode),
+      [v] "+f"(v)
+      :  // inputs
+      :  // clobbers
+  );
+  return v;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket,
+                                    const Packet2d& elsePacket) {
+  Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2l mask = __builtin_msa_ceqi_d((Packet2l)select, 0);
+  return (Packet2d)__builtin_msa_bsel_v((v16u8)mask, (v16u8)thenPacket, (v16u8)elsePacket);
+}
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_PACKET_MATH_MSA_H
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index 27f65f672..653e979b1 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -94,7 +94,7 @@
   // certain common platform (compiler+architecture combinations) to avoid these problems.
   // Only static alignment is really problematic (relies on nonstandard compiler extensions),
   // try to keep heap alignment even when we have to disable static alignment.
-  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64 || EIGEN_ARCH_MIPS)
   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
   #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
   // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h
index 5d37e5d04..f1afb4db9 100644
--- a/Eigen/src/Core/util/Constants.h
+++ b/Eigen/src/Core/util/Constants.h
@@ -468,6 +468,7 @@ namespace Architecture
     AltiVec = 0x2,
     VSX = 0x3,
     NEON = 0x4,
+    MSA = 0x5,
 #if defined EIGEN_VECTORIZE_SSE
     Target = SSE
 #elif defined EIGEN_VECTORIZE_ALTIVEC
@@ -476,6 +477,8 @@ namespace Architecture
     Target = VSX
 #elif defined EIGEN_VECTORIZE_NEON
     Target = NEON
+#elif defined EIGEN_VECTORIZE_MSA
+    Target = MSA
 #else
     Target = Generic
 #endif
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 1d4486c05..7712f109a 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -452,6 +452,12 @@ macro(ei_testing_print_summary)
       message(STATUS "VSX:               Using architecture defaults")
     endif()
 
+    if(EIGEN_TEST_MSA)
+      message(STATUS "MIPS MSA:          ON")
+    else()
+      message(STATUS "MIPS MSA:          Using architecture defaults")
+    endif()
+
     if(EIGEN_TEST_NEON)
       message(STATUS "ARM NEON:          ON")
     else()
@@ -655,6 +661,8 @@ macro(ei_get_cxxflags VAR)
     set(${VAR} SSE3)
   elseif(EIGEN_TEST_SSE2 OR IS_64BIT_ENV)
     set(${VAR} SSE2)
+  elseif(EIGEN_TEST_MSA)
+    set(${VAR} MSA)
   endif()
 
   if(EIGEN_TEST_OPENMP)

From 308725c3c94af57b7ba10e3c78ccd2355767fff9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 13:51:36 +0200
Subject: [PATCH 447/680] More clearly disable the inclusion of
 src/Core/arch/CUDA/Complex.h without CUDA

---
 Eigen/Core | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/Core b/Eigen/Core
index 2ab44fb3c..23676311f 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -206,7 +206,9 @@ using std::ptrdiff_t;
 
 // Specialized functors to enable the processing of complex numbers
 // on CUDA devices
+#ifdef EIGEN_CUDACC
 #include "src/Core/arch/CUDA/Complex.h"
+#endif
 
 #include "src/Core/util/IndexedViewHelper.h"
 #include "src/Core/ArithmeticSequence.h"

From a503fc87254aca625d7b128a2260599d40dc0889 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 23:26:13 +0200
Subject: [PATCH 448/680] bug #1575: fix regression introduced in bug #1573
 patch. Move ctor/assignment should not be defaulted.

---
 Eigen/src/Geometry/Quaternion.h | 34 ++++++++++++++-------------------
 1 file changed, 14 insertions(+), 20 deletions(-)

diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h
index 6fad6564a..8b203dc71 100644
--- a/Eigen/src/Geometry/Quaternion.h
+++ b/Eigen/src/Geometry/Quaternion.h
@@ -105,22 +105,6 @@ class QuaternionBase : public RotationBase<Derived, 3>
   EIGEN_DEVICE_FUNC Derived& operator=(const AngleAxisType& aa);
   template<class OtherDerived> EIGEN_DEVICE_FUNC Derived& operator=(const MatrixBase<OtherDerived>& m);
 
-#if EIGEN_HAS_RVALUE_REFERENCES
-  // Because we have a user-defined copy assignment operator, we don't get an implicit move constructor or assignment operator.
-  /** Default move constructor */
-  EIGEN_DEVICE_FUNC inline QuaternionBase(QuaternionBase&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value) = default;
-
-  /** Default move assignment operator */
-  EIGEN_DEVICE_FUNC QuaternionBase& operator=(QuaternionBase&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) = default;
-
-  // And now because we declared a constructor, we don't get a default constructor or copy constructor. Say we want them.
-  /** Default constructor */
-  EIGEN_DEVICE_FUNC QuaternionBase() = default;
-
-  /** Copy constructor */
-  EIGEN_DEVICE_FUNC QuaternionBase(const QuaternionBase& other) = default;
-#endif
-
   /** \returns a quaternion representing an identity rotation
     * \sa MatrixBase::Identity()
     */
@@ -295,14 +279,24 @@ public:
 #if EIGEN_HAS_RVALUE_REFERENCES
   // We define a copy constructor, which means we don't get an implicit move constructor or assignment operator.
   /** Default move constructor */
-  EIGEN_DEVICE_FUNC inline Quaternion(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value) = default;
+  EIGEN_DEVICE_FUNC inline Quaternion(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
+  {
+    m_coeffs(std::move(other.coeffs()));
+  }
 
   /** Default move assignment operator */
-  EIGEN_DEVICE_FUNC Quaternion& operator=(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) = default;
+  EIGEN_DEVICE_FUNC Quaternion& operator=(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
+  {
+    m_coeffs = std::move(other.coeffs());
+    return *this;
+  }
 
   // And now because we declared a constructor, we don't get an implicit copy constructor. Say we want one.
   /** Default copy constructor */
-  EIGEN_DEVICE_FUNC Quaternion(const Quaternion& other) = default;
+  EIGEN_DEVICE_FUNC Quaternion(const Quaternion& other)
+  {
+    m_coeffs = other.coeffs();
+  }
 #endif
 
   EIGEN_DEVICE_FUNC static Quaternion UnitRandom();
@@ -657,7 +651,7 @@ EIGEN_DEVICE_FUNC Quaternion<Scalar,Options> Quaternion<Scalar,Options>::UnitRan
   const Scalar u1 = internal::random<Scalar>(0, 1),
                u2 = internal::random<Scalar>(0, 2*EIGEN_PI),
                u3 = internal::random<Scalar>(0, 2*EIGEN_PI);
-  const Scalar a = sqrt(1 - u1),
+  const Scalar a = sqrt(Scalar(1) - u1),
                b = sqrt(u1);
   return Quaternion (a * sin(u2), a * cos(u2), b * sin(u3), b * cos(u3));
 }

From 053ed97c72543d513e92cdef1b14839844664c34 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 23:27:37 +0200
Subject: [PATCH 449/680] Generalize ScalarWithExceptions to a full
 non-copyable and trowing scalar type to be used in other unit tests.

---
 test/AnnoyingScalar.h | 151 ++++++++++++++++++++++++++++++++++++++++++
 test/exceptions.cpp   |  90 ++++---------------------
 2 files changed, 164 insertions(+), 77 deletions(-)
 create mode 100644 test/AnnoyingScalar.h

diff --git a/test/AnnoyingScalar.h b/test/AnnoyingScalar.h
new file mode 100644
index 000000000..08854a114
--- /dev/null
+++ b/test/AnnoyingScalar.h
@@ -0,0 +1,151 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2018 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TEST_ANNOYING_SCALAR_H
+#define EIGEN_TEST_ANNOYING_SCALAR_H
+
+#include <ostream>
+
+#ifndef EIGEN_TEST_ANNOYING_SCALAR_DONT_THROW
+struct my_exception
+{
+  my_exception() {}
+  ~my_exception() {}
+};
+#endif
+
+// An AnnoyingScalar is a pseudo scalar type that:
+// - can randomly through an exception in operator +
+// - randomly allocate on the heap or initialize a reference to itself making it non trivially copyable, nor movable.
+
+class AnnoyingScalar
+{
+  public:
+    AnnoyingScalar()                { init(); *v = 0;  }
+    AnnoyingScalar(long double _v)  { init(); *v = _v; }
+    AnnoyingScalar(double _v)       { init(); *v = _v; }
+    AnnoyingScalar(float _v)        { init(); *v = _v; }
+    AnnoyingScalar(int _v)          { init(); *v = _v; }
+    AnnoyingScalar(long _v)         { init(); *v = _v; }
+    AnnoyingScalar(const AnnoyingScalar& other) { init(); *v = *(other.v); }
+    ~AnnoyingScalar() {
+      if(v!=&data)
+        delete v;
+      instances--;
+    }
+
+    void init() {
+      if(internal::random<bool>())
+        v = new float;
+      else
+        v = &data;
+      instances++;
+    }
+
+    AnnoyingScalar operator+(const AnnoyingScalar& other) const
+    {
+      #ifndef EIGEN_TEST_ANNOYING_SCALAR_DONT_THROW
+      countdown--;
+      if(countdown<=0 && !dont_throw)
+        throw my_exception();
+      #endif
+      return AnnoyingScalar(*v+*other.v);
+    }
+
+    AnnoyingScalar operator-() const
+    { return AnnoyingScalar(-*v); }
+    
+    AnnoyingScalar operator-(const AnnoyingScalar& other) const
+    { return AnnoyingScalar(*v-*other.v); }
+    
+    AnnoyingScalar operator*(const AnnoyingScalar& other) const
+    { return AnnoyingScalar((*v)*(*other.v)); }
+
+    AnnoyingScalar operator/(const AnnoyingScalar& other) const
+    { return AnnoyingScalar((*v)/(*other.v)); }
+    
+    AnnoyingScalar& operator+=(const AnnoyingScalar& other) { *v += *other.v; return *this; }
+    AnnoyingScalar& operator-=(const AnnoyingScalar& other) { *v -= *other.v; return *this; }
+    AnnoyingScalar& operator*=(const AnnoyingScalar& other) { *v *= *other.v; return *this; }
+    AnnoyingScalar& operator/=(const AnnoyingScalar& other) { *v /= *other.v; return *this; }
+    AnnoyingScalar& operator= (const AnnoyingScalar& other) { *v  = *other.v; return *this; }
+  
+    bool operator==(const AnnoyingScalar& other) const { return *v == *other.v; }
+    bool operator!=(const AnnoyingScalar& other) const { return *v != *other.v; }
+    bool operator<=(const AnnoyingScalar& other) const { return *v <= *other.v; }
+    bool operator< (const AnnoyingScalar& other) const { return *v <  *other.v; }
+    bool operator>=(const AnnoyingScalar& other) const { return *v >= *other.v; }
+    bool operator> (const AnnoyingScalar& other) const { return *v >  *other.v; }
+    
+    float* v;
+    float data;
+    static int instances;
+#ifndef EIGEN_TEST_ANNOYING_SCALAR_DONT_THROW
+    static int countdown;
+    static bool dont_throw;
+#endif
+};
+
+AnnoyingScalar real(const AnnoyingScalar &x) { return x; }
+AnnoyingScalar imag(const AnnoyingScalar & ) { return 0; }
+AnnoyingScalar conj(const AnnoyingScalar &x) { return x; }
+AnnoyingScalar sqrt(const AnnoyingScalar &x) { return std::sqrt(*x.v); }
+AnnoyingScalar abs (const AnnoyingScalar &x) { return std::abs(*x.v); }
+AnnoyingScalar cos (const AnnoyingScalar &x) { return std::cos(*x.v); }
+AnnoyingScalar sin (const AnnoyingScalar &x) { return std::sin(*x.v); }
+AnnoyingScalar acos(const AnnoyingScalar &x) { return std::acos(*x.v); }
+AnnoyingScalar atan2(const AnnoyingScalar &y,const AnnoyingScalar &x) { return std::atan2(*y.v,*x.v); }
+
+std::ostream& operator<<(std::ostream& stream,const AnnoyingScalar& x) {
+  stream << (*(x.v));
+  return stream;
+}
+
+int AnnoyingScalar::instances = 0;
+
+#ifndef EIGEN_TEST_ANNOYING_SCALAR_DONT_THROW
+int AnnoyingScalar::countdown = 0;
+bool AnnoyingScalar::dont_throw = false;
+#endif
+
+namespace Eigen {
+template<>
+struct NumTraits<AnnoyingScalar> : NumTraits<float>
+{
+  enum {
+    RequireInitialization = true
+  };
+  typedef AnnoyingScalar Real;
+  typedef AnnoyingScalar Nested;
+  typedef AnnoyingScalar Literal;
+  typedef AnnoyingScalar NonInteger;
+};
+
+template<> inline AnnoyingScalar test_precision<AnnoyingScalar>() { return test_precision<float>(); }
+
+namespace internal {
+  template<> double cast(const AnnoyingScalar& x) { return double(*x.v); }
+  template<> float  cast(const AnnoyingScalar& x) { return *x.v; }
+}
+
+}
+
+AnnoyingScalar get_test_precision(const AnnoyingScalar&)
+{ return Eigen::test_precision<AnnoyingScalar>(); }
+
+AnnoyingScalar test_relative_error(AnnoyingScalar a, AnnoyingScalar b)
+{ return test_relative_error(*a.v, *b.v); }
+
+inline bool test_isApprox(AnnoyingScalar a, AnnoyingScalar b)
+{ return internal::isApprox(*a.v, *b.v, test_precision<float>()); }
+
+inline bool test_isMuchSmallerThan(AnnoyingScalar a, AnnoyingScalar b)
+{ return test_isMuchSmallerThan(*a.v, *b.v); }
+
+#endif // EIGEN_TEST_ANNOYING_SCALAR_H
diff --git a/test/exceptions.cpp b/test/exceptions.cpp
index 182082e39..3d93060ab 100644
--- a/test/exceptions.cpp
+++ b/test/exceptions.cpp
@@ -8,93 +8,34 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 
-// Various sanity tests with exceptions:
+// Various sanity tests with exceptions and non trivially copyable scalar type.
 //  - no memory leak when a custom scalar type trow an exceptions
 //  - todo: complete the list of tests!
 
 #define EIGEN_STACK_ALLOCATION_LIMIT 100000000
 
 #include "main.h"
-
-struct my_exception
-{
-  my_exception() {}
-  ~my_exception() {}
-};
-    
-class ScalarWithExceptions
-{
-  public:
-    ScalarWithExceptions() { init(); }
-    ScalarWithExceptions(const float& _v) { init(); *v = _v; }
-    ScalarWithExceptions(const ScalarWithExceptions& other) { init(); *v = *(other.v); }
-    ~ScalarWithExceptions() {
-      delete v;
-      instances--;
-    }
-
-    void init() {
-      v = new float;
-      instances++;
-    }
-
-    ScalarWithExceptions operator+(const ScalarWithExceptions& other) const
-    {
-      countdown--;
-      if(countdown<=0)
-        throw my_exception();
-      return ScalarWithExceptions(*v+*other.v);
-    }
-    
-    ScalarWithExceptions operator-(const ScalarWithExceptions& other) const
-    { return ScalarWithExceptions(*v-*other.v); }
-    
-    ScalarWithExceptions operator*(const ScalarWithExceptions& other) const
-    { return ScalarWithExceptions((*v)*(*other.v)); }
-    
-    ScalarWithExceptions& operator+=(const ScalarWithExceptions& other)
-    { *v+=*other.v; return *this; }
-    ScalarWithExceptions& operator-=(const ScalarWithExceptions& other)
-    { *v-=*other.v; return *this; }
-    ScalarWithExceptions& operator=(const ScalarWithExceptions& other)
-    { *v = *(other.v); return *this; }
-  
-    bool operator==(const ScalarWithExceptions& other) const
-    { return *v==*other.v; }
-    bool operator!=(const ScalarWithExceptions& other) const
-    { return *v!=*other.v; }
-    
-    float* v;
-    static int instances;
-    static int countdown;
-};
-
-ScalarWithExceptions real(const ScalarWithExceptions &x) { return x; }
-ScalarWithExceptions imag(const ScalarWithExceptions & ) { return 0; }
-ScalarWithExceptions conj(const ScalarWithExceptions &x) { return x; }
-
-int ScalarWithExceptions::instances = 0;
-int ScalarWithExceptions::countdown = 0;
-
+#include "AnnoyingScalar.h"
 
 #define CHECK_MEMLEAK(OP) {                                 \
-    ScalarWithExceptions::countdown = 100;                  \
-    int before = ScalarWithExceptions::instances;           \
-    bool exception_thrown = false;                         \
-    try { OP; }                              \
+    AnnoyingScalar::countdown = 100;                        \
+    int before = AnnoyingScalar::instances;                 \
+    bool exception_thrown = false;                          \
+    try { OP; }                                             \
     catch (my_exception) {                                  \
       exception_thrown = true;                              \
-      VERIFY(ScalarWithExceptions::instances==before && "memory leak detected in " && EIGEN_MAKESTRING(OP)); \
+      VERIFY(AnnoyingScalar::instances==before && "memory leak detected in " && EIGEN_MAKESTRING(OP)); \
     } \
-    VERIFY(exception_thrown && " no exception thrown in " && EIGEN_MAKESTRING(OP)); \
+    VERIFY( (AnnoyingScalar::dont_throw) || (exception_thrown && " no exception thrown in " && EIGEN_MAKESTRING(OP)) ); \
   }
 
-void memoryleak()
+EIGEN_DECLARE_TEST(exceptions)
 {
-  typedef Eigen::Matrix<ScalarWithExceptions,Dynamic,1> VectorType;
-  typedef Eigen::Matrix<ScalarWithExceptions,Dynamic,Dynamic> MatrixType;
+  typedef Eigen::Matrix<AnnoyingScalar,Dynamic,1> VectorType;
+  typedef Eigen::Matrix<AnnoyingScalar,Dynamic,Dynamic> MatrixType;
   
   {
+    AnnoyingScalar::dont_throw = false;
     int n = 50;
     VectorType v0(n), v1(n);
     MatrixType m0(n,n), m1(n,n), m2(n,n);
@@ -104,10 +45,5 @@ void memoryleak()
     CHECK_MEMLEAK(m2 = m0 * m1 * m2);
     CHECK_MEMLEAK((v0+v1).dot(v0+v1));
   }
-  VERIFY(ScalarWithExceptions::instances==0 && "global memory leak detected in " && EIGEN_MAKESTRING(OP)); \
-}
-
-EIGEN_DECLARE_TEST(exceptions)
-{
-  CALL_SUBTEST( memoryleak() );
+  VERIFY(AnnoyingScalar::instances==0 && "global memory leak detected in " && EIGEN_MAKESTRING(OP));
 }

From 863580fe881c32af82eb106817e71dc560d9e775 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 23:33:07 +0200
Subject: [PATCH 450/680] bug #1432: fix conservativeResize for non-relocatable
 scalar types. For those we need to by-pass realloc routines and fall-back to
 allocate as new - copy - delete. The remaining problem is that we don't have
 any mechanism to accurately determine whether a type is relocatable or not,
 so currently let's be super conservative using either RequireInitialization
 or std::is_trivially_copyable

---
 Eigen/src/Core/PlainObjectBase.h | 29 ++++++++++++++++++++++-------
 test/AnnoyingScalar.h            |  2 +-
 test/conservative_resize.cpp     | 31 ++++++++++++++++++++++++++++++-
 test/sparse_block.cpp            |  9 +++++++--
 4 files changed, 60 insertions(+), 11 deletions(-)

diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index 6c0a42ec7..016e76a53 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -921,13 +921,19 @@ namespace internal {
 template <typename Derived, typename OtherDerived, bool IsVector>
 struct conservative_resize_like_impl
 {
+  #if EIGEN_HAS_TYPE_TRAITS
+  static const bool IsRelocatable = std::is_trivially_copyable<typename Derived::Scalar>::value;
+  #else
+  static const bool IsRelocatable = NumTraits<typename Derived::Scalar>::RequireInitialization;
+  #endif
   static void run(DenseBase<Derived>& _this, Index rows, Index cols)
   {
     if (_this.rows() == rows && _this.cols() == cols) return;
     EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived)
 
-    if ( ( Derived::IsRowMajor && _this.cols() == cols) || // row-major and we change only the number of rows
-         (!Derived::IsRowMajor && _this.rows() == rows) )  // column-major and we change only the number of columns
+    if ( IsRelocatable
+          && (( Derived::IsRowMajor && _this.cols() == cols) ||  // row-major and we change only the number of rows
+              (!Derived::IsRowMajor && _this.rows() == rows) ))  // column-major and we change only the number of columns
     {
       internal::check_rows_cols_for_overflow<Derived::MaxSizeAtCompileTime>::run(rows, cols);
       _this.derived().m_storage.conservativeResize(rows*cols,rows,cols);
@@ -955,8 +961,9 @@ struct conservative_resize_like_impl
     EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived)
     EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(OtherDerived)
 
-    if ( ( Derived::IsRowMajor && _this.cols() == other.cols()) || // row-major and we change only the number of rows
-         (!Derived::IsRowMajor && _this.rows() == other.rows()) )  // column-major and we change only the number of columns
+    if ( IsRelocatable &&
+          (( Derived::IsRowMajor && _this.cols() == other.cols()) ||  // row-major and we change only the number of rows
+           (!Derived::IsRowMajor && _this.rows() == other.rows()) ))  // column-major and we change only the number of columns
     {
       const Index new_rows = other.rows() - _this.rows();
       const Index new_cols = other.cols() - _this.cols();
@@ -984,13 +991,18 @@ template <typename Derived, typename OtherDerived>
 struct conservative_resize_like_impl<Derived,OtherDerived,true>
   : conservative_resize_like_impl<Derived,OtherDerived,false>
 {
-  using conservative_resize_like_impl<Derived,OtherDerived,false>::run;
+  typedef conservative_resize_like_impl<Derived,OtherDerived,false> Base;
+  using Base::run;
+  using Base::IsRelocatable;
   
   static void run(DenseBase<Derived>& _this, Index size)
   {
     const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size;
     const Index new_cols = Derived::RowsAtCompileTime==1 ? size : 1;
-    _this.derived().m_storage.conservativeResize(size,new_rows,new_cols);
+    if(IsRelocatable)
+      _this.derived().m_storage.conservativeResize(size,new_rows,new_cols);
+    else
+      Base::run(_this.derived(), new_rows, new_cols);
   }
 
   static void run(DenseBase<Derived>& _this, const DenseBase<OtherDerived>& other)
@@ -1001,7 +1013,10 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 
     const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : other.rows();
     const Index new_cols = Derived::RowsAtCompileTime==1 ? other.cols() : 1;
-    _this.derived().m_storage.conservativeResize(other.size(),new_rows,new_cols);
+    if(IsRelocatable)
+      _this.derived().m_storage.conservativeResize(other.size(),new_rows,new_cols);
+    else
+      Base::run(_this.derived(), new_rows, new_cols);
 
     if (num_new_elements > 0)
       _this.tail(num_new_elements) = other.tail(num_new_elements);
diff --git a/test/AnnoyingScalar.h b/test/AnnoyingScalar.h
index 08854a114..b5b1476fe 100644
--- a/test/AnnoyingScalar.h
+++ b/test/AnnoyingScalar.h
@@ -22,7 +22,7 @@ struct my_exception
 
 // An AnnoyingScalar is a pseudo scalar type that:
 // - can randomly through an exception in operator +
-// - randomly allocate on the heap or initialize a reference to itself making it non trivially copyable, nor movable.
+// - randomly allocate on the heap or initialize a reference to itself making it non trivially copyable, nor movable, nor relocatable.
 
 class AnnoyingScalar
 {
diff --git a/test/conservative_resize.cpp b/test/conservative_resize.cpp
index d3d0a5a31..5dc500068 100644
--- a/test/conservative_resize.cpp
+++ b/test/conservative_resize.cpp
@@ -10,6 +10,7 @@
 #include "main.h"
 
 #include <Eigen/Core>
+#include "AnnoyingScalar.h"
 
 using namespace Eigen;
 
@@ -109,6 +110,30 @@ void run_vector_tests()
   }
 }
 
+// Basic memory leak check with a non-copyable scalar type
+template<int> void noncopyable()
+{
+  typedef Eigen::Matrix<AnnoyingScalar,Dynamic,1> VectorType;
+  typedef Eigen::Matrix<AnnoyingScalar,Dynamic,Dynamic> MatrixType;
+  
+  {
+    AnnoyingScalar::dont_throw = true;
+    int n = 50;
+    VectorType v0(n), v1(n);
+    MatrixType m0(n,n), m1(n,n), m2(n,n);
+    v0.setOnes(); v1.setOnes();
+    m0.setOnes(); m1.setOnes(); m2.setOnes();
+    VERIFY(m0==m1);
+    m0.conservativeResize(2*n,2*n);
+    VERIFY(m0.topLeftCorner(n,n) == m1);
+    
+    VERIFY(v0.head(n) == v1);
+    v0.conservativeResize(2*n);
+    VERIFY(v0.head(n) == v1);
+  }
+  VERIFY(AnnoyingScalar::instances==0 && "global memory leak detected in noncopyable");
+}
+
 EIGEN_DECLARE_TEST(conservative_resize)
 {
   for(int i=0; i<g_repeat; ++i)
@@ -122,12 +147,16 @@ EIGEN_DECLARE_TEST(conservative_resize)
     CALL_SUBTEST_4((run_matrix_tests<std::complex<float>, Eigen::RowMajor>()));
     CALL_SUBTEST_4((run_matrix_tests<std::complex<float>, Eigen::ColMajor>()));
     CALL_SUBTEST_5((run_matrix_tests<std::complex<double>, Eigen::RowMajor>()));
-    CALL_SUBTEST_6((run_matrix_tests<std::complex<double>, Eigen::ColMajor>()));
+    CALL_SUBTEST_5((run_matrix_tests<std::complex<double>, Eigen::ColMajor>()));
 
     CALL_SUBTEST_1((run_vector_tests<int>()));
     CALL_SUBTEST_2((run_vector_tests<float>()));
     CALL_SUBTEST_3((run_vector_tests<double>()));
     CALL_SUBTEST_4((run_vector_tests<std::complex<float> >()));
     CALL_SUBTEST_5((run_vector_tests<std::complex<double> >()));
+
+    AnnoyingScalar::dont_throw = true;
+    CALL_SUBTEST_6(( run_vector_tests<AnnoyingScalar>() ));
+    CALL_SUBTEST_6(( noncopyable<0>() ));
   }
 }
diff --git a/test/sparse_block.cpp b/test/sparse_block.cpp
index e7303eab0..f9668102c 100644
--- a/test/sparse_block.cpp
+++ b/test/sparse_block.cpp
@@ -8,6 +8,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #include "sparse.h"
+#include "AnnoyingScalar.h"
 
 template<typename T>
 typename Eigen::internal::enable_if<(T::Flags&RowMajorBit)==RowMajorBit, typename T::RowXpr>::type
@@ -31,6 +32,7 @@ template<typename SparseMatrixType> void sparse_block(const SparseMatrixType& re
   const Index outer = ref.outerSize();
 
   typedef typename SparseMatrixType::Scalar Scalar;
+  typedef typename SparseMatrixType::RealScalar RealScalar;
   typedef typename SparseMatrixType::StorageIndex StorageIndex;
 
   double density = (std::max)(8./(rows*cols), 0.01);
@@ -164,14 +166,14 @@ template<typename SparseMatrixType> void sparse_block(const SparseMatrixType& re
     {
       VERIFY(j==numext::real(m3.innerVector(j).nonZeros()));
       if(j>0)
-        VERIFY(j==numext::real(m3.innerVector(j).lastCoeff()));
+        VERIFY(RealScalar(j)==numext::real(m3.innerVector(j).lastCoeff()));
     }
     m3.makeCompressed();
     for(Index j=0; j<(std::min)(outer, inner); ++j)
     {
       VERIFY(j==numext::real(m3.innerVector(j).nonZeros()));
       if(j>0)
-        VERIFY(j==numext::real(m3.innerVector(j).lastCoeff()));
+        VERIFY(RealScalar(j)==numext::real(m3.innerVector(j).lastCoeff()));
     }
 
     VERIFY(m3.innerVector(j0).nonZeros() == m3.transpose().innerVector(j0).nonZeros());
@@ -313,5 +315,8 @@ EIGEN_DECLARE_TEST(sparse_block)
     
     CALL_SUBTEST_4(( sparse_block(SparseMatrix<double,ColMajor,short int>(short(r), short(c))) ));
     CALL_SUBTEST_4(( sparse_block(SparseMatrix<double,RowMajor,short int>(short(r), short(c))) ));
+
+    AnnoyingScalar::dont_throw = true;
+    CALL_SUBTEST_5((  sparse_block(SparseMatrix<AnnoyingScalar>(r,c)) ));
   }
 }

From 6e5a3b898fde197d4748315694103f577c0f503f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 18 Jul 2018 23:34:34 +0200
Subject: [PATCH 451/680] Add regression for bugs #1573 and #1575

---
 test/geo_quaternion.cpp | 33 ++++++++++++++++++++++++++-------
 1 file changed, 26 insertions(+), 7 deletions(-)

diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp
index 3541b8b25..ed801c71b 100644
--- a/test/geo_quaternion.cpp
+++ b/test/geo_quaternion.cpp
@@ -12,6 +12,7 @@
 #include <Eigen/Geometry>
 #include <Eigen/LU>
 #include <Eigen/SVD>
+#include "AnnoyingScalar.h"
 
 template<typename T> T bounded_acos(T v)
 {
@@ -85,7 +86,7 @@ template<typename Scalar, int Options> void quaternion(void)
   if (refangle>Scalar(EIGEN_PI))
     refangle = Scalar(2)*Scalar(EIGEN_PI) - refangle;
 
-  if((q1.coeffs()-q2.coeffs()).norm() > 10*largeEps)
+  if((q1.coeffs()-q2.coeffs()).norm() > Scalar(10)*largeEps)
   {
     VERIFY_IS_MUCH_SMALLER_THAN(abs(q1.angularDistance(q2) - refangle), Scalar(1));
   }
@@ -113,7 +114,7 @@ template<typename Scalar, int Options> void quaternion(void)
 
   // Do not execute the test if the rotation angle is almost zero, or
   // the rotation axis and v1 are almost parallel.
-  if (abs(aa.angle()) > 5*test_precision<Scalar>()
+  if (abs(aa.angle()) > Scalar(5)*test_precision<Scalar>()
       && (aa.axis() - v1.normalized()).norm() < Scalar(1.99)
       && (aa.axis() + v1.normalized()).norm() < Scalar(1.99))
   {
@@ -285,18 +286,36 @@ template<typename PlainObjectType> void check_const_correctness(const PlainObjec
   VERIFY( !(Map<ConstPlainObjectType, Aligned>::Flags & LvalueBit) );
 }
 
+#if EIGEN_HAS_RVALUE_REFERENCES
+
+// Regression for bug 1573
+struct MovableClass {
+  MovableClass() = default;
+  MovableClass(const MovableClass&) = default;
+  MovableClass(MovableClass&&) noexcept = default;
+  MovableClass& operator=(const MovableClass&) = default;
+  MovableClass& operator=(MovableClass&&) = default;
+  Quaternionf m_quat;
+};
+
+#endif
+
 EIGEN_DECLARE_TEST(geo_quaternion)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1(( quaternion<float,AutoAlign>() ));
     CALL_SUBTEST_1( check_const_correctness(Quaternionf()) );
+    CALL_SUBTEST_1(( quaternion<float,DontAlign>() ));
+    CALL_SUBTEST_1(( quaternionAlignment<float>() ));
+    CALL_SUBTEST_1( mapQuaternion<float>() );
+
     CALL_SUBTEST_2(( quaternion<double,AutoAlign>() ));
     CALL_SUBTEST_2( check_const_correctness(Quaterniond()) );
-    CALL_SUBTEST_3(( quaternion<float,DontAlign>() ));
-    CALL_SUBTEST_4(( quaternion<double,DontAlign>() ));
-    CALL_SUBTEST_5(( quaternionAlignment<float>() ));
-    CALL_SUBTEST_6(( quaternionAlignment<double>() ));
-    CALL_SUBTEST_1( mapQuaternion<float>() );
+    CALL_SUBTEST_2(( quaternion<double,DontAlign>() ));
+    CALL_SUBTEST_2(( quaternionAlignment<double>() ));
     CALL_SUBTEST_2( mapQuaternion<double>() );
+
+    AnnoyingScalar::dont_throw = true;
+    CALL_SUBTEST_3(( quaternion<AnnoyingScalar,AutoAlign>() ));
   }
 }

From 79d4129cce4055955e7a340822df31b6f8d25edc Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 18 Jul 2018 14:21:01 -0700
Subject: [PATCH 452/680] Specify default output kernel for TensorContractionOp

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index c7c443a59..0e69cd40c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -182,7 +182,7 @@ struct NoOpOutputKernel {
       Index num_cols) const {}
 };
 
-template<typename Indices, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+template<typename Indices, typename LhsXprType, typename RhsXprType, typename OutputKernelType = const NoOpOutputKernel>
 class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXprType, RhsXprType, OutputKernelType>, ReadOnlyAccessors>
 {
   public:
@@ -507,7 +507,7 @@ struct TensorContractionEvaluatorBase
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  #if !defined(EIGEN_HIPCC)    
+  #if !defined(EIGEN_HIPCC)
   EIGEN_DEVICE_FUNC
   #endif
   void evalGemv(Scalar* buffer) const {
@@ -556,7 +556,7 @@ struct TensorContractionEvaluatorBase
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  #if !defined(EIGEN_HIPCC)    
+  #if !defined(EIGEN_HIPCC)
   EIGEN_DEVICE_FUNC
   #endif
   void evalGemm(Scalar* buffer) const {

From 086ded5c85457f813cddaec03f6a53c561f8652d Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 18 Jul 2018 16:32:55 -0700
Subject: [PATCH 453/680] Disable type traits for GCC < 5.1.0

---
 Eigen/src/Core/util/Macros.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index adf25ee9b..985406953 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -396,7 +396,7 @@
   #define EIGEN_HIPCC __HIPCC__
 
   // We need to include hip_runtime.h here because it pulls in
-  // ++ hip_common.h which contains the define for  __HIP_DEVICE_COMPILE__ 
+  // ++ hip_common.h which contains the define for  __HIP_DEVICE_COMPILE__
   // ++ host_defines.h which contains the defines for the __host__ and __device__ macros
   #include <hip/hip_runtime.h>
 
@@ -541,7 +541,9 @@
 
 // Does the compiler support type_traits?
 #ifndef EIGEN_HAS_TYPE_TRAITS
-#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
+#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700) \
+  && (!EIGEN_COMP_GNUC || EIGEN_GNUC_AT_LEAST(5, 1))
+    // ^^ Full support of type traits was added only to GCC 5.1.0.
 #define EIGEN_HAS_TYPE_TRAITS 1
 #define EIGEN_INCLUDE_TYPE_TRAITS
 #else

From e3c2d61739ebf09ad4bf86752a9878b7cf4e045f Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 18 Jul 2018 14:34:22 -0700
Subject: [PATCH 454/680] Assert that no output kernel is defined for GPU
 contraction

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index 6d3aa24c8..b5e186d21 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -1219,6 +1219,9 @@ template<typename Indices, typename LeftArgType, typename RightArgType, typename
 struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> :
     public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> > {
 
+  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
+                "GPU tensor contraction does not support output kernels.");
+
   typedef GpuDevice Device;
 
   typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;

From 1f523e73042074fcaea6d9991bc2bb6e4f96a1ab Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Wed, 18 Jul 2018 12:27:50 -0700
Subject: [PATCH 455/680] Add MIPS changes missing from previous merge.

---
 Eigen/Core                                   |  4 ++++
 Eigen/src/Core/util/ConfigureVectorization.h | 17 +++++++++++++++++
 2 files changed, 21 insertions(+)

diff --git a/Eigen/Core b/Eigen/Core
index 23676311f..f336d407b 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -183,6 +183,10 @@ using std::ptrdiff_t;
   #include "src/Core/arch/ZVector/PacketMath.h"
   #include "src/Core/arch/ZVector/MathFunctions.h"
   #include "src/Core/arch/ZVector/Complex.h"
+#elif defined EIGEN_VECTORIZE_MSA
+  #include "src/Core/arch/MSA/PacketMath.h"
+  #include "src/Core/arch/MSA/MathFunctions.h"
+  #include "src/Core/arch/MSA/Complex.h"
 #endif
 
 // Half float support
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index 653e979b1..f30503b33 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -339,6 +339,21 @@
     #define EIGEN_VECTORIZE_ZVECTOR
     #include <vecintrin.h>
 
+  #elif defined __mips_msa
+
+    // Limit MSA optimizations to little-endian CPUs for now.
+    // TODO: Perhaps, eventually support MSA optimizations on big-endian CPUs?
+    #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+      #if defined(__LP64__)
+        #define EIGEN_MIPS_64
+      #else
+        #define EIGEN_MIPS_32
+      #endif
+      #define EIGEN_VECTORIZE
+      #define EIGEN_VECTORIZE_MSA
+      #include <msa.h>
+    #endif
+
   #endif
 #endif
 
@@ -404,6 +419,8 @@ inline static const char *SimdInstructionSetsInUse(void) {
   return "ARM NEON";
 #elif defined(EIGEN_VECTORIZE_ZVECTOR)
   return "S390X ZVECTOR";
+#elif defined(EIGEN_VECTORIZE_MSA)
+  return "MIPS MSA";
 #else
   return "None";
 #endif

From 7ccb623746ea36013689dbdf61f6ce50948e6c29 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 19 Jul 2018 13:15:40 +0200
Subject: [PATCH 456/680] bug #1569: fix Tensor<half>::mean() on AVX with
 respective unit test.

---
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   |  2 +-
 unsupported/test/cxx11_tensor_reduction.cpp   | 41 ++++++++++---------
 2 files changed, 22 insertions(+), 21 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 0efcd6253..a9027e97d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -140,7 +140,7 @@ struct reducer_traits<SumReducer<T>, Device> {
 
 template <typename T> struct MeanReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasAdd && !NumTraits<T>::IsInteger;
+  static const bool PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv && !NumTraits<T>::IsInteger;
   static const bool IsStateful = true;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp
index 4c8a95c34..ff8e18c07 100644
--- a/unsupported/test/cxx11_tensor_reduction.cpp
+++ b/unsupported/test/cxx11_tensor_reduction.cpp
@@ -53,20 +53,20 @@ static void test_trivial_reductions() {
   }
 }
 
-template <int DataLayout>
+template <typename Scalar,int DataLayout>
 static void test_simple_reductions() {
-  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
+  Tensor<Scalar, 4, DataLayout> tensor(2, 3, 5, 7);
   tensor.setRandom();
   array<ptrdiff_t, 2> reduction_axis2;
   reduction_axis2[0] = 1;
   reduction_axis2[1] = 3;
 
-  Tensor<float, 2, DataLayout> result = tensor.sum(reduction_axis2);
+  Tensor<Scalar, 2, DataLayout> result = tensor.sum(reduction_axis2);
   VERIFY_IS_EQUAL(result.dimension(0), 2);
   VERIFY_IS_EQUAL(result.dimension(1), 5);
   for (int i = 0; i < 2; ++i) {
     for (int j = 0; j < 5; ++j) {
-      float sum = 0.0f;
+      Scalar sum = Scalar(0.0f);
       for (int k = 0; k < 3; ++k) {
         for (int l = 0; l < 7; ++l) {
           sum += tensor(i, k, j, l);
@@ -77,7 +77,7 @@ static void test_simple_reductions() {
   }
 
   {
-    Tensor<float, 0, DataLayout> sum1 = tensor.sum();
+    Tensor<Scalar, 0, DataLayout> sum1 = tensor.sum();
     VERIFY_IS_EQUAL(sum1.rank(), 0);
 
     array<ptrdiff_t, 4> reduction_axis4;
@@ -85,7 +85,7 @@ static void test_simple_reductions() {
     reduction_axis4[1] = 1;
     reduction_axis4[2] = 2;
     reduction_axis4[3] = 3;
-    Tensor<float, 0, DataLayout> sum2 = tensor.sum(reduction_axis4);
+    Tensor<Scalar, 0, DataLayout> sum2 = tensor.sum(reduction_axis4);
     VERIFY_IS_EQUAL(sum2.rank(), 0);
 
     VERIFY_IS_APPROX(sum1(), sum2());
@@ -98,7 +98,7 @@ static void test_simple_reductions() {
   VERIFY_IS_EQUAL(result.dimension(1), 7);
   for (int i = 0; i < 3; ++i) {
     for (int j = 0; j < 7; ++j) {
-      float prod = 1.0f;
+      Scalar prod = Scalar(1.0f);
       for (int k = 0; k < 2; ++k) {
         for (int l = 0; l < 5; ++l) {
           prod *= tensor(k, i, l, j);
@@ -109,7 +109,7 @@ static void test_simple_reductions() {
   }
 
   {
-    Tensor<float, 0, DataLayout> prod1 = tensor.prod();
+    Tensor<Scalar, 0, DataLayout> prod1 = tensor.prod();
     VERIFY_IS_EQUAL(prod1.rank(), 0);
 
     array<ptrdiff_t, 4> reduction_axis4;
@@ -117,7 +117,7 @@ static void test_simple_reductions() {
     reduction_axis4[1] = 1;
     reduction_axis4[2] = 2;
     reduction_axis4[3] = 3;
-    Tensor<float, 0, DataLayout> prod2 = tensor.prod(reduction_axis4);
+    Tensor<Scalar, 0, DataLayout> prod2 = tensor.prod(reduction_axis4);
     VERIFY_IS_EQUAL(prod2.rank(), 0);
 
     VERIFY_IS_APPROX(prod1(), prod2());
@@ -130,7 +130,7 @@ static void test_simple_reductions() {
   VERIFY_IS_EQUAL(result.dimension(1), 7);
   for (int i = 0; i < 3; ++i) {
     for (int j = 0; j < 7; ++j) {
-      float max_val = std::numeric_limits<float>::lowest();
+      Scalar max_val = std::numeric_limits<Scalar>::lowest();
       for (int k = 0; k < 2; ++k) {
         for (int l = 0; l < 5; ++l) {
           max_val = (std::max)(max_val, tensor(k, i, l, j));
@@ -141,7 +141,7 @@ static void test_simple_reductions() {
   }
 
   {
-    Tensor<float, 0, DataLayout> max1 = tensor.maximum();
+    Tensor<Scalar, 0, DataLayout> max1 = tensor.maximum();
     VERIFY_IS_EQUAL(max1.rank(), 0);
 
     array<ptrdiff_t, 4> reduction_axis4;
@@ -149,7 +149,7 @@ static void test_simple_reductions() {
     reduction_axis4[1] = 1;
     reduction_axis4[2] = 2;
     reduction_axis4[3] = 3;
-    Tensor<float, 0, DataLayout> max2 = tensor.maximum(reduction_axis4);
+    Tensor<Scalar, 0, DataLayout> max2 = tensor.maximum(reduction_axis4);
     VERIFY_IS_EQUAL(max2.rank(), 0);
 
     VERIFY_IS_APPROX(max1(), max2());
@@ -162,7 +162,7 @@ static void test_simple_reductions() {
   VERIFY_IS_EQUAL(result.dimension(1), 7);
   for (int i = 0; i < 5; ++i) {
     for (int j = 0; j < 7; ++j) {
-      float min_val = (std::numeric_limits<float>::max)();
+      Scalar min_val = (std::numeric_limits<Scalar>::max)();
       for (int k = 0; k < 2; ++k) {
         for (int l = 0; l < 3; ++l) {
           min_val = (std::min)(min_val, tensor(k, l, i, j));
@@ -173,7 +173,7 @@ static void test_simple_reductions() {
   }
 
   {
-    Tensor<float, 0, DataLayout> min1 = tensor.minimum();
+    Tensor<Scalar, 0, DataLayout> min1 = tensor.minimum();
     VERIFY_IS_EQUAL(min1.rank(), 0);
 
     array<ptrdiff_t, 4> reduction_axis4;
@@ -181,7 +181,7 @@ static void test_simple_reductions() {
     reduction_axis4[1] = 1;
     reduction_axis4[2] = 2;
     reduction_axis4[3] = 3;
-    Tensor<float, 0, DataLayout> min2 = tensor.minimum(reduction_axis4);
+    Tensor<Scalar, 0, DataLayout> min2 = tensor.minimum(reduction_axis4);
     VERIFY_IS_EQUAL(min2.rank(), 0);
 
     VERIFY_IS_APPROX(min1(), min2());
@@ -194,7 +194,7 @@ static void test_simple_reductions() {
   VERIFY_IS_EQUAL(result.dimension(1), 7);
   for (int i = 0; i < 5; ++i) {
     for (int j = 0; j < 7; ++j) {
-      float sum = 0.0f;
+      Scalar sum = Scalar(0.0f);
       int count = 0;
       for (int k = 0; k < 2; ++k) {
         for (int l = 0; l < 3; ++l) {
@@ -207,7 +207,7 @@ static void test_simple_reductions() {
   }
 
   {
-    Tensor<float, 0, DataLayout> mean1 = tensor.mean();
+    Tensor<Scalar, 0, DataLayout> mean1 = tensor.mean();
     VERIFY_IS_EQUAL(mean1.rank(), 0);
 
     array<ptrdiff_t, 4> reduction_axis4;
@@ -215,7 +215,7 @@ static void test_simple_reductions() {
     reduction_axis4[1] = 1;
     reduction_axis4[2] = 2;
     reduction_axis4[3] = 3;
-    Tensor<float, 0, DataLayout> mean2 = tensor.mean(reduction_axis4);
+    Tensor<Scalar, 0, DataLayout> mean2 = tensor.mean(reduction_axis4);
     VERIFY_IS_EQUAL(mean2.rank(), 0);
 
     VERIFY_IS_APPROX(mean1(), mean2());
@@ -487,8 +487,9 @@ static void test_reduce_middle_dims() {
 EIGEN_DECLARE_TEST(cxx11_tensor_reduction) {
   CALL_SUBTEST(test_trivial_reductions<ColMajor>());
   CALL_SUBTEST(test_trivial_reductions<RowMajor>());
-  CALL_SUBTEST(test_simple_reductions<ColMajor>());
-  CALL_SUBTEST(test_simple_reductions<RowMajor>());
+  CALL_SUBTEST(( test_simple_reductions<float,ColMajor>() ));
+  CALL_SUBTEST(( test_simple_reductions<float,RowMajor>() ));
+  CALL_SUBTEST(( test_simple_reductions<Eigen::half,ColMajor>() ));
   CALL_SUBTEST(test_reductions_in_expr<ColMajor>());
   CALL_SUBTEST(test_reductions_in_expr<RowMajor>());
   CALL_SUBTEST(test_full_reductions<ColMajor>());

From 6e654f33794df0b329160d011e6d3ec035e679ae Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Mon, 16 Jul 2018 14:26:39 -0700
Subject: [PATCH 457/680] Reduce number of allocations in
 TensorContractionThreadPool.

---
 .../CXX11/src/Tensor/TensorContractionThreadPool.h   | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index c774d15e2..8b86d7aaf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -598,11 +598,13 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         else
           pack_lhs(start, k);
       } else {
-        Index mid = (start + end) / 2;
-        device_.enqueueNoNotification(
-            [=]() { enqueue_packing_helper(mid, end, k, rhs); });
-        device_.enqueueNoNotification(
-            [=]() { enqueue_packing_helper(start, mid, k, rhs); });
+        while (end - start > 1) {
+          Index mid = (start + end) / 2;
+          device_.enqueueNoNotification(
+              [=]() { enqueue_packing_helper(mid, end, k, rhs); });
+          end = mid;
+        }
+        enqueue_packing_helper(start, end, k, rhs);
       }
     }
 

From d908afe35f2bf521499abb009602e7029ed65f2d Mon Sep 17 00:00:00 2001
From: David Hyde <dabh@stanford.edu>
Date: Sun, 8 Jul 2018 22:06:38 -0700
Subject: [PATCH 458/680] bug #1558: fix a corner case in MINRES when both
 v_new and w_new vanish.

---
 unsupported/Eigen/src/IterativeSolvers/MINRES.h | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h
index 256990c1a..3a5c73eaf 100644
--- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h
@@ -3,6 +3,7 @@
 //
 // Copyright (C) 2012 Giacomo Po <gpo@ucla.edu>
 // Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2018 David Hyde <dabh@stanford.edu>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -64,8 +65,6 @@ namespace Eigen {
             eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
             RealScalar beta_new(sqrt(beta_new2));
             const RealScalar beta_one(beta_new);
-            v_new /= beta_new;
-            w_new /= beta_new;
             // Initialize other variables
             RealScalar c(1.0); // the cosine of the Givens rotation
             RealScalar c_old(1.0);
@@ -83,18 +82,18 @@ namespace Eigen {
                 /* Note that there are 4 variants on the Lanczos algorithm. These are
                  * described in Paige, C. C. (1972). Computational variants of
                  * the Lanczos method for the eigenproblem. IMA Journal of Applied
-                 * Mathematics, 10(3), 373–381. The current implementation corresponds 
+                 * Mathematics, 10(3), 373-381. The current implementation corresponds 
                  * to the case A(2,7) in the paper. It also corresponds to 
-                 * algorithm 6.14 in Y. Saad, Iterative Methods for Sparse Linear
+                 * algorithm 6.14 in Y. Saad, Iterative Methods for Sparse Linear
                  * Systems, 2003 p.173. For the preconditioned version see 
                  * A. Greenbaum, Iterative Methods for Solving Linear Systems, SIAM (1987).
                  */
                 const RealScalar beta(beta_new);
                 v_old = v; // update: at first time step, this makes v_old = 0 so value of beta doesn't matter
-//                const VectorType v_old(v); // NOT SURE IF CREATING v_old EVERY ITERATION IS EFFICIENT
+                v_new /= beta_new; // overwrite v_new for next iteration
+                w_new /= beta_new; // overwrite w_new for next iteration
                 v = v_new; // update
                 w = w_new; // update
-//                const VectorType w(w_new); // NOT SURE IF CREATING w EVERY ITERATION IS EFFICIENT
                 v_new.noalias() = mat*w - beta*v_old; // compute v_new
                 const RealScalar alpha = v_new.dot(w);
                 v_new -= alpha*v; // overwrite v_new
@@ -102,8 +101,6 @@ namespace Eigen {
                 beta_new2 = v_new.dot(w_new); // compute beta_new
                 eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
                 beta_new = sqrt(beta_new2); // compute beta_new
-                v_new /= beta_new; // overwrite v_new for next iteration
-                w_new /= beta_new; // overwrite w_new for next iteration
                 
                 // Givens rotation
                 const RealScalar r2 =s*alpha+c*c_old*beta; // s, s_old, c and c_old are still from previous iteration
@@ -117,7 +114,6 @@ namespace Eigen {
                 
                 // Update solution
                 p_oold = p_old;
-//                const VectorType p_oold(p_old); // NOT SURE IF CREATING p_oold EVERY ITERATION IS EFFICIENT
                 p_old = p;
                 p.noalias()=(w-r2*p_old-r3*p_oold) /r1; // IS NOALIAS REQUIRED?
                 x += beta_one*c*eta*p;
@@ -286,4 +282,3 @@ namespace Eigen {
 } // end namespace Eigen
 
 #endif // EIGEN_MINRES_H
-

From 86d9c0255c75c340bd03faa1a6ae5fe5c16ccada Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 13:06:44 +0200
Subject: [PATCH 459/680] Forward declaring std::array does not work with all
 std libs, so let's just include <array>

---
 Eigen/Core                 | 4 ++++
 Eigen/src/Core/util/Meta.h | 7 -------
 2 files changed, 4 insertions(+), 7 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index fd6edc018..2ab44fb3c 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -80,6 +80,10 @@
 // for min/max:
 #include <algorithm>
 
+#if EIGEN_HAS_CXX11
+#include <array>
+#endif
+
 // for std::is_nothrow_move_assignable
 #ifdef EIGEN_INCLUDE_TYPE_TRAITS
 #include <type_traits>
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 33dc0034c..748f24b1e 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -360,13 +360,6 @@ template<typename T, int N> struct array_size<T (&)[N]> {
 };
 
 #if EIGEN_HAS_CXX11
-}}
-namespace std {
-  template<class T, std::size_t N> struct array;
-}
-namespace Eigen {
-namespace internal {
-
 template<typename T, std::size_t N> struct array_size<const std::array<T,N> > {
   enum { value = N };
 };

From 5fd03ddbfb91a6d641903229ed1428bc82756c4f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 16:03:14 +0200
Subject: [PATCH 460/680] Make EIGEN_TEST_CUDA_CLANG more friendly with OSX

---
 cmake/EigenTesting.cmake | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 7712f109a..58f3697d3 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -24,17 +24,23 @@ macro(ei_add_test_internal testname testname_with_suffix)
       hip_add_executable(${targetname} ${filename} HIPCC_OPTIONS "-DEIGEN_USE_HIP ${ARGV2}")
     elseif(EIGEN_TEST_CUDA_CLANG)
       set_source_files_properties(${filename} PROPERTIES LANGUAGE CXX)
-      if(CUDA_64_BIT_DEVICE_CODE)
+      
+      if(CUDA_64_BIT_DEVICE_CODE AND (EXISTS "${CUDA_TOOLKIT_ROOT_DIR}/lib64"))
         link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib64")
       else()
         link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib")
       endif()
+
       if (${ARGC} GREATER 2)
         add_executable(${targetname} ${filename})
       else()
         add_executable(${targetname} ${filename} OPTIONS ${ARGV2})
       endif()
-      target_link_libraries(${targetname} "cudart_static" "cuda" "dl" "rt" "pthread")
+      set(CUDA_CLANG_LINK_LIBRARIES "cudart_static" "cuda" "dl" "pthread")
+      if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
+      set(CUDA_CLANG_LINK_LIBRARIES ${CUDA_CLANG_LINK_LIBRARIES} "rt")
+      endif()
+      target_link_libraries(${targetname} ${CUDA_CLANG_LINK_LIBRARIES})
     else()
       if (${ARGC} GREATER 2)
         cuda_add_executable(${targetname} ${filename} OPTIONS ${ARGV2})

From 06eb24cf4d7d54e56abfb37ea062a7cb0c887550 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 16:04:27 +0200
Subject: [PATCH 461/680] Introduce gpu_assert for assertion in device-code,
 and disable them with clang-cuda.

---
 .../CXX11/src/Tensor/TensorConvolution.h      |  6 +--
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |  6 +--
 .../Eigen/CXX11/src/Tensor/TensorDeviceGpu.h  | 40 +++++++++----------
 .../src/Tensor/TensorGpuHipCudaDefines.h      |  9 +++--
 .../Eigen/CXX11/src/Tensor/TensorRandom.h     |  2 +-
 .../CXX11/src/Tensor/TensorReductionGpu.h     | 24 +++++------
 6 files changed, 44 insertions(+), 43 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 3110887e1..25131600d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -903,7 +903,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
         }
 
         const int shared_mem = block_size.y * (maxX + kernel_size - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
+        gpu_assert(shared_mem <= maxSharedMem);
 
         const int num_x_blocks = ceil(numX, maxX);
         const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
@@ -960,7 +960,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
         block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxP);
 
         const int shared_mem = block_size.z * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
+        gpu_assert(shared_mem <= maxSharedMem);
 
         const int num_x_blocks = ceil(numX, maxX);
         const int num_y_blocks = ceil(numY, maxY);
@@ -1040,7 +1040,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
         dim3 num_blocks(ceil(numX, maxX), ceil(numY, maxY), ceil(numZ, maxZ));
 
         const int shared_mem = (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * (maxZ + kernel_size_z - 1) * sizeof(Scalar);
-        assert(shared_mem <= maxSharedMem);
+        gpu_assert(shared_mem <= maxSharedMem);
 
         //cout << "launching 3D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z  << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
         const array<Index, 3> indices(m_indices[idxX], m_indices[idxY],
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index da88bcb3b..65403905a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -352,7 +352,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           size_t range_x, GRange_x, tileSize_x, range_y, GRange_y, tileSize_y;
           m_device.parallel_for_setup(numX, numP, tileSize_x,tileSize_y,range_x,range_y, GRange_x, GRange_y );
           const size_t shared_mem =(tileSize_x +kernel_size -1)*(tileSize_y);
-          assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
+          gpu_assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
           auto global_range=cl::sycl::range<2>(GRange_x, GRange_y);  // global range
           auto local_range=cl::sycl::range<2>(tileSize_x, tileSize_y);  // local range
           InputLocalAcc local_acc(cl::sycl::range<1>(shared_mem), cgh);
@@ -377,7 +377,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           size_t range_x, GRange_x, tileSize_x, range_y, GRange_y, tileSize_y, range_z, GRange_z, tileSize_z;
           m_device.parallel_for_setup(numX, numY, numP, tileSize_x, tileSize_y, tileSize_z, range_x, range_y, range_z, GRange_x, GRange_y, GRange_z );
           const size_t shared_mem =(tileSize_x +kernel_size_x -1)*(tileSize_y +kernel_size_y -1) * tileSize_z;
-          assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
+          gpu_assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
           auto global_range=cl::sycl::range<3>(GRange_x, GRange_y, GRange_z);  // global range
           auto local_range=cl::sycl::range<3>(tileSize_x, tileSize_y, tileSize_z);  // local range
           InputLocalAcc local_acc(cl::sycl::range<1>(shared_mem), cgh);
@@ -408,7 +408,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
           size_t range_x, GRange_x, tileSize_x, range_y, GRange_y, tileSize_y, range_z, GRange_z, tileSize_z;
           m_device.parallel_for_setup(numX, numY, numZ, tileSize_x, tileSize_y, tileSize_z, range_x, range_y, range_z, GRange_x, GRange_y, GRange_z );
           const size_t shared_mem =(tileSize_x +kernel_size_x -1)*(tileSize_y +kernel_size_y -1) * (tileSize_z +kernel_size_y -1);
-          assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
+          gpu_assert(static_cast<unsigned long>(shared_mem) <= m_device.sharedMemPerBlock());
           auto global_range=cl::sycl::range<3>(GRange_x, GRange_y, GRange_z);  // global range
           auto local_range=cl::sycl::range<3>(tileSize_x, tileSize_y, tileSize_z);  // local range
           InputLocalAcc local_acc(cl::sycl::range<1>(shared_mem), cgh);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
index 64ef32793..0c036833f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
@@ -68,7 +68,7 @@ static void initializeDeviceProp() {
         std::cerr << "Failed to get the number of GPU devices: "
                   << gpuGetErrorString(status)
                   << std::endl;
-        assert(status == gpuSuccess);
+        gpu_assert(status == gpuSuccess);
       }
       m_deviceProperties = new gpuDeviceProp_t[num_devices];
       for (int i = 0; i < num_devices; ++i) {
@@ -79,7 +79,7 @@ static void initializeDeviceProp() {
                     << ": "
                     << gpuGetErrorString(status)
                     << std::endl;
-          assert(status == gpuSuccess);
+          gpu_assert(status == gpuSuccess);
         }
       }
 
@@ -124,8 +124,8 @@ class GpuStreamDevice : public StreamInterface {
       int num_devices;
       gpuError_t err = gpuGetDeviceCount(&num_devices);
       EIGEN_UNUSED_VARIABLE(err)
-      assert(err == gpuSuccess);
-      assert(device < num_devices);
+      gpu_assert(err == gpuSuccess);
+      gpu_assert(device < num_devices);
       device_ = device;
     }
     initializeDeviceProp();
@@ -144,20 +144,20 @@ class GpuStreamDevice : public StreamInterface {
   virtual void* allocate(size_t num_bytes) const {
     gpuError_t err = gpuSetDevice(device_);
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
     void* result;
     err = gpuMalloc(&result, num_bytes);
-    assert(err == gpuSuccess);
-    assert(result != NULL);
+    gpu_assert(err == gpuSuccess);
+    gpu_assert(result != NULL);
     return result;
   }
   virtual void deallocate(void* buffer) const {
     gpuError_t err = gpuSetDevice(device_);
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
-    assert(buffer != NULL);
+    gpu_assert(err == gpuSuccess);
+    gpu_assert(buffer != NULL);
     err = gpuFree(buffer);
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
   }
 
   virtual void* scratchpad() const {
@@ -173,7 +173,7 @@ class GpuStreamDevice : public StreamInterface {
       semaphore_ = reinterpret_cast<unsigned int*>(scratch);
       gpuError_t err = gpuMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
       EIGEN_UNUSED_VARIABLE(err)
-      assert(err == gpuSuccess);
+      gpu_assert(err == gpuSuccess);
     }
     return semaphore_;
   }
@@ -220,7 +220,7 @@ struct GpuDevice {
     gpuError_t err = gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToDevice,
                                       stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
 #else
     EIGEN_UNUSED_VARIABLE(dst);
     EIGEN_UNUSED_VARIABLE(src);
@@ -233,21 +233,21 @@ struct GpuDevice {
     gpuError_t err =
         gpuMemcpyAsync(dst, src, n, gpuMemcpyHostToDevice, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
   }
 
   EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
     gpuError_t err =
         gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToHost, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
 #ifndef EIGEN_GPU_COMPILE_PHASE
     gpuError_t err = gpuMemsetAsync(buffer, c, n, stream_->stream());
     EIGEN_UNUSED_VARIABLE(err)
-    assert(err == gpuSuccess);
+    gpu_assert(err == gpuSuccess);
 #else
   eigen_assert(false && "The default device should be used instead to generate kernel code");
 #endif
@@ -276,10 +276,10 @@ struct GpuDevice {
       std::cerr << "Error detected in GPU stream: "
                 << gpuGetErrorString(err)
                 << std::endl;
-      assert(err == gpuSuccess);
+      gpu_assert(err == gpuSuccess);
     }
 #else
-    assert(false && "The default device should be used instead to generate kernel code");
+    gpu_assert(false && "The default device should be used instead to generate kernel code");
 #endif
   }
 
@@ -326,13 +326,13 @@ struct GpuDevice {
 
 #define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
   hipLaunchKernelGGL(kernel, dim3(gridsize), dim3(blocksize), (sharedmem), (device).stream(), __VA_ARGS__); \
-  assert(hipGetLastError() == hipSuccess);
+  gpu_assert(hipGetLastError() == hipSuccess);
 
 #else
  
 #define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
   (kernel) <<< (gridsize), (blocksize), (sharedmem), (device).stream() >>> (__VA_ARGS__);   \
-  assert(cudaGetLastError() == cudaSuccess);
+  gpu_assert(cudaGetLastError() == cudaSuccess);
 
 #endif
  
@@ -342,7 +342,7 @@ static EIGEN_DEVICE_FUNC inline void setGpuSharedMemConfig(gpuSharedMemConfig co
 #ifndef EIGEN_GPU_COMPILE_PHASE
   gpuError_t status = gpuDeviceSetSharedMemConfig(config);
   EIGEN_UNUSED_VARIABLE(status)
-  assert(status == gpuSuccess);
+  gpu_assert(status == gpuSuccess);
 #else
   EIGEN_UNUSED_VARIABLE(config)
 #endif
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
index 9966955f7..5438ebe71 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
@@ -78,10 +78,11 @@
 
 #endif
 
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
-// HIPCC does not support the use of assert on the GPU side.
-#undef assert
-#define assert(COND)
+#if defined(EIGEN_HIP_DEVICE_COMPILE) || (defined(EIGEN_CUDACC) && (EIGEN_CUDACC_VER==0))
+// clang-cuda and HIPCC do not support the use of assert on the GPU side.
+#define gpu_assert(COND)
+#else
+#define gpu_assert(COND) assert(COND)
 #endif
 
 #endif  // EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
index 5a547141a..787cbd031 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
@@ -19,7 +19,7 @@ EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
 #if defined(EIGEN_GPU_COMPILE_PHASE)
   // We don't support 3d kernels since we currently only use 1 and
   // 2d kernels.
-  assert(threadIdx.z == 0);
+  gpu_assert(threadIdx.z == 0);
   return clock64() +
       blockIdx.x * blockDim.x + threadIdx.x +
       gridDim.x * blockDim.x * (blockIdx.y * blockDim.y + threadIdx.y);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
index a691e530a..cd20df505 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
@@ -60,10 +60,10 @@ __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer)
     }
   }
   else {
-    assert(0 && "Wordsize not supported");
+    gpu_assert(0 && "Wordsize not supported");
   }
 #else // EIGEN_CUDA_ARCH >= 300
-  assert(0 && "Shouldn't be called on unsupported device");
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
 #endif // EIGEN_CUDA_ARCH >= 300
 }
 
@@ -105,7 +105,7 @@ __device__ inline void atomicReduce(float* output, float accum, SumReducer<float
 #if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
   atomicAdd(output, accum);
 #else // EIGEN_CUDA_ARCH >= 300
-  assert(0 && "Shouldn't be called on unsupported device");
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
 #endif // EIGEN_CUDA_ARCH >= 300
 }
 
@@ -196,7 +196,7 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
 #endif
   }
 #else // EIGEN_CUDA_ARCH >= 300
-  assert(0 && "Shouldn't be called on unsupported device");
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
 #endif // EIGEN_CUDA_ARCH >= 300
 }
 
@@ -304,7 +304,7 @@ __global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct FullReductionLauncher {
   static void run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index) {
-    assert(false && "Should only be called on doubles, floats and half floats");
+    gpu_assert(false && "Should only be called on doubles, floats and half floats");
   }
 };
 
@@ -337,7 +337,7 @@ struct FullReductionLauncher<
 template <typename Self, typename Op>
 struct FullReductionLauncher<Self, Op, Eigen::half, false> {
   static void run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index) {
-    assert(false && "Should not be called since there is no packet accessor");
+    gpu_assert(false && "Should not be called since there is no packet accessor");
   }
 };
 
@@ -388,7 +388,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
 
   template <typename OutputType>
   static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
-    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    gpu_assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
     const Index num_coeffs = array_prod(self.m_impl.dimensions());
     // Don't crash when we're called with an input tensor of size 0.
     if (num_coeffs == 0) {
@@ -479,7 +479,7 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
     }
   }
 #else // EIGEN_CUDA_ARCH >= 300
-  assert(0 && "Shouldn't be called on unsupported device");
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
 #endif // EIGEN_CUDA_ARCH >= 300
 }
 
@@ -601,7 +601,7 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
 template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
 struct InnerReductionLauncher {
   static EIGEN_DEVICE_FUNC bool run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device");
+    gpu_assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device");
     return true;
   }
 };
@@ -648,7 +648,7 @@ struct InnerReductionLauncher<
 template <typename Self, typename Op>
 struct InnerReductionLauncher<Self, Op, Eigen::half, false> {
   static bool run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should not be called since there is no packet accessor");
+    gpu_assert(false && "Should not be called since there is no packet accessor");
     return true;
   }
 };
@@ -709,7 +709,7 @@ struct InnerReducer<Self, Op, GpuDevice> {
 
   template <typename OutputType>
   static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
-    assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    gpu_assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
     const Index num_coeffs = array_prod(self.m_impl.dimensions());
     // Don't crash when we're called with an input tensor of size 0.
     if (num_coeffs == 0) {
@@ -777,7 +777,7 @@ struct OuterReducer<Self, Op, GpuDevice> {
     EIGEN_DEVICE_FUNC
     #endif
     bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Should only be called to reduce doubles or floats on a gpu device");
+    gpu_assert(false && "Should only be called to reduce doubles or floats on a gpu device");
     return true;
   }
 

From 195c9c054b6beab9d8bf0c71e0556b69c5e61548 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 16:05:07 +0200
Subject: [PATCH 462/680] Print more debug info in gpu_basic

---
 test/gpu_common.h | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/test/gpu_common.h b/test/gpu_common.h
index 60a78ccd7..3aac49e96 100644
--- a/test/gpu_common.h
+++ b/test/gpu_common.h
@@ -100,12 +100,47 @@ void run_and_compare_to_gpu(const Kernel& ker, int n, const Input& in, Output& o
   #endif
 }
 
+struct compile_time_device_info {
+  EIGEN_DEVICE_FUNC
+  void operator()(int /*i*/, const int* /*in*/, int* info) const
+  {
+    #if defined(__CUDA_ARCH__)
+    info[0] = int(__CUDA_ARCH__ +0);
+    #endif
+    #if defined(EIGEN_HIP_DEVICE_COMPILE)
+    info[1] = int(EIGEN_HIP_DEVICE_COMPILE +0);
+    #endif
+  }
+};
 
 void ei_test_init_gpu()
 {
   int device = 0;
   gpuDeviceProp_t deviceProp;
   gpuGetDeviceProperties(&deviceProp, device);
+
+  ArrayXi dummy(1), info(10);
+  info = -1;
+  run_on_gpu(compile_time_device_info(),10,dummy,info);
+
+
+  std::cout << "GPU compile-time info:\n";
+  
+  #ifdef EIGEN_CUDACC
+  std::cout << "  EIGEN_CUDACC:                 " << int(EIGEN_CUDACC) << "\n";
+  #endif
+  
+  #ifdef EIGEN_CUDACC_VER
+  std::cout << "  EIGEN_CUDACC_VER:             " << int(EIGEN_CUDACC_VER) << "\n";
+  #endif
+
+  #ifdef EIGEN_HIPCC
+  std::cout << "  EIGEN_HIPCC:                 " << int(EIGEN_HIPCC) << "\n";
+  #endif
+
+  std::cout << "  EIGEN_CUDA_ARCH:             " << info[0] << "\n";  
+  std::cout << "  EIGEN_HIP_DEVICE_COMPILE:    " << info[1] << "\n";
+
   std::cout << "GPU device info:\n";
   std::cout << "  name:                        " << deviceProp.name << "\n";
   std::cout << "  capability:                  " << deviceProp.major << "." << deviceProp.minor << "\n";

From 1920129d7184d86603e15667989d69ee0e40d07d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 16:05:35 +0200
Subject: [PATCH 463/680] Remove clang warning

---
 unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index fbbc87661..dbcc9d8ac 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -646,6 +646,7 @@ struct igammac_cf_impl {
       case DERIVATIVE:
         return ans * dax_da + dans_da * ax;
       case SAMPLE_DERIVATIVE:
+      default: // this is needed to suppress clang warning
         return -(dans_da + ans * dlogax_da) * x;
     }
   }
@@ -707,6 +708,7 @@ struct igamma_series_impl {
       case DERIVATIVE:
         return ans * dax_da + dans_da * ax;
       case SAMPLE_DERIVATIVE:
+      default: // this is needed to suppress clang warning
         return -(dans_da + ans * dlogax_da) * x / a;
     }
   }

From 20991c3203818e685dbfd89e20408e5f1b254346 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 13 Jul 2018 17:47:28 +0200
Subject: [PATCH 464/680] bug #1571: fix is_convertible<from,to> with "from" a
 reference.

---
 Eigen/src/Core/util/Meta.h |  2 +-
 test/meta.cpp              | 10 +++++++++-
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 748f24b1e..658cfa9eb 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -161,7 +161,7 @@ private:
   static no  test(any_conversion, ...);
 
 public:
-  static From* ms_from;
+  static typename internal::remove_reference<From>::type* ms_from;
 #ifdef __INTEL_COMPILER
   #pragma warning push
   #pragma warning ( disable : 2259 )
diff --git a/test/meta.cpp b/test/meta.cpp
index 4904f3ba4..34c5520a6 100644
--- a/test/meta.cpp
+++ b/test/meta.cpp
@@ -80,7 +80,15 @@ void test_meta()
   STATIC_CHECK(( internal::is_convertible<double,std::complex<double> >::value ));
   STATIC_CHECK((!internal::is_convertible<std::complex<double>,double>::value ));
   STATIC_CHECK(( internal::is_convertible<Array33f,Matrix3f>::value ));
-//   VERIFY((!internal::is_convertible<Matrix3f,Matrix3d>::value )); //does not work because the conversion is prevented by a static assertion
+  STATIC_CHECK(( internal::is_convertible<Matrix3f&,Matrix3f>::value ));
+  STATIC_CHECK(( internal::is_convertible<Matrix3f&,Matrix3f&>::value ));
+  STATIC_CHECK(( internal::is_convertible<Matrix3f&,const Matrix3f&>::value ));
+  STATIC_CHECK(( internal::is_convertible<const Matrix3f&,Matrix3f>::value ));
+  STATIC_CHECK(( internal::is_convertible<const Matrix3f&,const Matrix3f&>::value ));
+  STATIC_CHECK((!internal::is_convertible<const Matrix3f&,Matrix3f&>::value ));
+  STATIC_CHECK((!internal::is_convertible<const Matrix3f,Matrix3f&>::value ));
+  STATIC_CHECK(( internal::is_convertible<Matrix3f,Matrix3f&>::value )); // std::is_convertible returns false here though Matrix3f from; Matrix3f& to = from; is valid.
+  //STATIC_CHECK((!internal::is_convertible<Matrix3f,Matrix3d>::value )); //does not work because the conversion is prevented by a static assertion
   STATIC_CHECK((!internal::is_convertible<Array33f,int>::value ));
   STATIC_CHECK((!internal::is_convertible<MatrixXf,float>::value ));
   {

From 02a9443db9b11ea959a0f65005506e21404ef933 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 13 Jul 2018 11:03:04 -0700
Subject: [PATCH 465/680] Clang produces incorrect Thumb2 assembler when using
 alloca. Don't define EIGEN_ALLOCA when generating Thumb with clang.

---
 Eigen/src/Core/util/Memory.h | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 85bc75da6..1805041c7 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -77,7 +77,7 @@ inline void throw_std_bad_alloc()
     // "no overloaded function has restriction specifiers that are compatible with the ambient context"
     //
     // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
-    // the same on "operator new" 
+    // the same on "operator new"
     // Reverting code back to the old version in this #if block for the hipcc compiler
     //
     new int[huge];
@@ -169,13 +169,13 @@ EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
 
   void *result;
   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
-  
+
     #if defined(EIGEN_HIP_DEVICE_COMPILE)
     result = ::malloc(size);
     #else
     result = std::malloc(size);
     #endif
-    
+
     #if EIGEN_DEFAULT_ALIGN_BYTES==16
     eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator.");
     #endif
@@ -199,7 +199,7 @@ EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
     #else
     std::free(ptr);
     #endif
-    
+
   #else
     handmade_aligned_free(ptr);
   #endif
@@ -248,7 +248,7 @@ template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std:
   #else
   void *result = std::malloc(size);
   #endif
-  
+
   if(!result && size)
     throw_std_bad_alloc();
   return result;
@@ -588,6 +588,15 @@ template<typename T> struct smart_memmove_helper<T,false> {
   #endif
 #endif
 
+// With clang -Oz -mthumb, alloca changes the stack pointer in a way that is
+// not allowed in Thumb2. -DEIGEN_STACK_ALLOCATION_LIMIT=0 doesn't work because
+// the compiler still emits bad code because stack allocation checks use "<=".
+// TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
+// is fixed.
+#if defined(__clang__) && defined(__ANDROID__) && defined(__thumb__)
+  #undef EIGEN_ALLOCA
+#endif
+
 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
 template<typename T> class aligned_stack_memory_handler : noncopyable
@@ -701,7 +710,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 } // end namespace internal
 
 /** \internal
-  * 
+  *
   * The macro ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) declares, allocates,
   * and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
   * if the size in bytes is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
@@ -716,7 +725,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
   * }
   * \endcode
   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
-  * 
+  *
   * The macro ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) is analogue to
   * \code
   *   typename internal::nested_eval<XPRT_T,N>::type NAME(XPR);

From 4a3952fd55a213baefeeebcae50366226952acce Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 13 Jul 2018 11:24:07 -0700
Subject: [PATCH 466/680] Relax the condition to not only work on Android.

---
 Eigen/src/Core/util/Memory.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 1805041c7..956f64d1d 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -593,7 +593,7 @@ template<typename T> struct smart_memmove_helper<T,false> {
 // the compiler still emits bad code because stack allocation checks use "<=".
 // TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
 // is fixed.
-#if defined(__clang__) && defined(__ANDROID__) && defined(__thumb__)
+#if defined(__clang__) && defined(__thumb__)
   #undef EIGEN_ALLOCA
 #endif
 

From 4222550e173c344ef60e14cb2a18fc0bf8968a51 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 13 Jul 2018 16:12:38 -0700
Subject: [PATCH 467/680] Optimize the case where broadcasting is a no-op.

---
 .../CXX11/src/Tensor/TensorBroadcasting.h     | 30 ++++++++++++++-----
 1 file changed, 23 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index b35b36475..3b934a5fd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -105,7 +105,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
-  bool nByOne = false, oneByN = false;
+  bool isCopy= false, nByOne = false, oneByN = false;
 
   enum {
     IsAligned = true,
@@ -123,9 +123,13 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
     const InputDimensions& input_dims = m_impl.dimensions();
     const Broadcast& broadcast = op.broadcast();
+    isCopy = true;
     for (int i = 0; i < NumDims; ++i) {
       eigen_assert(input_dims[i] > 0);
-      m_dimensions[i] = input_dims[i] * broadcast[i];
+      m_dimensions[i] = input_dims[i] * m_broadcast[i];
+      if (m_broadcast[i] != 1) {
+        isCopy = false;
+      }
     }
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
@@ -197,9 +201,17 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     }
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      return coeffColMajor(index);
+      if (isCopy) {
+        return m_impl.coeff(index);
+      } else {
+        return coeffColMajor(index);
+      }
     } else {
-      return coeffRowMajor(index);
+      if (isCopy) {
+        return m_impl.coeff(index);
+      } else {
+        return coeffRowMajor(index);
+      }
     }
   }
 
@@ -272,7 +284,9 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     }
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      if (oneByN && !nByOne) {
+      if (isCopy) {
+        return m_impl.template packet<LoadMode>(index);
+      } else if (oneByN && !nByOne) {
         return packetNByOne<LoadMode>(index);
       } else if (!oneByN && nByOne) {
         return packetOneByN<LoadMode>(index);
@@ -282,7 +296,9 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
         return packetColMajor<LoadMode>(index);
       }
     } else {
-      if (oneByN && !nByOne) {
+      if (isCopy) {
+        return m_impl.template packet<LoadMode>(index);
+      } else if (oneByN && !nByOne) {
         return packetOneByN<LoadMode>(index);
       } else if (!oneByN && nByOne) {
         return packetNByOne<LoadMode>(index);
@@ -516,7 +532,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
   costPerCoeff(bool vectorized) const {
     double compute_cost = TensorOpCost::AddCost<Index>();
-    if (NumDims > 0) {
+    if (!isCopy && NumDims > 0) {
       for (int i = NumDims - 1; i > 0; --i) {
         compute_cost += TensorOpCost::DivCost<Index>();
         if (internal::index_statically_eq<Broadcast>(i, 1)) {

From 3a9cf4e290c99588a22fcc9d288968f669809306 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 13 Jul 2018 16:24:48 -0700
Subject: [PATCH 468/680] Get rid of alias for m_broadcast.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 3b934a5fd..278689915 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -122,7 +122,6 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     // tensor with N >= 1 of 1 element first and then broadcast.
     EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
     const InputDimensions& input_dims = m_impl.dimensions();
-    const Broadcast& broadcast = op.broadcast();
     isCopy = true;
     for (int i = 0; i < NumDims; ++i) {
       eigen_assert(input_dims[i] > 0);
@@ -151,7 +150,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     if (input_dims[0] == 1) {
       oneByN = true;
       for (int i = 1; i < NumDims; ++i) {
-        if (broadcast[i] != 1) {
+        if (m_broadcast[i] != 1) {
           oneByN = false;
           break;
         }
@@ -159,7 +158,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     } else if (input_dims[NumDims-1] == 1) {
       nByOne = true;
       for (int i = 0; i < NumDims-1; ++i) {
-        if (broadcast[i] != 1) {
+        if (m_broadcast[i] != 1) {
           nByOne = false;
           break;
         }
@@ -173,7 +172,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
         nByOne = true;
         oneByN = true;
         for (int i = 1; i < NumDims-1; ++i) {
-          if (broadcast[i] != 1) {
+          if (m_broadcast[i] != 1) {
             nByOne = false;
             oneByN = false;
             break;

From a87cff20df5129c364c4675f05ceb6d451188f6c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 14 Jul 2018 09:38:49 +0200
Subject: [PATCH 469/680] Fix GeneralizedEigenSolver when requesting for
 eigenvalues only.

---
 Eigen/src/Eigenvalues/GeneralizedEigenSolver.h | 5 ++---
 test/eigensolver_generalized_real.cpp          | 7 +++++++
 2 files changed, 9 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
index 36a91dffc..87d789b3f 100644
--- a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
+++ b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
@@ -311,7 +311,6 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp
     // Aliases:
     Map<VectorType> v(reinterpret_cast<Scalar*>(m_tmp.data()), size);
     ComplexVectorType &cv = m_tmp;
-    const MatrixType &mZ = m_realQZ.matrixZ();
     const MatrixType &mS = m_realQZ.matrixS();
     const MatrixType &mT = m_realQZ.matrixT();
 
@@ -351,7 +350,7 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp
               }
             }
           }
-          m_eivec.col(i).real().noalias() = mZ.transpose() * v;
+          m_eivec.col(i).real().noalias() = m_realQZ.matrixZ().transpose() * v;
           m_eivec.col(i).real().normalize();
           m_eivec.col(i).imag().setConstant(0);
         }
@@ -400,7 +399,7 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp
                               / (alpha*mT.coeffRef(j,j) - static_cast<Scalar>(beta*mS.coeffRef(j,j)));
             }
           }
-          m_eivec.col(i+1).noalias() = (mZ.transpose() * cv);
+          m_eivec.col(i+1).noalias() = (m_realQZ.matrixZ().transpose() * cv);
           m_eivec.col(i+1).normalize();
           m_eivec.col(i) = m_eivec.col(i+1).conjugate();
         }
diff --git a/test/eigensolver_generalized_real.cpp b/test/eigensolver_generalized_real.cpp
index 12a15ab81..9dd44c89d 100644
--- a/test/eigensolver_generalized_real.cpp
+++ b/test/eigensolver_generalized_real.cpp
@@ -76,6 +76,13 @@ template<typename MatrixType> void generalized_eigensolver_real(const MatrixType
     GeneralizedEigenSolver<MatrixType> eig2(a.adjoint() * a,b.adjoint() * b);
     eig2.compute(a.adjoint() * a,b.adjoint() * b);
   }
+
+  // check without eigenvectors
+  {
+    GeneralizedEigenSolver<MatrixType> eig1(spdA, spdB, true);
+    GeneralizedEigenSolver<MatrixType> eig2(spdA, spdB, false);
+    VERIFY_IS_APPROX(eig1.eigenvalues(), eig2.eigenvalues());
+  }
 }
 
 void test_eigensolver_generalized_real()

From f2b52f9946720927fdb086da6e6d85dbed7a848d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 16 Jul 2018 17:59:30 +0200
Subject: [PATCH 470/680] Add the cmake option "EIGEN_DASHBOARD_BUILD_TARGET"
 to control the build target in dashboard mode (e.g., ctest -D Experimental)

---
 cmake/EigenConfigureTesting.cmake | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake
index a2a4f54b9..b02729fc8 100644
--- a/cmake/EigenConfigureTesting.cmake
+++ b/cmake/EigenConfigureTesting.cmake
@@ -11,13 +11,14 @@ add_custom_target(buildtests)
 add_custom_target(check COMMAND "ctest")
 add_dependencies(check buildtests)
 
-# check whether /bin/bash exists
-find_file(EIGEN_BIN_BASH_EXISTS "/bin/bash" PATHS "/" NO_DEFAULT_PATH)
+# check whether /bin/bash exists (disabled as not used anymore)
+# find_file(EIGEN_BIN_BASH_EXISTS "/bin/bash" PATHS "/" NO_DEFAULT_PATH)
 
 # This call activates testing and generates the DartConfiguration.tcl
 include(CTest)
 
 set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command of unit tests")
+set(EIGEN_DASHBOARD_BUILD_TARGET "buildtests" CACHE STRING "Target to be built in dashboard mode, default is buildtests")
 
 # Overwrite default DartConfiguration.tcl such that ctest can build our unit tests.
 # Recall that our unit tests are not in the "all" target, so we have to explicitly ask ctest to build our custom 'buildtests' target.
@@ -28,7 +29,7 @@ string(REGEX MATCH "MakeCommand:.*-- (.*)\nDefaultCTestConfigurationType" EIGEN_
 if(NOT CMAKE_MATCH_1)
 string(REGEX MATCH "MakeCommand:.*[^c]make (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE})
 endif()
-string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config \"\${CTEST_CONFIGURATION_TYPE}\" -- ${CMAKE_MATCH_1} ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType"
+string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target ${EIGEN_DASHBOARD_BUILD_TARGET} --config \"\${CTEST_CONFIGURATION_TYPE}\" -- ${CMAKE_MATCH_1} ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType"
        EIGEN_DART_CONFIG_FILE2 ${EIGEN_DART_CONFIG_FILE})
 file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" ${EIGEN_DART_CONFIG_FILE2})
 

From 901c7d31f026cce75816cf5d954e42f22f6f5691 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 16 Jul 2018 18:35:05 +0200
Subject: [PATCH 471/680] Fix usage of EIGEN_SPLIT_LARGE_TESTS=ON: some unit
 tests, such as indexed_view have to be split unconditionally.

---
 cmake/EigenTesting.cmake        | 13 ++++++++++---
 test/CMakeLists.txt             |  2 +-
 unsupported/test/CMakeLists.txt |  2 +-
 3 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 58f3697d3..5ad57ad6c 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -108,7 +108,7 @@ macro(ei_add_test_internal testname testname_with_suffix)
 
   add_test(${testname_with_suffix} "${targetname}")
 
-  # Specify target and test labels accoirding to EIGEN_CURRENT_SUBPROJECT
+  # Specify target and test labels according to EIGEN_CURRENT_SUBPROJECT
   get_property(current_subproject GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT)
   if ((current_subproject) AND (NOT (current_subproject STREQUAL "")))
     set_property(TARGET ${targetname} PROPERTY LABELS "Build${current_subproject}")
@@ -276,12 +276,19 @@ macro(ei_add_test testname)
   endif()
 
   file(READ "${filename}" test_source)
-  set(parts 0)
   string(REGEX MATCHALL "CALL_SUBTEST_[0-9]+|EIGEN_TEST_PART_[0-9]+|EIGEN_SUFFIXES(;[0-9]+)+"
          occurrences "${test_source}")
   string(REGEX REPLACE "CALL_SUBTEST_|EIGEN_TEST_PART_|EIGEN_SUFFIXES" "" suffixes "${occurrences}")
   list(REMOVE_DUPLICATES suffixes)
-  if(EIGEN_SPLIT_LARGE_TESTS AND suffixes)
+  set(explicit_suffixes "")
+  if( (NOT EIGEN_SPLIT_LARGE_TESTS) AND suffixes)
+    # Check whether we have EIGEN_TEST_PART_* statements, in which case we likely must enforce splitting.
+    # For instance, indexed_view activate a different c++ version for each part.
+    string(REGEX MATCHALL "EIGEN_TEST_PART_[0-9]+" occurrences "${test_source}")
+    string(REGEX REPLACE "EIGEN_TEST_PART_" "" explicit_suffixes "${occurrences}")
+    list(REMOVE_DUPLICATES explicit_suffixes)
+  endif()
+  if( (EIGEN_SPLIT_LARGE_TESTS AND suffixes) OR explicit_suffixes)
     add_custom_target(${testname})
     foreach(suffix ${suffixes})
       ei_add_test_internal(${testname} ${testname}_${suffix}
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index d312d16e4..cabd2a109 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -4,7 +4,7 @@ if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
   file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
   foreach(i RANGE 1 999)
     file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h
-      "#ifdef EIGEN_TEST_PART_${i}\n"
+      "#if defined(EIGEN_TEST_PART_${i}) || defined(EIGEN_TEST_PART_ALL)\n"
       "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n"
       "#else\n"
       "#define CALL_SUBTEST_${i}(FUNC)\n"
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 76d6f5e5b..02fb890c6 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -4,7 +4,7 @@ if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
   file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
   foreach(i RANGE 1 999)
     file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h
-      "#ifdef EIGEN_TEST_PART_${i}\n"
+      "#if defined(EIGEN_TEST_PART_${i}) || defined(EIGEN_TEST_PART_ALL)\n"
       "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n"
       "#else\n"
       "#define CALL_SUBTEST_${i}(FUNC)\n"

From add57574881f7389ce4a3f82934907420945b0f4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 16 Jul 2018 18:55:40 +0200
Subject: [PATCH 472/680] Simplify handling and non-splitted tests and include
 split_test_helper.h instead of re-generating it. This also allows us to
 modify it without breaking existing build folder.

---
 cmake/EigenTesting.cmake                |   11 +-
 scripts/eigen_gen_split_test_help.cmake |   11 +
 test/CMakeLists.txt                     |   17 +-
 test/split_test_helper.h                | 5994 +++++++++++++++++++++++
 unsupported/test/CMakeLists.txt         |   17 +-
 5 files changed, 6016 insertions(+), 34 deletions(-)
 create mode 100644 scripts/eigen_gen_split_test_help.cmake
 create mode 100644 test/split_test_helper.h

diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index 5ad57ad6c..d60873554 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -295,14 +295,9 @@ macro(ei_add_test testname)
         "${ARGV1} -DEIGEN_TEST_PART_${suffix}=1" "${ARGV2}")
       add_dependencies(${testname} ${testname}_${suffix})
     endforeach(suffix)
-  else(EIGEN_SPLIT_LARGE_TESTS AND suffixes)
-    set(symbols_to_enable_all_parts "")
-    foreach(suffix ${suffixes})
-      set(symbols_to_enable_all_parts
-        "${symbols_to_enable_all_parts} -DEIGEN_TEST_PART_${suffix}=1")
-    endforeach(suffix)
-    ei_add_test_internal(${testname} ${testname} "${ARGV1} ${symbols_to_enable_all_parts}" "${ARGV2}")
-  endif(EIGEN_SPLIT_LARGE_TESTS AND suffixes)
+  else()
+    ei_add_test_internal(${testname} ${testname} "${ARGV1} -DEIGEN_TEST_PART_ALL=1" "${ARGV2}")
+  endif()
 endmacro(ei_add_test)
 
 macro(ei_add_test_sycl testname)
diff --git a/scripts/eigen_gen_split_test_help.cmake b/scripts/eigen_gen_split_test_help.cmake
new file mode 100644
index 000000000..e43f5aabe
--- /dev/null
+++ b/scripts/eigen_gen_split_test_help.cmake
@@ -0,0 +1,11 @@
+#!cmake -P
+file(WRITE split_test_helper.h "")
+foreach(i RANGE 1 999)
+  file(APPEND split_test_helper.h
+    "#if defined(EIGEN_TEST_PART_${i}) || defined(EIGEN_TEST_PART_ALL)\n"
+    "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n"
+    "#else\n"
+    "#define CALL_SUBTEST_${i}(FUNC)\n"
+    "#endif\n\n"
+  )
+endforeach()
\ No newline at end of file
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index cabd2a109..aacfa22bb 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -1,16 +1,7 @@
-# generate split test header file only if it does not yet exist
-# in order to prevent a rebuild every time cmake is configured
-if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)  
-  file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
-  foreach(i RANGE 1 999)
-    file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h
-      "#if defined(EIGEN_TEST_PART_${i}) || defined(EIGEN_TEST_PART_ALL)\n"
-      "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n"
-      "#else\n"
-      "#define CALL_SUBTEST_${i}(FUNC)\n"
-      "#endif\n\n"
-    )
-  endforeach()
+# The file split_test_helper.h was generated at first run,
+# it is now included in test/
+if(EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
+  file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
 endif()
 
 # check if we have a Fortran compiler
diff --git a/test/split_test_helper.h b/test/split_test_helper.h
new file mode 100644
index 000000000..82e82aaef
--- /dev/null
+++ b/test/split_test_helper.h
@@ -0,0 +1,5994 @@
+#if defined(EIGEN_TEST_PART_1) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_1(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_1(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_2) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_2(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_2(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_3) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_3(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_3(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_4) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_4(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_4(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_5) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_5(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_5(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_6) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_6(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_6(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_7) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_7(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_7(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_8) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_8(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_8(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_9) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_9(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_9(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_10) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_10(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_10(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_11) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_11(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_11(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_12) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_12(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_12(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_13) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_13(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_13(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_14) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_14(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_14(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_15) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_15(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_15(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_16) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_16(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_16(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_17) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_17(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_17(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_18) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_18(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_18(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_19) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_19(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_19(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_20) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_20(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_20(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_21) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_21(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_21(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_22) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_22(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_22(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_23) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_23(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_23(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_24) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_24(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_24(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_25) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_25(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_25(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_26) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_26(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_26(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_27) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_27(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_27(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_28) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_28(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_28(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_29) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_29(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_29(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_30) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_30(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_30(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_31) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_31(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_31(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_32) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_32(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_32(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_33) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_33(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_33(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_34) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_34(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_34(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_35) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_35(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_35(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_36) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_36(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_36(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_37) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_37(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_37(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_38) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_38(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_38(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_39) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_39(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_39(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_40) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_40(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_40(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_41) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_41(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_41(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_42) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_42(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_42(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_43) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_43(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_43(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_44) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_44(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_44(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_45) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_45(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_45(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_46) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_46(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_46(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_47) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_47(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_47(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_48) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_48(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_48(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_49) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_49(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_49(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_50) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_50(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_50(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_51) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_51(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_51(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_52) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_52(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_52(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_53) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_53(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_53(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_54) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_54(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_54(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_55) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_55(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_55(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_56) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_56(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_56(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_57) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_57(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_57(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_58) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_58(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_58(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_59) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_59(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_59(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_60) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_60(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_60(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_61) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_61(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_61(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_62) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_62(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_62(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_63) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_63(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_63(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_64) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_64(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_64(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_65) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_65(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_65(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_66) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_66(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_66(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_67) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_67(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_67(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_68) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_68(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_68(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_69) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_69(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_69(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_70) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_70(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_70(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_71) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_71(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_71(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_72) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_72(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_72(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_73) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_73(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_73(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_74) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_74(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_74(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_75) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_75(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_75(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_76) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_76(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_76(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_77) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_77(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_77(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_78) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_78(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_78(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_79) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_79(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_79(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_80) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_80(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_80(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_81) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_81(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_81(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_82) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_82(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_82(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_83) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_83(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_83(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_84) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_84(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_84(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_85) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_85(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_85(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_86) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_86(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_86(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_87) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_87(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_87(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_88) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_88(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_88(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_89) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_89(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_89(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_90) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_90(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_90(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_91) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_91(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_91(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_92) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_92(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_92(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_93) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_93(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_93(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_94) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_94(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_94(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_95) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_95(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_95(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_96) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_96(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_96(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_97) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_97(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_97(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_98) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_98(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_98(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_99) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_99(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_99(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_100) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_100(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_100(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_101) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_101(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_101(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_102) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_102(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_102(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_103) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_103(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_103(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_104) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_104(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_104(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_105) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_105(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_105(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_106) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_106(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_106(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_107) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_107(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_107(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_108) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_108(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_108(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_109) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_109(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_109(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_110) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_110(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_110(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_111) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_111(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_111(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_112) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_112(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_112(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_113) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_113(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_113(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_114) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_114(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_114(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_115) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_115(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_115(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_116) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_116(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_116(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_117) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_117(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_117(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_118) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_118(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_118(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_119) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_119(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_119(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_120) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_120(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_120(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_121) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_121(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_121(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_122) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_122(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_122(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_123) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_123(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_123(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_124) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_124(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_124(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_125) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_125(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_125(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_126) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_126(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_126(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_127) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_127(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_127(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_128) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_128(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_128(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_129) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_129(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_129(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_130) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_130(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_130(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_131) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_131(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_131(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_132) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_132(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_132(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_133) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_133(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_133(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_134) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_134(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_134(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_135) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_135(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_135(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_136) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_136(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_136(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_137) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_137(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_137(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_138) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_138(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_138(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_139) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_139(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_139(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_140) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_140(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_140(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_141) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_141(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_141(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_142) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_142(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_142(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_143) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_143(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_143(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_144) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_144(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_144(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_145) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_145(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_145(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_146) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_146(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_146(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_147) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_147(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_147(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_148) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_148(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_148(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_149) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_149(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_149(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_150) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_150(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_150(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_151) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_151(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_151(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_152) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_152(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_152(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_153) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_153(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_153(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_154) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_154(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_154(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_155) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_155(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_155(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_156) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_156(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_156(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_157) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_157(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_157(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_158) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_158(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_158(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_159) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_159(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_159(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_160) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_160(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_160(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_161) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_161(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_161(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_162) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_162(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_162(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_163) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_163(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_163(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_164) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_164(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_164(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_165) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_165(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_165(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_166) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_166(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_166(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_167) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_167(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_167(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_168) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_168(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_168(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_169) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_169(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_169(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_170) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_170(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_170(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_171) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_171(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_171(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_172) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_172(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_172(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_173) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_173(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_173(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_174) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_174(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_174(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_175) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_175(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_175(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_176) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_176(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_176(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_177) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_177(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_177(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_178) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_178(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_178(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_179) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_179(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_179(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_180) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_180(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_180(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_181) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_181(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_181(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_182) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_182(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_182(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_183) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_183(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_183(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_184) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_184(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_184(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_185) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_185(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_185(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_186) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_186(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_186(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_187) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_187(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_187(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_188) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_188(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_188(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_189) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_189(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_189(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_190) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_190(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_190(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_191) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_191(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_191(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_192) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_192(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_192(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_193) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_193(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_193(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_194) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_194(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_194(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_195) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_195(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_195(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_196) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_196(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_196(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_197) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_197(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_197(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_198) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_198(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_198(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_199) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_199(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_199(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_200) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_200(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_200(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_201) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_201(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_201(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_202) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_202(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_202(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_203) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_203(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_203(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_204) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_204(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_204(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_205) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_205(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_205(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_206) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_206(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_206(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_207) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_207(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_207(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_208) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_208(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_208(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_209) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_209(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_209(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_210) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_210(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_210(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_211) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_211(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_211(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_212) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_212(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_212(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_213) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_213(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_213(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_214) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_214(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_214(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_215) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_215(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_215(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_216) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_216(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_216(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_217) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_217(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_217(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_218) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_218(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_218(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_219) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_219(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_219(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_220) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_220(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_220(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_221) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_221(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_221(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_222) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_222(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_222(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_223) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_223(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_223(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_224) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_224(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_224(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_225) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_225(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_225(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_226) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_226(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_226(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_227) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_227(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_227(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_228) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_228(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_228(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_229) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_229(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_229(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_230) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_230(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_230(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_231) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_231(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_231(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_232) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_232(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_232(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_233) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_233(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_233(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_234) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_234(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_234(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_235) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_235(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_235(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_236) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_236(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_236(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_237) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_237(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_237(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_238) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_238(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_238(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_239) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_239(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_239(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_240) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_240(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_240(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_241) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_241(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_241(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_242) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_242(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_242(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_243) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_243(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_243(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_244) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_244(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_244(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_245) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_245(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_245(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_246) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_246(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_246(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_247) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_247(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_247(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_248) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_248(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_248(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_249) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_249(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_249(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_250) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_250(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_250(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_251) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_251(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_251(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_252) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_252(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_252(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_253) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_253(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_253(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_254) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_254(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_254(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_255) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_255(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_255(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_256) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_256(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_256(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_257) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_257(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_257(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_258) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_258(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_258(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_259) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_259(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_259(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_260) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_260(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_260(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_261) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_261(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_261(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_262) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_262(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_262(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_263) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_263(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_263(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_264) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_264(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_264(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_265) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_265(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_265(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_266) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_266(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_266(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_267) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_267(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_267(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_268) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_268(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_268(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_269) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_269(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_269(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_270) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_270(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_270(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_271) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_271(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_271(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_272) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_272(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_272(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_273) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_273(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_273(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_274) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_274(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_274(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_275) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_275(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_275(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_276) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_276(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_276(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_277) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_277(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_277(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_278) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_278(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_278(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_279) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_279(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_279(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_280) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_280(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_280(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_281) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_281(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_281(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_282) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_282(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_282(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_283) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_283(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_283(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_284) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_284(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_284(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_285) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_285(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_285(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_286) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_286(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_286(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_287) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_287(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_287(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_288) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_288(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_288(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_289) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_289(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_289(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_290) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_290(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_290(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_291) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_291(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_291(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_292) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_292(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_292(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_293) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_293(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_293(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_294) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_294(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_294(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_295) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_295(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_295(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_296) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_296(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_296(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_297) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_297(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_297(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_298) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_298(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_298(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_299) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_299(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_299(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_300) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_300(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_300(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_301) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_301(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_301(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_302) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_302(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_302(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_303) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_303(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_303(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_304) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_304(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_304(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_305) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_305(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_305(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_306) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_306(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_306(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_307) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_307(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_307(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_308) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_308(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_308(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_309) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_309(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_309(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_310) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_310(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_310(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_311) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_311(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_311(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_312) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_312(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_312(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_313) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_313(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_313(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_314) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_314(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_314(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_315) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_315(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_315(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_316) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_316(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_316(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_317) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_317(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_317(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_318) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_318(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_318(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_319) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_319(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_319(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_320) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_320(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_320(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_321) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_321(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_321(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_322) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_322(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_322(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_323) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_323(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_323(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_324) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_324(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_324(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_325) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_325(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_325(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_326) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_326(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_326(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_327) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_327(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_327(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_328) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_328(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_328(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_329) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_329(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_329(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_330) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_330(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_330(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_331) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_331(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_331(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_332) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_332(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_332(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_333) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_333(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_333(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_334) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_334(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_334(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_335) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_335(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_335(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_336) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_336(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_336(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_337) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_337(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_337(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_338) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_338(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_338(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_339) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_339(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_339(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_340) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_340(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_340(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_341) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_341(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_341(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_342) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_342(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_342(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_343) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_343(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_343(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_344) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_344(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_344(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_345) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_345(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_345(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_346) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_346(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_346(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_347) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_347(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_347(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_348) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_348(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_348(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_349) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_349(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_349(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_350) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_350(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_350(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_351) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_351(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_351(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_352) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_352(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_352(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_353) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_353(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_353(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_354) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_354(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_354(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_355) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_355(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_355(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_356) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_356(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_356(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_357) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_357(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_357(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_358) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_358(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_358(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_359) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_359(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_359(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_360) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_360(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_360(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_361) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_361(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_361(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_362) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_362(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_362(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_363) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_363(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_363(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_364) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_364(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_364(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_365) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_365(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_365(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_366) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_366(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_366(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_367) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_367(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_367(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_368) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_368(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_368(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_369) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_369(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_369(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_370) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_370(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_370(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_371) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_371(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_371(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_372) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_372(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_372(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_373) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_373(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_373(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_374) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_374(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_374(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_375) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_375(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_375(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_376) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_376(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_376(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_377) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_377(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_377(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_378) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_378(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_378(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_379) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_379(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_379(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_380) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_380(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_380(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_381) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_381(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_381(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_382) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_382(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_382(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_383) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_383(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_383(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_384) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_384(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_384(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_385) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_385(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_385(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_386) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_386(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_386(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_387) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_387(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_387(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_388) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_388(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_388(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_389) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_389(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_389(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_390) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_390(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_390(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_391) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_391(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_391(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_392) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_392(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_392(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_393) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_393(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_393(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_394) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_394(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_394(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_395) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_395(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_395(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_396) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_396(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_396(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_397) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_397(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_397(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_398) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_398(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_398(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_399) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_399(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_399(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_400) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_400(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_400(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_401) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_401(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_401(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_402) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_402(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_402(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_403) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_403(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_403(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_404) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_404(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_404(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_405) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_405(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_405(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_406) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_406(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_406(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_407) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_407(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_407(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_408) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_408(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_408(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_409) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_409(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_409(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_410) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_410(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_410(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_411) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_411(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_411(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_412) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_412(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_412(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_413) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_413(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_413(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_414) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_414(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_414(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_415) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_415(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_415(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_416) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_416(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_416(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_417) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_417(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_417(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_418) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_418(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_418(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_419) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_419(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_419(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_420) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_420(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_420(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_421) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_421(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_421(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_422) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_422(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_422(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_423) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_423(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_423(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_424) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_424(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_424(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_425) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_425(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_425(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_426) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_426(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_426(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_427) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_427(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_427(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_428) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_428(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_428(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_429) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_429(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_429(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_430) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_430(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_430(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_431) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_431(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_431(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_432) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_432(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_432(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_433) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_433(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_433(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_434) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_434(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_434(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_435) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_435(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_435(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_436) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_436(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_436(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_437) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_437(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_437(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_438) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_438(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_438(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_439) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_439(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_439(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_440) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_440(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_440(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_441) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_441(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_441(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_442) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_442(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_442(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_443) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_443(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_443(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_444) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_444(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_444(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_445) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_445(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_445(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_446) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_446(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_446(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_447) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_447(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_447(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_448) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_448(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_448(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_449) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_449(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_449(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_450) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_450(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_450(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_451) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_451(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_451(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_452) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_452(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_452(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_453) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_453(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_453(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_454) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_454(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_454(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_455) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_455(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_455(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_456) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_456(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_456(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_457) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_457(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_457(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_458) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_458(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_458(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_459) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_459(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_459(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_460) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_460(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_460(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_461) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_461(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_461(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_462) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_462(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_462(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_463) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_463(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_463(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_464) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_464(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_464(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_465) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_465(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_465(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_466) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_466(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_466(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_467) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_467(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_467(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_468) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_468(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_468(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_469) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_469(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_469(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_470) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_470(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_470(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_471) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_471(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_471(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_472) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_472(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_472(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_473) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_473(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_473(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_474) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_474(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_474(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_475) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_475(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_475(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_476) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_476(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_476(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_477) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_477(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_477(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_478) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_478(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_478(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_479) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_479(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_479(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_480) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_480(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_480(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_481) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_481(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_481(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_482) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_482(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_482(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_483) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_483(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_483(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_484) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_484(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_484(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_485) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_485(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_485(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_486) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_486(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_486(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_487) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_487(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_487(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_488) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_488(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_488(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_489) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_489(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_489(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_490) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_490(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_490(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_491) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_491(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_491(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_492) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_492(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_492(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_493) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_493(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_493(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_494) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_494(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_494(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_495) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_495(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_495(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_496) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_496(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_496(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_497) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_497(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_497(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_498) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_498(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_498(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_499) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_499(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_499(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_500) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_500(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_500(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_501) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_501(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_501(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_502) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_502(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_502(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_503) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_503(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_503(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_504) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_504(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_504(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_505) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_505(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_505(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_506) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_506(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_506(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_507) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_507(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_507(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_508) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_508(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_508(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_509) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_509(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_509(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_510) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_510(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_510(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_511) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_511(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_511(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_512) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_512(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_512(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_513) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_513(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_513(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_514) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_514(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_514(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_515) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_515(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_515(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_516) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_516(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_516(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_517) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_517(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_517(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_518) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_518(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_518(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_519) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_519(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_519(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_520) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_520(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_520(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_521) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_521(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_521(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_522) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_522(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_522(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_523) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_523(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_523(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_524) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_524(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_524(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_525) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_525(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_525(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_526) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_526(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_526(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_527) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_527(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_527(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_528) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_528(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_528(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_529) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_529(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_529(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_530) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_530(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_530(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_531) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_531(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_531(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_532) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_532(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_532(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_533) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_533(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_533(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_534) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_534(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_534(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_535) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_535(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_535(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_536) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_536(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_536(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_537) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_537(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_537(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_538) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_538(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_538(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_539) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_539(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_539(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_540) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_540(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_540(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_541) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_541(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_541(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_542) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_542(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_542(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_543) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_543(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_543(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_544) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_544(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_544(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_545) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_545(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_545(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_546) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_546(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_546(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_547) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_547(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_547(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_548) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_548(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_548(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_549) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_549(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_549(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_550) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_550(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_550(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_551) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_551(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_551(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_552) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_552(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_552(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_553) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_553(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_553(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_554) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_554(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_554(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_555) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_555(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_555(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_556) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_556(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_556(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_557) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_557(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_557(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_558) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_558(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_558(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_559) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_559(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_559(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_560) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_560(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_560(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_561) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_561(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_561(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_562) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_562(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_562(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_563) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_563(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_563(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_564) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_564(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_564(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_565) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_565(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_565(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_566) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_566(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_566(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_567) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_567(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_567(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_568) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_568(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_568(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_569) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_569(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_569(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_570) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_570(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_570(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_571) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_571(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_571(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_572) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_572(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_572(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_573) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_573(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_573(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_574) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_574(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_574(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_575) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_575(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_575(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_576) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_576(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_576(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_577) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_577(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_577(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_578) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_578(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_578(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_579) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_579(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_579(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_580) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_580(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_580(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_581) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_581(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_581(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_582) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_582(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_582(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_583) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_583(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_583(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_584) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_584(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_584(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_585) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_585(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_585(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_586) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_586(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_586(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_587) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_587(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_587(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_588) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_588(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_588(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_589) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_589(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_589(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_590) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_590(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_590(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_591) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_591(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_591(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_592) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_592(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_592(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_593) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_593(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_593(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_594) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_594(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_594(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_595) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_595(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_595(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_596) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_596(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_596(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_597) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_597(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_597(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_598) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_598(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_598(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_599) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_599(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_599(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_600) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_600(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_600(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_601) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_601(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_601(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_602) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_602(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_602(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_603) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_603(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_603(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_604) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_604(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_604(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_605) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_605(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_605(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_606) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_606(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_606(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_607) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_607(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_607(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_608) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_608(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_608(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_609) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_609(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_609(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_610) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_610(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_610(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_611) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_611(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_611(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_612) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_612(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_612(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_613) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_613(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_613(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_614) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_614(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_614(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_615) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_615(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_615(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_616) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_616(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_616(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_617) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_617(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_617(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_618) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_618(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_618(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_619) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_619(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_619(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_620) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_620(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_620(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_621) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_621(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_621(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_622) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_622(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_622(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_623) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_623(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_623(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_624) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_624(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_624(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_625) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_625(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_625(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_626) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_626(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_626(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_627) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_627(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_627(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_628) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_628(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_628(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_629) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_629(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_629(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_630) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_630(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_630(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_631) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_631(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_631(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_632) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_632(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_632(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_633) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_633(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_633(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_634) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_634(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_634(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_635) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_635(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_635(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_636) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_636(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_636(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_637) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_637(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_637(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_638) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_638(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_638(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_639) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_639(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_639(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_640) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_640(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_640(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_641) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_641(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_641(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_642) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_642(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_642(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_643) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_643(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_643(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_644) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_644(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_644(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_645) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_645(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_645(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_646) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_646(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_646(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_647) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_647(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_647(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_648) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_648(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_648(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_649) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_649(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_649(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_650) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_650(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_650(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_651) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_651(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_651(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_652) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_652(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_652(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_653) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_653(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_653(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_654) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_654(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_654(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_655) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_655(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_655(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_656) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_656(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_656(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_657) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_657(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_657(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_658) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_658(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_658(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_659) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_659(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_659(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_660) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_660(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_660(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_661) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_661(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_661(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_662) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_662(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_662(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_663) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_663(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_663(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_664) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_664(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_664(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_665) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_665(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_665(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_666) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_666(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_666(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_667) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_667(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_667(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_668) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_668(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_668(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_669) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_669(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_669(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_670) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_670(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_670(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_671) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_671(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_671(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_672) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_672(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_672(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_673) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_673(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_673(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_674) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_674(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_674(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_675) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_675(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_675(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_676) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_676(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_676(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_677) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_677(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_677(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_678) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_678(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_678(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_679) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_679(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_679(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_680) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_680(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_680(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_681) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_681(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_681(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_682) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_682(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_682(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_683) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_683(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_683(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_684) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_684(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_684(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_685) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_685(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_685(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_686) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_686(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_686(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_687) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_687(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_687(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_688) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_688(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_688(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_689) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_689(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_689(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_690) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_690(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_690(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_691) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_691(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_691(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_692) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_692(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_692(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_693) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_693(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_693(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_694) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_694(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_694(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_695) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_695(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_695(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_696) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_696(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_696(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_697) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_697(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_697(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_698) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_698(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_698(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_699) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_699(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_699(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_700) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_700(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_700(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_701) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_701(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_701(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_702) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_702(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_702(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_703) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_703(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_703(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_704) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_704(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_704(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_705) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_705(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_705(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_706) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_706(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_706(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_707) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_707(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_707(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_708) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_708(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_708(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_709) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_709(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_709(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_710) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_710(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_710(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_711) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_711(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_711(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_712) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_712(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_712(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_713) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_713(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_713(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_714) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_714(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_714(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_715) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_715(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_715(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_716) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_716(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_716(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_717) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_717(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_717(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_718) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_718(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_718(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_719) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_719(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_719(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_720) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_720(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_720(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_721) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_721(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_721(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_722) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_722(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_722(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_723) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_723(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_723(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_724) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_724(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_724(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_725) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_725(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_725(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_726) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_726(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_726(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_727) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_727(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_727(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_728) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_728(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_728(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_729) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_729(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_729(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_730) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_730(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_730(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_731) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_731(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_731(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_732) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_732(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_732(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_733) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_733(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_733(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_734) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_734(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_734(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_735) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_735(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_735(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_736) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_736(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_736(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_737) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_737(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_737(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_738) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_738(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_738(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_739) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_739(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_739(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_740) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_740(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_740(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_741) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_741(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_741(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_742) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_742(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_742(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_743) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_743(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_743(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_744) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_744(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_744(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_745) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_745(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_745(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_746) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_746(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_746(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_747) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_747(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_747(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_748) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_748(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_748(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_749) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_749(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_749(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_750) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_750(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_750(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_751) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_751(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_751(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_752) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_752(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_752(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_753) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_753(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_753(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_754) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_754(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_754(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_755) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_755(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_755(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_756) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_756(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_756(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_757) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_757(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_757(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_758) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_758(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_758(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_759) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_759(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_759(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_760) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_760(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_760(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_761) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_761(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_761(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_762) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_762(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_762(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_763) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_763(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_763(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_764) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_764(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_764(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_765) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_765(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_765(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_766) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_766(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_766(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_767) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_767(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_767(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_768) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_768(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_768(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_769) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_769(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_769(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_770) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_770(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_770(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_771) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_771(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_771(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_772) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_772(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_772(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_773) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_773(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_773(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_774) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_774(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_774(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_775) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_775(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_775(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_776) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_776(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_776(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_777) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_777(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_777(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_778) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_778(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_778(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_779) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_779(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_779(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_780) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_780(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_780(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_781) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_781(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_781(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_782) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_782(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_782(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_783) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_783(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_783(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_784) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_784(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_784(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_785) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_785(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_785(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_786) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_786(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_786(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_787) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_787(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_787(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_788) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_788(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_788(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_789) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_789(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_789(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_790) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_790(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_790(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_791) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_791(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_791(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_792) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_792(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_792(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_793) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_793(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_793(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_794) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_794(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_794(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_795) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_795(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_795(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_796) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_796(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_796(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_797) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_797(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_797(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_798) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_798(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_798(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_799) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_799(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_799(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_800) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_800(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_800(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_801) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_801(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_801(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_802) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_802(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_802(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_803) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_803(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_803(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_804) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_804(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_804(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_805) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_805(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_805(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_806) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_806(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_806(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_807) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_807(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_807(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_808) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_808(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_808(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_809) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_809(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_809(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_810) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_810(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_810(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_811) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_811(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_811(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_812) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_812(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_812(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_813) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_813(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_813(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_814) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_814(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_814(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_815) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_815(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_815(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_816) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_816(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_816(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_817) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_817(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_817(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_818) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_818(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_818(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_819) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_819(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_819(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_820) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_820(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_820(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_821) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_821(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_821(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_822) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_822(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_822(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_823) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_823(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_823(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_824) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_824(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_824(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_825) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_825(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_825(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_826) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_826(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_826(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_827) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_827(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_827(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_828) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_828(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_828(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_829) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_829(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_829(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_830) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_830(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_830(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_831) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_831(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_831(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_832) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_832(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_832(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_833) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_833(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_833(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_834) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_834(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_834(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_835) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_835(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_835(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_836) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_836(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_836(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_837) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_837(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_837(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_838) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_838(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_838(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_839) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_839(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_839(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_840) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_840(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_840(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_841) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_841(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_841(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_842) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_842(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_842(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_843) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_843(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_843(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_844) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_844(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_844(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_845) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_845(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_845(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_846) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_846(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_846(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_847) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_847(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_847(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_848) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_848(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_848(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_849) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_849(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_849(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_850) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_850(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_850(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_851) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_851(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_851(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_852) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_852(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_852(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_853) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_853(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_853(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_854) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_854(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_854(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_855) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_855(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_855(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_856) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_856(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_856(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_857) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_857(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_857(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_858) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_858(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_858(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_859) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_859(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_859(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_860) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_860(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_860(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_861) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_861(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_861(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_862) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_862(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_862(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_863) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_863(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_863(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_864) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_864(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_864(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_865) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_865(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_865(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_866) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_866(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_866(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_867) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_867(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_867(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_868) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_868(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_868(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_869) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_869(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_869(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_870) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_870(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_870(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_871) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_871(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_871(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_872) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_872(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_872(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_873) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_873(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_873(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_874) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_874(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_874(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_875) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_875(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_875(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_876) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_876(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_876(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_877) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_877(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_877(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_878) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_878(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_878(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_879) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_879(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_879(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_880) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_880(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_880(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_881) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_881(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_881(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_882) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_882(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_882(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_883) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_883(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_883(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_884) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_884(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_884(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_885) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_885(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_885(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_886) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_886(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_886(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_887) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_887(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_887(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_888) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_888(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_888(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_889) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_889(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_889(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_890) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_890(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_890(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_891) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_891(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_891(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_892) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_892(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_892(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_893) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_893(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_893(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_894) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_894(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_894(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_895) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_895(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_895(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_896) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_896(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_896(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_897) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_897(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_897(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_898) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_898(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_898(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_899) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_899(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_899(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_900) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_900(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_900(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_901) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_901(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_901(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_902) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_902(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_902(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_903) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_903(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_903(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_904) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_904(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_904(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_905) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_905(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_905(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_906) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_906(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_906(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_907) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_907(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_907(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_908) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_908(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_908(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_909) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_909(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_909(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_910) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_910(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_910(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_911) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_911(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_911(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_912) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_912(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_912(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_913) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_913(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_913(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_914) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_914(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_914(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_915) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_915(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_915(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_916) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_916(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_916(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_917) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_917(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_917(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_918) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_918(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_918(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_919) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_919(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_919(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_920) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_920(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_920(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_921) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_921(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_921(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_922) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_922(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_922(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_923) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_923(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_923(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_924) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_924(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_924(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_925) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_925(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_925(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_926) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_926(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_926(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_927) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_927(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_927(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_928) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_928(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_928(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_929) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_929(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_929(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_930) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_930(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_930(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_931) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_931(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_931(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_932) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_932(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_932(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_933) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_933(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_933(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_934) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_934(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_934(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_935) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_935(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_935(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_936) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_936(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_936(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_937) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_937(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_937(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_938) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_938(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_938(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_939) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_939(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_939(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_940) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_940(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_940(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_941) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_941(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_941(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_942) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_942(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_942(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_943) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_943(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_943(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_944) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_944(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_944(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_945) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_945(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_945(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_946) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_946(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_946(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_947) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_947(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_947(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_948) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_948(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_948(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_949) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_949(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_949(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_950) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_950(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_950(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_951) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_951(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_951(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_952) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_952(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_952(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_953) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_953(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_953(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_954) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_954(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_954(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_955) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_955(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_955(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_956) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_956(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_956(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_957) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_957(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_957(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_958) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_958(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_958(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_959) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_959(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_959(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_960) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_960(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_960(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_961) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_961(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_961(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_962) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_962(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_962(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_963) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_963(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_963(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_964) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_964(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_964(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_965) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_965(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_965(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_966) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_966(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_966(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_967) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_967(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_967(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_968) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_968(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_968(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_969) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_969(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_969(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_970) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_970(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_970(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_971) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_971(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_971(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_972) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_972(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_972(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_973) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_973(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_973(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_974) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_974(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_974(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_975) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_975(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_975(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_976) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_976(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_976(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_977) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_977(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_977(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_978) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_978(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_978(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_979) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_979(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_979(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_980) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_980(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_980(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_981) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_981(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_981(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_982) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_982(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_982(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_983) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_983(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_983(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_984) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_984(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_984(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_985) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_985(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_985(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_986) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_986(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_986(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_987) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_987(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_987(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_988) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_988(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_988(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_989) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_989(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_989(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_990) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_990(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_990(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_991) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_991(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_991(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_992) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_992(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_992(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_993) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_993(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_993(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_994) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_994(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_994(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_995) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_995(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_995(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_996) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_996(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_996(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_997) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_997(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_997(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_998) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_998(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_998(FUNC)
+#endif
+
+#if defined(EIGEN_TEST_PART_999) || defined(EIGEN_TEST_PART_ALL)
+#define CALL_SUBTEST_999(FUNC) CALL_SUBTEST(FUNC)
+#else
+#define CALL_SUBTEST_999(FUNC)
+#endif
+
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 02fb890c6..73befbca1 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -1,16 +1,7 @@
-# generate split test header file only if it does not yet exist
-# in order to prevent a rebuild every time cmake is configured
-if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
-  file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "")
-  foreach(i RANGE 1 999)
-    file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h
-      "#if defined(EIGEN_TEST_PART_${i}) || defined(EIGEN_TEST_PART_ALL)\n"
-      "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n"
-      "#else\n"
-      "#define CALL_SUBTEST_${i}(FUNC)\n"
-      "#endif\n\n"
-    )
-  endforeach()
+# The file split_test_helper.h was generated at first run,
+# it is now included in test/
+if(EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
+  file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h)
 endif()
 
 set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Unsupported")

From 40797dbea334405fb49ea1ba2cd972a23a7dbe02 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 17 Jul 2018 00:11:20 +0200
Subject: [PATCH 473/680] bug #1572: use c++11 atomic instead of volatile if
 c++11 is available, and disable multi-threaded GEMM on non-x86 without c++11.

---
 CMakeLists.txt                                |  5 +++--
 Eigen/src/Core/products/GeneralMatrixMatrix.h |  4 +++-
 Eigen/src/Core/products/Parallelizer.h        | 18 +++++++++++++++++-
 Eigen/src/Core/util/Macros.h                  | 12 +++++++++++-
 test/CMakeLists.txt                           |  2 --
 5 files changed, 34 insertions(+), 7 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0547ee681..dbf0999ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -8,6 +8,7 @@ if(${CMAKE_SOURCE_DIR} STREQUAL ${CMAKE_BINARY_DIR})
   message(FATAL_ERROR "In-source builds not allowed. Please make a new directory (called a build directory) and run CMake from there. You may need to remove CMakeCache.txt. ")
 endif()
 
+
 # Alias Eigen_*_DIR to Eigen3_*_DIR:
 
 set(Eigen_SOURCE_DIR ${Eigen3_SOURCE_DIR})
@@ -107,7 +108,7 @@ if(NOT WIN32 OR NOT CMAKE_HOST_SYSTEM_NAME MATCHES Windows)
   option(EIGEN_BUILD_PKGCONFIG "Build pkg-config .pc file for Eigen" ON)
 endif()
 
-set(CMAKE_INCLUDE_CURRENT_DIR ON)
+set(CMAKE_INCLUDE_CURRENT_DIR OFF)
 
 option(EIGEN_SPLIT_LARGE_TESTS "Split large tests into smaller executables" ON)
 
@@ -377,7 +378,7 @@ option(EIGEN_TEST_CXX11 "Enable testing with C++11 and C++11 features (e.g. Tens
 
 set(EIGEN_CUDA_COMPUTE_ARCH 30 CACHE STRING "The CUDA compute architecture level to target when compiling CUDA code")
 
-include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 # Backward compatibility support for EIGEN_INCLUDE_INSTALL_DIR
 if(EIGEN_INCLUDE_INSTALL_DIR)
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index dc47cf7cf..bd7b6ff2a 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -108,7 +108,7 @@ static void run(Index rows, Index cols, Index depth,
       // i.e., we test that info[tid].users equals 0.
       // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
       while(info[tid].users!=0) {}
-      info[tid].users += threads;
+      info[tid].users = threads;
 
       pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
 
@@ -146,7 +146,9 @@ static void run(Index rows, Index cols, Index depth,
       // Release all the sub blocks A'_i of A' for the current thread,
       // i.e., we simply decrement the number of users by 1
       for(Index i=0; i<threads; ++i)
+#if !EIGEN_HAS_CXX11_ATOMIC
         #pragma omp atomic
+#endif
         info[i].users -= 1;
     }
   }
diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h
index 15b5c5f94..92e9b0d9f 100644
--- a/Eigen/src/Core/products/Parallelizer.h
+++ b/Eigen/src/Core/products/Parallelizer.h
@@ -10,6 +10,10 @@
 #ifndef EIGEN_PARALLELIZER_H
 #define EIGEN_PARALLELIZER_H
 
+#if EIGEN_HAS_CXX11_ATOMIC
+#include <atomic>
+#endif
+
 namespace Eigen {
 
 namespace internal {
@@ -75,8 +79,17 @@ template<typename Index> struct GemmParallelInfo
 {
   GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
 
+  // volatile is not enough on all architectures (see bug 1572)
+  // to guarantee that when thread A says to thread B that it is
+  // done with packing a block, then all writes have been really
+  // carried out... C++11 memory model+atomic guarantees this.
+#if EIGEN_HAS_CXX11_ATOMIC
+  std::atomic<Index> sync;
+  std::atomic<int> users;
+#else
   Index volatile sync;
   int volatile users;
+#endif
 
   Index lhs_start;
   Index lhs_length;
@@ -87,7 +100,10 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth,
 {
   // TODO when EIGEN_USE_BLAS is defined,
   // we should still enable OMP for other scalar types
-#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
+  // Without C++11, we have to disable GEMM's parallelization on
+  // non x86 architectures because there volatile is not enough for our purpose.
+  // See bug 1572.
+#if (! defined(EIGEN_HAS_OPENMP)) || defined(EIGEN_USE_BLAS) || ((!EIGEN_HAS_CXX11_ATOMIC) && !(EIGEN_ARCH_i386_OR_x86_64))
   // FIXME the transpose variable is only needed to properly split
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 46ca0193a..fc4c0815c 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -537,7 +537,7 @@
 #endif
 #endif
 
-// Does the compiler support type_trais?
+// Does the compiler support type_traits?
 #ifndef EIGEN_HAS_TYPE_TRAITS
 #if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
 #define EIGEN_HAS_TYPE_TRAITS 1
@@ -617,6 +617,16 @@
   #endif
 #endif
 
+#ifndef EIGEN_HAS_CXX11_ATOMIC
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         (__has_feature(cxx_atomic) \
+      || (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_MSVC==0 || EIGEN_COMP_MSVC >= 1700)))
+    #define EIGEN_HAS_CXX11_ATOMIC 1
+  #else
+    #define EIGEN_HAS_CXX11_ATOMIC 0
+  #endif
+#endif
 
 #if defined(EIGEN_CUDACC) && EIGEN_HAS_CONSTEXPR
   // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index aacfa22bb..b4730cff0 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -387,7 +387,6 @@ if(CUDA_FOUND)
   if(EIGEN_TEST_CUDA_CLANG)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 --cuda-gpu-arch=sm_30")
   endif()
-  cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR})
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
   
   ei_add_test(gpu_basic)
@@ -416,7 +415,6 @@ if (EIGEN_TEST_HIP)
 
       if (${HIP_PLATFORM} STREQUAL "hcc")
 
-	include_directories(${CMAKE_CURRENT_BINARY_DIR})
 	include_directories(${HIP_PATH}/include)
 
 	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")

From b3e7c9132d41da75c0a6af783300cb11101010db Mon Sep 17 00:00:00 2001
From: Paul Tucker <tucker@google.com>
Date: Mon, 16 Jul 2018 17:26:05 -0700
Subject: [PATCH 474/680] Add optional Allocator argument to ThreadPoolDevice
 constructor. When supplied, this allocator will be used in place of
 internal::aligned_malloc.  This permits e.g. use of a NUMA-node specific
 allocator where the thread-pool is also restricted a single NUMA-node.

---
 .../CXX11/src/Tensor/TensorDeviceThreadPool.h    | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index 90fd99027..be397e1b6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -95,14 +95,20 @@ static EIGEN_STRONG_INLINE void wait_until_ready(SyncType* n) {
 // Build a thread pool device on top the an existing pool of threads.
 struct ThreadPoolDevice {
   // The ownership of the thread pool remains with the caller.
-  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores) : pool_(pool), num_threads_(num_cores) { }
+  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores)
+      : pool_(pool), num_threads_(num_cores), allocator_(nullptr) { }
 
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
-    return internal::aligned_malloc(num_bytes);
+    return allocator_ ? allocator_->allocate(num_bytes)
+        : internal::aligned_malloc(num_bytes);
   }
 
   EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
-    internal::aligned_free(buffer);
+    if (allocator_) {
+      allocator_->deallocate(buffer);
+    } else {
+      internal::aligned_free(buffer);
+    }
   }
 
   EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
@@ -267,9 +273,13 @@ struct ThreadPoolDevice {
   // Thread pool accessor.
   ThreadPoolInterface* getPool() const { return pool_; }
 
+  // Allocator accessor.
+  Allocator* getAllocator() const { return allocator_; }
+
  private:
   ThreadPoolInterface* pool_;
   int num_threads_;
+  Allocator* allocator_;
 };
 
 

From 4e9848fa8600be69dfb51405606eafa1dba8d0bf Mon Sep 17 00:00:00 2001
From: Paul Tucker <tucker@google.com>
Date: Mon, 16 Jul 2018 17:53:36 -0700
Subject: [PATCH 475/680] Actually add optional Allocator* arg to
 ThreadPoolDevice().

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index be397e1b6..c9534d400 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -95,8 +95,8 @@ static EIGEN_STRONG_INLINE void wait_until_ready(SyncType* n) {
 // Build a thread pool device on top the an existing pool of threads.
 struct ThreadPoolDevice {
   // The ownership of the thread pool remains with the caller.
-  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores)
-      : pool_(pool), num_threads_(num_cores), allocator_(nullptr) { }
+  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores, Allocator* allocator = nullptr)
+      : pool_(pool), num_threads_(num_cores), allocator_(allocator) { }
 
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
     return allocator_ ? allocator_->allocate(num_bytes)

From 5e5987996f4a5e9cfda7122c5a0ee09cf4d3874b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 19 Jul 2018 18:33:53 +0200
Subject: [PATCH 476/680] Fix stupid error in Quaternion move ctor

---
 Eigen/src/Geometry/Quaternion.h | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h
index 8b203dc71..faea62f17 100644
--- a/Eigen/src/Geometry/Quaternion.h
+++ b/Eigen/src/Geometry/Quaternion.h
@@ -280,9 +280,8 @@ public:
   // We define a copy constructor, which means we don't get an implicit move constructor or assignment operator.
   /** Default move constructor */
   EIGEN_DEVICE_FUNC inline Quaternion(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
-  {
-    m_coeffs(std::move(other.coeffs()));
-  }
+    : m_coeffs(std::move(other.coeffs()))
+  {}
 
   /** Default move assignment operator */
   EIGEN_DEVICE_FUNC Quaternion& operator=(Quaternion&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
@@ -294,9 +293,8 @@ public:
   // And now because we declared a constructor, we don't get an implicit copy constructor. Say we want one.
   /** Default copy constructor */
   EIGEN_DEVICE_FUNC Quaternion(const Quaternion& other)
-  {
-    m_coeffs = other.coeffs();
-  }
+    : m_coeffs(other.coeffs())
+  {}
 #endif
 
   EIGEN_DEVICE_FUNC static Quaternion UnitRandom();

From 2ca25920090724f2547989a96f7449db17cf3315 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 19 Jul 2018 18:47:18 +0200
Subject: [PATCH 477/680] Fix determination of EIGEN_HAS_TYPE_TRAITS

---
 Eigen/src/Core/util/Macros.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 985406953..ac3805732 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -542,7 +542,7 @@
 // Does the compiler support type_traits?
 #ifndef EIGEN_HAS_TYPE_TRAITS
 #if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700) \
-  && (!EIGEN_COMP_GNUC || EIGEN_GNUC_AT_LEAST(5, 1))
+  && ((!EIGEN_COMP_GNUC_STRICT) || EIGEN_GNUC_AT_LEAST(5, 1))
     // ^^ Full support of type traits was added only to GCC 5.1.0.
 #define EIGEN_HAS_TYPE_TRAITS 1
 #define EIGEN_INCLUDE_TYPE_TRAITS

From 509a5fa77f1e2241123a0f22a2ea75a0e9267251 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 19 Jul 2018 18:47:38 +0200
Subject: [PATCH 478/680] Fix IsRelocatable without C++11

---
 Eigen/src/Core/PlainObjectBase.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index 016e76a53..da329fd4f 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -924,7 +924,7 @@ struct conservative_resize_like_impl
   #if EIGEN_HAS_TYPE_TRAITS
   static const bool IsRelocatable = std::is_trivially_copyable<typename Derived::Scalar>::value;
   #else
-  static const bool IsRelocatable = NumTraits<typename Derived::Scalar>::RequireInitialization;
+  static const bool IsRelocatable = !NumTraits<typename Derived::Scalar>::RequireInitialization;
   #endif
   static void run(DenseBase<Derived>& _this, Index rows, Index cols)
   {

From 2424e3b7aca514783fea4026775068827b8ee915 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 19 Jul 2018 18:48:19 +0200
Subject: [PATCH 479/680] Pass by const ref.

---
 test/AnnoyingScalar.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/test/AnnoyingScalar.h b/test/AnnoyingScalar.h
index b5b1476fe..7d1df2253 100644
--- a/test/AnnoyingScalar.h
+++ b/test/AnnoyingScalar.h
@@ -139,13 +139,13 @@ namespace internal {
 AnnoyingScalar get_test_precision(const AnnoyingScalar&)
 { return Eigen::test_precision<AnnoyingScalar>(); }
 
-AnnoyingScalar test_relative_error(AnnoyingScalar a, AnnoyingScalar b)
+AnnoyingScalar test_relative_error(const AnnoyingScalar &a, const AnnoyingScalar &b)
 { return test_relative_error(*a.v, *b.v); }
 
-inline bool test_isApprox(AnnoyingScalar a, AnnoyingScalar b)
+inline bool test_isApprox(const AnnoyingScalar &a, const AnnoyingScalar &b)
 { return internal::isApprox(*a.v, *b.v, test_precision<float>()); }
 
-inline bool test_isMuchSmallerThan(AnnoyingScalar a, AnnoyingScalar b)
+inline bool test_isMuchSmallerThan(const AnnoyingScalar &a, const AnnoyingScalar &b)
 { return test_isMuchSmallerThan(*a.v, *b.v); }
 
 #endif // EIGEN_TEST_ANNOYING_SCALAR_H

From c58b874727717c88f5026d29e3eadc9d1b19653d Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 19 Jul 2018 17:37:03 -0700
Subject: [PATCH 480/680] PR430: Convert count to the reducer type in
 MeanReducer
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Without explicit conversion Tensorflow fails to compile, pset1 template deduction fails.

cannot convert '((const Eigen::internal::MeanReducer<Eigen::half>*)this)
  ->Eigen::internal::MeanReducer<Eigen::half>::packetCount_'
(type 'const DenseIndex {aka const long int}')
to type 'const type& {aka const Eigen::half&}'
     return pdiv(vaccum, pset1<Packet>(packetCount_));
Honestly I’m not sure why it works in Eigen tests, because Eigen::half constructor is explicit, and why it stopped working in TF, I didn’t find any relevant changes since previous Eigen upgrade.

static_cast<T>(packetCount_) - breaks cxx11_tensor_reductions test for Eigen::half, also quite surprising.
---
 unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index a9027e97d..3d0e4035a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -171,7 +171,7 @@ template <typename T> struct MeanReducer
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
-    return pdiv(vaccum, pset1<Packet>(packetCount_));
+    return pdiv(vaccum, pset1<Packet>(T(packetCount_)));
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {

From d4afccde5a9553ddfb48b0f5fad0115cd8bf791a Mon Sep 17 00:00:00 2001
From: Paul Tucker <tucker@google.com>
Date: Thu, 19 Jul 2018 17:43:44 -0700
Subject: [PATCH 481/680] Add test coverage for ThreadPoolDevice optional
 allocator.

---
 .../CXX11/src/Tensor/TensorDeviceThreadPool.h |  7 +++
 unsupported/test/cxx11_tensor_thread_pool.cpp | 45 +++++++++++++++++--
 2 files changed, 48 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index c9534d400..f4123b71d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -91,6 +91,13 @@ static EIGEN_STRONG_INLINE void wait_until_ready(SyncType* n) {
   }
 }
 
+// An abstract interface to a device specific memory allocator.
+class Allocator {
+ public:
+  virtual ~Allocator() {}
+  EIGEN_DEVICE_FUNC virtual void* allocate(size_t num_bytes) const = 0;
+  EIGEN_DEVICE_FUNC virtual void deallocate(void* buffer) const = 0;
+};
 
 // Build a thread pool device on top the an existing pool of threads.
 struct ThreadPoolDevice {
diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index 2ef665f30..200664740 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -16,6 +16,25 @@
 
 using Eigen::Tensor;
 
+class TestAllocator : public Allocator {
+ public:
+  ~TestAllocator() override {}
+  EIGEN_DEVICE_FUNC void* allocate(size_t num_bytes) const override {
+    const_cast<TestAllocator*>(this)->alloc_count_++;
+    return internal::aligned_malloc(num_bytes);
+  }
+  EIGEN_DEVICE_FUNC void deallocate(void* buffer) const override {
+    const_cast<TestAllocator*>(this)->dealloc_count_++;
+    internal::aligned_free(buffer);
+  }
+
+  int alloc_count() const { return alloc_count_; }
+  int dealloc_count() const { return dealloc_count_; }
+
+ private:
+  int alloc_count_ = 0;
+  int dealloc_count_ = 0;
+};
 
 void test_multithread_elementwise()
 {
@@ -320,14 +339,14 @@ void test_multithread_random()
 }
 
 template<int DataLayout>
-void test_multithread_shuffle()
+void test_multithread_shuffle(Allocator* allocator)
 {
   Tensor<float, 4, DataLayout> tensor(17,5,7,11);
   tensor.setRandom();
 
   const int num_threads = internal::random<int>(2, 11);
   ThreadPool threads(num_threads);
-  Eigen::ThreadPoolDevice device(&threads, num_threads);
+  Eigen::ThreadPoolDevice device(&threads, num_threads, allocator);
 
   Tensor<float, 4, DataLayout> shuffle(7,5,11,17);
   array<ptrdiff_t, 4> shuffles = {{2,1,3,0}};
@@ -344,6 +363,21 @@ void test_multithread_shuffle()
   }
 }
 
+void test_threadpool_allocate(TestAllocator* allocator)
+{
+  const int num_threads = internal::random<int>(2, 11);
+  const int num_allocs = internal::random<int>(2, 11);
+  ThreadPool threads(num_threads);
+  Eigen::ThreadPoolDevice device(&threads, num_threads, allocator);
+
+  for (int a = 0; a < num_allocs; ++a) {
+    void* ptr = device.allocate(512);
+    device.deallocate(ptr);
+  }
+  VERIFY(allocator != nullptr);
+  VERIFY_IS_EQUAL(allocator->alloc_count(), num_allocs);
+  VERIFY_IS_EQUAL(allocator->dealloc_count(), num_allocs);
+}
 
 void test_cxx11_tensor_thread_pool()
 {
@@ -368,6 +402,9 @@ void test_cxx11_tensor_thread_pool()
 
   CALL_SUBTEST_6(test_memcpy());
   CALL_SUBTEST_6(test_multithread_random());
-  CALL_SUBTEST_6(test_multithread_shuffle<ColMajor>());
-  CALL_SUBTEST_6(test_multithread_shuffle<RowMajor>());
+
+  TestAllocator test_allocator;
+  CALL_SUBTEST_6(test_multithread_shuffle<ColMajor>(nullptr));
+  CALL_SUBTEST_6(test_multithread_shuffle<RowMajor>(&test_allocator));
+  CALL_SUBTEST_6(test_threadpool_allocate(&test_allocator));
 }

From 56a750b6ccf52e59874e991a2373265fc3ab4b8e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 20 Jul 2018 08:36:38 -0700
Subject: [PATCH 482/680] Disable optimization for sparse_product unit test
 with MSVC 2013, otherwise it takes several hours to build.

---
 test/sparse_product.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp
index d6b7638b1..a44d39d1d 100644
--- a/test/sparse_product.cpp
+++ b/test/sparse_product.cpp
@@ -7,6 +7,12 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
+#if EIGEN_COMP_MSVC==1800
+// This unit test takes forever to compile in Release mode with MSVC 2013,
+// multiple hours. So let's switch off optimization for this one.
+#pragma optimize("",off)
+#endif
+
 static long int nb_temporaries;
 
 inline void on_temporary_creation() {

From de70671937827dd61a63d2432e0bea4ca7b00300 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 20 Jul 2018 17:51:17 +0200
Subject: [PATCH 483/680] Oopps, EIGEN_COMP_MSVC is not available before
 including Eigen.

---
 test/sparse_product.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp
index a44d39d1d..db1b0e833 100644
--- a/test/sparse_product.cpp
+++ b/test/sparse_product.cpp
@@ -7,7 +7,7 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#if EIGEN_COMP_MSVC==1800
+#if defined(_MSC_VER) && (_MSC_VER==1800)
 // This unit test takes forever to compile in Release mode with MSVC 2013,
 // multiple hours. So let's switch off optimization for this one.
 #pragma optimize("",off)

From 2bf864f1eb1b2497d28747c7f77ec3e2f1e8d9d2 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 20 Jul 2018 10:11:44 -0700
Subject: [PATCH 484/680] Disable type traits for stdlibc++ <= 4.9.3

---
 Eigen/src/Core/util/Macros.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index ac3805732..a1ac9162b 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -542,8 +542,8 @@
 // Does the compiler support type_traits?
 #ifndef EIGEN_HAS_TYPE_TRAITS
 #if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700) \
-  && ((!EIGEN_COMP_GNUC_STRICT) || EIGEN_GNUC_AT_LEAST(5, 1))
-    // ^^ Full support of type traits was added only to GCC 5.1.0.
+  && ((!EIGEN_COMP_GNUC_STRICT) || EIGEN_GNUC_AT_LEAST(5, 1)) \ // Full support of type traits was added only to GCC 5.1.0.
+  && ((!defined(__GLIBCXX__))   || __GLIBCXX__ > 20150626)      // The last release of 4.x libstdc++
 #define EIGEN_HAS_TYPE_TRAITS 1
 #define EIGEN_INCLUDE_TYPE_TRAITS
 #else

From 02eaaacbc5b4f51718e9a82b7ede57659bdb514a Mon Sep 17 00:00:00 2001
From: Gustavo Lima Chaves <gustavo.lima.chaves@intel.com>
Date: Fri, 20 Jul 2018 16:08:40 -0700
Subject: [PATCH 485/680] Move cxx11_tensor_uint128 test under an
 EIGEN_TEST_CXX11 guarded  block

Builds configured without the -DEIGEN_TEST_CXX11=ON flag would fail
right away without this, as this test seems to rely on those language
features. The skip under compilation with MSVC was kept.
---
 unsupported/test/CMakeLists.txt | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 0fc864eca..55b86a32f 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -141,10 +141,6 @@ ei_add_test(cxx11_tensor_sugar)
 ei_add_test(cxx11_tensor_roundings)
 ei_add_test(cxx11_tensor_layout_swap)
 ei_add_test(cxx11_tensor_io)
-if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
-  # This test requires __uint128_t which is only available on 64bit systems
-  ei_add_test(cxx11_tensor_uint128)
-endif()
 endif()
 
 if(EIGEN_TEST_CXX11)
@@ -226,6 +222,10 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_scan)
   ei_add_test(cxx11_tensor_trace)
   ei_add_test(cxx11_tensor_move)
+if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8" AND NOT CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+  # This test requires __uint128_t which is only available on 64bit systems
+  ei_add_test(cxx11_tensor_uint128)
+endif()
 
 endif()
 

From c747cde69a9ab286984a09223c3f477b782a6617 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Jul 2018 16:20:25 +0200
Subject: [PATCH 486/680] Add lastN shorcuts to seq/seqN.

---
 Eigen/src/Core/ArithmeticSequence.h | 34 +++++++++++++++++++++++++++--
 test/indexed_view.cpp               |  8 +++++++
 2 files changed, 40 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/ArithmeticSequence.h b/Eigen/src/Core/ArithmeticSequence.h
index ada1571f1..2083cf178 100644
--- a/Eigen/src/Core/ArithmeticSequence.h
+++ b/Eigen/src/Core/ArithmeticSequence.h
@@ -225,7 +225,8 @@ auto seq(FirstType f, LastType l, IncrType incr)
                -typename internal::cleanup_index_type<FirstType>::type(f)+CleanedIncrType(incr)) / CleanedIncrType(incr),
               CleanedIncrType(incr));
 }
-#else
+
+#else // EIGEN_HAS_CXX11
 
 template<typename FirstType,typename LastType>
 typename internal::enable_if<!(Symbolic::is_symbolic<FirstType>::value || Symbolic::is_symbolic<LastType>::value),
@@ -314,10 +315,39 @@ seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, const Symbolic::BaseExpr<Last
   typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
   return seqN(f.derived(),(l.derived()-f.derived()+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
 }
-#endif
+#endif // EIGEN_HAS_CXX11
 
 #endif // EIGEN_PARSED_BY_DOXYGEN
 
+
+#if EIGEN_HAS_CXX11
+/** \cpp11
+  * \returns a symbolic ArithmeticSequence representing the last \a size elements with increment \a incr.
+  *
+  * It is a shortcut for: \code seqN(last-(size-fix<1>)*incr, size, incr) \endcode
+  * 
+  * \sa lastN(SizeType), seqN(FirstType,SizeType), seq(FirstType,LastType,IncrType) */
+template<typename SizeType,typename IncrType>
+auto lastN(SizeType size, IncrType incr)
+-> decltype(seqN(Eigen::placeholders::last-(size-fix<1>())*incr, size, incr))
+{
+  return seqN(Eigen::placeholders::last-(size-fix<1>())*incr, size, incr);
+}
+
+/** \cpp11
+  * \returns a symbolic ArithmeticSequence representing the last \a size elements with a unit increment.
+  *
+  *  It is a shortcut for: \code seq(last+fix<1>-size, last) \endcode
+  * 
+  * \sa lastN(SizeType,IncrType, seqN(FirstType,SizeType), seq(FirstType,LastType) */
+template<typename SizeType>
+auto lastN(SizeType size)
+-> decltype(seqN(Eigen::placeholders::last+fix<1>()-size, size))
+{
+  return seqN(Eigen::placeholders::last+fix<1>()-size, size);
+}
+#endif
+
 namespace internal {
 
 // Convert a symbolic span into a usable one (i.e., remove last/end "keywords")
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 4cc0954f7..7219c777f 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -293,6 +293,14 @@ void check_indexed_view()
   }
 
 #if EIGEN_HAS_CXX11
+  // check lastN
+  VERIFY_IS_APPROX( a(lastN(3)), a.tail(3) );
+  VERIFY( MATCH( a(lastN(3)), "7\n8\n9" ) );
+  VERIFY_IS_APPROX( a(lastN(fix<3>())), a.tail<3>() );
+  VERIFY( MATCH( a(lastN(3,2)), "5\n7\n9" ) );
+  VERIFY( MATCH( a(lastN(3,fix<2>())), "5\n7\n9" ) );
+  VERIFY( a(lastN(fix<3>())).SizeAtCompileTime == 3 );
+
   VERIFY( (A(all, std::array<int,4>{{1,3,2,4}})).ColsAtCompileTime == 4);
 
   VERIFY_IS_APPROX( (A(std::array<int,3>{{1,3,5}}, std::array<int,4>{{9,6,3,0}})), A(seqN(1,3,2), seqN(9,4,-3)) );

From 4ca3e48f42adab8f907c18c64107dd23d432ccdc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 23 Jul 2018 16:51:57 +0200
Subject: [PATCH 487/680] fix typo

---
 Eigen/src/Core/util/Macros.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index a1ac9162b..73a9883b9 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -540,10 +540,12 @@
 #endif
 
 // Does the compiler support type_traits?
+// - full support of type traits was added only to GCC 5.1.0.
+// - 20150626 corresponds to the last release of 4.x libstdc++
 #ifndef EIGEN_HAS_TYPE_TRAITS
 #if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700) \
-  && ((!EIGEN_COMP_GNUC_STRICT) || EIGEN_GNUC_AT_LEAST(5, 1)) \ // Full support of type traits was added only to GCC 5.1.0.
-  && ((!defined(__GLIBCXX__))   || __GLIBCXX__ > 20150626)      // The last release of 4.x libstdc++
+  && ((!EIGEN_COMP_GNUC_STRICT) || EIGEN_GNUC_AT_LEAST(5, 1)) \
+  && ((!defined(__GLIBCXX__))   || __GLIBCXX__ > 20150626)
 #define EIGEN_HAS_TYPE_TRAITS 1
 #define EIGEN_INCLUDE_TYPE_TRAITS
 #else

From 34a75c3c5cec4e2bfe5c68164f8c3372f6ae5ecb Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 20 Jul 2018 17:37:20 -0700
Subject: [PATCH 488/680] Initial support of TensorBlock

---
 unsupported/Eigen/CXX11/Tensor                |   1 +
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 412 ++++++++++++++++++
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h |   6 +
 unsupported/test/CMakeLists.txt               |  17 +-
 .../test/cxx11_tensor_block_access.cpp        | 182 ++++++++
 5 files changed, 610 insertions(+), 8 deletions(-)
 create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
 create mode 100644 unsupported/test/cxx11_tensor_block_access.cpp

diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index ddbbcfba2..397d55f76 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -118,6 +118,7 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorReduction.h"
 #include "src/Tensor/TensorReductionGpu.h"
 #include "src/Tensor/TensorArgMax.h"
+#include "src/Tensor/TensorBlock.h"
 #include "src/Tensor/TensorConcatenation.h"
 #include "src/Tensor/TensorContractionMapper.h"
 #include "src/Tensor/TensorContractionBlocking.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
new file mode 100644
index 000000000..59535cd91
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -0,0 +1,412 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Andy Davis <andydavis@google.com>
+// Copyright (C) 2018 Eugene Zhulenev <ezhulenev@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
+#define EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \class TensorBlockShapeType
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block shape type.
+ *
+ * Tensor block shape type defines what are the shape preference for the blocks
+ * extracted from the larger tensor.
+ *
+ * Example:
+ *
+ * We want to extract blocks of 100 elements from the large 100x100 tensor:
+ *  - tensor: 100x100
+ *  - target_block_size: 100
+ *
+ * TensorBlockShapeType:
+ *  - kUniformAllDims: 100 blocks of size 10x10
+ *  - kSkewedInnerDims: 100 blocks of size 100x1 (or 1x100 depending on a column
+ *                      or row major layout)
+ */
+enum class TensorBlockShapeType {
+  kUniformAllDims,
+  kSkewedInnerDims,
+};
+
+/**
+ * \class TensorBlock
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block class.
+ *
+ * This class represents a tensor block specified by the index of the
+ * first block coefficient, and the size of the block in each dimension.
+ */
+template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+class TensorBlock {
+ public:
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  TensorBlock(const Index first_coeff_index, const Dimensions& block_sizes,
+              const Dimensions& block_strides, const Dimensions& tensor_strides,
+              Scalar* data)
+      : m_first_coeff_index(first_coeff_index),
+        m_block_sizes(block_sizes),
+        m_block_strides(block_strides),
+        m_tensor_strides(tensor_strides),
+        m_data(data) {}
+
+  Index first_coeff_index() const { return m_first_coeff_index; }
+
+  const Dimensions& block_sizes() const { return m_block_sizes; }
+
+  const Dimensions& block_strides() const { return m_block_strides; }
+
+  const Dimensions& tensor_strides() const { return m_tensor_strides; }
+
+  Scalar* data() { return m_data; }
+
+  const Scalar* data() const { return m_data; }
+
+ private:
+  Index m_first_coeff_index;
+  Dimensions m_block_sizes;
+  Dimensions m_block_strides;
+  Dimensions m_tensor_strides;
+  Scalar* m_data;  // Not owned.
+};
+
+/**
+ * \class TensorBlockMapper
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block mapper class.
+ *
+ * This class is responsible for iterating over the blocks of a tensor.
+ */
+template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+class TensorBlockMapper {
+ public:
+  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      TensorBlock;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  TensorBlockMapper(const Dimensions& dims,
+                    const TensorBlockShapeType block_shape,
+                    size_t min_target_size)
+      : m_dimensions(dims),
+        m_block_dim_sizes(BlockDimensions(dims, block_shape, min_target_size)) {
+    // Calculate block counts by dimension and total block count.
+    DSizes<Index, NumDims> block_count;
+    for (size_t i = 0; i < block_count.rank(); ++i) {
+      block_count[i] = divup(m_dimensions[i], m_block_dim_sizes[i]);
+    }
+    m_total_block_count = array_prod(block_count);
+
+    // Calculate block strides (used for enumerating blocks).
+    if (NumDims > 0) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_block_strides[0] = 1;
+        m_tensor_strides[0] = 1;
+        for (int i = 1; i < NumDims; ++i) {
+          m_block_strides[i] = m_block_strides[i - 1] * block_count[i - 1];
+          m_tensor_strides[i] = m_tensor_strides[i - 1] * m_dimensions[i - 1];
+        }
+      } else {
+        m_block_strides[NumDims - 1] = 1;
+        m_tensor_strides[NumDims - 1] = 1;
+        for (int i = NumDims - 2; i >= 0; --i) {
+          m_block_strides[i] = m_block_strides[i + 1] * block_count[i + 1];
+          m_tensor_strides[i] = m_tensor_strides[i + 1] * m_dimensions[i + 1];
+        }
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  GetBlockForIndex(Index block_index, Scalar* data) const {
+    Index first_coeff_index = 0;
+    DSizes<Index, NumDims> coords;
+    DSizes<Index, NumDims> sizes;
+    DSizes<Index, NumDims> strides;
+    if (NumDims > 0) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        for (int i = NumDims - 1; i > 0; --i) {
+          const Index idx = block_index / m_block_strides[i];
+          coords[i] = idx * m_block_dim_sizes[i];
+          sizes[i] =
+              numext::mini((m_dimensions[i] - coords[i]), m_block_dim_sizes[i]);
+          block_index -= idx * m_block_strides[i];
+          first_coeff_index += coords[i] * m_tensor_strides[i];
+        }
+        coords[0] = block_index * m_block_dim_sizes[0];
+        sizes[0] =
+            numext::mini((m_dimensions[0] - coords[0]), m_block_dim_sizes[0]);
+        first_coeff_index += coords[0] * m_tensor_strides[0];
+
+        strides[0] = 1;
+        for (int i = 1; i < NumDims; ++i) {
+          strides[i] = strides[i - 1] * sizes[i - 1];
+        }
+      } else {
+        for (int i = 0; i < NumDims - 1; ++i) {
+          const Index idx = block_index / m_block_strides[i];
+          coords[i] = idx * m_block_dim_sizes[i];
+          sizes[i] =
+              numext::mini((m_dimensions[i] - coords[i]), m_block_dim_sizes[i]);
+          block_index -= idx * m_block_strides[i];
+          first_coeff_index += coords[i] * m_tensor_strides[i];
+        }
+        coords[NumDims - 1] = block_index * m_block_dim_sizes[NumDims - 1];
+        sizes[NumDims - 1] =
+            numext::mini((m_dimensions[NumDims - 1] - coords[NumDims - 1]),
+                         m_block_dim_sizes[NumDims - 1]);
+        first_coeff_index +=
+            coords[NumDims - 1] * m_tensor_strides[NumDims - 1];
+
+        strides[NumDims - 1] = 1;
+        for (int i = NumDims - 2; i >= 0; --i) {
+          strides[i] = strides[i + 1] * sizes[i + 1];
+        }
+      }
+    }
+
+    return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
+                       data);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index total_block_count() const {
+    return m_total_block_count;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index block_dims_total_size() const {
+    return m_block_dim_sizes.TotalSize();
+  }
+
+ private:
+  static int InnerDimIndex(Index i) {
+    return Layout == static_cast<int>(ColMajor) ? i : NumDims - i - 1;
+  }
+
+  static Dimensions BlockDimensions(const Dimensions& tensor_dims,
+                                    const TensorBlockShapeType block_shape,
+                                    size_t min_target_size) {
+    min_target_size = numext::maxi<size_t>(1, min_target_size);
+
+    // If tensor fully fits into the target size, we'll treat it a single block.
+    Dimensions block_dim_sizes = tensor_dims;
+
+    if (tensor_dims.TotalSize() == 0) {
+      // Corner case: one of the dimensions is zero. Logic below is too complex
+      // to handle this case on a general basis, just use unit block size.
+      // Note: we must not yield blocks with zero dimensions (recipe for
+      // overflows/underflows, divisions by zero and NaNs later).
+      for (int i = 0; i < NumDims; ++i) {
+        block_dim_sizes[i] = 1;
+      }
+    } else if (block_dim_sizes.TotalSize() > min_target_size) {
+      if (block_shape == TensorBlockShapeType::kUniformAllDims) {
+        // Tensor will not fit within 'min_target_size' budget: calculate tensor
+        // block dimension sizes based on "square" dimension size target.
+        const size_t dim_size_target = static_cast<const size_t>(
+            std::pow(static_cast<float>(min_target_size),
+                     1.0 / static_cast<float>(block_dim_sizes.rank())));
+        for (size_t i = 0; i < block_dim_sizes.rank(); ++i) {
+          // TODO(andydavis) Adjust the inner most 'block_dim_size' to make it
+          // a multiple of the packet size. Note that reducing
+          // 'block_dim_size' in this manner can increase the number of
+          // blocks, and so will amplify any per-block overhead.
+          block_dim_sizes[i] = numext::mini(
+              dim_size_target, static_cast<size_t>(tensor_dims[i]));
+        }
+        // Add any un-allocated coefficients to inner dimension(s).
+        Index total_size = block_dim_sizes.TotalSize();
+        for (int i = 0; i < NumDims; ++i) {
+          const int dim = InnerDimIndex(i);
+          if (block_dim_sizes[dim] < tensor_dims[dim]) {
+            const Index total_size_other_dims =
+                total_size / block_dim_sizes[dim];
+            const Index alloc_avail =
+                divup<Index>(min_target_size, total_size_other_dims);
+            if (alloc_avail == block_dim_sizes[dim]) {
+              // Insufficient excess coefficients to allocate.
+              break;
+            }
+            block_dim_sizes[dim] = numext::mini(tensor_dims[dim], alloc_avail);
+            total_size = total_size_other_dims * block_dim_sizes[dim];
+          }
+        }
+      } else if (block_shape == TensorBlockShapeType::kSkewedInnerDims) {
+        Index coeff_to_allocate = min_target_size;
+        for (int i = 0; i < NumDims; ++i) {
+          const int dim = InnerDimIndex(i);
+          block_dim_sizes[dim] =
+              numext::mini(coeff_to_allocate, tensor_dims[dim]);
+          coeff_to_allocate =
+              divup(coeff_to_allocate,
+                    numext::maxi(static_cast<Index>(1), block_dim_sizes[dim]));
+        }
+        eigen_assert(coeff_to_allocate == 1);
+      } else {
+        eigen_assert(false);  // someone added new block shape type
+      }
+    }
+
+    eigen_assert(
+        block_dim_sizes.TotalSize() >=
+        numext::mini<size_t>(min_target_size, tensor_dims.TotalSize()));
+
+    return block_dim_sizes;
+  }
+
+  Dimensions m_dimensions;
+  Dimensions m_block_dim_sizes;
+  Dimensions m_block_strides;
+  Dimensions m_tensor_strides;
+  Index m_total_block_count;
+};
+
+/**
+ * \class TensorSliceBlockMapper
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor slice block mapper class.
+ *
+ * This class is responsible for iterating over the blocks of
+ * a slice of a tensor. Supports shuffling of the block strides
+ * for callers that want to reduce strides for dimensions to be
+ * processed together.
+ *
+ */
+template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+class TensorSliceBlockMapper {
+ public:
+  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      TensorBlock;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  TensorSliceBlockMapper(const Dimensions& tensor_dims,
+                         const Dimensions& tensor_slice_offsets,
+                         const Dimensions& tensor_slice_extents,
+                         const Dimensions& block_dim_sizes,
+                         const Dimensions& block_stride_order)
+      : m_tensor_dimensions(tensor_dims),
+        m_tensor_slice_offsets(tensor_slice_offsets),
+        m_tensor_slice_extents(tensor_slice_extents),
+        m_block_dim_sizes(block_dim_sizes),
+        m_block_stride_order(block_stride_order),
+        m_total_block_count(1) {
+    // Calculate block counts by dimension and total block count.
+    DSizes<Index, NumDims> block_count;
+    for (size_t i = 0; i < block_count.rank(); ++i) {
+      block_count[i] = divup(m_tensor_slice_extents[i], m_block_dim_sizes[i]);
+    }
+    m_total_block_count = array_prod(block_count);
+
+    // Calculate block strides (used for enumerating blocks).
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_block_strides[0] = 1;
+      m_tensor_strides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_block_strides[i] = m_block_strides[i - 1] * block_count[i - 1];
+        m_tensor_strides[i] =
+            m_tensor_strides[i - 1] * m_tensor_dimensions[i - 1];
+      }
+    } else {
+      m_block_strides[NumDims - 1] = 1;
+      m_tensor_strides[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_block_strides[i] = m_block_strides[i + 1] * block_count[i + 1];
+        m_tensor_strides[i] =
+            m_tensor_strides[i + 1] * m_tensor_dimensions[i + 1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  GetBlockForIndex(Index block_index, Scalar* data) const {
+    Index first_coeff_index = 0;
+    DSizes<Index, NumDims> coords;
+    DSizes<Index, NumDims> sizes;
+    DSizes<Index, NumDims> strides;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = block_index / m_block_strides[i];
+        coords[i] = m_tensor_slice_offsets[i] + idx * m_block_dim_sizes[i];
+        sizes[i] = numext::mini(
+            m_tensor_slice_offsets[i] + m_tensor_slice_extents[i] - coords[i],
+            m_block_dim_sizes[i]);
+        block_index -= idx * m_block_strides[i];
+        first_coeff_index += coords[i] * m_tensor_strides[i];
+      }
+      coords[0] =
+          m_tensor_slice_offsets[0] + block_index * m_block_dim_sizes[0];
+      sizes[0] = numext::mini(
+          m_tensor_slice_offsets[0] + m_tensor_slice_extents[0] - coords[0],
+          m_block_dim_sizes[0]);
+      first_coeff_index += coords[0] * m_tensor_strides[0];
+
+      Index prev_dim = m_block_stride_order[0];
+      strides[prev_dim] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        const Index curr_dim = m_block_stride_order[i];
+        strides[curr_dim] = strides[prev_dim] * sizes[prev_dim];
+        prev_dim = curr_dim;
+      }
+    } else {
+      for (int i = 0; i < static_cast<int>(NumDims) - 1; ++i) {
+        const Index idx = block_index / m_block_strides[i];
+        coords[i] = m_tensor_slice_offsets[i] + idx * m_block_dim_sizes[i];
+        sizes[i] = numext::mini(
+            m_tensor_slice_offsets[i] + m_tensor_slice_extents[i] - coords[i],
+            m_block_dim_sizes[i]);
+        block_index -= idx * m_block_strides[i];
+        first_coeff_index += coords[i] * m_tensor_strides[i];
+      }
+      coords[NumDims - 1] = m_tensor_slice_offsets[NumDims - 1] +
+                            block_index * m_block_dim_sizes[NumDims - 1];
+      sizes[NumDims - 1] = numext::mini(
+          m_tensor_slice_offsets[NumDims - 1] +
+              m_tensor_slice_extents[NumDims - 1] - coords[NumDims - 1],
+          m_block_dim_sizes[NumDims - 1]);
+      first_coeff_index += coords[NumDims - 1] * m_tensor_strides[NumDims - 1];
+
+      Index prev_dim = m_block_stride_order[NumDims - 1];
+      strides[prev_dim] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        const Index curr_dim = m_block_stride_order[i];
+        strides[curr_dim] = strides[prev_dim] * sizes[prev_dim];
+        prev_dim = curr_dim;
+      }
+    }
+
+    return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
+                       data);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index total_block_count() const {
+    return m_total_block_count;
+  }
+
+ private:
+  Dimensions m_tensor_dimensions;
+  Dimensions m_tensor_slice_offsets;
+  Dimensions m_tensor_slice_extents;
+  Dimensions m_tensor_strides;
+  Dimensions m_block_dim_sizes;
+  Dimensions m_block_stride_order;
+  Dimensions m_block_strides;
+  Index m_total_block_count;
+};
+
+}  // namespace internal
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 86405e69b..192d4aa7b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -284,6 +284,12 @@ struct DSizes : array<DenseIndex, NumDims> {
     (*this)[0] = i0;
   }
 
+  EIGEN_DEVICE_FUNC DSizes(const DimensionList<DenseIndex, NumDims>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   template<typename... IndexTypes> EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE explicit DSizes(DenseIndex firstDimension, DenseIndex secondDimension, IndexTypes... otherDimensions) : Base({{firstDimension, secondDimension, otherDimensions...}}) {
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 55b86a32f..fa19b2159 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -130,6 +130,7 @@ if (NOT CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
 ei_add_test(cxx11_tensor_dimension)
 ei_add_test(cxx11_tensor_map)
 ei_add_test(cxx11_tensor_assign)
+ei_add_test(cxx11_tensor_block_access)
 ei_add_test(cxx11_tensor_comparisons)
 ei_add_test(cxx11_tensor_forced_eval)
 ei_add_test(cxx11_tensor_math)
@@ -291,14 +292,14 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
   unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 endif()
 
-# Add HIP specific tests 
+# Add HIP specific tests
 if (EIGEN_TEST_HIP)
 
   set(HIP_PATH "/opt/rocm/hip" CACHE STRING "Path to the HIP installation.")
 
   if (EXISTS ${HIP_PATH})
-    
-    list(APPEND CMAKE_MODULE_PATH ${HIP_PATH}/cmake) 
+
+    list(APPEND CMAKE_MODULE_PATH ${HIP_PATH}/cmake)
 
     find_package(HIP REQUIRED)
     if (HIP_FOUND)
@@ -328,22 +329,22 @@ if (EIGEN_TEST_HIP)
 	ei_add_test(cxx11_tensor_contract_gpu)
 	ei_add_test(cxx11_tensor_of_float16_gpu)
 	ei_add_test(cxx11_tensor_random_gpu)
-	
+
 	unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
-	
+
       elseif (${HIP_PLATFORM} STREQUAL "nvcc")
 	message(FATAL_ERROR "HIP_PLATFORM = nvcc is not supported within Eigen")
       else ()
 	message(FATAL_ERROR "Unknown HIP_PLATFORM = ${HIP_PLATFORM}")
       endif()
-      
+
     endif(HIP_FOUND)
 
   else ()
 
     message(FATAL_ERROR "EIGEN_TEST_HIP is ON, but the specified HIP_PATH (${HIP_PATH}) does not exist")
-    
+
   endif()
-    
+
 endif(EIGEN_TEST_HIP)
 
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
new file mode 100644
index 000000000..66e61aef1
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -0,0 +1,182 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Andy Davis <andydavis@google.com>
+// Copyright (C) 2018 Eugene Zhulenev <ezhulenev@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+
+#include <set>
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+using Eigen::Index;
+using Eigen::RowMajor;
+using Eigen::ColMajor;
+
+template<typename T>
+static const T& choose(int layout, const T& col, const T& row) {
+  return layout == ColMajor ? col : row;
+}
+
+template <int Layout>
+static void test_block_mapper_sanity()
+{
+  using T = int;
+  using TensorBlock = internal::TensorBlock<T, Index, 2, Layout>;
+  using TensorBlockMapper = internal::TensorBlockMapper<T, Index, 2, Layout>;
+
+  DSizes<Index, 2> tensor_dims(100, 100);
+
+  // Test uniform blocks.
+  TensorBlockMapper uniform_block_mapper(
+      tensor_dims, internal::TensorBlockShapeType::kUniformAllDims, 100);
+
+  VERIFY_IS_EQUAL(uniform_block_mapper.total_block_count(), 100);
+  VERIFY_IS_EQUAL(uniform_block_mapper.block_dims_total_size(), 100);
+
+  // 10x10 blocks
+  auto uniform_b0 = uniform_block_mapper.GetBlockForIndex(0, nullptr);
+  VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(0), 10);
+  VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(1), 10);
+  // Depending on a layout we stride by cols rows.
+  VERIFY_IS_EQUAL(uniform_b0.block_strides().at(0), choose(Layout, 1, 10));
+  VERIFY_IS_EQUAL(uniform_b0.block_strides().at(1), choose(Layout, 10, 1));
+  // Tensor strides depend only on a layout and not on the block size.
+  VERIFY_IS_EQUAL(uniform_b0.tensor_strides().at(0), choose(Layout, 1, 100));
+  VERIFY_IS_EQUAL(uniform_b0.tensor_strides().at(1), choose(Layout, 100, 1));
+
+  // Test skewed to inner dims blocks.
+  TensorBlockMapper skewed_block_mapper(
+      tensor_dims, internal::TensorBlockShapeType::kSkewedInnerDims, 100);
+
+  VERIFY_IS_EQUAL(skewed_block_mapper.total_block_count(), 100);
+  VERIFY_IS_EQUAL(skewed_block_mapper.block_dims_total_size(), 100);
+
+  // 1x100 (100x1) rows/cols depending on a tensor layout.
+  auto skewed_b0 = skewed_block_mapper.GetBlockForIndex(0, nullptr);
+  VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(0), choose(Layout, 100, 1));
+  VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(1), choose(Layout, 1, 100));
+  // Depending on a layout we stride by cols rows.
+  VERIFY_IS_EQUAL(skewed_b0.block_strides().at(0), choose(Layout, 1, 100));
+  VERIFY_IS_EQUAL(skewed_b0.block_strides().at(1), choose(Layout, 100, 1));
+  // Tensor strides depend only on a layout and not on the block size.
+  VERIFY_IS_EQUAL(skewed_b0.tensor_strides().at(0), choose(Layout, 1, 100));
+  VERIFY_IS_EQUAL(skewed_b0.tensor_strides().at(1), choose(Layout, 100, 1));
+}
+
+// Given a TensorBlock "visit" every element accessible though it, and a keep an
+// index in the visited set. Verify that every coeff accessed only once.
+template <typename T, int Layout, int NumDims>
+static void UpdateCoeffSet(
+    const internal::TensorBlock<T, Index, 4, Layout>& block,
+    Index first_coeff_index,
+    int dim_index,
+    std::set<Index>* visited_coeffs) {
+  const DSizes<Index, NumDims> block_sizes = block.block_sizes();
+  const DSizes<Index, NumDims> tensor_strides = block.tensor_strides();
+
+  for (int i = 0; i < block_sizes[dim_index]; ++i) {
+    if (tensor_strides[dim_index] == 1) {
+      auto inserted = visited_coeffs->insert(first_coeff_index + i);
+      VERIFY_IS_EQUAL(inserted.second, true);
+    } else {
+      int next_dim_index = dim_index + choose(Layout, -1, 1);
+      UpdateCoeffSet<T, Layout, NumDims>(block, first_coeff_index,
+                                         next_dim_index, visited_coeffs);
+      first_coeff_index += tensor_strides[dim_index];
+    }
+  }
+}
+
+template <int Layout>
+static void test_block_mapper_maps_every_element()
+{
+  using T = int;
+  using TensorBlock = internal::TensorBlock<T, Index, 4, Layout>;
+  using TensorBlockMapper = internal::TensorBlockMapper<T, Index, 4, Layout>;
+
+  DSizes<Index, 4> dims(5, 7, 11, 17);
+
+  auto total_coeffs = static_cast<int>(dims.TotalSize());
+
+  // Keep track of elements indices available via block access.
+  std::set<Index> coeff_set;
+
+  // Try different combinations of block types and sizes.
+  auto block_shape_type =
+      internal::random<bool>()
+          ? internal::TensorBlockShapeType::kUniformAllDims
+          : internal::TensorBlockShapeType::kSkewedInnerDims;
+  auto block_target_size = internal::random<int>(1, total_coeffs);
+  TensorBlockMapper block_mapper(dims, block_shape_type, block_target_size);
+
+  for (int i = 0; i < block_mapper.total_block_count(); ++i) {
+    TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
+    UpdateCoeffSet<T, Layout, 4>(block, block.first_coeff_index(),
+                                 choose(Layout, 3, 0), &coeff_set);
+  }
+
+  // Verify that every coefficient in the original Tensor is accessible through
+  // TensorBlock only once.
+  VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
+  VERIFY_IS_EQUAL(*coeff_set.begin(), static_cast<Index>(0));
+  VERIFY_IS_EQUAL(*coeff_set.rbegin(), static_cast<Index>(total_coeffs - 1));
+}
+
+template <int Layout>
+static void test_slice_block_mapper_maps_every_element()
+{
+  using T = int;
+  using TensorBlock = internal::TensorBlock<T, Index, 4, Layout>;
+  using TensorSliceBlockMapper =
+      internal::TensorSliceBlockMapper<T, Index, 4, Layout>;
+
+  DSizes<Index, 4> tensor_dims(5,7,11,17);
+  DSizes<Index, 4> tensor_slice_offsets(1,3,5,7);
+  DSizes<Index, 4> tensor_slice_extents(3,2,4,5);
+
+  // Keep track of elements indices available via block access.
+  std::set<Index> coeff_set;
+
+  auto total_coeffs = static_cast<int>(tensor_slice_extents.TotalSize());
+
+  // Try different combinations of block types and sizes.
+  auto block_shape_type =
+      internal::random<bool>()
+      ? internal::TensorBlockShapeType::kUniformAllDims
+      : internal::TensorBlockShapeType::kSkewedInnerDims;
+  auto block_target_size = internal::random<int>(1, total_coeffs);
+
+  // Pick a random dimension sizes for the tensor blocks.
+  DSizes<Index, 4> block_sizes;
+  for (int i = 0; i < 4; ++i) {
+    block_sizes[i] = internal::random<int>(1, tensor_slice_extents[i]);
+  }
+
+  TensorSliceBlockMapper block_mapper(tensor_dims, tensor_slice_offsets,
+                                      tensor_slice_extents, block_sizes,
+                                      DimensionList<Index, 4>());
+
+  for (int i = 0; i < block_mapper.total_block_count(); ++i) {
+    TensorBlock block = block_mapper.GetBlockForIndex(i, NULL);
+    UpdateCoeffSet<T, Layout, 4>(block, block.first_coeff_index(),
+                                 choose(Layout, 3, 0), &coeff_set);
+  }
+
+  VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
+}
+
+EIGEN_DECLARE_TEST(cxx11_tensor_assign) {
+  CALL_SUBTEST(test_block_mapper_sanity<ColMajor>());
+  CALL_SUBTEST(test_block_mapper_sanity<RowMajor>());
+  CALL_SUBTEST(test_block_mapper_maps_every_element<ColMajor>());
+  CALL_SUBTEST(test_block_mapper_maps_every_element<RowMajor>());
+  CALL_SUBTEST(test_slice_block_mapper_maps_every_element<ColMajor>());
+  CALL_SUBTEST(test_slice_block_mapper_maps_every_element<RowMajor>());
+}

From d55efa6f0f9ab9ec758c6b40204be476c01b7528 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Mon, 23 Jul 2018 15:50:55 -0700
Subject: [PATCH 489/680] TensorBlockIO

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 546 +++++++++++-
 .../test/cxx11_tensor_block_access.cpp        | 793 +++++++++++++++++-
 2 files changed, 1304 insertions(+), 35 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 59535cd91..8ffc9d093 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -14,6 +14,32 @@
 namespace Eigen {
 namespace internal {
 
+namespace {
+
+// Helper template to choose between ColMajor and RowMajor values.
+template <int Layout>
+struct cond;
+
+template <>
+struct cond<ColMajor> {
+  template <typename T>
+  EIGEN_STRONG_INLINE const T& operator()(const T& col,
+                                          const T& /*row*/) const {
+    return col;
+  }
+};
+
+template <>
+struct cond<RowMajor> {
+  template <typename T>
+  EIGEN_STRONG_INLINE const T& operator()(const T& /*col*/,
+                                          const T& row) const {
+    return row;
+  }
+};
+
+}  // namespace
+
 /**
  * \class TensorBlockShapeType
  * \ingroup CXX11_Tensor_Module
@@ -82,6 +108,512 @@ class TensorBlock {
   Scalar* m_data;  // Not owned.
 };
 
+template <typename Scalar, typename Index, bool Vectorizable>
+struct TensorBlockCopyOp {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Index num_coeff_to_copy, const Index dst_index,
+      const Index dst_stride, Scalar* EIGEN_RESTRICT dst_data,
+      const Index src_index, const Index src_stride,
+      const Scalar* EIGEN_RESTRICT src_data) {
+    for (Index i = 0; i < num_coeff_to_copy; ++i) {
+      dst_data[dst_index + i * dst_stride] =
+          src_data[src_index + i * src_stride];
+    }
+  }
+};
+
+// NOTE: Benchmarks run on an implementation of this that broke each of the
+// loops in these conditionals into it's own template specialization (to
+// avoid conditionals in the caller's loop) did not show an improvement.
+template <typename Scalar, typename Index>
+struct TensorBlockCopyOp<Scalar, Index, true> {
+  typedef typename packet_traits<Scalar>::type Packet;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Index num_coeff_to_copy, const Index dst_index,
+      const Index dst_stride, Scalar* EIGEN_RESTRICT dst_data,
+      const Index src_index, const Index src_stride,
+      const Scalar* EIGEN_RESTRICT src_data) {
+    if (src_stride == 1) {
+      const Index packet_size = internal::unpacket_traits<Packet>::size;
+      const Index vectorized_size =
+          (num_coeff_to_copy / packet_size) * packet_size;
+      if (dst_stride == 1) {
+        // LINEAR
+        for (Index i = 0; i < vectorized_size; i += packet_size) {
+          Packet p = internal::ploadu<Packet>(src_data + src_index + i);
+          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
+        }
+        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i] = src_data[src_index + i];
+        }
+      } else {
+        // SCATTER
+        for (Index i = 0; i < vectorized_size; i += packet_size) {
+          Packet p = internal::ploadu<Packet>(src_data + src_index + i);
+          internal::pscatter<Scalar, Packet>(
+              dst_data + dst_index + i * dst_stride, p, dst_stride);
+        }
+        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i * dst_stride] = src_data[src_index + i];
+        }
+      }
+    } else if (src_stride == 0) {
+      const Index packet_size = internal::unpacket_traits<Packet>::size;
+      const Index vectorized_size =
+          (num_coeff_to_copy / packet_size) * packet_size;
+      if (dst_stride == 1) {
+        // LINEAR
+        for (Index i = 0; i < vectorized_size; i += packet_size) {
+          Packet p = internal::pload1<Packet>(src_data + src_index);
+          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
+        }
+        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i] = src_data[src_index];
+        }
+      } else {
+        // SCATTER
+        for (Index i = 0; i < vectorized_size; i += packet_size) {
+          Packet p = internal::pload1<Packet>(src_data + src_index);
+          internal::pscatter<Scalar, Packet>(
+              dst_data + dst_index + i * dst_stride, p, dst_stride);
+        }
+        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i * dst_stride] = src_data[src_index];
+        }
+      }
+    } else {
+      if (dst_stride == 1) {
+        // GATHER
+        const Index packet_size = internal::unpacket_traits<Packet>::size;
+        const Index vectorized_size =
+            (num_coeff_to_copy / packet_size) * packet_size;
+        for (Index i = 0; i < vectorized_size; i += packet_size) {
+          Packet p = internal::pgather<Scalar, Packet>(
+              src_data + src_index + i * src_stride, src_stride);
+          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
+        }
+        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i] = src_data[src_index + i * src_stride];
+        }
+      } else {
+        // RANDOM
+        for (Index i = 0; i < num_coeff_to_copy; ++i) {
+          dst_data[dst_index + i * dst_stride] =
+              src_data[src_index + i * src_stride];
+        }
+      }
+    }
+  }
+};
+
+/**
+ * \class TensorBlockIO
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block IO class.
+ *
+ * This class is responsible for copying data between a tensor and a tensor
+ * block.
+ */
+template <typename Scalar, typename Index, int NumDims, int Layout,
+          bool Vectorizable, bool BlockRead>
+class TensorBlockIO {
+ public:
+  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      TensorBlock;
+  typedef typename internal::TensorBlockCopyOp<Scalar, Index, Vectorizable>
+      TensorBlockCopyOp;
+
+ protected:
+  struct BlockIteratorState {
+    Index input_stride;
+    Index output_stride;
+    Index input_span;
+    Index output_span;
+    Index size;
+    Index count;
+  };
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(
+      const TensorBlock& block, Index first_coeff_index,
+      const array<Index, NumDims>& tensor_to_block_dim_map,
+      const array<Index, NumDims>& tensor_strides, const Scalar* src_data,
+      Scalar* dst_data) {
+    // Find the innermost tensor dimension whose size is not 1. This is the
+    // effective inner dim. If all dimensions are of size 1, then fallback to
+    // using the actual innermost dim to avoid out-of-bound access.
+    Index num_size_one_inner_dims = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      if (block.block_sizes()[tensor_to_block_dim_map[dim]] != 1) {
+        num_size_one_inner_dims = i;
+        break;
+      }
+    }
+    // Calculate strides and dimensions.
+    const Index tensor_stride1_dim = cond<Layout>()(
+        num_size_one_inner_dims, NumDims - num_size_one_inner_dims - 1);
+    const Index block_dim_for_tensor_stride1_dim =
+        NumDims == 0 ? 1 : tensor_to_block_dim_map[tensor_stride1_dim];
+    size_t block_inner_dim_size =
+        NumDims == 0 ? 1
+                     : block.block_sizes()[block_dim_for_tensor_stride1_dim];
+    for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      const Index block_stride =
+          block.block_strides()[tensor_to_block_dim_map[dim]];
+      if (block_inner_dim_size == block_stride &&
+          block_stride == tensor_strides[dim]) {
+        block_inner_dim_size *=
+            block.block_sizes()[tensor_to_block_dim_map[dim]];
+        ++num_size_one_inner_dims;
+      } else {
+        break;
+      }
+    }
+
+    Index inputIndex;
+    Index outputIndex;
+    Index input_stride;
+    Index output_stride;
+
+    // Setup strides to read/write along the tensor's stride1 dimension.
+    if (BlockRead) {
+      inputIndex = first_coeff_index;
+      outputIndex = 0;
+      input_stride = NumDims == 0 ? 1 : tensor_strides[tensor_stride1_dim];
+      output_stride =
+          NumDims == 0
+              ? 1
+              : block.block_strides()[block_dim_for_tensor_stride1_dim];
+    } else {
+      inputIndex = 0;
+      outputIndex = first_coeff_index;
+      input_stride =
+          NumDims == 0
+              ? 1
+              : block.block_strides()[block_dim_for_tensor_stride1_dim];
+      output_stride = NumDims == 0 ? 1 : tensor_strides[tensor_stride1_dim];
+    }
+
+    const int at_least_1_dim = NumDims <= 1 ? 1 : NumDims - 1;
+    array<BlockIteratorState, at_least_1_dim> block_iter_state;
+
+    // Initialize block iterator state. Squeeze away any dimension of size 1.
+    int num_squeezed_dims = 0;
+    for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
+      const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
+      const Index size = block.block_sizes()[tensor_to_block_dim_map[dim]];
+      if (size == 1) {
+        continue;
+      }
+      block_iter_state[num_squeezed_dims].size = size;
+      if (BlockRead) {
+        block_iter_state[num_squeezed_dims].input_stride = tensor_strides[dim];
+        block_iter_state[num_squeezed_dims].output_stride =
+            block.block_strides()[tensor_to_block_dim_map[dim]];
+      } else {
+        block_iter_state[num_squeezed_dims].input_stride =
+            block.block_strides()[tensor_to_block_dim_map[dim]];
+        block_iter_state[num_squeezed_dims].output_stride = tensor_strides[dim];
+      }
+      block_iter_state[num_squeezed_dims].input_span =
+          block_iter_state[num_squeezed_dims].input_stride *
+          (block_iter_state[num_squeezed_dims].size - 1);
+      block_iter_state[num_squeezed_dims].output_span =
+          block_iter_state[num_squeezed_dims].output_stride *
+          (block_iter_state[num_squeezed_dims].size - 1);
+      block_iter_state[num_squeezed_dims].count = 0;
+      ++num_squeezed_dims;
+    }
+
+    // Iterate copying data from src to dst.
+    const Index block_total_size =
+        NumDims == 0 ? 1 : block.block_sizes().TotalSize();
+    for (Index i = 0; i < block_total_size; i += block_inner_dim_size) {
+      TensorBlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
+                             dst_data, inputIndex, input_stride, src_data);
+      // Update index.
+      for (int j = 0; j < num_squeezed_dims; ++j) {
+        if (++block_iter_state[j].count < block_iter_state[j].size) {
+          inputIndex += block_iter_state[j].input_stride;
+          outputIndex += block_iter_state[j].output_stride;
+          break;
+        }
+        block_iter_state[j].count = 0;
+        inputIndex -= block_iter_state[j].input_span;
+        outputIndex -= block_iter_state[j].output_span;
+      }
+    }
+  }
+};
+
+/**
+ * \class TensorBlockReader
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block reader class.
+ *
+ * This class is responsible for reading a tensor block.
+ *
+ */
+template <typename Scalar, typename Index, int NumDims, int Layout,
+          bool Vectorizable>
+class TensorBlockReader
+    : public TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, true> {
+ public:
+  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      TensorBlock;
+  typedef TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, true>
+      Base;
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      TensorBlock* block, const Scalar* src_data) {
+    array<Index, NumDims> tensor_to_block_dim_map;
+    for (int i = 0; i < NumDims; ++i) {
+      tensor_to_block_dim_map[i] = i;
+    }
+    Base::Copy(*block, block->first_coeff_index(), tensor_to_block_dim_map,
+               block->tensor_strides(), src_data, block->data());
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      TensorBlock* block, Index first_coeff_index,
+      const array<Index, NumDims>& tensor_to_block_dim_map,
+      const array<Index, NumDims>& tensor_strides, const Scalar* src_data) {
+    Base::Copy(*block, first_coeff_index, tensor_to_block_dim_map,
+               tensor_strides, src_data, block->data());
+  }
+};
+
+/**
+ * \class TensorBlockWriter
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block writer class.
+ *
+ * This class is responsible for writing a tensor block.
+ *
+ */
+template <typename Scalar, typename Index, int NumDims, int Layout,
+          bool Vectorizable>
+class TensorBlockWriter : public TensorBlockIO<Scalar, Index, NumDims, Layout,
+                                               Vectorizable, false> {
+ public:
+  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      TensorBlock;
+  typedef TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, false>
+      Base;
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const TensorBlock& block, Scalar* dst_data) {
+    array<Index, NumDims> tensor_to_block_dim_map;
+    for (int i = 0; i < NumDims; ++i) {
+      tensor_to_block_dim_map[i] = i;
+    }
+    Base::Copy(block, block.first_coeff_index(), tensor_to_block_dim_map,
+               block.tensor_strides(), block.data(), dst_data);
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const TensorBlock& block, Index first_coeff_index,
+      const array<Index, NumDims>& tensor_to_block_dim_map,
+      const array<Index, NumDims>& tensor_strides, Scalar* dst_data) {
+    Base::Copy(block, first_coeff_index, tensor_to_block_dim_map,
+               tensor_strides, block.data(), dst_data);
+  }
+};
+
+/**
+ * \class TensorBlockCwiseBinaryOp
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Carries out a cwise binary op on a number of coefficients.
+ *
+ * This class reads strided inputs from left and right operands, and writes the
+ * result of the cwise binary op to the strided output array.
+ *
+ */
+template <bool Vectorizable>
+struct TensorBlockCwiseBinaryOp {
+  template <typename Index, typename BinaryFunctor, typename OutputScalar,
+            typename LeftScalar, typename RightScalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const BinaryFunctor& functor, const Index num_coeff,
+      const Index output_index, const Index output_stride,
+      OutputScalar* output_data, const Index left_index,
+      const Index left_stride, const LeftScalar* left_data,
+      const Index right_index, const Index right_stride,
+      const RightScalar* right_data) {
+    for (Index i = 0; i < num_coeff; ++i) {
+      output_data[output_index + i * output_stride] =
+          functor(left_data[left_index + i * left_stride],
+                  right_data[right_index + i * right_stride]);
+    }
+  }
+};
+
+template <>
+struct TensorBlockCwiseBinaryOp<true> {
+  template <typename Index, typename BinaryFunctor, typename OutputScalar,
+            typename LeftScalar, typename RightScalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const BinaryFunctor& functor, const Index num_coeff,
+      const Index output_index, const Index output_stride,
+      OutputScalar* output_data, const Index left_index,
+      const Index left_stride, const LeftScalar* left_data,
+      const Index right_index, const Index right_stride,
+      const RightScalar* right_data) {
+    EIGEN_STATIC_ASSERT(functor_traits<BinaryFunctor>::PacketAccess,
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+    typedef typename packet_traits<OutputScalar>::type OutputPacket;
+    typedef typename packet_traits<LeftScalar>::type LeftPacket;
+    typedef typename packet_traits<RightScalar>::type RightPacket;
+    const Index packet_size = unpacket_traits<OutputPacket>::size;
+    EIGEN_STATIC_ASSERT(unpacket_traits<LeftPacket>::size == packet_size,
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT(unpacket_traits<RightPacket>::size == packet_size,
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+    const Index vectorized_size = (num_coeff / packet_size) * packet_size;
+    if (output_stride != 1 || left_stride != 1 || right_stride != 1) {
+      TensorBlockCwiseBinaryOp<false>::Run(
+          functor, num_coeff, output_index, output_stride, output_data,
+          left_index, left_stride, left_data, right_index, right_stride,
+          right_data);
+      return;
+    }
+    // Vectorization for the most common case.
+    for (Index i = 0; i < vectorized_size; i += packet_size) {
+      LeftPacket l = internal::ploadu<LeftPacket>(left_data + left_index + i);
+      RightPacket r =
+          internal::ploadu<RightPacket>(right_data + right_index + i);
+      OutputPacket p = functor.packetOp(l, r);
+      internal::pstoreu<OutputScalar, OutputPacket>(
+          output_data + output_index + i, p);
+    }
+    for (Index i = vectorized_size; i < num_coeff; ++i) {
+      output_data[output_index + i] =
+          functor(left_data[left_index + i], right_data[right_index + i]);
+    }
+  }
+};
+
+/**
+ * \class TensorBlockCwiseBinaryIO
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block IO class for carrying out cwise binary ops.
+ *
+ * This class carries out the binary op on given blocks.
+ *
+ */
+template <typename BinaryFunctor, typename Index, typename OutputScalar,
+          int NumDims, int Layout>
+struct TensorBlockCwiseBinaryIO {
+  typedef typename internal::TensorBlock<OutputScalar, Index, NumDims,
+                                         Layout>::Dimensions Dimensions;
+  typedef internal::TensorBlockCwiseBinaryOp<
+      functor_traits<BinaryFunctor>::PacketAccess>
+      TensorBlockCwiseBinaryOp;
+
+  struct BlockIteratorState {
+    Index output_stride, output_span;
+    Index left_stride, left_span;
+    Index right_stride, right_span;
+    Index size, count;
+  };
+
+  template <typename LeftScalar, typename RightScalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const BinaryFunctor& functor, const Dimensions& block_sizes,
+      const Dimensions& block_strides, OutputScalar* output_data,
+      const array<Index, NumDims>& left_strides, const LeftScalar* left_data,
+      const array<Index, NumDims>& right_strides,
+      const RightScalar* right_data) {
+    // Find the innermost dimension whose size is not 1. This is the effective
+    // inner dim. If all dimensions are of size 1, fallback to using the actual
+    // innermost dim to avoid out-of-bound access.
+    int num_size_one_inner_dims = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      if (block_sizes[dim] != 1) {
+        num_size_one_inner_dims = i;
+        break;
+      }
+    }
+    // Calculate strides and dimensions.
+    const int inner_dim =
+        NumDims == 0 ? 1
+                     : cond<Layout>()(num_size_one_inner_dims,
+                                      NumDims - num_size_one_inner_dims - 1);
+    Index inner_dim_size = NumDims == 0 ? 1 : block_sizes[inner_dim];
+    for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      // Merge multiple inner dims into one for larger inner dim size (i.e.
+      // fewer calls to TensorBlockCwiseBinaryOp::Run()).
+      if (inner_dim_size == block_strides[dim] &&
+          block_strides[dim] == left_strides[dim] &&
+          block_strides[dim] == right_strides[dim]) {
+        inner_dim_size *= block_sizes[dim];
+        ++num_size_one_inner_dims;
+      } else {
+        break;
+      }
+    }
+
+    Index output_index = 0, left_index = 0, right_index = 0;
+    const Index output_stride = NumDims == 0 ? 1 : block_strides[inner_dim];
+    const Index left_stride = NumDims == 0 ? 1 : left_strides[inner_dim];
+    const Index right_stride = NumDims == 0 ? 1 : right_strides[inner_dim];
+
+    const int at_least_1_dim = NumDims <= 1 ? 1 : NumDims - 1;
+    array<BlockIteratorState, at_least_1_dim> block_iter_state;
+
+    // Initialize block iterator state. Squeeze away any dimension of size 1.
+    int num_squeezed_dims = 0;
+    for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
+      const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
+      const Index size = block_sizes[dim];
+      if (size == 1) {
+        continue;
+      }
+      auto& state = block_iter_state[num_squeezed_dims];
+      state.output_stride = block_strides[dim];
+      state.left_stride = left_strides[dim];
+      state.right_stride = right_strides[dim];
+      state.size = size;
+      state.output_span = state.output_stride * (size - 1);
+      state.left_span = state.left_stride * (size - 1);
+      state.right_span = state.right_stride * (size - 1);
+      state.count = 0;
+      ++num_squeezed_dims;
+    }
+
+    // Compute cwise binary op.
+    const Index block_total_size = NumDims == 0 ? 1 : block_sizes.TotalSize();
+    for (Index i = 0; i < block_total_size; i += inner_dim_size) {
+      TensorBlockCwiseBinaryOp::Run(functor, inner_dim_size, output_index,
+                                    output_stride, output_data, left_index,
+                                    left_stride, left_data, right_index,
+                                    right_stride, right_data);
+      // Update index.
+      for (int j = 0; j < num_squeezed_dims; ++j) {
+        auto& state = block_iter_state[j];
+        if (++state.count < state.size) {
+          output_index += state.output_stride;
+          left_index += state.left_stride;
+          right_index += state.right_stride;
+          break;
+        }
+        state.count = 0;
+        output_index -= state.output_span;
+        left_index -= state.left_span;
+        right_index -= state.right_span;
+      }
+    }
+  }
+};
+
 /**
  * \class TensorBlockMapper
  * \ingroup CXX11_Tensor_Module
@@ -90,7 +622,7 @@ class TensorBlock {
  *
  * This class is responsible for iterating over the blocks of a tensor.
  */
-template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+template <typename Scalar, typename Index, int NumDims, int Layout>
 class TensorBlockMapper {
  public:
   typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
@@ -190,10 +722,6 @@ class TensorBlockMapper {
   }
 
  private:
-  static int InnerDimIndex(Index i) {
-    return Layout == static_cast<int>(ColMajor) ? i : NumDims - i - 1;
-  }
-
   static Dimensions BlockDimensions(const Dimensions& tensor_dims,
                                     const TensorBlockShapeType block_shape,
                                     size_t min_target_size) {
@@ -228,7 +756,7 @@ class TensorBlockMapper {
         // Add any un-allocated coefficients to inner dimension(s).
         Index total_size = block_dim_sizes.TotalSize();
         for (int i = 0; i < NumDims; ++i) {
-          const int dim = InnerDimIndex(i);
+          const int dim = cond<Layout>()(i, NumDims - i - 1);
           if (block_dim_sizes[dim] < tensor_dims[dim]) {
             const Index total_size_other_dims =
                 total_size / block_dim_sizes[dim];
@@ -245,7 +773,7 @@ class TensorBlockMapper {
       } else if (block_shape == TensorBlockShapeType::kSkewedInnerDims) {
         Index coeff_to_allocate = min_target_size;
         for (int i = 0; i < NumDims; ++i) {
-          const int dim = InnerDimIndex(i);
+          const int dim = cond<Layout>()(i, NumDims - i - 1);
           block_dim_sizes[dim] =
               numext::mini(coeff_to_allocate, tensor_dims[dim]);
           coeff_to_allocate =
@@ -284,7 +812,7 @@ class TensorBlockMapper {
  * processed together.
  *
  */
-template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+template <typename Scalar, typename Index, int NumDims, int Layout>
 class TensorSliceBlockMapper {
  public:
   typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
@@ -360,7 +888,7 @@ class TensorSliceBlockMapper {
         prev_dim = curr_dim;
       }
     } else {
-      for (int i = 0; i < static_cast<int>(NumDims) - 1; ++i) {
+      for (int i = 0; i < NumDims - 1; ++i) {
         const Index idx = block_index / m_block_strides[i];
         coords[i] = m_tensor_slice_offsets[i] + idx * m_block_dim_sizes[i];
         sizes[i] = numext::mini(
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 66e61aef1..15f2392a3 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -19,11 +19,33 @@ using Eigen::Index;
 using Eigen::RowMajor;
 using Eigen::ColMajor;
 
+using internal::TensorBlockShapeType;
+
 template<typename T>
 static const T& choose(int layout, const T& col, const T& row) {
   return layout == ColMajor ? col : row;
 }
 
+static const TensorBlockShapeType RandomShape() {
+  return internal::random<bool>()
+             ? internal::TensorBlockShapeType::kUniformAllDims
+             : internal::TensorBlockShapeType::kSkewedInnerDims;
+}
+
+template <int NumDims>
+static std::size_t RandomTargetSize(const DSizes<Index, NumDims>& dims) {
+  return internal::random<int>(1, dims.TotalSize());
+}
+
+template <typename T>
+static T* GenerateRandomData(const Index& size) {
+  T* data = new T[size];
+  for (int i = 0; i < size; ++i) {
+    data[i] = internal::random<T>();
+  }
+  return data;
+}
+
 template <int Layout>
 static void test_block_mapper_sanity()
 {
@@ -75,9 +97,7 @@ static void test_block_mapper_sanity()
 template <typename T, int Layout, int NumDims>
 static void UpdateCoeffSet(
     const internal::TensorBlock<T, Index, 4, Layout>& block,
-    Index first_coeff_index,
-    int dim_index,
-    std::set<Index>* visited_coeffs) {
+    Index first_coeff_index, int dim_index, std::set<Index>* visited_coeffs) {
   const DSizes<Index, NumDims> block_sizes = block.block_sizes();
   const DSizes<Index, NumDims> tensor_strides = block.tensor_strides();
 
@@ -103,18 +123,11 @@ static void test_block_mapper_maps_every_element()
 
   DSizes<Index, 4> dims(5, 7, 11, 17);
 
-  auto total_coeffs = static_cast<int>(dims.TotalSize());
-
   // Keep track of elements indices available via block access.
   std::set<Index> coeff_set;
 
   // Try different combinations of block types and sizes.
-  auto block_shape_type =
-      internal::random<bool>()
-          ? internal::TensorBlockShapeType::kUniformAllDims
-          : internal::TensorBlockShapeType::kSkewedInnerDims;
-  auto block_target_size = internal::random<int>(1, total_coeffs);
-  TensorBlockMapper block_mapper(dims, block_shape_type, block_target_size);
+  TensorBlockMapper block_mapper(dims, RandomShape(), RandomTargetSize(dims));
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
     TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
@@ -124,6 +137,7 @@ static void test_block_mapper_maps_every_element()
 
   // Verify that every coefficient in the original Tensor is accessible through
   // TensorBlock only once.
+  auto total_coeffs = static_cast<int>(dims.TotalSize());
   VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
   VERIFY_IS_EQUAL(*coeff_set.begin(), static_cast<Index>(0));
   VERIFY_IS_EQUAL(*coeff_set.rbegin(), static_cast<Index>(total_coeffs - 1));
@@ -146,13 +160,6 @@ static void test_slice_block_mapper_maps_every_element()
 
   auto total_coeffs = static_cast<int>(tensor_slice_extents.TotalSize());
 
-  // Try different combinations of block types and sizes.
-  auto block_shape_type =
-      internal::random<bool>()
-      ? internal::TensorBlockShapeType::kUniformAllDims
-      : internal::TensorBlockShapeType::kSkewedInnerDims;
-  auto block_target_size = internal::random<int>(1, total_coeffs);
-
   // Pick a random dimension sizes for the tensor blocks.
   DSizes<Index, 4> block_sizes;
   for (int i = 0; i < 4; ++i) {
@@ -164,7 +171,7 @@ static void test_slice_block_mapper_maps_every_element()
                                       DimensionList<Index, 4>());
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
-    TensorBlock block = block_mapper.GetBlockForIndex(i, NULL);
+    TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
     UpdateCoeffSet<T, Layout, 4>(block, block.first_coeff_index(),
                                  choose(Layout, 3, 0), &coeff_set);
   }
@@ -172,11 +179,745 @@ static void test_slice_block_mapper_maps_every_element()
   VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
 }
 
-EIGEN_DECLARE_TEST(cxx11_tensor_assign) {
-  CALL_SUBTEST(test_block_mapper_sanity<ColMajor>());
-  CALL_SUBTEST(test_block_mapper_sanity<RowMajor>());
-  CALL_SUBTEST(test_block_mapper_maps_every_element<ColMajor>());
-  CALL_SUBTEST(test_block_mapper_maps_every_element<RowMajor>());
-  CALL_SUBTEST(test_slice_block_mapper_maps_every_element<ColMajor>());
-  CALL_SUBTEST(test_slice_block_mapper_maps_every_element<RowMajor>());
+template <int Layout>
+static void test_block_io_copy_data_from_source_to_target()
+{
+  using T = float;
+
+  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
+
+  typedef internal::TensorBlockReader<T, Index, 5, Layout, true>
+      TensorBlockReader;
+  typedef internal::TensorBlockWriter<T, Index, 5, Layout, true>
+      TensorBlockWriter;
+
+  typedef std::vector<T, aligned_allocator<T>> DataVector;
+
+  DSizes<Index, 5> input_tensor_dims(5, 7, 11, 17, 3);
+  const auto input_tensor_size = input_tensor_dims.TotalSize();
+  DataVector input_data(input_tensor_size, 0);
+  for (int i = 0; i < input_tensor_size; ++i) {
+    input_data[i] = internal::random<T>();
+  }
+
+  DataVector output_data(input_tensor_size, 0);
+
+  TensorBlockMapper block_mapper(input_tensor_dims, RandomShape(),
+                                 RandomTargetSize(input_tensor_dims));
+
+  DataVector block_data(block_mapper.block_dims_total_size(), 0);
+  for (int i = 0; i < block_mapper.total_block_count(); ++i) {
+    TensorBlock block = block_mapper.GetBlockForIndex(i, block_data.data());
+    TensorBlockReader::Run(&block, input_data.data());
+    TensorBlockWriter::Run(block, output_data.data());
+  }
+
+  for (int i = 0; i < input_tensor_size; ++i) {
+    VERIFY_IS_EQUAL(input_data[i], output_data[i]);
+  }
 }
+
+template <int Layout, int NumDims>
+static int GetInputIndex(Index output_index,
+                         const array<Index, NumDims>& output_to_input_dim_map,
+                         const array<Index, NumDims>& input_strides,
+                         const array<Index, NumDims>& output_strides) {
+  int input_index = 0;
+  if (Layout == ColMajor) {
+    for (int i = NumDims - 1; i > 0; --i) {
+      const int idx = output_index / output_strides[i];
+      input_index += idx * input_strides[output_to_input_dim_map[i]];
+      output_index -= idx * output_strides[i];
+    }
+    return input_index +
+           output_index * input_strides[output_to_input_dim_map[0]];
+  } else {
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const int idx = output_index / output_strides[i];
+      input_index += idx * input_strides[output_to_input_dim_map[i]];
+      output_index -= idx * output_strides[i];
+    }
+    return input_index +
+           output_index * input_strides[output_to_input_dim_map[NumDims - 1]];
+  }
+}
+
+template <int Layout, int NumDims>
+static array<Index, NumDims> ComputeStrides(
+    const array<Index, NumDims>& sizes) {
+  array<Index, NumDims> strides;
+  if (Layout == ColMajor) {
+    strides[0] = 1;
+    for (int i = 1; i < NumDims; ++i) {
+      strides[i] = strides[i - 1] * sizes[i - 1];
+    }
+  } else {
+    strides[NumDims - 1] = 1;
+    for (int i = NumDims - 2; i >= 0; --i) {
+      strides[i] = strides[i + 1] * sizes[i + 1];
+    }
+  }
+  return strides;
+}
+
+template <int Layout>
+static void test_block_io_copy_using_reordered_dimensions() {
+  typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<float, Index, 5, Layout>
+      TensorBlockMapper;
+
+  typedef internal::TensorBlockReader<float, Index, 5, Layout, false>
+      TensorBlockReader;
+  typedef internal::TensorBlockWriter<float, Index, 5, Layout, false>
+      TensorBlockWriter;
+
+  DSizes<Index, 5> input_tensor_dims(5, 7, 11, 17, 3);
+  const auto input_tensor_size = input_tensor_dims.TotalSize();
+
+  // Create a random input tensor.
+  auto* input_data = GenerateRandomData<float>(input_tensor_size);
+
+  // Create a random dimension re-ordering/shuffle.
+  std::vector<Index> shuffle = {0, 1, 2, 3, 4};
+  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
+
+  DSizes<Index, 5> output_tensor_dims;
+  array<Index, 5> input_to_output_dim_map;
+  array<Index, 5> output_to_input_dim_map;
+  for (Index i = 0; i < 5; ++i) {
+    output_tensor_dims[shuffle[i]] = input_tensor_dims[i];
+    input_to_output_dim_map[i] = shuffle[i];
+    output_to_input_dim_map[shuffle[i]] = i;
+  }
+
+  // Random block shape and size.
+  TensorBlockMapper block_mapper(output_tensor_dims, RandomShape(),
+                                 RandomTargetSize(input_tensor_dims));
+
+  auto* block_data = new float[block_mapper.block_dims_total_size()];
+  auto* output_data = new float[input_tensor_size];
+
+  array<Index, 5> input_tensor_strides =
+      ComputeStrides<Layout, 5>(input_tensor_dims);
+  array<Index, 5> output_tensor_strides =
+      ComputeStrides<Layout, 5>(output_tensor_dims);
+
+  for (Index i = 0; i < block_mapper.total_block_count(); ++i) {
+    TensorBlock block = block_mapper.GetBlockForIndex(i, block_data);
+    const Index first_coeff_index = GetInputIndex<Layout, 5>(
+        block.first_coeff_index(), output_to_input_dim_map,
+        input_tensor_strides, output_tensor_strides);
+    TensorBlockReader::Run(&block, first_coeff_index, input_to_output_dim_map,
+                           input_tensor_strides, input_data);
+    TensorBlockWriter::Run(block, first_coeff_index, input_to_output_dim_map,
+                           input_tensor_strides, output_data);
+  }
+
+  for (int i = 0; i < input_tensor_size; ++i) {
+    VERIFY_IS_EQUAL(input_data[i], output_data[i]);
+  }
+
+  delete[] input_data;
+  delete[] block_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_block_io_zero_stride()
+{
+  typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockReader<float, Index, 5, Layout, true>
+      TensorBlockReader;
+  typedef internal::TensorBlockWriter<float, Index, 5, Layout, true>
+      TensorBlockWriter;
+
+  DSizes<Index, 5> input_tensor_dims(1, 2, 1, 3, 1);
+  const auto input_tensor_size = input_tensor_dims.TotalSize();
+
+  // Create a random input tensor.
+  auto* input_data = GenerateRandomData<float>(input_tensor_size);
+
+  DSizes<Index, 5> output_tensor_dims(3, 2, 3, 3, 2);
+
+  DSizes<Index, 5> input_tensor_strides(
+      ComputeStrides<Layout, 5>(input_tensor_dims));
+  DSizes<Index, 5> output_tensor_strides(
+      ComputeStrides<Layout, 5>(output_tensor_dims));
+
+  DSizes<Index, 5> input_tensor_strides_with_zeros(input_tensor_strides);
+  input_tensor_strides_with_zeros[0] = 0;
+  input_tensor_strides_with_zeros[2] = 0;
+  input_tensor_strides_with_zeros[4] = 0;
+
+  // Verify that data was correctly read/written from/into the block.
+  const auto verify_is_equal = [&](const float* output_data) {
+    for (int i = 0; i < output_tensor_dims[0]; ++i) {
+      for (int j = 0; j < output_tensor_dims[1]; ++j) {
+        for (int k = 0; k < output_tensor_dims[2]; ++k) {
+          for (int l = 0; l < output_tensor_dims[3]; ++l) {
+            for (int m = 0; m < output_tensor_dims[4]; ++m) {
+              const Index output_offset =
+                  i * output_tensor_strides[0] + j * output_tensor_strides[1] +
+                  k * output_tensor_strides[2] + l * output_tensor_strides[3] +
+                  m * output_tensor_strides[4];
+              const Index input_offset =
+                  i % input_tensor_dims[0] * input_tensor_strides[0] +
+                  j % input_tensor_dims[1] * input_tensor_strides[1] +
+                  k % input_tensor_dims[2] * input_tensor_strides[2] +
+                  l % input_tensor_dims[3] * input_tensor_strides[3] +
+                  m % input_tensor_dims[4] * input_tensor_strides[4];
+              VERIFY_IS_EQUAL(output_data[output_offset],
+                              input_data[input_offset]);
+            }
+          }
+        }
+      }
+    }
+  };
+
+  {
+    auto* output_data = new float[output_tensor_dims.TotalSize()];
+    TensorBlock read_block(0, output_tensor_dims, output_tensor_strides,
+                           input_tensor_strides_with_zeros, output_data);
+    TensorBlockReader::Run(&read_block, input_data);
+    verify_is_equal(output_data);
+    delete[] output_data;
+  }
+
+  {
+    auto* output_data = new float[output_tensor_dims.TotalSize()];
+    TensorBlock write_block(0, output_tensor_dims,
+                            input_tensor_strides_with_zeros,
+                            output_tensor_strides, input_data);
+    TensorBlockWriter::Run(write_block, output_data);
+    verify_is_equal(output_data);
+    delete[] output_data;
+  }
+
+  delete[] input_data;
+}
+
+template <int Layout>
+static void test_block_io_squeeze_ones() {
+  typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockReader<float, Index, 5, Layout, true>
+      TensorBlockReader;
+  typedef internal::TensorBlockWriter<float, Index, 5, Layout, true>
+      TensorBlockWriter;
+
+  // Total size > 1.
+  {
+    DSizes<Index, 5> block_sizes(1, 2, 1, 2, 1);
+    const auto total_size = block_sizes.TotalSize();
+
+    // Create a random input tensor.
+    auto* input_data = GenerateRandomData<float>(total_size);
+    DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+
+    {
+      auto* output_data = new float[block_sizes.TotalSize()];
+      TensorBlock read_block(0, block_sizes, strides, strides, output_data);
+      TensorBlockReader::Run(&read_block, input_data);
+      for (int i = 0; i < total_size; ++i) {
+        VERIFY_IS_EQUAL(output_data[i], input_data[i]);
+      }
+      delete[] output_data;
+    }
+
+    {
+      auto* output_data = new float[block_sizes.TotalSize()];
+      TensorBlock write_block(0, block_sizes, strides, strides, input_data);
+      TensorBlockWriter::Run(write_block, output_data);
+      for (int i = 0; i < total_size; ++i) {
+        VERIFY_IS_EQUAL(output_data[i], input_data[i]);
+      }
+      delete[] output_data;
+    }
+  }
+
+  // Total size == 1.
+  {
+    DSizes<Index, 5> block_sizes(1, 1, 1, 1, 1);
+    const auto total_size = block_sizes.TotalSize();
+
+    // Create a random input tensor.
+    auto* input_data = GenerateRandomData<float>(total_size);
+    DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+
+    {
+      auto* output_data = new float[block_sizes.TotalSize()];
+      TensorBlock read_block(0, block_sizes, strides, strides, output_data);
+      TensorBlockReader::Run(&read_block, input_data);
+      for (int i = 0; i < total_size; ++i) {
+        VERIFY_IS_EQUAL(output_data[i], input_data[i]);
+      }
+      delete[] output_data;
+    }
+
+    {
+      auto* output_data = new float[block_sizes.TotalSize()];
+      TensorBlock write_block(0, block_sizes, strides, strides, input_data);
+      TensorBlockWriter::Run(write_block, output_data);
+      for (int i = 0; i < total_size; ++i) {
+        VERIFY_IS_EQUAL(output_data[i], input_data[i]);
+      }
+      delete[] output_data;
+    }
+  }
+}
+
+template <int Layout>
+static void test_block_cwise_binary_io_basic() {
+  typedef internal::scalar_sum_op<float> BinaryFunctor;
+  typedef internal::TensorBlockCwiseBinaryIO<BinaryFunctor, Index, float, 5,
+                                             Layout>
+      TensorBlockCwiseBinaryIO;
+
+  DSizes<Index, 5> block_sizes(2, 3, 5, 7, 11);
+  DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+
+  const auto total_size = block_sizes.TotalSize();
+
+  // Create a random input tensors.
+  auto* left_data = GenerateRandomData<float>(total_size);
+  auto* right_data = GenerateRandomData<float>(total_size);
+
+  auto* output_data = new float[total_size];
+  BinaryFunctor functor;
+  TensorBlockCwiseBinaryIO::Run(functor, block_sizes, strides, output_data,
+                                strides, left_data, strides, right_data);
+  for (int i = 0; i < total_size; ++i) {
+    VERIFY_IS_EQUAL(output_data[i], functor(left_data[i], right_data[i]));
+  }
+
+  delete[] left_data;
+  delete[] right_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_block_cwise_binary_io_squeeze_ones() {
+  typedef internal::scalar_sum_op<float> BinaryFunctor;
+  typedef internal::TensorBlockCwiseBinaryIO<BinaryFunctor, Index, float, 5,
+                                             Layout>
+      TensorBlockCwiseBinaryIO;
+
+  DSizes<Index, 5> block_sizes(1, 2, 1, 3, 1);
+  DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+
+  const auto total_size = block_sizes.TotalSize();
+
+  // Create a random input tensors.
+  auto* left_data = GenerateRandomData<float>(total_size);
+  auto* right_data = GenerateRandomData<float>(total_size);
+
+  auto* output_data = new float[total_size];
+  BinaryFunctor functor;
+  TensorBlockCwiseBinaryIO::Run(functor, block_sizes, strides, output_data,
+                                strides, left_data, strides, right_data);
+  for (int i = 0; i < total_size; ++i) {
+    VERIFY_IS_EQUAL(output_data[i], functor(left_data[i], right_data[i]));
+  }
+
+  delete[] left_data;
+  delete[] right_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_block_cwise_binary_io_zero_strides() {
+  typedef internal::scalar_sum_op<float> BinaryFunctor;
+  typedef internal::TensorBlockCwiseBinaryIO<BinaryFunctor, Index, float, 5,
+                                             Layout>
+      TensorBlockCwiseBinaryIO;
+
+  DSizes<Index, 5> left_sizes(1, 3, 1, 7, 1);
+  DSizes<Index, 5> left_strides(ComputeStrides<Layout, 5>(left_sizes));
+  left_strides[0] = 0;
+  left_strides[2] = 0;
+  left_strides[4] = 0;
+
+  DSizes<Index, 5> right_sizes(2, 1, 5, 1, 11);
+  DSizes<Index, 5> right_strides(ComputeStrides<Layout, 5>(right_sizes));
+  right_strides[1] = 0;
+  right_strides[3] = 0;
+
+  // Generate random data.
+  auto* left_data = GenerateRandomData<float>(left_sizes.TotalSize());
+  auto* right_data = GenerateRandomData<float>(right_sizes.TotalSize());
+
+  DSizes<Index, 5> output_sizes(2, 3, 5, 7, 11);
+  DSizes<Index, 5> output_strides(ComputeStrides<Layout, 5>(output_sizes));
+
+  const auto output_total_size = output_sizes.TotalSize();
+  auto* output_data = new float[output_total_size];
+
+  BinaryFunctor functor;
+  TensorBlockCwiseBinaryIO::Run(functor, output_sizes, output_strides,
+                                output_data, left_strides, left_data,
+                                right_strides, right_data);
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 3; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        for (int l = 0; l < 7; ++l) {
+          for (int m = 0; m < 11; ++m) {
+            Index output_index = i * output_strides[0] + j * output_strides[1] +
+                                 k * output_strides[2] + l * output_strides[3] +
+                                 m * output_strides[4];
+            Index left_index = i * left_strides[0] + j * left_strides[1] +
+                               k * left_strides[2] + l * left_strides[3] +
+                               m * left_strides[4];
+            Index right_index = i * right_strides[0] + j * right_strides[1] +
+                                k * right_strides[2] + l * right_strides[3] +
+                                m * right_strides[4];
+            VERIFY_IS_EQUAL(
+                output_data[output_index],
+                functor(left_data[left_index], right_data[right_index]));
+          }
+        }
+      }
+    }
+  }
+
+  delete[] left_data;
+  delete[] right_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_uniform_block_shape()
+{
+  using T = int;
+  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
+
+  {
+    // Test shape 'UniformAllDims' with uniform 'max_coeff count'.
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 5;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    for (int i = 0; i < 5; ++i) {
+      VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'UniformAllDims' with larger 'max_coeff count' which spills
+  // partially into first inner-most dimension.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 7 * 5 * 5 * 5 * 5;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
+    for (int i = 1; i < 5; ++i) {
+      VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 6;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[4]);
+    for (int i = 3; i >= 0; --i) {
+      VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'UniformAllDims' with larger 'max_coeff count' which spills
+  // fully into first inner-most dimension.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 5 * 5 * 5 * 5;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    for (int i = 1; i < 5; ++i) {
+      VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    for (int i = 3; i >= 0; --i) {
+      VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'UniformAllDims' with larger 'max_coeff count' which spills
+  // fully into first few inner-most dimensions.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(7, 5, 6, 17, 7);
+    const size_t max_coeff_count = 7 * 5 * 6 * 7 * 5;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[4]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(7, 5, 6, 9, 7);
+    const size_t max_coeff_count = 5 * 5 * 5 * 6 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[0]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'UniformAllDims' with full allocation to all dims.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(7, 5, 6, 17, 7);
+    const size_t max_coeff_count = 7 * 5 * 6 * 17 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(7, 5, 6, 9, 7);
+    const size_t max_coeff_count = 7 * 5 * 6 * 9 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY_IS_EQUAL(9, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+}
+
+template <int Layout>
+static void test_skewed_inner_dim_block_shape()
+{
+  using T = int;
+  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
+
+  // Test shape 'SkewedInnerDims' with partial allocation to inner-most dim.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 10 * 1 * 1 * 1 * 1;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(10, block.block_sizes()[0]);
+    for (int i = 1; i < 5; ++i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 1 * 1 * 1 * 1 * 6;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[4]);
+    for (int i = 3; i >= 0; --i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'SkewedInnerDims' with full allocation to inner-most dim.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 1 * 1 * 1 * 1;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    for (int i = 1; i < 5; ++i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 1 * 1 * 1 * 1 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    for (int i = 3; i >= 0; --i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'SkewedInnerDims' with full allocation to inner-most dim,
+  // and partial allocation to second inner-dim.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 3 * 1 * 1 * 1;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    VERIFY_IS_EQUAL(3, block.block_sizes()[1]);
+    for (int i = 2; i < 5; ++i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 1 * 1 * 1 * 15 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY_IS_EQUAL(15, block.block_sizes()[3]);
+    for (int i = 2; i >= 0; --i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'SkewedInnerDims' with full allocation to inner-most dim,
+  // and partial allocation to third inner-dim.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 5 * 5 * 1 * 1;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
+    for (int i = 3; i < 5; ++i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 1 * 1 * 5 * 17 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
+    for (int i = 1; i >= 0; --i) {
+      VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
+    }
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+
+  // Test shape 'SkewedInnerDims' with full allocation to all dims.
+  if (Layout == ColMajor) {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  } else {
+    DSizes<Index, 5> dims(11, 5, 6, 17, 7);
+    const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
+    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+                                   max_coeff_count);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
+    VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
+    VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
+    VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
+    VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
+    VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
+  }
+}
+
+template <int Layout>
+static void test_empty_dims(const internal::TensorBlockShapeType block_shape)
+{
+  using T = int;
+
+  // Test blocking of tensors with zero dimensions:
+  //  - we must not crash on asserts and divisions by zero
+  //  - we must not return block with zero dimensions
+  //    (recipe for overflows/underflows, divisions by zero and NaNs later)
+  //  - total block count must be zero
+  {
+    typedef internal::TensorBlockMapper<T, Index, 1, Layout> TensorBlockMapper;
+    DSizes<Index, 1> dims(0);
+    for (int max_coeff_count = 0; max_coeff_count < 2; ++max_coeff_count) {
+      TensorBlockMapper block_mapper(dims, block_shape, max_coeff_count);
+      VERIFY_IS_EQUAL(block_mapper.total_block_count(), 0);
+      VERIFY(block_mapper.block_dims_total_size() >= 1);
+    }
+  }
+
+  {
+    typedef internal::TensorBlockMapper<T, Index, 2, Layout> TensorBlockMapper;
+    for (int dim1 = 0; dim1 < 3; ++dim1) {
+      for (int dim2 = 0; dim2 < 3; ++dim2) {
+        DSizes<Index, 2> dims(dim1, dim2);
+        for (int max_coeff_count = 0; max_coeff_count < 2; ++max_coeff_count) {
+          TensorBlockMapper block_mapper(dims, block_shape, max_coeff_count);
+          if (dim1 * dim2 == 0) {
+            VERIFY_IS_EQUAL(block_mapper.total_block_count(), 0);
+          }
+          VERIFY(block_mapper.block_dims_total_size() >= 1);
+        }
+      }
+    }
+  }
+}
+
+#define CALL_SUBTEST_LAYOUTS(NAME) \
+  CALL_SUBTEST(NAME<ColMajor>()); \
+  CALL_SUBTEST(NAME<RowMajor>())
+
+#define CALL_SUBTEST_LAYOUTS_WITH_ARG(NAME, ARG) \
+  CALL_SUBTEST(NAME<ColMajor>(ARG)); \
+  CALL_SUBTEST(NAME<RowMajor>(ARG))
+
+EIGEN_DECLARE_TEST(cxx11_tensor_assign) {
+  CALL_SUBTEST_LAYOUTS(test_block_mapper_sanity);
+  CALL_SUBTEST_LAYOUTS(test_block_mapper_maps_every_element);
+  CALL_SUBTEST_LAYOUTS(test_slice_block_mapper_maps_every_element);
+  CALL_SUBTEST_LAYOUTS(test_block_io_copy_data_from_source_to_target);
+  CALL_SUBTEST_LAYOUTS(test_block_io_copy_using_reordered_dimensions);
+  CALL_SUBTEST_LAYOUTS(test_block_io_zero_stride);
+  CALL_SUBTEST_LAYOUTS(test_block_io_squeeze_ones);
+  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_basic);
+  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_squeeze_ones);
+  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_zero_strides);
+  CALL_SUBTEST_LAYOUTS(test_uniform_block_shape);
+  CALL_SUBTEST_LAYOUTS(test_skewed_inner_dim_block_shape);
+
+  CALL_SUBTEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kUniformAllDims);
+  CALL_SUBTEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kSkewedInnerDims);
+}
+
+#undef CALL_SUBTEST_LAYOUTS
+#undef CALL_SUBTEST_LAYOUTS_WITH_ARG
\ No newline at end of file

From 636126ef406f03ad830ff9bb9927b7794b19973d Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 24 Jul 2018 20:12:49 +0200
Subject: [PATCH 490/680] Allow to filter out build-error messages

---
 CTestCustom.cmake.in              | 1 +
 cmake/EigenConfigureTesting.cmake | 1 +
 2 files changed, 2 insertions(+)

diff --git a/CTestCustom.cmake.in b/CTestCustom.cmake.in
index 9fed9d327..89e487f05 100644
--- a/CTestCustom.cmake.in
+++ b/CTestCustom.cmake.in
@@ -1,3 +1,4 @@
 
 set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_WARNINGS "2000")
 set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_ERRORS   "2000")
+list(APPEND CTEST_CUSTOM_ERROR_EXCEPTION    @EIGEN_CTEST_ERROR_EXCEPTION@)
diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake
index b02729fc8..ba88228a0 100644
--- a/cmake/EigenConfigureTesting.cmake
+++ b/cmake/EigenConfigureTesting.cmake
@@ -19,6 +19,7 @@ include(CTest)
 
 set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command of unit tests")
 set(EIGEN_DASHBOARD_BUILD_TARGET "buildtests" CACHE STRING "Target to be built in dashboard mode, default is buildtests")
+set(EIGEN_CTEST_ERROR_EXCEPTION "" CACHE STRING "Regular expression for build error messages to be filtered out")
 
 # Overwrite default DartConfiguration.tcl such that ctest can build our unit tests.
 # Recall that our unit tests are not in the "all" target, so we have to explicitly ask ctest to build our custom 'buildtests' target.

From fd4fe7cbc5559b0928661dab6d178921b50a15de Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 24 Jul 2018 22:56:15 +0200
Subject: [PATCH 491/680] Fixed issue which made documentation not getting
 built anymore

---
 Eigen/src/Core/util/Macros.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 73a9883b9..b15819f7d 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -518,6 +518,8 @@
 #endif
 
 // Does the compiler support C99?
+// Need to include <cmath> to make sure _GLIBCXX_USE_C99 gets defined
+#include <cmath>
 #ifndef EIGEN_HAS_C99_MATH
 #if EIGEN_MAX_CPP_VER>=11 && \
     ((defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901))       \

From 705f66a9caa90a32d424270d3bc99ae31925c9f2 Mon Sep 17 00:00:00 2001
From: Gustavo Lima Chaves <gustavo.lima.chaves@intel.com>
Date: Mon, 23 Jul 2018 16:29:09 -0700
Subject: [PATCH 492/680] Account for missing change on commit "Remove
 SimpleThreadPool and..."

"... always use {NonBlocking}ThreadPool". It seems the non-blocking
implementation was me the default/only one, but a reference to the old
name was left unmodified. Fix that.
---
 unsupported/test/cxx11_non_blocking_thread_pool.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/test/cxx11_non_blocking_thread_pool.cpp b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
index 5a8080ea3..e73a034b1 100644
--- a/unsupported/test/cxx11_non_blocking_thread_pool.cpp
+++ b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
@@ -18,7 +18,7 @@ static void test_create_destroy_empty_pool()
   // Just create and destroy the pool. This will wind up and tear down worker
   // threads. Ensure there are no issues in that logic.
   for (int i = 0; i < 16; ++i) {
-    NonBlockingThreadPool tp(i);
+    ThreadPool tp(i);
   }
 }
 
@@ -27,7 +27,7 @@ static void test_parallelism(bool allow_spinning)
 {
   // Test we never-ever fail to match available tasks with idle threads.
   const int kThreads = 16;  // code below expects that this is a multiple of 4
-  NonBlockingThreadPool tp(kThreads, allow_spinning);
+  ThreadPool tp(kThreads, allow_spinning);
   VERIFY_IS_EQUAL(tp.NumThreads(), kThreads);
   VERIFY_IS_EQUAL(tp.CurrentThreadId(), -1);
   for (int iter = 0; iter < 100; ++iter) {
@@ -104,7 +104,7 @@ static void test_parallelism(bool allow_spinning)
 
 static void test_cancel()
 {
-  NonBlockingThreadPool tp(2);
+  ThreadPool tp(2);
 
   // Schedule a large number of closure that each sleeps for one second. This
   // will keep the thread pool busy for much longer than the default test timeout.

From 44ee201337113eeebb1018ba8bebf110afada796 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Wed, 25 Jul 2018 20:26:15 +0200
Subject: [PATCH 493/680] Rename variable which shadows class name

---
 unsupported/Eigen/src/Polynomials/Companion.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/src/Polynomials/Companion.h b/unsupported/Eigen/src/Polynomials/Companion.h
index 41a4efc2f..126be783b 100644
--- a/unsupported/Eigen/src/Polynomials/Companion.h
+++ b/unsupported/Eigen/src/Polynomials/Companion.h
@@ -89,13 +89,13 @@ class companion
     {
       const Index deg   = m_monic.size();
       const Index deg_1 = deg-1;
-      DenseCompanionMatrixType companion(deg,deg);
-      companion <<
+      DenseCompanionMatrixType companMat(deg,deg);
+      companMat <<
         ( LeftBlock(deg,deg_1)
           << LeftBlockFirstRow::Zero(1,deg_1),
           BottomLeftBlock::Identity(deg-1,deg-1)*m_bl_diag.asDiagonal() ).finished()
         , m_monic;
-      return companion;
+      return companMat;
     }
 
 

From 5f79b7f9a9ec8addba78a28a120a4ab84e8164c3 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Wed, 25 Jul 2018 21:47:45 +0200
Subject: [PATCH 494/680] Removed several shadowing types and use global Index
 typedef everywhere

---
 .../src/MatrixFunctions/MatrixExponential.h   |  1 -
 .../src/MatrixFunctions/MatrixFunction.h      | 22 +++++--------------
 .../src/MatrixFunctions/MatrixLogarithm.h     |  4 +---
 .../Eigen/src/MatrixFunctions/MatrixPower.h   |  5 -----
 .../src/MatrixFunctions/MatrixSquareRoot.h    | 16 +++++---------
 5 files changed, 12 insertions(+), 36 deletions(-)

diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
index 03356998b..54037d58d 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
@@ -395,7 +395,6 @@ void matrix_exp_compute(const ArgType& arg, ResultType &result, false_type) // d
 template<typename Derived> struct MatrixExponentialReturnValue
 : public ReturnByValue<MatrixExponentialReturnValue<Derived> >
 {
-    typedef typename Derived::Index Index;
   public:
     /** \brief Constructor.
       *
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
index ef50c46a9..133d78625 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h
@@ -53,7 +53,7 @@ template <typename MatrixType>
 typename NumTraits<typename MatrixType::Scalar>::Real matrix_function_compute_mu(const MatrixType& A)
 {
   typedef typename plain_col_type<MatrixType>::type VectorType;
-  typename MatrixType::Index rows = A.rows();
+  Index rows = A.rows();
   const MatrixType N = MatrixType::Identity(rows, rows) - A;
   VectorType e = VectorType::Ones(rows);
   N.template triangularView<Upper>().solveInPlace(e);
@@ -65,7 +65,6 @@ MatrixType MatrixFunctionAtomic<MatrixType>::compute(const MatrixType& A)
 {
   // TODO: Use that A is upper triangular
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  typedef typename MatrixType::Index Index;
   Index rows = A.rows();
   Scalar avgEival = A.trace() / Scalar(RealScalar(rows));
   MatrixType Ashifted = A - avgEival * MatrixType::Identity(rows, rows);
@@ -131,7 +130,6 @@ typename ListOfClusters::iterator matrix_function_find_cluster(Index key, ListOf
 template <typename EivalsType, typename Cluster>
 void matrix_function_partition_eigenvalues(const EivalsType& eivals, std::list<Cluster>& clusters)
 {
-  typedef typename EivalsType::Index Index;
   typedef typename EivalsType::RealScalar RealScalar;
   for (Index i=0; i<eivals.rows(); ++i) {
     // Find cluster containing i-th ei'val, adding a new cluster if necessary
@@ -179,7 +177,7 @@ void matrix_function_compute_block_start(const VectorType& clusterSize, VectorTy
 {
   blockStart.resize(clusterSize.rows());
   blockStart(0) = 0;
-  for (typename VectorType::Index i = 1; i < clusterSize.rows(); i++) {
+  for (Index i = 1; i < clusterSize.rows(); i++) {
     blockStart(i) = blockStart(i-1) + clusterSize(i-1);
   }
 }
@@ -188,7 +186,6 @@ void matrix_function_compute_block_start(const VectorType& clusterSize, VectorTy
 template <typename EivalsType, typename ListOfClusters, typename VectorType>
 void matrix_function_compute_map(const EivalsType& eivals, const ListOfClusters& clusters, VectorType& eivalToCluster)
 {
-  typedef typename EivalsType::Index Index;
   eivalToCluster.resize(eivals.rows());
   Index clusterIndex = 0;
   for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) {
@@ -205,7 +202,6 @@ void matrix_function_compute_map(const EivalsType& eivals, const ListOfClusters&
 template <typename DynVectorType, typename VectorType>
 void matrix_function_compute_permutation(const DynVectorType& blockStart, const DynVectorType& eivalToCluster, VectorType& permutation)
 {
-  typedef typename VectorType::Index Index;
   DynVectorType indexNextEntry = blockStart;
   permutation.resize(eivalToCluster.rows());
   for (Index i = 0; i < eivalToCluster.rows(); i++) {
@@ -219,7 +215,6 @@ void matrix_function_compute_permutation(const DynVectorType& blockStart, const
 template <typename VectorType, typename MatrixType>
 void matrix_function_permute_schur(VectorType& permutation, MatrixType& U, MatrixType& T)
 {
-  typedef typename VectorType::Index Index;
   for (Index i = 0; i < permutation.rows() - 1; i++) {
     Index j;
     for (j = i; j < permutation.rows(); j++) {
@@ -247,7 +242,7 @@ template <typename MatrixType, typename AtomicType, typename VectorType>
 void matrix_function_compute_block_atomic(const MatrixType& T, AtomicType& atomic, const VectorType& blockStart, const VectorType& clusterSize, MatrixType& fT)
 { 
   fT.setZero(T.rows(), T.cols());
-  for (typename VectorType::Index i = 0; i < clusterSize.rows(); ++i) {
+  for (Index i = 0; i < clusterSize.rows(); ++i) {
     fT.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i))
       = atomic.compute(T.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i)));
   }
@@ -285,7 +280,6 @@ MatrixType matrix_function_solve_triangular_sylvester(const MatrixType& A, const
   eigen_assert(C.rows() == A.rows());
   eigen_assert(C.cols() == B.rows());
 
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
 
   Index m = A.rows();
@@ -330,11 +324,8 @@ void matrix_function_compute_above_diagonal(const MatrixType& T, const VectorTyp
 { 
   typedef internal::traits<MatrixType> Traits;
   typedef typename MatrixType::Scalar Scalar;
-  typedef typename MatrixType::Index Index;
-  static const int RowsAtCompileTime = Traits::RowsAtCompileTime;
-  static const int ColsAtCompileTime = Traits::ColsAtCompileTime;
   static const int Options = MatrixType::Options;
-  typedef Matrix<Scalar, Dynamic, Dynamic, Options, RowsAtCompileTime, ColsAtCompileTime> DynMatrixType;
+  typedef Matrix<Scalar, Dynamic, Dynamic, Options, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
 
   for (Index k = 1; k < clusterSize.rows(); k++) {
     for (Index i = 0; i < clusterSize.rows() - k; i++) {
@@ -481,7 +472,6 @@ template<typename Derived> class MatrixFunctionReturnValue
 {
   public:
     typedef typename Derived::Scalar Scalar;
-    typedef typename Derived::Index Index;
     typedef typename internal::stem_function<Scalar>::type StemFunction;
 
   protected:
@@ -506,10 +496,8 @@ template<typename Derived> class MatrixFunctionReturnValue
       typedef typename internal::nested_eval<Derived, 10>::type NestedEvalType;
       typedef typename internal::remove_all<NestedEvalType>::type NestedEvalTypeClean;
       typedef internal::traits<NestedEvalTypeClean> Traits;
-      static const int RowsAtCompileTime = Traits::RowsAtCompileTime;
-      static const int ColsAtCompileTime = Traits::ColsAtCompileTime;
       typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
-      typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, RowsAtCompileTime, ColsAtCompileTime> DynMatrixType;
+      typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
 
       typedef internal::MatrixFunctionAtomic<DynMatrixType> AtomicType;
       AtomicType atomic(m_f);
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
index ff8f6e732..a8d879a12 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h
@@ -332,10 +332,8 @@ public:
     typedef typename internal::nested_eval<Derived, 10>::type DerivedEvalType;
     typedef typename internal::remove_all<DerivedEvalType>::type DerivedEvalTypeClean;
     typedef internal::traits<DerivedEvalTypeClean> Traits;
-    static const int RowsAtCompileTime = Traits::RowsAtCompileTime;
-    static const int ColsAtCompileTime = Traits::ColsAtCompileTime;
     typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
-    typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, RowsAtCompileTime, ColsAtCompileTime> DynMatrixType;
+    typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
     typedef internal::MatrixLogarithmAtomic<DynMatrixType> AtomicType;
     AtomicType atomic;
     
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
index 33609aea9..1ceb5cf39 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h
@@ -40,7 +40,6 @@ class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParen
 {
   public:
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
 
     /**
      * \brief Constructor.
@@ -94,7 +93,6 @@ class MatrixPowerAtomic : internal::noncopyable
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
     typedef std::complex<RealScalar> ComplexScalar;
-    typedef typename MatrixType::Index Index;
     typedef Block<MatrixType,Dynamic,Dynamic> ResultType;
 
     const MatrixType& m_A;
@@ -340,7 +338,6 @@ class MatrixPower : internal::noncopyable
   private:
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
 
   public:
     /**
@@ -600,7 +597,6 @@ class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue<Deri
   public:
     typedef typename Derived::PlainObject PlainObject;
     typedef typename Derived::RealScalar RealScalar;
-    typedef typename Derived::Index Index;
 
     /**
      * \brief Constructor.
@@ -648,7 +644,6 @@ class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerRe
   public:
     typedef typename Derived::PlainObject PlainObject;
     typedef typename std::complex<typename Derived::RealScalar> ComplexScalar;
-    typedef typename Derived::Index Index;
 
     /**
      * \brief Constructor.
diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h
index afd88ec4d..34bf78913 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h
@@ -17,7 +17,7 @@ namespace internal {
 // pre:  T.block(i,i,2,2) has complex conjugate eigenvalues
 // post: sqrtT.block(i,i,2,2) is square root of T.block(i,i,2,2)
 template <typename MatrixType, typename ResultType>
-void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, typename MatrixType::Index i, ResultType& sqrtT)
+void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, Index i, ResultType& sqrtT)
 {
   // TODO: This case (2-by-2 blocks with complex conjugate eigenvalues) is probably hidden somewhere
   //       in EigenSolver. If we expose it, we could call it directly from here.
@@ -32,7 +32,7 @@ void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, typena
 //       all blocks of sqrtT to left of and below (i,j) are correct
 // post: sqrtT(i,j) has the correct value
 template <typename MatrixType, typename ResultType>
-void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT)
+void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
 {
   typedef typename traits<MatrixType>::Scalar Scalar;
   Scalar tmp = (sqrtT.row(i).segment(i+1,j-i-1) * sqrtT.col(j).segment(i+1,j-i-1)).value();
@@ -41,7 +41,7 @@ void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, ty
 
 // similar to compute1x1offDiagonalBlock()
 template <typename MatrixType, typename ResultType>
-void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT)
+void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
 {
   typedef typename traits<MatrixType>::Scalar Scalar;
   Matrix<Scalar,1,2> rhs = T.template block<1,2>(i,j);
@@ -54,7 +54,7 @@ void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, ty
 
 // similar to compute1x1offDiagonalBlock()
 template <typename MatrixType, typename ResultType>
-void matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT)
+void matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
 {
   typedef typename traits<MatrixType>::Scalar Scalar;
   Matrix<Scalar,2,1> rhs = T.template block<2,1>(i,j);
@@ -101,7 +101,7 @@ void matrix_sqrt_quasi_triangular_solve_auxiliary_equation(MatrixType& X, const
 
 // similar to compute1x1offDiagonalBlock()
 template <typename MatrixType, typename ResultType>
-void matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT)
+void matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
 {
   typedef typename traits<MatrixType>::Scalar Scalar;
   Matrix<Scalar,2,2> A = sqrtT.template block<2,2>(i,i);
@@ -120,7 +120,6 @@ template <typename MatrixType, typename ResultType>
 void matrix_sqrt_quasi_triangular_diagonal(const MatrixType& T, ResultType& sqrtT)
 {
   using std::sqrt;
-  typedef typename MatrixType::Index Index;
   const Index size = T.rows();
   for (Index i = 0; i < size; i++) {
     if (i == size - 1 || T.coeff(i+1, i) == 0) {
@@ -139,7 +138,6 @@ void matrix_sqrt_quasi_triangular_diagonal(const MatrixType& T, ResultType& sqrt
 template <typename MatrixType, typename ResultType>
 void matrix_sqrt_quasi_triangular_off_diagonal(const MatrixType& T, ResultType& sqrtT)
 {
-  typedef typename MatrixType::Index Index;
   const Index size = T.rows();
   for (Index j = 1; j < size; j++) {
       if (T.coeff(j, j-1) != 0)  // if T(j-1:j, j-1:j) is a 2-by-2 block
@@ -206,8 +204,7 @@ template <typename MatrixType, typename ResultType>
 void matrix_sqrt_triangular(const MatrixType &arg, ResultType &result)
 {
   using std::sqrt;
-  typedef typename MatrixType::Index Index;
-      typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::Scalar Scalar;
 
   eigen_assert(arg.rows() == arg.cols());
 
@@ -318,7 +315,6 @@ template<typename Derived> class MatrixSquareRootReturnValue
 : public ReturnByValue<MatrixSquareRootReturnValue<Derived> >
 {
   protected:
-    typedef typename Derived::Index Index;
     typedef typename internal::ref_selector<Derived>::type DerivedNested;
 
   public:

From 5e79402b4a742ef33574a568689c70be4f3d8549 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 24 Jul 2018 21:59:15 +0200
Subject: [PATCH 495/680] fix warnings for doc-eigen-prerequisites

---
 doc/snippets/MatrixBase_cwiseEqual.cpp    | 2 +-
 doc/snippets/MatrixBase_cwiseNotEqual.cpp | 2 +-
 unsupported/doc/examples/FFT.cpp          | 6 +++---
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/doc/snippets/MatrixBase_cwiseEqual.cpp b/doc/snippets/MatrixBase_cwiseEqual.cpp
index eb3656f4c..469af642c 100644
--- a/doc/snippets/MatrixBase_cwiseEqual.cpp
+++ b/doc/snippets/MatrixBase_cwiseEqual.cpp
@@ -3,5 +3,5 @@ m << 1, 0,
      1, 1;
 cout << "Comparing m with identity matrix:" << endl;
 cout << m.cwiseEqual(MatrixXi::Identity(2,2)) << endl;
-int count = m.cwiseEqual(MatrixXi::Identity(2,2)).count();
+Index count = m.cwiseEqual(MatrixXi::Identity(2,2)).count();
 cout << "Number of coefficients that are equal: " << count << endl;
diff --git a/doc/snippets/MatrixBase_cwiseNotEqual.cpp b/doc/snippets/MatrixBase_cwiseNotEqual.cpp
index 6a2e4fb6c..7f0a105d6 100644
--- a/doc/snippets/MatrixBase_cwiseNotEqual.cpp
+++ b/doc/snippets/MatrixBase_cwiseNotEqual.cpp
@@ -3,5 +3,5 @@ m << 1, 0,
      1, 1;
 cout << "Comparing m with identity matrix:" << endl;
 cout << m.cwiseNotEqual(MatrixXi::Identity(2,2)) << endl;
-int count = m.cwiseNotEqual(MatrixXi::Identity(2,2)).count();
+Index count = m.cwiseNotEqual(MatrixXi::Identity(2,2)).count();
 cout << "Number of coefficients that are not equal: " << count << endl;
diff --git a/unsupported/doc/examples/FFT.cpp b/unsupported/doc/examples/FFT.cpp
index fcbf81276..85e8a0241 100644
--- a/unsupported/doc/examples/FFT.cpp
+++ b/unsupported/doc/examples/FFT.cpp
@@ -61,14 +61,14 @@ template <typename T>
 void RandomFill(std::vector<T> & vec)
 {
     for (size_t k=0;k<vec.size();++k)
-        vec[k] = T( rand() )/T(RAND_MAX) - .5;
+        vec[k] = T( rand() )/T(RAND_MAX) - T(.5);
 }
 
 template <typename T>
 void RandomFill(std::vector<std::complex<T> > & vec)
 {
     for (size_t k=0;k<vec.size();++k)
-        vec[k] = std::complex<T> ( T( rand() )/T(RAND_MAX) - .5, T( rand() )/T(RAND_MAX) - .5);
+        vec[k] = std::complex<T> ( T( rand() )/T(RAND_MAX) - T(.5), T( rand() )/T(RAND_MAX) - T(.5));
 }
 
 template <typename T_time,typename T_freq>
@@ -85,7 +85,7 @@ void fwd_inv(size_t nfft)
     vector<T_time> timebuf2;
     fft.inv(timebuf2,freqbuf);
 
-    long double rmse = mag2(timebuf - timebuf2) / mag2(timebuf);
+    T_time rmse = mag2(timebuf - timebuf2) / mag2(timebuf);
     cout << "roundtrip rmse: " << rmse << endl;
 }
 

From 397b0547e1e3151baa64c9677bef5882ad24d1ea Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 26 Jul 2018 00:01:24 +0200
Subject: [PATCH 496/680] DIsable static assertions only when necessary and
 disable double-promotion warnings in that case as well

---
 test/mixingtypes.cpp | 58 ++++++++++++++++++++++++++++++++------------
 1 file changed, 43 insertions(+), 15 deletions(-)

diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp
index 38f062f1e..aad63ec2b 100644
--- a/test/mixingtypes.cpp
+++ b/test/mixingtypes.cpp
@@ -8,13 +8,27 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-// work around "uninitialized" warnings and give that option some testing
-#define EIGEN_INITIALIZE_MATRICES_BY_ZERO
+#if defined(EIGEN_TEST_PART_7)
 
 #ifndef EIGEN_NO_STATIC_ASSERT
 #define EIGEN_NO_STATIC_ASSERT // turn static asserts into runtime asserts in order to check them
 #endif
 
+// ignore double-promotion diagnostic for clang and gcc, if we check for static assertion anyway:
+// TODO do the same for MSVC?
+#if defined(__clang__)
+#  if (__clang_major__ * 100 + __clang_minor__) >= 308
+#    pragma clang diagnostic ignored "-Wdouble-promotion"
+#  endif
+#elif defined(__GNUC__)
+  // TODO is there a minimal GCC version for this? At least g++-4.7 seems to be fine with this.
+#  pragma GCC diagnostic ignored "-Wdouble-promotion"
+#endif
+
+#endif
+
+
+
 #if defined(EIGEN_TEST_PART_1) || defined(EIGEN_TEST_PART_2) || defined(EIGEN_TEST_PART_3)
 
 #ifndef EIGEN_DONT_VECTORIZE
@@ -35,6 +49,28 @@ using namespace std;
   VERIFY_IS_APPROX(XPR,REF); \
   VERIFY( g_called && #XPR" not properly optimized");
 
+template<int SizeAtCompileType>
+void raise_assertion(Index size = SizeAtCompileType)
+{
+  // VERIFY_RAISES_ASSERT(mf+md); // does not even compile
+  Matrix<float, SizeAtCompileType, 1> vf; vf.setRandom(size);
+  Matrix<double, SizeAtCompileType, 1> vd; vd.setRandom(size);
+  VERIFY_RAISES_ASSERT(vf=vd);
+  VERIFY_RAISES_ASSERT(vf+=vd);
+  VERIFY_RAISES_ASSERT(vf-=vd);
+  VERIFY_RAISES_ASSERT(vd=vf);
+  VERIFY_RAISES_ASSERT(vd+=vf);
+  VERIFY_RAISES_ASSERT(vd-=vf);
+
+  //   vd.asDiagonal() * mf;    // does not even compile
+  //   vcd.asDiagonal() * mf;   // does not even compile
+
+#if 0 // we get other compilation errors here than just static asserts
+  VERIFY_RAISES_ASSERT(vd.dot(vf));
+#endif
+}
+
+
 template<int SizeAtCompileType> void mixingtypes(int size = SizeAtCompileType)
 {
   typedef std::complex<float>   CF;
@@ -73,13 +109,6 @@ template<int SizeAtCompileType> void mixingtypes(int size = SizeAtCompileType)
   while(std::abs(scf)<epsf) scf = internal::random<CF>();
   while(std::abs(scd)<epsd) scd = internal::random<CD>();
 
-//   VERIFY_RAISES_ASSERT(mf+md); // does not even compile
-
-#ifdef EIGEN_DONT_VECTORIZE
-  VERIFY_RAISES_ASSERT(vf=vd);
-  VERIFY_RAISES_ASSERT(vf+=vd);
-#endif
-  
   // check scalar products
   VERIFY_MIX_SCALAR(vcf * sf , vcf * complex<float>(sf));
   VERIFY_MIX_SCALAR(sd * vcd , complex<double>(sd) * vcd);
@@ -119,9 +148,6 @@ template<int SizeAtCompileType> void mixingtypes(int size = SizeAtCompileType)
 
   // check dot product
   vf.dot(vf);
-#if 0 // we get other compilation errors here than just static asserts
-  VERIFY_RAISES_ASSERT(vd.dot(vf));
-#endif
   VERIFY_IS_APPROX(vcf.dot(vf), vcf.dot(vf.template cast<complex<float> >()));
 
   // check diagonal product
@@ -130,9 +156,6 @@ template<int SizeAtCompileType> void mixingtypes(int size = SizeAtCompileType)
   VERIFY_IS_APPROX(mcf * vf.asDiagonal(), mcf * vf.template cast<complex<float> >().asDiagonal());
   VERIFY_IS_APPROX(md * vcd.asDiagonal(), md.template cast<complex<double> >() * vcd.asDiagonal());
 
-//   vd.asDiagonal() * mf;    // does not even compile
-//   vcd.asDiagonal() * mf;   // does not even compile
-
   // check inner product
   VERIFY_IS_APPROX((vf.transpose() * vcf).value(), (vf.template cast<complex<float> >().transpose() * vcf).value());
 
@@ -296,5 +319,10 @@ EIGEN_DECLARE_TEST(mixingtypes)
     CALL_SUBTEST_4(mixingtypes<3>());
     CALL_SUBTEST_5(mixingtypes<4>());
     CALL_SUBTEST_6(mixingtypes<Dynamic>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE)));
+    CALL_SUBTEST_7(raise_assertion<Dynamic>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE)));
   }
+  CALL_SUBTEST_7(raise_assertion<0>());
+  CALL_SUBTEST_7(raise_assertion<3>());
+  CALL_SUBTEST_7(raise_assertion<4>());
+  CALL_SUBTEST_7(raise_assertion<Dynamic>(0));
 }

From 2ebcb911b27174c5402e4c7af3d2738fd042a5e2 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 26 Jul 2018 14:28:48 -0700
Subject: [PATCH 497/680] Add pcast packet op for NEON.

---
 Eigen/Core                             |  1 +
 Eigen/src/Core/arch/NEON/TypeCasting.h | 48 ++++++++++++++++++++++++++
 2 files changed, 49 insertions(+)
 create mode 100644 Eigen/src/Core/arch/NEON/TypeCasting.h

diff --git a/Eigen/Core b/Eigen/Core
index f336d407b..864bde551 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -179,6 +179,7 @@ using std::ptrdiff_t;
   #include "src/Core/arch/NEON/PacketMath.h"
   #include "src/Core/arch/NEON/MathFunctions.h"
   #include "src/Core/arch/NEON/Complex.h"
+  #include "src/Core/arch/NEON/TypeCasting.h"
 #elif defined EIGEN_VECTORIZE_ZVECTOR
   #include "src/Core/arch/ZVector/PacketMath.h"
   #include "src/Core/arch/ZVector/MathFunctions.h"
diff --git a/Eigen/src/Core/arch/NEON/TypeCasting.h b/Eigen/src/Core/arch/NEON/TypeCasting.h
new file mode 100644
index 000000000..95d1fd0e4
--- /dev/null
+++ b/Eigen/src/Core/arch/NEON/TypeCasting.h
@@ -0,0 +1,48 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_NEON_H
+#define EIGEN_TYPE_CASTING_NEON_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
+  return vcvtq_s32_f32(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
+  return vcvtq_f32_s32(a);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_NEON_H

From e4785326255c536214d2cead384477c35e3bdcc6 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 27 Jul 2018 12:36:34 -0700
Subject: [PATCH 498/680] Reduce the number of template specializations of
 classes related to tensor contraction to reduce binary size.

---
 .../CXX11/src/Tensor/TensorContraction.h      | 105 +++++++++-------
 .../src/Tensor/TensorContractionBlocking.h    |   7 +-
 .../src/Tensor/TensorContractionThreadPool.h  | 118 ++++++++----------
 3 files changed, 109 insertions(+), 121 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 0e69cd40c..57b5339d1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -177,9 +177,9 @@ struct NoOpOutputKernel {
    */
   template <typename Index, typename Scalar>
   EIGEN_ALWAYS_INLINE void operator()(
-      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
-      const TensorContractionParams& params, Index i, Index j, Index num_rows,
-      Index num_cols) const {}
+      const OutputKernel::OutputMapper<Index, Scalar>& /*output_mapper*/,
+      const TensorContractionParams& /*params*/, Index /*i*/,
+      Index /*j*/, Index /*num_rows*/, Index /*num_cols*/) const {}
 };
 
 template<typename Indices, typename LhsXprType, typename RhsXprType, typename OutputKernelType = const NoOpOutputKernel>
@@ -467,42 +467,58 @@ struct TensorContractionEvaluatorBase
     }
   }
 
-  EIGEN_DEVICE_FUNC void evalTo(Scalar* buffer) const {
-    if (this->m_lhs_inner_dim_contiguous) {
-      if (this->m_rhs_inner_dim_contiguous) {
-        if (this->m_rhs_inner_dim_reordered) {
-          static_cast<const Derived*>(this)->template evalProduct<true, true, true, Unaligned>(buffer);
-        }
-        else {
-          static_cast<const Derived*>(this)->template evalProduct<true, true, false, Unaligned>(buffer);
-        }
-      }
-      else {
-       if (this->m_rhs_inner_dim_reordered) {
-          static_cast<const Derived*>(this)->template evalProduct<true, false, true, Unaligned>(buffer);
-        }
-        else {
-          static_cast<const Derived*>(this)->template evalProduct<true, false, false, Unaligned>(buffer);
-        }
-      }
+#define TENSOR_CONTRACTION_DISPATCH(METHOD, ALIGNMENT, ARGS)   \
+    if (this->m_lhs_inner_dim_contiguous) { \
+      if (this->m_rhs_inner_dim_contiguous) { \
+        if (this->m_rhs_inner_dim_reordered) { \
+          METHOD<true, true, true, ALIGNMENT>ARGS;    \
+        } \
+        else { \
+          METHOD<true, true, false, ALIGNMENT>ARGS; \
+        } \
+      } \
+      else { \
+       if (this->m_rhs_inner_dim_reordered) { \
+          METHOD<true, false, true, ALIGNMENT>ARGS; \
+        } \
+        else { \
+          METHOD<true, false, false, ALIGNMENT>ARGS; \
+        } \
+      } \
+    } \
+    else { \
+      if (this->m_rhs_inner_dim_contiguous) { \
+        if (this->m_rhs_inner_dim_reordered) { \
+          METHOD<false, true, true, ALIGNMENT>ARGS; \
+        } \
+        else { \
+          METHOD<false, true, false, ALIGNMENT>ARGS; \
+        } \
+      } \
+      else { \
+       if (this->m_rhs_inner_dim_reordered) { \
+          METHOD<false, false, true, ALIGNMENT>ARGS; \
+        } \
+        else { \
+          METHOD<false, false, false, ALIGNMENT>ARGS; \
+        } \
+      } \
     }
-    else {
-      if (this->m_rhs_inner_dim_contiguous) {
-        if (this->m_rhs_inner_dim_reordered) {
-          static_cast<const Derived*>(this)->template evalProduct<false, true, true, Unaligned>(buffer);
-        }
-        else {
-          static_cast<const Derived*>(this)->template evalProduct<false, true, false, Unaligned>(buffer);
-        }
-      }
-      else {
-       if (this->m_rhs_inner_dim_reordered) {
-          static_cast<const Derived*>(this)->template evalProduct<false, false, true, Unaligned>(buffer);
-        }
-        else {
-          static_cast<const Derived*>(this)->template evalProduct<false, false, false, Unaligned>(buffer);
-        }
-      }
+
+  EIGEN_DEVICE_FUNC void evalTo(Scalar* buffer) const {
+   static_cast<const Derived*>(this)->template evalProduct<Unaligned>(buffer);
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
+            bool rhs_inner_dim_reordered, int Alignment>
+  void evalProductSequential(Scalar* buffer) const {
+    if (this->m_j_size == 1) {
+      this->template evalGemv<lhs_inner_dim_contiguous,
+                              rhs_inner_dim_contiguous, rhs_inner_dim_reordered,
+                              Alignment>(buffer);
+    } else {
+      this->template evalGemm<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous,
+                              rhs_inner_dim_reordered, Alignment>(buffer);
     }
   }
 
@@ -623,7 +639,7 @@ struct TensorContractionEvaluatorBase
     OutputMapper output(buffer, m);
 
     // Sizes of the blocks to load in cache. See the Goto paper for details.
-    internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index, internal::ShardByCol> blocking(k, m, n, 1);
+    internal::TensorContractionBlocking<LhsScalar, RhsScalar, Index, internal::ShardByCol> blocking(k, m, n, 1);
     const Index kc = blocking.kc();
     const Index mc = numext::mini(m, blocking.mc());
     const Index nc = numext::mini(n, blocking.nc());
@@ -976,14 +992,9 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) :
       Base(op, device) { }
 
-  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  EIGEN_DEVICE_FUNC void evalProduct(Scalar* buffer) const {
-    if (this->m_j_size == 1) {
-      this->template evalGemv<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
-      return;
-    }
-
-    this->template evalGemm<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
+  template <int Alignment>
+  void evalProduct(Scalar* buffer) const {
+    TENSOR_CONTRACTION_DISPATCH(this->template evalProductSequential, Alignment, (buffer));
   }
 };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
index 8c1af1da8..cf281192c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -21,13 +21,10 @@ enum {
 
 
 // Default Blocking Strategy
-template <typename LhsMapper, typename RhsMapper, typename Index, int ShardingType=ShardByCol>
+template <typename LhsScalar, typename RhsScalar, typename Index, int ShardingType=ShardByCol>
 class TensorContractionBlocking {
  public:
 
-  typedef typename LhsMapper::Scalar LhsScalar;
-  typedef typename RhsMapper::Scalar RhsScalar;
-
  /*
    adding EIGEN_DEVICE_FUNC unconditionally to 'TensorContractionBlocking' constructor in `TensorContractionBlocking.h`
      requires adding EIGEN_DEVICE_FUNC to `computeProductBlockingSizes` in `GeneralBlockPanelKernel.h`
@@ -41,7 +38,7 @@ class TensorContractionBlocking {
    ../Eigen/src/Core/products/GeneralBlockPanelKernel.h(57): error #2901:
       dynamic initialization is not supported for function-scope static variables within a __device__/__global__ function
  */
- 
+
   #if !defined(EIGEN_HIPCC)
   EIGEN_DEVICE_FUNC
   #endif
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 8b86d7aaf..182c5f7f9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -71,8 +71,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
   TensorEvaluator(const XprType& op, const Device& device) :
       Base(op, device) {}
 
-  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
-            bool rhs_inner_dim_reordered, int Alignment>
+  template <int Alignment>
   void evalProduct(Scalar* buffer) const {
     const Index m = this->m_i_size;
     const Index n = this->m_j_size;
@@ -96,39 +95,6 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     }
 #endif
 
-    typedef
-        typename internal::remove_const<typename EvalLeftArgType::Scalar>::type
-            LhsScalar;
-    typedef
-        typename internal::remove_const<typename EvalRightArgType::Scalar>::type
-            RhsScalar;
-    typedef typename internal::gebp_traits<LhsScalar, RhsScalar> Traits;
-    typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
-    typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
-    typedef internal::TensorContractionInputMapper<
-        LhsScalar, Index, internal::Lhs, LeftEvaluator, left_nocontract_t,
-        contract_t, internal::packet_traits<LhsScalar>::size,
-        lhs_inner_dim_contiguous, false, Unaligned>
-        LhsMapper;
-    typedef internal::TensorContractionInputMapper<
-        RhsScalar, Index, internal::Rhs, RightEvaluator, right_nocontract_t,
-        contract_t, internal::packet_traits<RhsScalar>::size,
-        rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Unaligned>
-        RhsMapper;
-    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
-    typedef internal::gemm_pack_lhs<LhsScalar, Index,
-                                    typename LhsMapper::SubMapper, Traits::mr,
-                                    Traits::LhsProgress, ColMajor>
-        LhsPacker;
-    typedef internal::gemm_pack_rhs<
-        RhsScalar, Index, typename RhsMapper::SubMapper, Traits::nr, ColMajor>
-        RhsPacker;
-    typedef internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper,
-                                  Traits::mr, Traits::nr, false, false>
-        GebpKernel;
-
-
-
     // Compute a set of algorithm parameters:
     // - kernel block sizes (bm, bn, bk)
     // - task grain sizes (number of kernels executed per task: gm, gn)
@@ -158,14 +124,14 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     // Again, we don't know number of threads yet, so we use 2.
     Index bm, bn, bk;
     if (shard_by_col) {
-      internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index,
+      internal::TensorContractionBlocking<LhsScalar, RhsScalar, Index,
                                           internal::ShardByCol>
           blocking(k, m, n, 2);
       bm = blocking.mc();
       bn = blocking.nc();
       bk = blocking.kc();
     } else {
-      internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index,
+      internal::TensorContractionBlocking<LhsScalar, RhsScalar, Index,
                                           internal::ShardByRow>
           blocking(k, m, n, 2);
       bm = blocking.mc();
@@ -187,29 +153,22 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     if (n == 1) num_threads = 1;
 
     if (num_threads == 1) {
-      // The single-threaded algorithm should be faster in this case.
-      if (n == 1)
-        this->template evalGemv<lhs_inner_dim_contiguous,
-                                rhs_inner_dim_contiguous,
-                                rhs_inner_dim_reordered, Alignment>(buffer);
-      else
-        this->template evalGemm<lhs_inner_dim_contiguous,
-                                rhs_inner_dim_contiguous,
-                                rhs_inner_dim_reordered, Alignment>(buffer);
+      TENSOR_CONTRACTION_DISPATCH(this->template evalProductSequential,
+                                  Unaligned, (buffer));
       return;
     }
 
     // Now that we know number of threads, recalculate sharding and blocking.
     shard_by_col = shardByCol(m, n, num_threads);
     if (shard_by_col) {
-      internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index,
+      internal::TensorContractionBlocking<LhsScalar, RhsScalar, Index,
                                           internal::ShardByCol>
           blocking(k, m, n, num_threads);
       bm = blocking.mc();
       bn = blocking.nc();
       bk = blocking.kc();
     } else {
-      internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index,
+      internal::TensorContractionBlocking<LhsScalar, RhsScalar, Index,
                                           internal::ShardByRow>
           blocking(k, m, n, num_threads);
       bm = blocking.mc();
@@ -257,34 +216,55 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     // more important in this case.
     if ((shard_by_col ? nm : nn) == 1) parallel_pack = false;
 
-    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides,
-                  this->m_i_strides, this->m_left_contracting_strides,
-                  this->m_k_strides);
+    #define CONTEXT_ARGS                                                        \
+  (this, num_threads, buffer, m, n, k, bm, bn, bk, nm, nn, nk, gm, gn, nm0, \
+   nn0, shard_by_col, parallel_pack)                                        \
+      .run()
 
-    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides,
-                  this->m_j_strides, this->m_right_contracting_strides,
-                  this->m_k_strides);
+    TENSOR_CONTRACTION_DISPATCH(Context, Alignment, CONTEXT_ARGS);
+
+#undef CONTEXT_ARGS
 
-    Context<LhsPacker, RhsPacker, GebpKernel, LhsMapper, RhsMapper,
-            OutputMapper>(this, num_threads, lhs, rhs, buffer, m, n,
-                          k, bm, bn, bk, nm, nn, nk, gm, gn, nm0, nn0,
-                          shard_by_col, parallel_pack)
-        .run();
   }
 
   // Context coordinates a single parallel gemm operation.
-  template <typename LhsPacker, typename RhsPacker, typename GebpKernel,
-            typename LhsMapper, typename RhsMapper, typename OutputMapper>
+ template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
+            bool rhs_inner_dim_reordered, int Alignment>
   class Context {
    public:
-    Context(const Self* self, int num_threads, LhsMapper& lhs,
-            RhsMapper& rhs, Scalar* buffer, Index tm, Index tn, Index tk, Index bm,
-            Index bn, Index bk, Index nm, Index nn, Index nk, Index gm,
-            Index gn, Index nm0, Index nn0, bool shard_by_col,
+    typedef internal::TensorContractionInputMapper<
+        LhsScalar, Index, internal::Lhs, LeftEvaluator, left_nocontract_t,
+        contract_t, internal::packet_traits<LhsScalar>::size,
+        lhs_inner_dim_contiguous, false, Unaligned>
+        LhsMapper;
+    typedef internal::TensorContractionInputMapper<
+        RhsScalar, Index, internal::Rhs, RightEvaluator, right_nocontract_t,
+        contract_t, internal::packet_traits<RhsScalar>::size,
+        rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Unaligned>
+        RhsMapper;
+    typedef internal::gemm_pack_lhs<LhsScalar, Index,
+                                    typename LhsMapper::SubMapper, Traits::mr,
+                                    Traits::LhsProgress, ColMajor>
+        LhsPacker;
+    typedef internal::gemm_pack_rhs<
+        RhsScalar, Index, typename RhsMapper::SubMapper, Traits::nr, ColMajor>
+        RhsPacker;
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+    typedef internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper,
+                                  Traits::mr, Traits::nr, false, false>
+        GebpKernel;
+
+    Context(const Self* self, int num_threads, Scalar* buffer, Index tm, Index tn,
+            Index tk, Index bm, Index bn, Index bk, Index nm, Index nn, Index nk,
+            Index gm, Index gn, Index nm0, Index nn0, bool shard_by_col,
             bool parallel_pack)
         : device_(self->m_device),
-          lhs_(lhs),
-          rhs_(rhs),
+          lhs_(self->m_leftImpl, self->m_left_nocontract_strides,
+               self->m_i_strides, self->m_left_contracting_strides,
+               self->m_k_strides),
+          rhs_(self->m_rightImpl, self->m_right_nocontract_strides,
+               self->m_j_strides, self->m_right_contracting_strides,
+               self->m_k_strides),
           buffer_(buffer),
           output_(buffer, tm),
           output_kernel_(self->m_output_kernel),
@@ -376,8 +356,8 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
    private:
     Notification done_;
     const Device& device_;
-    LhsMapper& lhs_;
-    RhsMapper& rhs_;
+    LhsMapper lhs_;
+    RhsMapper rhs_;
     Scalar* const buffer_;
     OutputMapper output_;
     OutputKernelType output_kernel_;

From 96b030a8e45a69d91fbd1f9ee1d9a48721f20e28 Mon Sep 17 00:00:00 2001
From: Mark D Ryan <mark.d.ryan@intel.com>
Date: Mon, 30 Jul 2018 10:19:51 +0100
Subject: [PATCH 499/680] Re-enable FMA for fast sqrt functions

This commit re-enables the use of FMA for the FAST sqrt functions.
Doing so improves the performance of both algorithms.  The float32
version is now 88% the speed of the original function, while the
double version is 90%.

From bc615e458559480a95a6b6de32295b4aa646c72f Mon Sep 17 00:00:00 2001
From: Mark D Ryan <mark.d.ryan@intel.com>
Date: Mon, 30 Jul 2018 13:21:00 +0200
Subject: [PATCH 500/680] Re-enable FMA for fast sqrt functions

---
 Eigen/src/Core/arch/AVX512/MathFunctions.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 81a3b4f62..ba1246722 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -258,7 +258,7 @@ pexp<Packet8d>(const Packet8d& _x) {
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
 psqrt<Packet16f>(const Packet16f& _x) {
-  Packet16f half = pmul(_x, pset1<Packet16f>(.5f));
+  Packet16f neg_half = pmul(_x, pset1<Packet16f>(-.5f));
   __mmask16 denormal_mask = _mm512_kand(
       _mm512_cmp_ps_mask(_x, pset1<Packet16f>((std::numeric_limits<float>::min)()),
                         _CMP_LT_OQ),
@@ -267,7 +267,7 @@ psqrt<Packet16f>(const Packet16f& _x) {
   Packet16f x = _mm512_rsqrt14_ps(_x);
 
   // Do a single step of Newton's iteration.
-  x = pmul(x, psub(pset1<Packet16f>(1.5f), pmul(half, pmul(x,x))));
+  x = pmul(x, pmadd(neg_half, pmul(x, x), pset1<Packet16f>(1.5f)));
 
   // Flush results for denormals to zero.
   return _mm512_mask_blend_ps(denormal_mask, pmul(_x,x), _mm512_setzero_ps());
@@ -276,7 +276,7 @@ psqrt<Packet16f>(const Packet16f& _x) {
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
 psqrt<Packet8d>(const Packet8d& _x) {
-  Packet8d half = pmul(_x, pset1<Packet8d>(.5f));
+  Packet8d neg_half = pmul(_x, pset1<Packet8d>(-.5f));
   __mmask16 denormal_mask = _mm512_kand(
       _mm512_cmp_pd_mask(_x, pset1<Packet8d>((std::numeric_limits<double>::min)()),
                         _CMP_LT_OQ),
@@ -285,10 +285,10 @@ psqrt<Packet8d>(const Packet8d& _x) {
   Packet8d x = _mm512_rsqrt14_pd(_x);
 
   // Do a single step of Newton's iteration.
-  x = pmul(x, psub(pset1<Packet8d>(1.5f), pmul(half, pmul(x,x))));
+  x = pmul(x, pmadd(neg_half, pmul(x, x), pset1<Packet8d>(1.5f)));
 
   // Do a second step of Newton's iteration.
-  x = pmul(x, psub(pset1<Packet8d>(1.5f), pmul(half, pmul(x,x))));
+  x = pmul(x, pmadd(neg_half, pmul(x, x), pset1<Packet8d>(1.5f)));
 
   return _mm512_mask_blend_pd(denormal_mask, pmul(_x,x), _mm512_setzero_pd());
 }

From f5cace5e9fc734a9f26b9cf14e806bd1b115e443 Mon Sep 17 00:00:00 2001
From: Patrik Huber <patrikhuber@gmail.com>
Date: Thu, 26 Jul 2018 19:55:19 +0000
Subject: [PATCH 501/680] Fix two small typos in the documentation

---
 doc/FunctionsTakingEigenTypes.dox | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/FunctionsTakingEigenTypes.dox b/doc/FunctionsTakingEigenTypes.dox
index e054714f9..6b4e49214 100644
--- a/doc/FunctionsTakingEigenTypes.dox
+++ b/doc/FunctionsTakingEigenTypes.dox
@@ -79,7 +79,7 @@ These examples are just intended to give the reader a first impression of how fu
 
 \section TopicUsingRefClass How to write generic, but non-templated function?
 
-In all the previous examples, the functions had to be template functions. This approach allows to write very generic code, but it is often desirable to write non templated function and still keep some level of genericity to avoid stupid copies of the arguments. The typical example is to write functions accepting both a MatrixXf or a block of a MatrixXf. This exactly the purpose of the Ref class. Here is a simple example:
+In all the previous examples, the functions had to be template functions. This approach allows to write very generic code, but it is often desirable to write non templated functions and still keep some level of genericity to avoid stupid copies of the arguments. The typical example is to write functions accepting both a MatrixXf or a block of a MatrixXf. This is exactly the purpose of the Ref class. Here is a simple example:
 
 <table class="example">
 <tr><th>Example:</th><th>Output:</th></tr>

From 7b91c11207ee94e6a113c2a2c36f4a0797fc1c3e Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Tue, 24 Jul 2018 18:36:44 -0700
Subject: [PATCH 502/680] bug #1578: Improve prefetching in matrix
 multiplication on MIPS.

---
 Eigen/src/Core/products/GeneralBlockPanelKernel.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 1890efd4d..3ec8eb082 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -972,7 +972,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
               EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \
               EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
               internal::prefetch(blA+(3*K+16)*LhsProgress); \
-              if (EIGEN_ARCH_ARM) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \
+              if (EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \
               traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
               traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
               traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \

From 6913221c43c6ad41b1fbfc0d263d2764abd11ad2 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 25 Jul 2018 13:51:10 -0700
Subject: [PATCH 503/680] Add tiled evaluation support to TensorExecutor

---
 unsupported/Eigen/CXX11/Tensor                |   2 +-
 .../Eigen/CXX11/src/Tensor/TensorAssign.h     |  44 ++-
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 113 +++++++-
 .../CXX11/src/Tensor/TensorBroadcasting.h     |   2 +
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   |   2 +
 .../CXX11/src/Tensor/TensorConcatenation.h    |   2 +
 .../CXX11/src/Tensor/TensorContraction.h      |   1 +
 .../Eigen/CXX11/src/Tensor/TensorConversion.h |   1 +
 .../CXX11/src/Tensor/TensorConvolution.h      |  10 +-
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h |  16 ++
 .../Eigen/CXX11/src/Tensor/TensorEvalTo.h     |   1 +
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  | 101 ++++++-
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   | 266 +++++++++++++-----
 .../Eigen/CXX11/src/Tensor/TensorFixedSize.h  |   2 +
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |   1 +
 .../src/Tensor/TensorForwardDeclarations.h    |   8 +-
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |   1 +
 .../Eigen/CXX11/src/Tensor/TensorLayoutSwap.h |   2 +
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |   8 +-
 .../Eigen/CXX11/src/Tensor/TensorPadding.h    |   1 +
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |   1 +
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorRef.h        |   3 +
 .../Eigen/CXX11/src/Tensor/TensorReverse.h    |   2 +
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  |   2 +
 .../Eigen/CXX11/src/Tensor/TensorStriding.h   |   2 +
 .../Eigen/CXX11/src/Tensor/TensorTrace.h      |   7 +-
 unsupported/test/CMakeLists.txt               |   1 +
 .../test/cxx11_tensor_block_access.cpp        |   2 +-
 .../cxx11_tensor_complex_cwise_ops_gpu.cu     |   2 +-
 unsupported/test/cxx11_tensor_complex_gpu.cu  |   2 +-
 unsupported/test/cxx11_tensor_executor.cpp    |  81 ++++++
 33 files changed, 598 insertions(+), 93 deletions(-)
 create mode 100644 unsupported/test/cxx11_tensor_executor.cpp

diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 397d55f76..47514703a 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -112,13 +112,13 @@ typedef unsigned __int64 uint64_t;
 #include "src/Tensor/TensorGlobalFunctions.h"
 
 #include "src/Tensor/TensorBase.h"
+#include "src/Tensor/TensorBlock.h"
 
 #include "src/Tensor/TensorEvaluator.h"
 #include "src/Tensor/TensorExpr.h"
 #include "src/Tensor/TensorReduction.h"
 #include "src/Tensor/TensorReductionGpu.h"
 #include "src/Tensor/TensorArgMax.h"
-#include "src/Tensor/TensorBlock.h"
 #include "src/Tensor/TensorConcatenation.h"
 #include "src/Tensor/TensorContractionMapper.h"
 #include "src/Tensor/TensorContractionBlocking.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
index 027305586..199ddb123 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
@@ -68,6 +68,8 @@ class TensorAssignOp : public TensorBase<TensorAssignOp<LhsXprType, RhsXprType>
   typedef typename Eigen::internal::traits<TensorAssignOp>::StorageKind StorageKind;
   typedef typename Eigen::internal::traits<TensorAssignOp>::Index Index;
 
+  static const int NumDims = Eigen::internal::traits<TensorAssignOp>::NumDimensions;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorAssignOp(LhsXprType& lhs, const RhsXprType& rhs)
       : m_lhs_xpr(lhs), m_rhs_xpr(rhs) {}
 
@@ -95,20 +97,33 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef typename TensorEvaluator<RightArgType, Device>::Dimensions Dimensions;
+
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int NumDims = XprType::NumDims;
 
   enum {
-    IsAligned = TensorEvaluator<LeftArgType, Device>::IsAligned & TensorEvaluator<RightArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
-    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
-    RawAccess = TensorEvaluator<LeftArgType, Device>::RawAccess
+    IsAligned    = TensorEvaluator<LeftArgType, Device>::IsAligned &
+                   TensorEvaluator<RightArgType, Device>::IsAligned,
+    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess &
+                   TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess  = TensorEvaluator<LeftArgType, Device>::BlockAccess &
+                   TensorEvaluator<RightArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<LeftArgType, Device>::Layout,
+    RawAccess    = TensorEvaluator<LeftArgType, Device>::RawAccess
   };
 
+  typedef typename internal::TensorBlock<
+      typename internal::remove_const<Scalar>::type, Index, NumDims, Layout>
+      TensorBlock;
+
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) :
       m_leftImpl(op.lhsExpression(), device),
       m_rightImpl(op.rhsExpression(), device)
   {
-    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT(
+        (static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
+         static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),
+        YOU_MADE_A_PROGRAMMING_MISTAKE);
   }
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
@@ -164,6 +179,25 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
            TensorOpCost(0, sizeof(CoeffReturnType), 0, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    m_leftImpl.getResourceRequirements(resources);
+    m_rightImpl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalBlock(TensorBlock* block) {
+    if (TensorEvaluator<LeftArgType, Device>::RawAccess &&
+        m_leftImpl.data() != nullptr) {
+      TensorBlock left_block(block->first_coeff_index(), block->block_sizes(),
+                             block->tensor_strides(), block->tensor_strides(),
+                             m_leftImpl.data() + block->first_coeff_index());
+      m_rightImpl.block(&left_block);
+    } else {
+      m_rightImpl.block(block);
+      m_leftImpl.writeBlock(*block);
+    }
+  }
+
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<LeftArgType, Device>& left_impl() const { return m_leftImpl; }
   /// required by sycl in order to extract the accessor
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 8ffc9d093..5321acecf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -65,6 +65,40 @@ enum class TensorBlockShapeType {
   kSkewedInnerDims,
 };
 
+struct TensorOpResourceRequirements {
+  TensorBlockShapeType block_shape;
+  std::size_t block_total_size;
+  // TODO(andydavis) Add 'target_num_threads' to support communication of
+  // thread-resource requirements. This will allow ops deep in the
+  // expression tree (like reductions) to communicate resources
+  // requirements based on local state (like the total number of reductions
+  // to be computed).
+  TensorOpResourceRequirements(internal::TensorBlockShapeType shape,
+                               const std::size_t size)
+      : block_shape(shape), block_total_size(size) {}
+};
+
+// Tries to merge multiple resource requirements.
+EIGEN_STRONG_INLINE void MergeResourceRequirements(
+    const std::vector<TensorOpResourceRequirements>& resources,
+    TensorBlockShapeType* block_shape, std::size_t* block_total_size) {
+  if (resources.empty()) {
+    return;
+  }
+  // TODO(andydavis) Implement different policies (i.e. revert to a default
+  // policy if block shapes/sizes conflict).
+  *block_shape = resources[0].block_shape;
+  *block_total_size = resources[0].block_total_size;
+  for (int i = 1; i < resources.size(); ++i) {
+    if (resources[i].block_shape == TensorBlockShapeType::kSkewedInnerDims &&
+        *block_shape != TensorBlockShapeType::kSkewedInnerDims) {
+      *block_shape = TensorBlockShapeType::kSkewedInnerDims;
+    }
+    *block_total_size =
+        numext::maxi(*block_total_size, resources[i].block_total_size);
+  }
+}
+
 /**
  * \class TensorBlock
  * \ingroup CXX11_Tensor_Module
@@ -74,7 +108,7 @@ enum class TensorBlockShapeType {
  * This class represents a tensor block specified by the index of the
  * first block coefficient, and the size of the block in each dimension.
  */
-template <typename Scalar, typename Index, std::size_t NumDims, int Layout>
+template <typename Scalar, typename Index, int NumDims, int Layout>
 class TensorBlock {
  public:
   typedef DSizes<Index, NumDims> Dimensions;
@@ -614,6 +648,83 @@ struct TensorBlockCwiseBinaryIO {
   }
 };
 
+/**
+ * \class TensorBlockView
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Read-only view into a block of data.
+ *
+ * This class provides read-only access to a block of data in impl. It may need
+ * to allocate space for holding the intermediate result.
+ *
+ */
+template <class ArgType, class Device>
+struct TensorBlockView {
+  typedef TensorEvaluator<ArgType, Device> Impl;
+  typedef typename Impl::Index Index;
+  typedef typename remove_const<typename Impl::Scalar>::type Scalar;
+  static const int NumDims = array_size<typename Impl::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  // Constructs a TensorBlockView for `impl`. `block` is only used for for
+  // specifying the start offset, shape, and strides of the block.
+  template <typename OtherTensorBlock>
+  TensorBlockView(const Device& device,
+                  const TensorEvaluator<ArgType, Device>& impl,
+                  const OtherTensorBlock& block)
+      : m_device(device),
+        m_block_sizes(block.block_sizes()),
+        m_data(NULL),
+        m_allocated_data(NULL) {
+    if (Impl::RawAccess && impl.data() != NULL) {
+      m_data = impl.data() + block.first_coeff_index();
+      m_block_strides = block.tensor_strides();
+    } else {
+      // Actually make a copy.
+
+      // TODO(wuke): This sometimes put a lot pressure on the heap allocator.
+      // Consider allowing ops to request additional temporary block memory in
+      // TensorOpResourceRequirements.
+      m_allocated_data = static_cast<Scalar*>(
+          m_device.allocate(m_block_sizes.TotalSize() * sizeof(Scalar)));
+      m_data = m_allocated_data;
+      if (NumDims > 0) {
+        if (static_cast<int>(Impl::Layout) == static_cast<int>(ColMajor)) {
+          m_block_strides[0] = 1;
+          for (int i = 1; i < NumDims; ++i) {
+            m_block_strides[i] = m_block_strides[i - 1] * m_block_sizes[i - 1];
+          }
+        } else {
+          m_block_strides[NumDims - 1] = 1;
+          for (int i = NumDims - 2; i >= 0; --i) {
+            m_block_strides[i] = m_block_strides[i + 1] * m_block_sizes[i + 1];
+          }
+        }
+      }
+      TensorBlock<Scalar, Index, NumDims, Impl::Layout> input_block(
+          block.first_coeff_index(), m_block_sizes, m_block_strides,
+          block.tensor_strides(), m_allocated_data);
+      impl.block(&input_block);
+    }
+  }
+
+  ~TensorBlockView() {
+    if (m_allocated_data != NULL) {
+      m_device.deallocate(m_allocated_data);
+    }
+  }
+
+  const Dimensions& block_sizes() const { return m_block_sizes; }
+  const Dimensions& block_strides() const { return m_block_strides; }
+  const Scalar* data() const { return m_data; }
+
+ private:
+  const Device& m_device;
+  Dimensions m_block_sizes, m_block_strides;
+  const Scalar* m_data;      // Not owned.
+  Scalar* m_allocated_data;  // Owned.
+};
+
 /**
  * \class TensorBlockMapper
  * \ingroup CXX11_Tensor_Module
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 278689915..7ff0d323b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -1,4 +1,5 @@
 // This file is part of Eigen, a lightweight C++ template library
+// This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
 // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
@@ -110,6 +111,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   enum {
     IsAligned = true,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 21ffa2872..085c05f3d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -146,6 +146,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     // slice offsets.
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -343,6 +344,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     RawAccess = false
   };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
index a7c1380b8..9f0321880 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@@ -122,6 +122,7 @@ struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgTy
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     RawAccess = false
   };
@@ -306,6 +307,7 @@ template<typename Axis, typename LeftArgType, typename RightArgType, typename De
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 0e69cd40c..12cfa8df3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -240,6 +240,7 @@ struct TensorContractionEvaluatorBase
   enum {
     IsAligned = true,
     PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    BlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
index 182bef918..e0cbbb315 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
@@ -195,6 +195,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = true,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 25131600d..1ec5819a7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -307,6 +307,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   enum {
     IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<InputArgType, Device>::PacketAccess & TensorEvaluator<KernelArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<InputArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -577,11 +578,11 @@ __global__ void EigenConvolutionKernel1D(
     const float* __restrict kernel, const int numPlanes, const int numX,
     const int maxX, const int kernelSize, float* buffer) {
 #if defined(EIGEN_HIPCC)
-  HIP_DYNAMIC_SHARED(float, s)	    
+  HIP_DYNAMIC_SHARED(float, s)
 #else
   extern __shared__ float s[];
 #endif
-  
+
   const int first_x = blockIdx.x * maxX;
   const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
   const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSize>()(kernelSize);
@@ -630,7 +631,7 @@ __global__ void EigenConvolutionKernel2D(
     const int maxX, const int numY, const int maxY, const int kernelSizeX,
     const int kernelSizeY, float* buffer) {
 #if defined(EIGEN_HIPCC)
-  HIP_DYNAMIC_SHARED(float, s)	    
+  HIP_DYNAMIC_SHARED(float, s)
 #else
   extern __shared__ float s[];
 #endif
@@ -702,7 +703,7 @@ __global__ void EigenConvolutionKernel3D(
     const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY,
     const size_t kernelSizeZ, float* buffer) {
 #if defined(EIGEN_HIPCC)
-  HIP_DYNAMIC_SHARED(float, s)	    
+  HIP_DYNAMIC_SHARED(float, s)
 #else
   extern __shared__ float s[];
 #endif
@@ -778,6 +779,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   enum {
     IsAligned = TensorEvaluator<InputArgType, GpuDevice>::IsAligned & TensorEvaluator<KernelArgType, GpuDevice>::IsAligned,
     PacketAccess = false,
+    BlockAccess = false,
     Layout = TensorEvaluator<InputArgType, GpuDevice>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index 65403905a..d301d0c01 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -242,6 +242,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   enum {
     IsAligned = TensorEvaluator<InputArgType, const Eigen::SyclDevice>::IsAligned & TensorEvaluator<KernelArgType, const Eigen::SyclDevice>::IsAligned,
     PacketAccess = false,
+    BlockAccess = false,
     Layout = TensorEvaluator<InputArgType, const Eigen::SyclDevice>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 192d4aa7b..5ca47cca7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -290,6 +290,22 @@ struct DSizes : array<DenseIndex, NumDims> {
     }
   }
 
+#ifndef EIGEN_EMULATE_CXX11_META_H
+  template <typename std::ptrdiff_t... Indices>
+  EIGEN_DEVICE_FUNC DSizes(const Sizes<Indices...>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+#else
+  template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5>
+  EIGEN_DEVICE_FUNC DSizes(const Sizes<V1, V2, V3, V4, V5>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+#endif
+
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   template<typename... IndexTypes> EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE explicit DSizes(DenseIndex firstDimension, DenseIndex secondDimension, IndexTypes... otherDimensions) : Base({{firstDimension, secondDimension, otherDimensions...}}) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
index d0c027890..af39daa91 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -107,6 +107,7 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType, MakePointer_>, Device>
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index fe62ff1ea..ba02802d2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -41,11 +41,24 @@ struct TensorEvaluator
   enum {
     IsAligned = Derived::IsAligned,
     PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
     RawAccess = true
   };
 
+  typedef typename internal::TensorBlock<
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
+      TensorBlock;
+  typedef typename internal::TensorBlockReader<
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
+      PacketAccess>
+      TensorBlockReader;
+  typedef typename internal::TensorBlockWriter<
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
+      PacketAccess>
+      TensorBlockWriter;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device& device)
       : m_data(const_cast<typename internal::traits<Derived>::template MakePointer<Scalar>::Type>(m.data())), m_dims(m.dimensions()), m_device(device), m_impl(m)
   { }
@@ -113,6 +126,20 @@ struct TensorEvaluator
                         internal::unpacket_traits<PacketReturnType>::size);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(TensorBlock* block) const {
+    assert(m_data != NULL);
+    TensorBlockReader::Run(block, m_data);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlock& block) {
+    assert(m_data != NULL);
+    TensorBlockWriter::Run(block, m_data);
+  }
+
   EIGEN_DEVICE_FUNC typename internal::traits<Derived>::template MakePointer<Scalar>::Type data() const { return m_data; }
 
   /// required by sycl in order to construct sycl buffer from raw pointer
@@ -167,11 +194,20 @@ struct TensorEvaluator<const Derived, Device>
   enum {
     IsAligned = Derived::IsAligned,
     PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
     RawAccess = true
   };
 
+  typedef typename internal::TensorBlock<
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
+      TensorBlock;
+  typedef typename internal::TensorBlockReader<
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
+      PacketAccess>
+      TensorBlockReader;
+
   // Used for accessor extraction in SYCL Managed TensorMap:
   const Derived& derived() const { return m_impl; }
 
@@ -219,6 +255,14 @@ struct TensorEvaluator<const Derived, Device>
                         internal::unpacket_traits<PacketReturnType>::size);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(TensorBlock* block) const {
+    assert(m_data != NULL);
+    TensorBlockReader::Run(block, m_data);
+  }
+
   EIGEN_DEVICE_FUNC typename internal::traits<Derived>::template MakePointer<const Scalar>::Type data() const { return m_data; }
 
   /// added for sycl in order to construct the buffer from the sycl device
@@ -244,6 +288,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
   enum {
     IsAligned = true,
     PacketAccess = internal::functor_traits<NullaryOp>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -308,7 +353,9 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
 
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess & internal::functor_traits<UnaryOp>::PacketAccess,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess &
+                   internal::functor_traits<UnaryOp>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -375,16 +422,21 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   typedef TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType> XprType;
 
   enum {
-    IsAligned = TensorEvaluator<LeftArgType, Device>::IsAligned & TensorEvaluator<RightArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess &
+    IsAligned    = TensorEvaluator<LeftArgType, Device>::IsAligned &
+                   TensorEvaluator<RightArgType, Device>::IsAligned,
+    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess &
+                   TensorEvaluator<RightArgType, Device>::PacketAccess &
                    internal::functor_traits<BinaryOp>::PacketAccess,
-    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<LeftArgType, Device>::BlockAccess &
+                   TensorEvaluator<RightArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<LeftArgType, Device>::Layout,
+    CoordAccess  = false,  // to be implemented
+    RawAccess    = false
   };
 
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
-    : m_functor(op.functor()),
+    : m_device(device),
+      m_functor(op.functor()),
       m_leftImpl(op.lhsExpression(), device),
       m_rightImpl(op.rhsExpression(), device)
   {
@@ -399,6 +451,14 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
   typedef typename TensorEvaluator<LeftArgType, Device>::Dimensions Dimensions;
 
+  static const int NumDims = internal::array_size<
+      typename TensorEvaluator<LeftArgType, Device>::Dimensions>::value;
+
+  typedef internal::TensorBlock<
+      typename internal::remove_const<Scalar>::type, Index, NumDims,
+      TensorEvaluator<LeftArgType, Device>::Layout>
+      TensorBlock;
+
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
   {
     // TODO: use right impl instead if right impl dimensions are known at compile time.
@@ -433,6 +493,30 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
            TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    m_leftImpl.getResourceRequirements(resources);
+    m_rightImpl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      TensorBlock* output_block) const {
+    if (NumDims <= 0) {
+      output_block->data()[0] = coeff(0);
+      return;
+    }
+    internal::TensorBlockView<LeftArgType, Device> left_block(
+        m_device, m_leftImpl, *output_block);
+    internal::TensorBlockView<RightArgType, Device> right_block(
+        m_device, m_rightImpl, *output_block);
+    internal::TensorBlockCwiseBinaryIO<
+        BinaryOp, Index, typename internal::remove_const<Scalar>::type, NumDims,
+        Layout>::Run(m_functor, output_block->block_sizes(),
+                     output_block->block_strides(), output_block->data(),
+                     left_block.block_strides(), left_block.data(),
+                     right_block.block_strides(), right_block.data());
+  }
+
   EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<LeftArgType, Device>& left_impl() const { return m_leftImpl; }
@@ -442,6 +526,7 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   BinaryOp functor() const { return m_functor; }
 
  private:
+  const Device& m_device;
   const BinaryOp m_functor;
   TensorEvaluator<LeftArgType, Device> m_leftImpl;
   TensorEvaluator<RightArgType, Device> m_rightImpl;
@@ -458,6 +543,7 @@ struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type,
     IsAligned = TensorEvaluator<Arg1Type, Device>::IsAligned & TensorEvaluator<Arg2Type, Device>::IsAligned & TensorEvaluator<Arg3Type, Device>::IsAligned,
     PacketAccess = TensorEvaluator<Arg1Type, Device>::PacketAccess & TensorEvaluator<Arg2Type, Device>::PacketAccess & TensorEvaluator<Arg3Type, Device>::PacketAccess &
                    internal::functor_traits<TernaryOp>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<Arg1Type, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -562,6 +648,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>
     IsAligned = TensorEvaluator<ThenArgType, Device>::IsAligned & TensorEvaluator<ElseArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ThenArgType, Device>::PacketAccess & TensorEvaluator<ElseArgType, Device>::PacketAccess &
                    internal::packet_traits<Scalar>::HasBlend,
+    BlockAccess = false,
     Layout = TensorEvaluator<IfArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 53640c6aa..024de3696 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -12,29 +12,37 @@
 
 namespace Eigen {
 
-/** \class TensorExecutor
-  * \ingroup CXX11_Tensor_Module
-  *
-  * \brief The tensor executor class.
-  *
-  * This class is responsible for launch the evaluation of the expression on
-  * the specified computing device.
-  */
+/**
+ * \class TensorExecutor
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief The tensor executor class.
+ *
+ * This class is responsible for launch the evaluation of the expression on
+ * the specified computing device.
+ *
+ * @tparam Vectorizable can use packet math (SSE/AVX/etc... registers and
+ *                      instructions)
+ * @tparam Tileable     can use block based tensor evaluation
+ *                      (see TensorBlock.h)
+ */
 namespace internal {
 
-// Default strategy: the expression is evaluated with a single cpu thread.
-template<typename Expression, typename Device, bool Vectorizable>
-class TensorExecutor
-{
+/**
+ * Default strategy: the expression is evaluated sequentially with a single cpu
+ * thread, without vectorization and block evaluation.
+ */
+template <typename Expression, typename Device, bool Vectorizable,
+          bool Tileable>
+class TensorExecutor {
  public:
   typedef typename Expression::Index Index;
   EIGEN_DEVICE_FUNC
-  static inline void run(const Expression& expr, const Device& device = Device())
-  {
+  static inline void run(const Expression& expr,
+                         const Device& device = Device()) {
     TensorEvaluator<Expression, Device> evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
-    if (needs_assign)
-    {
+    if (needs_assign) {
       const Index size = array_prod(evaluator.dimensions());
       for (Index i = 0; i < size; ++i) {
         evaluator.evalScalar(i);
@@ -44,12 +52,14 @@ class TensorExecutor
   }
 };
 
-
-template<typename Expression>
-class TensorExecutor<Expression, DefaultDevice, true>
-{
+/**
+ * Process all the data with a single cpu thread, using vectorized instructions.
+ */
+template <typename Expression>
+class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true, /*Tilable*/ false> {
  public:
   typedef typename Expression::Index Index;
+
   EIGEN_DEVICE_FUNC
   static inline void run(const Expression& expr, const DefaultDevice& device = DefaultDevice())
   {
@@ -58,9 +68,11 @@ class TensorExecutor<Expression, DefaultDevice, true>
     if (needs_assign)
     {
       const Index size = array_prod(evaluator.dimensions());
-      const int PacketSize = unpacket_traits<typename TensorEvaluator<Expression, DefaultDevice>::PacketReturnType>::size;
-      // Give the compiler a strong hint to unroll the loop. But don't insist
-      // on unrolling, because if the function is expensive the compiler should not
+      const int PacketSize = unpacket_traits<typename TensorEvaluator<
+          Expression, DefaultDevice>::PacketReturnType>::size;
+
+      // Give compiler a strong possibility to unroll the loop. But don't insist
+      // on unrolling, because if the function is expensive compiler should not
       // unroll the loop at the expense of inlining.
       const Index UnrolledSize = (size / (4 * PacketSize)) * 4 * PacketSize;
       for (Index i = 0; i < UnrolledSize; i += 4*PacketSize) {
@@ -80,9 +92,75 @@ class TensorExecutor<Expression, DefaultDevice, true>
   }
 };
 
+/**
+ * Process all the data with a single cpu thread, using blocks of data. By
+ * sizing a block to fit L1 cache we get better cache performance.
+ */
+template <typename Expression, bool Vectorizable>
+class TensorExecutor<Expression, DefaultDevice, Vectorizable, /*Tilable*/ true> {
+ public:
+  typedef typename Expression::Index Index;
 
+  EIGEN_DEVICE_FUNC
+  static inline void run(const Expression& expr,
+                         const DefaultDevice& device = DefaultDevice()) {
+    using Evaluator = TensorEvaluator<Expression, DefaultDevice>;
 
-// Multicore strategy: the index space is partitioned and each partition is executed on a single core
+    using Index = typename traits<Expression>::Index;
+    const int NumDims = traits<Expression>::NumDimensions;
+
+    using Scalar = typename traits<Expression>::Scalar;
+    using ScalarNoConst = typename remove_const<Scalar>::type;
+
+    using TensorBlock =
+        TensorBlock<ScalarNoConst, Index, NumDims, Evaluator::Layout>;
+    using TensorBlockMapper =
+        TensorBlockMapper<ScalarNoConst, Index, NumDims, Evaluator::Layout>;
+
+    Evaluator evaluator(expr, device);
+    std::size_t total_size = array_prod(evaluator.dimensions());
+    std::size_t cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
+
+    if (total_size < cache_size) {
+      // TODO(andydavis) Reduce block management overhead for small tensors.
+      // TODO(wuke) Do not do this when evaluating TensorBroadcastingOp.
+      internal::TensorExecutor<Expression, DefaultDevice, Vectorizable,
+                               false>::run(expr, device);
+      return;
+    }
+
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
+    if (needs_assign) {
+      // Size tensor blocks to fit in cache (or requested target block size).
+      size_t block_total_size = numext::mini(cache_size, total_size);
+      TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
+      // Query expression tree for desired block size/shape.
+      std::vector<TensorOpResourceRequirements> resources;
+      evaluator.getResourceRequirements(&resources);
+      MergeResourceRequirements(resources, &block_shape, &block_total_size);
+
+      TensorBlockMapper block_mapper(evaluator.dimensions(), block_shape,
+                                     block_total_size);
+      block_total_size = block_mapper.block_dims_total_size();
+
+      Scalar* data = static_cast<Scalar*>(
+          device.allocate(block_total_size * sizeof(Scalar)));
+
+      const Index total_block_count = block_mapper.total_block_count();
+      for (Index i = 0; i < total_block_count; ++i) {
+        TensorBlock block = block_mapper.GetBlockForIndex(i, data);
+        evaluator.evalBlock(&block);
+      }
+      device.deallocate(data);
+    }
+    evaluator.cleanup();
+  }
+};
+
+/**
+ * Multicore strategy: the index space is partitioned and each partition is
+ * executed on a single core.
+ */
 #ifdef EIGEN_USE_THREADS
 template <typename Evaluator, typename Index, bool Vectorizable>
 struct EvalRange {
@@ -100,7 +178,7 @@ struct EvalRange {
 };
 
 template <typename Evaluator, typename Index>
-struct EvalRange<Evaluator, Index, true> {
+struct EvalRange<Evaluator, Index, /*Vectorizable*/ true> {
   static const int PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
 
   static void run(Evaluator* evaluator_in, const Index first, const Index last) {
@@ -110,8 +188,8 @@ struct EvalRange<Evaluator, Index, true> {
     if (last - first >= PacketSize) {
       eigen_assert(first % PacketSize == 0);
       Index last_chunk_offset = last - 4 * PacketSize;
-      // Give the compiler a strong hint to unroll the loop. But don't insist
-      // on unrolling, because if the function is expensive the compiler should not
+      // Give compiler a strong possibility to unroll the loop. But don't insist
+      // on unrolling, because if the function is expensive compiler should not
       // unroll the loop at the expense of inlining.
       for (; i <= last_chunk_offset; i += 4*PacketSize) {
         for (Index j = 0; j < 4; j++) {
@@ -138,55 +216,113 @@ struct EvalRange<Evaluator, Index, true> {
   }
 };
 
-template <typename Expression, bool Vectorizable>
-class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> {
+template <typename Expression, bool Vectorizable, bool Tileable>
+class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
  public:
   typedef typename Expression::Index Index;
-  static inline void run(const Expression& expr, const ThreadPoolDevice& device)
-  {
-    typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
-    Evaluator evaluator(expr, device);
-    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
-    if (needs_assign)
-    {
-      const Index size = array_prod(evaluator.dimensions());
-      size_t num_threads = device.numThreads();
-      if (num_threads > 1) {
-        num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
-            size, evaluator.costPerCoeff(Vectorizable), num_threads);
-      }
-      if (num_threads == 1) {
-        EvalRange<Evaluator, Index, Vectorizable>::run(&evaluator, 0, size);
-      } else {
-        const Index PacketSize = Vectorizable ? unpacket_traits<typename Evaluator::PacketReturnType>::size : 1;
-        Index blocksz = std::ceil<Index>(static_cast<float>(size)/num_threads) + PacketSize - 1;
-        const Index blocksize = numext::maxi<Index>(PacketSize, (blocksz - (blocksz % PacketSize)));
-        const Index numblocks = size / blocksize;
 
-        Barrier barrier(numblocks);
-        for (int i = 0; i < numblocks; ++i) {
-          device.enqueue_with_barrier(
-              &barrier, &EvalRange<Evaluator, Index, Vectorizable>::run,
-              &evaluator, i * blocksize, (i + 1) * blocksize);
-        }
-        if (numblocks * blocksize < size) {
-          EvalRange<Evaluator, Index, Vectorizable>::run(
-              &evaluator, numblocks * blocksize, size);
-        }
-        barrier.Wait();
-      }
+  static inline void run(const Expression& expr,
+                         const ThreadPoolDevice& device) {
+    typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+    typedef EvalRange<Evaluator, Index, Vectorizable> EvalRange;
+
+    Evaluator evaluator(expr, device);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    if (needs_assign) {
+      const Index PacketSize =
+          Vectorizable
+              ? unpacket_traits<typename Evaluator::PacketReturnType>::size
+              : 1;
+      const Index size = array_prod(evaluator.dimensions());
+      device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
+                         EvalRange::alignBlockSize,
+                         [&evaluator](Index first, Index last) {
+                           EvalRange::run(&evaluator, first, last);
+                         });
     }
     evaluator.cleanup();
   }
 };
+
+template <typename Expression, bool Vectorizable>
+class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ true> {
+ public:
+  typedef typename Expression::Index Index;
+
+  static inline void run(const Expression& expr,
+                         const ThreadPoolDevice& device) {
+    typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+    typedef typename internal::remove_const<
+        typename traits<Expression>::Scalar>::type Scalar;
+    typedef typename traits<Expression>::Index Index;
+
+    static const int NumDims = traits<Expression>::NumDimensions;
+
+    typedef TensorBlock<Scalar, Index, NumDims, Evaluator::Layout> TensorBlock;
+    typedef TensorBlockMapper<Scalar, Index, NumDims, Evaluator::Layout>
+        TensorBlockMapper;
+
+    Evaluator evaluator(expr, device);
+    std::size_t total_size = array_prod(evaluator.dimensions());
+    std::size_t cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
+    if (total_size < cache_size) {
+      // TODO(andydavis) Reduce block management overhead for small tensors.
+      internal::TensorExecutor<Expression, ThreadPoolDevice, Vectorizable,
+                               false>::run(expr, device);
+      evaluator.cleanup();
+      return;
+    }
+
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    if (needs_assign) {
+      TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
+      size_t block_total_size = 0;
+      // Query expression tree for desired block size/shape.
+      std::vector<internal::TensorOpResourceRequirements> resources;
+      evaluator.getResourceRequirements(&resources);
+      MergeResourceRequirements(resources, &block_shape, &block_total_size);
+      int num_threads = device.numThreads();
+
+      // Estimate minimum block size based on cost.
+      TensorOpCost cost = evaluator.costPerCoeff(Vectorizable);
+      double taskSize = TensorCostModel<ThreadPoolDevice>::taskSize(1, cost);
+      size_t block_size = static_cast<size_t>(1.0 / taskSize);
+      TensorBlockMapper block_mapper(evaluator.dimensions(), block_shape,
+                                     block_size);
+      block_size = block_mapper.block_dims_total_size();
+      const size_t aligned_blocksize =
+          EIGEN_MAX_ALIGN_BYTES *
+          divup<size_t>(block_size * sizeof(Scalar), EIGEN_MAX_ALIGN_BYTES);
+      void* buf = device.allocate((num_threads + 1) * aligned_blocksize);
+      device.parallelFor(
+          block_mapper.total_block_count(), cost * block_size,
+          [=, &device, &evaluator, &block_mapper](Index first, Index last) {
+            // currentThreadId() returns -1 if called from a thread not in the
+            // threadpool, such as the main thread dispatching Eigen
+            // expressions.
+            const int thread_idx = device.currentThreadId();
+            eigen_assert(thread_idx >= -1 && thread_idx < num_threads);
+            Scalar* thread_buf = reinterpret_cast<Scalar*>(
+                static_cast<char*>(buf) + aligned_blocksize * (thread_idx + 1));
+            for (Index i = first; i < last; ++i) {
+              auto block = block_mapper.GetBlockForIndex(i, thread_buf);
+              evaluator.evalBlock(&block);
+            }
+          });
+      device.deallocate(buf);
+    }
+    evaluator.cleanup();
+  }
+};
+
 #endif  // EIGEN_USE_THREADS
 
 
 // GPU: the evaluation of the expression is offloaded to a GPU.
 #if defined(EIGEN_USE_GPU)
 
-template <typename Expression, bool Vectorizable>
-class TensorExecutor<Expression, GpuDevice, Vectorizable> {
+template <typename Expression, bool Vectorizable, bool Tileable>
+class TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable> {
  public:
   typedef typename Expression::Index Index;
   static void run(const Expression& expr, const GpuDevice& device);
@@ -236,8 +372,8 @@ EigenMetaKernel(Evaluator eval, Index size) {
 }
 
 /*static*/
-template <typename Expression, bool Vectorizable>
-inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
+template <typename Expression, bool Vectorizable, bool Tileable>
+inline void TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable>::run(
     const Expression& expr, const GpuDevice& device) {
   TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
   const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
index e943757ad..1342e47a6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
@@ -40,6 +40,8 @@ class TensorFixedSize : public TensorBase<TensorFixedSize<Scalar_, Dimensions_,
 
     enum {
       IsAligned = bool(EIGEN_MAX_ALIGN_BYTES>0),
+      PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+      BlockAccess = false,
       Layout = Options_ & RowMajor ? RowMajor : ColMajor,
       CoordAccess = true,
       RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index b8f0bc798..fdb31928f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -98,6 +98,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
   enum {
     IsAligned = true,
     PacketAccess = (PacketSize > 1),
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = true
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 19e456e19..8ed1796df 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -129,8 +129,14 @@ struct IsVectorizable<GpuDevice, Expression> {
                             TensorEvaluator<Expression, GpuDevice>::IsAligned;
 };
 
+template <typename Device, typename Expression>
+struct IsTileable {
+  static const bool value = TensorEvaluator<Expression, Device>::BlockAccess;
+};
+
 template <typename Expression, typename Device,
-          bool Vectorizable = IsVectorizable<Device, Expression>::value>
+          bool Vectorizable = IsVectorizable<Device, Expression>::value,
+          bool Tileable = IsTileable<Device, Expression>::value>
 class TensorExecutor;
 
 }  // end namespace internal
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index f0f7c7826..72cb2d15f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -186,6 +186,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
index 4e384f9b9..e3165fa10 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
@@ -119,6 +119,7 @@ struct TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
     CoordAccess = false,  // to be implemented
     RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@@ -181,6 +182,7 @@ template<typename ArgType, typename Device>
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
     CoordAccess = false  // to be implemented
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index cda49f8fe..498488649 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -105,6 +105,7 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@@ -170,6 +171,7 @@ template<typename NewDimensions, typename ArgType, typename Device>
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@@ -325,6 +327,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
     // slice offsets and sizes.
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
@@ -557,6 +560,7 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
   enum {
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
@@ -716,7 +720,6 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
   static const int NumDims = internal::array_size<Strides>::value;
   typedef typename XprType::Index Index;
   typedef typename XprType::Scalar Scalar;
-  typedef typename internal::remove_const<Scalar>::type ScalarNonConst;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef Strides Dimensions;
@@ -858,7 +861,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
     }
     return inputIndex;
   }
-  
+
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index clamp(Index value, Index min, Index max) {
 #ifndef __SYCL_DEVICE_ONLY__
     return numext::maxi(min, numext::mini(max,value));
@@ -907,7 +910,6 @@ struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Stride
 
   typedef typename XprType::Index Index;
   typedef typename XprType::Scalar Scalar;
-  typedef typename internal::remove_const<Scalar>::type ScalarNonConst;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef Strides Dimensions;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
index 5956e513d..ffa22f31e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
@@ -96,6 +96,7 @@ struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device
   enum {
     IsAligned = true,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = true,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 9e0a20abf..950ac32af 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -94,6 +94,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index ce573d730..375fc0802 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -412,6 +412,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   enum {
     IsAligned = false,
     PacketAccess = Self::InputPacketAccess && Op::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
index b2b4fd8d3..a6cade50f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
@@ -136,6 +136,7 @@ template<typename PlainObjectType> class TensorRef : public TensorBase<TensorRef
     enum {
       IsAligned = false,
       PacketAccess = false,
+      BlockAccess = false,
       Layout = PlainObjectType::Layout,
       CoordAccess = false,  // to be implemented
       RawAccess = false
@@ -364,6 +365,7 @@ struct TensorEvaluator<const TensorRef<Derived>, Device>
   enum {
     IsAligned = false,
     PacketAccess = false,
+    BlockAccess = false,
     Layout = TensorRef<Derived>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -411,6 +413,7 @@ struct TensorEvaluator<TensorRef<Derived>, Device> : public TensorEvaluator<cons
   enum {
     IsAligned = false,
     PacketAccess = false,
+    BlockAccess = false,
     RawAccess = false
   };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
index 14a50a029..bb2768ab1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
@@ -113,6 +113,7 @@ struct TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>, Device
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -253,6 +254,7 @@ struct TensorEvaluator<TensorReverseOp<ReverseDimensions, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 0697fd1ce..6b54f40ad 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -112,6 +112,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -240,6 +241,7 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    BlockAccess = false,
     RawAccess = false
   };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
index a7eea99b6..c09513c10 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
@@ -112,6 +112,7 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
   enum {
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -273,6 +274,7 @@ struct TensorEvaluator<TensorStridingOp<Strides, ArgType>, Device>
   enum {
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
index 2b1968de1..c8b2fad1e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@@ -95,6 +95,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
   enum {
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
@@ -110,7 +111,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
     for (int i = 0; i < NumInputDims; ++i) {
       m_reduced[i] = false;
     }
-    
+
     const Dims& op_dims = op.dims();
     for (int i = 0; i < NumReducedDims; ++i) {
       eigen_assert(op_dims[i] >= 0);
@@ -128,7 +129,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
 
     eigen_assert(num_distinct_reduce_dims == NumReducedDims);
 
-    // Compute the dimensions of the result. 
+    // Compute the dimensions of the result.
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
 
     int output_index = 0;
@@ -229,7 +230,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
         result += m_impl.coeff(cur_index);
         cur_index += index_stride;
     }
-      
+
     return result;
   }
 
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index fa19b2159..239a80926 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -213,6 +213,7 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_striding)
   ei_add_test(cxx11_tensor_notification "-pthread" "${CMAKE_THREAD_LIBS_INIT}")
   ei_add_test(cxx11_tensor_thread_pool "-pthread" "${CMAKE_THREAD_LIBS_INIT}")
+  ei_add_test(cxx11_tensor_executor "-pthread" "${CMAKE_THREAD_LIBS_INIT}")
   ei_add_test(cxx11_tensor_ref)
   ei_add_test(cxx11_tensor_random)
   ei_add_test(cxx11_tensor_generator)
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 15f2392a3..416b686e4 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -901,7 +901,7 @@ static void test_empty_dims(const internal::TensorBlockShapeType block_shape)
   CALL_SUBTEST(NAME<ColMajor>(ARG)); \
   CALL_SUBTEST(NAME<RowMajor>(ARG))
 
-EIGEN_DECLARE_TEST(cxx11_tensor_assign) {
+EIGEN_DECLARE_TEST(cxx11_tensor_block_access) {
   CALL_SUBTEST_LAYOUTS(test_block_mapper_sanity);
   CALL_SUBTEST_LAYOUTS(test_block_mapper_maps_every_element);
   CALL_SUBTEST_LAYOUTS(test_slice_block_mapper_maps_every_element);
diff --git a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
index aa28457b1..f2a2a6cfa 100644
--- a/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_cwise_ops_gpu.cu
@@ -93,7 +93,7 @@ void test_cuda_complex_cwise_ops() {
 }
 
 
-void test_cxx11_tensor_complex_cwise_ops()
+EIGEN_DECLARE_TEST(test_cxx11_tensor_complex_cwise_ops)
 {
   CALL_SUBTEST(test_cuda_complex_cwise_ops<float>());
   CALL_SUBTEST(test_cuda_complex_cwise_ops<double>());
diff --git a/unsupported/test/cxx11_tensor_complex_gpu.cu b/unsupported/test/cxx11_tensor_complex_gpu.cu
index 7cf06aa7a..f8b8ae704 100644
--- a/unsupported/test/cxx11_tensor_complex_gpu.cu
+++ b/unsupported/test/cxx11_tensor_complex_gpu.cu
@@ -177,7 +177,7 @@ static void test_cuda_product_reductions() {
 }
 
 
-void test_cxx11_tensor_complex()
+EIGEN_DECLARE_TEST(test_cxx11_tensor_complex)
 {
   CALL_SUBTEST(test_cuda_nullary());
   CALL_SUBTEST(test_cuda_sum_reductions());
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
new file mode 100644
index 000000000..5ae45ac5b
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -0,0 +1,81 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Eugene Zhulenev <ezhulenev@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_USE_THREADS
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Index;
+using Eigen::Tensor;
+using Eigen::RowMajor;
+using Eigen::ColMajor;
+
+// A set of tests to verify that different TensorExecutor strategies yields the
+// same results for all the ops, supporting tiled execution.
+
+template <typename Device, bool Vectorizable, bool Tileable, int Layout>
+static void test_execute_binary_expr(Device d) {
+  // Pick a large enough tensor size to bypass small tensor block evaluation
+  // optimization.
+  Tensor<float, 3> lhs(840, 390, 37);
+  Tensor<float, 3> rhs(840, 390, 37);
+  Tensor<float, 3> dst(840, 390, 37);
+
+  lhs.setRandom();
+  rhs.setRandom();
+
+  const auto expr = lhs + rhs;
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (int i = 0; i < 840; ++i) {
+    for (int j = 0; j < 390; ++j) {
+      for (int k = 0; k < 37; ++k) {
+        float sum = lhs(i, j, k) + rhs(i, j, k);
+        VERIFY_IS_EQUAL(sum, dst(i, j, k));
+      }
+    }
+  }
+}
+
+#define CALL_SUBTEST_COMBINATIONS(NAME)                                        \
+  CALL_SUBTEST((NAME<DefaultDevice, false, false, ColMajor>(default_device))); \
+  CALL_SUBTEST((NAME<DefaultDevice, false, true, ColMajor>(default_device)));  \
+  CALL_SUBTEST((NAME<DefaultDevice, true, false, ColMajor>(default_device)));  \
+  CALL_SUBTEST((NAME<DefaultDevice, true, true, ColMajor>(default_device)));   \
+  CALL_SUBTEST((NAME<DefaultDevice, false, false, RowMajor>(default_device))); \
+  CALL_SUBTEST((NAME<DefaultDevice, false, true, RowMajor>(default_device)));  \
+  CALL_SUBTEST((NAME<DefaultDevice, true, false, RowMajor>(default_device)));  \
+  CALL_SUBTEST((NAME<DefaultDevice, true, true, RowMajor>(default_device)));   \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, false, false, ColMajor>(tp_device)));   \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, false, true, ColMajor>(tp_device)));    \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, true, false, ColMajor>(tp_device)));    \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, true, true, ColMajor>(tp_device)));     \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, false, false, RowMajor>(tp_device)));   \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, false, true, RowMajor>(tp_device)));    \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, true, false, RowMajor>(tp_device)));    \
+  CALL_SUBTEST((NAME<ThreadPoolDevice, true, true, RowMajor>(tp_device)))
+
+EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
+  Eigen::DefaultDevice default_device;
+
+  const auto num_threads = internal::random<int>(1, 24);
+  Eigen::ThreadPool tp(num_threads);
+  Eigen::ThreadPoolDevice tp_device(&tp, num_threads);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr);
+}
+
+#undef CALL_SUBTEST_COMBINATIONS

From 966c2a7bb62a8b5b9ecd349730ffcd3b5719837d Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 27 Jul 2018 12:45:17 -0700
Subject: [PATCH 504/680] Rename Index to StorageIndex + use Eigen::Array and
 Eigen::Map when possible

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 418 +++++++-----------
 .../CXX11/src/Tensor/TensorBroadcasting.h     |   1 -
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  |   9 +-
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   | 185 ++++----
 .../test/cxx11_tensor_block_access.cpp        | 292 +++++++-----
 unsupported/test/cxx11_tensor_executor.cpp    |  20 +-
 6 files changed, 447 insertions(+), 478 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 5321acecf..84cf6d216 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -67,21 +67,21 @@ enum class TensorBlockShapeType {
 
 struct TensorOpResourceRequirements {
   TensorBlockShapeType block_shape;
-  std::size_t block_total_size;
+  Index block_total_size;
   // TODO(andydavis) Add 'target_num_threads' to support communication of
   // thread-resource requirements. This will allow ops deep in the
   // expression tree (like reductions) to communicate resources
   // requirements based on local state (like the total number of reductions
   // to be computed).
   TensorOpResourceRequirements(internal::TensorBlockShapeType shape,
-                               const std::size_t size)
+                               const Index size)
       : block_shape(shape), block_total_size(size) {}
 };
 
 // Tries to merge multiple resource requirements.
 EIGEN_STRONG_INLINE void MergeResourceRequirements(
     const std::vector<TensorOpResourceRequirements>& resources,
-    TensorBlockShapeType* block_shape, std::size_t* block_total_size) {
+    TensorBlockShapeType* block_shape, Index* block_total_size) {
   if (resources.empty()) {
     return;
   }
@@ -108,12 +108,12 @@ EIGEN_STRONG_INLINE void MergeResourceRequirements(
  * This class represents a tensor block specified by the index of the
  * first block coefficient, and the size of the block in each dimension.
  */
-template <typename Scalar, typename Index, int NumDims, int Layout>
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlock {
  public:
-  typedef DSizes<Index, NumDims> Dimensions;
+  typedef DSizes<StorageIndex, NumDims> Dimensions;
 
-  TensorBlock(const Index first_coeff_index, const Dimensions& block_sizes,
+  TensorBlock(const StorageIndex first_coeff_index, const Dimensions& block_sizes,
               const Dimensions& block_strides, const Dimensions& tensor_strides,
               Scalar* data)
       : m_first_coeff_index(first_coeff_index),
@@ -122,7 +122,7 @@ class TensorBlock {
         m_tensor_strides(tensor_strides),
         m_data(data) {}
 
-  Index first_coeff_index() const { return m_first_coeff_index; }
+  StorageIndex first_coeff_index() const { return m_first_coeff_index; }
 
   const Dimensions& block_sizes() const { return m_block_sizes; }
 
@@ -135,108 +135,33 @@ class TensorBlock {
   const Scalar* data() const { return m_data; }
 
  private:
-  Index m_first_coeff_index;
+  StorageIndex m_first_coeff_index;
   Dimensions m_block_sizes;
   Dimensions m_block_strides;
   Dimensions m_tensor_strides;
   Scalar* m_data;  // Not owned.
 };
 
-template <typename Scalar, typename Index, bool Vectorizable>
+template <typename Scalar, typename StorageIndex>
 struct TensorBlockCopyOp {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const Index num_coeff_to_copy, const Index dst_index,
-      const Index dst_stride, Scalar* EIGEN_RESTRICT dst_data,
-      const Index src_index, const Index src_stride,
+      const StorageIndex num_coeff_to_copy, const StorageIndex dst_index,
+      const StorageIndex dst_stride, Scalar* EIGEN_RESTRICT dst_data,
+      const StorageIndex src_index, const StorageIndex src_stride,
       const Scalar* EIGEN_RESTRICT src_data) {
-    for (Index i = 0; i < num_coeff_to_copy; ++i) {
-      dst_data[dst_index + i * dst_stride] =
-          src_data[src_index + i * src_stride];
-    }
-  }
-};
+    const Scalar* src_base = &src_data[src_index];
+    Scalar* dst_base = &dst_data[dst_index];
 
-// NOTE: Benchmarks run on an implementation of this that broke each of the
-// loops in these conditionals into it's own template specialization (to
-// avoid conditionals in the caller's loop) did not show an improvement.
-template <typename Scalar, typename Index>
-struct TensorBlockCopyOp<Scalar, Index, true> {
-  typedef typename packet_traits<Scalar>::type Packet;
-  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const Index num_coeff_to_copy, const Index dst_index,
-      const Index dst_stride, Scalar* EIGEN_RESTRICT dst_data,
-      const Index src_index, const Index src_stride,
-      const Scalar* EIGEN_RESTRICT src_data) {
-    if (src_stride == 1) {
-      const Index packet_size = internal::unpacket_traits<Packet>::size;
-      const Index vectorized_size =
-          (num_coeff_to_copy / packet_size) * packet_size;
-      if (dst_stride == 1) {
-        // LINEAR
-        for (Index i = 0; i < vectorized_size; i += packet_size) {
-          Packet p = internal::ploadu<Packet>(src_data + src_index + i);
-          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
-        }
-        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i] = src_data[src_index + i];
-        }
-      } else {
-        // SCATTER
-        for (Index i = 0; i < vectorized_size; i += packet_size) {
-          Packet p = internal::ploadu<Packet>(src_data + src_index + i);
-          internal::pscatter<Scalar, Packet>(
-              dst_data + dst_index + i * dst_stride, p, dst_stride);
-        }
-        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i * dst_stride] = src_data[src_index + i];
-        }
-      }
-    } else if (src_stride == 0) {
-      const Index packet_size = internal::unpacket_traits<Packet>::size;
-      const Index vectorized_size =
-          (num_coeff_to_copy / packet_size) * packet_size;
-      if (dst_stride == 1) {
-        // LINEAR
-        for (Index i = 0; i < vectorized_size; i += packet_size) {
-          Packet p = internal::pload1<Packet>(src_data + src_index);
-          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
-        }
-        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i] = src_data[src_index];
-        }
-      } else {
-        // SCATTER
-        for (Index i = 0; i < vectorized_size; i += packet_size) {
-          Packet p = internal::pload1<Packet>(src_data + src_index);
-          internal::pscatter<Scalar, Packet>(
-              dst_data + dst_index + i * dst_stride, p, dst_stride);
-        }
-        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i * dst_stride] = src_data[src_index];
-        }
-      }
-    } else {
-      if (dst_stride == 1) {
-        // GATHER
-        const Index packet_size = internal::unpacket_traits<Packet>::size;
-        const Index vectorized_size =
-            (num_coeff_to_copy / packet_size) * packet_size;
-        for (Index i = 0; i < vectorized_size; i += packet_size) {
-          Packet p = internal::pgather<Scalar, Packet>(
-              src_data + src_index + i * src_stride, src_stride);
-          internal::pstoreu<Scalar, Packet>(dst_data + dst_index + i, p);
-        }
-        for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i] = src_data[src_index + i * src_stride];
-        }
-      } else {
-        // RANDOM
-        for (Index i = 0; i < num_coeff_to_copy; ++i) {
-          dst_data[dst_index + i * dst_stride] =
-              src_data[src_index + i * src_stride];
-        }
-      }
-    }
+    using Src = const Eigen::Array<Scalar, Dynamic, 1>;
+    using Dst = Eigen::Array<Scalar, Dynamic, 1>;
+
+    using SrcMap = Eigen::Map<Src, 0, InnerStride<>>;
+    using DstMap = Eigen::Map<Dst, 0, InnerStride<>>;
+
+    const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
+    DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
+
+    dst = src;
   }
 };
 
@@ -249,34 +174,34 @@ struct TensorBlockCopyOp<Scalar, Index, true> {
  * This class is responsible for copying data between a tensor and a tensor
  * block.
  */
-template <typename Scalar, typename Index, int NumDims, int Layout,
-          bool Vectorizable, bool BlockRead>
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout,
+          bool BlockRead>
 class TensorBlockIO {
  public:
-  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef typename internal::TensorBlockCopyOp<Scalar, Index, Vectorizable>
+  typedef typename internal::TensorBlockCopyOp<Scalar, StorageIndex>
       TensorBlockCopyOp;
 
  protected:
   struct BlockIteratorState {
-    Index input_stride;
-    Index output_stride;
-    Index input_span;
-    Index output_span;
-    Index size;
-    Index count;
+    StorageIndex input_stride;
+    StorageIndex output_stride;
+    StorageIndex input_span;
+    StorageIndex output_span;
+    StorageIndex size;
+    StorageIndex count;
   };
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(
-      const TensorBlock& block, Index first_coeff_index,
-      const array<Index, NumDims>& tensor_to_block_dim_map,
-      const array<Index, NumDims>& tensor_strides, const Scalar* src_data,
+      const TensorBlock& block, StorageIndex first_coeff_index,
+      const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
+      const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data,
       Scalar* dst_data) {
     // Find the innermost tensor dimension whose size is not 1. This is the
     // effective inner dim. If all dimensions are of size 1, then fallback to
     // using the actual innermost dim to avoid out-of-bound access.
-    Index num_size_one_inner_dims = 0;
+    StorageIndex num_size_one_inner_dims = 0;
     for (int i = 0; i < NumDims; ++i) {
       const int dim = cond<Layout>()(i, NumDims - i - 1);
       if (block.block_sizes()[tensor_to_block_dim_map[dim]] != 1) {
@@ -285,16 +210,16 @@ class TensorBlockIO {
       }
     }
     // Calculate strides and dimensions.
-    const Index tensor_stride1_dim = cond<Layout>()(
+    const StorageIndex tensor_stride1_dim = cond<Layout>()(
         num_size_one_inner_dims, NumDims - num_size_one_inner_dims - 1);
-    const Index block_dim_for_tensor_stride1_dim =
+    const StorageIndex block_dim_for_tensor_stride1_dim =
         NumDims == 0 ? 1 : tensor_to_block_dim_map[tensor_stride1_dim];
     size_t block_inner_dim_size =
         NumDims == 0 ? 1
                      : block.block_sizes()[block_dim_for_tensor_stride1_dim];
     for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
       const int dim = cond<Layout>()(i, NumDims - i - 1);
-      const Index block_stride =
+      const StorageIndex block_stride =
           block.block_strides()[tensor_to_block_dim_map[dim]];
       if (block_inner_dim_size == block_stride &&
           block_stride == tensor_strides[dim]) {
@@ -306,10 +231,10 @@ class TensorBlockIO {
       }
     }
 
-    Index inputIndex;
-    Index outputIndex;
-    Index input_stride;
-    Index output_stride;
+    StorageIndex inputIndex;
+    StorageIndex outputIndex;
+    StorageIndex input_stride;
+    StorageIndex output_stride;
 
     // Setup strides to read/write along the tensor's stride1 dimension.
     if (BlockRead) {
@@ -337,7 +262,7 @@ class TensorBlockIO {
     int num_squeezed_dims = 0;
     for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
       const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
-      const Index size = block.block_sizes()[tensor_to_block_dim_map[dim]];
+      const StorageIndex size = block.block_sizes()[tensor_to_block_dim_map[dim]];
       if (size == 1) {
         continue;
       }
@@ -362,9 +287,9 @@ class TensorBlockIO {
     }
 
     // Iterate copying data from src to dst.
-    const Index block_total_size =
+    const StorageIndex block_total_size =
         NumDims == 0 ? 1 : block.block_sizes().TotalSize();
-    for (Index i = 0; i < block_total_size; i += block_inner_dim_size) {
+    for (StorageIndex i = 0; i < block_total_size; i += block_inner_dim_size) {
       TensorBlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
                              dst_data, inputIndex, input_stride, src_data);
       // Update index.
@@ -391,19 +316,18 @@ class TensorBlockIO {
  * This class is responsible for reading a tensor block.
  *
  */
-template <typename Scalar, typename Index, int NumDims, int Layout,
-          bool Vectorizable>
-class TensorBlockReader
-    : public TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, true> {
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
+class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
+                                               Layout, /*BlockRead=*/true> {
  public:
-  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, true>
+  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true>
       Base;
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
       TensorBlock* block, const Scalar* src_data) {
-    array<Index, NumDims> tensor_to_block_dim_map;
+    array<StorageIndex, NumDims> tensor_to_block_dim_map;
     for (int i = 0; i < NumDims; ++i) {
       tensor_to_block_dim_map[i] = i;
     }
@@ -412,9 +336,9 @@ class TensorBlockReader
   }
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      TensorBlock* block, Index first_coeff_index,
-      const array<Index, NumDims>& tensor_to_block_dim_map,
-      const array<Index, NumDims>& tensor_strides, const Scalar* src_data) {
+      TensorBlock* block, StorageIndex first_coeff_index,
+      const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
+      const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data) {
     Base::Copy(*block, first_coeff_index, tensor_to_block_dim_map,
                tensor_strides, src_data, block->data());
   }
@@ -429,19 +353,18 @@ class TensorBlockReader
  * This class is responsible for writing a tensor block.
  *
  */
-template <typename Scalar, typename Index, int NumDims, int Layout,
-          bool Vectorizable>
-class TensorBlockWriter : public TensorBlockIO<Scalar, Index, NumDims, Layout,
-                                               Vectorizable, false> {
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
+class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
+                                               Layout, /*BlockRead=*/false> {
  public:
-  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef TensorBlockIO<Scalar, Index, NumDims, Layout, Vectorizable, false>
+  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false>
       Base;
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
       const TensorBlock& block, Scalar* dst_data) {
-    array<Index, NumDims> tensor_to_block_dim_map;
+    array<StorageIndex, NumDims> tensor_to_block_dim_map;
     for (int i = 0; i < NumDims; ++i) {
       tensor_to_block_dim_map[i] = i;
     }
@@ -450,9 +373,9 @@ class TensorBlockWriter : public TensorBlockIO<Scalar, Index, NumDims, Layout,
   }
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const TensorBlock& block, Index first_coeff_index,
-      const array<Index, NumDims>& tensor_to_block_dim_map,
-      const array<Index, NumDims>& tensor_strides, Scalar* dst_data) {
+      const TensorBlock& block, StorageIndex first_coeff_index,
+      const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
+      const array<StorageIndex, NumDims>& tensor_strides, Scalar* dst_data) {
     Base::Copy(block, first_coeff_index, tensor_to_block_dim_map,
                tensor_strides, block.data(), dst_data);
   }
@@ -468,67 +391,34 @@ class TensorBlockWriter : public TensorBlockIO<Scalar, Index, NumDims, Layout,
  * result of the cwise binary op to the strided output array.
  *
  */
-template <bool Vectorizable>
 struct TensorBlockCwiseBinaryOp {
-  template <typename Index, typename BinaryFunctor, typename OutputScalar,
+  template <typename StorageIndex, typename BinaryFunctor, typename OutputScalar,
             typename LeftScalar, typename RightScalar>
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const BinaryFunctor& functor, const Index num_coeff,
-      const Index output_index, const Index output_stride,
-      OutputScalar* output_data, const Index left_index,
-      const Index left_stride, const LeftScalar* left_data,
-      const Index right_index, const Index right_stride,
+      const BinaryFunctor& functor, const StorageIndex num_coeff,
+      const StorageIndex output_index, const StorageIndex output_stride,
+      OutputScalar* output_data, const StorageIndex left_index,
+      const StorageIndex left_stride, const LeftScalar* left_data,
+      const StorageIndex right_index, const StorageIndex right_stride,
       const RightScalar* right_data) {
-    for (Index i = 0; i < num_coeff; ++i) {
-      output_data[output_index + i * output_stride] =
-          functor(left_data[left_index + i * left_stride],
-                  right_data[right_index + i * right_stride]);
-    }
-  }
-};
+    using Lhs = const Eigen::Array<LeftScalar, Dynamic, 1>;
+    using Rhs = const Eigen::Array<RightScalar, Dynamic, 1>;
+    using Out = Eigen::Array<OutputScalar, Dynamic, 1>;
 
-template <>
-struct TensorBlockCwiseBinaryOp<true> {
-  template <typename Index, typename BinaryFunctor, typename OutputScalar,
-            typename LeftScalar, typename RightScalar>
-  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const BinaryFunctor& functor, const Index num_coeff,
-      const Index output_index, const Index output_stride,
-      OutputScalar* output_data, const Index left_index,
-      const Index left_stride, const LeftScalar* left_data,
-      const Index right_index, const Index right_stride,
-      const RightScalar* right_data) {
-    EIGEN_STATIC_ASSERT(functor_traits<BinaryFunctor>::PacketAccess,
-                        YOU_MADE_A_PROGRAMMING_MISTAKE);
-    typedef typename packet_traits<OutputScalar>::type OutputPacket;
-    typedef typename packet_traits<LeftScalar>::type LeftPacket;
-    typedef typename packet_traits<RightScalar>::type RightPacket;
-    const Index packet_size = unpacket_traits<OutputPacket>::size;
-    EIGEN_STATIC_ASSERT(unpacket_traits<LeftPacket>::size == packet_size,
-                        YOU_MADE_A_PROGRAMMING_MISTAKE);
-    EIGEN_STATIC_ASSERT(unpacket_traits<RightPacket>::size == packet_size,
-                        YOU_MADE_A_PROGRAMMING_MISTAKE);
-    const Index vectorized_size = (num_coeff / packet_size) * packet_size;
-    if (output_stride != 1 || left_stride != 1 || right_stride != 1) {
-      TensorBlockCwiseBinaryOp<false>::Run(
-          functor, num_coeff, output_index, output_stride, output_data,
-          left_index, left_stride, left_data, right_index, right_stride,
-          right_data);
-      return;
-    }
-    // Vectorization for the most common case.
-    for (Index i = 0; i < vectorized_size; i += packet_size) {
-      LeftPacket l = internal::ploadu<LeftPacket>(left_data + left_index + i);
-      RightPacket r =
-          internal::ploadu<RightPacket>(right_data + right_index + i);
-      OutputPacket p = functor.packetOp(l, r);
-      internal::pstoreu<OutputScalar, OutputPacket>(
-          output_data + output_index + i, p);
-    }
-    for (Index i = vectorized_size; i < num_coeff; ++i) {
-      output_data[output_index + i] =
-          functor(left_data[left_index + i], right_data[right_index + i]);
-    }
+    using LhsMap = Eigen::Map<Lhs, 0, InnerStride<>>;
+    using RhsMap = Eigen::Map<Rhs, 0, InnerStride<>>;
+    using OutMap = Eigen::Map<Out, 0, InnerStride<>>;
+
+    const LeftScalar* lhs_base = &left_data[left_index];
+    const RightScalar* rhs_base = &right_data[right_index];
+    OutputScalar* out_base = &output_data[output_index];
+
+    const LhsMap lhs(lhs_base, num_coeff, InnerStride<>(left_stride));
+    const RhsMap rhs(rhs_base, num_coeff, InnerStride<>(right_stride));
+    OutMap out(out_base, num_coeff, InnerStride<>(output_stride));
+
+    out =
+        Eigen::CwiseBinaryOp<BinaryFunctor, LhsMap, RhsMap>(lhs, rhs, functor);
   }
 };
 
@@ -541,28 +431,26 @@ struct TensorBlockCwiseBinaryOp<true> {
  * This class carries out the binary op on given blocks.
  *
  */
-template <typename BinaryFunctor, typename Index, typename OutputScalar,
+template <typename BinaryFunctor, typename StorageIndex, typename OutputScalar,
           int NumDims, int Layout>
 struct TensorBlockCwiseBinaryIO {
-  typedef typename internal::TensorBlock<OutputScalar, Index, NumDims,
+  typedef typename internal::TensorBlock<OutputScalar, StorageIndex, NumDims,
                                          Layout>::Dimensions Dimensions;
-  typedef internal::TensorBlockCwiseBinaryOp<
-      functor_traits<BinaryFunctor>::PacketAccess>
-      TensorBlockCwiseBinaryOp;
 
   struct BlockIteratorState {
-    Index output_stride, output_span;
-    Index left_stride, left_span;
-    Index right_stride, right_span;
-    Index size, count;
+    StorageIndex output_stride, output_span;
+    StorageIndex left_stride, left_span;
+    StorageIndex right_stride, right_span;
+    StorageIndex size, count;
   };
 
   template <typename LeftScalar, typename RightScalar>
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
       const BinaryFunctor& functor, const Dimensions& block_sizes,
       const Dimensions& block_strides, OutputScalar* output_data,
-      const array<Index, NumDims>& left_strides, const LeftScalar* left_data,
-      const array<Index, NumDims>& right_strides,
+      const array<StorageIndex, NumDims>& left_strides,
+      const LeftScalar* left_data,
+      const array<StorageIndex, NumDims>& right_strides,
       const RightScalar* right_data) {
     // Find the innermost dimension whose size is not 1. This is the effective
     // inner dim. If all dimensions are of size 1, fallback to using the actual
@@ -580,7 +468,7 @@ struct TensorBlockCwiseBinaryIO {
         NumDims == 0 ? 1
                      : cond<Layout>()(num_size_one_inner_dims,
                                       NumDims - num_size_one_inner_dims - 1);
-    Index inner_dim_size = NumDims == 0 ? 1 : block_sizes[inner_dim];
+    StorageIndex inner_dim_size = NumDims == 0 ? 1 : block_sizes[inner_dim];
     for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
       const int dim = cond<Layout>()(i, NumDims - i - 1);
       // Merge multiple inner dims into one for larger inner dim size (i.e.
@@ -595,10 +483,12 @@ struct TensorBlockCwiseBinaryIO {
       }
     }
 
-    Index output_index = 0, left_index = 0, right_index = 0;
-    const Index output_stride = NumDims == 0 ? 1 : block_strides[inner_dim];
-    const Index left_stride = NumDims == 0 ? 1 : left_strides[inner_dim];
-    const Index right_stride = NumDims == 0 ? 1 : right_strides[inner_dim];
+    StorageIndex output_index = 0, left_index = 0, right_index = 0;
+    const StorageIndex output_stride =
+        NumDims == 0 ? 1 : block_strides[inner_dim];
+    const StorageIndex left_stride = NumDims == 0 ? 1 : left_strides[inner_dim];
+    const StorageIndex right_stride =
+        NumDims == 0 ? 1 : right_strides[inner_dim];
 
     const int at_least_1_dim = NumDims <= 1 ? 1 : NumDims - 1;
     array<BlockIteratorState, at_least_1_dim> block_iter_state;
@@ -607,7 +497,7 @@ struct TensorBlockCwiseBinaryIO {
     int num_squeezed_dims = 0;
     for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
       const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
-      const Index size = block_sizes[dim];
+      const StorageIndex size = block_sizes[dim];
       if (size == 1) {
         continue;
       }
@@ -624,8 +514,9 @@ struct TensorBlockCwiseBinaryIO {
     }
 
     // Compute cwise binary op.
-    const Index block_total_size = NumDims == 0 ? 1 : block_sizes.TotalSize();
-    for (Index i = 0; i < block_total_size; i += inner_dim_size) {
+    const StorageIndex block_total_size =
+        NumDims == 0 ? 1 : block_sizes.TotalSize();
+    for (StorageIndex i = 0; i < block_total_size; i += inner_dim_size) {
       TensorBlockCwiseBinaryOp::Run(functor, inner_dim_size, output_index,
                                     output_stride, output_data, left_index,
                                     left_stride, left_data, right_index,
@@ -661,10 +552,10 @@ struct TensorBlockCwiseBinaryIO {
 template <class ArgType, class Device>
 struct TensorBlockView {
   typedef TensorEvaluator<ArgType, Device> Impl;
-  typedef typename Impl::Index Index;
+  typedef typename Impl::Index StorageIndex;
   typedef typename remove_const<typename Impl::Scalar>::type Scalar;
   static const int NumDims = array_size<typename Impl::Dimensions>::value;
-  typedef DSizes<Index, NumDims> Dimensions;
+  typedef DSizes<StorageIndex, NumDims> Dimensions;
 
   // Constructs a TensorBlockView for `impl`. `block` is only used for for
   // specifying the start offset, shape, and strides of the block.
@@ -701,7 +592,7 @@ struct TensorBlockView {
           }
         }
       }
-      TensorBlock<Scalar, Index, NumDims, Impl::Layout> input_block(
+      TensorBlock<Scalar, StorageIndex, NumDims, Impl::Layout> input_block(
           block.first_coeff_index(), m_block_sizes, m_block_strides,
           block.tensor_strides(), m_allocated_data);
       impl.block(&input_block);
@@ -733,21 +624,21 @@ struct TensorBlockView {
  *
  * This class is responsible for iterating over the blocks of a tensor.
  */
-template <typename Scalar, typename Index, int NumDims, int Layout>
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockMapper {
  public:
-  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef DSizes<Index, NumDims> Dimensions;
+  typedef DSizes<StorageIndex, NumDims> Dimensions;
 
   TensorBlockMapper(const Dimensions& dims,
                     const TensorBlockShapeType block_shape,
-                    size_t min_target_size)
+                    Index min_target_size)
       : m_dimensions(dims),
         m_block_dim_sizes(BlockDimensions(dims, block_shape, min_target_size)) {
     // Calculate block counts by dimension and total block count.
-    DSizes<Index, NumDims> block_count;
-    for (size_t i = 0; i < block_count.rank(); ++i) {
+    DSizes<StorageIndex, NumDims> block_count;
+    for (Index i = 0; i < block_count.rank(); ++i) {
       block_count[i] = divup(m_dimensions[i], m_block_dim_sizes[i]);
     }
     m_total_block_count = array_prod(block_count);
@@ -773,15 +664,15 @@ class TensorBlockMapper {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
-  GetBlockForIndex(Index block_index, Scalar* data) const {
-    Index first_coeff_index = 0;
-    DSizes<Index, NumDims> coords;
-    DSizes<Index, NumDims> sizes;
-    DSizes<Index, NumDims> strides;
+  GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
+    StorageIndex first_coeff_index = 0;
+    DSizes<StorageIndex, NumDims> coords;
+    DSizes<StorageIndex, NumDims> sizes;
+    DSizes<StorageIndex, NumDims> strides;
     if (NumDims > 0) {
       if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
         for (int i = NumDims - 1; i > 0; --i) {
-          const Index idx = block_index / m_block_strides[i];
+          const StorageIndex idx = block_index / m_block_strides[i];
           coords[i] = idx * m_block_dim_sizes[i];
           sizes[i] =
               numext::mini((m_dimensions[i] - coords[i]), m_block_dim_sizes[i]);
@@ -799,7 +690,7 @@ class TensorBlockMapper {
         }
       } else {
         for (int i = 0; i < NumDims - 1; ++i) {
-          const Index idx = block_index / m_block_strides[i];
+          const StorageIndex idx = block_index / m_block_strides[i];
           coords[i] = idx * m_block_dim_sizes[i];
           sizes[i] =
               numext::mini((m_dimensions[i] - coords[i]), m_block_dim_sizes[i]);
@@ -824,19 +715,20 @@ class TensorBlockMapper {
                        data);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index total_block_count() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {
     return m_total_block_count;
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index block_dims_total_size() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex
+  block_dims_total_size() const {
     return m_block_dim_sizes.TotalSize();
   }
 
  private:
   static Dimensions BlockDimensions(const Dimensions& tensor_dims,
                                     const TensorBlockShapeType block_shape,
-                                    size_t min_target_size) {
-    min_target_size = numext::maxi<size_t>(1, min_target_size);
+                                    Index min_target_size) {
+    min_target_size = numext::maxi<Index>(1, min_target_size);
 
     // If tensor fully fits into the target size, we'll treat it a single block.
     Dimensions block_dim_sizes = tensor_dims;
@@ -865,14 +757,14 @@ class TensorBlockMapper {
               dim_size_target, static_cast<size_t>(tensor_dims[i]));
         }
         // Add any un-allocated coefficients to inner dimension(s).
-        Index total_size = block_dim_sizes.TotalSize();
+        StorageIndex total_size = block_dim_sizes.TotalSize();
         for (int i = 0; i < NumDims; ++i) {
           const int dim = cond<Layout>()(i, NumDims - i - 1);
           if (block_dim_sizes[dim] < tensor_dims[dim]) {
-            const Index total_size_other_dims =
+            const StorageIndex total_size_other_dims =
                 total_size / block_dim_sizes[dim];
-            const Index alloc_avail =
-                divup<Index>(min_target_size, total_size_other_dims);
+            const StorageIndex alloc_avail =
+                divup<StorageIndex>(min_target_size, total_size_other_dims);
             if (alloc_avail == block_dim_sizes[dim]) {
               // Insufficient excess coefficients to allocate.
               break;
@@ -882,14 +774,14 @@ class TensorBlockMapper {
           }
         }
       } else if (block_shape == TensorBlockShapeType::kSkewedInnerDims) {
-        Index coeff_to_allocate = min_target_size;
+        StorageIndex coeff_to_allocate = min_target_size;
         for (int i = 0; i < NumDims; ++i) {
           const int dim = cond<Layout>()(i, NumDims - i - 1);
           block_dim_sizes[dim] =
               numext::mini(coeff_to_allocate, tensor_dims[dim]);
-          coeff_to_allocate =
-              divup(coeff_to_allocate,
-                    numext::maxi(static_cast<Index>(1), block_dim_sizes[dim]));
+          coeff_to_allocate = divup(
+              coeff_to_allocate,
+              numext::maxi(static_cast<StorageIndex>(1), block_dim_sizes[dim]));
         }
         eigen_assert(coeff_to_allocate == 1);
       } else {
@@ -908,7 +800,7 @@ class TensorBlockMapper {
   Dimensions m_block_dim_sizes;
   Dimensions m_block_strides;
   Dimensions m_tensor_strides;
-  Index m_total_block_count;
+  StorageIndex m_total_block_count;
 };
 
 /**
@@ -923,12 +815,12 @@ class TensorBlockMapper {
  * processed together.
  *
  */
-template <typename Scalar, typename Index, int NumDims, int Layout>
+template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorSliceBlockMapper {
  public:
-  typedef typename internal::TensorBlock<Scalar, Index, NumDims, Layout>
+  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef DSizes<Index, NumDims> Dimensions;
+  typedef DSizes<StorageIndex, NumDims> Dimensions;
 
   TensorSliceBlockMapper(const Dimensions& tensor_dims,
                          const Dimensions& tensor_slice_offsets,
@@ -942,7 +834,7 @@ class TensorSliceBlockMapper {
         m_block_stride_order(block_stride_order),
         m_total_block_count(1) {
     // Calculate block counts by dimension and total block count.
-    DSizes<Index, NumDims> block_count;
+    DSizes<StorageIndex, NumDims> block_count;
     for (size_t i = 0; i < block_count.rank(); ++i) {
       block_count[i] = divup(m_tensor_slice_extents[i], m_block_dim_sizes[i]);
     }
@@ -969,11 +861,11 @@ class TensorSliceBlockMapper {
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
-  GetBlockForIndex(Index block_index, Scalar* data) const {
-    Index first_coeff_index = 0;
-    DSizes<Index, NumDims> coords;
-    DSizes<Index, NumDims> sizes;
-    DSizes<Index, NumDims> strides;
+  GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
+    StorageIndex first_coeff_index = 0;
+    DSizes<StorageIndex, NumDims> coords;
+    DSizes<StorageIndex, NumDims> sizes;
+    DSizes<StorageIndex, NumDims> strides;
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int i = NumDims - 1; i > 0; --i) {
         const Index idx = block_index / m_block_strides[i];
@@ -991,16 +883,16 @@ class TensorSliceBlockMapper {
           m_block_dim_sizes[0]);
       first_coeff_index += coords[0] * m_tensor_strides[0];
 
-      Index prev_dim = m_block_stride_order[0];
+      StorageIndex prev_dim = m_block_stride_order[0];
       strides[prev_dim] = 1;
       for (int i = 1; i < NumDims; ++i) {
-        const Index curr_dim = m_block_stride_order[i];
+        const StorageIndex curr_dim = m_block_stride_order[i];
         strides[curr_dim] = strides[prev_dim] * sizes[prev_dim];
         prev_dim = curr_dim;
       }
     } else {
       for (int i = 0; i < NumDims - 1; ++i) {
-        const Index idx = block_index / m_block_strides[i];
+        const StorageIndex idx = block_index / m_block_strides[i];
         coords[i] = m_tensor_slice_offsets[i] + idx * m_block_dim_sizes[i];
         sizes[i] = numext::mini(
             m_tensor_slice_offsets[i] + m_tensor_slice_extents[i] - coords[i],
@@ -1016,10 +908,10 @@ class TensorSliceBlockMapper {
           m_block_dim_sizes[NumDims - 1]);
       first_coeff_index += coords[NumDims - 1] * m_tensor_strides[NumDims - 1];
 
-      Index prev_dim = m_block_stride_order[NumDims - 1];
+      StorageIndex prev_dim = m_block_stride_order[NumDims - 1];
       strides[prev_dim] = 1;
       for (int i = NumDims - 2; i >= 0; --i) {
-        const Index curr_dim = m_block_stride_order[i];
+        const StorageIndex curr_dim = m_block_stride_order[i];
         strides[curr_dim] = strides[prev_dim] * sizes[prev_dim];
         prev_dim = curr_dim;
       }
@@ -1029,7 +921,7 @@ class TensorSliceBlockMapper {
                        data);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index total_block_count() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {
     return m_total_block_count;
   }
 
@@ -1041,7 +933,7 @@ class TensorSliceBlockMapper {
   Dimensions m_block_dim_sizes;
   Dimensions m_block_stride_order;
   Dimensions m_block_strides;
-  Index m_total_block_count;
+  StorageIndex m_total_block_count;
 };
 
 }  // namespace internal
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 7ff0d323b..343ab6269 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -1,5 +1,4 @@
 // This file is part of Eigen, a lightweight C++ template library
-// This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
 // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index ba02802d2..f9a1bd68c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -51,12 +51,10 @@ struct TensorEvaluator
       typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
       TensorBlock;
   typedef typename internal::TensorBlockReader<
-      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
-      PacketAccess>
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
       TensorBlockReader;
   typedef typename internal::TensorBlockWriter<
-      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
-      PacketAccess>
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
       TensorBlockWriter;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device& device)
@@ -204,8 +202,7 @@ struct TensorEvaluator<const Derived, Device>
       typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
       TensorBlock;
   typedef typename internal::TensorBlockReader<
-      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout,
-      PacketAccess>
+      typename internal::remove_const<Scalar>::type, Index, NumCoords, Layout>
       TensorBlockReader;
 
   // Used for accessor extraction in SYCL Managed TensorMap:
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 024de3696..ac5afd891 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -36,15 +36,16 @@ template <typename Expression, typename Device, bool Vectorizable,
           bool Tileable>
 class TensorExecutor {
  public:
-  typedef typename Expression::Index Index;
+  using StorageIndex = typename Expression::Index;
+
   EIGEN_DEVICE_FUNC
   static inline void run(const Expression& expr,
                          const Device& device = Device()) {
     TensorEvaluator<Expression, Device> evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
-      const Index size = array_prod(evaluator.dimensions());
-      for (Index i = 0; i < size; ++i) {
+      const StorageIndex size = array_prod(evaluator.dimensions());
+      for (StorageIndex i = 0; i < size; ++i) {
         evaluator.evalScalar(i);
       }
     }
@@ -56,35 +57,36 @@ class TensorExecutor {
  * Process all the data with a single cpu thread, using vectorized instructions.
  */
 template <typename Expression>
-class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true, /*Tilable*/ false> {
+class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true,
+                     /*Tileable*/ false> {
  public:
-  typedef typename Expression::Index Index;
+  using StorageIndex = typename Expression::Index;
 
   EIGEN_DEVICE_FUNC
-  static inline void run(const Expression& expr, const DefaultDevice& device = DefaultDevice())
-  {
+  static inline void run(const Expression& expr,
+                         const DefaultDevice& device = DefaultDevice()) {
     TensorEvaluator<Expression, DefaultDevice> evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
-    if (needs_assign)
-    {
-      const Index size = array_prod(evaluator.dimensions());
+    if (needs_assign) {
+      const StorageIndex size = array_prod(evaluator.dimensions());
       const int PacketSize = unpacket_traits<typename TensorEvaluator<
           Expression, DefaultDevice>::PacketReturnType>::size;
 
       // Give compiler a strong possibility to unroll the loop. But don't insist
       // on unrolling, because if the function is expensive compiler should not
       // unroll the loop at the expense of inlining.
-      const Index UnrolledSize = (size / (4 * PacketSize)) * 4 * PacketSize;
-      for (Index i = 0; i < UnrolledSize; i += 4*PacketSize) {
-        for (Index j = 0; j < 4; j++) {
+      const StorageIndex UnrolledSize =
+          (size / (4 * PacketSize)) * 4 * PacketSize;
+      for (StorageIndex i = 0; i < UnrolledSize; i += 4 * PacketSize) {
+        for (StorageIndex j = 0; j < 4; j++) {
           evaluator.evalPacket(i + j * PacketSize);
         }
       }
-      const Index VectorizedSize = (size / PacketSize) * PacketSize;
-      for (Index i = UnrolledSize; i < VectorizedSize; i += PacketSize) {
+      const StorageIndex VectorizedSize = (size / PacketSize) * PacketSize;
+      for (StorageIndex i = UnrolledSize; i < VectorizedSize; i += PacketSize) {
         evaluator.evalPacket(i);
       }
-      for (Index i = VectorizedSize; i < size; ++i) {
+      for (StorageIndex i = VectorizedSize; i < size; ++i) {
         evaluator.evalScalar(i);
       }
     }
@@ -97,42 +99,41 @@ class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true, /*Tilable
  * sizing a block to fit L1 cache we get better cache performance.
  */
 template <typename Expression, bool Vectorizable>
-class TensorExecutor<Expression, DefaultDevice, Vectorizable, /*Tilable*/ true> {
+class TensorExecutor<Expression, DefaultDevice, Vectorizable,
+                     /*Tileable*/ true> {
  public:
-  typedef typename Expression::Index Index;
+  using Scalar = typename traits<Expression>::Scalar;
+  using ScalarNoConst = typename remove_const<Scalar>::type;
+
+  using Evaluator = TensorEvaluator<Expression, DefaultDevice>;
+  using StorageIndex = typename traits<Expression>::Index;
+
+  static const int NumDims = traits<Expression>::NumDimensions;
 
   EIGEN_DEVICE_FUNC
   static inline void run(const Expression& expr,
                          const DefaultDevice& device = DefaultDevice()) {
-    using Evaluator = TensorEvaluator<Expression, DefaultDevice>;
-
-    using Index = typename traits<Expression>::Index;
-    const int NumDims = traits<Expression>::NumDimensions;
-
-    using Scalar = typename traits<Expression>::Scalar;
-    using ScalarNoConst = typename remove_const<Scalar>::type;
-
     using TensorBlock =
-        TensorBlock<ScalarNoConst, Index, NumDims, Evaluator::Layout>;
-    using TensorBlockMapper =
-        TensorBlockMapper<ScalarNoConst, Index, NumDims, Evaluator::Layout>;
+        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
+    using TensorBlockMapper = TensorBlockMapper<ScalarNoConst, StorageIndex,
+                                                NumDims, Evaluator::Layout>;
 
     Evaluator evaluator(expr, device);
-    std::size_t total_size = array_prod(evaluator.dimensions());
-    std::size_t cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
+    Index total_size = array_prod(evaluator.dimensions());
+    Index cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
 
     if (total_size < cache_size) {
       // TODO(andydavis) Reduce block management overhead for small tensors.
       // TODO(wuke) Do not do this when evaluating TensorBroadcastingOp.
       internal::TensorExecutor<Expression, DefaultDevice, Vectorizable,
-                               false>::run(expr, device);
+                               /*Tileable*/ false>::run(expr, device);
       return;
     }
 
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
       // Size tensor blocks to fit in cache (or requested target block size).
-      size_t block_total_size = numext::mini(cache_size, total_size);
+      Index block_total_size = numext::mini(cache_size, total_size);
       TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
       // Query expression tree for desired block size/shape.
       std::vector<TensorOpResourceRequirements> resources;
@@ -146,8 +147,8 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable, /*Tilable*/ true>
       Scalar* data = static_cast<Scalar*>(
           device.allocate(block_total_size * sizeof(Scalar)));
 
-      const Index total_block_count = block_mapper.total_block_count();
-      for (Index i = 0; i < total_block_count; ++i) {
+      const StorageIndex total_block_count = block_mapper.total_block_count();
+      for (StorageIndex i = 0; i < total_block_count; ++i) {
         TensorBlock block = block_mapper.GetBlockForIndex(i, data);
         evaluator.evalBlock(&block);
       }
@@ -162,37 +163,38 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable, /*Tilable*/ true>
  * executed on a single core.
  */
 #ifdef EIGEN_USE_THREADS
-template <typename Evaluator, typename Index, bool Vectorizable>
+template <typename Evaluator, typename StorageIndex, bool Vectorizable>
 struct EvalRange {
-  static void run(Evaluator* evaluator_in, const Index first, const Index last) {
+  static void run(Evaluator* evaluator_in, const StorageIndex first,
+                  const StorageIndex last) {
     Evaluator evaluator = *evaluator_in;
     eigen_assert(last >= first);
-    for (Index i = first; i < last; ++i) {
+    for (StorageIndex i = first; i < last; ++i) {
       evaluator.evalScalar(i);
     }
   }
 
-  static Index alignBlockSize(Index size) {
-    return size;
-  }
+  static StorageIndex alignBlockSize(StorageIndex size) { return size; }
 };
 
-template <typename Evaluator, typename Index>
-struct EvalRange<Evaluator, Index, /*Vectorizable*/ true> {
-  static const int PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
+template <typename Evaluator, typename StorageIndex>
+struct EvalRange<Evaluator, StorageIndex, /*Vectorizable*/ true> {
+  static const int PacketSize =
+      unpacket_traits<typename Evaluator::PacketReturnType>::size;
 
-  static void run(Evaluator* evaluator_in, const Index first, const Index last) {
+  static void run(Evaluator* evaluator_in, const StorageIndex first,
+                  const StorageIndex last) {
     Evaluator evaluator = *evaluator_in;
     eigen_assert(last >= first);
-    Index i = first;
+    StorageIndex i = first;
     if (last - first >= PacketSize) {
       eigen_assert(first % PacketSize == 0);
-      Index last_chunk_offset = last - 4 * PacketSize;
+      StorageIndex last_chunk_offset = last - 4 * PacketSize;
       // Give compiler a strong possibility to unroll the loop. But don't insist
       // on unrolling, because if the function is expensive compiler should not
       // unroll the loop at the expense of inlining.
-      for (; i <= last_chunk_offset; i += 4*PacketSize) {
-        for (Index j = 0; j < 4; j++) {
+      for (; i <= last_chunk_offset; i += 4 * PacketSize) {
+        for (StorageIndex j = 0; j < 4; j++) {
           evaluator.evalPacket(i + j * PacketSize);
         }
       }
@@ -206,7 +208,7 @@ struct EvalRange<Evaluator, Index, /*Vectorizable*/ true> {
     }
   }
 
-  static Index alignBlockSize(Index size) {
+  static StorageIndex alignBlockSize(StorageIndex size) {
     // Align block size to packet size and account for unrolling in run above.
     if (size >= 16 * PacketSize) {
       return (size + 4 * PacketSize - 1) & ~(4 * PacketSize - 1);
@@ -219,24 +221,24 @@ struct EvalRange<Evaluator, Index, /*Vectorizable*/ true> {
 template <typename Expression, bool Vectorizable, bool Tileable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
  public:
-  typedef typename Expression::Index Index;
+  using StorageIndex = typename Expression::Index;
 
   static inline void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
     typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
-    typedef EvalRange<Evaluator, Index, Vectorizable> EvalRange;
+    typedef EvalRange<Evaluator, StorageIndex, Vectorizable> EvalRange;
 
     Evaluator evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
     if (needs_assign) {
-      const Index PacketSize =
+      const StorageIndex PacketSize =
           Vectorizable
               ? unpacket_traits<typename Evaluator::PacketReturnType>::size
               : 1;
-      const Index size = array_prod(evaluator.dimensions());
+      const StorageIndex size = array_prod(evaluator.dimensions());
       device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
                          EvalRange::alignBlockSize,
-                         [&evaluator](Index first, Index last) {
+                         [&evaluator](StorageIndex first, StorageIndex last) {
                            EvalRange::run(&evaluator, first, last);
                          });
     }
@@ -247,24 +249,24 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
 template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ true> {
  public:
-  typedef typename Expression::Index Index;
+  using Scalar = typename traits<Expression>::Scalar;
+  using ScalarNoConst = typename remove_const<Scalar>::type;
+
+  using Evaluator = TensorEvaluator<Expression, ThreadPoolDevice>;
+  using StorageIndex = typename traits<Expression>::Index;
+
+  static const int NumDims = traits<Expression>::NumDimensions;
 
   static inline void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
-    typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
-    typedef typename internal::remove_const<
-        typename traits<Expression>::Scalar>::type Scalar;
-    typedef typename traits<Expression>::Index Index;
-
-    static const int NumDims = traits<Expression>::NumDimensions;
-
-    typedef TensorBlock<Scalar, Index, NumDims, Evaluator::Layout> TensorBlock;
-    typedef TensorBlockMapper<Scalar, Index, NumDims, Evaluator::Layout>
-        TensorBlockMapper;
+    using TensorBlock =
+        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
+    using TensorBlockMapper =
+        TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
 
     Evaluator evaluator(expr, device);
-    std::size_t total_size = array_prod(evaluator.dimensions());
-    std::size_t cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
+    StorageIndex total_size = array_prod(evaluator.dimensions());
+    StorageIndex cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
     if (total_size < cache_size) {
       // TODO(andydavis) Reduce block management overhead for small tensors.
       internal::TensorExecutor<Expression, ThreadPoolDevice, Vectorizable,
@@ -276,7 +278,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
     if (needs_assign) {
       TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
-      size_t block_total_size = 0;
+      Index block_total_size = 0;
       // Query expression tree for desired block size/shape.
       std::vector<internal::TensorOpResourceRequirements> resources;
       evaluator.getResourceRequirements(&resources);
@@ -296,15 +298,16 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
       void* buf = device.allocate((num_threads + 1) * aligned_blocksize);
       device.parallelFor(
           block_mapper.total_block_count(), cost * block_size,
-          [=, &device, &evaluator, &block_mapper](Index first, Index last) {
+          [=, &device, &evaluator, &block_mapper](StorageIndex first,
+                                                  StorageIndex last) {
             // currentThreadId() returns -1 if called from a thread not in the
-            // threadpool, such as the main thread dispatching Eigen
+            // thread pool, such as the main thread dispatching Eigen
             // expressions.
             const int thread_idx = device.currentThreadId();
             eigen_assert(thread_idx >= -1 && thread_idx < num_threads);
             Scalar* thread_buf = reinterpret_cast<Scalar*>(
                 static_cast<char*>(buf) + aligned_blocksize * (thread_idx + 1));
-            for (Index i = first; i < last; ++i) {
+            for (StorageIndex i = first; i < last; ++i) {
               auto block = block_mapper.GetBlockForIndex(i, thread_buf);
               evaluator.evalBlock(&block);
             }
@@ -324,51 +327,51 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
 template <typename Expression, bool Vectorizable, bool Tileable>
 class TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable> {
  public:
-  typedef typename Expression::Index Index;
+  typedef typename Expression::Index StorageIndex;
   static void run(const Expression& expr, const GpuDevice& device);
 };
 
 
 #if defined(EIGEN_GPUCC)
-template <typename Evaluator, typename Index, bool Vectorizable>
+template <typename Evaluator, typename StorageIndex, bool Vectorizable>
 struct EigenMetaKernelEval {
   static __device__ EIGEN_ALWAYS_INLINE
-  void run(Evaluator& eval, Index first, Index last, Index step_size) {
-    for (Index i = first; i < last; i += step_size) {
+  void run(Evaluator& eval, StorageIndex first, StorageIndex last, StorageIndex step_size) {
+    for (StorageIndex i = first; i < last; i += step_size) {
       eval.evalScalar(i);
     }
   }
 };
 
-template <typename Evaluator, typename Index>
-struct EigenMetaKernelEval<Evaluator, Index, true> {
+template <typename Evaluator, typename StorageIndex>
+struct EigenMetaKernelEval<Evaluator, StorageIndex, true> {
   static __device__ EIGEN_ALWAYS_INLINE
-  void run(Evaluator& eval, Index first, Index last, Index step_size) {
-    const Index PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
-    const Index vectorized_size = (last / PacketSize) * PacketSize;
-    const Index vectorized_step_size = step_size * PacketSize;
+  void run(Evaluator& eval, StorageIndex first, StorageIndex last, StorageIndex step_size) {
+    const StorageIndex PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
+    const StorageIndex vectorized_size = (last / PacketSize) * PacketSize;
+    const StorageIndex vectorized_step_size = step_size * PacketSize;
 
     // Use the vector path
-    for (Index i = first * PacketSize; i < vectorized_size;
+    for (StorageIndex i = first * PacketSize; i < vectorized_size;
          i += vectorized_step_size) {
       eval.evalPacket(i);
     }
-    for (Index i = vectorized_size + first; i < last; i += step_size) {
+    for (StorageIndex i = vectorized_size + first; i < last; i += step_size) {
       eval.evalScalar(i);
     }
   }
 };
 
-template <typename Evaluator, typename Index>
+template <typename Evaluator, typename StorageIndex>
 __global__ void
 __launch_bounds__(1024)
-EigenMetaKernel(Evaluator eval, Index size) {
+EigenMetaKernel(Evaluator eval, StorageIndex size) {
 
-  const Index first_index = blockIdx.x * blockDim.x + threadIdx.x;
-  const Index step_size = blockDim.x * gridDim.x;
+  const StorageIndex first_index = blockIdx.x * blockDim.x + threadIdx.x;
+  const StorageIndex step_size = blockDim.x * gridDim.x;
 
   const bool vectorizable = Evaluator::PacketAccess & Evaluator::IsAligned;
-  EigenMetaKernelEval<Evaluator, Index, vectorizable>::run(eval, first_index, size, step_size);
+  EigenMetaKernelEval<Evaluator, StorageIndex, vectorizable>::run(eval, first_index, size, step_size);
 }
 
 /*static*/
@@ -382,12 +385,12 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable>::run(
     const int block_size = device.maxGpuThreadsPerBlock();
     const int max_blocks = device.getNumGpuMultiProcessors() *
                            device.maxGpuThreadsPerMultiProcessor() / block_size;
-    const Index size = array_prod(evaluator.dimensions());
+    const StorageIndex size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
     const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
 
     LAUNCH_GPU_KERNEL(
-        (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),
+        (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, StorageIndex>),
         num_blocks, block_size, 0, device, evaluator, size);
   }
   evaluator.cleanup();
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 416b686e4..6feeff231 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -37,6 +37,31 @@ static std::size_t RandomTargetSize(const DSizes<Index, NumDims>& dims) {
   return internal::random<int>(1, dims.TotalSize());
 }
 
+template <int NumDims>
+static DSizes<Index, NumDims> RandomDims() {
+  array<Index, NumDims> dims;
+  for (int i = 0; i < NumDims; ++i) {
+    dims[i] = internal::random<int>(1, 20);
+  }
+  return DSizes<Index, NumDims>(dims);
+};
+
+/** Dummy data type to test TensorBlock copy ops. */
+struct Data {
+  Data() : Data(0) {}
+  explicit Data(int v) { value = v; }
+  int value;
+};
+
+bool operator==(const Data& lhs, const Data& rhs) {
+  return lhs.value == rhs.value;
+}
+
+std::ostream& operator<<(std::ostream& os, const Data& d) {
+  os << "Data: value=" << d.value;
+  return os;
+}
+
 template <typename T>
 static T* GenerateRandomData(const Index& size) {
   T* data = new T[size];
@@ -46,6 +71,23 @@ static T* GenerateRandomData(const Index& size) {
   return data;
 }
 
+template <>
+Data* GenerateRandomData(const Index& size) {
+  Data* data = new Data[size];
+  for (int i = 0; i < size; ++i) {
+    data[i] = Data(internal::random<int>(1, 100));
+  }
+  return data;
+}
+
+template <int NumDims>
+static void Debug(DSizes<Index, NumDims> dims) {
+  for (int i = 0; i < NumDims; ++i) {
+    std::cout << dims[i] << "; ";
+  }
+  std::cout << std::endl;
+}
+
 template <int Layout>
 static void test_block_mapper_sanity()
 {
@@ -96,7 +138,7 @@ static void test_block_mapper_sanity()
 // index in the visited set. Verify that every coeff accessed only once.
 template <typename T, int Layout, int NumDims>
 static void UpdateCoeffSet(
-    const internal::TensorBlock<T, Index, 4, Layout>& block,
+    const internal::TensorBlock<T, Index, NumDims, Layout>& block,
     Index first_coeff_index, int dim_index, std::set<Index>* visited_coeffs) {
   const DSizes<Index, NumDims> block_sizes = block.block_sizes();
   const DSizes<Index, NumDims> tensor_strides = block.tensor_strides();
@@ -114,14 +156,13 @@ static void UpdateCoeffSet(
   }
 }
 
-template <int Layout>
-static void test_block_mapper_maps_every_element()
-{
-  using T = int;
-  using TensorBlock = internal::TensorBlock<T, Index, 4, Layout>;
-  using TensorBlockMapper = internal::TensorBlockMapper<T, Index, 4, Layout>;
+template <typename T, int NumDims, int Layout>
+static void test_block_mapper_maps_every_element() {
+  using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
+  using TensorBlockMapper =
+      internal::TensorBlockMapper<T, Index, NumDims, Layout>;
 
-  DSizes<Index, 4> dims(5, 7, 11, 17);
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>();
 
   // Keep track of elements indices available via block access.
   std::set<Index> coeff_set;
@@ -131,29 +172,36 @@ static void test_block_mapper_maps_every_element()
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
     TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
-    UpdateCoeffSet<T, Layout, 4>(block, block.first_coeff_index(),
-                                 choose(Layout, 3, 0), &coeff_set);
+    UpdateCoeffSet<T, Layout, NumDims>(block, block.first_coeff_index(),
+                                       choose(Layout, NumDims - 1, 0),
+                                       &coeff_set);
   }
 
   // Verify that every coefficient in the original Tensor is accessible through
   // TensorBlock only once.
-  auto total_coeffs = static_cast<int>(dims.TotalSize());
+  Index total_coeffs = dims.TotalSize();
   VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
-  VERIFY_IS_EQUAL(*coeff_set.begin(), static_cast<Index>(0));
-  VERIFY_IS_EQUAL(*coeff_set.rbegin(), static_cast<Index>(total_coeffs - 1));
+  VERIFY_IS_EQUAL(*coeff_set.begin(), 0);
+  VERIFY_IS_EQUAL(*coeff_set.rbegin(), total_coeffs - 1);
 }
 
-template <int Layout>
-static void test_slice_block_mapper_maps_every_element()
-{
-  using T = int;
-  using TensorBlock = internal::TensorBlock<T, Index, 4, Layout>;
+template <typename T, int NumDims, int Layout>
+static void test_slice_block_mapper_maps_every_element() {
+  using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
   using TensorSliceBlockMapper =
-      internal::TensorSliceBlockMapper<T, Index, 4, Layout>;
+      internal::TensorSliceBlockMapper<T, Index, NumDims, Layout>;
 
-  DSizes<Index, 4> tensor_dims(5,7,11,17);
-  DSizes<Index, 4> tensor_slice_offsets(1,3,5,7);
-  DSizes<Index, 4> tensor_slice_extents(3,2,4,5);
+  DSizes<Index, NumDims> tensor_dims = RandomDims<NumDims>();
+  DSizes<Index, NumDims> tensor_slice_offsets = RandomDims<NumDims>();
+  DSizes<Index, NumDims> tensor_slice_extents = RandomDims<NumDims>();
+
+  // Make sure that tensor offsets + extents do not overflow.
+  for (int i = 0; i < NumDims; ++i) {
+    tensor_slice_offsets[i] =
+        numext::mini(tensor_dims[i] - 1, tensor_slice_offsets[i]);
+    tensor_slice_extents[i] = numext::mini(
+        tensor_slice_extents[i], tensor_dims[i] - tensor_slice_offsets[i]);
+  }
 
   // Keep track of elements indices available via block access.
   std::set<Index> coeff_set;
@@ -161,61 +209,59 @@ static void test_slice_block_mapper_maps_every_element()
   auto total_coeffs = static_cast<int>(tensor_slice_extents.TotalSize());
 
   // Pick a random dimension sizes for the tensor blocks.
-  DSizes<Index, 4> block_sizes;
-  for (int i = 0; i < 4; ++i) {
+  DSizes<Index, NumDims> block_sizes;
+  for (int i = 0; i < NumDims; ++i) {
     block_sizes[i] = internal::random<int>(1, tensor_slice_extents[i]);
   }
 
   TensorSliceBlockMapper block_mapper(tensor_dims, tensor_slice_offsets,
                                       tensor_slice_extents, block_sizes,
-                                      DimensionList<Index, 4>());
+                                      DimensionList<Index, NumDims>());
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
     TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
-    UpdateCoeffSet<T, Layout, 4>(block, block.first_coeff_index(),
-                                 choose(Layout, 3, 0), &coeff_set);
+    UpdateCoeffSet<T, Layout, NumDims>(block, block.first_coeff_index(),
+                                       choose(Layout, NumDims - 1, 0),
+                                       &coeff_set);
   }
 
   VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
 }
 
-template <int Layout>
-static void test_block_io_copy_data_from_source_to_target()
-{
-  using T = float;
+template <typename T, int NumDims, int Layout>
+static void test_block_io_copy_data_from_source_to_target() {
+  typedef internal::TensorBlock<T, Index, NumDims, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, NumDims, Layout>
+      TensorBlockMapper;
 
-  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
-
-  typedef internal::TensorBlockReader<T, Index, 5, Layout, true>
+  typedef internal::TensorBlockReader<T, Index, NumDims, Layout>
       TensorBlockReader;
-  typedef internal::TensorBlockWriter<T, Index, 5, Layout, true>
+  typedef internal::TensorBlockWriter<T, Index, NumDims, Layout>
       TensorBlockWriter;
 
-  typedef std::vector<T, aligned_allocator<T>> DataVector;
-
-  DSizes<Index, 5> input_tensor_dims(5, 7, 11, 17, 3);
+  DSizes<Index, NumDims> input_tensor_dims = RandomDims<NumDims>();
   const auto input_tensor_size = input_tensor_dims.TotalSize();
-  DataVector input_data(input_tensor_size, 0);
-  for (int i = 0; i < input_tensor_size; ++i) {
-    input_data[i] = internal::random<T>();
-  }
 
-  DataVector output_data(input_tensor_size, 0);
+  T* input_data = GenerateRandomData<T>(input_tensor_size);
+  T* output_data = new T[input_tensor_size];
 
   TensorBlockMapper block_mapper(input_tensor_dims, RandomShape(),
                                  RandomTargetSize(input_tensor_dims));
+  T* block_data = new T[block_mapper.block_dims_total_size()];
 
-  DataVector block_data(block_mapper.block_dims_total_size(), 0);
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
-    TensorBlock block = block_mapper.GetBlockForIndex(i, block_data.data());
-    TensorBlockReader::Run(&block, input_data.data());
-    TensorBlockWriter::Run(block, output_data.data());
+    TensorBlock block = block_mapper.GetBlockForIndex(i, block_data);
+    TensorBlockReader::Run(&block, input_data);
+    TensorBlockWriter::Run(block, output_data);
   }
 
   for (int i = 0; i < input_tensor_size; ++i) {
     VERIFY_IS_EQUAL(input_data[i], output_data[i]);
   }
+
+  delete[] input_data;
+  delete[] output_data;
+  delete[] block_data;
 }
 
 template <int Layout, int NumDims>
@@ -261,31 +307,32 @@ static array<Index, NumDims> ComputeStrides(
   return strides;
 }
 
-template <int Layout>
+template <typename T, int NumDims, int Layout>
 static void test_block_io_copy_using_reordered_dimensions() {
-  typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockMapper<float, Index, 5, Layout>
+  typedef internal::TensorBlock<T, Index, NumDims, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, NumDims, Layout>
       TensorBlockMapper;
 
-  typedef internal::TensorBlockReader<float, Index, 5, Layout, false>
+  typedef internal::TensorBlockReader<T, Index, NumDims, Layout>
       TensorBlockReader;
-  typedef internal::TensorBlockWriter<float, Index, 5, Layout, false>
+  typedef internal::TensorBlockWriter<T, Index, NumDims, Layout>
       TensorBlockWriter;
 
-  DSizes<Index, 5> input_tensor_dims(5, 7, 11, 17, 3);
+  DSizes<Index, NumDims> input_tensor_dims = RandomDims<NumDims>();
   const auto input_tensor_size = input_tensor_dims.TotalSize();
 
   // Create a random input tensor.
-  auto* input_data = GenerateRandomData<float>(input_tensor_size);
+  T* input_data = GenerateRandomData<T>(input_tensor_size);
 
   // Create a random dimension re-ordering/shuffle.
-  std::vector<Index> shuffle = {0, 1, 2, 3, 4};
+  std::vector<Index> shuffle;
+  for (int i = 0; i < NumDims; ++i) shuffle.push_back(i);
   std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
 
-  DSizes<Index, 5> output_tensor_dims;
-  array<Index, 5> input_to_output_dim_map;
-  array<Index, 5> output_to_input_dim_map;
-  for (Index i = 0; i < 5; ++i) {
+  DSizes<Index, NumDims> output_tensor_dims;
+  array<Index, NumDims> input_to_output_dim_map;
+  array<Index, NumDims> output_to_input_dim_map;
+  for (Index i = 0; i < NumDims; ++i) {
     output_tensor_dims[shuffle[i]] = input_tensor_dims[i];
     input_to_output_dim_map[i] = shuffle[i];
     output_to_input_dim_map[shuffle[i]] = i;
@@ -295,17 +342,17 @@ static void test_block_io_copy_using_reordered_dimensions() {
   TensorBlockMapper block_mapper(output_tensor_dims, RandomShape(),
                                  RandomTargetSize(input_tensor_dims));
 
-  auto* block_data = new float[block_mapper.block_dims_total_size()];
-  auto* output_data = new float[input_tensor_size];
+  auto* block_data = new T[block_mapper.block_dims_total_size()];
+  auto* output_data = new T[input_tensor_size];
 
-  array<Index, 5> input_tensor_strides =
-      ComputeStrides<Layout, 5>(input_tensor_dims);
-  array<Index, 5> output_tensor_strides =
-      ComputeStrides<Layout, 5>(output_tensor_dims);
+  array<Index, NumDims> input_tensor_strides =
+      ComputeStrides<Layout, NumDims>(input_tensor_dims);
+  array<Index, NumDims> output_tensor_strides =
+      ComputeStrides<Layout, NumDims>(output_tensor_dims);
 
   for (Index i = 0; i < block_mapper.total_block_count(); ++i) {
     TensorBlock block = block_mapper.GetBlockForIndex(i, block_data);
-    const Index first_coeff_index = GetInputIndex<Layout, 5>(
+    const Index first_coeff_index = GetInputIndex<Layout, NumDims>(
         block.first_coeff_index(), output_to_input_dim_map,
         input_tensor_strides, output_tensor_strides);
     TensorBlockReader::Run(&block, first_coeff_index, input_to_output_dim_map,
@@ -327,18 +374,21 @@ template <int Layout>
 static void test_block_io_zero_stride()
 {
   typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockReader<float, Index, 5, Layout, true>
+  typedef internal::TensorBlockReader<float, Index, 5, Layout>
       TensorBlockReader;
-  typedef internal::TensorBlockWriter<float, Index, 5, Layout, true>
+  typedef internal::TensorBlockWriter<float, Index, 5, Layout>
       TensorBlockWriter;
 
-  DSizes<Index, 5> input_tensor_dims(1, 2, 1, 3, 1);
-  const auto input_tensor_size = input_tensor_dims.TotalSize();
+  DSizes<Index, 5> rnd_dims = RandomDims<5>();
 
-  // Create a random input tensor.
+  DSizes<Index, 5> input_tensor_dims = rnd_dims;
+  input_tensor_dims[0] = 1;
+  input_tensor_dims[2] = 1;
+  input_tensor_dims[4] = 1;
+  const auto input_tensor_size = input_tensor_dims.TotalSize();
   auto* input_data = GenerateRandomData<float>(input_tensor_size);
 
-  DSizes<Index, 5> output_tensor_dims(3, 2, 3, 3, 2);
+  DSizes<Index, 5> output_tensor_dims = rnd_dims;
 
   DSizes<Index, 5> input_tensor_strides(
       ComputeStrides<Layout, 5>(input_tensor_dims));
@@ -401,9 +451,9 @@ static void test_block_io_zero_stride()
 template <int Layout>
 static void test_block_io_squeeze_ones() {
   typedef internal::TensorBlock<float, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockReader<float, Index, 5, Layout, true>
+  typedef internal::TensorBlockReader<float, Index, 5, Layout>
       TensorBlockReader;
-  typedef internal::TensorBlockWriter<float, Index, 5, Layout, true>
+  typedef internal::TensorBlockWriter<float, Index, 5, Layout>
       TensorBlockWriter;
 
   // Total size > 1.
@@ -467,23 +517,23 @@ static void test_block_io_squeeze_ones() {
   }
 }
 
-template <int Layout>
+template <typename T, int NumDims, int Layout>
 static void test_block_cwise_binary_io_basic() {
-  typedef internal::scalar_sum_op<float> BinaryFunctor;
-  typedef internal::TensorBlockCwiseBinaryIO<BinaryFunctor, Index, float, 5,
+  typedef internal::scalar_sum_op<T> BinaryFunctor;
+  typedef internal::TensorBlockCwiseBinaryIO<BinaryFunctor, Index, T, NumDims,
                                              Layout>
       TensorBlockCwiseBinaryIO;
 
-  DSizes<Index, 5> block_sizes(2, 3, 5, 7, 11);
-  DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+  DSizes<Index, NumDims> block_sizes = RandomDims<NumDims>();
+  DSizes<Index, NumDims> strides(ComputeStrides<Layout, NumDims>(block_sizes));
 
   const auto total_size = block_sizes.TotalSize();
 
   // Create a random input tensors.
-  auto* left_data = GenerateRandomData<float>(total_size);
-  auto* right_data = GenerateRandomData<float>(total_size);
+  T* left_data = GenerateRandomData<T>(total_size);
+  T* right_data = GenerateRandomData<T>(total_size);
 
-  auto* output_data = new float[total_size];
+  T* output_data = new T[total_size];
   BinaryFunctor functor;
   TensorBlockCwiseBinaryIO::Run(functor, block_sizes, strides, output_data,
                                 strides, left_data, strides, right_data);
@@ -532,13 +582,22 @@ static void test_block_cwise_binary_io_zero_strides() {
                                              Layout>
       TensorBlockCwiseBinaryIO;
 
-  DSizes<Index, 5> left_sizes(1, 3, 1, 7, 1);
+  DSizes<Index, 5> rnd_dims = RandomDims<5>();
+
+  DSizes<Index, 5> left_sizes = rnd_dims;
+  left_sizes[0] = 1;
+  left_sizes[2] = 1;
+  left_sizes[4] = 1;
+
   DSizes<Index, 5> left_strides(ComputeStrides<Layout, 5>(left_sizes));
   left_strides[0] = 0;
   left_strides[2] = 0;
   left_strides[4] = 0;
 
-  DSizes<Index, 5> right_sizes(2, 1, 5, 1, 11);
+  DSizes<Index, 5> right_sizes = rnd_dims;
+  right_sizes[1] = 0;
+  right_sizes[3] = 0;
+
   DSizes<Index, 5> right_strides(ComputeStrides<Layout, 5>(right_sizes));
   right_strides[1] = 0;
   right_strides[3] = 0;
@@ -547,7 +606,7 @@ static void test_block_cwise_binary_io_zero_strides() {
   auto* left_data = GenerateRandomData<float>(left_sizes.TotalSize());
   auto* right_data = GenerateRandomData<float>(right_sizes.TotalSize());
 
-  DSizes<Index, 5> output_sizes(2, 3, 5, 7, 11);
+  DSizes<Index, 5> output_sizes = rnd_dims;
   DSizes<Index, 5> output_strides(ComputeStrides<Layout, 5>(output_sizes));
 
   const auto output_total_size = output_sizes.TotalSize();
@@ -557,11 +616,11 @@ static void test_block_cwise_binary_io_zero_strides() {
   TensorBlockCwiseBinaryIO::Run(functor, output_sizes, output_strides,
                                 output_data, left_strides, left_data,
                                 right_strides, right_data);
-  for (int i = 0; i < 2; ++i) {
-    for (int j = 0; j < 3; ++j) {
-      for (int k = 0; k < 5; ++k) {
-        for (int l = 0; l < 7; ++l) {
-          for (int m = 0; m < 11; ++m) {
+  for (int i = 0; i < rnd_dims[0]; ++i) {
+    for (int j = 0; j < rnd_dims[1]; ++j) {
+      for (int k = 0; k < rnd_dims[2]; ++k) {
+        for (int l = 0; l < rnd_dims[3]; ++l) {
+          for (int m = 0; m < rnd_dims[4]; ++m) {
             Index output_index = i * output_strides[0] + j * output_strides[1] +
                                  k * output_strides[2] + l * output_strides[3] +
                                  m * output_strides[4];
@@ -893,31 +952,44 @@ static void test_empty_dims(const internal::TensorBlockShapeType block_shape)
   }
 }
 
-#define CALL_SUBTEST_LAYOUTS(NAME) \
+#define TEST_LAYOUTS(NAME) \
   CALL_SUBTEST(NAME<ColMajor>()); \
   CALL_SUBTEST(NAME<RowMajor>())
 
-#define CALL_SUBTEST_LAYOUTS_WITH_ARG(NAME, ARG) \
+#define TEST_LAYOUTS_AND_DIMS(TYPE, NAME)    \
+  CALL_SUBTEST((NAME<TYPE, 1, ColMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 1, RowMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 2, ColMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 2, RowMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 3, ColMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 3, RowMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 4, ColMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 4, RowMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 5, ColMajor>())); \
+  CALL_SUBTEST((NAME<TYPE, 5, RowMajor>()))
+
+#define TEST_LAYOUTS_WITH_ARG(NAME, ARG) \
   CALL_SUBTEST(NAME<ColMajor>(ARG)); \
   CALL_SUBTEST(NAME<RowMajor>(ARG))
 
 EIGEN_DECLARE_TEST(cxx11_tensor_block_access) {
-  CALL_SUBTEST_LAYOUTS(test_block_mapper_sanity);
-  CALL_SUBTEST_LAYOUTS(test_block_mapper_maps_every_element);
-  CALL_SUBTEST_LAYOUTS(test_slice_block_mapper_maps_every_element);
-  CALL_SUBTEST_LAYOUTS(test_block_io_copy_data_from_source_to_target);
-  CALL_SUBTEST_LAYOUTS(test_block_io_copy_using_reordered_dimensions);
-  CALL_SUBTEST_LAYOUTS(test_block_io_zero_stride);
-  CALL_SUBTEST_LAYOUTS(test_block_io_squeeze_ones);
-  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_basic);
-  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_squeeze_ones);
-  CALL_SUBTEST_LAYOUTS(test_block_cwise_binary_io_zero_strides);
-  CALL_SUBTEST_LAYOUTS(test_uniform_block_shape);
-  CALL_SUBTEST_LAYOUTS(test_skewed_inner_dim_block_shape);
-
-  CALL_SUBTEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kUniformAllDims);
-  CALL_SUBTEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kSkewedInnerDims);
+  TEST_LAYOUTS(test_block_mapper_sanity);
+  TEST_LAYOUTS_AND_DIMS(float, test_block_mapper_maps_every_element);
+  TEST_LAYOUTS_AND_DIMS(float, test_slice_block_mapper_maps_every_element);
+  TEST_LAYOUTS_AND_DIMS(float, test_block_io_copy_data_from_source_to_target);
+  TEST_LAYOUTS_AND_DIMS(Data, test_block_io_copy_data_from_source_to_target);
+  TEST_LAYOUTS_AND_DIMS(float, test_block_io_copy_using_reordered_dimensions);
+  TEST_LAYOUTS_AND_DIMS(Data, test_block_io_copy_using_reordered_dimensions);
+  TEST_LAYOUTS(test_block_io_zero_stride);
+  TEST_LAYOUTS(test_block_io_squeeze_ones);
+  TEST_LAYOUTS_AND_DIMS(float, test_block_cwise_binary_io_basic);
+  TEST_LAYOUTS(test_block_cwise_binary_io_squeeze_ones);
+  TEST_LAYOUTS(test_block_cwise_binary_io_zero_strides);
+  TEST_LAYOUTS(test_uniform_block_shape);
+  TEST_LAYOUTS(test_skewed_inner_dim_block_shape);
+  TEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kUniformAllDims);
+  TEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kSkewedInnerDims);
 }
 
-#undef CALL_SUBTEST_LAYOUTS
-#undef CALL_SUBTEST_LAYOUTS_WITH_ARG
\ No newline at end of file
+#undef TEST_LAYOUTS
+#undef TEST_LAYOUTS_WITH_ARG
\ No newline at end of file
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 5ae45ac5b..274f901ce 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -13,7 +13,6 @@
 
 #include <Eigen/CXX11/Tensor>
 
-using Eigen::Index;
 using Eigen::Tensor;
 using Eigen::RowMajor;
 using Eigen::ColMajor;
@@ -25,9 +24,16 @@ template <typename Device, bool Vectorizable, bool Tileable, int Layout>
 static void test_execute_binary_expr(Device d) {
   // Pick a large enough tensor size to bypass small tensor block evaluation
   // optimization.
-  Tensor<float, 3> lhs(840, 390, 37);
-  Tensor<float, 3> rhs(840, 390, 37);
-  Tensor<float, 3> dst(840, 390, 37);
+  int d0 = internal::random<int>(100, 200);
+  int d1 = internal::random<int>(100, 200);
+  int d2 = internal::random<int>(100, 200);
+
+  static constexpr int Options = 0;
+  using IndexType = int;
+
+  Tensor<float, 3, Options, IndexType> lhs(d0, d1, d2);
+  Tensor<float, 3, Options, IndexType> rhs(d0, d1, d2);
+  Tensor<float, 3, Options, IndexType> dst(d0, d1, d2);
 
   lhs.setRandom();
   rhs.setRandom();
@@ -40,9 +46,9 @@ static void test_execute_binary_expr(Device d) {
 
   Executor::run(Assign(dst, expr), d);
 
-  for (int i = 0; i < 840; ++i) {
-    for (int j = 0; j < 390; ++j) {
-      for (int k = 0; k < 37; ++k) {
+  for (int i = 0; i < d0; ++i) {
+    for (int j = 0; j < d1; ++j) {
+      for (int k = 0; k < d2; ++k) {
         float sum = lhs(i, j, k) + rhs(i, j, k);
         VERIFY_IS_EQUAL(sum, dst(i, j, k));
       }

From 723856dec1b5d5be0e35be0612e188a30bfa594b Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 30 Jul 2018 14:52:15 +0200
Subject: [PATCH 505/680] bug #1577: fix msvc compilation of unit test, msvc
 defines ptrdiff_t as long long

---
 test/AnnoyingScalar.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/test/AnnoyingScalar.h b/test/AnnoyingScalar.h
index 7d1df2253..2b6544a6a 100644
--- a/test/AnnoyingScalar.h
+++ b/test/AnnoyingScalar.h
@@ -33,6 +33,9 @@ class AnnoyingScalar
     AnnoyingScalar(float _v)        { init(); *v = _v; }
     AnnoyingScalar(int _v)          { init(); *v = _v; }
     AnnoyingScalar(long _v)         { init(); *v = _v; }
+    #if EIGEN_HAS_CXX11
+    AnnoyingScalar(long long _v)    { init(); *v = _v; }
+    #endif
     AnnoyingScalar(const AnnoyingScalar& other) { init(); *v = *(other.v); }
     ~AnnoyingScalar() {
       if(v!=&data)

From 679eece8760ce9b9ff09e48b6ee8673afcf94caa Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 31 Jul 2018 10:10:14 +0200
Subject: [PATCH 506/680] Speedup trivial tensor broadcasting on GPU by
 enforcing unaligned loads. See PR 437.

---
 .../Eigen/CXX11/src/Tensor/TensorBroadcasting.h       | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 278689915..e647b3609 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -284,7 +284,13 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
 
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       if (isCopy) {
+        #ifdef EIGEN_GPU_COMPILE_PHASE
+        // See PR 437: on NVIDIA P100 and K20m we observed a x3-4 speed up by enforcing
+        // unaligned loads here. The reason is unclear though.
+        return m_impl.template packet<Unaligned>(index);
+        #else
         return m_impl.template packet<LoadMode>(index);
+        #endif
       } else if (oneByN && !nByOne) {
         return packetNByOne<LoadMode>(index);
       } else if (!oneByN && nByOne) {
@@ -296,7 +302,12 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       }
     } else {
       if (isCopy) {
+        #ifdef EIGEN_GPU_COMPILE_PHASE
+        // See above.
+        return m_impl.template packet<Unaligned>(index);
+        #else
         return m_impl.template packet<LoadMode>(index);
+        #endif
       } else if (oneByN && !nByOne) {
         return packetOneByN<LoadMode>(index);
       } else if (!oneByN && nByOne) {

From d6568425f8b384ef36da1f16be76a3977fb90846 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 31 Jul 2018 08:13:43 +0000
Subject: [PATCH 507/680] Close branch tiling_3.


From 385f7b8d0ca926d00c71987ab308202511e5c753 Mon Sep 17 00:00:00 2001
From: Paul Tucker <tucker@google.com>
Date: Tue, 31 Jul 2018 13:52:18 -0700
Subject: [PATCH 508/680] Change getAllocator() to allocator() in
 ThreadPoolDevice.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index f4123b71d..f8188ffde 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -281,7 +281,7 @@ struct ThreadPoolDevice {
   ThreadPoolInterface* getPool() const { return pool_; }
 
   // Allocator accessor.
-  Allocator* getAllocator() const { return allocator_; }
+  Allocator* allocator() const { return allocator_; }
 
  private:
   ThreadPoolInterface* pool_;

From 83c0a16baf5ecac6288cd9b74536a82de8985b31 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 31 Jul 2018 15:56:31 -0700
Subject: [PATCH 509/680] Add block evaluation support to TensorOps

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      |  14 +-
 .../CXX11/src/Tensor/TensorBroadcasting.h     | 333 +++++++++++-
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   | 144 +++++-
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h | 218 +++++++-
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   | 333 ++++++++++--
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 437 +++++++++++++++-
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  | 237 ++++++++-
 unsupported/test/cxx11_tensor_executor.cpp    | 480 ++++++++++++++++--
 8 files changed, 2047 insertions(+), 149 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 84cf6d216..33c4ef5b7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -214,7 +214,7 @@ class TensorBlockIO {
         num_size_one_inner_dims, NumDims - num_size_one_inner_dims - 1);
     const StorageIndex block_dim_for_tensor_stride1_dim =
         NumDims == 0 ? 1 : tensor_to_block_dim_map[tensor_stride1_dim];
-    size_t block_inner_dim_size =
+    Index block_inner_dim_size =
         NumDims == 0 ? 1
                      : block.block_sizes()[block_dim_for_tensor_stride1_dim];
     for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
@@ -745,16 +745,15 @@ class TensorBlockMapper {
       if (block_shape == TensorBlockShapeType::kUniformAllDims) {
         // Tensor will not fit within 'min_target_size' budget: calculate tensor
         // block dimension sizes based on "square" dimension size target.
-        const size_t dim_size_target = static_cast<const size_t>(
+        const Index dim_size_target = static_cast<Index>(
             std::pow(static_cast<float>(min_target_size),
                      1.0 / static_cast<float>(block_dim_sizes.rank())));
-        for (size_t i = 0; i < block_dim_sizes.rank(); ++i) {
+        for (Index i = 0; i < block_dim_sizes.rank(); ++i) {
           // TODO(andydavis) Adjust the inner most 'block_dim_size' to make it
           // a multiple of the packet size. Note that reducing
           // 'block_dim_size' in this manner can increase the number of
           // blocks, and so will amplify any per-block overhead.
-          block_dim_sizes[i] = numext::mini(
-              dim_size_target, static_cast<size_t>(tensor_dims[i]));
+          block_dim_sizes[i] = numext::mini(dim_size_target, tensor_dims[i]);
         }
         // Add any un-allocated coefficients to inner dimension(s).
         StorageIndex total_size = block_dim_sizes.TotalSize();
@@ -789,9 +788,8 @@ class TensorBlockMapper {
       }
     }
 
-    eigen_assert(
-        block_dim_sizes.TotalSize() >=
-        numext::mini<size_t>(min_target_size, tensor_dims.TotalSize()));
+    eigen_assert(block_dim_sizes.TotalSize() >=
+                 numext::mini<Index>(min_target_size, tensor_dims.TotalSize()));
 
     return block_dim_sizes;
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 343ab6269..a851e7f55 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -108,16 +108,29 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   bool isCopy= false, nByOne = false, oneByN = false;
 
   enum {
-    IsAligned = true,
+    IsAligned    = true,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess    = false
   };
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-    : m_broadcast(op.broadcast()),m_impl(op.expression(), device)
-  {
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  // Block based access to the XprType (input) tensor.
+  using TensorBlock                = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  using TensorBlockReader          = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
+  // We do block based broadcasting using a a trick with 2x tensor rank and 0
+  // strides. See block method implementation for details.
+  using BroadcastDimensions        = DSizes<Index, 2 * NumDims>;
+  using BroadcastTensorBlock       = internal::TensorBlock<ScalarNoConst, Index, 2 * NumDims, Layout>;
+  using BroadcastTensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, 2 * NumDims, Layout>;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
+                                                        const Device& device)
+      : m_device(device),
+        m_broadcast(op.broadcast()),
+        m_impl(op.expression(), device) {
     // The broadcasting op doesn't change the rank of the tensor. One can't broadcast a scalar
     // and store the result in a scalar. Instead one should reshape the scalar into a a N-D
     // tensor with N >= 1 of 1 element first and then broadcast.
@@ -216,8 +229,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   }
 
   // TODO: attempt to speed this up. The integer divisions and modulo are slow
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffColMajor(Index index) const
-  {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index indexColMajor(Index index) const {
     Index inputIndex = 0;
     for (int i = NumDims - 1; i > 0; --i) {
       const Index idx = index / m_outputStrides[i];
@@ -243,11 +255,15 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
         inputIndex += (index % m_impl.dimensions()[0]);
       }
     }
-    return m_impl.coeff(inputIndex);
+    return inputIndex;
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffRowMajor(Index index) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffColMajor(Index index) const
   {
+    return m_impl.coeff(indexColMajor(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index indexRowMajor(Index index) const {
     Index inputIndex = 0;
     for (int i = 0; i < NumDims - 1; ++i) {
       const Index idx = index / m_outputStrides[i];
@@ -263,17 +279,22 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       }
       index -= idx * m_outputStrides[i];
     }
-    if (internal::index_statically_eq<Broadcast>(NumDims-1, 1)) {
-      eigen_assert(index < m_impl.dimensions()[NumDims-1]);
+    if (internal::index_statically_eq<Broadcast>(NumDims - 1, 1)) {
+      eigen_assert(index < m_impl.dimensions()[NumDims - 1]);
       inputIndex += index;
     } else {
-      if (internal::index_statically_eq<InputDimensions>(NumDims-1, 1)) {
-        eigen_assert(index % m_impl.dimensions()[NumDims-1] == 0);
+      if (internal::index_statically_eq<InputDimensions>(NumDims - 1, 1)) {
+        eigen_assert(index % m_impl.dimensions()[NumDims - 1] == 0);
       } else {
-        inputIndex += (index % m_impl.dimensions()[NumDims-1]);
+        inputIndex += (index % m_impl.dimensions()[NumDims - 1]);
       }
     }
-    return m_impl.coeff(inputIndex);
+    return inputIndex;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffRowMajor(Index index) const
+  {
+    return m_impl.coeff(indexRowMajor(index));
   }
 
   template<int LoadMode>
@@ -553,13 +574,291 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    // TODO(wuke): Targeting L1 size is 30% faster than targeting L{-1} on large
+    // tensors. But this might need further tuning.
+    Index l1_cache_scalars = m_device.firstLevelCacheSize() / sizeof(Scalar);
+    Index block_total_size_max = numext::maxi(Index(1), l1_cache_scalars);
+
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kSkewedInnerDims,
+        block_total_size_max));
+
+    m_impl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      TensorBlock* output_block) const {
+    if (NumDims <= 0) {
+      output_block->data()[0] = m_impl.coeff(0);
+      return;
+    }
+
+    // Because we only support kSkewedInnerDims blocking, block size should be
+    // equal to m_dimensions for inner dims, a smaller than m_dimensions[i] size
+    // for the first outer dim, and 1 for other outer dims. This is guaranteed
+    // by MergeResourceRequirements() in TensorBlock.h.
+    const auto& output_block_sizes = output_block->block_sizes();
+    const auto& output_block_strides = output_block->block_strides();
+
+    // Find where outer dims start.
+    int outer_dim_start = 0;
+    Index outer_dim_size = 1, inner_dim_size = 1;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                          ? i
+                          : NumDims - i - 1;
+      if (i > outer_dim_start) {
+        eigen_assert(output_block_sizes[dim] == 1);
+      } else if (output_block_sizes[dim] != m_dimensions[dim]) {
+        eigen_assert(output_block_sizes[dim] < m_dimensions[dim]);
+        outer_dim_size = output_block_sizes[dim];
+      } else {
+        inner_dim_size *= output_block_sizes[dim];
+        ++outer_dim_start;
+      }
+    }
+
+    if (inner_dim_size == 0 || outer_dim_size == 0) {
+      return;
+    }
+
+    const auto& input_dims = m_impl.dimensions();
+
+    // Pre-fill input_block_sizes, broadcast_block_sizes,
+    // broadcast_block_strides, and broadcast_tensor_strides. Later on we will
+    // only modify the outer_dim_start-th dimension on these arrays.
+
+    // Calculate the input block size for looking into the input.
+    Dimensions input_block_sizes;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < outer_dim_start; ++i) {
+        input_block_sizes[i] = input_dims[i];
+      }
+      for (int i = outer_dim_start; i < NumDims; ++i) {
+        input_block_sizes[i] = 1;
+      }
+    } else {
+      for (int i = 0; i < outer_dim_start; ++i) {
+        input_block_sizes[NumDims - i - 1] = input_dims[NumDims - i - 1];
+      }
+      for (int i = outer_dim_start; i < NumDims; ++i) {
+        input_block_sizes[NumDims - i - 1] = 1;
+      }
+    }
+
+    // Broadcast with the 0-stride trick: Create 1 extra dim for each
+    // broadcast, set the input stride to 0.
+    //
+    // When ColMajor:
+    // - broadcast_block_sizes is [d_0, b_0, d_1, b_1, ...].
+    //
+    // - broadcast_block_strides is [output_block_strides[0],
+    //                               output_block_strides[0] * d_0,
+    //                               output_block_strides[1],
+    //                               output_block_strides[1] * d_1,
+    //                               ...].
+    //
+    // - broadcast_tensor_strides is [output_block_strides[0],
+    //                                0,
+    //                                output_block_strides[1],
+    //                                0,
+    //                                ...].
+    BroadcastDimensions broadcast_block_sizes, broadcast_block_strides,
+        broadcast_tensor_strides;
+
+    for (int i = 0; i < outer_dim_start; ++i) {
+      const int dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                          ? i
+                          : NumDims - i - 1;
+      const int copy_dim =
+          static_cast<int>(Layout) == static_cast<int>(ColMajor)
+              ? 2 * i
+              : 2 * NumDims - 2 * i - 1;
+      const int broadcast_dim =
+          static_cast<int>(Layout) == static_cast<int>(ColMajor) ? copy_dim + 1
+                                                                 : copy_dim - 1;
+      broadcast_block_sizes[copy_dim] = input_dims[dim];
+      broadcast_block_sizes[broadcast_dim] = m_broadcast[dim];
+      broadcast_block_strides[copy_dim] = output_block_strides[dim];
+      broadcast_block_strides[broadcast_dim] =
+          output_block_strides[dim] * input_dims[dim];
+      broadcast_tensor_strides[copy_dim] = m_inputStrides[dim];
+      broadcast_tensor_strides[broadcast_dim] = 0;
+    }
+    for (int i = 2 * outer_dim_start; i < 2 * NumDims; ++i) {
+      const int dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                          ? i
+                          : 2 * NumDims - i - 1;
+      broadcast_block_sizes[dim] = 1;
+      broadcast_block_strides[dim] = 0;
+      broadcast_tensor_strides[dim] = 0;
+    }
+
+    const int outer_dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? outer_dim_start
+                              : NumDims - outer_dim_start - 1;
+
+    if (outer_dim_size == 1) {
+      // We just need one block read using the ready-set values above.
+      BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                     broadcast_block_strides, broadcast_tensor_strides, 0,
+                     output_block);
+    } else if (input_dims[outer_dim] == 1) {
+      // Broadcast outer_dim_start-th dimension (< NumDims) by outer_dim_size.
+      const int broadcast_outer_dim =
+          static_cast<int>(Layout) == static_cast<int>(ColMajor)
+              ? 2 * outer_dim_start + 1
+              : 2 * NumDims - 2 * outer_dim_start - 2;
+      broadcast_block_sizes[broadcast_outer_dim] = outer_dim_size;
+      broadcast_tensor_strides[broadcast_outer_dim] = 0;
+      broadcast_block_strides[broadcast_outer_dim] =
+          output_block_strides[outer_dim];
+      BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                     broadcast_block_strides, broadcast_tensor_strides, 0,
+                     output_block);
+    } else {
+      // The general case. Let's denote the output block as x[...,
+      // a:a+outer_dim_size, :, ..., :], where a:a+outer_dim_size is a slice on
+      // the outer_dim_start-th dimension (< NumDims). We need to split the
+      // a:a+outer_dim_size into possibly 3 sub-blocks:
+      //
+      // (1) a:b, where b is the smallest multiple of
+      // input_dims[outer_dim_start] in [a, a+outer_dim_size].
+      //
+      // (2) b:c, where c is the largest multiple of input_dims[outer_dim_start]
+      // in [a, a+outer_dim_size].
+      //
+      // (3) c:a+outer_dim_size .
+      //
+      // Or, when b and c do not exist, we just need to process the whole block
+      // together.
+
+      // Find a.
+      const Index outer_dim_left_index =
+          output_block->first_coeff_index() / m_outputStrides[outer_dim];
+
+      // Find b and c.
+      const Index input_outer_dim_size = input_dims[outer_dim];
+
+      // First multiple after a. This is b when <= outer_dim_left_index +
+      // outer_dim_size.
+      const Index first_multiple =
+          divup<Index>(outer_dim_left_index, input_outer_dim_size) *
+          input_outer_dim_size;
+
+      if (first_multiple <= outer_dim_left_index + outer_dim_size) {
+        // b exists, so does c. Find it.
+        const Index last_multiple = (outer_dim_left_index + outer_dim_size) /
+                                    input_outer_dim_size * input_outer_dim_size;
+        const int copy_outer_dim =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                ? 2 * outer_dim_start
+                : 2 * NumDims - 2 * outer_dim_start - 1;
+        const int broadcast_outer_dim =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                ? 2 * outer_dim_start + 1
+                : 2 * NumDims - 2 * outer_dim_start - 2;
+        if (first_multiple > outer_dim_left_index) {
+          const Index head_size = first_multiple - outer_dim_left_index;
+          input_block_sizes[outer_dim] = head_size;
+          broadcast_block_sizes[copy_outer_dim] = head_size;
+          broadcast_tensor_strides[copy_outer_dim] = m_inputStrides[outer_dim];
+          broadcast_block_strides[copy_outer_dim] =
+              output_block_strides[outer_dim];
+          broadcast_block_sizes[broadcast_outer_dim] = 1;
+          broadcast_tensor_strides[broadcast_outer_dim] = 0;
+          broadcast_block_strides[broadcast_outer_dim] =
+              output_block_strides[outer_dim] * input_dims[outer_dim];
+          BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                         broadcast_block_strides, broadcast_tensor_strides, 0,
+                         output_block);
+        }
+        if (first_multiple < last_multiple) {
+          input_block_sizes[outer_dim] = input_outer_dim_size;
+          broadcast_block_sizes[copy_outer_dim] = input_outer_dim_size;
+          broadcast_tensor_strides[copy_outer_dim] = m_inputStrides[outer_dim];
+          broadcast_block_strides[copy_outer_dim] =
+              output_block_strides[outer_dim];
+          broadcast_block_sizes[broadcast_outer_dim] =
+              (last_multiple - first_multiple) / input_outer_dim_size;
+          broadcast_tensor_strides[broadcast_outer_dim] = 0;
+          broadcast_block_strides[broadcast_outer_dim] =
+              output_block_strides[outer_dim] * input_dims[outer_dim];
+          const Index offset = (first_multiple - outer_dim_left_index) *
+                               m_outputStrides[outer_dim];
+          BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                         broadcast_block_strides, broadcast_tensor_strides,
+                         offset, output_block);
+        }
+        if (last_multiple < outer_dim_left_index + outer_dim_size) {
+          const Index tail_size =
+              outer_dim_left_index + outer_dim_size - last_multiple;
+          input_block_sizes[outer_dim] = tail_size;
+          broadcast_block_sizes[copy_outer_dim] = tail_size;
+          broadcast_tensor_strides[copy_outer_dim] = m_inputStrides[outer_dim];
+          broadcast_block_strides[copy_outer_dim] =
+              output_block_strides[outer_dim];
+          broadcast_block_sizes[broadcast_outer_dim] = 1;
+          broadcast_tensor_strides[broadcast_outer_dim] = 0;
+          broadcast_block_strides[broadcast_outer_dim] =
+              output_block_strides[outer_dim] * input_dims[outer_dim];
+          const Index offset = (last_multiple - outer_dim_left_index) *
+                               m_outputStrides[outer_dim];
+          BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                         broadcast_block_strides, broadcast_tensor_strides,
+                         offset, output_block);
+        }
+      } else {
+        // b and c do not exist.
+        const int copy_outer_dim =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                ? 2 * outer_dim_start
+                : 2 * NumDims - 2 * outer_dim_start - 1;
+        input_block_sizes[outer_dim] = outer_dim_size;
+        broadcast_block_sizes[copy_outer_dim] = outer_dim_size;
+        broadcast_tensor_strides[copy_outer_dim] = m_inputStrides[outer_dim];
+        broadcast_block_strides[copy_outer_dim] =
+            output_block_strides[outer_dim];
+        BroadcastBlock(input_block_sizes, broadcast_block_sizes,
+                       broadcast_block_strides, broadcast_tensor_strides, 0,
+                       output_block);
+      }
+    }
+  }
+
   EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
 
   Broadcast functor() const { return m_broadcast; }
 
+ private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void BroadcastBlock(
+      const Dimensions& input_block_sizes,
+      const BroadcastDimensions& broadcast_block_sizes,
+      const BroadcastDimensions& broadcast_block_strides,
+      const BroadcastDimensions& broadcast_tensor_strides, Index offset,
+      TensorBlock* output_block) const {
+    TensorBlock input_view_block(
+        static_cast<int>(Layout) == static_cast<int>(ColMajor)
+            ? indexColMajor(output_block->first_coeff_index() + offset)
+            : indexRowMajor(output_block->first_coeff_index() + offset),
+        input_block_sizes, Dimensions(m_inputStrides),
+        Dimensions(m_inputStrides), NULL);
+
+    internal::TensorBlockView<ArgType, Device> input_block(m_device, m_impl,
+                                                           input_view_block);
+    BroadcastTensorBlock broadcast_block(
+        0, broadcast_block_sizes, broadcast_block_strides,
+        broadcast_tensor_strides, output_block->data() + offset);
+
+    BroadcastTensorBlockReader::Run(&broadcast_block, input_block.data());
+  }
+
  protected:
+  const Device& m_device;
   const Broadcast m_broadcast;
   Dimensions m_dimensions;
   array<Index, NumDims> m_outputStrides;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 085c05f3d..a0d039e64 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -144,14 +144,19 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   enum {
     // Alignment can't be guaranteed at compile time since it depends on the
     // slice offsets.
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,  // to be implemented
+    RawAccess    = false
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using InputTensorBlock  = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
+  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device), m_offset(op.offset())
   {
@@ -184,6 +189,23 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     }
     m_inputStride *= input_dims[m_dim.actualDim()];
     m_inputOffset = m_stride * op.offset();
+
+    if (BlockAccess) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_inputStrides[0] = 1;
+        for (int i = 1; i < NumInputDims; ++i) {
+          m_inputStrides[i] = m_inputStrides[i - 1] * input_dims[i - 1];
+        }
+      } else {
+        m_inputStrides[NumInputDims - 1] = 1;
+        for (int i = NumInputDims - 2; i >= 0; --i) {
+          m_inputStrides[i] = m_inputStrides[i + 1] * input_dims[i + 1];
+        }
+      }
+
+      m_block_total_size_max =
+          numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
+    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -266,6 +288,60 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
            TensorOpCost(0, 0, cost, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kSkewedInnerDims,
+        m_block_total_size_max));
+    m_impl.getResourceRequirements(resources);
+  }
+
+  // TODO(andydavis) Reduce the overhead of this function (experiment with
+  // using a fixed block size).
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      OutputTensorBlock* output_block) const {
+    // Calculate input block sizes.
+    const DSizes<Index, NumDims>& output_block_sizes =
+        output_block->block_sizes();
+    const DSizes<Index, NumDims>& output_block_strides =
+        output_block->block_strides();
+    const Index chip_dim = m_dim.actualDim();
+    DSizes<Index, NumInputDims> input_block_sizes;
+    DSizes<Index, NumInputDims> input_block_strides;
+    for (Index i = 0; i < NumInputDims; ++i) {
+      if (i < chip_dim) {
+        input_block_sizes[i] = output_block_sizes[i];
+        input_block_strides[i] = output_block_strides[i];
+      } else if (i > chip_dim) {
+        input_block_sizes[i] = output_block_sizes[i - 1];
+        input_block_strides[i] = output_block_strides[i - 1];
+      } else {
+        input_block_sizes[i] = 1;
+      }
+    }
+    // Fix up input_block_stride for chip dimension.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      if (chip_dim == 0) {
+        input_block_strides[chip_dim] = 1;
+      } else {
+        input_block_strides[chip_dim] =
+            input_block_strides[chip_dim - 1] * input_block_sizes[chip_dim - 1];
+      }
+    } else {
+      if (chip_dim == NumInputDims - 1) {
+        input_block_strides[chip_dim] = 1;
+      } else {
+        input_block_strides[chip_dim] =
+            input_block_strides[chip_dim + 1] * input_block_sizes[chip_dim + 1];
+      }
+    }
+    // Instantiate and read input block from input tensor.
+    InputTensorBlock input_block(srcCoeff(output_block->first_coeff_index()),
+                                 input_block_sizes, input_block_strides,
+                                 m_inputStrides, output_block->data());
+    m_impl.block(&input_block);
+  }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
     CoeffReturnType* result = const_cast<CoeffReturnType*>(m_impl.data());
     if (((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumDims) ||
@@ -316,6 +392,8 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   Index m_stride;
   Index m_inputOffset;
   Index m_inputStride;
+  Index m_block_total_size_max;
+  DSizes<Index, NumInputDims> m_inputStrides;
   TensorEvaluator<ArgType, Device> m_impl;
   const internal::DimensionId<DimId> m_dim;
   const Device& m_device;
@@ -342,12 +420,18 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
 
   enum {
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess    = false
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using InputTensorBlock  = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
+  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
     : Base(op, device)
     { }
@@ -395,6 +479,50 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
       }
     }
   }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const OutputTensorBlock& output_block) {
+    // Calculate input block sizes.
+    const DSizes<Index, NumDims>& output_block_sizes =
+        output_block.block_sizes();
+    const DSizes<Index, NumDims>& output_block_strides =
+        output_block.block_strides();
+    const Index chip_dim = this->m_dim.actualDim();
+    DSizes<Index, NumInputDims> input_block_sizes;
+    DSizes<Index, NumInputDims> input_block_strides;
+    for (Index i = 0; i < NumInputDims; ++i) {
+      if (i < chip_dim) {
+        input_block_sizes[i] = output_block_sizes[i];
+        input_block_strides[i] = output_block_strides[i];
+      } else if (i > chip_dim) {
+        input_block_sizes[i] = output_block_sizes[i - 1];
+        input_block_strides[i] = output_block_strides[i - 1];
+      } else {
+        input_block_sizes[i] = 1;
+      }
+    }
+    // Fix up input_block_stride for chip dimension.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      if (chip_dim == 0) {
+        input_block_strides[chip_dim] = 1;
+      } else {
+        input_block_strides[chip_dim] =
+            input_block_strides[chip_dim - 1] * input_block_sizes[chip_dim - 1];
+      }
+    } else {
+      if (chip_dim == NumInputDims - 1) {
+        input_block_strides[chip_dim] = 1;
+      } else {
+        input_block_strides[chip_dim] =
+            input_block_strides[chip_dim + 1] * input_block_sizes[chip_dim + 1];
+      }
+    }
+    // Write input block.
+    this->m_impl.writeBlock(InputTensorBlock(
+        this->srcCoeff(output_block.first_coeff_index()), input_block_sizes,
+        input_block_strides, this->m_inputStrides,
+        const_cast<ScalarNoConst*>(output_block.data())));
+  }
 };
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 72cb2d15f..4987b898b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -54,6 +54,66 @@ struct nested<TensorImagePatchOp<Rows, Cols, XprType>, 1, typename eval<TensorIm
   typedef TensorImagePatchOp<Rows, Cols, XprType> type;
 };
 
+template <typename Self, bool Vectorizable>
+struct ImagePatchCopyOp {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::Impl Impl;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Self& self, const Index num_coeff_to_copy, const Index dst_index,
+      Scalar* dst_data, const Index src_index) {
+    const Impl& impl = self.impl();
+    for (Index i = 0; i < num_coeff_to_copy; ++i) {
+      dst_data[dst_index + i] = impl.coeff(src_index + i);
+    }
+  }
+};
+
+template <typename Self>
+struct ImagePatchCopyOp<Self, true> {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::Impl Impl;
+  typedef typename packet_traits<Scalar>::type Packet;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Self& self, const Index num_coeff_to_copy, const Index dst_index,
+      Scalar* dst_data, const Index src_index) {
+    const Impl& impl = self.impl();
+    const Index packet_size = internal::unpacket_traits<Packet>::size;
+    const Index vectorized_size =
+        (num_coeff_to_copy / packet_size) * packet_size;
+    for (Index i = 0; i < vectorized_size; i += packet_size) {
+      Packet p = impl.template packet<Unaligned>(src_index + i);
+      internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i, p);
+    }
+    for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+      dst_data[dst_index + i] = impl.coeff(src_index + i);
+    }
+  }
+};
+
+template <typename Self>
+struct ImagePatchPaddingOp {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename packet_traits<Scalar>::type Packet;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Index num_coeff_to_pad, const Scalar padding_value,
+      const Index dst_index, Scalar* dst_data) {
+    const Index packet_size = internal::unpacket_traits<Packet>::size;
+    const Packet padded_packet = internal::pset1<Packet>(padding_value);
+    const Index vectorized_size =
+        (num_coeff_to_pad / packet_size) * packet_size;
+    for (Index i = 0; i < vectorized_size; i += packet_size) {
+      internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i,
+                                                   padded_packet);
+    }
+    for (Index i = vectorized_size; i < num_coeff_to_pad; ++i) {
+      dst_data[dst_index + i] = padding_value;
+    }
+  }
+};
+
 }  // end namespace internal
 
 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
@@ -184,15 +244,17 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
 
   enum {
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,
-    RawAccess = false
+    BlockAccess  = true,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,
+    RawAccess    = false
   };
 
-  #ifdef __SYCL_DEVICE_ONLY__
+  using OutputTensorBlock = internal::TensorBlock<Scalar, Index, NumDims, Layout>;
+
+#ifdef __SYCL_DEVICE_ONLY__
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType op, const Device& device)
   #else
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
@@ -342,6 +404,9 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
     } else {
       m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
     }
+
+    m_block_total_size_max =
+        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -484,6 +549,146 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kSkewedInnerDims,
+        m_block_total_size_max));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      OutputTensorBlock* output_block) const {
+    using ImagePatchCopyOp = internal::ImagePatchCopyOp<Self, PacketAccess>;
+    using ImagePatchPaddingOp = internal::ImagePatchPaddingOp<Self>;
+
+    // Calculate loop limits and various input/output dim sizes.
+    const DSizes<Index, NumDims>& block_sizes = output_block->block_sizes();
+    const bool col_major =
+        static_cast<int>(Layout) == static_cast<int>(ColMajor);
+    const Index depth_dim_size = block_sizes[col_major ? 0 : NumDims - 1];
+    const Index output_depth_dim_size =
+        m_dimensions[col_major ? 0 : NumDims - 1];
+    const Index row_dim_size = block_sizes[col_major ? 1 : NumDims - 2];
+    const Index output_row_dim_size = m_dimensions[col_major ? 1 : NumDims - 2];
+    const Index col_dim_size = block_sizes[col_major ? 2 : NumDims - 3];
+    const Index block_col_stride = row_dim_size * depth_dim_size;
+    const Index patch_index_dim_size = block_sizes[col_major ? 3 : NumDims - 4];
+    const Index outer_dim_size =
+        block_sizes.TotalSize() /
+        (depth_dim_size * row_dim_size * col_dim_size * patch_index_dim_size);
+
+    const Index patch_size = row_dim_size * col_dim_size * depth_dim_size;
+    const Index batch_size = patch_size * patch_index_dim_size;
+
+    Index output_index = output_block->first_coeff_index();
+
+    // Loop through outer dimensions.
+    for (Index outer_dim_index = 0; outer_dim_index < outer_dim_size;
+         ++outer_dim_index) {
+      const Index outer_output_base_index = outer_dim_index * batch_size;
+      // Find the offset of the element wrt the location of the first element.
+      const Index patchIndexStart = output_index / m_fastPatchStride;
+      const Index patchOffset =
+          (output_index - patchIndexStart * m_patchStride) / m_fastOutputDepth;
+      const Index colOffsetStart = patchOffset / m_fastColStride;
+      // Other ways to index this element.
+      const Index otherIndex =
+          (NumDims == 4) ? 0 : output_index / m_fastOtherStride;
+      const Index patch2DIndexStart =
+          (NumDims == 4)
+              ? 0
+              : (output_index - otherIndex * m_otherStride) / m_fastPatchStride;
+      // Calculate starting depth index.
+      const Index depth = output_index - (output_index / m_fastOutputDepth) *
+                                             output_depth_dim_size;
+      const Index patch_input_base_index =
+          depth + otherIndex * m_patchInputStride;
+
+      // Loop through patches.
+      for (Index patch_index_dim_index = 0;
+           patch_index_dim_index < patch_index_dim_size;
+           ++patch_index_dim_index) {
+        const Index patch_output_base_index =
+            outer_output_base_index + patch_index_dim_index * patch_size;
+        // Patch index corresponding to the passed in index.
+        const Index patchIndex = patchIndexStart + patch_index_dim_index;
+        const Index patch2DIndex =
+            (NumDims == 4) ? patchIndex
+                           : patch2DIndexStart + patch_index_dim_index;
+        const Index colIndex = patch2DIndex / m_fastOutputRows;
+        const Index input_col_base = colIndex * m_col_strides;
+        const Index row_offset_base =
+            (patch2DIndex - colIndex * m_outputRows) * m_row_strides -
+            m_rowPaddingTop;
+
+        // Loop through columns.
+        for (Index col_dim_index = 0; col_dim_index < col_dim_size;
+             ++col_dim_index) {
+          const Index col_output_base_index =
+              patch_output_base_index + col_dim_index * block_col_stride;
+
+          // Calculate col index in the input original tensor.
+          Index colOffset = colOffsetStart + col_dim_index;
+          Index inputCol =
+              input_col_base + colOffset * m_in_col_strides - m_colPaddingLeft;
+          Index origInputCol =
+              (m_col_inflate_strides == 1)
+                  ? inputCol
+                  : ((inputCol >= 0) ? (inputCol / m_fastInflateColStride) : 0);
+
+          bool pad_column = false;
+          if (inputCol < 0 || inputCol >= m_input_cols_eff ||
+              ((m_col_inflate_strides != 1) &&
+               (inputCol != origInputCol * m_col_inflate_strides))) {
+            pad_column = true;
+          }
+
+          const Index col_input_base_index =
+              patch_input_base_index + origInputCol * m_colInputStride;
+          const Index input_row_base =
+              row_offset_base +
+              ((patchOffset + col_dim_index * output_row_dim_size) -
+               colOffset * m_colStride) *
+                  m_in_row_strides;
+          // Loop through rows.
+          for (Index row_dim_index = 0; row_dim_index < row_dim_size;
+               ++row_dim_index) {
+            const Index output_base_index =
+                col_output_base_index + row_dim_index * depth_dim_size;
+            bool pad_row = false;
+            Index inputIndex;
+            if (!pad_column) {
+              Index inputRow =
+                  input_row_base + row_dim_index * m_in_row_strides;
+              Index origInputRow =
+                  (m_row_inflate_strides == 1)
+                      ? inputRow
+                      : ((inputRow >= 0) ? (inputRow / m_fastInflateRowStride)
+                                         : 0);
+              if (inputRow < 0 || inputRow >= m_input_rows_eff ||
+                  ((m_row_inflate_strides != 1) &&
+                   (inputRow != origInputRow * m_row_inflate_strides))) {
+                pad_row = true;
+              } else {
+                inputIndex =
+                    col_input_base_index + origInputRow * m_rowInputStride;
+              }
+            }
+            // Copy (or pad) along depth dimension.
+            if (pad_column || pad_row) {
+              ImagePatchPaddingOp::Run(depth_dim_size, Scalar(m_paddingValue),
+                                       output_base_index, output_block->data());
+            } else {
+              ImagePatchCopyOp::Run(*this, depth_dim_size, output_base_index,
+                                    output_block->data(), inputIndex);
+            }
+          }
+        }
+      }
+      output_index += m_otherStride;
+    }
+  }
+
  protected:
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
   {
@@ -538,6 +743,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   internal::TensorIntDivisor<Index> m_fastOutputDepth;
 
   Scalar m_paddingValue;
+  Index m_block_total_size_max;
 
   TensorEvaluator<ArgType, Device> m_impl;
   #ifdef EIGEN_USE_SYCL
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 498488649..3747bff9e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -102,27 +102,64 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
   typedef TensorReshapingOp<NewDimensions, ArgType> XprType;
   typedef NewDimensions Dimensions;
 
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  static const int NumOutputDims = internal::array_size<Dimensions>::value;
+  static const int NumInputDims  = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+
   enum {
-    IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
+    IsAligned    = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
+    // TODO(andydavis, wuke) Enable BlockAccess for the general case when the
+    // performance issue with block-based reshape is resolved.
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess &&
+                   TensorEvaluator<ArgType, Device>::RawAccess &&
+                   NumInputDims > 0 && NumOutputDims > 0,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,  // to be implemented
+    RawAccess    = TensorEvaluator<ArgType, Device>::RawAccess
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using InputTensorBlock        = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
+  using OutputTensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>;
+  using OutputTensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, NumOutputDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_dimensions(op.dimensions())
   {
     // The total size of the reshaped tensor must be equal to the total size
     // of the input tensor.
     eigen_assert(internal::array_prod(m_impl.dimensions()) == internal::array_prod(op.dimensions()));
-  }
 
-  typedef typename XprType::Index Index;
-  typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
-  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+    if (BlockAccess) {
+      const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims =
+          m_impl.dimensions();
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_outputStrides[0] = 1;
+        for (int i = 1; i < NumOutputDims; ++i) {
+          m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+        }
+        m_inputStrides[0] = 1;
+        for (int i = 1; i < NumInputDims; ++i) {
+          m_inputStrides[i] = m_inputStrides[i - 1] * input_dims[i - 1];
+        }
+      } else {
+        m_outputStrides[NumOutputDims - 1] = 1;
+        for (int i = NumOutputDims - 2; i >= 0; --i) {
+          m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+        }
+        m_inputStrides[NumInputDims - 1] = 1;
+        for (int i = NumInputDims - 2; i >= 0; --i) {
+          m_inputStrides[i] = m_inputStrides[i + 1] * input_dims[i + 1];
+        }
+      }
+    }
+  }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
@@ -148,6 +185,140 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
     return m_impl.costPerCoeff(vectorized);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    m_impl.getResourceRequirements(resources);
+  }
+
+  // TODO(andydavis) Reduce the overhead of this function.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      OutputTensorBlock* output_block) const {
+    if (m_impl.data() != NULL) {
+      OutputTensorBlockReader::Run(output_block, m_impl.data());
+      return;
+    }
+
+    // Calculate output block unit-stride inner dimension length.
+    const DSizes<Index, NumOutputDims>& output_block_sizes =
+        output_block->block_sizes();
+    Index output_inner_dim_size = 1;
+    Index output_outer_dim_start = NumOutputDims;
+    for (Index i = 0; i < NumOutputDims; ++i) {
+      const Index dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                        ? i : NumOutputDims - i - 1;
+      output_inner_dim_size *= output_block_sizes[dim];
+      if (output_block_sizes[dim] < m_dimensions[dim]) {
+        output_outer_dim_start = i + 1;
+        break;
+      }
+    }
+
+    // Initialize output block iterator state.
+    struct BlockIteratorState {
+      Index stride;
+      Index span;
+      Index size;
+      Index count;
+    };
+    array<BlockIteratorState, NumOutputDims> block_iter_state;
+
+    for (Index i = 0; i < NumOutputDims; ++i) {
+      const Index dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                        ? i : NumOutputDims - i - 1;
+      block_iter_state[i].size = output_block_sizes[dim];
+      block_iter_state[i].stride = m_outputStrides[dim];
+      block_iter_state[i].span =
+          block_iter_state[i].stride * (block_iter_state[i].size - 1);
+      block_iter_state[i].count = 0;
+    }
+
+    const Index output_outer_dim_size = output_block_sizes.TotalSize() /
+        output_inner_dim_size;
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims =
+        m_impl.dimensions();
+
+    Index index = output_block->first_coeff_index();
+    for (Index outer_idx = 0; outer_idx < output_outer_dim_size; ++outer_idx) {
+      Index inner_idx = 0;
+      while (inner_idx < output_inner_dim_size) {
+        // Calculate input coords based on 'index'.
+        array<Index, NumInputDims> input_coords;
+        Index idx = index;
+        if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+          for (int i = NumInputDims - 1; i > 0; --i) {
+            input_coords[i] = idx / m_inputStrides[i];
+            idx -= input_coords[i] * m_inputStrides[i];
+          }
+          input_coords[0] = idx;
+        } else {
+          for (int i = 0; i < NumInputDims - 1; ++i) {
+            input_coords[i] = idx / m_inputStrides[i];
+            idx -= input_coords[i] * m_inputStrides[i];
+          }
+          input_coords[NumInputDims - 1] = idx;
+        }
+
+        // Calculate target input block shape, using at most
+        // 'output_inner_dim_size' coefficients along the input block's inner
+        // dimensions.
+        DSizes<Index, NumInputDims> input_block_sizes;
+        Index num_to_allocate = output_inner_dim_size - inner_idx;
+        for (Index i = 0; i < NumInputDims; ++i) {
+          const Index dim =
+              static_cast<int>(Layout) == static_cast<int>(ColMajor)
+              ? i : NumInputDims - i - 1;
+          input_block_sizes[dim] = numext::mini(
+              num_to_allocate, (static_cast<Index>(input_dims[dim]) -
+                  input_coords[dim]));
+          if (input_coords[dim] == 0) {
+            num_to_allocate /= input_block_sizes[dim];
+          } else {
+            num_to_allocate = 1;
+          }
+        }
+
+        // Calculate input block strides.
+        DSizes<Index, NumInputDims> input_block_strides;
+        if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+          input_block_strides[0] = 1;
+          for (int i = 1; i < NumInputDims; ++i) {
+            input_block_strides[i] = input_block_strides[i - 1] *
+                input_block_sizes[i - 1];
+          }
+        } else {
+          input_block_strides[NumInputDims - 1] = 1;
+          for (int i = NumInputDims - 2; i >= 0; --i) {
+            input_block_strides[i] = input_block_strides[i + 1] *
+                input_block_sizes[i + 1];
+          }
+        }
+
+        // Instantiate and read input block from input tensor.
+        InputTensorBlock input_block(index, input_block_sizes,
+                                     input_block_strides, m_inputStrides,
+                                     output_block->data() + outer_idx *
+                                         output_inner_dim_size + inner_idx);
+
+        m_impl.block(&input_block);
+
+        const Index input_block_total_size = input_block_sizes.TotalSize();
+        index += input_block_total_size;
+        inner_idx += input_block_total_size;
+      }
+      eigen_assert(inner_idx == output_inner_dim_size);
+      index -= output_inner_dim_size;
+      // Update index.
+      for (Index i = output_outer_dim_start; i < NumOutputDims; ++i) {
+        if (++block_iter_state[i].count < block_iter_state[i].size) {
+          index += block_iter_state[i].stride;
+          break;
+        }
+        block_iter_state[i].count = 0;
+        index -= block_iter_state[i].span;
+      }
+    }
+  }
+
   EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return const_cast<Scalar*>(m_impl.data()); }
 
   EIGEN_DEVICE_FUNC const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
@@ -155,6 +326,8 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
  protected:
   TensorEvaluator<ArgType, Device> m_impl;
   NewDimensions m_dimensions;
+  DSizes<Index, NumOutputDims> m_outputStrides;
+  DSizes<Index, NumInputDims> m_inputStrides;
 };
 
 
@@ -322,17 +495,27 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   typedef TensorSlicingOp<StartIndices, Sizes, ArgType> XprType;
   static const int NumDims = internal::array_size<Sizes>::value;
 
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Sizes Dimensions;
+
   enum {
     // Alignment can't be guaranteed at compile time since it depends on the
     // slice offsets and sizes.
-    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
+    IsAligned    = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,
+    RawAccess    = false
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using TensorBlock =  internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices())
   {
@@ -340,6 +523,16 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
       eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]);
     }
 
+    m_is_identity = true;
+    for (int i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+      eigen_assert(m_impl.dimensions()[i] >=
+                   op.sizes()[i] + op.startIndices()[i]);
+      if (m_impl.dimensions()[i] != op.sizes()[i] ||
+          op.startIndices()[i] != 0) {
+        m_is_identity = false;
+      }
+    }
+
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
     const Sizes& output_dims = op.sizes();
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
@@ -367,13 +560,10 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
         m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
       }
     }
-  }
 
-  typedef typename XprType::Index Index;
-  typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::CoeffReturnType CoeffReturnType;
-  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  typedef Sizes Dimensions;
+    m_block_total_size_max =
+        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
+  }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
@@ -417,7 +607,11 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
   {
-    return m_impl.coeff(srcCoeff(index));
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
   }
 
   template<int LoadMode>
@@ -427,6 +621,10 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
     EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
     eigen_assert(index+packetSize-1 < internal::array_prod(dimensions()));
 
+    if (m_is_identity) {
+      return m_impl.template packet<LoadMode>(index);
+    }
+
     Index inputIndices[] = {0, 0};
     Index indices[] = {index, index + packetSize - 1};
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
@@ -469,9 +667,26 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
-    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, NumDims);
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, m_is_identity ? 1 : NumDims);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kSkewedInnerDims,
+        m_block_total_size_max));
+    m_impl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      TensorBlock* output_block) const {
+    TensorBlock input_block(srcCoeff(output_block->first_coeff_index()),
+                            output_block->block_sizes(),
+                            output_block->block_strides(),
+                            Dimensions(m_inputStrides),
+                            output_block->data());
+    m_impl.block(&input_block);
+  }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
     Scalar* result = m_impl.data();
@@ -544,7 +759,9 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   TensorEvaluator<ArgType, Device> m_impl;
   const Device& m_device;
   Dimensions m_dimensions;
+  bool m_is_identity;
   const StartIndices m_offsets;
+  Index m_block_total_size_max;
 };
 
 
@@ -557,33 +774,46 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
   typedef TensorSlicingOp<StartIndices, Sizes, ArgType> XprType;
   static const int NumDims = internal::array_size<Sizes>::value;
 
-  enum {
-    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,
-    RawAccess = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
-  };
-
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-    : Base(op, device)
-    { }
-
   typedef typename XprType::Index Index;
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef Sizes Dimensions;
 
+  enum {
+    IsAligned    = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,
+    RawAccess    = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
+  };
+
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using TensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+    { }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
   {
-    return this->m_impl.coeffRef(this->srcCoeff(index));
+    if (this->m_is_identity) {
+      return this->m_impl.coeffRef(index);
+    } else {
+      return this->m_impl.coeffRef(this->srcCoeff(index));
+    }
   }
 
   template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketReturnType& x)
   {
+    if (this->m_is_identity) {
+      this->m_impl.template writePacket<StoreMode>(index, x);
+      return;
+    }
+
     const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
     Index inputIndices[] = {0, 0};
     Index indices[] = {index, index + packetSize - 1};
@@ -623,6 +853,14 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
       }
     }
   }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlock& block) {
+    this->m_impl.writeBlock(TensorBlock(
+        this->srcCoeff(block.first_coeff_index()), block.block_sizes(),
+        block.block_strides(), Dimensions(this->m_inputStrides),
+        const_cast<ScalarNoConst*>(block.data())));
+  }
 };
 
 
@@ -739,7 +977,13 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
   {
     // Handle degenerate intervals by gracefully clamping and allowing m_dimensions to be zero
     DSizes<Index,NumDims> startIndicesClamped, stopIndicesClamped;
+    m_is_identity = true;
     for (size_t i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+      if (m_strides[i] != 1 || op.startIndices()[i] != 0 ||
+          op.stopIndices()[i] != (m_impl.dimensions()[i] - 1)) {
+        m_is_identity = false;
+      }
+
       eigen_assert(m_strides[i] != 0 && "0 stride is invalid");
       if(m_strides[i]>0){
         startIndicesClamped[i] = clamp(op.startIndices()[i], 0, m_impl.dimensions()[i]);
@@ -822,11 +1066,15 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
   {
-    return m_impl.coeff(srcCoeff(index));
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
-    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, NumDims);
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, m_is_identity ? 1 : NumDims);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
@@ -873,6 +1121,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
   array<Index, NumDims> m_outputStrides;
   array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
   array<Index, NumDims> m_inputStrides;
+  bool m_is_identity;
   TensorEvaluator<ArgType, Device> m_impl;
   const Device& m_device;
   DSizes<Index, NumDims> m_startIndices; // clamped startIndices
@@ -916,7 +1165,11 @@ struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Stride
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
   {
-    return this->m_impl.coeffRef(this->srcCoeff(index));
+    if (this->m_is_identity) {
+      return this->m_impl.coeffRef(index);
+    } else {
+      return this->m_impl.coeffRef(this->srcCoeff(index));
+    }
   }
 };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 375fc0802..05c0990dc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -356,6 +356,70 @@ template <int NPT, typename S, typename R, typename I>
 __global__ void OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
 #endif
 
+template <typename Self, typename Op,
+          bool Vectorizable =
+              (Self::InputPacketAccess & Self::ReducerTraits::PacketAccess)>
+class BlockReducer {
+ public:
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::CoeffReturnType CoeffReturnType;
+  typedef typename Self::PacketReturnType PacketReturnType;
+  explicit BlockReducer(const Op& reducer) : op_(reducer) {
+    accum_ = op_.initialize();
+  }
+  void Reduce(Index index, Index num_values_to_reduce, Scalar* data) {
+    for (Index i = 0; i < num_values_to_reduce; ++i) {
+      op_.reduce(data[index + i], &accum_);
+    }
+  }
+  CoeffReturnType Finalize() { return op_.finalize(accum_); }
+  PacketReturnType FinalizePacket() {
+    // TODO(andydavis) This function should not be called for Scalar
+    // reductions: clean this up or add an assert here.
+    return PacketReturnType();
+  }
+
+ private:
+  CoeffReturnType accum_;
+  Op op_;
+};
+
+template <typename Self, typename Op>
+class BlockReducer<Self, Op, true> {
+ public:
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::CoeffReturnType CoeffReturnType;
+  typedef typename Self::PacketReturnType PacketReturnType;
+  static const Index PacketSize =
+      internal::unpacket_traits<PacketReturnType>::size;
+
+  explicit BlockReducer(const Op& reducer) : op_(reducer) {
+    vaccum_ = op_.template initializePacket<PacketReturnType>();
+    accum_ = op_.initialize();
+  }
+  void Reduce(Index index, Index num_values_to_reduce, Scalar* data) {
+    const Index vectorized_size =
+        (num_values_to_reduce / PacketSize) * PacketSize;
+    for (Index i = 0; i < vectorized_size; i += PacketSize) {
+      op_.reducePacket(
+          internal::ploadt<PacketReturnType, Unaligned>(&data[index + i]),
+          &vaccum_);
+    }
+    for (Index i = vectorized_size; i < num_values_to_reduce; ++i) {
+      op_.reduce(data[index + i], &accum_);
+    }
+  }
+  CoeffReturnType Finalize() { return op_.finalizeBoth(accum_, vaccum_); }
+  PacketReturnType FinalizePacket() { return op_.finalizePacket(vaccum_); }
+
+ private:
+  PacketReturnType vaccum_;
+  CoeffReturnType accum_;
+  Op op_;
+};
+
 }  // end namespace internal
 
 
@@ -394,6 +458,7 @@ class TensorReductionOp : public TensorBase<TensorReductionOp<Op, Dims, XprType,
 template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>
 struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>
 {
+  typedef internal::reducer_traits<Op, Device> ReducerTraits;
   typedef TensorReductionOp<Op, Dims, ArgType, MakePointer_> XprType;
   typedef typename XprType::Index Index;
   typedef ArgType ChildType;
@@ -410,14 +475,19 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
 
   enum {
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = Self::InputPacketAccess && Op::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,  // to be implemented
+    RawAccess    = false
   };
 
+  using ScalarNoConst =  typename internal::remove_const<Scalar>::type;
+
+  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>;
+  using InputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
+
   static const bool ReducingInnerMostDims = internal::are_inner_most_dims<Dims, NumInputDims, Layout>::value;
   static const bool PreservingInnerMostDims = internal::preserve_inner_most_dims<Dims, NumInputDims, Layout>::value;
   static const bool RunningFullReduction = (NumOutputDims==0);
@@ -451,11 +521,13 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         m_outputStrides[0] = 1;
         for (int i = 1; i < NumOutputDims; ++i) {
           m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+          m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
         }
       } else {
-        m_outputStrides.back() = 1;
+        m_outputStrides[NumOutputDims - 1] = 1;
         for (int i = NumOutputDims - 2; i >= 0; --i) {
           m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+          m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
         }
       }
     }
@@ -483,6 +555,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
           ++reduceIndex;
         } else {
           m_preservedStrides[outputIndex] = input_strides[i];
+          m_output_to_input_dim_map[outputIndex] = i;
           ++outputIndex;
         }
       }
@@ -492,6 +565,16 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     if (NumOutputDims == 0) {
       m_preservedStrides[0] = internal::array_prod(input_dims);
     }
+
+    m_numValuesToReduce =
+        NumOutputDims == 0
+            ? internal::array_prod(input_dims)
+            : (static_cast<int>(Layout) == static_cast<int>(ColMajor))
+                  ? m_preservedStrides[0]
+                  : m_preservedStrides[NumOutputDims - 1];
+
+    m_block_total_size_max =
+        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -686,6 +769,265 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     }
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kSkewedInnerDims,
+        m_block_total_size_max));
+    m_impl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void block(
+      OutputTensorBlock* output_block) const {
+    // Special case full reductions to avoid input block copy below.
+    if (NumInputDims == NumReducedDims) {
+      eigen_assert(output_block->first_coeff_index() == 0);
+      eigen_assert(output_block->block_sizes().TotalSize() == 1);
+      Op reducer(m_reducer);
+      output_block->data()[0] = internal::InnerMostDimReducer<Self, Op>::reduce(
+          *this, 0, m_numValuesToReduce, reducer);
+      return;
+    }
+
+    // Calculate input tensor 'slice' required to reduce output block coeffs.
+    DSizes<Index, NumInputDims> input_slice_sizes(m_impl.dimensions());
+    for (int i = 0; i < NumOutputDims; ++i) {
+      // Clip preserved input dimensions by output block size.
+      input_slice_sizes[m_output_to_input_dim_map[i]] =
+          output_block->block_sizes()[i];
+    }
+
+    // Shard input tensor slice into blocks (because it could be large if we
+    // need to reduce along several dimensions to calculate required output
+    // coefficients).
+    const Index max_coeff_count =
+        numext::mini<Index>(((m_device.firstLevelCacheSize()) / sizeof(Scalar)),
+                            input_slice_sizes.TotalSize());
+
+    // Calculate max output shard size needed to keep working set of reducers
+    // in L1, while leaving enough space for reducer overhead and 'PacketSize'
+    // reductions.
+    DSizes<Index, NumInputDims> target_input_block_sizes;
+    CalculateTargetInputBlockShape(max_coeff_count, input_slice_sizes,
+                                   &target_input_block_sizes);
+    // Calculate indices for first preserved dimension.
+    const Index first_preserved_dim_output_index =
+        static_cast<int>(Layout) == static_cast<int>(ColMajor)
+            ? 0
+            : NumOutputDims - 1;
+    const Index first_preserved_dim_input_index =
+        m_output_to_input_dim_map[first_preserved_dim_output_index];
+    const bool inner_most_dim_preserved =
+        first_preserved_dim_input_index ==
+            (static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                 ? 0
+                 : NumInputDims - 1) |
+        PreservingInnerMostDims;
+
+    // Calculate output block inner/outer dimension sizes.
+    const Index output_block_inner_dim_size =
+        output_block->block_sizes()[first_preserved_dim_output_index];
+    const Index output_block_outer_dim_size =
+        output_block->block_sizes().TotalSize() / output_block_inner_dim_size;
+    // Calculate shard size for first preserved dimension.
+    const Index output_shard_size =
+        target_input_block_sizes[first_preserved_dim_input_index];
+    const Index num_output_shards =
+        (output_block_inner_dim_size + output_shard_size - 1) /
+        output_shard_size;
+
+    // Initialize 'tensor_slice_offsets' from input coords of output index.
+    DSizes<Index, NumInputDims> tensor_slice_offsets;
+    GetInputCoordsForOutputIndex(output_block->first_coeff_index(),
+                                 &tensor_slice_offsets);
+
+    // Store tensor slice offset in first preserved dimension to be used
+    // to update tensor slice extents in loop below.
+    const Index first_preserved_dim_offset_start =
+        tensor_slice_offsets[first_preserved_dim_input_index];
+
+    array<BlockIteratorState, NumOutputDims> block_iter_state;
+
+    // Initialize state used to iterate through output coefficients
+    // and update 'tensor_slice_offsets' in outer preserved dims.
+    for (int i = 0; i < NumOutputDims - 1; ++i) {
+      const int dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                          ? i + 1
+                          : NumOutputDims - i - 2;
+      block_iter_state[i].input_dim = m_output_to_input_dim_map[dim];
+      block_iter_state[i].output_size = output_block->block_sizes()[dim];
+      block_iter_state[i].output_count = 0;
+    }
+
+    // Allocate input block memory.
+    ScalarNoConst* input_block_data = static_cast<ScalarNoConst*>(
+        m_device.allocate(max_coeff_count * sizeof(Scalar)));
+    // Allocate reducer memory.
+    const bool packet_reductions_enabled =
+        (Self::InputPacketAccess & Self::ReducerTraits::PacketAccess);
+    const Index num_reducers =
+        (inner_most_dim_preserved && packet_reductions_enabled)
+            ? (output_shard_size / PacketSize + output_shard_size % PacketSize +
+               PacketSize)
+            : output_shard_size;
+    typedef internal::BlockReducer<Self, Op> BlockReducer;
+    BlockReducer* reducers = static_cast<BlockReducer*>(
+        m_device.allocate(num_reducers * sizeof(BlockReducer)));
+
+    InputDimensions input_tensor_dims(m_impl.dimensions());
+    for (Index output_outer_index = 0;
+         output_outer_index < output_block_outer_dim_size;
+         ++output_outer_index) {
+      for (Index output_shard_index = 0; output_shard_index < num_output_shards;
+           ++output_shard_index) {
+        // Initialize 'tensor_slice_extents' for this output shard.
+        DSizes<Index, NumInputDims> tensor_slice_extents(input_slice_sizes);
+        for (int i = 0; i < NumInputDims; ++i) {
+          if (i == first_preserved_dim_input_index) {
+            // Clip first preserved dim size to output shard size.
+            tensor_slice_extents[i] = numext::mini(
+                output_shard_size,
+                input_slice_sizes[i] - (tensor_slice_offsets[i] -
+                                        first_preserved_dim_offset_start));
+
+          } else if (!m_reduced[i]) {
+            // Clip outer preserved dims to size 1, so that we reduce a
+            // contiguous set of output coefficients.
+            tensor_slice_extents[i] = 1;
+          }
+        }
+
+        // Intialize output coefficient reducers.
+        for (int i = 0; i < num_reducers; ++i) {
+          new (&reducers[i]) BlockReducer(m_reducer);
+        }
+
+        using TensorSliceBlockMapper =
+            internal::TensorSliceBlockMapper<ScalarNoConst, Index, NumInputDims,
+                                             Layout>;
+
+        // TODO(andydavis) Consider removing 'input_block_stride_order' if we
+        // find that scattered reads are not worth supporting in
+        // TensorSliceBlockMapper.
+        TensorSliceBlockMapper block_mapper(
+            input_tensor_dims, tensor_slice_offsets, tensor_slice_extents,
+            target_input_block_sizes, DimensionList<Index, NumInputDims>());
+
+        const Index num_outputs_to_update =
+            tensor_slice_extents[first_preserved_dim_input_index];
+        const Index preserved_dim_vector_reducer_count =
+            (inner_most_dim_preserved && packet_reductions_enabled)
+                ? num_outputs_to_update / PacketSize
+                : 0;
+        const Index preserved_dim_vector_coeff_count =
+            inner_most_dim_preserved
+                ? preserved_dim_vector_reducer_count * PacketSize
+                : 0;
+        const Index preserved_dim_reducer_limit =
+            (inner_most_dim_preserved && packet_reductions_enabled)
+                ? (preserved_dim_vector_reducer_count +
+                   num_outputs_to_update % PacketSize)
+                : num_outputs_to_update;
+
+        const Index total_block_count = block_mapper.total_block_count();
+        for (Index b = 0; b < total_block_count; ++b) {
+          InputTensorBlock input_block =
+              block_mapper.GetBlockForIndex(b, input_block_data);
+          // Read.
+          m_impl.block(&input_block);
+
+          Index num_values_to_reduce = 1;
+          for (Index i = 0; i < NumInputDims; ++i) {
+            if (m_reduced[i]) {
+              num_values_to_reduce *= input_block.block_sizes()[i];
+            }
+          }
+          // Reduce.
+          if (inner_most_dim_preserved) {
+            const Index input_outer_dim_size =
+                input_block.block_sizes().TotalSize() / num_outputs_to_update;
+            for (Index input_outer_dim_index = 0;
+                 input_outer_dim_index < input_outer_dim_size;
+                 ++input_outer_dim_index) {
+              const Index input_outer_dim_base =
+                  input_outer_dim_index * num_outputs_to_update;
+              for (Index i = 0; i < preserved_dim_vector_reducer_count; ++i) {
+                reducers[i].Reduce(input_outer_dim_base + i * PacketSize,
+                                   PacketSize, input_block.data());
+              }
+              const Index scalar_reducer_base =
+                  input_outer_dim_base + preserved_dim_vector_coeff_count;
+              for (Index i = preserved_dim_vector_reducer_count;
+                   i < preserved_dim_reducer_limit; ++i) {
+                reducers[i].Reduce(scalar_reducer_base + i -
+                                       preserved_dim_vector_reducer_count,
+                                   1, input_block.data());
+              }
+            }
+          } else {
+            for (Index i = 0; i < num_outputs_to_update; ++i) {
+              reducers[i].Reduce(i * num_values_to_reduce, num_values_to_reduce,
+                                 input_block.data());
+            }
+          }
+        }
+
+        // Finalize all reducers for this output shard.
+        const Index output_base_index =
+            output_outer_index * output_block_inner_dim_size +
+            output_shard_index * output_shard_size;
+        if (inner_most_dim_preserved) {
+          EIGEN_ALIGN_MAX
+              typename internal::remove_const<CoeffReturnType>::type
+                  values[PacketSize];
+          for (Index i = 0; i < preserved_dim_vector_reducer_count; ++i) {
+            const Index reducer_base = output_base_index + i * PacketSize;
+            internal::pstore<CoeffReturnType, PacketReturnType>(
+                values, reducers[i].FinalizePacket());
+            for (Index j = 0; j < PacketSize; ++j) {
+              output_block->data()[reducer_base + j] = values[j];
+            }
+          }
+          const Index scalar_reducer_base =
+              output_base_index + preserved_dim_vector_coeff_count;
+
+          for (Index i = preserved_dim_vector_reducer_count;
+               i < preserved_dim_reducer_limit; ++i) {
+            output_block->data()[scalar_reducer_base + i -
+                                 preserved_dim_vector_reducer_count] =
+                reducers[i].Finalize();
+          }
+        } else {
+          for (int i = 0; i < num_outputs_to_update; ++i) {
+            output_block->data()[output_base_index + i] =
+                reducers[i].Finalize();
+          }
+        }
+
+        // Update 'tensor_slice_offsets' by num outputs for this output shard.
+        tensor_slice_offsets[first_preserved_dim_input_index] +=
+            num_outputs_to_update;
+      }
+      // Update slice offset for inner preserved dim.
+      tensor_slice_offsets[first_preserved_dim_input_index] -=
+          output_block_inner_dim_size;
+      // Update slice offsets for remaining output dims.
+      for (int i = 0; i < NumOutputDims - 1; ++i) {
+        BlockIteratorState& b = block_iter_state[i];
+        if (++b.output_count < b.output_size) {
+          ++tensor_slice_offsets[b.input_dim];
+          break;
+        }
+        b.output_count = 0;
+        tensor_slice_offsets[b.input_dim] -= b.output_size - 1;
+      }
+    }
+
+    // Free memory.
+    m_device.deallocate(input_block_data);
+    m_device.deallocate(reducers);
+  }
+
   EIGEN_DEVICE_FUNC typename MakePointer_<CoeffReturnType>::Type data() const { return m_result; }
 
 #if defined(EIGEN_USE_SYCL)
@@ -722,6 +1064,12 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   template <typename S, typename O, typename D> friend struct internal::InnerReducer;
 
+  struct BlockIteratorState {
+    Index input_dim;
+    Index output_size;
+    Index output_count;
+  };
+
   // Returns the Index in the input tensor of the first value that needs to be
   // used to compute the reduction at output index "index".
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
@@ -764,16 +1112,90 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     return startInput;
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void GetInputCoordsForOutputIndex(
+      Index index,
+      DSizes<Index, NumInputDims>* coords) const {
+    for (int i = 0; i < NumInputDims; ++i) {
+      (*coords)[i] = 0;
+    }
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumOutputDims - 1; i > 0; --i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        (*coords)[m_output_to_input_dim_map[i]] = idx;
+        index -= idx * m_outputStrides[i];
+      }
+      (*coords)[m_output_to_input_dim_map[0]] = index;
+    } else {
+      for (int i = 0; i < NumOutputDims - 1; ++i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        (*coords)[m_output_to_input_dim_map[i]] = idx;
+        index -= idx * m_outputStrides[i];
+      }
+      (*coords)[m_output_to_input_dim_map[NumOutputDims-1]] = index;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void CalculateTargetInputBlockShape(
+      const Index max_coeff_count,
+      const DSizes<Index, NumInputDims>& input_slice_sizes,
+      DSizes<Index, NumInputDims>* target_input_block_sizes) const {
+    typedef typename internal::packet_traits<Scalar>::type Packet;
+    typedef internal::BlockReducer<Self, Op> BlockReducer;
+    // TODO(andydavis) Compute reducer overhead correctly for the case where
+    // we are preserving the inner most dimension, and a single reducer
+    // reduces a packet's worth of output coefficients.
+    const Index reducer_overhead = sizeof(BlockReducer) / sizeof(Scalar);
+
+    Index coeff_to_allocate = max_coeff_count;
+    bool first_preserved_dim_allocated = false;
+    bool first_reduced_dim_allocated = false;
+    for (int i = 0; i < NumInputDims; ++i) {
+      const int dim = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                          ? i
+                          : NumInputDims - i - 1;
+      (*target_input_block_sizes)[dim] = 1;
+      if (m_reduced[dim]) {
+        // TODO(andydavis) Consider allocating to multiple reduced dimensions.
+        // Watch out for cases where reduced dimensions are not contiguous,
+        // which induces scattered reads.
+        if (!first_reduced_dim_allocated) {
+          (*target_input_block_sizes)[dim] =
+              numext::mini(input_slice_sizes[dim], coeff_to_allocate);
+          coeff_to_allocate /= (*target_input_block_sizes)[dim];
+          first_reduced_dim_allocated = true;
+        }
+      } else if (!first_preserved_dim_allocated) {
+        // TODO(andydavis) Include output block size in this L1 working set
+        // calculation.
+        const Index allocated = max_coeff_count - coeff_to_allocate;
+        const Index alloc_size = numext::maxi(
+            static_cast<Index>(1), coeff_to_allocate / reducer_overhead);
+        (*target_input_block_sizes)[dim] =
+            numext::mini(input_slice_sizes[dim], alloc_size);
+        coeff_to_allocate = numext::maxi(
+            static_cast<Index>(1),
+            coeff_to_allocate /
+                ((*target_input_block_sizes)[dim] * reducer_overhead));
+        first_preserved_dim_allocated = true;
+      }
+    }
+  }
+
   // Bitmap indicating if an input dimension is reduced or not.
   array<bool, NumInputDims> m_reduced;
   // Dimensions of the output of the operation.
   Dimensions m_dimensions;
   // Precomputed strides for the output tensor.
   array<Index, NumOutputDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumOutputDims> m_fastOutputStrides;
   // Subset of strides of the input tensor for the non-reduced dimensions.
   // Indexed by output dimensions.
   static const int NumPreservedStrides = max_n_1<NumOutputDims>::size;
   array<Index, NumPreservedStrides> m_preservedStrides;
+  // Map from output to input dimension index.
+  array<Index, NumOutputDims> m_output_to_input_dim_map;
+  // How many values go into each reduction
+  Index m_numValuesToReduce;
 
   // Subset of strides of the input tensor for the reduced dimensions.
   // Indexed by reduced dimensions.
@@ -782,6 +1204,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   // Indexed by reduced dimensions.
   array<Index, NumReducedDims> m_reducedDims;
 
+  // Block size for tiled (aka TensorBlock) evaluation.
+  Index m_block_total_size_max;
+
   // Evaluator for the input expression.
   TensorEvaluator<ArgType, Device> m_impl;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 6b54f40ad..3add9dac0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -100,6 +100,7 @@ class TensorShufflingOp : public TensorBase<TensorShufflingOp<Shuffle, XprType>
 template<typename Shuffle, typename ArgType, typename Device>
 struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
 {
+  typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Self;
   typedef TensorShufflingOp<Shuffle, ArgType> XprType;
   typedef typename XprType::Index Index;
   static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
@@ -110,44 +111,60 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
 
   enum {
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess  = false,  // to be implemented
+    RawAccess    = false
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using TensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  using TensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_shuffle(op.shufflePermutation())
   {
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
     const Shuffle& shuffle = op.shufflePermutation();
+    m_is_identity = true;
     for (int i = 0; i < NumDims; ++i) {
       m_dimensions[i] = input_dims[shuffle[i]];
+      m_inverseShuffle[shuffle[i]] = i;
+      if (m_is_identity && shuffle[i] != i) {
+        m_is_identity = false;
+      }
     }
 
-    array<Index, NumDims> inputStrides;
-
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      inputStrides[0] = 1;
+      m_unshuffledInputStrides[0] = 1;
       m_outputStrides[0] = 1;
+
       for (int i = 1; i < NumDims; ++i) {
-        inputStrides[i] = inputStrides[i - 1] * input_dims[i - 1];
+        m_unshuffledInputStrides[i] =
+            m_unshuffledInputStrides[i - 1] * input_dims[i - 1];
         m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
       }
     } else {
-      inputStrides[NumDims - 1] = 1;
+      m_unshuffledInputStrides[NumDims - 1] = 1;
       m_outputStrides[NumDims - 1] = 1;
       for (int i = NumDims - 2; i >= 0; --i) {
-        inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
+        m_unshuffledInputStrides[i] =
+            m_unshuffledInputStrides[i + 1] * input_dims[i + 1];
         m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
       }
     }
 
     for (int i = 0; i < NumDims; ++i) {
-      m_inputStrides[i] = inputStrides[shuffle[i]];
+      m_inputStrides[i] = m_unshuffledInputStrides[shuffle[i]];
     }
+
+    m_block_total_size_max =
+        numext::maxi<Index>(1, device.firstLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -162,29 +179,151 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
   {
-    return m_impl.coeff(srcCoeff(index));
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
   }
 
+  template <int LoadMode, typename Self, bool ImplPacketAccess>
+  struct PacketLoader {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    static PacketReturnType Run(const Self& self, Index index) {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      for (int i = 0; i < PacketSize; ++i) {
+        values[i] = self.coeff(index + i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  };
+
+  template<int LoadMode, typename Self>
+  struct PacketLoader<LoadMode, Self, true> {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    static PacketReturnType Run(const Self& self, Index index) {
+      if (self.m_is_identity) {
+        return self.m_impl.template packet<LoadMode>(index);
+      } else {
+        EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+        for (int i = 0; i < PacketSize; ++i) {
+          values[i] = self.coeff(index + i);
+        }
+        PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+        return rslt;
+      }
+    }
+  };
+
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
-    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
-    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+    EIGEN_STATIC_ASSERT(PacketSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+        eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
+    return PacketLoader<LoadMode, Self, TensorEvaluator<ArgType, Device>::PacketAccess>::Run(*this, index);
+  }
 
-    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
-    for (int i = 0; i < PacketSize; ++i) {
-      values[i] = coeff(index+i);
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    resources->push_back(internal::TensorOpResourceRequirements(
+        internal::TensorBlockShapeType::kUniformAllDims,
+        m_block_total_size_max));
+    m_impl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      TensorBlock* output_block) const {
+    if (m_impl.data() != NULL) {
+      // Fast path: we have direct access to the data, so shuffle as we read.
+      TensorBlockReader::Run(output_block,
+                             srcCoeff(output_block->first_coeff_index()),
+                             m_inverseShuffle,
+                             m_unshuffledInputStrides,
+                             m_impl.data());
+      return;
+    }
+
+    // Slow path: read unshuffled block from the input and shuffle in-place.
+    // Initialize input block sizes using input-to-output shuffle map.
+    DSizes<Index, NumDims> input_block_sizes;
+    for (Index i = 0; i < NumDims; ++i) {
+      input_block_sizes[i] = output_block->block_sizes()[m_inverseShuffle[i]];
+    }
+
+    // Calculate input block strides.
+    DSizes<Index, NumDims> input_block_strides;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      input_block_strides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        input_block_strides[i] =
+            input_block_strides[i - 1] * input_block_sizes[i - 1];
+      }
+    } else {
+      input_block_strides[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        input_block_strides[i] =
+            input_block_strides[i + 1] * input_block_sizes[i + 1];
+      }
+    }
+
+    // Read input block.
+    TensorBlock input_block(srcCoeff(output_block->first_coeff_index()),
+                            input_block_sizes,
+                            input_block_strides,
+                            Dimensions(m_unshuffledInputStrides),
+                            output_block->data());
+
+    m_impl.block(&input_block);
+
+    // Naive In-place shuffle: random IO but block size is O(L1 cache size).
+    // TODO(andydavis) Improve the performance of this in-place shuffle.
+    const Index total_size = input_block_sizes.TotalSize();
+    std::vector<bool> bitmap(total_size, false);
+    ScalarNoConst* data = const_cast<ScalarNoConst*>(output_block->data());
+    const DSizes<Index, NumDims>& output_block_strides =
+        output_block->block_strides();
+    for (Index input_index = 0; input_index < total_size; ++input_index) {
+      if (bitmap[input_index]) {
+        // Coefficient at this index has already been shuffled.
+        continue;
+      }
+
+      Index output_index = GetBlockOutputIndex(input_index, input_block_strides,
+                                               output_block_strides);
+      if (output_index == input_index) {
+        // Coefficient already in place.
+        bitmap[output_index] = true;
+        continue;
+      }
+
+      // The following loop starts at 'input_index', and shuffles
+      // coefficients into their shuffled location at 'output_index'.
+      // It skips through the array shuffling coefficients by following
+      // the shuffle cycle starting and ending a 'start_index'.
+      ScalarNoConst evicted_value;
+      ScalarNoConst shuffled_value = data[input_index];
+      do {
+        evicted_value = data[output_index];
+        data[output_index] = shuffled_value;
+        shuffled_value = evicted_value;
+        bitmap[output_index] = true;
+        output_index = GetBlockOutputIndex(output_index, input_block_strides,
+                                           output_block_strides);
+      } while (output_index != input_index);
+
+      data[output_index] = shuffled_value;
+      bitmap[output_index] = true;
     }
-    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
-    return rslt;
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
-    const double compute_cost = NumDims * (2 * TensorOpCost::AddCost<Index>() +
+    const double compute_cost = m_is_identity ? TensorOpCost::AddCost<Index>() :
+                                NumDims * (2 * TensorOpCost::AddCost<Index>() +
                                            2 * TensorOpCost::MulCost<Index>() +
                                            TensorOpCost::DivCost<Index>());
     return m_impl.costPerCoeff(vectorized) +
-           TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize);
+           TensorOpCost(0, 0, compute_cost, m_is_identity /* vectorized */, PacketSize);
   }
 
   EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
@@ -195,27 +334,57 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const {return m_impl;}
 
  protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index GetBlockOutputIndex(
+      Index input_index,
+      const DSizes<Index, NumDims>& input_block_strides,
+      const DSizes<Index, NumDims>& output_block_strides) const {
+    Index output_index = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = input_index / input_block_strides[i];
+        output_index += idx * output_block_strides[m_inverseShuffle[i]];
+        input_index -= idx * input_block_strides[i];
+      }
+      return output_index + input_index *
+          output_block_strides[m_inverseShuffle[0]];
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = input_index / input_block_strides[i];
+        output_index += idx * output_block_strides[m_inverseShuffle[i]];
+        input_index -= idx * input_block_strides[i];
+      }
+      return output_index + input_index *
+          output_block_strides[m_inverseShuffle[NumDims - 1]];
+    }
+  }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const {
     Index inputIndex = 0;
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
       for (int i = NumDims - 1; i > 0; --i) {
-        const Index idx = index / m_outputStrides[i];
+        const Index idx = index / m_fastOutputStrides[i];
         inputIndex += idx * m_inputStrides[i];
         index -= idx * m_outputStrides[i];
       }
       return inputIndex + index * m_inputStrides[0];
     } else {
       for (int i = 0; i < NumDims - 1; ++i) {
-        const Index idx = index / m_outputStrides[i];
+        const Index idx = index / m_fastOutputStrides[i];
         inputIndex += idx * m_inputStrides[i];
         index -= idx * m_outputStrides[i];
       }
       return inputIndex + index * m_inputStrides[NumDims - 1];
     }
   }
+
   Dimensions m_dimensions;
+  bool m_is_identity;
+  array<Index, NumDims> m_inverseShuffle;
   array<Index, NumDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
   array<Index, NumDims> m_inputStrides;
+  array<Index, NumDims> m_unshuffledInputStrides;
+  Index m_block_total_size_max;
   TensorEvaluator<ArgType, Device> m_impl;
   /// required by sycl
   Shuffle m_shuffle;
@@ -239,12 +408,18 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
 
   enum {
-    IsAligned = false,
+    IsAligned    = false,
     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
-    BlockAccess = false,
-    RawAccess = false
+    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
+    Layout       = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess    = false
   };
 
+  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+
+  using TensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  using TensorBlockWriter = internal::TensorBlockWriter<ScalarNoConst, Index, NumDims, Layout>;
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : Base(op, device)
   { }
@@ -265,6 +440,14 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
       this->coeffRef(index+i) = values[i];
     }
   }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlock& block) {
+    eigen_assert(this->m_impl.data() != NULL);
+    TensorBlockWriter::Run(block, this->srcCoeff(block.first_coeff_index()),
+                           this->m_inverseShuffle,
+                           this->m_unshuffledInputStrides, this->m_impl.data());
+  }
 };
 
 
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 274f901ce..448f47f1d 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -18,22 +18,30 @@ using Eigen::RowMajor;
 using Eigen::ColMajor;
 
 // A set of tests to verify that different TensorExecutor strategies yields the
-// same results for all the ops, supporting tiled execution.
+// same results for all the ops, supporting tiled evaluation.
+
+template <int NumDims>
+static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
+  array<Index, NumDims> dims;
+  for (int i = 0; i < NumDims; ++i) {
+    dims[i] = internal::random<int>(min_dim, max_dim);
+  }
+  return dims;
+};
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_binary_expr(Device d)
+{
+  static constexpr int Options = 0 | Layout;
 
-template <typename Device, bool Vectorizable, bool Tileable, int Layout>
-static void test_execute_binary_expr(Device d) {
   // Pick a large enough tensor size to bypass small tensor block evaluation
   // optimization.
-  int d0 = internal::random<int>(100, 200);
-  int d1 = internal::random<int>(100, 200);
-  int d2 = internal::random<int>(100, 200);
+  auto dims = RandomDims<NumDims>(50 / NumDims, 100 / NumDims);
 
-  static constexpr int Options = 0;
-  using IndexType = int;
-
-  Tensor<float, 3, Options, IndexType> lhs(d0, d1, d2);
-  Tensor<float, 3, Options, IndexType> rhs(d0, d1, d2);
-  Tensor<float, 3, Options, IndexType> dst(d0, d1, d2);
+  Tensor<T, NumDims, Options, Index> lhs(dims);
+  Tensor<T, NumDims, Options, Index> rhs(dims);
+  Tensor<T, NumDims, Options, Index> dst(dims);
 
   lhs.setRandom();
   rhs.setRandom();
@@ -46,33 +54,389 @@ static void test_execute_binary_expr(Device d) {
 
   Executor::run(Assign(dst, expr), d);
 
-  for (int i = 0; i < d0; ++i) {
-    for (int j = 0; j < d1; ++j) {
-      for (int k = 0; k < d2; ++k) {
-        float sum = lhs(i, j, k) + rhs(i, j, k);
-        VERIFY_IS_EQUAL(sum, dst(i, j, k));
-      }
-    }
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    T sum = lhs.coeff(i) + rhs.coeff(i);
+    VERIFY_IS_EQUAL(sum, dst.coeff(i));
   }
 }
 
-#define CALL_SUBTEST_COMBINATIONS(NAME)                                        \
-  CALL_SUBTEST((NAME<DefaultDevice, false, false, ColMajor>(default_device))); \
-  CALL_SUBTEST((NAME<DefaultDevice, false, true, ColMajor>(default_device)));  \
-  CALL_SUBTEST((NAME<DefaultDevice, true, false, ColMajor>(default_device)));  \
-  CALL_SUBTEST((NAME<DefaultDevice, true, true, ColMajor>(default_device)));   \
-  CALL_SUBTEST((NAME<DefaultDevice, false, false, RowMajor>(default_device))); \
-  CALL_SUBTEST((NAME<DefaultDevice, false, true, RowMajor>(default_device)));  \
-  CALL_SUBTEST((NAME<DefaultDevice, true, false, RowMajor>(default_device)));  \
-  CALL_SUBTEST((NAME<DefaultDevice, true, true, RowMajor>(default_device)));   \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, false, false, ColMajor>(tp_device)));   \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, false, true, ColMajor>(tp_device)));    \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, true, false, ColMajor>(tp_device)));    \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, true, true, ColMajor>(tp_device)));     \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, false, false, RowMajor>(tp_device)));   \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, false, true, RowMajor>(tp_device)));    \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, true, false, RowMajor>(tp_device)));    \
-  CALL_SUBTEST((NAME<ThreadPoolDevice, true, true, RowMajor>(tp_device)))
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_broadcasting(Device d)
+{
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(1, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  const auto broadcasts = RandomDims<NumDims>(1, 7);
+  const auto expr = src.broadcast(broadcasts);
+
+  // We assume that broadcasting on a default device is tested and correct, so
+  // we can rely on it to verify correctness of tensor executor and tiling.
+  Tensor<T, NumDims, Options, Index> golden;
+  golden = expr;
+
+  // Now do the broadcasting using configured tensor executor.
+  Tensor<T, NumDims, Options, Index> dst(golden.dimensions());
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+};
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_chipping_rvalue(Device d) {
+  auto dims = RandomDims<NumDims>(1, 10);
+  Tensor<T, NumDims, Layout, Index> src(dims);
+  src.setRandom();
+
+#define TEST_CHIPPING(CHIP_DIM)                                           \
+  if (NumDims > (CHIP_DIM)) {                                             \
+    const auto offset = internal::random<Index>(0, dims[(CHIP_DIM)] - 1); \
+    const auto expr = src.template chip<(CHIP_DIM)>(offset);              \
+                                                                          \
+    Tensor<T, NumDims - 1, Layout, Index> golden;                         \
+    golden = expr;                                                        \
+                                                                          \
+    Tensor<T, NumDims - 1, Layout, Index> dst(golden.dimensions());       \
+                                                                          \
+    using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;   \
+    using Executor = internal::TensorExecutor<const Assign, Device,       \
+                                              Vectorizable, Tileable>;    \
+                                                                          \
+    Executor::run(Assign(dst, expr), d);                                  \
+                                                                          \
+    for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {            \
+      VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));                     \
+    }                                                                     \
+  }
+
+  TEST_CHIPPING(0)
+  TEST_CHIPPING(1)
+  TEST_CHIPPING(2)
+  TEST_CHIPPING(3)
+  TEST_CHIPPING(4)
+  TEST_CHIPPING(5)
+
+#undef TEST_CHIPPING
+};
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+    bool Tileable, int Layout>
+static void test_execute_chipping_lvalue(Device d) {
+  auto dims = RandomDims<NumDims>(1, 10);
+
+#define TEST_CHIPPING(CHIP_DIM)                                             \
+  if (NumDims > (CHIP_DIM)) {                                               \
+    /* Generate random data that we'll assign to the chipped tensor dim. */ \
+    array<Index, NumDims - 1> src_dims;                                     \
+    for (int i = 0; i < NumDims - 1; ++i) {                                 \
+      int dim = i < (CHIP_DIM) ? i : i + 1;                                 \
+      src_dims[i] = dims[dim];                                              \
+    }                                                                       \
+                                                                            \
+    Tensor<T, NumDims - 1, Layout, Index> src(src_dims);                    \
+    src.setRandom();                                                        \
+                                                                            \
+    const auto offset = internal::random<Index>(0, dims[(CHIP_DIM)] - 1);   \
+                                                                            \
+    /* Generate random data to fill non-chipped dimensions*/                \
+    Tensor<T, NumDims, Layout, Index> random(dims);                         \
+    random.setRandom();                                                     \
+                                                                            \
+    Tensor<T, NumDims, Layout, Index> golden(dims);                         \
+    golden = random;                                                        \
+    golden.template chip<(CHIP_DIM)>(offset) = src;                         \
+                                                                            \
+    Tensor<T, NumDims, Layout, Index> dst(dims);                            \
+    dst = random;                                                           \
+    auto expr = dst.template chip<(CHIP_DIM)>(offset);                      \
+                                                                            \
+    using Assign = TensorAssignOp<decltype(expr), const decltype(src)>;     \
+    using Executor = internal::TensorExecutor<const Assign, Device,         \
+                                              Vectorizable, Tileable>;      \
+                                                                            \
+    Executor::run(Assign(expr, src), d);                                    \
+                                                                            \
+    for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {              \
+      VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));                       \
+    }                                                                       \
+  }
+
+  TEST_CHIPPING(0)
+  TEST_CHIPPING(1)
+  TEST_CHIPPING(2)
+  TEST_CHIPPING(3)
+  TEST_CHIPPING(4)
+  TEST_CHIPPING(5)
+
+#undef TEST_CHIPPING
+};
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_shuffle_rvalue(Device d) {
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(1, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Create a random dimension re-ordering/shuffle.
+  std::vector<Index> shuffle;
+  for (int i = 0; i < NumDims; ++i) shuffle.push_back(i);
+  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
+
+  const auto expr = src.shuffle(shuffle);
+
+  // We assume that shuffling on a default device is tested and correct, so
+  // we can rely on it to verify correctness of tensor executor and tiling.
+  Tensor<T, NumDims, Options, Index> golden;
+  golden = expr;
+
+  // Now do the shuffling using configured tensor executor.
+  Tensor<T, NumDims, Options, Index> dst(golden.dimensions());
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_shuffle_lvalue(Device d) {
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(5, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Create a random dimension re-ordering/shuffle.
+  std::vector<Index> shuffle;
+  for (int i = 0; i < NumDims; ++i) shuffle.push_back(i);
+  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
+
+  array<Index, NumDims> shuffled_dims;
+  for (int i = 0; i < NumDims; ++i) shuffled_dims[shuffle[i]] = dims[i];
+
+  // We assume that shuffling on a default device is tested and correct, so
+  // we can rely on it to verify correctness of tensor executor and tiling.
+  Tensor<T, NumDims, Options, Index> golden(shuffled_dims);
+  golden.shuffle(shuffle) = src;
+
+  // Now do the shuffling using configured tensor executor.
+  Tensor<T, NumDims, Options, Index> dst(shuffled_dims);
+
+  auto expr = dst.shuffle(shuffle);
+
+  using Assign = TensorAssignOp<decltype(expr), const decltype(src)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(expr, src), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_reduction(Device d)
+{
+  static_assert(NumDims >= 2);
+
+  static constexpr int ReducedDims = NumDims - 2;
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(5, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Pick two random and unique reduction dimensions.
+  int reduction0 = internal::random<int>(0, NumDims - 1);
+  int reduction1 = internal::random<int>(0, NumDims - 1);
+  while (reduction0 == reduction1) {
+    reduction1 = internal::random<int>(0, NumDims - 1);
+  }
+
+  DSizes<Index, 2> reduction_axis;
+  reduction_axis[0] = reduction0;
+  reduction_axis[1] = reduction1;
+
+  Tensor<T, ReducedDims, Options, Index> golden = src.sum(reduction_axis);
+
+  // Now do the reduction using configured tensor executor.
+  Tensor<T, ReducedDims, Options, Index> dst(golden.dimensions());
+
+  auto expr = src.sum(reduction_axis);
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+    bool Tileable, int Layout>
+static void test_execute_reshape(Device d)
+{
+  static_assert(NumDims >= 2);
+
+  static constexpr int ReshapedDims = NumDims - 1;
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(5, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Multiple 0th dimension and then shuffle.
+  std::vector<Index> shuffle;
+  for (int i = 0; i < ReshapedDims; ++i) shuffle.push_back(i);
+  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
+
+  DSizes<Index, ReshapedDims> reshaped_dims;
+  reshaped_dims[shuffle[0]] = dims[0] * dims[1];
+  for (int i = 2; i < NumDims; ++i) reshaped_dims[shuffle[i]] = dims[i];
+
+  Tensor<T, ReshapedDims, Options, Index> golden = src.reshape(reshaped_dims);
+
+  // Now reshape using configured tensor executor.
+  Tensor<T, ReshapedDims, Options, Index> dst(golden.dimensions());
+
+  auto expr = src.reshape(reshaped_dims);
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_slice_rvalue(Device d)
+{
+  static_assert(NumDims >= 2);
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(5, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Pick a random slice of src tensor.
+  auto slice_start = DSizes<Index, NumDims>(RandomDims<NumDims>());
+  auto slice_size = DSizes<Index, NumDims>(RandomDims<NumDims>());
+
+  // Make sure that slice start + size do not overflow tensor dims.
+  for (int i = 0; i < NumDims; ++i) {
+    slice_start[i] = numext::mini(dims[i] - 1, slice_start[i]);
+    slice_size[i] = numext::mini(slice_size[i], dims[i] - slice_start[i]);
+  }
+
+  Tensor<T, NumDims, Options, Index> golden =
+      src.slice(slice_start, slice_size);
+
+  // Now reshape using configured tensor executor.
+  Tensor<T, NumDims, Options, Index> dst(golden.dimensions());
+
+  auto expr = src.slice(slice_start, slice_size);
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+    bool Tileable, int Layout>
+static void test_execute_slice_lvalue(Device d)
+{
+  static_assert(NumDims >= 2);
+  static constexpr int Options = 0 | Layout;
+
+  auto dims = RandomDims<NumDims>(5, 10);
+  Tensor<T, NumDims, Options, Index> src(dims);
+  src.setRandom();
+
+  // Pick a random slice of src tensor.
+  auto slice_start = DSizes<Index, NumDims>(RandomDims<NumDims>(1, 10));
+  auto slice_size = DSizes<Index, NumDims>(RandomDims<NumDims>(1, 10));
+
+  // Make sure that slice start + size do not overflow tensor dims.
+  for (int i = 0; i < NumDims; ++i) {
+    slice_start[i] = numext::mini(dims[i] - 1, slice_start[i]);
+    slice_size[i] = numext::mini(slice_size[i], dims[i] - slice_start[i]);
+  }
+
+  Tensor<T, NumDims, Options, Index> slice(slice_size);
+  slice.setRandom();
+
+  // Asign a slice using default executor.
+  Tensor<T, NumDims, Options, Index> golden = src;
+  golden.slice(slice_start, slice_size) = slice;
+
+  // And using configured execution strategy.
+  Tensor<T, NumDims, Options, Index> dst = src;
+  auto expr = dst.slice(slice_start, slice_size);
+
+  using Assign = TensorAssignOp<decltype(expr), const decltype(slice)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(expr, slice), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
+  }
+}
+
+#define CALL_SUBTEST_COMBINATIONS(NAME, T, NUM_DIMS)                                           \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, false, ColMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  ColMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, ColMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  ColMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, false, RowMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  RowMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, RowMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  RowMajor>(default_device))); \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  ColMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, ColMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  ColMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  RowMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, RowMajor>(tp_device)));      \
+  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  RowMajor>(tp_device)))
 
 EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
   Eigen::DefaultDevice default_device;
@@ -81,7 +445,49 @@ EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
   Eigen::ThreadPool tp(num_threads);
   Eigen::ThreadPoolDevice tp_device(&tp, num_threads);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr);
+  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 2);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 2);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 2);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 5);
+
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 2);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 5);
 }
 
 #undef CALL_SUBTEST_COMBINATIONS

From 9e219bb3d3f0f3a3157dcf8c2a27895e9f85035b Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 10:47:49 +0100
Subject: [PATCH 510/680] Converting ad-hoc inline keyword to
 EIGEN_STRONG_INLINE MACRO.

---
 unsupported/Eigen/CXX11/src/Tensor/Tensor.h   |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 14 ++--
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   | 14 ++--
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   |  4 +-
 .../Eigen/CXX11/src/Tensor/TensorIndexList.h  |  8 +-
 unsupported/Eigen/CXX11/src/util/CXX11Meta.h  | 74 +++++++++----------
 .../SpecialFunctionsArrayAPI.h                | 18 ++---
 .../SpecialFunctionsFunctors.h                | 32 ++++----
 8 files changed, 83 insertions(+), 83 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
index e3f6e37f0..aed71b265 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
@@ -112,7 +112,7 @@ class Tensor : public TensorBase<Tensor<Scalar_, NumIndices_, Options_, IndexTyp
 
 #if EIGEN_HAS_VARIADIC_TEMPLATES
     template<typename... IndexTypes>
-    EIGEN_DEVICE_FUNC inline const Scalar& coeff(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
     {
       // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
       EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 5ca47cca7..4f973a5b7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -41,7 +41,7 @@ template<typename Index, std::size_t NumIndices, std::size_t n, bool RowMajor>
 struct fixed_size_tensor_index_linearization_helper
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
-  static inline Index run(array<Index, NumIndices> const& indices,
+  static EIGEN_STRONG_INLINE Index run(array<Index, NumIndices> const& indices,
                           const Dimensions& dimensions)
   {
     return array_get<RowMajor ? n - 1 : (NumIndices - n)>(indices) +
@@ -54,7 +54,7 @@ template<typename Index, std::size_t NumIndices, bool RowMajor>
 struct fixed_size_tensor_index_linearization_helper<Index, NumIndices, 0, RowMajor>
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
-  static inline Index run(array<Index, NumIndices> const&, const Dimensions&)
+  static EIGEN_STRONG_INLINE Index run(array<Index, NumIndices> const&, const Dimensions&)
   {
     return 0;
   }
@@ -64,7 +64,7 @@ template<typename Index, std::size_t n>
 struct fixed_size_tensor_index_extraction_helper
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
-  static inline Index run(const Index index,
+  static EIGEN_STRONG_INLINE Index run(const Index index,
                           const Dimensions& dimensions)
   {
     const Index mult = (index == n-1) ? 1 : 0;
@@ -77,7 +77,7 @@ template<typename Index>
 struct fixed_size_tensor_index_extraction_helper<Index, 0>
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
-  static inline Index run(const Index,
+  static EIGEN_STRONG_INLINE Index run(const Index,
                           const Dimensions&)
   {
     return 0;
@@ -421,20 +421,20 @@ template <std::size_t n, std::size_t V1, std::size_t V2, std::size_t V3, std::si
 
 template <typename Dims1, typename Dims2, size_t n, size_t m>
 struct sizes_match_below_dim {
-  static EIGEN_DEVICE_FUNC  inline bool run(Dims1&, Dims2&) {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1&, Dims2&) {
     return false;
   }
 };
 template <typename Dims1, typename Dims2, size_t n>
 struct sizes_match_below_dim<Dims1, Dims2, n, n> {
-  static EIGEN_DEVICE_FUNC  inline bool run(Dims1& dims1, Dims2& dims2) {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1& dims1, Dims2& dims2) {
     return (array_get<n-1>(dims1) == array_get<n-1>(dims2)) &
         sizes_match_below_dim<Dims1, Dims2, n-1, n-1>::run(dims1, dims2);
   }
 };
 template <typename Dims1, typename Dims2>
 struct sizes_match_below_dim<Dims1, Dims2, 0, 0> {
-  static EIGEN_DEVICE_FUNC  inline bool run(Dims1&, Dims2&) {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1&, Dims2&) {
     return true;
   }
 };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index ac5afd891..17008917a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -39,7 +39,7 @@ class TensorExecutor {
   using StorageIndex = typename Expression::Index;
 
   EIGEN_DEVICE_FUNC
-  static inline void run(const Expression& expr,
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const Device& device = Device()) {
     TensorEvaluator<Expression, Device> evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
@@ -63,7 +63,7 @@ class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true,
   using StorageIndex = typename Expression::Index;
 
   EIGEN_DEVICE_FUNC
-  static inline void run(const Expression& expr,
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const DefaultDevice& device = DefaultDevice()) {
     TensorEvaluator<Expression, DefaultDevice> evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
@@ -111,7 +111,7 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable,
   static const int NumDims = traits<Expression>::NumDimensions;
 
   EIGEN_DEVICE_FUNC
-  static inline void run(const Expression& expr,
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const DefaultDevice& device = DefaultDevice()) {
     using TensorBlock =
         TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
@@ -223,7 +223,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
  public:
   using StorageIndex = typename Expression::Index;
 
-  static inline void run(const Expression& expr,
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
     typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
     typedef EvalRange<Evaluator, StorageIndex, Vectorizable> EvalRange;
@@ -257,7 +257,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
 
   static const int NumDims = traits<Expression>::NumDimensions;
 
-  static inline void run(const Expression& expr,
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
     using TensorBlock =
         TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
@@ -376,7 +376,7 @@ EigenMetaKernel(Evaluator eval, StorageIndex size) {
 
 /*static*/
 template <typename Expression, bool Vectorizable, bool Tileable>
-inline void TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable>::run(
+EIGEN_STRONG_INLINE void TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable>::run(
     const Expression& expr, const GpuDevice& device) {
   TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
   const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
@@ -405,7 +405,7 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable, Tileable>::run(
 template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, SyclDevice, Vectorizable> {
 public:
-  static inline void run(const Expression &expr, const SyclDevice &device) {
+  static EIGEN_STRONG_INLINE void run(const Expression &expr, const SyclDevice &device) {
     // call TensorSYCL module
     TensorSycl::run(expr, device);
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 3d0e4035a..7ecd4d1ac 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -20,7 +20,7 @@ namespace internal {
 template <typename Scalar>
 struct scalar_mod_op {
   EIGEN_DEVICE_FUNC scalar_mod_op(const Scalar& divisor) : m_divisor(divisor) {}
-  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a % m_divisor; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a % m_divisor; }
   const Scalar m_divisor;
 };
 template <typename Scalar>
@@ -34,7 +34,7 @@ struct functor_traits<scalar_mod_op<Scalar> >
 template <typename Scalar>
 struct scalar_mod2_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_mod2_op)
-  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a, const Scalar& b) const { return a % b; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a, const Scalar& b) const { return a % b; }
 };
 template <typename Scalar>
 struct functor_traits<scalar_mod2_op<Scalar> >
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index 8810d78cf..98ad661ca 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -75,10 +75,10 @@ template<DenseIndex n> struct NumTraits<type2index<n> >
     MulCost = 1
   };
 
-  EIGEN_DEVICE_FUNC static inline Real epsilon() { return 0; }
-  EIGEN_DEVICE_FUNC static inline Real dummy_precision() { return 0; }
-  EIGEN_DEVICE_FUNC static inline Real highest() { return n; }
-  EIGEN_DEVICE_FUNC static inline Real lowest() { return n; }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Real epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Real dummy_precision() { return 0; }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Real highest() { return n; }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Real lowest() { return n; }
 };
 
 namespace internal {
diff --git a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
index 8de3bbcab..6c95d0a6c 100644
--- a/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
+++ b/unsupported/Eigen/CXX11/src/util/CXX11Meta.h
@@ -104,9 +104,9 @@ template<>                                  struct h_skip_helper_type<0>
 template<int n>
 struct h_skip {
   template<typename T, T... ii>
-  constexpr static inline typename h_skip_helper_numeric<T, n, ii...>::type helper(numeric_list<T, ii...>) { return typename h_skip_helper_numeric<T, n, ii...>::type(); }
+  constexpr static EIGEN_STRONG_INLINE typename h_skip_helper_numeric<T, n, ii...>::type helper(numeric_list<T, ii...>) { return typename h_skip_helper_numeric<T, n, ii...>::type(); }
   template<typename... tt>
-  constexpr static inline typename h_skip_helper_type<n, tt...>::type helper(type_list<tt...>) { return typename h_skip_helper_type<n, tt...>::type(); }
+  constexpr static EIGEN_STRONG_INLINE typename h_skip_helper_type<n, tt...>::type helper(type_list<tt...>) { return typename h_skip_helper_type<n, tt...>::type(); }
 };
 
 template<int n, typename a> struct skip { typedef decltype(h_skip<n>::helper(a())) type; };
@@ -268,7 +268,7 @@ template<
   typename Reducer
 > struct reduce<Reducer>
 {
-  EIGEN_DEVICE_FUNC constexpr static inline int run() { return Reducer::Identity; }
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE int run() { return Reducer::Identity; }
 };
 
 template<
@@ -276,7 +276,7 @@ template<
   typename A
 > struct reduce<Reducer, A>
 {
-  EIGEN_DEVICE_FUNC constexpr static inline A run(A a) { return a; }
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE A run(A a) { return a; }
 };
 
 template<
@@ -285,7 +285,7 @@ template<
   typename... Ts
 > struct reduce<Reducer, A, Ts...>
 {
-  EIGEN_DEVICE_FUNC constexpr static inline auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
     return Reducer::run(a, reduce<Reducer, Ts...>::run(ts...));
   }
 };
@@ -293,29 +293,29 @@ template<
 /* generic binary operations */
 
 struct sum_op           {
-  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static inline auto run(A a, B b) -> decltype(a + b)   { return a + b;   }
+  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a + b)   { return a + b;   }
   static constexpr int Identity = 0;
 };
 struct product_op       {
-  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static inline auto run(A a, B b) -> decltype(a * b)   { return a * b;   }
+  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a * b)   { return a * b;   }
   static constexpr int Identity = 1;
 };
 
-struct logical_and_op   { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a && b)  { return a && b;  } };
-struct logical_or_op    { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a || b)  { return a || b;  } };
+struct logical_and_op   { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a && b)  { return a && b;  } };
+struct logical_or_op    { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a || b)  { return a || b;  } };
 
-struct equal_op         { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a == b)  { return a == b;  } };
-struct not_equal_op     { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a != b)  { return a != b;  } };
-struct lesser_op        { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a < b)   { return a < b;   } };
-struct lesser_equal_op  { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a <= b)  { return a <= b;  } };
-struct greater_op       { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a > b)   { return a > b;   } };
-struct greater_equal_op { template<typename A, typename B> constexpr static inline auto run(A a, B b) -> decltype(a >= b)  { return a >= b;  } };
+struct equal_op         { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a == b)  { return a == b;  } };
+struct not_equal_op     { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a != b)  { return a != b;  } };
+struct lesser_op        { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a < b)   { return a < b;   } };
+struct lesser_equal_op  { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a <= b)  { return a <= b;  } };
+struct greater_op       { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a > b)   { return a > b;   } };
+struct greater_equal_op { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a >= b)  { return a >= b;  } };
 
 /* generic unary operations */
 
-struct not_op                { template<typename A> constexpr static inline auto run(A a) -> decltype(!a)      { return !a;      } };
-struct negation_op           { template<typename A> constexpr static inline auto run(A a) -> decltype(-a)      { return -a;      } };
-struct greater_equal_zero_op { template<typename A> constexpr static inline auto run(A a) -> decltype(a >= 0)  { return a >= 0;  } };
+struct not_op                { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(!a)      { return !a;      } };
+struct negation_op           { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(-a)      { return -a;      } };
+struct greater_equal_zero_op { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(a >= 0)  { return a >= 0;  } };
 
 
 /* reductions for lists */
@@ -324,13 +324,13 @@ struct greater_equal_zero_op { template<typename A> constexpr static inline auto
 // together in front... (13.0 doesn't work with array_prod/array_reduce/... anyway, but 13.1
 // does...
 template<typename... Ts>
-EIGEN_DEVICE_FUNC constexpr inline decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
 {
   return reduce<product_op, Ts...>::run(ts...);
 }
 
 template<typename... Ts>
-constexpr inline decltype(reduce<sum_op, Ts...>::run((*((Ts*)0))...)) arg_sum(Ts... ts)
+constexpr EIGEN_STRONG_INLINE decltype(reduce<sum_op, Ts...>::run((*((Ts*)0))...)) arg_sum(Ts... ts)
 {
   return reduce<sum_op, Ts...>::run(ts...);
 }
@@ -338,13 +338,13 @@ constexpr inline decltype(reduce<sum_op, Ts...>::run((*((Ts*)0))...)) arg_sum(Ts
 /* reverse arrays */
 
 template<typename Array, int... n>
-constexpr inline Array h_array_reverse(Array arr, numeric_list<int, n...>)
+constexpr EIGEN_STRONG_INLINE Array h_array_reverse(Array arr, numeric_list<int, n...>)
 {
   return {{array_get<sizeof...(n) - n - 1>(arr)...}};
 }
 
 template<typename T, std::size_t N>
-constexpr inline array<T, N> array_reverse(array<T, N> arr)
+constexpr EIGEN_STRONG_INLINE array<T, N> array_reverse(array<T, N> arr)
 {
   return h_array_reverse(arr, typename gen_numeric_list<int, N>::type());
 }
@@ -359,7 +359,7 @@ constexpr inline array<T, N> array_reverse(array<T, N> arr)
 // an infinite loop)
 template<typename Reducer, typename T, std::size_t N, std::size_t n = N - 1>
 struct h_array_reduce {
-  EIGEN_DEVICE_FUNC constexpr static inline auto run(array<T, N> arr, T identity) -> decltype(Reducer::run(h_array_reduce<Reducer, T, N, n - 1>::run(arr, identity), array_get<n>(arr)))
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(array<T, N> arr, T identity) -> decltype(Reducer::run(h_array_reduce<Reducer, T, N, n - 1>::run(arr, identity), array_get<n>(arr)))
   {
     return Reducer::run(h_array_reduce<Reducer, T, N, n - 1>::run(arr, identity), array_get<n>(arr));
   }
@@ -368,7 +368,7 @@ struct h_array_reduce {
 template<typename Reducer, typename T, std::size_t N>
 struct h_array_reduce<Reducer, T, N, 0>
 {
-  EIGEN_DEVICE_FUNC constexpr static inline T run(const array<T, N>& arr, T)
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE T run(const array<T, N>& arr, T)
   {
     return array_get<0>(arr);
   }
@@ -377,14 +377,14 @@ struct h_array_reduce<Reducer, T, N, 0>
 template<typename Reducer, typename T>
 struct h_array_reduce<Reducer, T, 0>
 {
-  EIGEN_DEVICE_FUNC constexpr static inline T run(const array<T, 0>&, T identity)
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE T run(const array<T, 0>&, T identity)
   {
     return identity;
   }
 };
 
 template<typename Reducer, typename T, std::size_t N>
-EIGEN_DEVICE_FUNC constexpr inline auto array_reduce(const array<T, N>& arr, T identity) -> decltype(h_array_reduce<Reducer, T, N>::run(arr, identity))
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_reduce(const array<T, N>& arr, T identity) -> decltype(h_array_reduce<Reducer, T, N>::run(arr, identity))
 {
   return h_array_reduce<Reducer, T, N>::run(arr, identity);
 }
@@ -392,13 +392,13 @@ EIGEN_DEVICE_FUNC constexpr inline auto array_reduce(const array<T, N>& arr, T i
 /* standard array reductions */
 
 template<typename T, std::size_t N>
-EIGEN_DEVICE_FUNC constexpr inline auto array_sum(const array<T, N>& arr) -> decltype(array_reduce<sum_op, T, N>(arr, static_cast<T>(0)))
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_sum(const array<T, N>& arr) -> decltype(array_reduce<sum_op, T, N>(arr, static_cast<T>(0)))
 {
   return array_reduce<sum_op, T, N>(arr, static_cast<T>(0));
 }
 
 template<typename T, std::size_t N>
-EIGEN_DEVICE_FUNC constexpr inline auto array_prod(const array<T, N>& arr) -> decltype(array_reduce<product_op, T, N>(arr, static_cast<T>(1)))
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_prod(const array<T, N>& arr) -> decltype(array_reduce<product_op, T, N>(arr, static_cast<T>(1)))
 {
   return array_reduce<product_op, T, N>(arr, static_cast<T>(1));
 }
@@ -414,13 +414,13 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const std::vector<t>& a) {
 /* zip an array */
 
 template<typename Op, typename A, typename B, std::size_t N, int... n>
-constexpr inline array<decltype(Op::run(A(), B())),N> h_array_zip(array<A, N> a, array<B, N> b, numeric_list<int, n...>)
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A(), B())),N> h_array_zip(array<A, N> a, array<B, N> b, numeric_list<int, n...>)
 {
   return array<decltype(Op::run(A(), B())),N>{{ Op::run(array_get<n>(a), array_get<n>(b))... }};
 }
 
 template<typename Op, typename A, typename B, std::size_t N>
-constexpr inline array<decltype(Op::run(A(), B())),N> array_zip(array<A, N> a, array<B, N> b)
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A(), B())),N> array_zip(array<A, N> a, array<B, N> b)
 {
   return h_array_zip<Op>(a, b, typename gen_numeric_list<int, N>::type());
 }
@@ -428,13 +428,13 @@ constexpr inline array<decltype(Op::run(A(), B())),N> array_zip(array<A, N> a, a
 /* zip an array and reduce the result */
 
 template<typename Reducer, typename Op, typename A, typename B, std::size_t N, int... n>
-constexpr inline auto h_array_zip_and_reduce(array<A, N> a, array<B, N> b, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A(), B()))>::type...>::run(Op::run(array_get<n>(a), array_get<n>(b))...))
+constexpr EIGEN_STRONG_INLINE auto h_array_zip_and_reduce(array<A, N> a, array<B, N> b, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A(), B()))>::type...>::run(Op::run(array_get<n>(a), array_get<n>(b))...))
 {
   return reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A(), B()))>::type...>::run(Op::run(array_get<n>(a), array_get<n>(b))...);
 }
 
 template<typename Reducer, typename Op, typename A, typename B, std::size_t N>
-constexpr inline auto array_zip_and_reduce(array<A, N> a, array<B, N> b) -> decltype(h_array_zip_and_reduce<Reducer, Op, A, B, N>(a, b, typename gen_numeric_list<int, N>::type()))
+constexpr EIGEN_STRONG_INLINE auto array_zip_and_reduce(array<A, N> a, array<B, N> b) -> decltype(h_array_zip_and_reduce<Reducer, Op, A, B, N>(a, b, typename gen_numeric_list<int, N>::type()))
 {
   return h_array_zip_and_reduce<Reducer, Op, A, B, N>(a, b, typename gen_numeric_list<int, N>::type());
 }
@@ -442,13 +442,13 @@ constexpr inline auto array_zip_and_reduce(array<A, N> a, array<B, N> b) -> decl
 /* apply stuff to an array */
 
 template<typename Op, typename A, std::size_t N, int... n>
-constexpr inline array<decltype(Op::run(A())),N> h_array_apply(array<A, N> a, numeric_list<int, n...>)
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A())),N> h_array_apply(array<A, N> a, numeric_list<int, n...>)
 {
   return array<decltype(Op::run(A())),N>{{ Op::run(array_get<n>(a))... }};
 }
 
 template<typename Op, typename A, std::size_t N>
-constexpr inline array<decltype(Op::run(A())),N> array_apply(array<A, N> a)
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A())),N> array_apply(array<A, N> a)
 {
   return h_array_apply<Op>(a, typename gen_numeric_list<int, N>::type());
 }
@@ -456,13 +456,13 @@ constexpr inline array<decltype(Op::run(A())),N> array_apply(array<A, N> a)
 /* apply stuff to an array and reduce */
 
 template<typename Reducer, typename Op, typename A, std::size_t N, int... n>
-constexpr inline auto h_array_apply_and_reduce(array<A, N> arr, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A()))>::type...>::run(Op::run(array_get<n>(arr))...))
+constexpr EIGEN_STRONG_INLINE auto h_array_apply_and_reduce(array<A, N> arr, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A()))>::type...>::run(Op::run(array_get<n>(arr))...))
 {
   return reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A()))>::type...>::run(Op::run(array_get<n>(arr))...);
 }
 
 template<typename Reducer, typename Op, typename A, std::size_t N>
-constexpr inline auto array_apply_and_reduce(array<A, N> a) -> decltype(h_array_apply_and_reduce<Reducer, Op, A, N>(a, typename gen_numeric_list<int, N>::type()))
+constexpr EIGEN_STRONG_INLINE auto array_apply_and_reduce(array<A, N> a) -> decltype(h_array_apply_and_reduce<Reducer, Op, A, N>(a, typename gen_numeric_list<int, N>::type()))
 {
   return h_array_apply_and_reduce<Reducer, Op, A, N>(a, typename gen_numeric_list<int, N>::type());
 }
@@ -476,7 +476,7 @@ template<int n>
 struct h_repeat
 {
   template<typename t, int... ii>
-  constexpr static inline array<t, n> run(t v, numeric_list<int, ii...>)
+  constexpr static EIGEN_STRONG_INLINE array<t, n> run(t v, numeric_list<int, ii...>)
   {
     return {{ typename id_numeric<int, ii, t>::type(v)... }};
   }
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
index 30cdf4751..ed6d83251 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
@@ -24,7 +24,7 @@ namespace Eigen {
   * \sa Eigen::igammac(), Eigen::lgamma()
   */
 template<typename Derived,typename ExponentDerived>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
 igamma(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x)
 {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
@@ -47,7 +47,7 @@ igamma(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerive
   * \sa Eigen::igamma(), Eigen::lgamma()
   */
 template <typename Derived, typename ExponentDerived>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
 igamma_der_a(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x) {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
     a.derived(),
@@ -68,7 +68,7 @@ igamma_der_a(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<Exponent
   * \sa Eigen::igamma(), Eigen::lgamma()
   */
 template <typename AlphaDerived, typename SampleDerived>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>
 gamma_sample_der_alpha(const Eigen::ArrayBase<AlphaDerived>& alpha, const Eigen::ArrayBase<SampleDerived>& sample) {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>(
       alpha.derived(),
@@ -86,7 +86,7 @@ gamma_sample_der_alpha(const Eigen::ArrayBase<AlphaDerived>& alpha, const Eigen:
   * \sa Eigen::igamma(), Eigen::lgamma()
   */
 template<typename Derived,typename ExponentDerived>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igammac_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igammac_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
 igammac(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x)
 {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igammac_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
@@ -108,7 +108,7 @@ igammac(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDeriv
 // * \warning Be careful with the order of the parameters: x.polygamma(n) is equivalent to polygamma(n,x)
 // * \sa ArrayBase::polygamma()
 template<typename DerivedN,typename DerivedX>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_polygamma_op<typename DerivedX::Scalar>, const DerivedN, const DerivedX>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_polygamma_op<typename DerivedX::Scalar>, const DerivedN, const DerivedX>
 polygamma(const Eigen::ArrayBase<DerivedN>& n, const Eigen::ArrayBase<DerivedX>& x)
 {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_polygamma_op<typename DerivedX::Scalar>, const DerivedN, const DerivedX>(
@@ -128,7 +128,7 @@ polygamma(const Eigen::ArrayBase<DerivedN>& n, const Eigen::ArrayBase<DerivedX>&
   * \sa Eigen::betainc(), Eigen::lgamma()
   */
 template<typename ArgADerived, typename ArgBDerived, typename ArgXDerived>
-inline const Eigen::CwiseTernaryOp<Eigen::internal::scalar_betainc_op<typename ArgXDerived::Scalar>, const ArgADerived, const ArgBDerived, const ArgXDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseTernaryOp<Eigen::internal::scalar_betainc_op<typename ArgXDerived::Scalar>, const ArgADerived, const ArgBDerived, const ArgXDerived>
 betainc(const Eigen::ArrayBase<ArgADerived>& a, const Eigen::ArrayBase<ArgBDerived>& b, const Eigen::ArrayBase<ArgXDerived>& x)
 {
   return Eigen::CwiseTernaryOp<Eigen::internal::scalar_betainc_op<typename ArgXDerived::Scalar>, const ArgADerived, const ArgBDerived, const ArgXDerived>(
@@ -152,7 +152,7 @@ betainc(const Eigen::ArrayBase<ArgADerived>& a, const Eigen::ArrayBase<ArgBDeriv
   * \sa ArrayBase::zeta()
   */
 template<typename DerivedX,typename DerivedQ>
-inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_zeta_op<typename DerivedX::Scalar>, const DerivedX, const DerivedQ>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_zeta_op<typename DerivedX::Scalar>, const DerivedX, const DerivedQ>
 zeta(const Eigen::ArrayBase<DerivedX>& x, const Eigen::ArrayBase<DerivedQ>& q)
 {
   return Eigen::CwiseBinaryOp<Eigen::internal::scalar_zeta_op<typename DerivedX::Scalar>, const DerivedX, const DerivedQ>(
@@ -176,7 +176,7 @@ zeta(const Eigen::ArrayBase<DerivedX>& x, const Eigen::ArrayBase<DerivedQ>& q)
   * \sa ArrayBase::i0e()
   */
 template <typename Derived>
-inline const Eigen::CwiseUnaryOp<
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
     Eigen::internal::scalar_i0e_op<typename Derived::Scalar>, const Derived>
 i0e(const Eigen::ArrayBase<Derived>& x) {
   return Eigen::CwiseUnaryOp<
@@ -199,7 +199,7 @@ i0e(const Eigen::ArrayBase<Derived>& x) {
   * \sa ArrayBase::i1e()
   */
 template <typename Derived>
-inline const Eigen::CwiseUnaryOp<
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
     Eigen::internal::scalar_i1e_op<typename Derived::Scalar>, const Derived>
 i1e(const Eigen::ArrayBase<Derived>& x) {
   return Eigen::CwiseUnaryOp<
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
index 3a63dcdd6..c6fac91bb 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsFunctors.h
@@ -155,11 +155,11 @@ struct functor_traits<scalar_betainc_op<Scalar> > {
  */
 template<typename Scalar> struct scalar_lgamma_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_lgamma_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
     using numext::lgamma; return lgamma(a);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plgamma(a); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::plgamma(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_lgamma_op<Scalar> >
@@ -177,11 +177,11 @@ struct functor_traits<scalar_lgamma_op<Scalar> >
  */
 template<typename Scalar> struct scalar_digamma_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_digamma_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
     using numext::digamma; return digamma(a);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pdigamma(a); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::pdigamma(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_digamma_op<Scalar> >
@@ -199,11 +199,11 @@ struct functor_traits<scalar_digamma_op<Scalar> >
  */
 template<typename Scalar> struct scalar_zeta_op {
     EIGEN_EMPTY_STRUCT_CTOR(scalar_zeta_op)
-    EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& x, const Scalar& q) const {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& x, const Scalar& q) const {
         using numext::zeta; return zeta(x, q);
     }
     typedef typename packet_traits<Scalar>::type Packet;
-    EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x, const Packet& q) const { return internal::pzeta(x, q); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x, const Packet& q) const { return internal::pzeta(x, q); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_zeta_op<Scalar> >
@@ -221,11 +221,11 @@ struct functor_traits<scalar_zeta_op<Scalar> >
  */
 template<typename Scalar> struct scalar_polygamma_op {
     EIGEN_EMPTY_STRUCT_CTOR(scalar_polygamma_op)
-    EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& n, const Scalar& x) const {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& n, const Scalar& x) const {
         using numext::polygamma; return polygamma(n, x);
     }
     typedef typename packet_traits<Scalar>::type Packet;
-    EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& n, const Packet& x) const { return internal::ppolygamma(n, x); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& n, const Packet& x) const { return internal::ppolygamma(n, x); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_polygamma_op<Scalar> >
@@ -244,11 +244,11 @@ struct functor_traits<scalar_polygamma_op<Scalar> >
  */
 template<typename Scalar> struct scalar_erf_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_erf_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
     using numext::erf; return erf(a);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::perf(a); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::perf(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_erf_op<Scalar> >
@@ -267,11 +267,11 @@ struct functor_traits<scalar_erf_op<Scalar> >
  */
 template<typename Scalar> struct scalar_erfc_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_erfc_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
     using numext::erfc; return erfc(a);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::perfc(a); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::perfc(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_erfc_op<Scalar> >
@@ -291,12 +291,12 @@ struct functor_traits<scalar_erfc_op<Scalar> >
 template <typename Scalar>
 struct scalar_i0e_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_i0e_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& x) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
     using numext::i0e;
     return i0e(x);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
     return internal::pi0e(x);
   }
 };
@@ -318,12 +318,12 @@ struct functor_traits<scalar_i0e_op<Scalar> > {
 template <typename Scalar>
 struct scalar_i1e_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_i1e_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& x) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
     using numext::i1e;
     return i1e(x);
   }
   typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
     return internal::pi1e(x);
   }
 };

From d7a84148483b1a11b993c037a2cea5b43f2c052f Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 11:56:30 +0100
Subject: [PATCH 511/680] Distinguishing between internal memory
 allocation/deallocation from explicit user memory allocation/deallocation.

---
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h        |  8 ++++----
 .../Eigen/CXX11/src/Tensor/TensorDeviceDefault.h   |  6 ++++++
 .../Eigen/CXX11/src/Tensor/TensorDeviceGpu.h       |  7 +++++++
 .../CXX11/src/Tensor/TensorDeviceThreadPool.h      |  8 ++++++++
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h      |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorReduction.h       | 14 +++++++-------
 6 files changed, 34 insertions(+), 13 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 0e4db46de..7b4d56e4b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -112,7 +112,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
       return false;
     } else {
       m_result = static_cast<CoeffReturnType*>(
-          m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)));
+          m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
       evalTo(m_result);
       return true;
     }
@@ -120,7 +120,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
     if (m_result != NULL) {
-      m_device.deallocate(m_result);
+      m_device.deallocate_temp(m_result);
       m_result = NULL;
     }
   }
@@ -273,7 +273,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
       evalTo(data);
       return false;
     } else {
-      m_result = static_cast<Scalar *>(m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)));
+      m_result = static_cast<Scalar *>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
       evalTo(m_result);
       return true;
     }
@@ -281,7 +281,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
     if (m_result != NULL) {
-      m_device.deallocate(m_result);
+      m_device.deallocate_temp(m_result);
       m_result = NULL;
     }
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
index 5c1c68912..8cb95f731 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
@@ -20,6 +20,12 @@ struct DefaultDevice {
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
     internal::aligned_free(buffer);
+  }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return allocate(num_bytes);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    deallocate(buffer);
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
     ::memcpy(dst, src, n);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
index 0c036833f..48bbd5e45 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
@@ -159,6 +159,13 @@ class GpuStreamDevice : public StreamInterface {
     err = gpuFree(buffer);
     gpu_assert(err == gpuSuccess);
   }
+   EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return stream_->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    stream_->deallocate(buffer);
+  }
 
   virtual void* scratchpad() const {
     if (scratch_ == NULL) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index 90fd99027..5a16ebe50 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -105,6 +105,14 @@ struct ThreadPoolDevice {
     internal::aligned_free(buffer);
   }
 
+    EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    deallocate(buffer);
+  }
+
   EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
     ::memcpy(dst, src, n);
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index fdb31928f..09f6f2067 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -115,7 +115,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
   #endif
   EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
     const Index numValues =  internal::array_prod(m_impl.dimensions());
-    m_buffer = (CoeffReturnType*)m_device.allocate(numValues * sizeof(CoeffReturnType));
+    m_buffer = (CoeffReturnType*)m_device.allocate_temp(numValues * sizeof(CoeffReturnType));
     // Should initialize the memory in case we're dealing with non POD types.
     if (NumTraits<CoeffReturnType>::RequireInitialization) {
       for (Index i = 0; i < numValues; ++i) {
@@ -129,7 +129,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     return true;
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
-    m_device.deallocate(m_buffer);
+    m_device.deallocate_temp(m_buffer);
     m_buffer = NULL;
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 375fc0802..c1cbdebc6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -513,7 +513,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
          !RunningOnGPU))) {
       bool need_assign = false;
       if (!data) {
-        m_result = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType)));
+        m_result = static_cast<CoeffReturnType*>(m_device.allocate_temp(sizeof(CoeffReturnType)));
         data = m_result;
         need_assign = true;
       }
@@ -525,7 +525,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
       const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
       if (!data) {
-        data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
+        data = static_cast<CoeffReturnType*>(m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
         m_result = data;
       }
       Op reducer(m_reducer);
@@ -549,7 +549,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
         if (!data) {
           if (num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 128) {
-            data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
+            data = static_cast<CoeffReturnType*>(m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
             m_result = data;
           }
           else {
@@ -559,7 +559,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         Op reducer(m_reducer);
         if (internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) {
           if (m_result) {
-            m_device.deallocate(m_result);
+            m_device.deallocate_temp(m_result);
             m_result = NULL;
           }
           return true;
@@ -582,7 +582,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
         if (!data) {
           if (num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 32) {
-            data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
+            data = static_cast<CoeffReturnType*>(m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve));
             m_result = data;
           }
           else {
@@ -592,7 +592,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         Op reducer(m_reducer);
         if (internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) {
           if (m_result) {
-            m_device.deallocate(m_result);
+            m_device.deallocate_temp(m_result);
             m_result = NULL;
           }
           return true;
@@ -607,7 +607,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
     m_impl.cleanup();
     if (m_result) {
-      m_device.deallocate(m_result);
+      m_device.deallocate_temp(m_result);
       m_result = NULL;
     }
   }

From 3a197a60e602ea0cd836438ab717810803dc9074 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 12:19:14 +0100
Subject: [PATCH 512/680] variadic version of assert which can take a parameter
 pack as its input.

---
 Eigen/src/Core/util/Macros.h                  | 24 +++++++++++++++++++
 .../Eigen/CXX11/src/Tensor/TensorMap.h        |  2 ++
 2 files changed, 26 insertions(+)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index b15819f7d..3255b8351 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1076,4 +1076,28 @@ namespace Eigen {
 #   endif
 #endif
 
+#ifdef EIGEN_HAS_VARIADIC_TEMPLATES
+// Provide a variadic version of assert which can take a parameter pack as its input
+// The eigen_assert macro used here might have been redefined to use other macros such as EIGEN_THROW, such as used in Eigen's test suite, therefore this needs to be defined after the other macros.
+// Note that this does not provide as nice a string to assert as a straight forward call to eigen_assert, so we add a message to the assert.
+#if defined(EIGEN_NO_DEBUG)
+#define eigen_variadic_assert(x)
+#else
+namespace Eigen {
+namespace internal {
+inline void variadic_assert(const char*) {}
+template<typename... Bools> inline void variadic_assert(const char* message, bool first, Bools ... others) {
+  eigen_assert(first && message);
+  variadic_assert(message, others...);
+  EIGEN_UNUSED_VARIABLE(first);
+}
+}
+}
+#define EIGEN_VARIADIC_ASSERT_MESSAGE(x) EIGEN_MAKESTRING(x) " in " __FILE__ ":" EIGEN_MAKESTRING(__LINE__)
+#define eigen_variadic_assert(x) \
+  do { Eigen::internal::variadic_assert(EIGEN_VARIADIC_ASSERT_MESSAGE(x), x); } while(false);
+#endif
+#endif
+
+
 #endif // EIGEN_MACROS_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
index a8e55757e..f69f8f24a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
@@ -150,6 +150,7 @@ template<typename PlainObjectType, int Options_, template <class> class MakePoin
     EIGEN_STRONG_INLINE const Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
     {
       EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      eigen_variadic_assert((Eigen::NumTraits<Index>::highest() >= otherIndices)...);
       if (PlainObjectType::Options&RowMajor) {
         const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
         return m_data[index];
@@ -237,6 +238,7 @@ template<typename PlainObjectType, int Options_, template <class> class MakePoin
     EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices)
     {
       static_assert(sizeof...(otherIndices) + 2 == NumIndices || NumIndices == Dynamic, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor.");
+      eigen_variadic_assert((Eigen::NumTraits<Index>::highest() >= otherIndices)...);
       const std::size_t NumDims = sizeof...(otherIndices) + 2;
       if (PlainObjectType::Options&RowMajor) {
         const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumDims>{{firstIndex, secondIndex, otherIndices...}});

From c84509d7cc5fa3e032da8cfdcd5e82b2897cc5d9 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 12:40:54 +0100
Subject: [PATCH 513/680] Adding new arch/SYCL headers, used for SYCL
 vectorization.

---
 Eigen/src/Core/arch/SYCL/InteropHeaders.h | 104 +++++
 Eigen/src/Core/arch/SYCL/MathFunctions.h  | 221 +++++++++++
 Eigen/src/Core/arch/SYCL/PacketMath.h     | 458 ++++++++++++++++++++++
 Eigen/src/Core/arch/SYCL/TypeCasting.h    |  89 +++++
 4 files changed, 872 insertions(+)
 create mode 100644 Eigen/src/Core/arch/SYCL/InteropHeaders.h
 create mode 100644 Eigen/src/Core/arch/SYCL/MathFunctions.h
 create mode 100644 Eigen/src/Core/arch/SYCL/PacketMath.h
 create mode 100644 Eigen/src/Core/arch/SYCL/TypeCasting.h

diff --git a/Eigen/src/Core/arch/SYCL/InteropHeaders.h b/Eigen/src/Core/arch/SYCL/InteropHeaders.h
new file mode 100644
index 000000000..c1da40d14
--- /dev/null
+++ b/Eigen/src/Core/arch/SYCL/InteropHeaders.h
@@ -0,0 +1,104 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * InteropHeaders.h
+ *
+ * \brief:
+ *  InteropHeaders
+ *
+*****************************************************************/
+
+#ifndef EIGEN_INTEROP_HEADERS_SYCL_H
+#define EIGEN_INTEROP_HEADERS_SYCL_H
+#if defined EIGEN_USE_SYCL
+namespace Eigen {
+
+namespace internal {
+#define SYCL_PACKET_TRAITS(packet_type, val, unpacket_type, lengths)\
+  template<> struct packet_traits<unpacket_type> : default_packet_traits\
+  {\
+    typedef packet_type type;\
+    typedef packet_type half;\
+    enum {\
+      Vectorizable = 1,\
+      AlignedOnScalar = 1,\
+      size=lengths,\
+      HasHalfPacket = 0,\
+      HasDiv  = 1,\
+      HasLog  = 1,\
+      HasExp  = 1,\
+      HasSqrt = 1,\
+      HasRsqrt = 1,\
+      HasSin    = 1,\
+      HasCos    = 1,\
+      HasTan    = 1,\
+      HasASin   = 1,\
+      HasACos   = 1,\
+      HasATan   = 1,\
+      HasSinh   = 1,\
+      HasCosh   = 1,\
+      HasTanh   = 1,\
+      HasLGamma = 0,\
+      HasDiGamma = 0,\
+      HasZeta = 0,\
+      HasPolygamma = 0,\
+      HasErf = 0,\
+      HasErfc = 0,\
+      HasIGamma = 0,\
+      HasIGammac = 0,\
+      HasBetaInc = 0,\
+      HasBlend = val,\
+      HasMax=1,\
+      HasMin=1,\
+      HasMul=1,\
+      HasAdd=1,\
+      HasFloor=1,\
+      HasRound=1,\
+      HasLog1p=1,\
+      HasExpm1=1,\
+      HasCeil=1,\
+    };\
+  };
+
+SYCL_PACKET_TRAITS(cl::sycl::cl_float4, 1, float, 4)
+SYCL_PACKET_TRAITS(cl::sycl::cl_float4, 1, const float, 4)
+SYCL_PACKET_TRAITS(cl::sycl::cl_double2, 0, double, 2)
+SYCL_PACKET_TRAITS(cl::sycl::cl_double2, 0, const double, 2)
+#undef SYCL_PACKET_TRAITS
+
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#define SYCL_ARITHMETIC(packet_type) template<> struct is_arithmetic<packet_type>  { enum { value = true }; };
+SYCL_ARITHMETIC(cl::sycl::cl_float4)
+SYCL_ARITHMETIC(cl::sycl::cl_double2)
+#undef SYCL_ARITHMETIC
+
+#define SYCL_UNPACKET_TRAITS(packet_type, unpacket_type, lengths)\
+template<> struct unpacket_traits<packet_type>  {\
+  typedef unpacket_type  type;\
+  enum {size=lengths, alignment=Aligned16};\
+  typedef packet_type half;\
+};
+SYCL_UNPACKET_TRAITS(cl::sycl::cl_float4, float, 4)
+SYCL_UNPACKET_TRAITS(cl::sycl::cl_double2, double, 2)
+
+#undef SYCL_UNPACKET_TRAITS
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_USE_SYCL
+#endif // EIGEN_INTEROP_HEADERS_SYCL_H
diff --git a/Eigen/src/Core/arch/SYCL/MathFunctions.h b/Eigen/src/Core/arch/SYCL/MathFunctions.h
new file mode 100644
index 000000000..422839c6c
--- /dev/null
+++ b/Eigen/src/Core/arch/SYCL/MathFunctions.h
@@ -0,0 +1,221 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * MathFunctions.h
+ *
+ * \brief:
+ *  MathFunctions
+ *
+*****************************************************************/
+
+#ifndef EIGEN_MATH_FUNCTIONS_SYCL_H
+#define EIGEN_MATH_FUNCTIONS_SYCL_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+//#if defined(__SYCL_DEVICE_ONLY__) && defined(EIGEN_USE_SYCL)
+#define SYCL_PLOG(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type plog<packet_type>(const packet_type& a) { return cl::sycl::log(a); }
+
+SYCL_PLOG(cl::sycl::cl_float4)
+SYCL_PLOG(cl::sycl::cl_double2)
+#undef SYCL_PLOG
+
+#define SYCL_PLOG1P(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type plog1p<packet_type>(const packet_type& a) { return cl::sycl::log1p(a); }
+
+SYCL_PLOG1P(cl::sycl::cl_float4)
+SYCL_PLOG1P(cl::sycl::cl_double2)
+#undef SYCL_PLOG1P
+
+#define SYCL_PLOG10(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type plog10<packet_type>(const packet_type& a) { return cl::sycl::log10(a); }
+
+SYCL_PLOG10(cl::sycl::cl_float4)
+SYCL_PLOG10(cl::sycl::cl_double2)
+#undef SYCL_PLOG10
+
+#define SYCL_PEXP(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pexp<packet_type>(const packet_type& a) { return cl::sycl::exp(a); }
+
+SYCL_PEXP(cl::sycl::cl_float4)
+SYCL_PEXP(cl::sycl::cl_double2)
+#undef SYCL_PEXP
+
+#define SYCL_PEXPM1(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pexpm1<packet_type>(const packet_type& a) { return cl::sycl::expm1(a); }
+
+SYCL_PEXPM1(cl::sycl::cl_float4)
+SYCL_PEXPM1(cl::sycl::cl_double2)
+#undef SYCL_PEXPM1
+
+#define SYCL_PSQRT(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type psqrt<packet_type>(const packet_type& a) { return cl::sycl::sqrt(a); }
+
+SYCL_PSQRT(cl::sycl::cl_float4)
+SYCL_PSQRT(cl::sycl::cl_double2)
+#undef SYCL_PSQRT
+
+
+#define SYCL_PRSQRT(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type prsqrt<packet_type>(const packet_type& a) { return cl::sycl::rsqrt(a); }
+
+SYCL_PRSQRT(cl::sycl::cl_float4)
+SYCL_PRSQRT(cl::sycl::cl_double2)
+#undef SYCL_PRSQRT
+
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+#define SYCL_PSIN(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type psin<packet_type>(const packet_type& a)  { return cl::sycl::sin(a); }
+
+SYCL_PSIN(cl::sycl::cl_float4)
+SYCL_PSIN(cl::sycl::cl_double2)
+#undef SYCL_PSIN
+
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+#define SYCL_PCOS(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pcos<packet_type>(const packet_type& a)  { return cl::sycl::cos(a); }
+
+SYCL_PCOS(cl::sycl::cl_float4)
+SYCL_PCOS(cl::sycl::cl_double2)
+#undef SYCL_PCOS
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+#define SYCL_PTAN(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type ptan<packet_type>(const packet_type& a) {  return cl::sycl::tan(a); }
+
+SYCL_PTAN(cl::sycl::cl_float4)
+SYCL_PTAN(cl::sycl::cl_double2)
+#undef SYCL_PTAN
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+#define SYCL_PASIN(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pasin<packet_type>(const packet_type& a)  { return cl::sycl::asin(a); }
+
+SYCL_PASIN(cl::sycl::cl_float4)
+SYCL_PASIN(cl::sycl::cl_double2)
+#undef SYCL_PASIN
+
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+#define SYCL_PACOS(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pacos<packet_type>(const packet_type& a)  { return cl::sycl::acos(a); }
+
+SYCL_PACOS(cl::sycl::cl_float4)
+SYCL_PACOS(cl::sycl::cl_double2)
+#undef SYCL_PACOS
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+#define SYCL_PATAN(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type patan<packet_type>(const packet_type& a) {  return cl::sycl::atan(a); }
+
+SYCL_PATAN(cl::sycl::cl_float4)
+SYCL_PATAN(cl::sycl::cl_double2)
+#undef SYCL_PATAN
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+#define SYCL_PSINH(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type psinh<packet_type>(const packet_type& a)  { return cl::sycl::sinh(a); }
+
+SYCL_PSINH(cl::sycl::cl_float4)
+SYCL_PSINH(cl::sycl::cl_double2)
+#undef SYCL_PSINH
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+#define SYCL_PCOSH(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pcosh<packet_type>(const packet_type& a)  { return cl::sycl::cosh(a); }
+
+SYCL_PCOSH(cl::sycl::cl_float4)
+SYCL_PCOSH(cl::sycl::cl_double2)
+#undef SYCL_PCOSH
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+#define SYCL_PTANH(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type ptanh<packet_type>(const packet_type& a) {  return cl::sycl::tanh(a); }
+
+SYCL_PTANH(cl::sycl::cl_float4)
+SYCL_PTANH(cl::sycl::cl_double2)
+#undef SYCL_PTANH
+
+#define SYCL_PCEIL(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pceil<packet_type>(const packet_type& a) { return cl::sycl::ceil(a); }
+
+SYCL_PCEIL(cl::sycl::cl_float4)
+SYCL_PCEIL(cl::sycl::cl_double2)
+#undef SYCL_PCEIL
+
+
+#define SYCL_PROUND(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pround<packet_type>(const packet_type& a) { return cl::sycl::round(a); }
+
+SYCL_PROUND(cl::sycl::cl_float4)
+SYCL_PROUND(cl::sycl::cl_double2)
+#undef SYCL_PROUND
+
+#define SYCL_FLOOR(packet_type) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pfloor<packet_type>(const packet_type& a) { return cl::sycl::floor(a); }
+
+SYCL_FLOOR(cl::sycl::cl_float4)
+SYCL_FLOOR(cl::sycl::cl_double2)
+#undef SYCL_FLOOR
+
+
+#define SYCL_PMIN(packet_type, expr) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pmin<packet_type>(const packet_type& a, const packet_type& b) { return expr; }
+
+SYCL_PMIN(cl::sycl::cl_float4, cl::sycl::fmin(a, b))
+SYCL_PMIN(cl::sycl::cl_double2, cl::sycl::fmin(a, b))
+#undef SYCL_PMIN
+
+#define SYCL_PMAX(packet_type, expr) \
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \
+packet_type pmax<packet_type>(const packet_type& a, const packet_type& b) { return expr; }
+
+SYCL_PMAX(cl::sycl::cl_float4, cl::sycl::fmax(a, b))
+SYCL_PMAX(cl::sycl::cl_double2, cl::sycl::fmax(a, b))
+#undef SYCL_PMAX
+
+//#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
diff --git a/Eigen/src/Core/arch/SYCL/PacketMath.h b/Eigen/src/Core/arch/SYCL/PacketMath.h
new file mode 100644
index 000000000..820a83311
--- /dev/null
+++ b/Eigen/src/Core/arch/SYCL/PacketMath.h
@@ -0,0 +1,458 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * PacketMath.h
+ *
+ * \brief:
+ *  PacketMath
+ *
+*****************************************************************/
+
+#ifndef EIGEN_PACKET_MATH_SYCL_H
+#define EIGEN_PACKET_MATH_SYCL_H
+#include <type_traits>
+#if defined EIGEN_USE_SYCL
+namespace Eigen {
+
+namespace internal {
+
+#define SYCL_PLOADT_RO(address_space_target)\
+template<typename packet_type, int Alignment>\
+ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type\
+ ploadt_ro(typename cl::sycl::multi_ptr<const typename unpacket_traits<packet_type>::type,\
+   cl::sycl::access::address_space::address_space_target>::pointer_t from) {\
+   typedef typename unpacket_traits<packet_type>::type scalar;\
+   typedef cl::sycl::multi_ptr<scalar, cl::sycl::access::address_space::address_space_target> multi_ptr;\
+   auto res=packet_type(static_cast<typename unpacket_traits<packet_type>::type>(0));\
+   res.load(0, multi_ptr(const_cast<typename multi_ptr::pointer_t>(from)));\
+   return res;\
+}
+
+SYCL_PLOADT_RO(global_space)
+SYCL_PLOADT_RO(local_space)
+
+#undef SYCL_PLOADT_RO
+
+
+#define SYCL_PLOAD(address_space_target, Alignment, AlignedType)\
+template<typename packet_type> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type\
+ pload##AlignedType(typename cl::sycl::multi_ptr<const typename unpacket_traits<packet_type>::type,\
+   cl::sycl::access::address_space::address_space_target>::pointer_t from) {\
+   return ploadt_ro<packet_type, Alignment>(from);\
+ }
+
+// global space
+SYCL_PLOAD(global_space, Unaligned, u)
+SYCL_PLOAD(global_space, Aligned, )
+
+// local space
+SYCL_PLOAD(local_space, Unaligned, u)
+SYCL_PLOAD(local_space, Aligned, )
+
+// private space
+//SYCL_PLOAD(private_space, Unaligned, u)
+//SYCL_PLOAD(private_space, Aligned, )
+
+#undef SYCL_PLOAD
+
+
+/** \internal \returns a packet version of \a *from.
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+#define SYCL_PLOADT(address_space_target)\
+template<typename packet_type, int Alignment>\
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type ploadt(\
+  typename cl::sycl::multi_ptr<const typename unpacket_traits<packet_type>::type,\
+  cl::sycl::access::address_space::address_space_target>::pointer_t from)\
+{\
+  if(Alignment >= unpacket_traits<packet_type>::alignment)\
+    return pload<packet_type>(from);\
+  else\
+    return ploadu<packet_type>(from);\
+}
+
+// global space
+SYCL_PLOADT(global_space)
+// local space
+SYCL_PLOADT(local_space)
+
+//private_space
+// There is no need to specialise it for private space as it can use the GenericPacketMath version
+
+#define SYCL_PLOADT_RO_SPECIAL(packet_type, Alignment)\
+ template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type\
+  ploadt_ro<packet_type, Alignment>(const typename unpacket_traits<packet_type>::type * from) { \
+    typedef typename unpacket_traits<packet_type>::type scalar;\
+   auto res=packet_type(static_cast<scalar>(0));\
+   res. template load<cl::sycl::access::address_space::private_space>(0, const_cast<scalar*>(from));\
+   return res;\
+  }
+
+SYCL_PLOADT_RO_SPECIAL(cl::sycl::cl_float4, Aligned)
+SYCL_PLOADT_RO_SPECIAL(cl::sycl::cl_double2, Aligned)
+SYCL_PLOADT_RO_SPECIAL(cl::sycl::cl_float4, Unaligned)
+SYCL_PLOADT_RO_SPECIAL(cl::sycl::cl_double2, Unaligned)
+
+
+#define SYCL_PLOAD_SPECIAL(packet_type, alignment_type)\
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type\
+  pload##alignment_type(const typename unpacket_traits<packet_type>::type * from) { \
+    typedef typename unpacket_traits<packet_type>::type scalar;\
+   auto res=packet_type(static_cast<scalar>(0));\
+   res. template load<cl::sycl::access::address_space::private_space>(0, const_cast<scalar*>(from));\
+   return res;\
+  }
+SYCL_PLOAD_SPECIAL(cl::sycl::cl_float4,)
+SYCL_PLOAD_SPECIAL(cl::sycl::cl_double2,)
+SYCL_PLOAD_SPECIAL(cl::sycl::cl_float4, u)
+SYCL_PLOAD_SPECIAL(cl::sycl::cl_double2, u)
+
+#undef SYCL_PLOAD_SPECIAL
+
+#define SYCL_PSTORE(scalar, packet_type, address_space_target, alignment)\
+template<>\
+ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstore##alignment( \
+   typename cl::sycl::multi_ptr<scalar, cl::sycl::access::address_space::address_space_target>::pointer_t to, \
+   const packet_type& from) {\
+     typedef cl::sycl::multi_ptr<scalar, cl::sycl::access::address_space::address_space_target> multi_ptr;\
+     from.store(0, multi_ptr(to));\
+}
+
+// global space
+SYCL_PSTORE(float, cl::sycl::cl_float4, global_space, )
+SYCL_PSTORE(float, cl::sycl::cl_float4, global_space, u)
+SYCL_PSTORE(double, cl::sycl::cl_double2, global_space, )
+SYCL_PSTORE(double, cl::sycl::cl_double2, global_space, u)
+
+SYCL_PSTORE(float, cl::sycl::cl_float4, local_space, )
+SYCL_PSTORE(float, cl::sycl::cl_float4, local_space, u)
+SYCL_PSTORE(double, cl::sycl::cl_double2, local_space, )
+SYCL_PSTORE(double, cl::sycl::cl_double2, local_space, u)
+
+SYCL_PSTORE(float, cl::sycl::cl_float4, private_space, )
+SYCL_PSTORE(float, cl::sycl::cl_float4, private_space, u)
+SYCL_PSTORE(double, cl::sycl::cl_double2, private_space, )
+SYCL_PSTORE(double, cl::sycl::cl_double2, private_space, u)
+
+
+#define SYCL_PSTORE_T(scalar, packet_type, Alignment)\
+template<>\
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret<scalar, packet_type, Alignment>(\
+  scalar* to,\
+  const packet_type& from) {\
+  if(Alignment)\
+  pstore(to, from);\
+  else\
+  pstoreu(to,from);\
+}
+
+
+SYCL_PSTORE_T(float, cl::sycl::cl_float4, Aligned)
+
+SYCL_PSTORE_T(float, cl::sycl::cl_float4, Unaligned)
+
+SYCL_PSTORE_T(double, cl::sycl::cl_double2, Aligned)
+
+SYCL_PSTORE_T(double, cl::sycl::cl_double2, Unaligned)
+
+
+#undef SYCL_PSTORE_T
+
+#define SYCL_PSET1(packet_type)\
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type pset1<packet_type>(\
+  const typename unpacket_traits<packet_type>::type&  from) {\
+  return packet_type(from);\
+}
+
+// global space
+SYCL_PSET1(cl::sycl::cl_float4)
+SYCL_PSET1(cl::sycl::cl_double2)
+
+#undef SYCL_PSET1
+
+
+template <typename packet_type> struct get_base_packet {
+template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE packet_type get_ploaddup(sycl_multi_pointer ) {}
+
+  template <typename sycl_multi_pointer>
+    static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE packet_type get_pgather(sycl_multi_pointer , Index ) {}
+};
+
+template <> struct get_base_packet <cl::sycl::cl_float4>  {
+  template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_float4 get_ploaddup(sycl_multi_pointer from) {
+    return cl::sycl::cl_float4(from[0], from[0], from[1], from[1]);
+  }
+  template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_float4 get_pgather(sycl_multi_pointer from, Index stride) {
+    return cl::sycl::cl_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
+  }
+
+  template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void set_pscatter(sycl_multi_pointer to , const cl::sycl::cl_float4& from, Index stride) {
+    auto tmp = stride;
+    to[0] = from.x();
+    to[tmp] = from.y();
+    to[tmp += stride] = from.z();
+    to[tmp += stride] = from.w();
+ }
+ static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_float4 set_plset(const float& a) {
+   return  cl::sycl::cl_float4(static_cast<float>(a), static_cast<float>(a+1), static_cast<float>(a+2), static_cast<float>(a+3));
+ }
+};
+
+template <> struct get_base_packet <cl::sycl::cl_double2>  {
+  template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_double2 get_ploaddup(const sycl_multi_pointer from) {
+    return cl::sycl::cl_double2(from[0], from[0]);
+  }
+
+  template <typename sycl_multi_pointer, typename Index>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_double2 get_pgather(const sycl_multi_pointer from, Index stride) {
+    return cl::sycl::cl_double2(from[0*stride], from[1*stride]);
+  }
+
+  template <typename sycl_multi_pointer>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void set_pscatter(sycl_multi_pointer to , const cl::sycl::cl_double2& from, Index stride) {
+    to[0] = from.x();
+    to[stride] = from.y();
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE cl::sycl::cl_double2 set_plset(const double& a) {
+    return  cl::sycl::cl_double2(static_cast<double>(a), static_cast<double>(a + 1));
+  }
+};
+
+#define SYCL_PLOAD_DUP(address_space_target)\
+template<typename packet_type> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE packet_type \
+ploaddup(typename cl::sycl::multi_ptr<const typename unpacket_traits<packet_type>::type,\
+  cl::sycl::access::address_space::address_space_target>::pointer_t from)\
+{\
+  return get_base_packet<packet_type>::get_ploaddup(from); \
+}
+
+// global space
+SYCL_PLOAD_DUP(global_space)
+// local_space
+SYCL_PLOAD_DUP(local_space)
+// private_space
+//SYCL_PLOAD_DUP(private_space)
+#undef SYCL_PLOAD_DUP
+
+#define SYCL_PLOAD_DUP_SPECILIZE(packet_type)\
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE packet_type \
+ploaddup<packet_type>(const typename unpacket_traits<packet_type>::type * from)\
+{ \
+  return get_base_packet<packet_type>::get_ploaddup(from); \
+}
+
+SYCL_PLOAD_DUP_SPECILIZE(cl::sycl::cl_float4)
+SYCL_PLOAD_DUP_SPECILIZE(cl::sycl::cl_double2)
+
+#undef SYCL_PLOAD_DUP_SPECILIZE
+
+#define SYCL_PLSET(packet_type)\
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type plset<packet_type>(const typename unpacket_traits<packet_type>::type& a) {\
+  return get_base_packet<packet_type>::set_plset(a);\
+}
+
+SYCL_PLSET(cl::sycl::cl_float4)
+SYCL_PLSET(cl::sycl::cl_double2)
+
+#undef SYCL_PLSET
+
+
+#define SYCL_PGATHER(address_space_target)\
+template<typename Scalar, typename packet_type> EIGEN_DEVICE_FUNC inline packet_type pgather(\
+  typename cl::sycl::multi_ptr<const typename unpacket_traits<packet_type>::type,\
+  cl::sycl::access::address_space::address_space_target>::pointer_t from, Index stride) {\
+    return get_base_packet<packet_type>::get_pgather(from, stride); \
+}
+
+// global space
+SYCL_PGATHER(global_space)
+// local space
+SYCL_PGATHER(local_space)
+// private space
+//SYCL_PGATHER(private_space)
+
+#undef SYCL_PGATHER
+
+
+#define SYCL_PGATHER_SPECILIZE(scalar, packet_type)\
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE packet_type \
+pgather<scalar, packet_type>(const typename unpacket_traits<packet_type>::type * from, Index stride)\
+{ \
+  return get_base_packet<packet_type>::get_pgather(from, stride); \
+}
+
+SYCL_PGATHER_SPECILIZE(float, cl::sycl::cl_float4)
+SYCL_PGATHER_SPECILIZE(double, cl::sycl::cl_double2)
+
+#undef SYCL_PGATHER_SPECILIZE
+
+#define SYCL_PSCATTER(address_space_target)\
+template<typename Scalar, typename packet_type> EIGEN_DEVICE_FUNC inline void pscatter(\
+    typename cl::sycl::multi_ptr<typename unpacket_traits<packet_type>::type,\
+    cl::sycl::access::address_space::address_space_target>::pointer_t to,\
+      const packet_type& from, Index stride) {\
+      get_base_packet<packet_type>::set_pscatter(to, from, stride);\
+}
+
+// global space
+SYCL_PSCATTER(global_space)
+// local space
+SYCL_PSCATTER(local_space)
+// private space
+//SYCL_PSCATTER(private_space)
+
+#undef SYCL_PSCATTER
+
+
+
+#define SYCL_PSCATTER_SPECILIZE(scalar, packet_type)\
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void \
+pscatter<scalar, packet_type>(typename unpacket_traits<packet_type>::type * to, const packet_type& from, Index stride)\
+{ \
+   get_base_packet<packet_type>::set_pscatter(to, from, stride);\
+}
+
+SYCL_PSCATTER_SPECILIZE(float, cl::sycl::cl_float4)
+SYCL_PSCATTER_SPECILIZE(double, cl::sycl::cl_double2)
+
+#undef SYCL_PSCATTER_SPECILIZE
+
+
+#define SYCL_PMAD(packet_type)\
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE packet_type pmadd( const packet_type& a,\
+  const packet_type& b, const packet_type& c){\
+  return cl::sycl::mad(a,b,c);\
+}
+
+SYCL_PMAD(cl::sycl::cl_float4)
+SYCL_PMAD(cl::sycl::cl_double2)
+#undef SYCL_PMAD
+
+
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float  pfirst<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return a.x();
+}
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double pfirst<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return a.x();
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float  predux<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return a.x() + a.y() + a.z() + a.w();
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double predux<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return a.x() + a.y();
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float  predux_max<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return cl::sycl::fmax(cl::sycl::fmax(a.x(), a.y()), cl::sycl::fmax(a.z(), a.w()));
+}
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double predux_max<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return cl::sycl::fmax(a.x(), a.y());
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float  predux_min<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return cl::sycl::fmin(cl::sycl::fmin(a.x(), a.y()), cl::sycl::fmin(a.z(), a.w()));
+}
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double predux_min<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return cl::sycl::fmin(a.x(), a.y());
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float  predux_mul<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return a.x() * a.y() * a.z() * a.w();
+}
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double predux_mul<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return a.x() * a.y();
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_float4  pabs<cl::sycl::cl_float4>(const cl::sycl::cl_float4& a) {
+  return cl::sycl::cl_float4(cl::sycl::fabs(a.x()), cl::sycl::fabs(a.y()), cl::sycl::fabs(a.z()), cl::sycl::fabs(a.w()));
+}
+template<>  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_double2 pabs<cl::sycl::cl_double2>(const cl::sycl::cl_double2& a) {
+  return cl::sycl::cl_double2(cl::sycl::fabs(a.x()), cl::sycl::fabs(a.y()));
+}
+
+ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void
+ptranspose(PacketBlock<cl::sycl::cl_float4,4>& kernel) {
+  float tmp = kernel.packet[0].y();
+  kernel.packet[0].y() = kernel.packet[1].x();
+  kernel.packet[1].x() = tmp;
+//  std::swap(kernel.packet[0].y(), kernel.packet[1].x());
+
+  tmp = kernel.packet[0].z();
+  kernel.packet[0].z() = kernel.packet[2].x();
+  kernel.packet[2].x() = tmp;
+  //std::swap(kernel.packet[0].z(), kernel.packet[2].x());
+
+  tmp = kernel.packet[0].w();
+  kernel.packet[0].w() = kernel.packet[3].x();
+  kernel.packet[3].x() = tmp;
+
+  //std::swap(kernel.packet[0].w(), kernel.packet[3].x());
+
+  tmp = kernel.packet[1].z();
+  kernel.packet[1].z() = kernel.packet[2].y();
+  kernel.packet[2].y() = tmp;
+//  std::swap(kernel.packet[1].z(), kernel.packet[2].y());
+
+  tmp = kernel.packet[1].w();
+  kernel.packet[1].w() = kernel.packet[3].y();
+  kernel.packet[3].y() = tmp;
+//  std::swap(kernel.packet[1].w(), kernel.packet[3].y());
+
+  tmp = kernel.packet[2].w();
+  kernel.packet[2].w() = kernel.packet[3].z();
+  kernel.packet[3].z() = tmp;
+//  std::swap(kernel.packet[2].w(), kernel.packet[3].z());
+
+}
+
+ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void
+ptranspose(PacketBlock<cl::sycl::cl_double2,2>& kernel) {
+  double tmp = kernel.packet[0].y();
+  kernel.packet[0].y() = kernel.packet[1].x();
+  kernel.packet[1].x() = tmp;
+//std::swap(kernel.packet[0].y(), kernel.packet[1].x());
+}
+
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_float4
+pblend(const Selector<unpacket_traits<cl::sycl::cl_float4>::size>& ifPacket,
+  const cl::sycl::cl_float4& thenPacket, const cl::sycl::cl_float4& elsePacket) {
+  cl::sycl::cl_int4 condition(ifPacket.select[0] ? 0 : -1,
+                              ifPacket.select[1] ? 0 : -1,
+                              ifPacket.select[2] ? 0 : -1,
+                              ifPacket.select[3] ? 0 : -1);
+  return cl::sycl::select(thenPacket, elsePacket, condition);
+}
+
+template<> inline cl::sycl::cl_double2
+pblend(const Selector<unpacket_traits<cl::sycl::cl_double2>::size>& ifPacket,
+  const cl::sycl::cl_double2& thenPacket, const cl::sycl::cl_double2& elsePacket) {
+  cl::sycl::cl_long2 condition(ifPacket.select[0] ? 0 : -1,
+                               ifPacket.select[1] ? 0 : -1);
+  return cl::sycl::select(thenPacket, elsePacket, condition);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_USE_SYCL
+#endif // EIGEN_PACKET_MATH_SYCL_H
diff --git a/Eigen/src/Core/arch/SYCL/TypeCasting.h b/Eigen/src/Core/arch/SYCL/TypeCasting.h
new file mode 100644
index 000000000..dedd5c84a
--- /dev/null
+++ b/Eigen/src/Core/arch/SYCL/TypeCasting.h
@@ -0,0 +1,89 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * TypeCasting.h
+ *
+ * \brief:
+ *  TypeCasting
+ *
+*****************************************************************/
+
+#ifndef EIGEN_TYPE_CASTING_SYCL_H
+#define EIGEN_TYPE_CASTING_SYCL_H
+
+namespace Eigen {
+
+namespace internal {
+#ifdef __SYCL_DEVICE_ONLY__
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_int4 pcast<cl::sycl::cl_float4, cl::sycl::cl_int4>(const cl::sycl::cl_float4& a) {
+  return a. template convert<cl::sycl::cl_int, cl::sycl::rounding_mode::automatic>();
+}
+
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_float4 pcast<cl::sycl::cl_int4, cl::sycl::cl_float4>(const cl::sycl::cl_int4& a) {
+  return a. template convert<cl::sycl::cl_float, cl::sycl::rounding_mode::automatic>();
+}
+
+template <>
+struct type_casting_traits<double, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_float4 pcast<cl::sycl::cl_double2, cl::sycl::cl_float4>(const cl::sycl::cl_double2& a, const cl::sycl::cl_double2& b) {
+  auto a1=a. template convert<cl::sycl::cl_float, cl::sycl::rounding_mode::automatic>();
+  auto b1=b. template convert<cl::sycl::cl_float, cl::sycl::rounding_mode::automatic>();
+  return cl::sycl::float4(a1.x(), a1.y(), b1.x(), b1.y());
+}
+
+template <>
+struct type_casting_traits<float, double> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE cl::sycl::cl_double2 pcast<cl::sycl::cl_float4, cl::sycl::cl_double2>(const cl::sycl::cl_float4& a) {
+  // Simply discard the second half of the input
+  return cl::sycl::cl_double2(a.x(), a.y());
+}
+
+#endif
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_SYCL_H

From b512a9536f4b6260fd7af1d39f337eea8c6932cb Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 13:39:13 +0100
Subject: [PATCH 514/680] Enabling per device specialisation of packetsize.

---
 .../Eigen/CXX11/src/Tensor/TensorAssign.h     |  2 +-
 .../CXX11/src/Tensor/TensorBroadcasting.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   |  4 ++--
 .../CXX11/src/Tensor/TensorConcatenation.h    |  4 ++--
 .../CXX11/src/Tensor/TensorContraction.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorConversion.h |  2 +-
 .../CXX11/src/Tensor/TensorConvolution.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   |  8 +++----
 .../Eigen/CXX11/src/Tensor/TensorEvalTo.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  | 24 ++++++++++---------
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorGenerator.h  |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorInflation.h  |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorPadding.h    |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorReverse.h    |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  |  8 +++----
 .../Eigen/CXX11/src/Tensor/TensorStriding.h   |  4 ++--
 .../CXX11/src/Tensor/TensorVolumePatch.h      |  2 +-
 23 files changed, 48 insertions(+), 46 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
index 199ddb123..f1f877c16 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
@@ -98,7 +98,7 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef typename TensorEvaluator<RightArgType, Device>::Dimensions Dimensions;
 
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   static const int NumDims = XprType::NumDims;
 
   enum {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 8fecbe657..b4a77b022 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -104,7 +104,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   bool isCopy= false, nByOne = false, oneByN = false;
 
   enum {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 085c05f3d..3ab0a0f49 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -138,7 +138,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
 
   enum {
@@ -339,7 +339,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
index 9f0321880..27c92d8f6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@@ -251,7 +251,7 @@ struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgTy
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
-    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
     EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
     eigen_assert(index + packetSize - 1 < dimensions().TotalSize());
 
@@ -354,7 +354,7 @@ template<typename Axis, typename LeftArgType, typename RightArgType, typename De
   template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketReturnType& x)
   {
-    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
     EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
     eigen_assert(index + packetSize - 1 < this->dimensions().TotalSize());
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 86602c27e..e1649fb47 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -239,7 +239,7 @@ struct TensorContractionEvaluatorBase
 
   enum {
     IsAligned = true,
-    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
index e0cbbb315..a7751eee1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
@@ -190,7 +190,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
   typedef typename internal::remove_all<typename internal::traits<ArgType>::Scalar>::type SrcType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef typename PacketType<SrcType, Device>::type PacketSourceType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 1ec5819a7..0d3ca966c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -302,7 +302,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 0e4db46de..571922073 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -87,11 +87,11 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
   typedef typename internal::remove_const<typename ArgType::Scalar>::type Scalar;
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
-    PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<XprType, Device>::Layout,
     CoordAccess = false,  // to be implemented
@@ -249,11 +249,11 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
   typedef typename XprType::Scalar Scalar;
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
-    PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<LhsXprType, Device>::Layout,
     CoordAccess = false,  // to be implemented
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
index af39daa91..256d499f2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -102,7 +102,7 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType, MakePointer_>, Device>
   typedef typename XprType::Index Index;
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index f9a1bd68c..8f7a81575 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -33,6 +33,7 @@ struct TensorEvaluator
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   typedef typename Derived::Dimensions Dimensions;
   typedef Derived XprType;
+  static const int PacketSize =  PacketType<CoeffReturnType, Device>::size;
 
   // NumDimensions is -1 for variable dim tensors
   static const int NumCoords = internal::traits<Derived>::NumDimensions > 0 ?
@@ -40,7 +41,7 @@ struct TensorEvaluator
 
   enum {
     IsAligned = Derived::IsAligned,
-    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
@@ -121,7 +122,7 @@ struct TensorEvaluator
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
-                        internal::unpacket_traits<PacketReturnType>::size);
+                        PacketType<CoeffReturnType, Device>::size);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
@@ -188,10 +189,11 @@ struct TensorEvaluator<const Derived, Device>
   // NumDimensions is -1 for variable dim tensors
   static const int NumCoords = internal::traits<Derived>::NumDimensions > 0 ?
                                internal::traits<Derived>::NumDimensions : 0;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = Derived::IsAligned,
-    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
@@ -249,7 +251,7 @@ struct TensorEvaluator<const Derived, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
-                        internal::unpacket_traits<PacketReturnType>::size);
+                        PacketType<CoeffReturnType, Device>::size);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
@@ -300,7 +302,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); }
@@ -322,7 +324,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
   costPerCoeff(bool vectorized) const {
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
-                        internal::unpacket_traits<PacketReturnType>::size);
+                        PacketType<CoeffReturnType, Device>::size);
   }
 
   EIGEN_DEVICE_FUNC  typename Eigen::internal::traits<XprType>::PointerType  data() const { return NULL; }
@@ -367,7 +369,7 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); }
@@ -445,7 +447,7 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   typedef typename XprType::Scalar Scalar;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<LeftArgType, Device>::Dimensions Dimensions;
 
   static const int NumDims = internal::array_size<
@@ -574,7 +576,7 @@ struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type,
   typedef typename XprType::Scalar Scalar;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<Arg1Type, Device>::Dimensions Dimensions;
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
@@ -644,7 +646,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>
   enum {
     IsAligned = TensorEvaluator<ThenArgType, Device>::IsAligned & TensorEvaluator<ElseArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ThenArgType, Device>::PacketAccess & TensorEvaluator<ElseArgType, Device>::PacketAccess &
-                   internal::packet_traits<Scalar>::HasBlend,
+                    PacketType<Scalar, Device>::HasBlend,
     BlockAccess = false,
     Layout = TensorEvaluator<IfArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
@@ -665,7 +667,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>
   typedef typename XprType::Index Index;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<IfArgType, Device>::Dimensions Dimensions;
 
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index fdb31928f..5a6555cde 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -93,11 +93,11 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
   typedef typename XprType::Index Index;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = true,
-    PacketAccess = (PacketSize > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index fa269b8c6..97c8d4a02 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -90,7 +90,7 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   enum {
     IsAligned = false,
-    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
@@ -137,7 +137,7 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
-    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
     EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
     eigen_assert(index+packetSize-1 < dimensions().TotalSize());
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 72cb2d15f..00e1186e5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -181,7 +181,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   typedef TensorEvaluator<ArgType, Device> Impl;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
index 6147fbdf1..64f2ad81f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
@@ -85,7 +85,7 @@ struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 498488649..9a6431f29 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -423,7 +423,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
-    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
     EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
     eigen_assert(index+packetSize-1 < internal::array_prod(dimensions()));
 
@@ -584,7 +584,7 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
   template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketReturnType& x)
   {
-    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
     Index inputIndices[] = {0, 0};
     Index indices[] = {index, index + packetSize - 1};
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
index ffa22f31e..aa1db3c73 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
@@ -91,7 +91,7 @@ struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = true,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 950ac32af..a0a1ad8f4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -88,7 +88,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
 
   enum {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 375fc0802..bc09d3699 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -407,7 +407,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   static const bool InputPacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess;
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
index bb2768ab1..9193bdd8e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
@@ -108,7 +108,7 @@ struct TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>, Device
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
@@ -266,7 +266,7 @@ struct TensorEvaluator<TensorReverseOp<ReverseDimensions, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   const Dimensions& dimensions() const { return this->m_dimensions; }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index 39717efaa..b1135f297 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -95,7 +95,7 @@ struct TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> {
 
   enum {
     IsAligned = false,
-    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 6b54f40ad..0fc49255d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -107,11 +107,11 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
-    PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
@@ -236,11 +236,11 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
-    PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
index c09513c10..4b69072f2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
@@ -107,7 +107,7 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
@@ -287,7 +287,7 @@ struct TensorEvaluator<TensorStridingOp<Strides, ArgType>, Device>
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
   {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index ef199bfb6..3c7d8bbc0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -194,7 +194,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,

From af96018b499be64ff0b262cafc7b31f1a907b4c8 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 16:04:44 +0100
Subject: [PATCH 515/680] Using the suggested modification.

---
 Eigen/src/Core/util/Macros.h                  | 21 +++++--------------
 .../Eigen/CXX11/src/Tensor/TensorMap.h        |  4 ++--
 2 files changed, 7 insertions(+), 18 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 3255b8351..f59b93608 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1077,26 +1077,15 @@ namespace Eigen {
 #endif
 
 #ifdef EIGEN_HAS_VARIADIC_TEMPLATES
-// Provide a variadic version of assert which can take a parameter pack as its input
-// The eigen_assert macro used here might have been redefined to use other macros such as EIGEN_THROW, such as used in Eigen's test suite, therefore this needs to be defined after the other macros.
-// Note that this does not provide as nice a string to assert as a straight forward call to eigen_assert, so we add a message to the assert.
-#if defined(EIGEN_NO_DEBUG)
-#define eigen_variadic_assert(x)
-#else
+// The all function is used to enable a variadic version of eigen_assert which can take a parameter pack as its input.
 namespace Eigen {
 namespace internal {
-inline void variadic_assert(const char*) {}
-template<typename... Bools> inline void variadic_assert(const char* message, bool first, Bools ... others) {
-  eigen_assert(first && message);
-  variadic_assert(message, others...);
-  EIGEN_UNUSED_VARIABLE(first);
+bool all(){ return true; }
+template<typename T, typename ...Ts>
+bool all(T t, Ts ... ts){ return t && all(ts...); }
+
 }
 }
-}
-#define EIGEN_VARIADIC_ASSERT_MESSAGE(x) EIGEN_MAKESTRING(x) " in " __FILE__ ":" EIGEN_MAKESTRING(__LINE__)
-#define eigen_variadic_assert(x) \
-  do { Eigen::internal::variadic_assert(EIGEN_VARIADIC_ASSERT_MESSAGE(x), x); } while(false);
-#endif
 #endif
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
index f69f8f24a..d1cc0593f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h
@@ -150,7 +150,7 @@ template<typename PlainObjectType, int Options_, template <class> class MakePoin
     EIGEN_STRONG_INLINE const Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
     {
       EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
-      eigen_variadic_assert((Eigen::NumTraits<Index>::highest() >= otherIndices)...);
+      eigen_assert(internal::all((Eigen::NumTraits<Index>::highest() >= otherIndices)...));
       if (PlainObjectType::Options&RowMajor) {
         const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
         return m_data[index];
@@ -238,7 +238,7 @@ template<typename PlainObjectType, int Options_, template <class> class MakePoin
     EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices)
     {
       static_assert(sizeof...(otherIndices) + 2 == NumIndices || NumIndices == Dynamic, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor.");
-      eigen_variadic_assert((Eigen::NumTraits<Index>::highest() >= otherIndices)...);
+       eigen_assert(internal::all((Eigen::NumTraits<Index>::highest() >= otherIndices)...));
       const std::size_t NumDims = sizeof...(otherIndices) + 2;
       if (PlainObjectType::Options&RowMajor) {
         const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumDims>{{firstIndex, secondIndex, otherIndices...}});

From c6a5c70712851cd696d7410579506fc299c04a05 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 1 Aug 2018 16:56:26 +0100
Subject: [PATCH 516/680] Correcting the position of
 allocate_temp/deallocate_temp in TensorDeviceGpu.h

---
 .../Eigen/CXX11/src/Tensor/TensorDeviceGpu.h     | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
index 48bbd5e45..b490433db 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceGpu.h
@@ -159,13 +159,6 @@ class GpuStreamDevice : public StreamInterface {
     err = gpuFree(buffer);
     gpu_assert(err == gpuSuccess);
   }
-   EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
-    return stream_->allocate(num_bytes);
-  }
-
-  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
-    stream_->deallocate(buffer);
-  }
 
   virtual void* scratchpad() const {
     if (scratch_ == NULL) {
@@ -214,6 +207,15 @@ struct GpuDevice {
     stream_->deallocate(buffer);
   }
 
+  EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return stream_->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    stream_->deallocate(buffer);
+  }
+
+
   EIGEN_STRONG_INLINE void* scratchpad() const {
     return stream_->scratchpad();
   }

From 64abdf1d7eb17174f571751346dd0cbadcf3bc52 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 1 Aug 2018 12:35:19 -0700
Subject: [PATCH 517/680] Fix typo + get rid of redundant member variables for
 block sizes

---
 .../CXX11/src/Tensor/TensorBroadcasting.h      |  6 +++---
 .../Eigen/CXX11/src/Tensor/TensorChipping.h    | 17 ++++++++---------
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h  | 11 +++++------
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h    | 12 +++---------
 .../Eigen/CXX11/src/Tensor/TensorReduction.h   | 10 +++-------
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h   | 18 ++++++++++--------
 6 files changed, 32 insertions(+), 42 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index b6dbe5a22..cca14aafd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -120,7 +120,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   // Block based access to the XprType (input) tensor.
   using TensorBlock                = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
   using TensorBlockReader          = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
-  // We do block based broadcasting using a a trick with 2x tensor rank and 0
+  // We do block based broadcasting using a trick with 2x tensor rank and 0
   // strides. See block method implementation for details.
   using BroadcastDimensions        = DSizes<Index, 2 * NumDims>;
   using BroadcastTensorBlock       = internal::TensorBlock<ScalarNoConst, Index, 2 * NumDims, Layout>;
@@ -589,8 +589,8 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
     // TODO(wuke): Targeting L1 size is 30% faster than targeting L{-1} on large
     // tensors. But this might need further tuning.
-    Index l1_cache_scalars = m_device.firstLevelCacheSize() / sizeof(Scalar);
-    Index block_total_size_max = numext::maxi(Index(1), l1_cache_scalars);
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.firstLevelCacheSize() / sizeof(Scalar));
 
     resources->push_back(internal::TensorOpResourceRequirements(
         internal::TensorBlockShapeType::kSkewedInnerDims,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 7579ab507..aca2ead12 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -202,9 +202,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
           m_inputStrides[i] = m_inputStrides[i + 1] * input_dims[i + 1];
         }
       }
-
-      m_block_total_size_max =
-          numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
     }
   }
 
@@ -290,9 +287,11 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
         internal::TensorBlockShapeType::kSkewedInnerDims,
-        m_block_total_size_max));
+        block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
@@ -370,13 +369,14 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   {
     Index inputIndex;
     if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == 0) ||
-  (static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
+        (static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims - 1)) {
       // m_stride is equal to 1, so let's avoid the integer division.
       eigen_assert(m_stride == 1);
       inputIndex = index * m_inputStride + m_inputOffset;
-    } else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims-1) ||
-         (static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
-      // m_stride is aways greater than index, so let's avoid the integer division.
+    } else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims - 1) ||
+               (static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
+      // m_stride is aways greater than index, so let's avoid the integer
+      // division.
       eigen_assert(m_stride > index);
       inputIndex = index + m_inputOffset;
     } else {
@@ -392,7 +392,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   Index m_stride;
   Index m_inputOffset;
   Index m_inputStride;
-  Index m_block_total_size_max;
   DSizes<Index, NumInputDims> m_inputStrides;
   TensorEvaluator<ArgType, Device> m_impl;
   const internal::DimensionId<DimId> m_dim;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 39759b6c3..a8247be90 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -259,7 +259,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   #else
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
   #endif
-      : m_impl(op.expression(), device)
+      : m_device(device), m_impl(op.expression(), device)
 #ifdef EIGEN_USE_SYCL
       , m_op(op)
 #endif
@@ -404,9 +404,6 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
     } else {
       m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
     }
-
-    m_block_total_size_max =
-        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -551,9 +548,11 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
         internal::TensorBlockShapeType::kSkewedInnerDims,
-        m_block_total_size_max));
+        block_total_size_max));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
@@ -743,8 +742,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   internal::TensorIntDivisor<Index> m_fastOutputDepth;
 
   Scalar m_paddingValue;
-  Index m_block_total_size_max;
 
+  const Device& m_device;
   TensorEvaluator<ArgType, Device> m_impl;
   #ifdef EIGEN_USE_SYCL
   // Required for SYCL in order to construct the expression tree on the device
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 2630311b8..6ddded0bd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -560,9 +560,6 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
         m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
       }
     }
-
-    m_block_total_size_max =
-        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -672,9 +669,11 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
         internal::TensorBlockShapeType::kSkewedInnerDims,
-        m_block_total_size_max));
+        block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
@@ -761,7 +760,6 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   Dimensions m_dimensions;
   bool m_is_identity;
   const StartIndices m_offsets;
-  Index m_block_total_size_max;
 };
 
 
@@ -1047,9 +1045,6 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
         m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(degenerate ? 1 : m_outputStrides[i]);
       }
     }
-    m_block_total_size_max = numext::maxi(static_cast<std::size_t>(1),
-                                          device.lastLevelCacheSize() /
-                                          sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -1128,7 +1123,6 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
   DSizes<Index, NumDims> m_dimensions;
   DSizes<Index, NumDims> m_offsets; // offset in a flattened shape
   const Strides m_strides;
-  std::size_t m_block_total_size_max;
   //use by sycl
   const StartIndices m_exprStartIndices;
   //use by sycl
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index c41783106..73675e7dd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -572,9 +572,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
             : (static_cast<int>(Layout) == static_cast<int>(ColMajor))
                   ? m_preservedStrides[0]
                   : m_preservedStrides[NumOutputDims - 1];
-
-    m_block_total_size_max =
-        numext::maxi<Index>(1, device.lastLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -771,9 +768,11 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
         internal::TensorBlockShapeType::kSkewedInnerDims,
-        m_block_total_size_max));
+        block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
@@ -1204,9 +1203,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   // Indexed by reduced dimensions.
   array<Index, NumReducedDims> m_reducedDims;
 
-  // Block size for tiled (aka TensorBlock) evaluation.
-  Index m_block_total_size_max;
-
   // Evaluator for the input expression.
   TensorEvaluator<ArgType, Device> m_impl;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 77f47bf64..f94c1380d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -124,8 +124,11 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   using TensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
   using TensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-      : m_impl(op.expression(), device), m_shuffle(op.shufflePermutation())
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
+                                                        const Device& device)
+      : m_device(device),
+        m_impl(op.expression(), device),
+        m_shuffle(op.shufflePermutation())
   {
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
     const Shuffle& shuffle = op.shufflePermutation();
@@ -162,9 +165,6 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
     for (int i = 0; i < NumDims; ++i) {
       m_inputStrides[i] = m_unshuffledInputStrides[shuffle[i]];
     }
-
-    m_block_total_size_max =
-        numext::maxi<Index>(1, device.firstLevelCacheSize() / sizeof(Scalar));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -226,9 +226,10 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    auto block_total_size_max = numext::maxi<Eigen::Index>(
+        1, m_device.firstLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kUniformAllDims,
-        m_block_total_size_max));
+        internal::TensorBlockShapeType::kUniformAllDims, block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
@@ -384,7 +385,8 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
   array<Index, NumDims> m_inputStrides;
   array<Index, NumDims> m_unshuffledInputStrides;
-  Index m_block_total_size_max;
+
+  const Device& m_device;
   TensorEvaluator<ArgType, Device> m_impl;
   /// required by sycl
   Shuffle m_shuffle;

From 62169419aba9048c6c66c71ce4910241e594cbe1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 1 Aug 2018 23:35:34 +0200
Subject: [PATCH 518/680] Fix two regressions introduced in previous merges:
 bad usage of EIGEN_HAS_VARIADIC_TEMPLATES and linking issue.

---
 Eigen/src/Core/util/Macros.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index f59b93608..0f3b428c9 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1076,11 +1076,13 @@ namespace Eigen {
 #   endif
 #endif
 
-#ifdef EIGEN_HAS_VARIADIC_TEMPLATES
+#if EIGEN_HAS_VARIADIC_TEMPLATES
 // The all function is used to enable a variadic version of eigen_assert which can take a parameter pack as its input.
 namespace Eigen {
 namespace internal {
-bool all(){ return true; }
+
+inline bool all(){ return true; }
+
 template<typename T, typename ...Ts>
 bool all(T t, Ts ... ts){ return t && all(ts...); }
 

From 40d6d020a098d9f828bad5e2a1accfc5ee1a1289 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 2 Aug 2018 13:34:53 +0100
Subject: [PATCH 519/680] Fixing ambigous constructor error for Clang compiler.

---
 unsupported/test/cxx11_tensor_concatenation.cpp | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp
index 9189a609b..f53515b4e 100644
--- a/unsupported/test/cxx11_tensor_concatenation.cpp
+++ b/unsupported/test/cxx11_tensor_concatenation.cpp
@@ -50,7 +50,13 @@ static void test_static_dimension_failure()
       .reshape(Tensor<int, 3>::Dimensions(2, 3, 1))
       .concatenate(right, 0);
   Tensor<int, 2, DataLayout> alternative = left
-      .concatenate(right.reshape(Tensor<int, 2>::Dimensions{{{2, 3}}}), 0);
+   // Clang compiler break with {{{}}} with an ambigous error on copy constructor
+  // the variadic DSize constructor added for #ifndef EIGEN_EMULATE_CXX11_META_H.
+  // Solution:
+  // either the code should change to 
+  //  Tensor<int, 2>::Dimensions{{2, 3}}
+  // or Tensor<int, 2>::Dimensions{Tensor<int, 2>::Dimensions{{2, 3}}}
+      .concatenate(right.reshape(Tensor<int, 2>::Dimensions{{2, 3}}), 0);
 }
 
 template<int DataLayout>

From 516d2621b96a0c41ff999781b8bbcaa527a7c6ee Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 2 Aug 2018 14:30:48 +0100
Subject: [PATCH 520/680] fixing compilation error for cxx11_tensor_trace.cpp
 error on Microsoft Visual Studio.

---
 unsupported/test/cxx11_tensor_trace.cpp | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_trace.cpp b/unsupported/test/cxx11_tensor_trace.cpp
index 1579bc1eb..0cb23060e 100644
--- a/unsupported/test/cxx11_tensor_trace.cpp
+++ b/unsupported/test/cxx11_tensor_trace.cpp
@@ -37,7 +37,7 @@ static void test_all_dimensions_trace() {
   VERIFY_IS_EQUAL(result1(), sum);
 
   Tensor<float, 5, DataLayout> tensor2(7, 7, 7, 7, 7);
-  array<ptrdiff_t, 5> dims({{2, 1, 0, 3, 4}});
+  array<ptrdiff_t, 5> dims = { { 2, 1, 0, 3, 4 } };
   Tensor<float, 0, DataLayout> result2 = tensor2.trace(dims);
   VERIFY_IS_EQUAL(result2.rank(), 0);
   sum = 0.0f;
@@ -52,7 +52,7 @@ template <int DataLayout>
 static void test_simple_trace() {
   Tensor<float, 3, DataLayout> tensor1(3, 5, 3);
   tensor1.setRandom();
-  array<ptrdiff_t, 2> dims1({{0, 2}});
+  array<ptrdiff_t, 2> dims1 = { { 0, 2 } };
   Tensor<float, 1, DataLayout> result1 = tensor1.trace(dims1);
   VERIFY_IS_EQUAL(result1.rank(), 1);
   VERIFY_IS_EQUAL(result1.dimension(0), 5);
@@ -67,7 +67,7 @@ static void test_simple_trace() {
 
   Tensor<float, 4, DataLayout> tensor2(5, 5, 7, 7);
   tensor2.setRandom();
-  array<ptrdiff_t, 2> dims2({{2, 3}});
+  array<ptrdiff_t, 2> dims2 = { { 2, 3 } };
   Tensor<float, 2, DataLayout> result2 = tensor2.trace(dims2);
   VERIFY_IS_EQUAL(result2.rank(), 2);
   VERIFY_IS_EQUAL(result2.dimension(0), 5);
@@ -82,7 +82,7 @@ static void test_simple_trace() {
     }
   }
 
-  array<ptrdiff_t, 2> dims3({{1, 0}});
+  array<ptrdiff_t, 2> dims3 = { { 1, 0 } };
   Tensor<float, 2, DataLayout> result3 = tensor2.trace(dims3);
   VERIFY_IS_EQUAL(result3.rank(), 2);
   VERIFY_IS_EQUAL(result3.dimension(0), 7);
@@ -99,7 +99,7 @@ static void test_simple_trace() {
 
   Tensor<float, 5, DataLayout> tensor3(3, 7, 3, 7, 3);
   tensor3.setRandom();
-  array<ptrdiff_t, 3> dims4({{0, 2, 4}});
+  array<ptrdiff_t, 3> dims4 = { { 0, 2, 4 } };
   Tensor<float, 2, DataLayout> result4 = tensor3.trace(dims4);
   VERIFY_IS_EQUAL(result4.rank(), 2);
   VERIFY_IS_EQUAL(result4.dimension(0), 7);
@@ -116,7 +116,7 @@ static void test_simple_trace() {
 
   Tensor<float, 5, DataLayout> tensor4(3, 7, 4, 7, 5);
   tensor4.setRandom();
-  array<ptrdiff_t, 2> dims5({{1, 3}});
+  array<ptrdiff_t, 2> dims5 = { { 1, 3 } };
   Tensor<float, 3, DataLayout> result5 = tensor4.trace(dims5);
   VERIFY_IS_EQUAL(result5.rank(), 3);
   VERIFY_IS_EQUAL(result5.dimension(0), 3);
@@ -140,7 +140,7 @@ template<int DataLayout>
 static void test_trace_in_expr() {
   Tensor<float, 4, DataLayout> tensor(2, 3, 5, 3);
   tensor.setRandom();
-  array<ptrdiff_t, 2> dims({{1, 3}});
+  array<ptrdiff_t, 2> dims = { { 1, 3 } };
   Tensor<float, 2, DataLayout> result(2, 5);
   result = result.constant(1.0f) - tensor.trace(dims);
   VERIFY_IS_EQUAL(result.rank(), 2);
@@ -168,4 +168,4 @@ EIGEN_DECLARE_TEST(cxx11_tensor_trace) {
   CALL_SUBTEST(test_simple_trace<RowMajor>());
   CALL_SUBTEST(test_trace_in_expr<ColMajor>());
   CALL_SUBTEST(test_trace_in_expr<RowMajor>());
-}
+}
\ No newline at end of file

From 2bf1cc8cf72396c8c0c8103a5e941121534cf858 Mon Sep 17 00:00:00 2001
From: Gustavo Lima Chaves <gustavo.lima.chaves@intel.com>
Date: Thu, 2 Aug 2018 15:55:36 -0700
Subject: [PATCH 521/680] Fix 256 bit packet size assumptions in unit tests.

Like in change 2606abed535744fcaa41b923c71338a06b8ed3fa
, we have hit the threshould again. With
AVX512 builds we would never have Vector8f packets aligned at 64
bytes (the new value of EIGEN_MAX_ALIGN_BYTES after change 405859f18dac56f324e1d93ca8721d5f7fd22c62
,
for AVX512-enabled builds).

This makes test/dynalloc.cpp pass for those builds.
---
 test/dynalloc.cpp | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/test/dynalloc.cpp b/test/dynalloc.cpp
index ceecd76e3..1c74866ba 100644
--- a/test/dynalloc.cpp
+++ b/test/dynalloc.cpp
@@ -15,6 +15,7 @@
 #define ALIGNMENT 1
 #endif
 
+typedef Matrix<float,16,1> Vector16f;
 typedef Matrix<float,8,1> Vector8f;
 
 void check_handmade_aligned_malloc()
@@ -70,7 +71,7 @@ struct MyStruct
 {
   EIGEN_MAKE_ALIGNED_OPERATOR_NEW
   char dummychar;
-  Vector8f avec;
+  Vector16f avec;
 };
 
 class MyClassA
@@ -78,7 +79,7 @@ class MyClassA
   public:
     EIGEN_MAKE_ALIGNED_OPERATOR_NEW
     char dummychar;
-    Vector8f avec;
+    Vector16f avec;
 };
 
 template<typename T> void check_dynaligned()
@@ -145,6 +146,7 @@ EIGEN_DECLARE_TEST(dynalloc)
     CALL_SUBTEST(check_dynaligned<Vector4d>() );
     CALL_SUBTEST(check_dynaligned<Vector4i>() );
     CALL_SUBTEST(check_dynaligned<Vector8f>() );
+    CALL_SUBTEST(check_dynaligned<Vector16f>() );
   }
 
   {

From 01358300d581f94c842bcdf35266b50f2795ded3 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Fri, 3 Aug 2018 16:59:15 +0100
Subject: [PATCH 522/680] Creating separate SYCL required PR for
 uncontroversial files.

---
 Eigen/Core                                    |  6 ++++
 Eigen/src/Core/arch/GPU/Half.h                |  7 +++-
 .../Eigen/CXX11/src/Tensor/TensorTrace.h      |  4 +--
 .../Eigen/CXX11/src/Tensor/TensorTraits.h     |  4 +++
 .../SpecialFunctions/SpecialFunctionsImpl.h   | 36 ++++++++++++++++---
 5 files changed, 50 insertions(+), 7 deletions(-)

diff --git a/Eigen/Core b/Eigen/Core
index 864bde551..080511fc9 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -200,6 +200,12 @@ using std::ptrdiff_t;
   #include "src/Core/arch/GPU/MathFunctions.h"
 #endif
 
+#if defined EIGEN_VECTORIZE_SYCL
+  #include "src/Core/arch/SYCL/InteropHeaders.h"
+  #include "src/Core/arch/SYCL/PacketMath.h"
+  #include "src/Core/arch/SYCL/MathFunctions.h"
+  #include "src/Core/arch/SYCL/TypeCasting.h"
+#endif
 #include "src/Core/arch/Default/Settings.h"
 
 #include "src/Core/functors/TernaryFunctors.h"
diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index ab9d27591..aca56fa72 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -83,7 +83,11 @@ struct __half_raw {
  #if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
 // In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
  typedef __half __half_raw;
- #endif
+ #endif // defined(EIGEN_HAS_CUDA_FP16)
+
+#elif defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+typedef cl::sycl::half __half_raw;
+
 #endif
 
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
@@ -200,6 +204,7 @@ struct half : public half_impl::half_base {
     x = other.x;
     return *this;
   }
+
 };
 
 } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
index c8b2fad1e..ea53bb04b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@@ -273,11 +273,11 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
 
   Dimensions m_dimensions;
   TensorEvaluator<ArgType, Device> m_impl;
+  // Initialize the size of the trace dimension
+  Index m_traceDim;
   const Device& m_device;
   array<bool, NumInputDims> m_reduced;
   array<Index, NumReducedDims> m_reducedDims;
-  // Initialize the size of the trace dimension
-  Index m_traceDim;
   array<Index, NumOutputDims> m_outputStrides;
   array<Index, NumReducedDims> m_reducedStrides;
   array<Index, NumOutputDims> m_preservedStrides;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
index 006b37921..0a394c88d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h
@@ -59,6 +59,7 @@ struct traits<Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
   template <typename T> struct MakePointer {
     typedef T* Type;
     typedef T& RefType;
+    typedef T ScalarType;
 
   };
   typedef typename MakePointer<Scalar>::Type PointerType;
@@ -80,6 +81,7 @@ struct traits<TensorFixedSize<Scalar_, Dimensions, Options_, IndexType_> >
   template <typename T> struct MakePointer {
     typedef T* Type;
     typedef T& RefType;
+    typedef T ScalarType;
 
   };
   typedef typename MakePointer<Scalar>::Type PointerType;
@@ -105,6 +107,8 @@ struct traits<TensorMap<PlainObjectType, Options_, MakePointer_> >
     typedef MakePointer_<T> MakePointerT;
     typedef typename MakePointerT::Type Type;
     typedef typename MakePointerT::RefType RefType;
+    typedef typename MakePointerT::ScalarType ScalarType;
+
 
   };
   typedef typename MakePointer<Scalar>::Type PointerType;
diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index dbcc9d8ac..5784cbc86 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -193,6 +193,8 @@ struct lgamma_impl<float> {
 #if !defined(EIGEN_GPU_COMPILE_PHASE) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) 
     int dummy;
     return ::lgammaf_r(x, &dummy);
+#elif defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::lgamma(x);
 #else
     return ::lgammaf(x);
 #endif
@@ -206,6 +208,8 @@ struct lgamma_impl<double> {
 #if !defined(EIGEN_GPU_COMPILE_PHASE) && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE)) && !defined(__APPLE__) 
     int dummy;
     return ::lgamma_r(x, &dummy);
+#elif defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::lgamma(x);
 #else
     return ::lgamma(x);
 #endif
@@ -423,13 +427,25 @@ struct erf_retval {
 template <>
 struct erf_impl<float> {
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE float run(float x) { return ::erff(x); }
+  static EIGEN_STRONG_INLINE float run(float x) {
+#if defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::erf(x);
+#else
+    return ::erff(x);
+#endif
+  }
 };
 
 template <>
 struct erf_impl<double> {
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE double run(double x) { return ::erf(x); }
+  static EIGEN_STRONG_INLINE double run(double x) {
+#if defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::erf(x);
+#else
+    return ::erf(x);
+#endif
+  }
 };
 #endif  // EIGEN_HAS_C99_MATH
 
@@ -456,13 +472,25 @@ struct erfc_retval {
 template <>
 struct erfc_impl<float> {
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE float run(const float x) { return ::erfcf(x); }
+  static EIGEN_STRONG_INLINE float run(const float x) {
+#if defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::erfc(x);
+#else
+    return ::erfcf(x);
+#endif
+  }
 };
 
 template <>
 struct erfc_impl<double> {
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE double run(const double x) { return ::erfc(x); }
+  static EIGEN_STRONG_INLINE double run(const double x) {
+#if defined(EIGEN_USE_SYCL) && defined(__SYCL_DEVICE_ONLY__)
+    return cl::sycl::erfc(x);
+#else
+    return ::erfc(x);
+#endif
+  }
 };
 #endif  // EIGEN_HAS_C99_MATH
 

From 3074b1ff9e61a14384accdbda7c8b3b520140360 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Fri, 3 Aug 2018 17:13:44 +0100
Subject: [PATCH 523/680] Fixing the compilation error.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 8ed1796df..ec1dc0fab 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -21,6 +21,7 @@ namespace Eigen {
 template<typename T> struct MakePointer {
   typedef T* Type;
   typedef T& RefType;
+  typedef T ScalarType;
 };
 
 namespace internal{

From bcb29f890ccdbf4922780ed5da0e23db65d7ae64 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 3 Aug 2018 10:18:53 -0700
Subject: [PATCH 524/680] Fix initialization order.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index e1649fb47..e604456e8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -274,8 +274,8 @@ struct TensorContractionEvaluatorBase
                           op.lhsExpression(), op.rhsExpression()), device),
     m_rightImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
                           op.rhsExpression(), op.lhsExpression()), device),
-        m_output_kernel(op.outputKernel()),
         m_device(device),
+        m_output_kernel(op.outputKernel()),
         m_result(NULL) {
     EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
          static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),

From 7f8b53fd0e359aa2b1a588489f31135a8c52b55f Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 1 Aug 2018 12:36:24 -0700
Subject: [PATCH 525/680] bug #1580: Fix cuda clang build. STL is not
 supported, so std::equal_to and std::not_equal breaks compilation. Update the
 definition of EIGEN_CONSTEXPR_ARE_DEVICE_FUNC to exclude clang. See also PR
 450.

---
 Eigen/src/Core/util/Macros.h | 20 +++++++++++---------
 Eigen/src/Core/util/Meta.h   |  2 +-
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 0f3b428c9..9a2f5ab05 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -571,20 +571,19 @@
 
 // Does the compiler fully support const expressions? (as in c++14)
 #ifndef EIGEN_HAS_CONSTEXPR
-
   #if defined(EIGEN_CUDACC)
   // Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
-  #if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500))
-    #define EIGEN_HAS_CONSTEXPR 1
-  #endif
+    #if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500))
+      #define EIGEN_HAS_CONSTEXPR 1
+    #endif
   #elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
     (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L)) || \
     (EIGEN_COMP_CLANG >= 306 && (__cplusplus > 199711L)))
-  #define EIGEN_HAS_CONSTEXPR 1
+    #define EIGEN_HAS_CONSTEXPR 1
   #endif
 
   #ifndef EIGEN_HAS_CONSTEXPR
-  #define EIGEN_HAS_CONSTEXPR 0
+    #define EIGEN_HAS_CONSTEXPR 0
   #endif
 
 #endif // EIGEN_HAS_CONSTEXPR
@@ -643,9 +642,12 @@
     #ifdef __CUDACC_RELAXED_CONSTEXPR__
       #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
     #endif
-  #elif defined(__clang__) && defined(__CUDA__)
-    // clang++ always considers constexpr functions as implicitly __host__ __device__
-    #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
+    // See bug 1580: clang/CUDA fails to make the following calls
+    // to constexpr bool std::equal_to::operator() even when
+    // EIGEN_CONSTEXPR_ARE_DEVICE_FUNC is defined in c++14 only.
+    // #elif defined(__clang__) && defined(__CUDA__) && EIGEN_HAS_CONSTEXPR == 1
+    //   // clang++ always considers constexpr functions as implicitly __host__ __device__
+    //   #define EIGEN_CONSTEXPR_ARE_DEVICE_FUNC
   #endif
 #endif
 
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index 658cfa9eb..f27b8e85d 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -569,7 +569,7 @@ template<typename T, typename U> struct scalar_product_traits
 } // end namespace internal
 
 namespace numext {
-  
+
 #if defined(EIGEN_GPU_COMPILE_PHASE)
 template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
 #else

From 1b0373ae10687ecc51ad9a0bfd46aa4ee116ade1 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Wed, 1 Aug 2018 15:55:46 -0700
Subject: [PATCH 526/680] Replace all using declarations with typedefs in
 Tensor ops

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 20 +++++-----
 .../CXX11/src/Tensor/TensorBroadcasting.h     | 17 +++++---
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   | 16 +++++---
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   | 40 ++++++++++---------
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |  7 ++--
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   | 22 ++++++----
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 14 ++++---
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  | 16 +++++---
 8 files changed, 88 insertions(+), 64 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 33c4ef5b7..1db8d6124 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -152,11 +152,11 @@ struct TensorBlockCopyOp {
     const Scalar* src_base = &src_data[src_index];
     Scalar* dst_base = &dst_data[dst_index];
 
-    using Src = const Eigen::Array<Scalar, Dynamic, 1>;
-    using Dst = Eigen::Array<Scalar, Dynamic, 1>;
+    typedef const Eigen::Array<Scalar, Dynamic, 1> Src;
+    typedef Eigen::Array<Scalar, Dynamic, 1> Dst;
 
-    using SrcMap = Eigen::Map<Src, 0, InnerStride<>>;
-    using DstMap = Eigen::Map<Dst, 0, InnerStride<>>;
+    typedef Eigen::Map<Src, 0, InnerStride<>> SrcMap;
+    typedef Eigen::Map<Dst, 0, InnerStride<>> DstMap;
 
     const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
     DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
@@ -401,13 +401,13 @@ struct TensorBlockCwiseBinaryOp {
       const StorageIndex left_stride, const LeftScalar* left_data,
       const StorageIndex right_index, const StorageIndex right_stride,
       const RightScalar* right_data) {
-    using Lhs = const Eigen::Array<LeftScalar, Dynamic, 1>;
-    using Rhs = const Eigen::Array<RightScalar, Dynamic, 1>;
-    using Out = Eigen::Array<OutputScalar, Dynamic, 1>;
+    typedef const Eigen::Array<LeftScalar, Dynamic, 1> Lhs;
+    typedef const Eigen::Array<RightScalar, Dynamic, 1> Rhs;
+    typedef Eigen::Array<OutputScalar, Dynamic, 1> Out;
 
-    using LhsMap = Eigen::Map<Lhs, 0, InnerStride<>>;
-    using RhsMap = Eigen::Map<Rhs, 0, InnerStride<>>;
-    using OutMap = Eigen::Map<Out, 0, InnerStride<>>;
+    typedef Eigen::Map<Lhs, 0, InnerStride<>> LhsMap;
+    typedef Eigen::Map<Rhs, 0, InnerStride<>> RhsMap;
+    typedef Eigen::Map<Out, 0, InnerStride<>> OutMap;
 
     const LeftScalar* lhs_base = &left_data[left_index];
     const RightScalar* rhs_base = &right_data[right_index];
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index cca14aafd..a4d750885 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -115,16 +115,21 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
   // Block based access to the XprType (input) tensor.
-  using TensorBlock                = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
-  using TensorBlockReader          = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
+  typedef internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlockReader;
+
   // We do block based broadcasting using a trick with 2x tensor rank and 0
   // strides. See block method implementation for details.
-  using BroadcastDimensions        = DSizes<Index, 2 * NumDims>;
-  using BroadcastTensorBlock       = internal::TensorBlock<ScalarNoConst, Index, 2 * NumDims, Layout>;
-  using BroadcastTensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, 2 * NumDims, Layout>;
+  typedef DSizes<Index, 2 * NumDims> BroadcastDimensions;
+  typedef internal::TensorBlock<ScalarNoConst, Index, 2 * NumDims, Layout>
+      BroadcastTensorBlock;
+  typedef internal::TensorBlockReader<ScalarNoConst, Index, 2 * NumDims, Layout>
+      BroadcastTensorBlockReader;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
                                                         const Device& device)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index aca2ead12..b4c4162ef 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -152,10 +152,12 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using InputTensorBlock  = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
-  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>
+      InputTensorBlock;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      OutputTensorBlock;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device), m_offset(op.offset())
@@ -426,10 +428,12 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using InputTensorBlock  = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
-  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>
+      InputTensorBlock;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      OutputTensorBlock;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
     : Base(op, device)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 17008917a..f11241f83 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -36,7 +36,7 @@ template <typename Expression, typename Device, bool Vectorizable,
           bool Tileable>
 class TensorExecutor {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
@@ -60,7 +60,7 @@ template <typename Expression>
 class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true,
                      /*Tileable*/ false> {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
@@ -102,21 +102,22 @@ template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, DefaultDevice, Vectorizable,
                      /*Tileable*/ true> {
  public:
-  using Scalar = typename traits<Expression>::Scalar;
-  using ScalarNoConst = typename remove_const<Scalar>::type;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
 
-  using Evaluator = TensorEvaluator<Expression, DefaultDevice>;
-  using StorageIndex = typename traits<Expression>::Index;
+  typedef TensorEvaluator<Expression, DefaultDevice> Evaluator;
+  typedef typename traits<Expression>::Index StorageIndex;
 
   static const int NumDims = traits<Expression>::NumDimensions;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const DefaultDevice& device = DefaultDevice()) {
-    using TensorBlock =
-        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
-    using TensorBlockMapper = TensorBlockMapper<ScalarNoConst, StorageIndex,
-                                                NumDims, Evaluator::Layout>;
+    typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>
+        TensorBlock;
+    typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims,
+                              Evaluator::Layout>
+        TensorBlockMapper;
 
     Evaluator evaluator(expr, device);
     Index total_size = array_prod(evaluator.dimensions());
@@ -221,7 +222,7 @@ struct EvalRange<Evaluator, StorageIndex, /*Vectorizable*/ true> {
 template <typename Expression, bool Vectorizable, bool Tileable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
@@ -249,20 +250,21 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
 template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ true> {
  public:
-  using Scalar = typename traits<Expression>::Scalar;
-  using ScalarNoConst = typename remove_const<Scalar>::type;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
 
-  using Evaluator = TensorEvaluator<Expression, ThreadPoolDevice>;
-  using StorageIndex = typename traits<Expression>::Index;
+  typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+  typedef typename traits<Expression>::Index StorageIndex;
 
   static const int NumDims = traits<Expression>::NumDimensions;
 
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
-    using TensorBlock =
-        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
-    using TensorBlockMapper =
-        TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
+    typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>
+        TensorBlock;
+    typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims,
+                              Evaluator::Layout>
+        TensorBlockMapper;
 
     Evaluator evaluator(expr, device);
     StorageIndex total_size = array_prod(evaluator.dimensions());
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index a8247be90..2b2f4a650 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -252,7 +252,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
     RawAccess    = false
   };
 
-  using OutputTensorBlock = internal::TensorBlock<Scalar, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<Scalar, Index, NumDims, Layout>
+      OutputTensorBlock;
 
 #ifdef __SYCL_DEVICE_ONLY__
     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType op, const Device& device)
@@ -557,8 +558,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
       OutputTensorBlock* output_block) const {
-    using ImagePatchCopyOp = internal::ImagePatchCopyOp<Self, PacketAccess>;
-    using ImagePatchPaddingOp = internal::ImagePatchPaddingOp<Self>;
+    typedef internal::ImagePatchCopyOp<Self, PacketAccess> ImagePatchCopyOp;
+    typedef internal::ImagePatchPaddingOp<Self> ImagePatchPaddingOp;
 
     // Calculate loop limits and various input/output dim sizes.
     const DSizes<Index, NumDims>& block_sizes = output_block->block_sizes();
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 6ddded0bd..d5b0c1237 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -123,11 +123,15 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
     RawAccess    = TensorEvaluator<ArgType, Device>::RawAccess
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using InputTensorBlock        = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
-  using OutputTensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>;
-  using OutputTensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, NumOutputDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>
+      InputTensorBlock;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>
+      OutputTensorBlock;
+  typedef internal::TensorBlockReader<ScalarNoConst, Index, NumOutputDims,
+                                      Layout>
+      OutputTensorBlockReader;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_dimensions(op.dimensions())
@@ -512,9 +516,10 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using TensorBlock =  internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices())
@@ -787,9 +792,10 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
     RawAccess    = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using TensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
     : Base(op, device)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 73675e7dd..80f179ba4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -483,10 +483,12 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     RawAccess    = false
   };
 
-  using ScalarNoConst =  typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using OutputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>;
-  using InputTensorBlock = internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumOutputDims, Layout>
+      OutputTensorBlock;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>
+      InputTensorBlock;
 
   static const bool ReducingInnerMostDims = internal::are_inner_most_dims<Dims, NumInputDims, Layout>::value;
   static const bool PreservingInnerMostDims = internal::preserve_inner_most_dims<Dims, NumInputDims, Layout>::value;
@@ -901,9 +903,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
           new (&reducers[i]) BlockReducer(m_reducer);
         }
 
-        using TensorSliceBlockMapper =
-            internal::TensorSliceBlockMapper<ScalarNoConst, Index, NumInputDims,
-                                             Layout>;
+        typedef internal::TensorSliceBlockMapper<ScalarNoConst, Index,
+                                                 NumInputDims, Layout>
+            TensorSliceBlockMapper;
 
         // TODO(andydavis) Consider removing 'input_block_stride_order' if we
         // find that scattered reads are not worth supporting in
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index f94c1380d..fbe69aabc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -119,10 +119,12 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using TensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
-  using TensorBlockReader = internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
+  typedef internal::TensorBlockReader<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlockReader;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
                                                         const Device& device)
@@ -417,10 +419,12 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
     RawAccess    = false
   };
 
-  using ScalarNoConst = typename internal::remove_const<Scalar>::type;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  using TensorBlock       = internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>;
-  using TensorBlockWriter = internal::TensorBlockWriter<ScalarNoConst, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
+  typedef internal::TensorBlockWriter<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlockWriter;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : Base(op, device)

From ec38f07b799e2e3b82921ac3a5839f90ea3333c4 Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Thu, 6 Sep 2018 15:40:09 -0700
Subject: [PATCH 527/680] bug #1595: Don't use C++11's std::isnan() in MIPS/MSA
 packet math.

This removes reliance on C++11 and improves generated code.
---
 Eigen/src/Core/arch/MSA/PacketMath.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/MSA/PacketMath.h b/Eigen/src/Core/arch/MSA/PacketMath.h
index 94c15d132..094c874ee 100644
--- a/Eigen/src/Core/arch/MSA/PacketMath.h
+++ b/Eigen/src/Core/arch/MSA/PacketMath.h
@@ -1178,7 +1178,7 @@ EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) {
   return v[0];
 #else
   double a0 = a[0], a1 = a[1];
-  return ((std::isnan)(a0) || a0 < a1) ? a0 : a1;
+  return ((numext::isnan)(a0) || a0 < a1) ? a0 : a1;
 #endif
 }
 
@@ -1193,7 +1193,7 @@ EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) {
   return v[0];
 #else
   double a0 = a[0], a1 = a[1];
-  return ((std::isnan)(a0) || a0 > a1) ? a0 : a1;
+  return ((numext::isnan)(a0) || a0 > a1) ? a0 : a1;
 #endif
 }
 

From edfb7962fdae447125e188a03bfea33b47df010a Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 7 Sep 2018 14:08:39 +0200
Subject: [PATCH 528/680] Use `static const int` instead of `enum` to avoid
 numerous `local-type-template-args` warnings in C++03 mode

---
 unsupported/Eigen/src/EulerAngles/EulerSystem.h | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/src/EulerAngles/EulerSystem.h b/unsupported/Eigen/src/EulerAngles/EulerSystem.h
index 65c2e94c7..88acabcf8 100644
--- a/unsupported/Eigen/src/EulerAngles/EulerSystem.h
+++ b/unsupported/Eigen/src/EulerAngles/EulerSystem.h
@@ -173,15 +173,14 @@ namespace Eigen
     EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT((unsigned)BetaAxisAbs != (unsigned)GammaAxisAbs,
       BETA_AXIS_CANT_BE_EQUAL_TO_GAMMA_AXIS);
 
-    enum
-    {
+    static const int
       // I, J, K are the pivot indexes permutation for the rotation matrix, that match this Euler system. 
       // They are used in this class converters.
       // They are always different from each other, and their possible values are: 0, 1, or 2.
       I = AlphaAxisAbs - 1,
       J = (AlphaAxisAbs - 1 + 1 + IsOdd)%3,
       K = (AlphaAxisAbs - 1 + 2 - IsOdd)%3
-    };
+    ;
     
     // TODO: Get @mat parameter in form that avoids double evaluation.
     template <typename Derived>

From 6dcd2642aae67d0bc46eee959a9f196e6ea51e7c Mon Sep 17 00:00:00 2001
From: Jiandong Ruan <jd_ruan@hotmail.com>
Date: Sat, 8 Sep 2018 12:05:33 -0700
Subject: [PATCH 529/680] bug #1526 - CUDA compilation fails on CUDA 9.x SDK
 when arch is set to compute_60 and/or above

---
 Eigen/src/Core/arch/GPU/Half.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index b44a4356a..aef1eb079 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -272,7 +272,11 @@ namespace half_impl {
 // conversion steps back and forth.
 
 EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
+#if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  return __hadd(::__half(a), ::__half(b));
+#else
   return __hadd(a, b);
+#endif
 }
 EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
   return __hmul(a, b);
@@ -281,9 +285,13 @@ EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
   return __hsub(a, b);
 }
 EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
+#if defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  return __hdiv(a, b);
+#else
   float num = __half2float(a);
   float denom = __half2float(b);
   return __float2half(num / denom);
+#endif
 }
 EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
   return __hneg(a);

From a80a290079499a12246939d79e955dbce5f6fc4a Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Mon, 10 Sep 2018 18:57:28 +0200
Subject: [PATCH 530/680] Fix 'template argument uses local type'-warnings
 (when compiled in C++03 mode)

---
 test/indexed_view.cpp  | 10 ++++++++++
 test/integer_types.cpp |  4 ++--
 2 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index 7219c777f..e6ffe5a74 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -77,6 +77,9 @@ is_same_seq_type(const T1& a, const T2& b)
 
 #define VERIFY_EQ_INT(A,B) VERIFY_IS_APPROX(int(A),int(B))
 
+// C++03 does not allow local or unnamed enums as index
+enum DummyEnum { XX=0, YY=1 };
+
 void check_indexed_view()
 {
   using Eigen::placeholders::all;
@@ -375,9 +378,16 @@ void check_indexed_view()
   }
 
   // Check compilation of enums as index type:
+  a(XX) = 1;
+  A(XX,YY) = 1;
+  // Anonymous enums only work with C++11
+#if EIGEN_HAS_CXX11
   enum { X=0, Y=1 };
   a(X) = 1;
   A(X,Y) = 1;
+  A(XX,Y) = 1;
+  A(X,YY) = 1;
+#endif
 
   // Check compilation of varying integer types as index types:
   Index i = n/2;
diff --git a/test/integer_types.cpp b/test/integer_types.cpp
index e9e514f12..3f9030d77 100644
--- a/test/integer_types.cpp
+++ b/test/integer_types.cpp
@@ -134,8 +134,8 @@ template<typename MatrixType> void integer_type_tests(const MatrixType& m)
 template<int>
 void integer_types_extra()
 {
-  VERIFY_IS_EQUAL(internal::scalar_div_cost<int>::value, 8);
-  VERIFY_IS_EQUAL(internal::scalar_div_cost<unsigned int>::value, 8);
+  VERIFY_IS_EQUAL(int(internal::scalar_div_cost<int>::value), 8);
+  VERIFY_IS_EQUAL(int(internal::scalar_div_cost<unsigned int>::value), 8);
   if(sizeof(long)>sizeof(int)) {
     VERIFY(int(internal::scalar_div_cost<long>::value) > int(internal::scalar_div_cost<int>::value));
     VERIFY(int(internal::scalar_div_cost<unsigned long>::value) > int(internal::scalar_div_cost<int>::value));

From 2cbd9dd49806d686a40841b6d888a83c816efccf Mon Sep 17 00:00:00 2001
From: Jean-Christophe Fillion-Robin <jchris.fillionr@kitware.com>
Date: Fri, 7 Sep 2018 15:50:19 -0400
Subject: [PATCH 531/680] [PATCH] cmake: Support source include with
 add_subdirectory and  find_package use This commit allows the sources of the
 project to be included in a parent project CMakeLists.txt and support use of
 "find_package(Eigen3 CONFIG REQUIRED)"

Here is an example allowing to test the changes. It is not particularly
useful in itself. This change will allow to support one of the scenario
allowing to create custom 3D Slicer application bundling associated plugins.

/tmp/eigen-git-mirror  # Eigen sources

/tmp/test/CMakeLists.txt:

  cmake_minimum_required(VERSION 3.12)
  project(test)
  add_subdirectory("/tmp/eigen-git-mirror" "eigen-git-mirror")
  find_package(Eigen3 CONFIG REQUIRED)

and configuring it using:

  mkdir /tmp/test-build && cd $_
  cmake \
    -DCMAKE_FIND_PACKAGE_NO_PACKAGE_REGISTRY:BOOL=1 \
    -DEigen3_DIR:PATH=/tmp/test-build/eigen-git-mirror \
    /tmp/test

Co-authored-by: Pablo Hernandez <pablo.hernandez@kitware.com>
---
 CMakeLists.txt              | 1 +
 cmake/Eigen3Config.cmake.in | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)
---
 CMakeLists.txt              | 1 +
 cmake/Eigen3Config.cmake.in | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index dbf0999ce..6d74709a3 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -538,6 +538,7 @@ if (NOT CMAKE_VERSION VERSION_LESS 3.0)
 
   # Imported target support
   add_library (eigen INTERFACE)
+  add_library (Eigen3::Eigen ALIAS eigen)
 
   target_compile_definitions (eigen INTERFACE ${EIGEN_DEFINITIONS})
   target_include_directories (eigen INTERFACE
diff --git a/cmake/Eigen3Config.cmake.in b/cmake/Eigen3Config.cmake.in
index c5c546887..0a1ac61c9 100644
--- a/cmake/Eigen3Config.cmake.in
+++ b/cmake/Eigen3Config.cmake.in
@@ -3,7 +3,9 @@
 
 @PACKAGE_INIT@
 
-include ("${CMAKE_CURRENT_LIST_DIR}/Eigen3Targets.cmake")
+if (NOT TARGET eigen)
+  include ("${CMAKE_CURRENT_LIST_DIR}/Eigen3Targets.cmake")
+endif ()
 
 # Legacy variables, do *not* use. May be removed in the future.
 

From 43fd42a33b484914ca92931ea63583b672c5e67b Mon Sep 17 00:00:00 2001
From: "luz.paz\"" <luzpaz@users.noreply.github.com>
Date: Wed, 1 Aug 2018 21:34:47 -0400
Subject: [PATCH 532/680] Fix doxy and misc. typos Found via `codespell -q 3 -I
 ../eigen-word-whitelist.txt` ---  Eigen/src/Core/ProductEvaluators.h |  4
 ++--  Eigen/src/Core/arch/GPU/Half.h     |  2 +- 
 Eigen/src/Core/util/Memory.h       |  2 +-  Eigen/src/Geometry/Hyperplane.h  
  |  2 +-  Eigen/src/Geometry/Transform.h     |  2 +- 
 Eigen/src/Geometry/Translation.h   | 12 ++++++------ 
 doc/PreprocessorDirectives.dox     |  2 +-  doc/TutorialGeometry.dox         
  |  2 +-  test/boostmultiprec.cpp            |  2 +-  test/triangular.cpp    
            |  2 +-  10 files changed, 16 insertions(+), 16 deletions(-)

---
 Eigen/src/Core/ProductEvaluators.h |  4 ++--
 Eigen/src/Core/arch/GPU/Half.h     |  2 +-
 Eigen/src/Core/util/Memory.h       |  2 +-
 Eigen/src/Geometry/Hyperplane.h    |  2 +-
 Eigen/src/Geometry/Transform.h     |  2 +-
 Eigen/src/Geometry/Translation.h   | 12 ++++++------
 doc/PreprocessorDirectives.dox     |  2 +-
 doc/TutorialGeometry.dox           |  2 +-
 test/boostmultiprec.cpp            |  2 +-
 test/triangular.cpp                |  2 +-
 10 files changed, 16 insertions(+), 16 deletions(-)

diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index 0330b5741..2787987e7 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -20,7 +20,7 @@ namespace internal {
 /** \internal
   * Evaluator of a product expression.
   * Since products require special treatments to handle all possible cases,
-  * we simply deffer the evaluation logic to a product_evaluator class
+  * we simply defer the evaluation logic to a product_evaluator class
   * which offers more partial specialization possibilities.
   * 
   * \sa class product_evaluator
@@ -128,7 +128,7 @@ protected:
   PlainObject m_result;
 };
 
-// The following three shortcuts are enabled only if the scalar types match excatly.
+// The following three shortcuts are enabled only if the scalar types match exactly.
 // TODO: we could enable them for different scalar types when the product is not vectorized.
 
 // Dense = Product
diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index aef1eb079..65b38bbfb 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -60,7 +60,7 @@ struct __half_raw {
  #if defined(EIGEN_HAS_OLD_HIP_FP16)
 // Make a __half_raw definition that is
 // ++ compatible with that of Eigen and
-// ++ add a implcit conversion to the native __half of the old HIP implementation.
+// ++ add an implicit conversion to the native __half of the old HIP implementation.
 //
 // Keeping ".x" as "unsigned short" keeps the interface the same between the Eigen and HIP implementation.
 //
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 8dbd4d93b..056070159 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -177,7 +177,7 @@ EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
     #endif
 
     #if EIGEN_DEFAULT_ALIGN_BYTES==16
-    eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator.");
+    eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade aligned memory allocator.");
     #endif
   #else
     result = handmade_aligned_malloc(size);
diff --git a/Eigen/src/Geometry/Hyperplane.h b/Eigen/src/Geometry/Hyperplane.h
index 05929b299..cebe03557 100644
--- a/Eigen/src/Geometry/Hyperplane.h
+++ b/Eigen/src/Geometry/Hyperplane.h
@@ -119,7 +119,7 @@ public:
     * If the dimension of the ambient space is greater than 2, then there isn't uniqueness,
     * so an arbitrary choice is made.
     */
-  // FIXME to be consitent with the rest this could be implemented as a static Through function ??
+  // FIXME to be consistent with the rest this could be implemented as a static Through function ??
   EIGEN_DEVICE_FUNC explicit Hyperplane(const ParametrizedLine<Scalar, AmbientDimAtCompileTime>& parametrized)
   {
     normal() = parametrized.direction().unitOrthogonal();
diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h
index 2d36dfadf..75991aaed 100644
--- a/Eigen/src/Geometry/Transform.h
+++ b/Eigen/src/Geometry/Transform.h
@@ -115,7 +115,7 @@ template<int Mode> struct transform_make_affine;
   * \end{array} \right) \f$
   *
   * Note that for a projective transformation the last row can be anything,
-  * and then the interpretation of different parts might be sightly different.
+  * and then the interpretation of different parts might be slightly different.
   *
   * However, unlike a plain matrix, the Transform class provides many features
   * simplifying both its assembly and usage. In particular, it can be composed
diff --git a/Eigen/src/Geometry/Translation.h b/Eigen/src/Geometry/Translation.h
index 51d9a82eb..23b19f74f 100644
--- a/Eigen/src/Geometry/Translation.h
+++ b/Eigen/src/Geometry/Translation.h
@@ -70,18 +70,18 @@ public:
   /** Constructs and initialize the translation transformation from a vector of translation coefficients */
   EIGEN_DEVICE_FUNC explicit inline Translation(const VectorType& vector) : m_coeffs(vector) {}
 
-  /** \brief Retruns the x-translation by value. **/
+  /** \brief Returns the x-translation by value. **/
   EIGEN_DEVICE_FUNC inline Scalar x() const { return m_coeffs.x(); }
-  /** \brief Retruns the y-translation by value. **/
+  /** \brief Returns the y-translation by value. **/
   EIGEN_DEVICE_FUNC inline Scalar y() const { return m_coeffs.y(); }
-  /** \brief Retruns the z-translation by value. **/
+  /** \brief Returns the z-translation by value. **/
   EIGEN_DEVICE_FUNC inline Scalar z() const { return m_coeffs.z(); }
 
-  /** \brief Retruns the x-translation as a reference. **/
+  /** \brief Returns the x-translation as a reference. **/
   EIGEN_DEVICE_FUNC inline Scalar& x() { return m_coeffs.x(); }
-  /** \brief Retruns the y-translation as a reference. **/
+  /** \brief Returns the y-translation as a reference. **/
   EIGEN_DEVICE_FUNC inline Scalar& y() { return m_coeffs.y(); }
-  /** \brief Retruns the z-translation as a reference. **/
+  /** \brief Returns the z-translation as a reference. **/
   EIGEN_DEVICE_FUNC inline Scalar& z() { return m_coeffs.z(); }
 
   EIGEN_DEVICE_FUNC const VectorType& vector() const { return m_coeffs; }
diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox
index b49f7d3cf..d801798c9 100644
--- a/doc/PreprocessorDirectives.dox
+++ b/doc/PreprocessorDirectives.dox
@@ -66,7 +66,7 @@ functions by defining EIGEN_HAS_C99_MATH=1.
    Automatic detection disabled if EIGEN_MAX_CPP_VER<11.
  - \b EIGEN_HAS_CXX11_MATH - controls the implementation of some functions such as round, logp1, isinf, isnan, etc.
    Automatic detection disabled if EIGEN_MAX_CPP_VER<11.
- - \b EIGEN_HAS_RVALUE_REFERENCES - defines whetehr rvalue references are supported
+ - \b EIGEN_HAS_RVALUE_REFERENCES - defines whether rvalue references are supported
    Automatic detection disabled if EIGEN_MAX_CPP_VER<11.
  - \b EIGEN_HAS_STD_RESULT_OF - defines whether std::result_of is supported
    Automatic detection disabled if EIGEN_MAX_CPP_VER<11.
diff --git a/doc/TutorialGeometry.dox b/doc/TutorialGeometry.dox
index 2e1420f98..86e9f1e72 100644
--- a/doc/TutorialGeometry.dox
+++ b/doc/TutorialGeometry.dox
@@ -111,7 +111,7 @@ rot3 = rot1.slerp(alpha,rot2);\endcode</td></tr>
 
 
 <a href="#" class="top">top</a>\section TutorialGeoTransform Affine transformations
-Generic affine transformations are represented by the Transform class which internaly
+Generic affine transformations are represented by the Transform class which internally
 is a (Dim+1)^2 matrix. In Eigen we have chosen to not distinghish between points and
 vectors such that all points are actually represented by displacement vectors from the
 origin ( \f$ \mathbf{p} \equiv \mathbf{p}-0 \f$ ). With that in mind, real points and
diff --git a/test/boostmultiprec.cpp b/test/boostmultiprec.cpp
index 1588eb86d..579a6fd25 100644
--- a/test/boostmultiprec.cpp
+++ b/test/boostmultiprec.cpp
@@ -156,7 +156,7 @@ EIGEN_DECLARE_TEST(boostmultiprec)
   std::cout << "NumTraits<Real>::highest()         = " << NumTraits<Real>::highest() << std::endl;
   std::cout << "NumTraits<Real>::digits10()        = " << NumTraits<Real>::digits10() << std::endl;
 
-  // chekc stream output
+  // check stream output
   {
     Mat A(10,10);
     A.setRandom();
diff --git a/test/triangular.cpp b/test/triangular.cpp
index 73b6fc46b..99ef1dcda 100644
--- a/test/triangular.cpp
+++ b/test/triangular.cpp
@@ -68,7 +68,7 @@ template<typename MatrixType> void triangular_square(const MatrixType& m)
     while (numext::abs2(m1(i,i))<RealScalar(1e-1)) m1(i,i) = internal::random<Scalar>();
 
   Transpose<MatrixType> trm4(m4);
-  // test back and forward subsitution with a vector as the rhs
+  // test back and forward substitution with a vector as the rhs
   m3 = m1.template triangularView<Upper>();
   VERIFY(v2.isApprox(m3.adjoint() * (m1.adjoint().template triangularView<Lower>().solve(v2)), largerEps));
   m3 = m1.template triangularView<Lower>();

From 09c81ac03335584674371369d0045510dc83aba4 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 4 Aug 2018 00:17:37 +0200
Subject: [PATCH 533/680] bug #1451: fix numeric_limits<AutoDiffScalar<Der>>
 with a reference as derivative type

---
 unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
index 279fe5cd3..13d959df4 100755
--- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
+++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
@@ -684,10 +684,15 @@ template<typename DerType> struct NumTraits<AutoDiffScalar<DerType> >
 }
 
 namespace std {
+
 template <typename T>
 class numeric_limits<Eigen::AutoDiffScalar<T> >
   : public numeric_limits<typename T::Scalar> {};
 
+template <typename T>
+class numeric_limits<Eigen::AutoDiffScalar<T&> >
+  : public numeric_limits<typename T::Scalar> {};
+
 }  // namespace std
 
 #endif // EIGEN_AUTODIFF_SCALAR_H

From fa68342ef80d72f52b340851a5ab18003259b2ed Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 3 Aug 2018 17:31:23 -0700
Subject: [PATCH 534/680] Move sigmoid functor to core.

---
 Eigen/src/Core/GlobalFunctions.h              | 11 ++++---
 Eigen/src/Core/functors/UnaryFunctors.h       | 32 +++++++++++++++++--
 Eigen/src/plugins/ArrayCwiseUnaryOps.h        | 10 ++++++
 test/array.cpp                                |  4 +++
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   | 30 -----------------
 5 files changed, 50 insertions(+), 37 deletions(-)

diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h
index 50406400b..4c9823d70 100644
--- a/Eigen/src/Core/GlobalFunctions.h
+++ b/Eigen/src/Core/GlobalFunctions.h
@@ -66,6 +66,7 @@ namespace Eigen
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op,hyperbolic sine,\sa ArrayBase::sinh)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op,hyperbolic cosine,\sa ArrayBase::cosh)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op,hyperbolic tangent,\sa ArrayBase::tanh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sigmoid,scalar_sigmoid_op,sigmoid function,\sa ArrayBase::sigmoid)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op,natural logarithm of the gamma function,\sa ArrayBase::lgamma)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(digamma,scalar_digamma_op,derivative of lgamma,\sa ArrayBase::digamma)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op,error function,\sa ArrayBase::erf)
@@ -89,7 +90,7 @@ namespace Eigen
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op,infinite value test,\sa Eigen::isnan DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isinf)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op,finite value test,\sa Eigen::isinf DOXCOMMA Eigen::isnan DOXCOMMA ArrayBase::isfinite)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sign,scalar_sign_op,sign (or 0),\sa ArrayBase::sign)
-  
+
   /** \returns an expression of the coefficient-wise power of \a x to the given constant \a exponent.
     *
     * \tparam ScalarExponent is the scalar type of \a exponent. It must be compatible with the scalar type of the given expression (\c Derived::Scalar).
@@ -124,21 +125,21 @@ namespace Eigen
     *
     * Example: \include Cwise_array_power_array.cpp
     * Output: \verbinclude Cwise_array_power_array.out
-    * 
+    *
     * \sa ArrayBase::pow()
     *
     * \relates ArrayBase
     */
   template<typename Derived,typename ExponentDerived>
   inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>
-  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) 
+  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents)
   {
     return Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>(
       x.derived(),
       exponents.derived()
     );
   }
-  
+
   /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents.
     *
     * This function computes the coefficient-wise power between a scalar and an array of exponents.
@@ -147,7 +148,7 @@ namespace Eigen
     *
     * Example: \include Cwise_scalar_power_array.cpp
     * Output: \verbinclude Cwise_scalar_power_array.out
-    * 
+    *
     * \sa ArrayBase::pow()
     *
     * \relates ArrayBase
diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index bfc046556..c350cdf98 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -701,7 +701,7 @@ template<typename Scalar> struct scalar_isnan_op {
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const {
 #if defined(__SYCL_DEVICE_ONLY__)
     return numext::isnan(a);
-#else  
+#else
     return (numext::isnan)(a);
 #endif
   }
@@ -815,7 +815,7 @@ struct scalar_sign_op<Scalar,true> {
 template<typename Scalar>
 struct functor_traits<scalar_sign_op<Scalar> >
 { enum {
-    Cost = 
+    Cost =
         NumTraits<Scalar>::IsComplex
         ? ( 8*NumTraits<Scalar>::MulCost  ) // roughly
         : ( 3*NumTraits<Scalar>::AddCost),
@@ -823,6 +823,34 @@ struct functor_traits<scalar_sign_op<Scalar> >
   };
 };
 
+/** \internal
+  * \brief Template functor to compute the sigmoid of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sigmoid()
+  */
+template <typename T>
+struct scalar_sigmoid_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sigmoid_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& x) const {
+    const T one = T(1);
+    return one / (one + numext::exp(-x));
+  }
+
+  template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(const Packet& x) const {
+    const Packet one = pset1<Packet>(T(1));
+    return pdiv(one, padd(one, pexp(pnegate(x))));
+  }
+};
+template <typename T>
+struct functor_traits<scalar_sigmoid_op<T> > {
+  enum {
+    Cost = NumTraits<T>::AddCost * 2 + NumTraits<T>::MulCost * 6,
+    PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv &&
+                   packet_traits<T>::HasNegate && packet_traits<T>::HasExp
+  };
+};
+
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h
index 43615bd56..c9c7e7e4f 100644
--- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h
+++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h
@@ -21,6 +21,7 @@ typedef CwiseUnaryOp<internal::scalar_acos_op<Scalar>, const Derived> AcosReturn
 typedef CwiseUnaryOp<internal::scalar_asin_op<Scalar>, const Derived> AsinReturnType;
 typedef CwiseUnaryOp<internal::scalar_atan_op<Scalar>, const Derived> AtanReturnType;
 typedef CwiseUnaryOp<internal::scalar_tanh_op<Scalar>, const Derived> TanhReturnType;
+typedef CwiseUnaryOp<internal::scalar_sigmoid_op<Scalar>, const Derived> SigmoidReturnType;
 typedef CwiseUnaryOp<internal::scalar_sinh_op<Scalar>, const Derived> SinhReturnType;
 typedef CwiseUnaryOp<internal::scalar_cosh_op<Scalar>, const Derived> CoshReturnType;
 typedef CwiseUnaryOp<internal::scalar_square_op<Scalar>, const Derived> SquareReturnType;
@@ -335,6 +336,15 @@ cosh() const
   return CoshReturnType(derived());
 }
 
+/** \returns an expression of the coefficient-wise sigmoid of *this.
+  */
+EIGEN_DEVICE_FUNC
+inline const SigmoidReturnType
+sigmoid() const
+{
+  return SigmoidReturnType(derived());
+}
+
 /** \returns an expression of the coefficient-wise inverse of *this.
   *
   * Example: \include Cwise_inverse.cpp
diff --git a/test/array.cpp b/test/array.cpp
index c01653668..1b5725fb4 100644
--- a/test/array.cpp
+++ b/test/array.cpp
@@ -231,6 +231,7 @@ template<typename ArrayType> void array_real(const ArrayType& m)
   VERIFY_IS_APPROX(m1.sinh(), sinh(m1));
   VERIFY_IS_APPROX(m1.cosh(), cosh(m1));
   VERIFY_IS_APPROX(m1.tanh(), tanh(m1));
+  VERIFY_IS_APPROX(m1.sigmoid(), sigmoid(m1));
 
   VERIFY_IS_APPROX(m1.arg(), arg(m1));
   VERIFY_IS_APPROX(m1.round(), round(m1));
@@ -266,6 +267,7 @@ template<typename ArrayType> void array_real(const ArrayType& m)
   VERIFY_IS_APPROX(sinh(m1), 0.5*(exp(m1)-exp(-m1)));
   VERIFY_IS_APPROX(cosh(m1), 0.5*(exp(m1)+exp(-m1)));
   VERIFY_IS_APPROX(tanh(m1), (0.5*(exp(m1)-exp(-m1)))/(0.5*(exp(m1)+exp(-m1))));
+  VERIFY_IS_APPROX(sigmoid(m1), (1.0/(1.0+exp(-m1))));
   VERIFY_IS_APPROX(arg(m1), ((m1<0).template cast<Scalar>())*std::acos(-1.0));
   VERIFY((round(m1) <= ceil(m1) && round(m1) >= floor(m1)).all());
   VERIFY((Eigen::isnan)((m1*0.0)/0.0).all());
@@ -345,6 +347,7 @@ template<typename ArrayType> void array_complex(const ArrayType& m)
   VERIFY_IS_APPROX(m1.sinh(), sinh(m1));
   VERIFY_IS_APPROX(m1.cosh(), cosh(m1));
   VERIFY_IS_APPROX(m1.tanh(), tanh(m1));
+  VERIFY_IS_APPROX(m1.sigmoid(), sigmoid(m1));
   VERIFY_IS_APPROX(m1.arg(), arg(m1));
   VERIFY((m1.isNaN() == (Eigen::isnan)(m1)).all());
   VERIFY((m1.isInf() == (Eigen::isinf)(m1)).all());
@@ -368,6 +371,7 @@ template<typename ArrayType> void array_complex(const ArrayType& m)
   VERIFY_IS_APPROX(sinh(m1), 0.5*(exp(m1)-exp(-m1)));
   VERIFY_IS_APPROX(cosh(m1), 0.5*(exp(m1)+exp(-m1)));
   VERIFY_IS_APPROX(tanh(m1), (0.5*(exp(m1)-exp(-m1)))/(0.5*(exp(m1)+exp(-m1))));
+  VERIFY_IS_APPROX(sigmoid(m1), (1.0/(1.0 + exp(-m1))));
 
   for (Index i = 0; i < m.rows(); ++i)
     for (Index j = 0; j < m.cols(); ++j)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 7ecd4d1ac..cd666c173 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -54,36 +54,6 @@ struct functor_traits<scalar_fmod_op<Scalar> > {
          PacketAccess = false };
 };
 
-
-/** \internal
-  * \brief Template functor to compute the sigmoid of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::sigmoid()
-  */
-template <typename T>
-struct scalar_sigmoid_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sigmoid_op)
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& x) const {
-    const T one = T(1);
-    return one / (one + numext::exp(-x));
-  }
-
-  template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  Packet packetOp(const Packet& x) const {
-    const Packet one = pset1<Packet>(T(1));
-    return pdiv(one, padd(one, pexp(pnegate(x))));
-  }
-};
-
-template <typename T>
-struct functor_traits<scalar_sigmoid_op<T> > {
-  enum {
-    Cost = NumTraits<T>::AddCost * 2 + NumTraits<T>::MulCost * 6,
-    PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv &&
-                   packet_traits<T>::HasNegate && packet_traits<T>::HasExp
-  };
-};
-
-
 template<typename Reducer, typename Device>
 struct reducer_traits {
   enum {

From 36e7e7dd8fc995c8a7cb8caa924663a56bc4cc3a Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 6 Aug 2018 13:16:32 -0700
Subject: [PATCH 535/680] Forward declare NoOpOutputKernel as struct rather
 than class to be consistent with implementation.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index ec1dc0fab..0dd524a30 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -98,7 +98,7 @@ template<typename XprType> class TensorForcedEvalOp;
 template<typename ExpressionType, typename DeviceType> class TensorDevice;
 template<typename Derived, typename Device> struct TensorEvaluator;
 
-class NoOpOutputKernel;
+struct NoOpOutputKernel;
 
 struct DefaultDevice;
 struct ThreadPoolDevice;

From d011d05fd6306de11fbc74c5b4fdaa929f2fe3ac Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 6 Aug 2018 13:40:51 -0700
Subject: [PATCH 536/680] Fixed compilation errors.

---
 .../Eigen/CXX11/src/Tensor/TensorBase.h       | 24 +++++++++----------
 .../CXX11/src/Tensor/TensorConvolution.h      |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  |  4 ++--
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |  4 ++--
 4 files changed, 18 insertions(+), 18 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 97f90f638..ab3731952 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -538,8 +538,8 @@ class TensorBase<Derived, ReadOnlyAccessors>
     // Fourier transforms
     template <int FFTDataType, int FFTDirection, typename FFT> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorFFTOp<const FFT, const Derived, FFTDataType, FFTDirection>
-    fft(const FFT& fft) const {
-      return TensorFFTOp<const FFT, const Derived, FFTDataType, FFTDirection>(derived(), fft);
+    fft(const FFT& dims) const {
+      return TensorFFTOp<const FFT, const Derived, FFTDataType, FFTDirection>(derived(), dims);
     }
 
     // Scan.
@@ -723,8 +723,8 @@ class TensorBase<Derived, ReadOnlyAccessors>
 
     template <typename Broadcast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorBroadcastingOp<const Broadcast, const Derived>
-    broadcast(const Broadcast& broadcast) const {
-      return TensorBroadcastingOp<const Broadcast, const Derived>(derived(), broadcast);
+    broadcast(const Broadcast& bcast) const {
+      return TensorBroadcastingOp<const Broadcast, const Derived>(derived(), bcast);
     }
 
     template <typename Axis, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -832,8 +832,8 @@ class TensorBase<Derived, ReadOnlyAccessors>
     }
     template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorShufflingOp<const Shuffle, const Derived>
-    shuffle(const Shuffle& shuffle) const {
-      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shuffle);
+    shuffle(const Shuffle& shfl) const {
+      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shfl);
     }
     template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorStridingOp<const Strides, const Derived>
@@ -1030,13 +1030,13 @@ class TensorBase : public TensorBase<Derived, ReadOnlyAccessors> {
 
     template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorShufflingOp<const Shuffle, const Derived>
-    shuffle(const Shuffle& shuffle) const {
-      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shuffle);
+    shuffle(const Shuffle& shfl) const {
+      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shfl);
     }
     template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     TensorShufflingOp<const Shuffle, Derived>
-    shuffle(const Shuffle& shuffle) {
-      return TensorShufflingOp<const Shuffle, Derived>(derived(), shuffle);
+    shuffle(const Shuffle& shfl) {
+      return TensorShufflingOp<const Shuffle, Derived>(derived(), shfl);
     }
 
     template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -1052,8 +1052,8 @@ class TensorBase : public TensorBase<Derived, ReadOnlyAccessors> {
 
     // Select the device on which to evaluate the expression.
     template <typename DeviceType>
-    TensorDevice<Derived, DeviceType> device(const DeviceType& device) {
-      return TensorDevice<Derived, DeviceType>(device, derived());
+    TensorDevice<Derived, DeviceType> device(const DeviceType& dev) {
+      return TensorDevice<Derived, DeviceType>(dev, derived());
     }
 
  protected:
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 0d3ca966c..e3e650fa5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -527,8 +527,8 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
       Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
       typedef TensorEvalToOp<const KernelArgType> EvalTo;
       EvalTo evalToTmp(local, m_kernelArg);
-      const bool PacketAccess = internal::IsVectorizable<Device, KernelArgType>::value;
-      internal::TensorExecutor<const EvalTo, Device, PacketAccess>::run(evalToTmp, m_device);
+      const bool Vectorize = internal::IsVectorizable<Device, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, Device, Vectorize>::run(evalToTmp, m_device);
 
       m_kernel = local;
       m_local_kernel = true;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index 8f7a81575..028902fea 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -126,7 +126,7 @@ struct TensorEvaluator
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
-      std::vector<internal::TensorOpResourceRequirements>* resources) const {}
+      std::vector<internal::TensorOpResourceRequirements>*) const {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(TensorBlock* block) const {
     assert(m_data != NULL);
@@ -255,7 +255,7 @@ struct TensorEvaluator<const Derived, Device>
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
-      std::vector<internal::TensorOpResourceRequirements>* resources) const {}
+      std::vector<internal::TensorOpResourceRequirements>*) const {}
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(TensorBlock* block) const {
     assert(m_data != NULL);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index a456f308b..2778bf5ec 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -124,8 +124,8 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     }
     typedef TensorEvalToOp< const typename internal::remove_const<ArgType>::type > EvalTo;
     EvalTo evalToTmp(m_buffer, m_op);
-    const bool PacketAccess = internal::IsVectorizable<Device, const ArgType>::value;
-    internal::TensorExecutor<const EvalTo, typename internal::remove_const<Device>::type, PacketAccess>::run(evalToTmp, m_device);
+    const bool Vectorize = internal::IsVectorizable<Device, const ArgType>::value;
+    internal::TensorExecutor<const EvalTo, typename internal::remove_const<Device>::type, Vectorize>::run(evalToTmp, m_device);
     return true;
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {

From 10d286f55b30bfcd45be3e83c44edde5ac184270 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 6 Aug 2018 16:00:29 -0700
Subject: [PATCH 537/680] Silenced a couple of compilation warnings.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h     | 2 +-
 unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 84cf6d216..dc8b5ae6d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -747,7 +747,7 @@ class TensorBlockMapper {
         // block dimension sizes based on "square" dimension size target.
         const size_t dim_size_target = static_cast<const size_t>(
             std::pow(static_cast<float>(min_target_size),
-                     1.0 / static_cast<float>(block_dim_sizes.rank())));
+                     1.0f / static_cast<float>(block_dim_sizes.rank())));
         for (size_t i = 0; i < block_dim_sizes.rank(); ++i) {
           // TODO(andydavis) Adjust the inner most 'block_dim_size' to make it
           // a multiple of the packet size. Note that reducing
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 0fc49255d..e25dd9cf8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -61,8 +61,8 @@ class TensorShufflingOp : public TensorBase<TensorShufflingOp<Shuffle, XprType>
   typedef typename Eigen::internal::traits<TensorShufflingOp>::StorageKind StorageKind;
   typedef typename Eigen::internal::traits<TensorShufflingOp>::Index Index;
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shuffle)
-      : m_xpr(expr), m_shuffle(shuffle) {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shfl)
+      : m_xpr(expr), m_shuffle(shfl) {}
 
     EIGEN_DEVICE_FUNC
     const Shuffle& shufflePermutation() const { return m_shuffle; }

From 693fb1d41e1b267ae149b8f368fc3008bd59aab9 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Tue, 7 Aug 2018 17:18:51 -0700
Subject: [PATCH 538/680] Fix init order.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index e3e650fa5..a07e32db0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -786,7 +786,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
   };
 
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const GpuDevice& device)
-      : m_inputImpl(op.inputExpression(), device), m_kernelArg(op.kernelExpression()), m_kernelImpl(op.kernelExpression(), device), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device)
+      : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device)
   {
     EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, GpuDevice>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, GpuDevice>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
 

From 908b906d79a30b357f82e9c8a1e0d238e2f3469a Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Aug 2018 10:01:10 +0100
Subject: [PATCH 539/680] Disabling assert inside SYCL kernel.

---
 test/main.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/test/main.h b/test/main.h
index 5d64bc736..de8a4865f 100644
--- a/test/main.h
+++ b/test/main.h
@@ -193,7 +193,7 @@ namespace Eigen
 
 #define EIGEN_DEFAULT_IO_FORMAT IOFormat(4, 0, "  ", "\n", "", "", "", "")
 
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__)
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__) && !defined(__SYCL_DEVICE_ONLY__)
   #define EIGEN_EXCEPTIONS
 #endif
 
@@ -272,7 +272,7 @@ namespace Eigen
       }
     #endif //EIGEN_EXCEPTIONS
 
-  #elif !defined(__CUDACC__) && !defined(__HIPCC__)// EIGEN_DEBUG_ASSERTS
+  #elif !defined(__CUDACC__) && !defined(__HIPCC__) && !defined(__SYCL_DEVICE_ONLY__) // EIGEN_DEBUG_ASSERTS
     // see bug 89. The copy_bool here is working around a bug in gcc <= 4.3
     #define eigen_assert(a) \
       if( (!Eigen::internal::copy_bool(a)) && (!no_more_assert) )\
@@ -329,7 +329,7 @@ namespace Eigen
     std::cout << "Can't VERIFY_RAISES_STATIC_ASSERT( " #a " ) with exceptions disabled\n";
 #endif
     
-  #if !defined(__CUDACC__) && !defined(__HIPCC__)
+  #if !defined(__CUDACC__) && !defined(__HIPCC__) && !defined(__SYCL_DEVICE_ONLY__)
   #define EIGEN_USE_CUSTOM_ASSERT
   #endif
 
@@ -845,4 +845,4 @@ int main(int argc, char *argv[])
 #ifdef _MSC_VER
   // 4503 - decorated name length exceeded, name was truncated
   #pragma warning( disable : 4503)
-#endif
+#endif
\ No newline at end of file

From 22031ab59a5b9c7b2612feaa12abe7bcef56a8e2 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Aug 2018 11:07:27 +0100
Subject: [PATCH 540/680] Adding EIGEN_UNROLL_LOOP macro.

---
 Eigen/src/Core/util/Macros.h | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 9a2f5ab05..81ca6b8fe 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1092,5 +1092,15 @@ bool all(T t, Ts ... ts){ return t && all(ts...); }
 }
 #endif
 
+// Wrapping #pragma unroll in a macro since it is required for SYCL
+#if defined(__SYCL_DEVICE_ONLY__)
+  #if defined(_MSC_VER)
+    #define EIGEN_UNROLL_LOOP __Pragma(unroll)
+  #else
+    #define EIGEN_UNROLL_LOOP _Pragma("unroll")
+  #endif
+#else
+  #define EIGEN_UNROLL_LOOP
+#endif
 
 #endif // EIGEN_MACROS_H

From 3055e3a7c213e7f1337cdc43d5b1953acd013897 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Aug 2018 11:19:02 +0100
Subject: [PATCH 541/680] Creating a pointer type in TensorCustomOp.h

---
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   | 25 ++++++++++---------
 1 file changed, 13 insertions(+), 12 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 47b5a5a5e..87d84a311 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -88,6 +88,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename internal::remove_all<typename Eigen::internal::traits<XprType>::PointerType>::type * PointerType;
 
   enum {
     IsAligned = false,
@@ -106,7 +107,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerType data) {
     if (data) {
       evalTo(data);
       return false;
@@ -139,23 +140,22 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC PointerType data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
 #endif
 
  protected:
-  EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
-    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(
-        data, m_dimensions);
+  EIGEN_DEVICE_FUNC void evalTo(PointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(data, m_dimensions);
     m_op.func().eval(m_op.expression(), result, m_device);
   }
 
   Dimensions m_dimensions;
   const ArgType m_op;
   const Device& m_device;
-  CoeffReturnType* m_result;
+  PointerType m_result;
 };
 
 
@@ -250,6 +250,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename internal::remove_all<typename Eigen::internal::traits<XprType>::PointerType>::type * PointerType;
 
   enum {
     IsAligned = false,
@@ -268,12 +269,12 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerType data) {
     if (data) {
       evalTo(data);
       return false;
     } else {
-      m_result = static_cast<Scalar *>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
+      m_result = static_cast<PointerType>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(CoeffReturnType)));
       evalTo(m_result);
       return true;
     }
@@ -300,22 +301,22 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename internal::traits<XprType>::PointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC PointerType data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
 #endif
 
  protected:
-  EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
-    TensorMap<Tensor<Scalar, NumDims, Layout> > result(data, m_dimensions);
+  EIGEN_DEVICE_FUNC void evalTo(PointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout> > result(data, m_dimensions);
     m_op.func().eval(m_op.lhsExpression(), m_op.rhsExpression(), result, m_device);
   }
 
   Dimensions m_dimensions;
   const XprType m_op;
   const Device& m_device;
-  CoeffReturnType* m_result;
+  PointerType m_result;
 };
 
 

From 67711eaa310b1f1db67fc866cc64f703faa4f788 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Aug 2018 11:38:10 +0100
Subject: [PATCH 542/680] Fixing typo.

---
 Eigen/src/Core/util/Macros.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 81ca6b8fe..bcdede61e 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -1095,7 +1095,7 @@ bool all(T t, Ts ... ts){ return t && all(ts...); }
 // Wrapping #pragma unroll in a macro since it is required for SYCL
 #if defined(__SYCL_DEVICE_ONLY__)
   #if defined(_MSC_VER)
-    #define EIGEN_UNROLL_LOOP __Pragma(unroll)
+    #define EIGEN_UNROLL_LOOP __pragma(unroll)
   #else
     #define EIGEN_UNROLL_LOOP _Pragma("unroll")
   #endif

From 532a0be05c0bc0f9a199f97122c88967494dd4ba Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Wed, 8 Aug 2018 12:12:26 +0100
Subject: [PATCH 543/680] Fixing compiler warning in TensorBlock.h as it was
 creating a lot of noise at compilation.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index dc8b5ae6d..45ddfdb39 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -89,7 +89,7 @@ EIGEN_STRONG_INLINE void MergeResourceRequirements(
   // policy if block shapes/sizes conflict).
   *block_shape = resources[0].block_shape;
   *block_total_size = resources[0].block_total_size;
-  for (int i = 1; i < resources.size(); ++i) {
+  for (std::vector<TensorOpResourceRequirements>::size_type i = 1; i < resources.size(); ++i) {
     if (resources[i].block_shape == TensorBlockShapeType::kSkewedInnerDims &&
         *block_shape != TensorBlockShapeType::kSkewedInnerDims) {
       *block_shape = TensorBlockShapeType::kSkewedInnerDims;

From 050bcf61261d5b3bcc86b2c5afc7e35d3fd16ff7 Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Wed, 8 Aug 2018 20:19:32 -0700
Subject: [PATCH 544/680] bug #1584: Improve random (avoid undefined behavior).

---
 Eigen/src/Core/MathFunctions.h | 31 +++++++++++++++++++------------
 Eigen/src/Core/util/Meta.h     | 21 +++++++++++++++++++++
 2 files changed, 40 insertions(+), 12 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index f16476a92..58cf7021a 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -684,20 +684,27 @@ struct random_default_impl<Scalar, false, true>
 {
   static inline Scalar run(const Scalar& x, const Scalar& y)
   {
-    typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
-    if(y<x)
+    if (y <= x)
       return x;
-    // the following difference might overflow on a 32 bits system,
-    // but since y>=x the result converted to an unsigned long is still correct.
-    std::size_t range = ScalarX(y)-ScalarX(x);
-    std::size_t offset = 0;
-    // rejection sampling
-    std::size_t divisor = 1;
-    std::size_t multiplier = 1;
-    if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1);
-    else               multiplier = 1 + range/(std::size_t(RAND_MAX)+1);
+    // ScalarU is the unsigned counterpart of Scalar, possibly Scalar itself.
+    typedef typename make_unsigned<Scalar>::type ScalarU;
+    // ScalarX is the widest of ScalarU and unsigned int.
+    // We'll deal only with ScalarX and unsigned int below thus avoiding signed
+    // types and arithmetic and signed overflows (which are undefined behavior).
+    typedef typename conditional<(ScalarU(-1) > unsigned(-1)), ScalarU, unsigned>::type ScalarX;
+    // The following difference doesn't overflow, provided our integer types are two's
+    // complement and have the same number of padding bits in signed and unsigned variants.
+    // This is the case in most modern implementations of C++.
+    ScalarX range = ScalarX(y) - ScalarX(x);
+    ScalarX offset = 0;
+    ScalarX divisor = 1;
+    ScalarX multiplier = 1;
+    const unsigned rand_max = RAND_MAX;
+    if (range <= rand_max) divisor = (rand_max + 1) / (range + 1);
+    else                   multiplier = 1 + range / (rand_max + 1);
+    // Rejection sampling.
     do {
-      offset = (std::size_t(std::rand()) * multiplier) / divisor;
+      offset = (unsigned(std::rand()) * multiplier) / divisor;
     } while (offset > range);
     return Scalar(ScalarX(x) + offset);
   }
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index f27b8e85d..1e4f95581 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -128,6 +128,27 @@ template<> struct is_integral<unsigned __int64>       { enum { value = true }; }
 #endif
 #endif
 
+#if EIGEN_HAS_CXX11
+using std::make_unsigned;
+#else
+// TODO: Possibly improve this implementation of make_unsigned.
+// It is currently used only by
+// template<typename Scalar> struct random_default_impl<Scalar, false, true>.
+template<typename> struct make_unsigned;
+template<> struct make_unsigned<char>             { typedef unsigned char type; };
+template<> struct make_unsigned<signed char>      { typedef unsigned char type; };
+template<> struct make_unsigned<unsigned char>    { typedef unsigned char type; };
+template<> struct make_unsigned<signed short>     { typedef unsigned short type; };
+template<> struct make_unsigned<unsigned short>   { typedef unsigned short type; };
+template<> struct make_unsigned<signed int>       { typedef unsigned int type; };
+template<> struct make_unsigned<unsigned int>     { typedef unsigned int type; };
+template<> struct make_unsigned<signed long>      { typedef unsigned long type; };
+template<> struct make_unsigned<unsigned long>    { typedef unsigned long type; };
+#if EIGEN_COMP_MSVC
+template<> struct make_unsigned<signed __int64>   { typedef unsigned __int64 type; };
+template<> struct make_unsigned<unsigned __int64> { typedef unsigned __int64 type; };
+#endif
+#endif
 
 template <typename T> struct add_const { typedef const T type; };
 template <typename T> struct add_const<T&> { typedef T& type; };

From 8c083bfd0e975fc2592ced1a066e1796550338a2 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 9 Aug 2018 13:57:43 +0100
Subject: [PATCH 545/680] Properly fixing the PointerType for TensorCustomOp.h.
 As the output type here should be based on CoeffreturnType not the Scalar
 type.  Therefore, Similar to reduction and evalTo function, it should have
 its own MakePointer class. In this case, for other device the type is
 defaulted to CoeffReturnType and no changes is required on users' code.
 However, in SYCL, on the device, we can recunstruct the device Type.

---
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   | 77 +++++++++++--------
 .../src/Tensor/TensorForwardDeclarations.h    |  4 +-
 2 files changed, 48 insertions(+), 33 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 87d84a311..39410e63d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -20,8 +20,8 @@ namespace Eigen {
   *
   */
 namespace internal {
-template<typename CustomUnaryFunc, typename XprType>
-struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_>
+struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_> >
 {
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::StorageKind StorageKind;
@@ -30,27 +30,35 @@ struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
-  typedef typename traits<XprType>::PointerType PointerType;
+
+   template <class T> struct MakePointer {
+    // Intermediate typedef to workaround MSVC issue.
+    typedef MakePointer_<T> MakePointerT;
+    typedef typename MakePointerT::Type Type;
+    typedef typename MakePointerT::RefType RefType;
+    typedef typename MakePointerT::ScalarType ScalarType;
+  };
+  typedef typename MakePointer<typename internal::remove_const<typename XprType::CoeffReturnType>::type>::Type PointerType;
 };
 
-template<typename CustomUnaryFunc, typename XprType>
-struct eval<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Eigen::Dense>
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_>
+struct eval<TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_>, Eigen::Dense>
 {
-  typedef const TensorCustomUnaryOp<CustomUnaryFunc, XprType>& type;
+  typedef const TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_>& type;
 };
 
-template<typename CustomUnaryFunc, typename XprType>
-struct nested<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_>
+struct nested<TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_> >
 {
-  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> type;
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_> type;
 };
 
 }  // end namespace internal
 
 
 
-template<typename CustomUnaryFunc, typename XprType>
-class TensorCustomUnaryOp : public TensorBase<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, ReadOnlyAccessors>
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_>
+class TensorCustomUnaryOp : public TensorBase<TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_>, ReadOnlyAccessors>
 {
   public:
   typedef typename internal::traits<TensorCustomUnaryOp>::Scalar Scalar;
@@ -77,10 +85,10 @@ class TensorCustomUnaryOp : public TensorBase<TensorCustomUnaryOp<CustomUnaryFun
 
 
 // Eval as rvalue
-template<typename CustomUnaryFunc, typename XprType, typename Device>
-struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Device>
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_, typename Device>
+struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_>, Device>
 {
-  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> ArgType;
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType, MakePointer_> ArgType;
   typedef typename internal::traits<ArgType>::Index Index;
   static const int NumDims = internal::traits<ArgType>::NumDimensions;
   typedef DSizes<Index, NumDims> Dimensions;
@@ -88,7 +96,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
-  typedef typename internal::remove_all<typename Eigen::internal::traits<XprType>::PointerType>::type * PointerType;
+  typedef typename Eigen::internal::traits<ArgType>::PointerType PointerType;
 
   enum {
     IsAligned = false,
@@ -112,7 +120,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
       evalTo(data);
       return false;
     } else {
-      m_result = static_cast<CoeffReturnType*>(
+      m_result = static_cast<PointerType>(
           m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
       evalTo(m_result);
       return true;
@@ -168,8 +176,8 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
   *
   */
 namespace internal {
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
-struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, template <class> class MakePointer_>
+struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_> >
 {
   typedef typename internal::promote_storage_type<typename LhsXprType::Scalar,
                                                   typename RhsXprType::Scalar>::ret Scalar;
@@ -185,28 +193,35 @@ struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
-  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
-                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
+
+ template <class T> struct MakePointer {
+    // Intermediate typedef to workaround MSVC issue.
+    typedef MakePointer_<T> MakePointerT;
+    typedef typename MakePointerT::Type Type;
+    typedef typename MakePointerT::RefType RefType;
+    typedef typename MakePointerT::ScalarType ScalarType;
+  };
+  typedef typename MakePointer<CoeffReturnType>::Type PointerType;
 };
 
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
-struct eval<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Eigen::Dense>
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, template <class> class MakePointer_>
+struct eval<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_>, Eigen::Dense>
 {
   typedef const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>& type;
 };
 
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
-struct nested<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, template <class> class MakePointer_>
+struct nested<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_> >
 {
-  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> type;
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_> type;
 };
 
 }  // end namespace internal
 
 
 
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
-class TensorCustomBinaryOp : public TensorBase<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, ReadOnlyAccessors>
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType,template <class> class MakePointer_>
+class TensorCustomBinaryOp : public TensorBase<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_>, ReadOnlyAccessors>
 {
   public:
   typedef typename internal::traits<TensorCustomBinaryOp>::Scalar Scalar;
@@ -239,10 +254,10 @@ class TensorCustomBinaryOp : public TensorBase<TensorCustomBinaryOp<CustomBinary
 
 
 // Eval as rvalue
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, typename Device>
-struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Device>
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, template <class> class MakePointer_, typename Device>
+struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_>, Device>
 {
-  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> XprType;
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType, MakePointer_> XprType;
   typedef typename internal::traits<XprType>::Index Index;
   static const int NumDims = internal::traits<XprType>::NumDimensions;
   typedef DSizes<Index, NumDims> Dimensions;
@@ -250,7 +265,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
-  typedef typename internal::remove_all<typename Eigen::internal::traits<XprType>::PointerType>::type * PointerType;
+  typedef typename Eigen::internal::traits<XprType>::PointerType PointerType;
 
   enum {
     IsAligned = false,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 0dd524a30..da0751039 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -89,8 +89,8 @@ template<typename LeftXprType, typename RightXprType> class TensorAssignOp;
 template<typename Op, typename XprType> class TensorScanOp;
 template<typename Dims, typename XprType> class TensorTraceOp;
 
-template<typename CustomUnaryFunc, typename XprType> class TensorCustomUnaryOp;
-template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType> class TensorCustomBinaryOp;
+template<typename CustomUnaryFunc, typename XprType, template <class> class MakePointer_ = MakePointer> class TensorCustomUnaryOp;
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, template <class> class MakePointer_ = MakePointer> class TensorCustomBinaryOp;
 
 template<typename XprType, template <class> class MakePointer_ = MakePointer> class TensorEvalToOp;
 template<typename XprType> class TensorForcedEvalOp;

From cfaedb38cd662def3b5684a20965b3bc1b0d6a3f Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 9 Aug 2018 09:44:07 -0700
Subject: [PATCH 546/680] Fix bug in a test + compilation errors

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 6 ++++--
 unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 8 ++++----
 unsupported/test/cxx11_tensor_executor.cpp          | 2 +-
 unsupported/test/cxx11_tensor_shuffling.cpp         | 8 ++++----
 4 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index f11241f83..4bc18bd97 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -115,6 +115,7 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable,
                          const DefaultDevice& device = DefaultDevice()) {
     typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>
         TensorBlock;
+    typedef typename TensorBlock::Dimensions TensorBlockDimensions;
     typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims,
                               Evaluator::Layout>
         TensorBlockMapper;
@@ -141,8 +142,9 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable,
       evaluator.getResourceRequirements(&resources);
       MergeResourceRequirements(resources, &block_shape, &block_total_size);
 
-      TensorBlockMapper block_mapper(evaluator.dimensions(), block_shape,
-                                     block_total_size);
+      TensorBlockMapper block_mapper(
+          TensorBlockDimensions(evaluator.dimensions()), block_shape,
+          block_total_size);
       block_total_size = block_mapper.block_dims_total_size();
 
       Scalar* data = static_cast<Scalar*>(
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index d5b0c1237..b25c1eabc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -520,6 +520,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
 
   typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
       TensorBlock;
+  typedef typename TensorBlock::Dimensions TensorBlockDimensions;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices())
@@ -687,7 +688,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
     TensorBlock input_block(srcCoeff(output_block->first_coeff_index()),
                             output_block->block_sizes(),
                             output_block->block_strides(),
-                            Dimensions(m_inputStrides),
+                            TensorBlockDimensions(m_inputStrides),
                             output_block->data());
     m_impl.block(&input_block);
   }
@@ -796,6 +797,7 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
 
   typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
       TensorBlock;
+  typedef typename TensorBlock::Dimensions TensorBlockDimensions;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
     : Base(op, device)
@@ -862,13 +864,11 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
       const TensorBlock& block) {
     this->m_impl.writeBlock(TensorBlock(
         this->srcCoeff(block.first_coeff_index()), block.block_sizes(),
-        block.block_strides(), Dimensions(this->m_inputStrides),
+        block.block_strides(), TensorBlockDimensions(this->m_inputStrides),
         const_cast<ScalarNoConst*>(block.data())));
   }
 };
 
-
-
 namespace internal {
 template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
 struct traits<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> > : public traits<XprType>
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 448f47f1d..d16ae4d11 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -317,7 +317,7 @@ static void test_execute_reshape(Device d)
 
   DSizes<Index, ReshapedDims> reshaped_dims;
   reshaped_dims[shuffle[0]] = dims[0] * dims[1];
-  for (int i = 2; i < NumDims; ++i) reshaped_dims[shuffle[i]] = dims[i];
+  for (int i = 1; i < ReshapedDims; ++i) reshaped_dims[shuffle[i]] = dims[i + 1];
 
   Tensor<T, ReshapedDims, Options, Index> golden = src.reshape(reshaped_dims);
 
diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp
index ab19b6e6b..467df39c7 100644
--- a/unsupported/test/cxx11_tensor_shuffling.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling.cpp
@@ -83,10 +83,10 @@ static void test_expr_shuffling()
 
   Tensor<float, 4, DataLayout> result(5,7,3,2);
 
-  array<int, 4> src_slice_dim{{2,3,1,7}};
-  array<int, 4> src_slice_start{{0,0,0,0}};
-  array<int, 4> dst_slice_dim{{1,7,3,2}};
-  array<int, 4> dst_slice_start{{0,0,0,0}};
+  array<ptrdiff_t, 4> src_slice_dim{{2,3,1,7}};
+  array<ptrdiff_t, 4> src_slice_start{{0,0,0,0}};
+  array<ptrdiff_t, 4> dst_slice_dim{{1,7,3,2}};
+  array<ptrdiff_t, 4> dst_slice_start{{0,0,0,0}};
 
   for (int i = 0; i < 5; ++i) {
     result.slice(dst_slice_start, dst_slice_dim) =

From 8603d80029581272a04f78959cb1636395fa47bc Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 9 Aug 2018 11:09:10 -0700
Subject: [PATCH 547/680] Cast diagonalSize() to Scalar before multiplication.
 Without this, automatic differentiation in Ceres breaks because Scalar is a
 custom type that does not support multiplication by Index.

---
 Eigen/src/LU/FullPivLU.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h
index 50d1bb41b..8541dfa02 100644
--- a/Eigen/src/LU/FullPivLU.h
+++ b/Eigen/src/LU/FullPivLU.h
@@ -53,7 +53,7 @@ template<typename _MatrixType> struct traits<FullPivLU<_MatrixType> >
   * Output: \verbinclude class_FullPivLU.out
   *
   * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
-  * 
+  *
   * \sa MatrixBase::fullPivLu(), MatrixBase::determinant(), MatrixBase::inverse()
   */
 template<typename _MatrixType> class FullPivLU
@@ -320,7 +320,7 @@ template<typename _MatrixType> class FullPivLU
       return m_usePrescribedThreshold ? m_prescribedThreshold
       // this formula comes from experimenting (see "LU precision tuning" thread on the list)
       // and turns out to be identical to Higham's formula used already in LDLt.
-                                      : NumTraits<Scalar>::epsilon() * m_lu.diagonalSize();
+          : NumTraits<Scalar>::epsilon() * Scalar(m_lu.diagonalSize());
     }
 
     /** \returns the rank of the matrix of which *this is the LU decomposition.

From bfc5091dd56a18e3f6ed1dd30509f45d0fba0828 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 9 Aug 2018 14:46:17 -0700
Subject: [PATCH 548/680] Cast to diagonalSize to RealScalar instead Scalar.

---
 Eigen/src/LU/FullPivLU.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h
index 8541dfa02..344ec8926 100644
--- a/Eigen/src/LU/FullPivLU.h
+++ b/Eigen/src/LU/FullPivLU.h
@@ -320,7 +320,7 @@ template<typename _MatrixType> class FullPivLU
       return m_usePrescribedThreshold ? m_prescribedThreshold
       // this formula comes from experimenting (see "LU precision tuning" thread on the list)
       // and turns out to be identical to Higham's formula used already in LDLt.
-          : NumTraits<Scalar>::epsilon() * Scalar(m_lu.diagonalSize());
+          : NumTraits<Scalar>::epsilon() * RealScalar(m_lu.diagonalSize());
     }
 
     /** \returns the rank of the matrix of which *this is the LU decomposition.

From c49e93440f85462728975b330e6534aeb37aa2d0 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 13 Aug 2018 15:53:31 -0700
Subject: [PATCH 549/680] SuiteSparse defines the macro SuiteSparse_long to
 control what type is used for 64bit integers. The default value of this macro
 on non-MSVC platforms is long and __int64 on MSVC. CholmodSupport defaults to
 using long for the long variants of CHOLMOD functions. This creates problems
 when SuiteSparse_long is different than long. So the correct thing to do here
 is to use SuiteSparse_long as the type instead of long.

---
 Eigen/src/CholmodSupport/CholmodSupport.h | 76 +++++++++++------------
 1 file changed, 38 insertions(+), 38 deletions(-)

diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h
index dc199ece6..adaf52858 100644
--- a/Eigen/src/CholmodSupport/CholmodSupport.h
+++ b/Eigen/src/CholmodSupport/CholmodSupport.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_CHOLMODSUPPORT_H
 #define EIGEN_CHOLMODSUPPORT_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -79,12 +79,12 @@ cholmod_sparse viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_StorageIndex> >
 
   res.dtype   = 0;
   res.stype   = -1;
-  
+
   if (internal::is_same<_StorageIndex,int>::value)
   {
     res.itype = CHOLMOD_INT;
   }
-  else if (internal::is_same<_StorageIndex,long>::value)
+  else if (internal::is_same<_StorageIndex,SuiteSparse_long>::value)
   {
     res.itype = CHOLMOD_LONG;
   }
@@ -95,9 +95,9 @@ cholmod_sparse viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_StorageIndex> >
 
   // setup res.xtype
   internal::cholmod_configure_matrix<_Scalar>::run(res);
-  
+
   res.stype = 0;
-  
+
   return res;
 }
 
@@ -121,7 +121,7 @@ template<typename _Scalar, int _Options, typename _Index, unsigned int UpLo>
 cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<const SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
 {
   cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.matrix().const_cast_derived()));
-  
+
   if(UpLo==Upper) res.stype =  1;
   if(UpLo==Lower) res.stype = -1;
   // swap stype for rowmajor matrices (only works for real matrices)
@@ -168,11 +168,11 @@ namespace internal {
 
 #define EIGEN_CHOLMOD_SPECIALIZE0(ret, name) \
     template<typename _StorageIndex> inline ret cm_ ## name       (cholmod_common &Common) { return cholmod_ ## name   (&Common); } \
-    template<>                       inline ret cm_ ## name<long> (cholmod_common &Common) { return cholmod_l_ ## name (&Common); }
+    template<>                       inline ret cm_ ## name<SuiteSparse_long> (cholmod_common &Common) { return cholmod_l_ ## name (&Common); }
 
 #define EIGEN_CHOLMOD_SPECIALIZE1(ret, name, t1, a1) \
     template<typename _StorageIndex> inline ret cm_ ## name       (t1& a1, cholmod_common &Common) { return cholmod_ ## name   (&a1, &Common); } \
-    template<>                       inline ret cm_ ## name<long> (t1& a1, cholmod_common &Common) { return cholmod_l_ ## name (&a1, &Common); }
+    template<>                       inline ret cm_ ## name<SuiteSparse_long> (t1& a1, cholmod_common &Common) { return cholmod_l_ ## name (&a1, &Common); }
 
 EIGEN_CHOLMOD_SPECIALIZE0(int, start)
 EIGEN_CHOLMOD_SPECIALIZE0(int, finish)
@@ -184,15 +184,15 @@ EIGEN_CHOLMOD_SPECIALIZE1(int, free_sparse, cholmod_sparse*, A)
 EIGEN_CHOLMOD_SPECIALIZE1(cholmod_factor*, analyze, cholmod_sparse, A)
 
 template<typename _StorageIndex> inline cholmod_dense*  cm_solve         (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_solve     (sys, &L, &B, &Common); }
-template<>                       inline cholmod_dense*  cm_solve<long>   (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_l_solve   (sys, &L, &B, &Common); }
+template<>                       inline cholmod_dense*  cm_solve<SuiteSparse_long>   (int sys, cholmod_factor& L, cholmod_dense&  B, cholmod_common &Common) { return cholmod_l_solve   (sys, &L, &B, &Common); }
 
 template<typename _StorageIndex> inline cholmod_sparse* cm_spsolve       (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_spsolve   (sys, &L, &B, &Common); }
-template<>                       inline cholmod_sparse* cm_spsolve<long> (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_l_spsolve (sys, &L, &B, &Common); }
+template<>                       inline cholmod_sparse* cm_spsolve<SuiteSparse_long> (int sys, cholmod_factor& L, cholmod_sparse& B, cholmod_common &Common) { return cholmod_l_spsolve (sys, &L, &B, &Common); }
 
 template<typename _StorageIndex>
 inline int  cm_factorize_p       (cholmod_sparse*  A, double beta[2], _StorageIndex* fset, std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_factorize_p   (A, beta, fset, fsize, L, &Common); }
 template<>
-inline int  cm_factorize_p<long> (cholmod_sparse*  A, double beta[2], long* fset,          std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_l_factorize_p (A, beta, fset, fsize, L, &Common); }
+inline int  cm_factorize_p<SuiteSparse_long> (cholmod_sparse*  A, double beta[2], SuiteSparse_long* fset,          std::size_t fsize, cholmod_factor* L, cholmod_common &Common) { return cholmod_l_factorize_p (A, beta, fset, fsize, L, &Common); }
 
 #undef EIGEN_CHOLMOD_SPECIALIZE0
 #undef EIGEN_CHOLMOD_SPECIALIZE1
@@ -254,10 +254,10 @@ class CholmodBase : public SparseSolverBase<Derived>
         internal::cm_free_factor<StorageIndex>(m_cholmodFactor, m_cholmod);
       internal::cm_finish<StorageIndex>(m_cholmod);
     }
-    
+
     inline StorageIndex cols() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
     inline StorageIndex rows() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
-    
+
     /** \brief Reports whether previous computation was successful.
       *
       * \returns \c Success if computation was successful,
@@ -276,11 +276,11 @@ class CholmodBase : public SparseSolverBase<Derived>
       factorize(matrix);
       return derived();
     }
-    
+
     /** Performs a symbolic decomposition on the sparsity pattern of \a matrix.
       *
       * This function is particularly useful when solving for several problems having the same structure.
-      * 
+      *
       * \sa factorize()
       */
     void analyzePattern(const MatrixType& matrix)
@@ -292,13 +292,13 @@ class CholmodBase : public SparseSolverBase<Derived>
       }
       cholmod_sparse A = viewAsCholmod(matrix.template selfadjointView<UpLo>());
       m_cholmodFactor = internal::cm_analyze<StorageIndex>(A, m_cholmod);
-      
+
       this->m_isInitialized = true;
       this->m_info = Success;
       m_analysisIsOk = true;
       m_factorizationIsOk = false;
     }
-    
+
     /** Performs a numeric decomposition of \a matrix
       *
       * The given matrix must have the same sparsity pattern as the matrix on which the symbolic decomposition has been performed.
@@ -315,11 +315,11 @@ class CholmodBase : public SparseSolverBase<Derived>
       this->m_info = (m_cholmodFactor->minor == m_cholmodFactor->n ? Success : NumericalIssue);
       m_factorizationIsOk = true;
     }
-    
+
     /** Returns a reference to the Cholmod's configuration structure to get a full control over the performed operations.
      *  See the Cholmod user guide for details. */
     cholmod_common& cholmod() { return m_cholmod; }
-    
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal */
     template<typename Rhs,typename Dest>
@@ -329,7 +329,7 @@ class CholmodBase : public SparseSolverBase<Derived>
       const Index size = m_cholmodFactor->n;
       EIGEN_UNUSED_VARIABLE(size);
       eigen_assert(size==b.rows());
-      
+
       // Cholmod needs column-major storage without inner-stride, which corresponds to the default behavior of Ref.
       Ref<const Matrix<typename Rhs::Scalar,Dynamic,Dynamic,ColMajor> > b_ref(b.derived());
 
@@ -345,7 +345,7 @@ class CholmodBase : public SparseSolverBase<Derived>
       dest = Matrix<Scalar,Dest::RowsAtCompileTime,Dest::ColsAtCompileTime>::Map(reinterpret_cast<Scalar*>(x_cd->x),b.rows(),b.cols());
       internal::cm_free_dense<StorageIndex>(x_cd, m_cholmod);
     }
-    
+
     /** \internal */
     template<typename RhsDerived, typename DestDerived>
     void _solve_impl(const SparseMatrixBase<RhsDerived> &b, SparseMatrixBase<DestDerived> &dest) const
@@ -370,8 +370,8 @@ class CholmodBase : public SparseSolverBase<Derived>
       internal::cm_free_sparse<StorageIndex>(x_cs, m_cholmod);
     }
     #endif // EIGEN_PARSED_BY_DOXYGEN
-    
-    
+
+
     /** Sets the shift parameter that will be used to adjust the diagonal coefficients during the numerical factorization.
       *
       * During the numerical factorization, an offset term is added to the diagonal coefficients:\n
@@ -386,7 +386,7 @@ class CholmodBase : public SparseSolverBase<Derived>
       m_shiftOffset[0] = double(offset);
       return derived();
     }
-    
+
     /** \returns the determinant of the underlying matrix from the current factorization */
     Scalar determinant() const
     {
@@ -441,7 +441,7 @@ class CholmodBase : public SparseSolverBase<Derived>
     template<typename Stream>
     void dumpMemory(Stream& /*s*/)
     {}
-    
+
   protected:
     mutable cholmod_common m_cholmod;
     cholmod_factor* m_cholmodFactor;
@@ -478,11 +478,11 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
 {
     typedef CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLLT> Base;
     using Base::m_cholmod;
-    
+
   public:
-    
+
     typedef _MatrixType MatrixType;
-    
+
     CholmodSimplicialLLT() : Base() { init(); }
 
     CholmodSimplicialLLT(const MatrixType& matrix) : Base()
@@ -529,11 +529,11 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
 {
     typedef CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLDLT> Base;
     using Base::m_cholmod;
-    
+
   public:
-    
+
     typedef _MatrixType MatrixType;
-    
+
     CholmodSimplicialLDLT() : Base() { init(); }
 
     CholmodSimplicialLDLT(const MatrixType& matrix) : Base()
@@ -578,11 +578,11 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
 {
     typedef CholmodBase<_MatrixType, _UpLo, CholmodSupernodalLLT> Base;
     using Base::m_cholmod;
-    
+
   public:
-    
+
     typedef _MatrixType MatrixType;
-    
+
     CholmodSupernodalLLT() : Base() { init(); }
 
     CholmodSupernodalLLT(const MatrixType& matrix) : Base()
@@ -629,11 +629,11 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
 {
     typedef CholmodBase<_MatrixType, _UpLo, CholmodDecomposition> Base;
     using Base::m_cholmod;
-    
+
   public:
-    
+
     typedef _MatrixType MatrixType;
-    
+
     CholmodDecomposition() : Base() { init(); }
 
     CholmodDecomposition(const MatrixType& matrix) : Base()
@@ -643,7 +643,7 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     }
 
     ~CholmodDecomposition() {}
-    
+
     void setMode(CholmodMode mode)
     {
       switch(mode)

From eabc7a4031cf0f5702fa16009ef33f87c1d51873 Mon Sep 17 00:00:00 2001
From: Justin Carpentier <jcarpent@laas.fr>
Date: Fri, 10 Aug 2018 14:30:06 +0200
Subject: [PATCH 550/680] PR 465: Fix issue in RowMajor assignment in
 plain_matrix_type_row_major::type

The type should be RowMajor
---
 Eigen/src/Core/util/XprHelper.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index e3231c712..836ff4711 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -405,7 +405,7 @@ template<typename T> struct plain_matrix_type_row_major
   typedef Matrix<typename traits<T>::Scalar,
                 Rows,
                 Cols,
-                (MaxCols==1&&MaxRows!=1) ? RowMajor : ColMajor,
+                (MaxCols==1&&MaxRows!=1) ? ColMajor : RowMajor,
                 MaxRows,
                 MaxCols
           > type;

From 4be42862241d4f976d8ed92a2cc0ad5b06520ef9 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 10 Aug 2018 11:32:58 -0700
Subject: [PATCH 551/680] Made the code compile with gcc 5.4.

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 22 +++++++++----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 45ddfdb39..cbf91013b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -60,7 +60,7 @@ struct cond<RowMajor> {
  *  - kSkewedInnerDims: 100 blocks of size 100x1 (or 1x100 depending on a column
  *                      or row major layout)
  */
-enum class TensorBlockShapeType {
+enum TensorBlockShapeType {
   kUniformAllDims,
   kSkewedInnerDims,
 };
@@ -152,11 +152,11 @@ struct TensorBlockCopyOp {
     const Scalar* src_base = &src_data[src_index];
     Scalar* dst_base = &dst_data[dst_index];
 
-    using Src = const Eigen::Array<Scalar, Dynamic, 1>;
-    using Dst = Eigen::Array<Scalar, Dynamic, 1>;
+    typedef const Eigen::Array<Scalar, Dynamic, 1> Src;
+    typedef Eigen::Array<Scalar, Dynamic, 1> Dst;
 
-    using SrcMap = Eigen::Map<Src, 0, InnerStride<>>;
-    using DstMap = Eigen::Map<Dst, 0, InnerStride<>>;
+    typedef Eigen::Map<Src, 0, InnerStride<>> SrcMap;
+    typedef Eigen::Map<Dst, 0, InnerStride<>> DstMap;
 
     const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
     DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
@@ -401,13 +401,13 @@ struct TensorBlockCwiseBinaryOp {
       const StorageIndex left_stride, const LeftScalar* left_data,
       const StorageIndex right_index, const StorageIndex right_stride,
       const RightScalar* right_data) {
-    using Lhs = const Eigen::Array<LeftScalar, Dynamic, 1>;
-    using Rhs = const Eigen::Array<RightScalar, Dynamic, 1>;
-    using Out = Eigen::Array<OutputScalar, Dynamic, 1>;
+    typedef const Eigen::Array<LeftScalar, Dynamic, 1> Lhs;
+    typedef const Eigen::Array<RightScalar, Dynamic, 1> Rhs;
+    typedef Eigen::Array<OutputScalar, Dynamic, 1> Out;
 
-    using LhsMap = Eigen::Map<Lhs, 0, InnerStride<>>;
-    using RhsMap = Eigen::Map<Rhs, 0, InnerStride<>>;
-    using OutMap = Eigen::Map<Out, 0, InnerStride<>>;
+    typedef Eigen::Map<Lhs, 0, InnerStride<>> LhsMap;
+    typedef Eigen::Map<Rhs, 0, InnerStride<>> RhsMap;
+    typedef Eigen::Map<Out, 0, InnerStride<>> OutMap;
 
     const LeftScalar* lhs_base = &left_data[left_index];
     const RightScalar* rhs_base = &right_data[right_index];

From c8ea39867573b36c72c5342393f4e3ee22ab6406 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 10 Aug 2018 13:02:41 -0700
Subject: [PATCH 552/680] Avoided language features that are only available in
 cxx11 mode.

---
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   | 34 ++++++++-----------
 1 file changed, 15 insertions(+), 19 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 17008917a..3f3b5685d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -36,7 +36,7 @@ template <typename Expression, typename Device, bool Vectorizable,
           bool Tileable>
 class TensorExecutor {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
@@ -60,7 +60,7 @@ template <typename Expression>
 class TensorExecutor<Expression, DefaultDevice, /*Vectorizable*/ true,
                      /*Tileable*/ false> {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
@@ -102,21 +102,19 @@ template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, DefaultDevice, Vectorizable,
                      /*Tileable*/ true> {
  public:
-  using Scalar = typename traits<Expression>::Scalar;
-  using ScalarNoConst = typename remove_const<Scalar>::type;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
 
-  using Evaluator = TensorEvaluator<Expression, DefaultDevice>;
-  using StorageIndex = typename traits<Expression>::Index;
+  typedef TensorEvaluator<Expression, DefaultDevice> Evaluator;
+  typedef typename traits<Expression>::Index StorageIndex;
 
   static const int NumDims = traits<Expression>::NumDimensions;
 
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const DefaultDevice& device = DefaultDevice()) {
-    using TensorBlock =
-        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
-    using TensorBlockMapper = TensorBlockMapper<ScalarNoConst, StorageIndex,
-                                                NumDims, Evaluator::Layout>;
+    typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlock;
+    typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlockMapper;
 
     Evaluator evaluator(expr, device);
     Index total_size = array_prod(evaluator.dimensions());
@@ -221,7 +219,7 @@ struct EvalRange<Evaluator, StorageIndex, /*Vectorizable*/ true> {
 template <typename Expression, bool Vectorizable, bool Tileable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
  public:
-  using StorageIndex = typename Expression::Index;
+  typedef typename Expression::Index StorageIndex;
 
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
@@ -249,20 +247,18 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
 template <typename Expression, bool Vectorizable>
 class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ true> {
  public:
-  using Scalar = typename traits<Expression>::Scalar;
-  using ScalarNoConst = typename remove_const<Scalar>::type;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
 
-  using Evaluator = TensorEvaluator<Expression, ThreadPoolDevice>;
-  using StorageIndex = typename traits<Expression>::Index;
+  typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+  typedef typename traits<Expression>::Index StorageIndex;
 
   static const int NumDims = traits<Expression>::NumDimensions;
 
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
-    using TensorBlock =
-        TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
-    using TensorBlockMapper =
-        TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout>;
+    typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlock;
+    typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlockMapper;
 
     Evaluator evaluator(expr, device);
     StorageIndex total_size = array_prod(evaluator.dimensions());

From f2209d06e428e0691de71f30fc2db4cb29191cd2 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 10 Aug 2018 16:53:36 -0700
Subject: [PATCH 553/680] Add block evaluationto CwiseUnaryOp and add
 PreferBlockAccess enum to all evaluators

---
 .../Eigen/CXX11/src/Tensor/TensorArgMax.h     |   2 +
 .../Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h |   1 +
 .../Eigen/CXX11/src/Tensor/TensorAssign.h     |  18 ++-
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 141 ++++++++++++++++++
 .../CXX11/src/Tensor/TensorBroadcasting.h     |  11 +-
 .../Eigen/CXX11/src/Tensor/TensorChipping.h   |  13 +-
 .../CXX11/src/Tensor/TensorConcatenation.h    |   2 +
 .../CXX11/src/Tensor/TensorContraction.h      |   1 +
 .../Eigen/CXX11/src/Tensor/TensorConversion.h |   1 +
 .../CXX11/src/Tensor/TensorConvolution.h      |   2 +
 .../CXX11/src/Tensor/TensorConvolutionSycl.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorCustomOp.h   |   2 +
 .../Eigen/CXX11/src/Tensor/TensorEvalTo.h     |   1 +
 .../Eigen/CXX11/src/Tensor/TensorEvaluator.h  |  76 +++++++---
 .../Eigen/CXX11/src/Tensor/TensorFFT.h        |   1 +
 .../Eigen/CXX11/src/Tensor/TensorFixedSize.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorForcedEval.h |   1 +
 .../src/Tensor/TensorForwardDeclarations.h    |   7 +-
 .../Eigen/CXX11/src/Tensor/TensorGenerator.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h |  13 +-
 .../Eigen/CXX11/src/Tensor/TensorInflation.h  |   1 +
 .../Eigen/CXX11/src/Tensor/TensorLayoutSwap.h |   2 +
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h   |  46 +++---
 .../Eigen/CXX11/src/Tensor/TensorPadding.h    |   1 +
 .../Eigen/CXX11/src/Tensor/TensorPatch.h      |   1 +
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  13 +-
 .../Eigen/CXX11/src/Tensor/TensorRef.h        |   3 +
 .../Eigen/CXX11/src/Tensor/TensorReverse.h    |   2 +
 .../Eigen/CXX11/src/Tensor/TensorScan.h       |   1 +
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h  |  24 +--
 .../Eigen/CXX11/src/Tensor/TensorStriding.h   |   2 +
 .../Eigen/CXX11/src/Tensor/TensorTrace.h      |   1 +
 .../CXX11/src/Tensor/TensorVolumePatch.h      |   1 +
 .../test/cxx11_tensor_block_access.cpp        | 111 ++++++++++++++
 unsupported/test/cxx11_tensor_executor.cpp    |  31 ++++
 35 files changed, 454 insertions(+), 82 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
index c0f33ba2d..ea3ab329d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@@ -87,6 +87,7 @@ struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
     PacketAccess = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -220,6 +221,7 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
     PacketAccess = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
index 442639868..5110e99ee 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMaxSycl.h
@@ -109,6 +109,7 @@ struct TensorEvaluator<const TensorTupleReducerDeviceOp<StrideDims, ArgType>, Sy
     IsAligned =  false,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, SyclKernelDevice>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
index f1f877c16..9ec1ec726 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
@@ -102,14 +102,16 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
   static const int NumDims = XprType::NumDims;
 
   enum {
-    IsAligned    = TensorEvaluator<LeftArgType, Device>::IsAligned &
-                   TensorEvaluator<RightArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess &
-                   TensorEvaluator<RightArgType, Device>::PacketAccess,
-    BlockAccess  = TensorEvaluator<LeftArgType, Device>::BlockAccess &
-                   TensorEvaluator<RightArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<LeftArgType, Device>::Layout,
-    RawAccess    = TensorEvaluator<LeftArgType, Device>::RawAccess
+    IsAligned         = TensorEvaluator<LeftArgType, Device>::IsAligned &
+                        TensorEvaluator<RightArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &
+                        TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<LeftArgType, Device>::BlockAccess &
+                        TensorEvaluator<RightArgType, Device>::BlockAccess,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess |
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    RawAccess         = TensorEvaluator<LeftArgType, Device>::RawAccess
   };
 
   typedef typename internal::TensorBlock<
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 877603421..4a3e1ac17 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -381,6 +381,147 @@ class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
   }
 };
 
+/**
+ * \class TensorBlockCwiseUnaryOp
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Carries out a cwise binary op on a number of coefficients.
+ *
+ * This class reads strided input from the argument, and writes the
+ * result of the cwise unary op to the strided output array.
+ *
+ */
+struct TensorBlockCwiseUnaryOp {
+  template <typename StorageIndex, typename UnaryFunctor,
+            typename OutputScalar, typename InputScalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const UnaryFunctor& functor, const StorageIndex num_coeff,
+      const StorageIndex output_index, const StorageIndex output_stride,
+      OutputScalar* output_data, const StorageIndex input_index,
+      const StorageIndex input_stride, const InputScalar* input_data) {
+    typedef const Eigen::Array<InputScalar, Dynamic, 1> Input;
+    typedef Eigen::Array<OutputScalar, Dynamic, 1> Output;
+
+    typedef Eigen::Map<Input, 0, InnerStride<>> InputMap;
+    typedef Eigen::Map<Output, 0, InnerStride<>> OutputMap;
+
+    const InputScalar* input_base = &input_data[input_index];
+    OutputScalar* output_base = &output_data[output_index];
+
+    const InputMap input(input_base, num_coeff, InnerStride<>(input_stride));
+    OutputMap output(output_base, num_coeff, InnerStride<>(output_stride));
+
+    output = Eigen::CwiseUnaryOp<UnaryFunctor, InputMap>(input, functor);
+  }
+};
+
+/**
+ * \class TensorBlockCwiseUnaryIO
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor block IO class for carrying out cwise unary ops.
+ *
+ * This class carries out the unary op on given blocks.
+ */
+template <typename UnaryFunctor, typename StorageIndex, typename OutputScalar,
+          int NumDims, int Layout>
+struct TensorBlockCwiseUnaryIO {
+  typedef typename internal::TensorBlock<OutputScalar, StorageIndex, NumDims,
+                                         Layout>::Dimensions Dimensions;
+
+  struct BlockIteratorState {
+    StorageIndex output_stride, output_span;
+    StorageIndex input_stride, input_span;
+    StorageIndex size, count;
+  };
+
+  template <typename InputScalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const UnaryFunctor& functor, const Dimensions& block_sizes,
+      const Dimensions& block_strides, OutputScalar* output_data,
+      const array<StorageIndex, NumDims>& input_strides,
+      const InputScalar* input_data) {
+    // Find the innermost dimension whose size is not 1. This is the effective
+    // inner dim. If all dimensions are of size 1, fallback to using the actual
+    // innermost dim to avoid out-of-bound access.
+    int num_size_one_inner_dims = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      if (block_sizes[dim] != 1) {
+        num_size_one_inner_dims = i;
+        break;
+      }
+    }
+    // Calculate strides and dimensions.
+    const int inner_dim =
+        NumDims == 0 ? 1
+                     : cond<Layout>()(num_size_one_inner_dims,
+                                      NumDims - num_size_one_inner_dims - 1);
+    StorageIndex inner_dim_size = NumDims == 0 ? 1 : block_sizes[inner_dim];
+    for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
+      const int dim = cond<Layout>()(i, NumDims - i - 1);
+      // Merge multiple inner dims into one for larger inner dim size (i.e.
+      // fewer calls to TensorBlockCwiseUnaryOp::Run()).
+      if (inner_dim_size == block_strides[dim] &&
+          block_strides[dim] == input_strides[dim]) {
+        inner_dim_size *= block_sizes[dim];
+        ++num_size_one_inner_dims;
+      } else {
+        break;
+      }
+    }
+
+    StorageIndex output_index = 0, input_index = 0;
+
+    const StorageIndex output_stride =
+        NumDims == 0 ? 1 : block_strides[inner_dim];
+    const StorageIndex input_stride =
+        NumDims == 0 ? 1 : input_strides[inner_dim];
+
+    const int at_least_1_dim = NumDims <= 1 ? 1 : NumDims - 1;
+    array<BlockIteratorState, at_least_1_dim> block_iter_state;
+
+    // Initialize block iterator state. Squeeze away any dimension of size 1.
+    int num_squeezed_dims = 0;
+    for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
+      const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
+      const StorageIndex size = block_sizes[dim];
+      if (size == 1) {
+        continue;
+      }
+      BlockIteratorState& state = block_iter_state[num_squeezed_dims];
+      state.output_stride = block_strides[dim];
+      state.input_stride = input_strides[dim];
+      state.size = size;
+      state.output_span = state.output_stride * (size - 1);
+      state.input_span = state.input_stride * (size - 1);
+      state.count = 0;
+      ++num_squeezed_dims;
+    }
+
+    // Compute cwise unary op.
+    const StorageIndex block_total_size =
+        NumDims == 0 ? 1 : block_sizes.TotalSize();
+    for (StorageIndex i = 0; i < block_total_size; i += inner_dim_size) {
+      TensorBlockCwiseUnaryOp::Run(functor, inner_dim_size, output_index,
+                                   output_stride, output_data, input_index,
+                                   input_stride, input_data);
+      // Update index.
+      for (int j = 0; j < num_squeezed_dims; ++j) {
+        auto& state = block_iter_state[j];
+        if (++state.count < state.size) {
+          output_index += state.output_stride;
+          input_index += state.input_stride;
+          break;
+        }
+        state.count = 0;
+        output_index -= state.output_span;
+        input_index -= state.input_span;
+      }
+    }
+  }
+};
+
 /**
  * \class TensorBlockCwiseBinaryOp
  * \ingroup CXX11_Tensor_Module
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index a4d750885..5e812b04d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -108,11 +108,12 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   bool isCopy= false, nByOne = false, oneByN = false;
 
   enum {
-    IsAligned    = true,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    RawAccess    = false
+    IsAligned         = true,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index b4c4162ef..76fab39e2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -144,12 +144,13 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   enum {
     // Alignment can't be guaranteed at compile time since it depends on the
     // slice offsets.
-    IsAligned    = false,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,  // to be implemented
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
index 27c92d8f6..3863ee8c3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@@ -123,6 +123,7 @@ struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgTy
     IsAligned = false,
     PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     RawAccess = false
   };
@@ -308,6 +309,7 @@ template<typename Axis, typename LeftArgType, typename RightArgType, typename De
     IsAligned = false,
     PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index e604456e8..c459fc649 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -241,6 +241,7 @@ struct TensorContractionEvaluatorBase
     IsAligned = true,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
index a7751eee1..1f613d3c7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
@@ -196,6 +196,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
     IsAligned = false,
     PacketAccess = true,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = false
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index a07e32db0..2d0e6599f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -308,6 +308,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
     IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<InputArgType, Device>::PacketAccess & TensorEvaluator<KernelArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<InputArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -780,6 +781,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
     IsAligned = TensorEvaluator<InputArgType, GpuDevice>::IsAligned & TensorEvaluator<KernelArgType, GpuDevice>::IsAligned,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<InputArgType, GpuDevice>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
index d301d0c01..e79958fc9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -243,6 +243,7 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
     IsAligned = TensorEvaluator<InputArgType, const Eigen::SyclDevice>::IsAligned & TensorEvaluator<KernelArgType, const Eigen::SyclDevice>::IsAligned,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<InputArgType, const Eigen::SyclDevice>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 47b5a5a5e..d71b2e34b 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -93,6 +93,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
     IsAligned = false,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<XprType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -255,6 +256,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
     IsAligned = false,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<LhsXprType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
index 256d499f2..554ee5f59 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -108,6 +108,7 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType, MakePointer_>, Device>
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index 028902fea..e30f9ad8e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -43,6 +43,7 @@ struct TensorEvaluator
     IsAligned = Derived::IsAligned,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
+    PreferBlockAccess = false,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
     RawAccess = true
@@ -195,6 +196,7 @@ struct TensorEvaluator<const Derived, Device>
     IsAligned = Derived::IsAligned,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
+    PreferBlockAccess = false,
     Layout = Derived::Layout,
     CoordAccess = NumCoords > 0,
     RawAccess = true
@@ -288,6 +290,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
     IsAligned = true,
     PacketAccess = internal::functor_traits<NullaryOp>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -351,27 +354,34 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
   typedef TensorCwiseUnaryOp<UnaryOp, ArgType> XprType;
 
   enum {
-    IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess &
-                   internal::functor_traits<UnaryOp>::PacketAccess,
-    BlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess = false,  // to be implemented
-    RawAccess = false
+    IsAligned          = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess       = TensorEvaluator<ArgType, Device>::PacketAccess &
+                         internal::functor_traits<UnaryOp>::PacketAccess,
+    BlockAccess        = TensorEvaluator<ArgType, Device>::PacketAccess,
+    PreferBlockAccess  = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout             = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess        = false,  // to be implemented
+    RawAccess          = false
   };
 
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
-    : m_functor(op.functor()),
+    : m_device(device),
+      m_functor(op.functor()),
       m_argImpl(op.nestedExpression(), device)
   { }
 
   typedef typename XprType::Index Index;
   typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
   typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
   typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
 
+  static const int NumDims = internal::array_size<Dimensions>::value;
+  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
+      TensorBlock;
+
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
@@ -399,6 +409,29 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
         TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>* resources) const {
+    m_argImpl.getResourceRequirements(resources);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
+      TensorBlock* output_block) const {
+    if (NumDims <= 0) {
+      output_block->data()[0] = coeff(0);
+      return;
+    }
+    internal::TensorBlockView<ArgType, Device> arg_block(m_device, m_argImpl,
+                                                         *output_block);
+    internal::TensorBlockCwiseUnaryIO<UnaryOp, Index, ScalarNoConst, NumDims,
+                                      Layout>::Run(m_functor,
+                                                   output_block->block_sizes(),
+                                                   output_block
+                                                       ->block_strides(),
+                                                   output_block->data(),
+                                                   arg_block.block_strides(),
+                                                   arg_block.data());
+  }
+
   EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
 
   /// required by sycl in order to extract the accessor
@@ -408,6 +441,7 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
 
 
  private:
+  const Device& m_device;
   const UnaryOp m_functor;
   TensorEvaluator<ArgType, Device> m_argImpl;
 };
@@ -421,16 +455,18 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   typedef TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType> XprType;
 
   enum {
-    IsAligned    = TensorEvaluator<LeftArgType, Device>::IsAligned &
-                   TensorEvaluator<RightArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess &
-                   TensorEvaluator<RightArgType, Device>::PacketAccess &
-                   internal::functor_traits<BinaryOp>::PacketAccess,
-    BlockAccess  = TensorEvaluator<LeftArgType, Device>::BlockAccess &
-                   TensorEvaluator<RightArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<LeftArgType, Device>::Layout,
-    CoordAccess  = false,  // to be implemented
-    RawAccess    = false
+    IsAligned         = TensorEvaluator<LeftArgType, Device>::IsAligned &
+                        TensorEvaluator<RightArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &
+                        TensorEvaluator<RightArgType, Device>::PacketAccess &
+                        internal::functor_traits<BinaryOp>::PacketAccess,
+    BlockAccess       = TensorEvaluator<LeftArgType, Device>::BlockAccess &
+                        TensorEvaluator<RightArgType, Device>::BlockAccess,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess |
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
   };
 
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
@@ -501,7 +537,7 @@ struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArg
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
       TensorBlock* output_block) const {
     if (NumDims <= 0) {
-      output_block->data()[0] = coeff(0);
+      output_block->data()[0] = coeff(Index(0));
       return;
     }
     internal::TensorBlockView<LeftArgType, Device> left_block(
@@ -543,6 +579,7 @@ struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type,
     PacketAccess = TensorEvaluator<Arg1Type, Device>::PacketAccess & TensorEvaluator<Arg2Type, Device>::PacketAccess & TensorEvaluator<Arg3Type, Device>::PacketAccess &
                    internal::functor_traits<TernaryOp>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<Arg1Type, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -648,6 +685,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>
     PacketAccess = TensorEvaluator<ThenArgType, Device>::PacketAccess & TensorEvaluator<ElseArgType, Device>::PacketAccess &
                     PacketType<Scalar, Device>::HasBlend,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<IfArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
index d6ab4d997..480cf1f39 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
@@ -136,6 +136,7 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
     IsAligned = false,
     PacketAccess = true,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
index 1342e47a6..71ba56773 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h
@@ -42,6 +42,7 @@ class TensorFixedSize : public TensorBase<TensorFixedSize<Scalar_, Dimensions_,
       IsAligned = bool(EIGEN_MAX_ALIGN_BYTES>0),
       PacketAccess = (internal::packet_traits<Scalar>::size > 1),
       BlockAccess = false,
+      PreferBlockAccess = false,
       Layout = Options_ & RowMajor ? RowMajor : ColMajor,
       CoordAccess = true,
       RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index 2778bf5ec..edf9f85d8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -99,6 +99,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     IsAligned = true,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = true
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 0dd524a30..b8a57ab99 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -132,7 +132,12 @@ struct IsVectorizable<GpuDevice, Expression> {
 
 template <typename Device, typename Expression>
 struct IsTileable {
-  static const bool value = TensorEvaluator<Expression, Device>::BlockAccess;
+  // Check that block evaluation is supported and it's a preferred option (at
+  // least one sub-expression has much faster block evaluation, e.g.
+  // broadcasting).
+  static const bool value =
+      TensorEvaluator<Expression, Device>::BlockAccess &
+      TensorEvaluator<Expression, Device>::PreferBlockAccess;
 };
 
 template <typename Expression, typename Device,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
index 97c8d4a02..95c9e6aee 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGenerator.h
@@ -92,6 +92,7 @@ struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
     IsAligned = false,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 2b2f4a650..1826d7022 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -244,12 +244,13 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
-    IsAligned    = false,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess  = true,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = true,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = false
   };
 
   typedef internal::TensorBlock<Scalar, Index, NumDims, Layout>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
index 64f2ad81f..e28565009 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h
@@ -91,6 +91,7 @@ struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
index e3165fa10..998757d14 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
@@ -120,6 +120,7 @@ struct TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
     CoordAccess = false,  // to be implemented
     RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@@ -183,6 +184,7 @@ template<typename ArgType, typename Device>
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
     CoordAccess = false  // to be implemented
   };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index b25c1eabc..6f9294ccf 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -111,16 +111,17 @@ struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
   static const int NumInputDims  = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
 
   enum {
-    IsAligned    = TensorEvaluator<ArgType, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    IsAligned         = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
     // TODO(andydavis, wuke) Enable BlockAccess for the general case when the
     // performance issue with block-based reshape is resolved.
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess &&
-                   TensorEvaluator<ArgType, Device>::RawAccess &&
-                   NumInputDims > 0 && NumOutputDims > 0,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,  // to be implemented
-    RawAccess    = TensorEvaluator<ArgType, Device>::RawAccess
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess &&
+                        TensorEvaluator<ArgType, Device>::RawAccess &&
+                        NumInputDims > 0 && NumOutputDims > 0,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = TensorEvaluator<ArgType, Device>::RawAccess
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
@@ -349,6 +350,7 @@ template<typename NewDimensions, typename ArgType, typename Device>
     IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@@ -508,12 +510,13 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   enum {
     // Alignment can't be guaranteed at compile time since it depends on the
     // slice offsets and sizes.
-    IsAligned    = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
@@ -785,12 +788,13 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
   typedef Sizes Dimensions;
 
   enum {
-    IsAligned    = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
-    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,
-    RawAccess    = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
@@ -972,6 +976,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
     IsAligned = false,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess = false
   };
@@ -1148,6 +1153,7 @@ struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Stride
     IsAligned = false,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = TensorEvaluator<ArgType, Device>::CoordAccess,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
index aa1db3c73..59c1704ed 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
@@ -97,6 +97,7 @@ struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device
     IsAligned = true,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = true,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index a0a1ad8f4..4292fe0c2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -95,6 +95,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 80f179ba4..3d534eaa2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -475,12 +475,13 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
-    IsAligned    = false,
-    PacketAccess = Self::InputPacketAccess && Op::PacketAccess,
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,  // to be implemented
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = Self::InputPacketAccess && Op::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
index a6cade50f..6e15e75f9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h
@@ -137,6 +137,7 @@ template<typename PlainObjectType> class TensorRef : public TensorBase<TensorRef
       IsAligned = false,
       PacketAccess = false,
       BlockAccess = false,
+      PreferBlockAccess = false,
       Layout = PlainObjectType::Layout,
       CoordAccess = false,  // to be implemented
       RawAccess = false
@@ -366,6 +367,7 @@ struct TensorEvaluator<const TensorRef<Derived>, Device>
     IsAligned = false,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorRef<Derived>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -414,6 +416,7 @@ struct TensorEvaluator<TensorRef<Derived>, Device> : public TensorEvaluator<cons
     IsAligned = false,
     PacketAccess = false,
     BlockAccess = false,
+    PreferBlockAccess = false,
     RawAccess = false
   };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
index 9193bdd8e..b7fb969f3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h
@@ -114,6 +114,7 @@ struct TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>, Device
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -255,6 +256,7 @@ struct TensorEvaluator<TensorReverseOp<ReverseDimensions, ArgType>, Device>
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
index b1135f297..641366d9d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
@@ -97,6 +97,7 @@ struct TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> {
     IsAligned = false,
     PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = true
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index 98f125408..a5b541a68 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -111,12 +111,13 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
-    IsAligned    = false,
-    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    CoordAccess  = false,  // to be implemented
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
@@ -412,11 +413,12 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
-    IsAligned    = false,
-    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
-    BlockAccess  = TensorEvaluator<ArgType, Device>::BlockAccess,
-    Layout       = TensorEvaluator<ArgType, Device>::Layout,
-    RawAccess    = false
+    IsAligned         = false,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = false
   };
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
index 4b69072f2..221dc96c9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
@@ -113,6 +113,7 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
@@ -275,6 +276,7 @@ struct TensorEvaluator<TensorStridingOp<Strides, ArgType>, Device>
     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
index ea53bb04b..9fc54a4c0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@@ -96,6 +96,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
index 3c7d8bbc0..c1b7a58ca 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@@ -200,6 +200,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
     IsAligned = false,
     PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
     BlockAccess = false,
+    PreferBlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,
     RawAccess = false
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 6feeff231..746f62511 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -517,6 +517,114 @@ static void test_block_io_squeeze_ones() {
   }
 }
 
+template <typename T, int NumDims, int Layout>
+static void test_block_cwise_unary_io_basic() {
+  typedef internal::scalar_square_op<T> UnaryFunctor;
+  typedef internal::TensorBlockCwiseUnaryIO<UnaryFunctor, Index, T, NumDims,
+                                            Layout>
+      TensorBlockCwiseUnaryIO;
+
+  DSizes<Index, NumDims> block_sizes = RandomDims<NumDims>();
+  DSizes<Index, NumDims> strides(ComputeStrides<Layout, NumDims>(block_sizes));
+
+  const auto total_size = block_sizes.TotalSize();
+
+  // Create a random input tensors.
+  T* input_data = GenerateRandomData<T>(total_size);
+
+  T* output_data = new T[total_size];
+  UnaryFunctor functor;
+  TensorBlockCwiseUnaryIO::Run(functor, block_sizes, strides, output_data,
+                               strides, input_data);
+  for (int i = 0; i < total_size; ++i) {
+    VERIFY_IS_EQUAL(output_data[i], functor(input_data[i]));
+  }
+
+  delete[] input_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_block_cwise_unary_io_squeeze_ones() {
+  typedef internal::scalar_square_op<float> UnaryFunctor;
+  typedef internal::TensorBlockCwiseUnaryIO<UnaryFunctor, Index, float, 5,
+                                            Layout>
+      TensorBlockCwiseUnaryIO;
+
+  DSizes<Index, 5> block_sizes(1, 2, 1, 3, 1);
+  DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
+
+  const auto total_size = block_sizes.TotalSize();
+
+  // Create a random input tensors.
+  auto* input_data = GenerateRandomData<float>(total_size);
+
+  auto* output_data = new float[total_size];
+  UnaryFunctor functor;
+  TensorBlockCwiseUnaryIO::Run(functor, block_sizes, strides, output_data,
+                               strides, input_data);
+  for (int i = 0; i < total_size; ++i) {
+    VERIFY_IS_EQUAL(output_data[i], functor(input_data[i]));
+  }
+
+  delete[] input_data;
+  delete[] output_data;
+}
+
+template <int Layout>
+static void test_block_cwise_unary_io_zero_strides() {
+  typedef internal::scalar_square_op<float> UnaryFunctor;
+  typedef internal::TensorBlockCwiseUnaryIO<UnaryFunctor, Index, float, 5,
+                                            Layout>
+      TensorBlockCwiseUnaryIO;
+
+  DSizes<Index, 5> rnd_dims = RandomDims<5>();
+
+  DSizes<Index, 5> input_sizes = rnd_dims;
+  input_sizes[0] = 1;
+  input_sizes[2] = 1;
+  input_sizes[4] = 1;
+
+  DSizes<Index, 5> input_strides(ComputeStrides<Layout, 5>(input_sizes));
+  input_strides[0] = 0;
+  input_strides[2] = 0;
+  input_strides[4] = 0;
+
+  // Generate random data.
+  auto* input_data = GenerateRandomData<float>(input_sizes.TotalSize());
+
+  DSizes<Index, 5> output_sizes = rnd_dims;
+  DSizes<Index, 5> output_strides(ComputeStrides<Layout, 5>(output_sizes));
+
+  const auto output_total_size = output_sizes.TotalSize();
+  auto* output_data = new float[output_total_size];
+
+  UnaryFunctor functor;
+  TensorBlockCwiseUnaryIO::Run(functor, output_sizes, output_strides,
+                               output_data, input_strides, input_data);
+  for (int i = 0; i < rnd_dims[0]; ++i) {
+    for (int j = 0; j < rnd_dims[1]; ++j) {
+      for (int k = 0; k < rnd_dims[2]; ++k) {
+        for (int l = 0; l < rnd_dims[3]; ++l) {
+          for (int m = 0; m < rnd_dims[4]; ++m) {
+            Index output_index = i * output_strides[0] + j * output_strides[1] +
+                                 k * output_strides[2] + l * output_strides[3] +
+                                 m * output_strides[4];
+            Index input_index = i * input_strides[0] + j * input_strides[1] +
+                                k * input_strides[2] + l * input_strides[3] +
+                                m * input_strides[4];
+            VERIFY_IS_EQUAL(output_data[output_index],
+                            functor(input_data[input_index]));
+          }
+        }
+      }
+    }
+  }
+
+  delete[] input_data;
+  delete[] output_data;
+}
+
 template <typename T, int NumDims, int Layout>
 static void test_block_cwise_binary_io_basic() {
   typedef internal::scalar_sum_op<T> BinaryFunctor;
@@ -982,6 +1090,9 @@ EIGEN_DECLARE_TEST(cxx11_tensor_block_access) {
   TEST_LAYOUTS_AND_DIMS(Data, test_block_io_copy_using_reordered_dimensions);
   TEST_LAYOUTS(test_block_io_zero_stride);
   TEST_LAYOUTS(test_block_io_squeeze_ones);
+  TEST_LAYOUTS_AND_DIMS(float, test_block_cwise_unary_io_basic);
+  TEST_LAYOUTS(test_block_cwise_unary_io_squeeze_ones);
+  TEST_LAYOUTS(test_block_cwise_unary_io_zero_strides);
   TEST_LAYOUTS_AND_DIMS(float, test_block_cwise_binary_io_basic);
   TEST_LAYOUTS(test_block_cwise_binary_io_squeeze_ones);
   TEST_LAYOUTS(test_block_cwise_binary_io_zero_strides);
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index d16ae4d11..1dc18220c 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -29,6 +29,33 @@ static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
   return dims;
 };
 
+template <typename T, int NumDims, typename Device, bool Vectorizable,
+          bool Tileable, int Layout>
+static void test_execute_unary_expr(Device d) {
+  static constexpr int Options = 0 | Layout;
+
+  // Pick a large enough tensor size to bypass small tensor block evaluation
+  // optimization.
+  auto dims = RandomDims<NumDims>(50 / NumDims, 100 / NumDims);
+
+  Tensor<T, NumDims, Options, Index> src(dims);
+  Tensor<T, NumDims, Options, Index> dst(dims);
+
+  src.setRandom();
+  const auto expr = src.square();
+
+  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
+  using Executor =
+      internal::TensorExecutor<const Assign, Device, Vectorizable, Tileable>;
+
+  Executor::run(Assign(dst, expr), d);
+
+  for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
+    T square = src.coeff(i) * src.coeff(i);
+    VERIFY_IS_EQUAL(square, dst.coeff(i));
+  }
+}
+
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
 static void test_execute_binary_expr(Device d)
@@ -445,6 +472,10 @@ EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
   Eigen::ThreadPool tp(num_threads);
   Eigen::ThreadPoolDevice tp_device(&tp, num_threads);
 
+  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 3);
+  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 4);
+  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 5);
+
   CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 3);
   CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 4);
   CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 5);

From 35d90e89600ff2524ec8bdd4ef4b95dd7c78b656 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 10 Aug 2018 17:37:58 -0700
Subject: [PATCH 554/680] Fix BlockAccess enum in CwiseUnaryOp evaluator

---
 unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h         | 2 +-
 .../Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h       | 5 ++---
 2 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index e30f9ad8e..4ca6b3d8c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -357,7 +357,7 @@ struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
     IsAligned          = TensorEvaluator<ArgType, Device>::IsAligned,
     PacketAccess       = TensorEvaluator<ArgType, Device>::PacketAccess &
                          internal::functor_traits<UnaryOp>::PacketAccess,
-    BlockAccess        = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess        = TensorEvaluator<ArgType, Device>::BlockAccess,
     PreferBlockAccess  = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
     Layout             = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess        = false,  // to be implemented
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index b8a57ab99..b7a0193fe 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -135,9 +135,8 @@ struct IsTileable {
   // Check that block evaluation is supported and it's a preferred option (at
   // least one sub-expression has much faster block evaluation, e.g.
   // broadcasting).
-  static const bool value =
-      TensorEvaluator<Expression, Device>::BlockAccess &
-      TensorEvaluator<Expression, Device>::PreferBlockAccess;
+  static const bool value = TensorEvaluator<Expression, Device>::BlockAccess &&
+                            TensorEvaluator<Expression, Device>::PreferBlockAccess;
 };
 
 template <typename Expression, typename Device,

From e6d5be811ddab928ae7ed73f76e1c4c8e18917e2 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 13 Aug 2018 10:29:21 -0700
Subject: [PATCH 555/680] Fixed syntax of nested templates chevrons to make it
 compatible with c++97 mode.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index cbf91013b..3904552a9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -155,8 +155,8 @@ struct TensorBlockCopyOp {
     typedef const Eigen::Array<Scalar, Dynamic, 1> Src;
     typedef Eigen::Array<Scalar, Dynamic, 1> Dst;
 
-    typedef Eigen::Map<Src, 0, InnerStride<>> SrcMap;
-    typedef Eigen::Map<Dst, 0, InnerStride<>> DstMap;
+    typedef Eigen::Map<Src, 0, InnerStride<> > SrcMap;
+    typedef Eigen::Map<Dst, 0, InnerStride<> > DstMap;
 
     const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
     DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
@@ -405,9 +405,9 @@ struct TensorBlockCwiseBinaryOp {
     typedef const Eigen::Array<RightScalar, Dynamic, 1> Rhs;
     typedef Eigen::Array<OutputScalar, Dynamic, 1> Out;
 
-    typedef Eigen::Map<Lhs, 0, InnerStride<>> LhsMap;
-    typedef Eigen::Map<Rhs, 0, InnerStride<>> RhsMap;
-    typedef Eigen::Map<Out, 0, InnerStride<>> OutMap;
+    typedef Eigen::Map<Lhs, 0, InnerStride<> > LhsMap;
+    typedef Eigen::Map<Rhs, 0, InnerStride<> > RhsMap;
+    typedef Eigen::Map<Out, 0, InnerStride<> > OutMap;
 
     const LeftScalar* lhs_base = &left_data[left_index];
     const RightScalar* rhs_base = &right_data[right_index];

From 3810ec228fbc9ff8fff23a997c09a490f319c902 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 13 Aug 2018 10:46:09 -0700
Subject: [PATCH 556/680] Don't use the auto keyword since it's not always
 supported properly.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 3904552a9..21a6b66e8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -501,7 +501,7 @@ struct TensorBlockCwiseBinaryIO {
       if (size == 1) {
         continue;
       }
-      auto& state = block_iter_state[num_squeezed_dims];
+      BlockIteratorState& state = block_iter_state[num_squeezed_dims];
       state.output_stride = block_strides[dim];
       state.left_stride = left_strides[dim];
       state.right_stride = right_strides[dim];
@@ -523,7 +523,7 @@ struct TensorBlockCwiseBinaryIO {
                                     right_stride, right_data);
       // Update index.
       for (int j = 0; j < num_squeezed_dims; ++j) {
-        auto& state = block_iter_state[j];
+        BlockIteratorState& state = block_iter_state[j];
         if (++state.count < state.size) {
           output_index += state.output_stride;
           left_index += state.left_stride;

From 26239ee580e5ffbdcad657c291bf4f49e6b297cf Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 13 Aug 2018 11:05:51 -0700
Subject: [PATCH 557/680] Use NULL instead of nullptr to avoid adding a cxx11
 requirement.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h           | 2 +-
 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h | 4 ++--
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h         | 4 ++--
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
index f1f877c16..bcaf5c97f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h
@@ -187,7 +187,7 @@ struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalBlock(TensorBlock* block) {
     if (TensorEvaluator<LeftArgType, Device>::RawAccess &&
-        m_leftImpl.data() != nullptr) {
+        m_leftImpl.data() != NULL) {
       TensorBlock left_block(block->first_coeff_index(), block->block_sizes(),
                              block->tensor_strides(), block->tensor_strides(),
                              m_leftImpl.data() + block->first_coeff_index());
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index cc134228a..3e3665efb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -102,7 +102,7 @@ class Allocator {
 // Build a thread pool device on top the an existing pool of threads.
 struct ThreadPoolDevice {
   // The ownership of the thread pool remains with the caller.
-  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores, Allocator* allocator = nullptr)
+  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores, Allocator* allocator = NULL)
       : pool_(pool), num_threads_(num_cores), allocator_(allocator) { }
 
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
@@ -282,7 +282,7 @@ struct ThreadPoolDevice {
   // Convenience wrapper for parallelFor that does not align blocks.
   void parallelFor(Index n, const TensorOpCost& cost,
                    std::function<void(Index, Index)> f) const {
-    parallelFor(n, cost, nullptr, std::move(f));
+    parallelFor(n, cost, NULL, std::move(f));
   }
 
   // Thread pool accessor.
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 3f3b5685d..0294aa62e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -227,7 +227,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
     typedef EvalRange<Evaluator, StorageIndex, Vectorizable> EvalRange;
 
     Evaluator evaluator(expr, device);
-    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
       const StorageIndex PacketSize =
           Vectorizable
@@ -271,7 +271,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
       return;
     }
 
-    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
       TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
       Index block_total_size = 0;

From d6e283ba9650faed0c48c36f475d5dd77b916427 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 13 Aug 2018 11:14:50 -0700
Subject: [PATCH 558/680] sigmoid -> logistic

---
 Eigen/src/Core/GlobalFunctions.h        |  2 +-
 Eigen/src/Core/functors/UnaryFunctors.h | 10 +++++-----
 Eigen/src/plugins/ArrayCwiseUnaryOps.h  | 10 +++++-----
 test/array.cpp                          |  8 ++++----
 4 files changed, 15 insertions(+), 15 deletions(-)

diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h
index 4c9823d70..563df6e84 100644
--- a/Eigen/src/Core/GlobalFunctions.h
+++ b/Eigen/src/Core/GlobalFunctions.h
@@ -66,7 +66,7 @@ namespace Eigen
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op,hyperbolic sine,\sa ArrayBase::sinh)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op,hyperbolic cosine,\sa ArrayBase::cosh)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op,hyperbolic tangent,\sa ArrayBase::tanh)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sigmoid,scalar_sigmoid_op,sigmoid function,\sa ArrayBase::sigmoid)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(logistic,scalar_logistic_op,logistic function,\sa ArrayBase::logistic)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op,natural logarithm of the gamma function,\sa ArrayBase::lgamma)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(digamma,scalar_digamma_op,derivative of lgamma,\sa ArrayBase::digamma)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op,error function,\sa ArrayBase::erf)
diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index c350cdf98..c1cc2ab3b 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -824,12 +824,12 @@ struct functor_traits<scalar_sign_op<Scalar> >
 };
 
 /** \internal
-  * \brief Template functor to compute the sigmoid of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::sigmoid()
+  * \brief Template functor to compute the logistic function of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::logistic()
   */
 template <typename T>
-struct scalar_sigmoid_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sigmoid_op)
+struct scalar_logistic_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_logistic_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& x) const {
     const T one = T(1);
     return one / (one + numext::exp(-x));
@@ -842,7 +842,7 @@ struct scalar_sigmoid_op {
   }
 };
 template <typename T>
-struct functor_traits<scalar_sigmoid_op<T> > {
+struct functor_traits<scalar_logistic_op<T> > {
   enum {
     Cost = NumTraits<T>::AddCost * 2 + NumTraits<T>::MulCost * 6,
     PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv &&
diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h
index c9c7e7e4f..e928db467 100644
--- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h
+++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h
@@ -21,7 +21,7 @@ typedef CwiseUnaryOp<internal::scalar_acos_op<Scalar>, const Derived> AcosReturn
 typedef CwiseUnaryOp<internal::scalar_asin_op<Scalar>, const Derived> AsinReturnType;
 typedef CwiseUnaryOp<internal::scalar_atan_op<Scalar>, const Derived> AtanReturnType;
 typedef CwiseUnaryOp<internal::scalar_tanh_op<Scalar>, const Derived> TanhReturnType;
-typedef CwiseUnaryOp<internal::scalar_sigmoid_op<Scalar>, const Derived> SigmoidReturnType;
+typedef CwiseUnaryOp<internal::scalar_logistic_op<Scalar>, const Derived> LogisticReturnType;
 typedef CwiseUnaryOp<internal::scalar_sinh_op<Scalar>, const Derived> SinhReturnType;
 typedef CwiseUnaryOp<internal::scalar_cosh_op<Scalar>, const Derived> CoshReturnType;
 typedef CwiseUnaryOp<internal::scalar_square_op<Scalar>, const Derived> SquareReturnType;
@@ -336,13 +336,13 @@ cosh() const
   return CoshReturnType(derived());
 }
 
-/** \returns an expression of the coefficient-wise sigmoid of *this.
+/** \returns an expression of the coefficient-wise logistic of *this.
   */
 EIGEN_DEVICE_FUNC
-inline const SigmoidReturnType
-sigmoid() const
+inline const LogisticReturnType
+logistic() const
 {
-  return SigmoidReturnType(derived());
+  return LogisticReturnType(derived());
 }
 
 /** \returns an expression of the coefficient-wise inverse of *this.
diff --git a/test/array.cpp b/test/array.cpp
index 1b5725fb4..d9c4626c0 100644
--- a/test/array.cpp
+++ b/test/array.cpp
@@ -231,7 +231,7 @@ template<typename ArrayType> void array_real(const ArrayType& m)
   VERIFY_IS_APPROX(m1.sinh(), sinh(m1));
   VERIFY_IS_APPROX(m1.cosh(), cosh(m1));
   VERIFY_IS_APPROX(m1.tanh(), tanh(m1));
-  VERIFY_IS_APPROX(m1.sigmoid(), sigmoid(m1));
+  VERIFY_IS_APPROX(m1.logistic(), logistic(m1));
 
   VERIFY_IS_APPROX(m1.arg(), arg(m1));
   VERIFY_IS_APPROX(m1.round(), round(m1));
@@ -267,7 +267,7 @@ template<typename ArrayType> void array_real(const ArrayType& m)
   VERIFY_IS_APPROX(sinh(m1), 0.5*(exp(m1)-exp(-m1)));
   VERIFY_IS_APPROX(cosh(m1), 0.5*(exp(m1)+exp(-m1)));
   VERIFY_IS_APPROX(tanh(m1), (0.5*(exp(m1)-exp(-m1)))/(0.5*(exp(m1)+exp(-m1))));
-  VERIFY_IS_APPROX(sigmoid(m1), (1.0/(1.0+exp(-m1))));
+  VERIFY_IS_APPROX(logistic(m1), (1.0/(1.0+exp(-m1))));
   VERIFY_IS_APPROX(arg(m1), ((m1<0).template cast<Scalar>())*std::acos(-1.0));
   VERIFY((round(m1) <= ceil(m1) && round(m1) >= floor(m1)).all());
   VERIFY((Eigen::isnan)((m1*0.0)/0.0).all());
@@ -347,7 +347,7 @@ template<typename ArrayType> void array_complex(const ArrayType& m)
   VERIFY_IS_APPROX(m1.sinh(), sinh(m1));
   VERIFY_IS_APPROX(m1.cosh(), cosh(m1));
   VERIFY_IS_APPROX(m1.tanh(), tanh(m1));
-  VERIFY_IS_APPROX(m1.sigmoid(), sigmoid(m1));
+  VERIFY_IS_APPROX(m1.logistic(), logistic(m1));
   VERIFY_IS_APPROX(m1.arg(), arg(m1));
   VERIFY((m1.isNaN() == (Eigen::isnan)(m1)).all());
   VERIFY((m1.isInf() == (Eigen::isinf)(m1)).all());
@@ -371,7 +371,7 @@ template<typename ArrayType> void array_complex(const ArrayType& m)
   VERIFY_IS_APPROX(sinh(m1), 0.5*(exp(m1)-exp(-m1)));
   VERIFY_IS_APPROX(cosh(m1), 0.5*(exp(m1)+exp(-m1)));
   VERIFY_IS_APPROX(tanh(m1), (0.5*(exp(m1)-exp(-m1)))/(0.5*(exp(m1)+exp(-m1))));
-  VERIFY_IS_APPROX(sigmoid(m1), (1.0/(1.0 + exp(-m1))));
+  VERIFY_IS_APPROX(logistic(m1), (1.0/(1.0 + exp(-m1))));
 
   for (Index i = 0; i < m.rows(); ++i)
     for (Index j = 0; j < m.cols(); ++j)

From 0f1b2e08a5d5ed9158435d12da60dfaa5e8b2803 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 13 Aug 2018 11:52:58 -0700
Subject: [PATCH 559/680] Call logistic functor from Tensor::sigmoid.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 97f90f638..7cc71d99e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -200,9 +200,9 @@ class TensorBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_sigmoid_op<Scalar>, const Derived>
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_logistic_op<Scalar>, const Derived>
     sigmoid() const {
-      return unaryExpr(internal::scalar_sigmoid_op<Scalar>());
+      return unaryExpr(internal::scalar_logistic_op<Scalar>());
     }
 
     EIGEN_DEVICE_FUNC

From 3d3711f22fd2b04026f04ce6f1fe7e888ea5a4da Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 13 Aug 2018 15:16:06 -0700
Subject: [PATCH 560/680] Fixed compilation errors.

---
 unsupported/test/cxx11_tensor_thread_pool.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index 5c3aae482..dd163c18a 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -428,7 +428,7 @@ void test_threadpool_allocate(TestAllocator* allocator)
     void* ptr = device.allocate(512);
     device.deallocate(ptr);
   }
-  VERIFY(allocator != nullptr);
+  VERIFY(allocator != NULL);
   VERIFY_IS_EQUAL(allocator->alloc_count(), num_allocs);
   VERIFY_IS_EQUAL(allocator->dealloc_count(), num_allocs);
 }
@@ -460,7 +460,7 @@ EIGEN_DECLARE_TEST(cxx11_tensor_thread_pool)
   CALL_SUBTEST_6(test_multithread_random());
 
   TestAllocator test_allocator;
-  CALL_SUBTEST_6(test_multithread_shuffle<ColMajor>(nullptr));
+  CALL_SUBTEST_6(test_multithread_shuffle<ColMajor>(NULL));
   CALL_SUBTEST_6(test_multithread_shuffle<RowMajor>(&test_allocator));
   CALL_SUBTEST_6(test_threadpool_allocate(&test_allocator));
 }

From 501be70b27e855d4921c5b510ec1a19783238c16 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 13 Aug 2018 15:16:40 -0700
Subject: [PATCH 561/680] Code cleanup

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 0294aa62e..0cefe42dd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -257,7 +257,6 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
 
   static EIGEN_STRONG_INLINE void run(const Expression& expr,
                          const ThreadPoolDevice& device) {
-    typedef TensorBlock<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlock;
     typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlockMapper;
 
     Evaluator evaluator(expr, device);

From 8278ae63137fb2ce2cdf6fc8117df3080e5cb2fe Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 13 Aug 2018 15:31:23 -0700
Subject: [PATCH 562/680] Add support for thread local support on platforms
 that do not support it through emulation using a hash map.

---
 test/main.h                                   |  6 +-
 unsupported/Eigen/CXX11/ThreadPool            | 10 ++-
 .../CXX11/src/Tensor/TensorDeviceThreadPool.h | 50 --------------
 .../src/ThreadPool/NonBlockingThreadPool.h    | 65 ++++++++++++++-----
 .../Eigen/CXX11/src/ThreadPool/ThreadLocal.h  | 48 ++++++++++++--
 5 files changed, 103 insertions(+), 76 deletions(-)

diff --git a/test/main.h b/test/main.h
index de8a4865f..36784b1f4 100644
--- a/test/main.h
+++ b/test/main.h
@@ -125,7 +125,7 @@ inline void on_temporary_creation(long int size) {
     if(nb_temporaries!=(N)) { std::cerr << "nb_temporaries == " << nb_temporaries << "\n"; }\
     VERIFY( (#XPR) && nb_temporaries==(N) ); \
   }
-  
+
 #endif
 
 #include "split_test_helper.h"
@@ -328,7 +328,7 @@ namespace Eigen
   #define VERIFY_RAISES_STATIC_ASSERT(a) \
     std::cout << "Can't VERIFY_RAISES_STATIC_ASSERT( " #a " ) with exceptions disabled\n";
 #endif
-    
+
   #if !defined(__CUDACC__) && !defined(__HIPCC__) && !defined(__SYCL_DEVICE_ONLY__)
   #define EIGEN_USE_CUSTOM_ASSERT
   #endif
@@ -845,4 +845,4 @@ int main(int argc, char *argv[])
 #ifdef _MSC_VER
   // 4503 - decorated name length exceeded, name was truncated
   #pragma warning( disable : 4503)
-#endif
\ No newline at end of file
+#endif
diff --git a/unsupported/Eigen/CXX11/ThreadPool b/unsupported/Eigen/CXX11/ThreadPool
index cbb3bbf2c..12aa07c7f 100644
--- a/unsupported/Eigen/CXX11/ThreadPool
+++ b/unsupported/Eigen/CXX11/ThreadPool
@@ -44,6 +44,14 @@
 #include <thread>
 #include <functional>
 #include <memory>
+#ifndef EIGEN_THREAD_LOCAL
+// There are non-parenthesized calls to "max" in the  <unordered_map> header,
+// which trigger a check in test/main.h causing compilation to fail.
+// We work around the check here by removing the check for max in
+// the case where we have to emulate thread_local.
+#undef max
+#include <unordered_map>
+#endif
 
 #include "src/util/CXX11Meta.h"
 #include "src/util/MaxSizeVector.h"
@@ -55,6 +63,7 @@
 #include "src/ThreadPool/RunQueue.h"
 #include "src/ThreadPool/ThreadPoolInterface.h"
 #include "src/ThreadPool/ThreadEnvironment.h"
+#include "src/ThreadPool/Barrier.h"
 #include "src/ThreadPool/NonBlockingThreadPool.h"
 
 #endif
@@ -62,4 +71,3 @@
 #include <Eigen/src/Core/util/ReenableStupidWarnings.h>
 
 #endif // EIGEN_CXX11_THREADPOOL_MODULE
-
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
index 3e3665efb..6fc6688d3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -12,56 +12,6 @@
 
 namespace Eigen {
 
-// Barrier is an object that allows one or more threads to wait until
-// Notify has been called a specified number of times.
-class Barrier {
- public:
-  Barrier(unsigned int count) : state_(count << 1), notified_(false) {
-    eigen_assert(((count << 1) >> 1) == count);
-  }
-  ~Barrier() {
-    eigen_assert((state_>>1) == 0);
-  }
-
-  void Notify() {
-    unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2;
-    if (v != 1) {
-      eigen_assert(((v + 2) & ~1) != 0);
-      return;  // either count has not dropped to 0, or waiter is not waiting
-    }
-    std::unique_lock<std::mutex> l(mu_);
-    eigen_assert(!notified_);
-    notified_ = true;
-    cv_.notify_all();
-  }
-
-  void Wait() {
-    unsigned int v = state_.fetch_or(1, std::memory_order_acq_rel);
-    if ((v >> 1) == 0) return;
-    std::unique_lock<std::mutex> l(mu_);
-    while (!notified_) {
-      cv_.wait(l);
-    }
-  }
-
- private:
-  std::mutex mu_;
-  std::condition_variable cv_;
-  std::atomic<unsigned int> state_;  // low bit is waiter flag
-  bool notified_;
-};
-
-
-// Notification is an object that allows a user to to wait for another
-// thread to signal a notification that an event has occurred.
-//
-// Multiple threads can wait on the same Notification object,
-// but only one caller must call Notify() on the object.
-struct Notification : Barrier {
-  Notification() : Barrier(1) {};
-};
-
-
 // Runs an arbitrary function and then calls Notify() on the passed in
 // Notification.
 template <typename Function, typename... Args> struct FunctionWrapperWithNotification
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index ecd49f382..ede70da8d 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -10,7 +10,6 @@
 #ifndef EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
 #define EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
 
-
 namespace Eigen {
 
 template <typename Environment>
@@ -23,7 +22,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       : ThreadPoolTempl(num_threads, true, env) {}
 
   ThreadPoolTempl(int num_threads, bool allow_spinning,
-                             Environment env = Environment())
+                  Environment env = Environment())
       : env_(env),
         num_threads_(num_threads),
         allow_spinning_(allow_spinning),
@@ -61,9 +60,17 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     for (int i = 0; i < num_threads_; i++) {
       queues_.push_back(new Queue());
     }
+#ifndef EIGEN_THREAD_LOCAL
+    init_barrier_.reset(new Barrier(num_threads_));
+#endif
     for (int i = 0; i < num_threads_; i++) {
       threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
     }
+#ifndef EIGEN_THREAD_LOCAL
+    // Wait for workers to initialize per_thread_map_. Otherwise we might race
+    // with them in Schedule or CurrentThreadId.
+    init_barrier_->Wait();
+#endif
   }
 
   ~ThreadPoolTempl() {
@@ -85,6 +92,9 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // Join threads explicitly to avoid destruction order issues.
     for (size_t i = 0; i < num_threads_; i++) delete threads_[i];
     for (size_t i = 0; i < num_threads_; i++) delete queues_[i];
+#ifndef EIGEN_THREAD_LOCAL
+    for (auto it : per_thread_map_) delete it.second;
+#endif
   }
 
   void Schedule(std::function<void()> fn) {
@@ -109,8 +119,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // this is kept alive while any threads can potentially be in Schedule.
     if (!t.f) {
       ec_.Notify(false);
-    }
-    else {
+    } else {
       env_.ExecuteTask(t);  // Push failed, execute directly.
     }
   }
@@ -130,13 +139,10 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     ec_.Notify(true);
   }
 
-  int NumThreads() const final {
-    return num_threads_;
-  }
+  int NumThreads() const final { return num_threads_; }
 
   int CurrentThreadId() const final {
-    const PerThread* pt =
-        const_cast<ThreadPoolTempl*>(this)->GetPerThread();
+    const PerThread* pt = const_cast<ThreadPoolTempl*>(this)->GetPerThread();
     if (pt->pool == this) {
       return pt->thread_id;
     } else {
@@ -148,10 +154,10 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   typedef typename Environment::EnvThread Thread;
 
   struct PerThread {
-    constexpr PerThread() : pool(NULL), rand(0), thread_id(-1) { }
+    constexpr PerThread() : pool(NULL), rand(0), thread_id(-1) {}
     ThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
-    uint64_t rand;  // Random generator state.
-    int thread_id;  // Worker thread index in pool.
+    uint64_t rand;          // Random generator state.
+    int thread_id;          // Worker thread index in pool.
   };
 
   Environment env_;
@@ -166,12 +172,26 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   std::atomic<bool> done_;
   std::atomic<bool> cancelled_;
   EventCount ec_;
+#ifndef EIGEN_THREAD_LOCAL
+  std::unique_ptr<Barrier> init_barrier_;
+  std::mutex mu;  // Protects per_thread_map_.
+  std::unordered_map<uint64_t, PerThread*> per_thread_map_;
+#endif
 
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
+#ifndef EIGEN_THREAD_LOCAL
+    PerThread* pt = new PerThread();
+    mu.lock();
+    per_thread_map_[GlobalThreadIdHash()] = pt;
+    mu.unlock();
+    init_barrier_->Notify();
+    init_barrier_->Wait();
+#else
     PerThread* pt = GetPerThread();
+#endif
     pt->pool = this;
-    pt->rand = std::hash<std::thread::id>()(std::this_thread::get_id());
+    pt->rand = GlobalThreadIdHash();
     pt->thread_id = thread_id;
     Queue* q = queues_[thread_id];
     EventCount::Waiter* waiter = &waiters_[thread_id];
@@ -322,10 +342,24 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     return -1;
   }
 
-  static EIGEN_STRONG_INLINE PerThread* GetPerThread() {
+  static EIGEN_STRONG_INLINE uint64_t GlobalThreadIdHash() {
+    return std::hash<std::thread::id>()(std::this_thread::get_id());
+  }
+
+  EIGEN_STRONG_INLINE PerThread* GetPerThread() {
+#ifndef EIGEN_THREAD_LOCAL
+    static PerThread dummy;
+    auto it = per_thread_map_.find(GlobalThreadIdHash());
+    if (it == per_thread_map_.end()) {
+      return &dummy;
+    } else {
+      return it->second;
+    }
+#else
     EIGEN_THREAD_LOCAL PerThread per_thread_;
     PerThread* pt = &per_thread_;
     return pt;
+#endif
   }
 
   static EIGEN_STRONG_INLINE unsigned Rand(uint64_t* state) {
@@ -333,7 +367,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // Update the internal state
     *state = current * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
     // Generate the random output (using the PCG-XSH-RS scheme)
-    return static_cast<unsigned>((current ^ (current >> 22)) >> (22 + (current >> 61)));
+    return static_cast<unsigned>((current ^ (current >> 22)) >>
+                                 (22 + (current >> 61)));
   }
 };
 
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
index cfa221732..f33759ba9 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
@@ -10,13 +10,47 @@
 #ifndef EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
 #define EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
 
-// Try to come up with a portable implementation of thread local variables
-#if EIGEN_COMP_GNUC && EIGEN_GNUC_AT_MOST(4, 7)
-#define EIGEN_THREAD_LOCAL static __thread
-#elif EIGEN_COMP_CLANG
-#define EIGEN_THREAD_LOCAL static __thread
-#else
-#define EIGEN_THREAD_LOCAL static thread_local
+#undef EIGEN_THREAD_LOCAL
+
+#if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_thread_local))
+  #define EIGEN_THREAD_LOCAL static thread_local
+#elif (EIGEN_COMP_GNUC && EIGEN_GNUC_AT_MOST(4, 7)) || EIGEN_COMP_CLANG
+  #define EIGEN_THREAD_LOCAL static __thread
 #endif
 
+// Disable TLS for Apple and Android builds with older toolchains.
+#if defined(__APPLE__)
+// Included for TARGET_OS_IPHONE, __IPHONE_OS_VERSION_MIN_REQUIRED,
+// __IPHONE_8_0.
+#include <Availability.h>
+#include <TargetConditionals.h>
+#endif
+// Checks whether C++11's `thread_local` storage duration specifier is
+// supported.
+#if defined(__apple_build_version__) &&     \
+    ((__apple_build_version__ < 8000042) || \
+     (TARGET_OS_IPHONE && __IPHONE_OS_VERSION_MIN_REQUIRED < __IPHONE_9_0))
+// Notes: Xcode's clang did not support `thread_local` until version
+// 8, and even then not for all iOS < 9.0.
+#undef EIGEN_THREAD_LOCAL
+
+#elif defined(__ANDROID__) && EIGEN_COMP_CLANG
+// There are platforms for which TLS should not be used even though the compiler
+// makes it seem like it's supported (Android NDK < r12b for example).
+// This is primarily because of linker problems and toolchain misconfiguration:
+// TLS isn't supported until NDK r12b per
+// https://developer.android.com/ndk/downloads/revision_history.html
+// Since NDK r16, `__NDK_MAJOR__` and `__NDK_MINOR__` are defined in
+// <android/ndk-version.h>. For NDK < r16, users should define these macros,
+// e.g. `-D__NDK_MAJOR__=11 -D__NKD_MINOR__=0` for NDK r11.
+#if __has_include(<android/ndk-version.h>)
+#include <android/ndk-version.h>
+#endif  // __has_include(<android/ndk-version.h>)
+#if defined(__ANDROID__) && defined(__clang__) && defined(__NDK_MAJOR__) && \
+    defined(__NDK_MINOR__) &&                                               \
+    ((__NDK_MAJOR__ < 12) || ((__NDK_MAJOR__ == 12) && (__NDK_MINOR__ < 1)))
+#undef EIGEN_THREAD_LOCAL
+#endif
+#endif  // defined(__ANDROID__) && defined(__clang__)
+
 #endif  // EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H

From 9bb75d8d31571f5513107080c8d3c85e27ff8430 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Mon, 13 Aug 2018 15:34:03 -0700
Subject: [PATCH 563/680] Add Barrier.h.

---
 .../Eigen/CXX11/src/ThreadPool/Barrier.h      | 67 +++++++++++++++++++
 1 file changed, 67 insertions(+)
 create mode 100644 unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h b/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h
new file mode 100644
index 000000000..c37fc1e65
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h
@@ -0,0 +1,67 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// Barrier is an object that allows one or more threads to wait until
+// Notify has been called a specified number of times.
+
+#ifndef EIGEN_CXX11_THREADPOOL_BARRIER_H
+#define EIGEN_CXX11_THREADPOOL_BARRIER_H
+
+namespace Eigen {
+
+class Barrier {
+ public:
+  Barrier(unsigned int count) : state_(count << 1), notified_(false) {
+    eigen_assert(((count << 1) >> 1) == count);
+  }
+  ~Barrier() {
+    eigen_assert((state_>>1) == 0);
+  }
+
+  void Notify() {
+    unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2;
+    if (v != 1) {
+      eigen_assert(((v + 2) & ~1) != 0);
+      return;  // either count has not dropped to 0, or waiter is not waiting
+    }
+    std::unique_lock<std::mutex> l(mu_);
+    eigen_assert(!notified_);
+    notified_ = true;
+    cv_.notify_all();
+  }
+
+  void Wait() {
+    unsigned int v = state_.fetch_or(1, std::memory_order_acq_rel);
+    if ((v >> 1) == 0) return;
+    std::unique_lock<std::mutex> l(mu_);
+    while (!notified_) {
+      cv_.wait(l);
+    }
+  }
+
+ private:
+  std::mutex mu_;
+  std::condition_variable cv_;
+  std::atomic<unsigned int> state_;  // low bit is waiter flag
+  bool notified_;
+};
+
+
+// Notification is an object that allows a user to to wait for another
+// thread to signal a notification that an event has occurred.
+//
+// Multiple threads can wait on the same Notification object,
+// but only one caller must call Notify() on the object.
+struct Notification : Barrier {
+  Notification() : Barrier(1) {};
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_BARRIER_H

From 59bba77ead210f71b61ee6c551207c6f062bc123 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 14 Aug 2018 10:54:48 -0700
Subject: [PATCH 564/680] Fixed compilation errors with gcc 4.7 and 4.8

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 26 +++++++++----------
 .../CXX11/src/Tensor/TensorContraction.h      |  2 +-
 2 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 21a6b66e8..322260011 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -73,7 +73,7 @@ struct TensorOpResourceRequirements {
   // expression tree (like reductions) to communicate resources
   // requirements based on local state (like the total number of reductions
   // to be computed).
-  TensorOpResourceRequirements(internal::TensorBlockShapeType shape,
+  TensorOpResourceRequirements(TensorBlockShapeType shape,
                                const Index size)
       : block_shape(shape), block_total_size(size) {}
 };
@@ -90,9 +90,9 @@ EIGEN_STRONG_INLINE void MergeResourceRequirements(
   *block_shape = resources[0].block_shape;
   *block_total_size = resources[0].block_total_size;
   for (std::vector<TensorOpResourceRequirements>::size_type i = 1; i < resources.size(); ++i) {
-    if (resources[i].block_shape == TensorBlockShapeType::kSkewedInnerDims &&
-        *block_shape != TensorBlockShapeType::kSkewedInnerDims) {
-      *block_shape = TensorBlockShapeType::kSkewedInnerDims;
+    if (resources[i].block_shape == kSkewedInnerDims &&
+        *block_shape ! kSkewedInnerDims) {
+      *block_shape = kSkewedInnerDims;
     }
     *block_total_size =
         numext::maxi(*block_total_size, resources[i].block_total_size);
@@ -178,9 +178,9 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout,
           bool BlockRead>
 class TensorBlockIO {
  public:
-  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
+  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
-  typedef typename internal::TensorBlockCopyOp<Scalar, StorageIndex>
+  typedef typename TensorBlockCopyOp<Scalar, StorageIndex>
       TensorBlockCopyOp;
 
  protected:
@@ -320,7 +320,7 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
                                                Layout, /*BlockRead=*/true> {
  public:
-  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
+  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
   typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true>
       Base;
@@ -357,7 +357,7 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
                                                Layout, /*BlockRead=*/false> {
  public:
-  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
+  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
   typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false>
       Base;
@@ -434,7 +434,7 @@ struct TensorBlockCwiseBinaryOp {
 template <typename BinaryFunctor, typename StorageIndex, typename OutputScalar,
           int NumDims, int Layout>
 struct TensorBlockCwiseBinaryIO {
-  typedef typename internal::TensorBlock<OutputScalar, StorageIndex, NumDims,
+  typedef typename TensorBlock<OutputScalar, StorageIndex, NumDims,
                                          Layout>::Dimensions Dimensions;
 
   struct BlockIteratorState {
@@ -627,7 +627,7 @@ struct TensorBlockView {
 template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockMapper {
  public:
-  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
+  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
   typedef DSizes<StorageIndex, NumDims> Dimensions;
 
@@ -742,7 +742,7 @@ class TensorBlockMapper {
         block_dim_sizes[i] = 1;
       }
     } else if (block_dim_sizes.TotalSize() > min_target_size) {
-      if (block_shape == TensorBlockShapeType::kUniformAllDims) {
+      if (block_shape == kUniformAllDims) {
         // Tensor will not fit within 'min_target_size' budget: calculate tensor
         // block dimension sizes based on "square" dimension size target.
         const size_t dim_size_target = static_cast<const size_t>(
@@ -773,7 +773,7 @@ class TensorBlockMapper {
             total_size = total_size_other_dims * block_dim_sizes[dim];
           }
         }
-      } else if (block_shape == TensorBlockShapeType::kSkewedInnerDims) {
+      } else if (block_shape == kSkewedInnerDims) {
         StorageIndex coeff_to_allocate = min_target_size;
         for (int i = 0; i < NumDims; ++i) {
           const int dim = cond<Layout>()(i, NumDims - i - 1);
@@ -818,7 +818,7 @@ class TensorBlockMapper {
 template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorSliceBlockMapper {
  public:
-  typedef typename internal::TensorBlock<Scalar, StorageIndex, NumDims, Layout>
+  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
       TensorBlock;
   typedef DSizes<StorageIndex, NumDims> Dimensions;
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index e604456e8..a023718c6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -155,7 +155,7 @@ struct TensorContractionParams {
 // See expected implementation in NoOpOutputKernel.
 struct OutputKernel {
   template <typename Index, typename Scalar>
-  using OutputMapper = internal::blas_data_mapper<Scalar, Index, ColMajor>;
+  typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
 };
 
 // Output kernel that does absolutely nothing.

From aebdb0642402e49ded58db98dd29c67cd76d204a Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Tue, 14 Aug 2018 12:06:39 -0700
Subject: [PATCH 565/680] Fix a few compiler warnings in CXX11 tests.

---
 unsupported/test/cxx11_tensor_contraction.cpp |  4 +--
 unsupported/test/cxx11_tensor_convolution.cpp |  2 +-
 unsupported/test/cxx11_tensor_index_list.cpp  |  1 -
 unsupported/test/kronecker_product.cpp        | 27 ++++++++++---------
 4 files changed, 17 insertions(+), 17 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp
index d4cfbd0da..2e918eb30 100644
--- a/unsupported/test/cxx11_tensor_contraction.cpp
+++ b/unsupported/test/cxx11_tensor_contraction.cpp
@@ -471,7 +471,7 @@ static void test_tensor_product()
   mat1.setRandom();
   mat2.setRandom();
 
-  Tensor<float, 4, DataLayout> result = mat1.contract(mat2, Eigen::array<DimPair, 0>{{}});
+  Tensor<float, 4, DataLayout> result = mat1.contract(mat2, Eigen::array<DimPair, 0>{});
 
   VERIFY_IS_EQUAL(result.dimension(0), 2);
   VERIFY_IS_EQUAL(result.dimension(1), 3);
@@ -553,7 +553,7 @@ static void test_large_contraction_with_output_kernel() {
 
   m_result = m_left * m_right;
 
-  for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
+  for (std::ptrdiff_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
     VERIFY(&t_result.data()[i] != &m_result.data()[i]);
     VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
   }
diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp
index 01bc77bc1..9fe980648 100644
--- a/unsupported/test/cxx11_tensor_convolution.cpp
+++ b/unsupported/test/cxx11_tensor_convolution.cpp
@@ -25,7 +25,7 @@ static void test_evals()
 
   Tensor<float, 2, DataLayout> result(2,3);
   result.setZero();
-  Eigen::array<Tensor<float, 2>::Index, 1> dims3{{0}};
+  Eigen::array<Tensor<float, 2>::Index, 1> dims3{0};
 
   typedef TensorEvaluator<decltype(input.convolve(kernel, dims3)), DefaultDevice> Evaluator;
   Evaluator eval(input.convolve(kernel, dims3), DefaultDevice());
diff --git a/unsupported/test/cxx11_tensor_index_list.cpp b/unsupported/test/cxx11_tensor_index_list.cpp
index e81fa5e40..294677a4d 100644
--- a/unsupported/test/cxx11_tensor_index_list.cpp
+++ b/unsupported/test/cxx11_tensor_index_list.cpp
@@ -170,7 +170,6 @@ static void test_type2indexpair_list()
   typedef Eigen::IndexPairList<Eigen::type2indexpair<0,10>, Eigen::IndexPair<DenseIndex>, Eigen::type2indexpair<2,12>> Dims2_b;
   typedef Eigen::IndexPairList<Eigen::IndexPair<DenseIndex>, Eigen::type2indexpair<1,11>, Eigen::IndexPair<DenseIndex>> Dims2_c;
 
-  Dims0 d0;
   Dims2_a d2_a;
 
   Dims2_b d2_b;
diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp
index 4f143b6de..1a936ed25 100644
--- a/unsupported/test/kronecker_product.cpp
+++ b/unsupported/test/kronecker_product.cpp
@@ -9,6 +9,9 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
+
+#include "main.h"
+
 #ifdef EIGEN_TEST_PART_1
 
 #include "sparse.h"
@@ -95,7 +98,7 @@ EIGEN_DECLARE_TEST(kronecker_product)
   SM_a.insert(1,0) = DM_a.coeffRef(1,0) = -0.9076572187376921;
   SM_a.insert(1,1) = DM_a.coeffRef(1,1) =  0.6469156566545853;
   SM_a.insert(1,2) = DM_a.coeffRef(1,2) = -0.3658010398782789;
- 
+
   MatrixXd             DM_b(3,2);
   SparseMatrix<double> SM_b(3,2);
   SM_b.insert(0,0) = DM_b.coeffRef(0,0) =  0.9004440976767099;
@@ -165,7 +168,7 @@ EIGEN_DECLARE_TEST(kronecker_product)
   SM_a.insert(0,3) = -0.2;
   SM_a.insert(2,4) =  0.3;
   SM_a.finalize();
-  
+
   SM_b.insert(0,0) =  0.4;
   SM_b.insert(2,1) = -0.5;
   SM_b.finalize();
@@ -183,7 +186,7 @@ EIGEN_DECLARE_TEST(kronecker_product)
   DM_b2.resize(4,8);
   DM_ab2 = kroneckerProduct(DM_a2,DM_b2);
   CALL_SUBTEST(check_dimension(DM_ab2,10*4,9*8));
-  
+
   for(int i = 0; i < g_repeat; i++)
   {
     double density = Eigen::internal::random<double>(0.01,0.5);
@@ -196,35 +199,35 @@ EIGEN_DECLARE_TEST(kronecker_product)
     MatrixXf dA(ra,ca), dB(rb,cb), dC;
     initSparse(density, dA, sA);
     initSparse(density, dB, sB);
-    
+
     sC = kroneckerProduct(sA,sB);
     dC = kroneckerProduct(dA,dB);
     VERIFY_IS_APPROX(MatrixXf(sC),dC);
-    
+
     sC = kroneckerProduct(sA.transpose(),sB);
     dC = kroneckerProduct(dA.transpose(),dB);
     VERIFY_IS_APPROX(MatrixXf(sC),dC);
-    
+
     sC = kroneckerProduct(sA.transpose(),sB.transpose());
     dC = kroneckerProduct(dA.transpose(),dB.transpose());
     VERIFY_IS_APPROX(MatrixXf(sC),dC);
-    
+
     sC = kroneckerProduct(sA,sB.transpose());
     dC = kroneckerProduct(dA,dB.transpose());
     VERIFY_IS_APPROX(MatrixXf(sC),dC);
-    
+
     sC2 = kroneckerProduct(sA,sB);
     dC = kroneckerProduct(dA,dB);
     VERIFY_IS_APPROX(MatrixXf(sC2),dC);
-    
+
     sC2 = kroneckerProduct(dA,sB);
     dC = kroneckerProduct(dA,dB);
     VERIFY_IS_APPROX(MatrixXf(sC2),dC);
-    
+
     sC2 = kroneckerProduct(sA,dB);
     dC = kroneckerProduct(dA,dB);
     VERIFY_IS_APPROX(MatrixXf(sC2),dC);
-    
+
     sC2 = kroneckerProduct(2*sA,sB);
     dC = kroneckerProduct(2*dA,dB);
     VERIFY_IS_APPROX(MatrixXf(sC2),dC);
@@ -236,8 +239,6 @@ EIGEN_DECLARE_TEST(kronecker_product)
 #ifdef EIGEN_TEST_PART_2
 
 // simply check that for a dense kronecker product, sparse module is not needed
-
-#include "main.h"
 #include <Eigen/KroneckerProduct>
 
 EIGEN_DECLARE_TEST(kronecker_product)

From 15d4f515e2d4982bd16f0a85a7bbb5343270deec Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Tue, 14 Aug 2018 12:17:46 -0700
Subject: [PATCH 566/680] Use plain_assert in destructors to avoid throwing in
 CXX11 tests where main.h owerwrites eigen_assert with a throwing version.

---
 unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h          | 7 ++-----
 unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h       | 5 +++--
 .../Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h    | 6 +++---
 unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h         | 8 +++-----
 4 files changed, 11 insertions(+), 15 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h b/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h
index c37fc1e65..ef5e9ff18 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/Barrier.h
@@ -20,9 +20,7 @@ class Barrier {
   Barrier(unsigned int count) : state_(count << 1), notified_(false) {
     eigen_assert(((count << 1) >> 1) == count);
   }
-  ~Barrier() {
-    eigen_assert((state_>>1) == 0);
-  }
+  ~Barrier() { eigen_plain_assert((state_ >> 1) == 0); }
 
   void Notify() {
     unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2;
@@ -52,14 +50,13 @@ class Barrier {
   bool notified_;
 };
 
-
 // Notification is an object that allows a user to to wait for another
 // thread to signal a notification that an event has occurred.
 //
 // Multiple threads can wait on the same Notification object,
 // but only one caller must call Notify() on the object.
 struct Notification : Barrier {
-  Notification() : Barrier(1) {};
+  Notification() : Barrier(1){};
 };
 
 }  // namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
index 0a7181102..22c952ae1 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
@@ -58,7 +58,7 @@ class EventCount {
 
   ~EventCount() {
     // Ensure there are no waiters.
-    eigen_assert((state_.load() & (kStackMask | kWaiterMask)) == kStackMask);
+    eigen_plain_assert((state_.load() & (kStackMask | kWaiterMask)) == kStackMask);
   }
 
   // Prewait prepares for waiting.
@@ -169,7 +169,8 @@ class EventCount {
 
   class Waiter {
     friend class EventCount;
-    // Align to 128 byte boundary to prevent false sharing with other Waiter objects in the same vector.
+    // Align to 128 byte boundary to prevent false sharing with other Waiter
+    // objects in the same vector.
     EIGEN_ALIGN_TO_BOUNDARY(128) std::atomic<Waiter*> next;
     std::mutex mu;
     std::condition_variable cv;
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index ede70da8d..354995be8 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -90,8 +90,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     }
 
     // Join threads explicitly to avoid destruction order issues.
-    for (size_t i = 0; i < num_threads_; i++) delete threads_[i];
-    for (size_t i = 0; i < num_threads_; i++) delete queues_[i];
+    for (int i = 0; i < num_threads_; i++) delete threads_[i];
+    for (int i = 0; i < num_threads_; i++) delete queues_[i];
 #ifndef EIGEN_THREAD_LOCAL
     for (auto it : per_thread_map_) delete it.second;
 #endif
@@ -298,7 +298,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // If we are shutting down and all worker threads blocked without work,
     // that's we are done.
     blocked_++;
-    if (done_ && blocked_ == num_threads_) {
+    if (done_ && blocked_ == static_cast<unsigned>(num_threads_)) {
       ec_.CancelWait(waiter);
       // Almost done, but need to re-check queues.
       // Consider that all queues are empty and all worker threads are preempted
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h b/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
index cb3690a2e..05c739aa1 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/RunQueue.h
@@ -10,7 +10,6 @@
 #ifndef EIGEN_CXX11_THREADPOOL_RUNQUEUE_H_
 #define EIGEN_CXX11_THREADPOOL_RUNQUEUE_H_
 
-
 namespace Eigen {
 
 // RunQueue is a fixed-size, partially non-blocking deque or Work items.
@@ -47,7 +46,7 @@ class RunQueue {
       array_[i].state.store(kEmpty, std::memory_order_relaxed);
   }
 
-  ~RunQueue() { eigen_assert(Size() == 0); }
+  ~RunQueue() { eigen_plain_assert(Size() == 0); }
 
   // PushFront inserts w at the beginning of the queue.
   // If queue is full returns w, otherwise returns default-constructed Work.
@@ -131,9 +130,8 @@ class RunQueue {
       Elem* e = &array_[mid & kMask];
       uint8_t s = e->state.load(std::memory_order_relaxed);
       if (n == 0) {
-        if (s != kReady ||
-            !e->state.compare_exchange_strong(s, kBusy,
-                                              std::memory_order_acquire))
+        if (s != kReady || !e->state.compare_exchange_strong(
+                               s, kBusy, std::memory_order_acquire))
           continue;
         start = mid;
       } else {

From ab3f481141a6bc72d2bbdc6300fb9dc157029ea9 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 14 Aug 2018 14:05:46 -0700
Subject: [PATCH 567/680] Cleaned up the code and make it compile with more
 compilers

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 76 ++++++++-----------
 1 file changed, 32 insertions(+), 44 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 322260011..24a6343e8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -91,7 +91,7 @@ EIGEN_STRONG_INLINE void MergeResourceRequirements(
   *block_total_size = resources[0].block_total_size;
   for (std::vector<TensorOpResourceRequirements>::size_type i = 1; i < resources.size(); ++i) {
     if (resources[i].block_shape == kSkewedInnerDims &&
-        *block_shape ! kSkewedInnerDims) {
+        *block_shape != kSkewedInnerDims) {
       *block_shape = kSkewedInnerDims;
     }
     *block_total_size =
@@ -152,11 +152,11 @@ struct TensorBlockCopyOp {
     const Scalar* src_base = &src_data[src_index];
     Scalar* dst_base = &dst_data[dst_index];
 
-    typedef const Eigen::Array<Scalar, Dynamic, 1> Src;
-    typedef Eigen::Array<Scalar, Dynamic, 1> Dst;
+    typedef const Array<Scalar, Dynamic, 1> Src;
+    typedef Array<Scalar, Dynamic, 1> Dst;
 
-    typedef Eigen::Map<Src, 0, InnerStride<> > SrcMap;
-    typedef Eigen::Map<Dst, 0, InnerStride<> > DstMap;
+    typedef Map<Src, 0, InnerStride<> > SrcMap;
+    typedef Map<Dst, 0, InnerStride<> > DstMap;
 
     const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
     DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
@@ -178,10 +178,8 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout,
           bool BlockRead>
 class TensorBlockIO {
  public:
-  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
-      TensorBlock;
-  typedef typename TensorBlockCopyOp<Scalar, StorageIndex>
-      TensorBlockCopyOp;
+  typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
+  typedef TensorBlockCopyOp<Scalar, StorageIndex> BlockCopyOp;
 
  protected:
   struct BlockIteratorState {
@@ -194,7 +192,7 @@ class TensorBlockIO {
   };
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(
-      const TensorBlock& block, StorageIndex first_coeff_index,
+      const Block& block, StorageIndex first_coeff_index,
       const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
       const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data,
       Scalar* dst_data) {
@@ -290,8 +288,8 @@ class TensorBlockIO {
     const StorageIndex block_total_size =
         NumDims == 0 ? 1 : block.block_sizes().TotalSize();
     for (StorageIndex i = 0; i < block_total_size; i += block_inner_dim_size) {
-      TensorBlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
-                             dst_data, inputIndex, input_stride, src_data);
+      BlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
+                       dst_data, inputIndex, input_stride, src_data);
       // Update index.
       for (int j = 0; j < num_squeezed_dims; ++j) {
         if (++block_iter_state[j].count < block_iter_state[j].size) {
@@ -320,13 +318,11 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
                                                Layout, /*BlockRead=*/true> {
  public:
-  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
-      TensorBlock;
-  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true>
-      Base;
+  typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
+  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true> Base;
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      TensorBlock* block, const Scalar* src_data) {
+      Block* block, const Scalar* src_data) {
     array<StorageIndex, NumDims> tensor_to_block_dim_map;
     for (int i = 0; i < NumDims; ++i) {
       tensor_to_block_dim_map[i] = i;
@@ -336,7 +332,7 @@ class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
   }
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      TensorBlock* block, StorageIndex first_coeff_index,
+      Block* block, StorageIndex first_coeff_index,
       const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
       const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data) {
     Base::Copy(*block, first_coeff_index, tensor_to_block_dim_map,
@@ -357,13 +353,11 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
                                                Layout, /*BlockRead=*/false> {
  public:
-  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
-      TensorBlock;
-  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false>
-      Base;
+  typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
+  typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false> Base;
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const TensorBlock& block, Scalar* dst_data) {
+      const Block& block, Scalar* dst_data) {
     array<StorageIndex, NumDims> tensor_to_block_dim_map;
     for (int i = 0; i < NumDims; ++i) {
       tensor_to_block_dim_map[i] = i;
@@ -373,7 +367,7 @@ class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
   }
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
-      const TensorBlock& block, StorageIndex first_coeff_index,
+      const Block& block, StorageIndex first_coeff_index,
       const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
       const array<StorageIndex, NumDims>& tensor_strides, Scalar* dst_data) {
     Base::Copy(block, first_coeff_index, tensor_to_block_dim_map,
@@ -401,13 +395,13 @@ struct TensorBlockCwiseBinaryOp {
       const StorageIndex left_stride, const LeftScalar* left_data,
       const StorageIndex right_index, const StorageIndex right_stride,
       const RightScalar* right_data) {
-    typedef const Eigen::Array<LeftScalar, Dynamic, 1> Lhs;
-    typedef const Eigen::Array<RightScalar, Dynamic, 1> Rhs;
-    typedef Eigen::Array<OutputScalar, Dynamic, 1> Out;
+    typedef const Array<LeftScalar, Dynamic, 1> Lhs;
+    typedef const Array<RightScalar, Dynamic, 1> Rhs;
+    typedef Array<OutputScalar, Dynamic, 1> Out;
 
-    typedef Eigen::Map<Lhs, 0, InnerStride<> > LhsMap;
-    typedef Eigen::Map<Rhs, 0, InnerStride<> > RhsMap;
-    typedef Eigen::Map<Out, 0, InnerStride<> > OutMap;
+    typedef Map<Lhs, 0, InnerStride<> > LhsMap;
+    typedef Map<Rhs, 0, InnerStride<> > RhsMap;
+    typedef Map<Out, 0, InnerStride<> > OutMap;
 
     const LeftScalar* lhs_base = &left_data[left_index];
     const RightScalar* rhs_base = &right_data[right_index];
@@ -417,8 +411,7 @@ struct TensorBlockCwiseBinaryOp {
     const RhsMap rhs(rhs_base, num_coeff, InnerStride<>(right_stride));
     OutMap out(out_base, num_coeff, InnerStride<>(output_stride));
 
-    out =
-        Eigen::CwiseBinaryOp<BinaryFunctor, LhsMap, RhsMap>(lhs, rhs, functor);
+    out = CwiseBinaryOp<BinaryFunctor, LhsMap, RhsMap>(lhs, rhs, functor);
   }
 };
 
@@ -434,8 +427,7 @@ struct TensorBlockCwiseBinaryOp {
 template <typename BinaryFunctor, typename StorageIndex, typename OutputScalar,
           int NumDims, int Layout>
 struct TensorBlockCwiseBinaryIO {
-  typedef typename TensorBlock<OutputScalar, StorageIndex, NumDims,
-                                         Layout>::Dimensions Dimensions;
+  typedef typename TensorBlock<OutputScalar, StorageIndex, NumDims, Layout>::Dimensions Dimensions;
 
   struct BlockIteratorState {
     StorageIndex output_stride, output_span;
@@ -627,8 +619,7 @@ struct TensorBlockView {
 template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorBlockMapper {
  public:
-  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
-      TensorBlock;
+  typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
   typedef DSizes<StorageIndex, NumDims> Dimensions;
 
   TensorBlockMapper(const Dimensions& dims,
@@ -663,7 +654,7 @@ class TensorBlockMapper {
     }
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Block
   GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
     StorageIndex first_coeff_index = 0;
     DSizes<StorageIndex, NumDims> coords;
@@ -711,8 +702,7 @@ class TensorBlockMapper {
       }
     }
 
-    return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
-                       data);
+    return Block(first_coeff_index, sizes, strides, m_tensor_strides, data);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {
@@ -818,8 +808,7 @@ class TensorBlockMapper {
 template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
 class TensorSliceBlockMapper {
  public:
-  typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
-      TensorBlock;
+  typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
   typedef DSizes<StorageIndex, NumDims> Dimensions;
 
   TensorSliceBlockMapper(const Dimensions& tensor_dims,
@@ -860,7 +849,7 @@ class TensorSliceBlockMapper {
     }
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Block
   GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
     StorageIndex first_coeff_index = 0;
     DSizes<StorageIndex, NumDims> coords;
@@ -917,8 +906,7 @@ class TensorSliceBlockMapper {
       }
     }
 
-    return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
-                       data);
+    return Block(first_coeff_index, sizes, strides, m_tensor_strides, data);
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {

From 43ec0082a6ca5e6098bdca1728094b0b7ce550c4 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 14 Aug 2018 14:08:36 -0700
Subject: [PATCH 568/680] Made the kronecker_product test compile again

---
 unsupported/test/kronecker_product.cpp | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp
index 1a936ed25..b5b764c65 100644
--- a/unsupported/test/kronecker_product.cpp
+++ b/unsupported/test/kronecker_product.cpp
@@ -10,8 +10,6 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 
-#include "main.h"
-
 #ifdef EIGEN_TEST_PART_1
 
 #include "sparse.h"
@@ -239,6 +237,7 @@ EIGEN_DECLARE_TEST(kronecker_product)
 #ifdef EIGEN_TEST_PART_2
 
 // simply check that for a dense kronecker product, sparse module is not needed
+#include "main.h"
 #include <Eigen/KroneckerProduct>
 
 EIGEN_DECLARE_TEST(kronecker_product)

From 6bb3f1b43ecad440fb7ad172657f0f7a0c804d29 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 14 Aug 2018 14:26:59 -0700
Subject: [PATCH 569/680] Made the tensor_block_access test compile again

---
 .../test/cxx11_tensor_block_access.cpp        | 120 +++++++++---------
 1 file changed, 58 insertions(+), 62 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 6feeff231..2631767e2 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -10,6 +10,8 @@
 
 #include "main.h"
 
+#include <algorithm>
+#include <random>
 #include <set>
 
 #include <Eigen/CXX11/Tensor>
@@ -19,17 +21,16 @@ using Eigen::Index;
 using Eigen::RowMajor;
 using Eigen::ColMajor;
 
-using internal::TensorBlockShapeType;
 
 template<typename T>
 static const T& choose(int layout, const T& col, const T& row) {
   return layout == ColMajor ? col : row;
 }
 
-static const TensorBlockShapeType RandomShape() {
+static internal::TensorBlockShapeType RandomShape() {
   return internal::random<bool>()
-             ? internal::TensorBlockShapeType::kUniformAllDims
-             : internal::TensorBlockShapeType::kSkewedInnerDims;
+             ? internal::kUniformAllDims
+             : internal::kSkewedInnerDims;
 }
 
 template <int NumDims>
@@ -44,7 +45,7 @@ static DSizes<Index, NumDims> RandomDims() {
     dims[i] = internal::random<int>(1, 20);
   }
   return DSizes<Index, NumDims>(dims);
-};
+}
 
 /** Dummy data type to test TensorBlock copy ops. */
 struct Data {
@@ -91,21 +92,19 @@ static void Debug(DSizes<Index, NumDims> dims) {
 template <int Layout>
 static void test_block_mapper_sanity()
 {
-  using T = int;
-  using TensorBlock = internal::TensorBlock<T, Index, 2, Layout>;
-  using TensorBlockMapper = internal::TensorBlockMapper<T, Index, 2, Layout>;
+  typedef internal::TensorBlockMapper<int, Index, 2, Layout> TensorBlockMapper;
 
   DSizes<Index, 2> tensor_dims(100, 100);
 
   // Test uniform blocks.
   TensorBlockMapper uniform_block_mapper(
-      tensor_dims, internal::TensorBlockShapeType::kUniformAllDims, 100);
+      tensor_dims, internal::kUniformAllDims, 100);
 
   VERIFY_IS_EQUAL(uniform_block_mapper.total_block_count(), 100);
   VERIFY_IS_EQUAL(uniform_block_mapper.block_dims_total_size(), 100);
 
   // 10x10 blocks
-  auto uniform_b0 = uniform_block_mapper.GetBlockForIndex(0, nullptr);
+  auto uniform_b0 = uniform_block_mapper.GetBlockForIndex(0, NULL);
   VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(0), 10);
   VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(1), 10);
   // Depending on a layout we stride by cols rows.
@@ -117,13 +116,13 @@ static void test_block_mapper_sanity()
 
   // Test skewed to inner dims blocks.
   TensorBlockMapper skewed_block_mapper(
-      tensor_dims, internal::TensorBlockShapeType::kSkewedInnerDims, 100);
+      tensor_dims, internal::kSkewedInnerDims, 100);
 
   VERIFY_IS_EQUAL(skewed_block_mapper.total_block_count(), 100);
   VERIFY_IS_EQUAL(skewed_block_mapper.block_dims_total_size(), 100);
 
   // 1x100 (100x1) rows/cols depending on a tensor layout.
-  auto skewed_b0 = skewed_block_mapper.GetBlockForIndex(0, nullptr);
+  auto skewed_b0 = skewed_block_mapper.GetBlockForIndex(0, NULL);
   VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(0), choose(Layout, 100, 1));
   VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(1), choose(Layout, 1, 100));
   // Depending on a layout we stride by cols rows.
@@ -171,7 +170,7 @@ static void test_block_mapper_maps_every_element() {
   TensorBlockMapper block_mapper(dims, RandomShape(), RandomTargetSize(dims));
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
-    TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(i, NULL);
     UpdateCoeffSet<T, Layout, NumDims>(block, block.first_coeff_index(),
                                        choose(Layout, NumDims - 1, 0),
                                        &coeff_set);
@@ -219,7 +218,7 @@ static void test_slice_block_mapper_maps_every_element() {
                                       DimensionList<Index, NumDims>());
 
   for (int i = 0; i < block_mapper.total_block_count(); ++i) {
-    TensorBlock block = block_mapper.GetBlockForIndex(i, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(i, NULL);
     UpdateCoeffSet<T, Layout, NumDims>(block, block.first_coeff_index(),
                                        choose(Layout, NumDims - 1, 0),
                                        &coeff_set);
@@ -647,17 +646,16 @@ static void test_block_cwise_binary_io_zero_strides() {
 template <int Layout>
 static void test_uniform_block_shape()
 {
-  using T = int;
-  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
+  typedef internal::TensorBlock<int, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<int, Index, 5, Layout> TensorBlockMapper;
 
   {
     // Test shape 'UniformAllDims' with uniform 'max_coeff count'.
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 5 * 5 * 5 * 5 * 5;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     for (int i = 0; i < 5; ++i) {
       VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
     }
@@ -669,9 +667,9 @@ static void test_uniform_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 7 * 5 * 5 * 5 * 5;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
     for (int i = 1; i < 5; ++i) {
       VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
@@ -680,9 +678,9 @@ static void test_uniform_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 5 * 5 * 5 * 5 * 6;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(6, block.block_sizes()[4]);
     for (int i = 3; i >= 0; --i) {
       VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
@@ -695,9 +693,9 @@ static void test_uniform_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 5 * 5 * 5 * 5;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
     for (int i = 1; i < 5; ++i) {
       VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
@@ -706,9 +704,9 @@ static void test_uniform_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 5 * 5 * 5 * 5 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     for (int i = 3; i >= 0; --i) {
       VERIFY_IS_EQUAL(5, block.block_sizes()[i]);
@@ -721,9 +719,9 @@ static void test_uniform_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(7, 5, 6, 17, 7);
     const size_t max_coeff_count = 7 * 5 * 6 * 7 * 5;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
@@ -733,9 +731,9 @@ static void test_uniform_block_shape()
   } else {
     DSizes<Index, 5> dims(7, 5, 6, 9, 7);
     const size_t max_coeff_count = 5 * 5 * 5 * 6 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[3]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
@@ -748,9 +746,9 @@ static void test_uniform_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(7, 5, 6, 17, 7);
     const size_t max_coeff_count = 7 * 5 * 6 * 17 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[0]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
@@ -760,9 +758,9 @@ static void test_uniform_block_shape()
   } else {
     DSizes<Index, 5> dims(7, 5, 6, 9, 7);
     const size_t max_coeff_count = 7 * 5 * 6 * 9 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kUniformAllDims,
+    TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     VERIFY_IS_EQUAL(9, block.block_sizes()[3]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
@@ -783,9 +781,9 @@ static void test_skewed_inner_dim_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 10 * 1 * 1 * 1 * 1;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(10, block.block_sizes()[0]);
     for (int i = 1; i < 5; ++i) {
       VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
@@ -794,9 +792,9 @@ static void test_skewed_inner_dim_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 1 * 1 * 1 * 1 * 6;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(6, block.block_sizes()[4]);
     for (int i = 3; i >= 0; --i) {
       VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
@@ -808,9 +806,9 @@ static void test_skewed_inner_dim_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 1 * 1 * 1 * 1;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
     for (int i = 1; i < 5; ++i) {
       VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
@@ -819,9 +817,9 @@ static void test_skewed_inner_dim_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 1 * 1 * 1 * 1 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     for (int i = 3; i >= 0; --i) {
       VERIFY_IS_EQUAL(1, block.block_sizes()[i]);
@@ -834,9 +832,9 @@ static void test_skewed_inner_dim_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 3 * 1 * 1 * 1;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
     VERIFY_IS_EQUAL(3, block.block_sizes()[1]);
     for (int i = 2; i < 5; ++i) {
@@ -846,9 +844,9 @@ static void test_skewed_inner_dim_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 1 * 1 * 1 * 15 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     VERIFY_IS_EQUAL(15, block.block_sizes()[3]);
     for (int i = 2; i >= 0; --i) {
@@ -862,9 +860,9 @@ static void test_skewed_inner_dim_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 5 * 5 * 1 * 1;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
@@ -875,9 +873,9 @@ static void test_skewed_inner_dim_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 1 * 1 * 5 * 17 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[2]);
@@ -891,9 +889,9 @@ static void test_skewed_inner_dim_block_shape()
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(11, block.block_sizes()[0]);
     VERIFY_IS_EQUAL(5, block.block_sizes()[1]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
@@ -903,9 +901,9 @@ static void test_skewed_inner_dim_block_shape()
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
     const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
-    TensorBlockMapper block_mapper(dims, TensorBlockShapeType::kSkewedInnerDims,
+    TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
-    TensorBlock block = block_mapper.GetBlockForIndex(0, nullptr);
+    TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
     VERIFY_IS_EQUAL(7, block.block_sizes()[4]);
     VERIFY_IS_EQUAL(17, block.block_sizes()[3]);
     VERIFY_IS_EQUAL(6, block.block_sizes()[2]);
@@ -918,15 +916,13 @@ static void test_skewed_inner_dim_block_shape()
 template <int Layout>
 static void test_empty_dims(const internal::TensorBlockShapeType block_shape)
 {
-  using T = int;
-
   // Test blocking of tensors with zero dimensions:
   //  - we must not crash on asserts and divisions by zero
   //  - we must not return block with zero dimensions
   //    (recipe for overflows/underflows, divisions by zero and NaNs later)
   //  - total block count must be zero
   {
-    typedef internal::TensorBlockMapper<T, Index, 1, Layout> TensorBlockMapper;
+    typedef internal::TensorBlockMapper<int, Index, 1, Layout> TensorBlockMapper;
     DSizes<Index, 1> dims(0);
     for (int max_coeff_count = 0; max_coeff_count < 2; ++max_coeff_count) {
       TensorBlockMapper block_mapper(dims, block_shape, max_coeff_count);
@@ -936,7 +932,7 @@ static void test_empty_dims(const internal::TensorBlockShapeType block_shape)
   }
 
   {
-    typedef internal::TensorBlockMapper<T, Index, 2, Layout> TensorBlockMapper;
+    typedef internal::TensorBlockMapper<int, Index, 2, Layout> TensorBlockMapper;
     for (int dim1 = 0; dim1 < 3; ++dim1) {
       for (int dim2 = 0; dim2 < 3; ++dim2) {
         DSizes<Index, 2> dims(dim1, dim2);
@@ -987,9 +983,9 @@ EIGEN_DECLARE_TEST(cxx11_tensor_block_access) {
   TEST_LAYOUTS(test_block_cwise_binary_io_zero_strides);
   TEST_LAYOUTS(test_uniform_block_shape);
   TEST_LAYOUTS(test_skewed_inner_dim_block_shape);
-  TEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kUniformAllDims);
-  TEST_LAYOUTS_WITH_ARG(test_empty_dims, TensorBlockShapeType::kSkewedInnerDims);
+  TEST_LAYOUTS_WITH_ARG(test_empty_dims, internal::kUniformAllDims);
+  TEST_LAYOUTS_WITH_ARG(test_empty_dims, internal::kSkewedInnerDims);
 }
 
 #undef TEST_LAYOUTS
-#undef TEST_LAYOUTS_WITH_ARG
\ No newline at end of file
+#undef TEST_LAYOUTS_WITH_ARG

From fbb834144df6190a93757098d097f230b167edc5 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 15 Aug 2018 08:52:58 -0700
Subject: [PATCH 570/680] Fixed more compilation errors

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 0cefe42dd..676645b0c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -132,7 +132,7 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable,
     if (needs_assign) {
       // Size tensor blocks to fit in cache (or requested target block size).
       Index block_total_size = numext::mini(cache_size, total_size);
-      TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
+      TensorBlockShapeType block_shape = kSkewedInnerDims;
       // Query expression tree for desired block size/shape.
       std::vector<TensorOpResourceRequirements> resources;
       evaluator.getResourceRequirements(&resources);
@@ -272,7 +272,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
 
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
-      TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
+      TensorBlockShapeType block_shape = kSkewedInnerDims;
       Index block_total_size = 0;
       // Query expression tree for desired block size/shape.
       std::vector<internal::TensorOpResourceRequirements> resources;

From b6f96cf7dd616ed1604919892f68f5b94d31fa5e Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 15 Aug 2018 08:54:31 -0700
Subject: [PATCH 571/680] Removed dependencies on cxx11 language features from
 the tensor_block_access test

---
 unsupported/test/cxx11_tensor_block_access.cpp | 15 ++++++---------
 1 file changed, 6 insertions(+), 9 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 2631767e2..417b72201 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -157,9 +157,8 @@ static void UpdateCoeffSet(
 
 template <typename T, int NumDims, int Layout>
 static void test_block_mapper_maps_every_element() {
-  using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
-  using TensorBlockMapper =
-      internal::TensorBlockMapper<T, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<T, Index, NumDims, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<T, Index, NumDims, Layout> TensorBlockMapper;
 
   DSizes<Index, NumDims> dims = RandomDims<NumDims>();
 
@@ -186,9 +185,8 @@ static void test_block_mapper_maps_every_element() {
 
 template <typename T, int NumDims, int Layout>
 static void test_slice_block_mapper_maps_every_element() {
-  using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
-  using TensorSliceBlockMapper =
-      internal::TensorSliceBlockMapper<T, Index, NumDims, Layout>;
+  typedef internal::TensorBlock<T, Index, NumDims, Layout> TensorBlock;
+  typedef internal::TensorSliceBlockMapper<T, Index, NumDims, Layout> TensorSliceBlockMapper;
 
   DSizes<Index, NumDims> tensor_dims = RandomDims<NumDims>();
   DSizes<Index, NumDims> tensor_slice_offsets = RandomDims<NumDims>();
@@ -773,9 +771,8 @@ static void test_uniform_block_shape()
 template <int Layout>
 static void test_skewed_inner_dim_block_shape()
 {
-  using T = int;
-  typedef internal::TensorBlock<T, Index, 5, Layout> TensorBlock;
-  typedef internal::TensorBlockMapper<T, Index, 5, Layout> TensorBlockMapper;
+  typedef internal::TensorBlock<int, Index, 5, Layout> TensorBlock;
+  typedef internal::TensorBlockMapper<int, Index, 5, Layout> TensorBlockMapper;
 
   // Test shape 'SkewedInnerDims' with partial allocation to inner-most dim.
   if (Layout == ColMajor) {

From 4181556907fd29d6328fb718fa42cf9ce4734133 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Wed, 15 Aug 2018 09:34:47 -0700
Subject: [PATCH 572/680] Fixed the tensor contraction code.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 10 ++--------
 unsupported/test/cxx11_tensor_contraction.cpp          |  2 +-
 unsupported/test/cxx11_tensor_thread_pool.cpp          |  2 +-
 3 files changed, 4 insertions(+), 10 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index a023718c6..5d619efd8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -152,13 +152,7 @@ struct TensorContractionParams {
 //   1. Elementwise Relu transformation following Conv2D.
 //   2. AddBias to the Conv2D output channels dimension.
 //
-// See expected implementation in NoOpOutputKernel.
-struct OutputKernel {
-  template <typename Index, typename Scalar>
-  typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
-};
-
-// Output kernel that does absolutely nothing.
+// The NoOpOutputKernel implements an output kernel that does absolutely nothing.
 struct NoOpOutputKernel {
   /**
    * Tensor contraction evaluator calls this kernel after finishing each block
@@ -177,7 +171,7 @@ struct NoOpOutputKernel {
    */
   template <typename Index, typename Scalar>
   EIGEN_ALWAYS_INLINE void operator()(
-      const OutputKernel::OutputMapper<Index, Scalar>& /*output_mapper*/,
+      const internal::blas_data_mapper<Scalar, Index, ColMajor>& /*output_mapper*/,
       const TensorContractionParams& /*params*/, Index /*i*/,
       Index /*j*/, Index /*num_rows*/, Index /*num_cols*/) const {}
 };
diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp
index 2e918eb30..928d20f6e 100644
--- a/unsupported/test/cxx11_tensor_contraction.cpp
+++ b/unsupported/test/cxx11_tensor_contraction.cpp
@@ -514,7 +514,7 @@ static void test_const_inputs()
 struct SqrtOutputKernel {
   template <typename Index, typename Scalar>
   EIGEN_ALWAYS_INLINE void operator()(
-      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
+      const internal::blas_data_mapper<Scalar, Index, ColMajor>& output_mapper,
       const TensorContractionParams&, Index, Index, Index num_rows,
       Index num_cols) const {
     for (int i = 0; i < num_rows; ++i) {
diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index dd163c18a..7606b0abf 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -255,7 +255,7 @@ void test_multithread_contraction_agrees_with_singlethread() {
 struct SqrtOutputKernel {
   template <typename Index, typename Scalar>
   EIGEN_ALWAYS_INLINE void operator()(
-      const OutputKernel::OutputMapper<Index, Scalar>& output_mapper,
+      const internal::blas_data_mapper<Scalar, Index, ColMajor>& output_mapper,
       const TensorContractionParams&, Index, Index, Index num_rows,
       Index num_cols) const {
     for (int i = 0; i < num_rows; ++i) {

From f197c3f55b3a04ab24dfee8057b1d510c7483fc3 Mon Sep 17 00:00:00 2001
From: Sameer Agarwal <sameeragarwal@google.com>
Date: Wed, 15 Aug 2018 11:24:57 -0700
Subject: [PATCH 573/680] Removed an used variable (PacketSize) from
 TensorExecutor

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 676645b0c..9b9587de5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -229,10 +229,6 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
     Evaluator evaluator(expr, device);
     const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
     if (needs_assign) {
-      const StorageIndex PacketSize =
-          Vectorizable
-              ? unpacket_traits<typename Evaluator::PacketReturnType>::size
-              : 1;
       const StorageIndex size = array_prod(evaluator.dimensions());
       device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
                          EvalRange::alignBlockSize,

From d0b01ebbf63f0594760b6e1568bec0228987157a Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 16 Aug 2018 13:21:36 +0100
Subject: [PATCH 574/680] Reverting the unitended delete from the code.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index ab5990c14..dde1b449e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -30,6 +30,7 @@ struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
   typedef typename remove_reference<Nested>::type _Nested;
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
 
 };
 
@@ -185,6 +186,8 @@ struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   static const int NumDimensions = traits<LhsXprType>::NumDimensions;
   static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
 
 };
 

From 80f1a76dec9a5fbe4305633ab0c8797a876e4ab5 Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Thu, 16 Aug 2018 13:33:24 +0100
Subject: [PATCH 575/680] removing the noises.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index dde1b449e..cbec5e9b4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -31,7 +31,6 @@ struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
   static const int NumDimensions = traits<XprType>::NumDimensions;
   static const int Layout = traits<XprType>::Layout;
   typedef typename traits<XprType>::PointerType PointerType;
-
 };
 
 template<typename CustomUnaryFunc, typename XprType>
@@ -188,7 +187,6 @@ struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
   static const int Layout = traits<LhsXprType>::Layout;
   typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
                                 typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
-
 };
 
 template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>

From e23c8c294e57d6600086ede5480c2e6c89db56ec Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 16 Aug 2018 10:41:01 -0700
Subject: [PATCH 576/680] Use actual types instead of the auto keyword to make
 the code more portable

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 5d619efd8..2ca91709f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -660,7 +660,7 @@ struct TensorContractionEvaluatorBase
 
           // call gebp (matrix kernel)
           // The parameters here are copied from Eigen's GEMM implementation
-          const auto output_mapper = output.getSubMapper(i2, j2);
+          const OutputMapper output_mapper = output.getSubMapper(i2, j2);
           gebp(output_mapper, blockA, blockB, actual_mc, actual_kc, actual_nc,
                Scalar(1), -1, -1, 0, 0);
 

From ede580ccdac3b964bdfcf12d55560a268c366c3c Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 16 Aug 2018 10:49:47 -0700
Subject: [PATCH 577/680] Avoid using the auto keyword to make the tensor block
 access test more portable

---
 .../test/cxx11_tensor_block_access.cpp        | 56 +++++++++----------
 1 file changed, 28 insertions(+), 28 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 417b72201..da093166b 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -104,7 +104,7 @@ static void test_block_mapper_sanity()
   VERIFY_IS_EQUAL(uniform_block_mapper.block_dims_total_size(), 100);
 
   // 10x10 blocks
-  auto uniform_b0 = uniform_block_mapper.GetBlockForIndex(0, NULL);
+  typename TensorBlockMapper::Block uniform_b0 = uniform_block_mapper.GetBlockForIndex(0, NULL);
   VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(0), 10);
   VERIFY_IS_EQUAL(uniform_b0.block_sizes().at(1), 10);
   // Depending on a layout we stride by cols rows.
@@ -122,7 +122,7 @@ static void test_block_mapper_sanity()
   VERIFY_IS_EQUAL(skewed_block_mapper.block_dims_total_size(), 100);
 
   // 1x100 (100x1) rows/cols depending on a tensor layout.
-  auto skewed_b0 = skewed_block_mapper.GetBlockForIndex(0, NULL);
+  typename TensorBlockMapper::Block skewed_b0 = skewed_block_mapper.GetBlockForIndex(0, NULL);
   VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(0), choose(Layout, 100, 1));
   VERIFY_IS_EQUAL(skewed_b0.block_sizes().at(1), choose(Layout, 1, 100));
   // Depending on a layout we stride by cols rows.
@@ -203,7 +203,7 @@ static void test_slice_block_mapper_maps_every_element() {
   // Keep track of elements indices available via block access.
   std::set<Index> coeff_set;
 
-  auto total_coeffs = static_cast<int>(tensor_slice_extents.TotalSize());
+  int total_coeffs = static_cast<int>(tensor_slice_extents.TotalSize());
 
   // Pick a random dimension sizes for the tensor blocks.
   DSizes<Index, NumDims> block_sizes;
@@ -237,7 +237,7 @@ static void test_block_io_copy_data_from_source_to_target() {
       TensorBlockWriter;
 
   DSizes<Index, NumDims> input_tensor_dims = RandomDims<NumDims>();
-  const auto input_tensor_size = input_tensor_dims.TotalSize();
+  const Index input_tensor_size = input_tensor_dims.TotalSize();
 
   T* input_data = GenerateRandomData<T>(input_tensor_size);
   T* output_data = new T[input_tensor_size];
@@ -316,7 +316,7 @@ static void test_block_io_copy_using_reordered_dimensions() {
       TensorBlockWriter;
 
   DSizes<Index, NumDims> input_tensor_dims = RandomDims<NumDims>();
-  const auto input_tensor_size = input_tensor_dims.TotalSize();
+  const Index input_tensor_size = input_tensor_dims.TotalSize();
 
   // Create a random input tensor.
   T* input_data = GenerateRandomData<T>(input_tensor_size);
@@ -339,8 +339,8 @@ static void test_block_io_copy_using_reordered_dimensions() {
   TensorBlockMapper block_mapper(output_tensor_dims, RandomShape(),
                                  RandomTargetSize(input_tensor_dims));
 
-  auto* block_data = new T[block_mapper.block_dims_total_size()];
-  auto* output_data = new T[input_tensor_size];
+  T* block_data = new T[block_mapper.block_dims_total_size()];
+  T* output_data = new T[input_tensor_size];
 
   array<Index, NumDims> input_tensor_strides =
       ComputeStrides<Layout, NumDims>(input_tensor_dims);
@@ -382,8 +382,8 @@ static void test_block_io_zero_stride()
   input_tensor_dims[0] = 1;
   input_tensor_dims[2] = 1;
   input_tensor_dims[4] = 1;
-  const auto input_tensor_size = input_tensor_dims.TotalSize();
-  auto* input_data = GenerateRandomData<float>(input_tensor_size);
+  const Index input_tensor_size = input_tensor_dims.TotalSize();
+  float* input_data = GenerateRandomData<float>(input_tensor_size);
 
   DSizes<Index, 5> output_tensor_dims = rnd_dims;
 
@@ -424,7 +424,7 @@ static void test_block_io_zero_stride()
   };
 
   {
-    auto* output_data = new float[output_tensor_dims.TotalSize()];
+    float* output_data = new float[output_tensor_dims.TotalSize()];
     TensorBlock read_block(0, output_tensor_dims, output_tensor_strides,
                            input_tensor_strides_with_zeros, output_data);
     TensorBlockReader::Run(&read_block, input_data);
@@ -433,7 +433,7 @@ static void test_block_io_zero_stride()
   }
 
   {
-    auto* output_data = new float[output_tensor_dims.TotalSize()];
+    float* output_data = new float[output_tensor_dims.TotalSize()];
     TensorBlock write_block(0, output_tensor_dims,
                             input_tensor_strides_with_zeros,
                             output_tensor_strides, input_data);
@@ -456,14 +456,14 @@ static void test_block_io_squeeze_ones() {
   // Total size > 1.
   {
     DSizes<Index, 5> block_sizes(1, 2, 1, 2, 1);
-    const auto total_size = block_sizes.TotalSize();
+    const Index total_size = block_sizes.TotalSize();
 
     // Create a random input tensor.
-    auto* input_data = GenerateRandomData<float>(total_size);
+    float* input_data = GenerateRandomData<float>(total_size);
     DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
 
     {
-      auto* output_data = new float[block_sizes.TotalSize()];
+      float* output_data = new float[block_sizes.TotalSize()];
       TensorBlock read_block(0, block_sizes, strides, strides, output_data);
       TensorBlockReader::Run(&read_block, input_data);
       for (int i = 0; i < total_size; ++i) {
@@ -473,7 +473,7 @@ static void test_block_io_squeeze_ones() {
     }
 
     {
-      auto* output_data = new float[block_sizes.TotalSize()];
+      float* output_data = new float[block_sizes.TotalSize()];
       TensorBlock write_block(0, block_sizes, strides, strides, input_data);
       TensorBlockWriter::Run(write_block, output_data);
       for (int i = 0; i < total_size; ++i) {
@@ -486,14 +486,14 @@ static void test_block_io_squeeze_ones() {
   // Total size == 1.
   {
     DSizes<Index, 5> block_sizes(1, 1, 1, 1, 1);
-    const auto total_size = block_sizes.TotalSize();
+    const Index total_size = block_sizes.TotalSize();
 
     // Create a random input tensor.
-    auto* input_data = GenerateRandomData<float>(total_size);
+    float* input_data = GenerateRandomData<float>(total_size);
     DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
 
     {
-      auto* output_data = new float[block_sizes.TotalSize()];
+      float* output_data = new float[block_sizes.TotalSize()];
       TensorBlock read_block(0, block_sizes, strides, strides, output_data);
       TensorBlockReader::Run(&read_block, input_data);
       for (int i = 0; i < total_size; ++i) {
@@ -503,7 +503,7 @@ static void test_block_io_squeeze_ones() {
     }
 
     {
-      auto* output_data = new float[block_sizes.TotalSize()];
+      float* output_data = new float[block_sizes.TotalSize()];
       TensorBlock write_block(0, block_sizes, strides, strides, input_data);
       TensorBlockWriter::Run(write_block, output_data);
       for (int i = 0; i < total_size; ++i) {
@@ -524,7 +524,7 @@ static void test_block_cwise_binary_io_basic() {
   DSizes<Index, NumDims> block_sizes = RandomDims<NumDims>();
   DSizes<Index, NumDims> strides(ComputeStrides<Layout, NumDims>(block_sizes));
 
-  const auto total_size = block_sizes.TotalSize();
+  const Index total_size = block_sizes.TotalSize();
 
   // Create a random input tensors.
   T* left_data = GenerateRandomData<T>(total_size);
@@ -553,13 +553,13 @@ static void test_block_cwise_binary_io_squeeze_ones() {
   DSizes<Index, 5> block_sizes(1, 2, 1, 3, 1);
   DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
 
-  const auto total_size = block_sizes.TotalSize();
+  const Index total_size = block_sizes.TotalSize();
 
   // Create a random input tensors.
-  auto* left_data = GenerateRandomData<float>(total_size);
-  auto* right_data = GenerateRandomData<float>(total_size);
+  float* left_data = GenerateRandomData<float>(total_size);
+  float* right_data = GenerateRandomData<float>(total_size);
 
-  auto* output_data = new float[total_size];
+  float* output_data = new float[total_size];
   BinaryFunctor functor;
   TensorBlockCwiseBinaryIO::Run(functor, block_sizes, strides, output_data,
                                 strides, left_data, strides, right_data);
@@ -600,14 +600,14 @@ static void test_block_cwise_binary_io_zero_strides() {
   right_strides[3] = 0;
 
   // Generate random data.
-  auto* left_data = GenerateRandomData<float>(left_sizes.TotalSize());
-  auto* right_data = GenerateRandomData<float>(right_sizes.TotalSize());
+  float* left_data = GenerateRandomData<float>(left_sizes.TotalSize());
+  float* right_data = GenerateRandomData<float>(right_sizes.TotalSize());
 
   DSizes<Index, 5> output_sizes = rnd_dims;
   DSizes<Index, 5> output_strides(ComputeStrides<Layout, 5>(output_sizes));
 
-  const auto output_total_size = output_sizes.TotalSize();
-  auto* output_data = new float[output_total_size];
+  const Index output_total_size = output_sizes.TotalSize();
+  float* output_data = new float[output_total_size];
 
   BinaryFunctor functor;
   TensorBlockCwiseBinaryIO::Run(functor, output_sizes, output_strides,

From f641cf1253bc2f7388b632a50b818b9d15b7588d Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 16 Aug 2018 11:24:37 -0700
Subject: [PATCH 578/680] Adding missing at method in Eigen::array

---
 unsupported/Eigen/CXX11/src/util/EmulateArray.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index d91662d96..32db51592 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -25,6 +25,11 @@ template <typename T, size_t n> class array {
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { return values[index]; }
 
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& at(size_t index) { eigen_assert(index < size()); return values[index]; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& at(size_t index) const { eigen_assert(index < size()); return values[index]; }
+
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE T& front() { return values[0]; }
   EIGEN_DEVICE_FUNC

From dbdeceabdd115293a2f6a9c17079940cf5b096dd Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 17 Aug 2018 16:26:11 +0200
Subject: [PATCH 579/680] Silence double-promotion warning (when converting
 double to complex<long double>)

---
 unsupported/Eigen/FFT | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/FFT b/unsupported/Eigen/FFT
index 2c45b3999..d8cf3e642 100644
--- a/unsupported/Eigen/FFT
+++ b/unsupported/Eigen/FFT
@@ -289,6 +289,7 @@ class FFT
     void inv( MatrixBase<OutputDerived> & dst, const MatrixBase<ComplexDerived> & src, Index nfft=-1)
     {
       typedef typename ComplexDerived::Scalar src_type;
+      typedef typename ComplexDerived::RealScalar real_type;
       typedef typename OutputDerived::Scalar dst_type;
       const bool realfft= (NumTraits<dst_type>::IsComplex == 0);
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OutputDerived)
@@ -329,9 +330,9 @@ class FFT
             tmp.head(nhead) = src.head(nhead);
             tmp.tail(ntail) = src.tail(ntail);
             if (resize_input<0) { //shrinking -- create the Nyquist bin as the average of the two bins that fold into it
-              tmp(nhead) = ( src(nfft/2) + src( src.size() - nfft/2 ) )*src_type(.5);
+              tmp(nhead) = ( src(nfft/2) + src( src.size() - nfft/2 ) )*real_type(.5);
             }else{ // expanding -- split the old Nyquist bin into two halves
-              tmp(nhead) = src(nhead) * src_type(.5);
+              tmp(nhead) = src(nhead) * real_type(.5);
               tmp(tmp.size()-nhead) = tmp(nhead);
             }
           }

From c9b25fbefa44f684f76e9a669be217c9d3e7734d Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 17 Aug 2018 16:28:28 +0200
Subject: [PATCH 580/680] Silence unused parameter warning

---
 unsupported/Eigen/src/BVH/KdBVH.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/src/BVH/KdBVH.h b/unsupported/Eigen/src/BVH/KdBVH.h
index 13f792cd0..2d5b76ad0 100644
--- a/unsupported/Eigen/src/BVH/KdBVH.h
+++ b/unsupported/Eigen/src/BVH/KdBVH.h
@@ -35,6 +35,7 @@ struct get_boxes_helper {
   {
     outBoxes.insert(outBoxes.end(), boxBegin, boxEnd);
     eigen_assert(outBoxes.size() == objects.size());
+    EIGEN_ONLY_USED_FOR_DEBUG(objects);
   }
 };
 

From 595cae9b09bf322a747f8ff5aade76448db58a17 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 17 Aug 2018 16:30:32 +0200
Subject: [PATCH 581/680] Silence logical-op-parentheses warning

---
 Eigen/src/SuperLUSupport/SuperLUSupport.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h
index 4bb95eb8b..354e33de5 100644
--- a/Eigen/src/SuperLUSupport/SuperLUSupport.h
+++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h
@@ -297,8 +297,8 @@ SluMatrix asSluMatrix(MatrixType& mat)
 template<typename Scalar, int Flags, typename Index>
 MappedSparseMatrix<Scalar,Flags,Index> map_superlu(SluMatrix& sluMat)
 {
-  eigen_assert((Flags&RowMajor)==RowMajor && sluMat.Stype == SLU_NR
-         || (Flags&ColMajor)==ColMajor && sluMat.Stype == SLU_NC);
+  eigen_assert(((Flags&RowMajor)==RowMajor && sluMat.Stype == SLU_NR)
+         || ((Flags&ColMajor)==ColMajor && sluMat.Stype == SLU_NC));
 
   Index outerSize = (Flags&RowMajor)==RowMajor ? sluMat.ncol : sluMat.nrow;
 

From 4713465eefbdb725ee3bb2cc3330bf77f51f1c6b Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 17 Aug 2018 16:39:43 +0200
Subject: [PATCH 582/680] Silence double-promotion warning

---
 unsupported/Eigen/OpenGLSupport | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/OpenGLSupport b/unsupported/Eigen/OpenGLSupport
index 11d99567e..489fd8354 100644
--- a/unsupported/Eigen/OpenGLSupport
+++ b/unsupported/Eigen/OpenGLSupport
@@ -184,7 +184,7 @@ inline void glRotate(const Rotation2D<float>& rot)
 }
 inline void glRotate(const Rotation2D<double>& rot)
 {
-  glRotated(rot.angle()*180.0/EIGEN_PI, 0.0, 0.0, 1.0);
+  glRotated(rot.angle()*180.0/double(EIGEN_PI), 0.0, 0.0, 1.0);
 }
 
 template<typename Derived> void glRotate(const RotationBase<Derived,3>& rot)

From 41f1cc67b8e55469367416151d6a82e3632cfda8 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 17 Aug 2018 16:42:53 +0200
Subject: [PATCH 583/680] Assertion depended on a not yet initialized value

---
 Eigen/src/Core/MapBase.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h
index 020f939ad..668922ffc 100644
--- a/Eigen/src/Core/MapBase.h
+++ b/Eigen/src/Core/MapBase.h
@@ -43,6 +43,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     enum {
       RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
       ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      InnerStrideAtCompileTime = internal::traits<Derived>::InnerStrideAtCompileTime,
       SizeAtCompileTime = Base::SizeAtCompileTime
     };
 
@@ -187,8 +188,11 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     void checkSanity(typename internal::enable_if<(internal::traits<T>::Alignment>0),void*>::type = 0) const
     {
 #if EIGEN_MAX_ALIGN_BYTES>0
+      // innerStride() is not set yet when this function is called, so we optimistically assume the lowest plausible value:
+      const Index minInnerStride = InnerStrideAtCompileTime == Dynamic ? 1 : Index(InnerStrideAtCompileTime);
+      EIGEN_ONLY_USED_FOR_DEBUG(minInnerStride);
       eigen_assert((   ((internal::UIntPtr(m_data) % internal::traits<Derived>::Alignment) == 0)
-                    || (cols() * rows() * innerStride() * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
+                    || (cols() * rows() * minInnerStride * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
 #endif
     }
 

From f76c802973a5e70309eb1d713adf59a5f9711ad1 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 17 Aug 2018 17:16:12 +0200
Subject: [PATCH 584/680] Add missing empty line

---
 test/main.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/main.h b/test/main.h
index de8a4865f..5d550969e 100644
--- a/test/main.h
+++ b/test/main.h
@@ -845,4 +845,4 @@ int main(int argc, char *argv[])
 #ifdef _MSC_VER
   // 4503 - decorated name length exceeded, name was truncated
   #pragma warning( disable : 4503)
-#endif
\ No newline at end of file
+#endif

From 43d9dd9b2844828de3996b762b2f760749df711b Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 17 Aug 2018 08:49:32 -0700
Subject: [PATCH 585/680] Removed more dependencies on cxx11.

---
 unsupported/test/cxx11_tensor_block_access.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index da093166b..7bf338522 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -11,7 +11,6 @@
 #include "main.h"
 
 #include <algorithm>
-#include <random>
 #include <set>
 
 #include <Eigen/CXX11/Tensor>
@@ -49,8 +48,8 @@ static DSizes<Index, NumDims> RandomDims() {
 
 /** Dummy data type to test TensorBlock copy ops. */
 struct Data {
-  Data() : Data(0) {}
-  explicit Data(int v) { value = v; }
+  Data() : value(0) {}
+  explicit Data(int v) : value(v) { }
   int value;
 };
 
@@ -324,7 +323,7 @@ static void test_block_io_copy_using_reordered_dimensions() {
   // Create a random dimension re-ordering/shuffle.
   std::vector<Index> shuffle;
   for (int i = 0; i < NumDims; ++i) shuffle.push_back(i);
-  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937());
+  std::random_shuffle(shuffle.begin(), shuffle.end());
 
   DSizes<Index, NumDims> output_tensor_dims;
   array<Index, NumDims> input_to_output_dim_map;

From ff8e0ecc2fa83ca4de14fc2f1049bd48907df3f6 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 17 Aug 2018 15:15:52 -0700
Subject: [PATCH 586/680] Updated one more line of code to avoid making the
 test dependent on cxx11 features.

---
 unsupported/test/cxx11_tensor_block_access.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 7bf338522..f572e496d 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -143,7 +143,8 @@ static void UpdateCoeffSet(
 
   for (int i = 0; i < block_sizes[dim_index]; ++i) {
     if (tensor_strides[dim_index] == 1) {
-      auto inserted = visited_coeffs->insert(first_coeff_index + i);
+      typedef std::pair<std::set<Index>::iterator, bool> ReturnType;
+      ReturnType inserted = visited_coeffs->insert(first_coeff_index + i);
       VERIFY_IS_EQUAL(inserted.second, true);
     } else {
       int next_dim_index = dim_index + choose(Layout, -1, 1);

From 39335cf51e7ea5edfe9113cb91034625a039ccbf Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 23 Aug 2018 19:37:56 +0200
Subject: [PATCH 587/680] Make MaxSizeVector leak-safe

---
 .../Eigen/CXX11/src/util/MaxSizeVector.h      | 44 +++++++++++++------
 1 file changed, 30 insertions(+), 14 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
index 4bc3dd1ba..bc5b3632c 100644
--- a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
+++ b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
@@ -35,7 +35,6 @@ class MaxSizeVector {
   explicit MaxSizeVector(size_t n)
       : reserve_(n), size_(0),
         data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
-    for (size_t i = 0; i < n; ++i) { new (&data_[i]) T; }
   }
 
   // Construct a new MaxSizeVector, reserve and resize to n.
@@ -44,35 +43,55 @@ class MaxSizeVector {
   MaxSizeVector(size_t n, const T& init)
       : reserve_(n), size_(n),
         data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
-    for (size_t i = 0; i < n; ++i) { new (&data_[i]) T(init); }
+    size_t i = 0;
+    EIGEN_TRY
+    {
+      for(; i < size_; ++i) { new (&data_[i]) T(init); }
+    }
+    EIGEN_CATCH(...)
+    {
+      // Construction failed, destruct in reverse order:
+      for(; (i+1) > 0; --i) { data_[i-1].~T(); }
+      internal::aligned_free(data_);
+      EIGEN_THROW;
+    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   ~MaxSizeVector() {
-    for (size_t i = 0; i < size_; ++i) {
-      data_[i].~T();
+    for (size_t i = size_; i > 0; --i) {
+      data_[i-1].~T();
     }
     internal::aligned_free(data_);
   }
 
   void resize(size_t n) {
     eigen_assert(n <= reserve_);
-    for (size_t i = size_; i < n; ++i) {
-      new (&data_[i]) T;
+    for (; size_ < n; ++size_) {
+      new (&data_[size_]) T;
     }
-    for (size_t i = n; i < size_; ++i) {
-      data_[i].~T();
+    for (; size_ > n; --size_) {
+      data_[size_-1].~T();
     }
-    size_ = n;
+    eigen_assert(size_ == n);
   }
 
   // Append new elements (up to reserved size).
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void push_back(const T& t) {
     eigen_assert(size_ < reserve_);
-    data_[size_++] = t;
+    new (&data_[size_++]) T(t);
   }
 
+  // For C++03 compatibility this only takes one argument
+  template<class X>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void emplace_back(const X& x) {
+    eigen_assert(size_ < reserve_);
+    new (&data_[size_++]) T(x);
+  }
+
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   const T& operator[] (size_t i) const {
     eigen_assert(i < size_);
@@ -99,11 +118,8 @@ class MaxSizeVector {
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void pop_back() {
-    // NOTE: This does not destroy the value at the end the way
-    // std::vector's version of pop_back() does.  That happens when
-    // the Vector is destroyed.
     eigen_assert(size_ > 0);
-    size_--;
+    data_[--size_].~T();
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE

From a709c8efb4927ebac338cb93865e8d0bdfcac85d Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 23 Aug 2018 19:41:59 +0200
Subject: [PATCH 588/680] Replace pointers by values or unique_ptr for better
 leak-safety

---
 .../src/ThreadPool/NonBlockingThreadPool.h    | 40 +++++++++----------
 1 file changed, 19 insertions(+), 21 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index ecd49f382..a93e22a76 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -58,11 +58,9 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
         coprimes_.push_back(i);
       }
     }
+    queues_.resize(num_threads_);
     for (int i = 0; i < num_threads_; i++) {
-      queues_.push_back(new Queue());
-    }
-    for (int i = 0; i < num_threads_; i++) {
-      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
+      threads_.emplace_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
     }
   }
 
@@ -78,13 +76,12 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       // Since we were cancelled, there might be entries in the queues.
       // Empty them to prevent their destructor from asserting.
       for (size_t i = 0; i < queues_.size(); i++) {
-        queues_[i]->Flush();
+        queues_[i].Flush();
       }
     }
 
     // Join threads explicitly to avoid destruction order issues.
-    for (size_t i = 0; i < num_threads_; i++) delete threads_[i];
-    for (size_t i = 0; i < num_threads_; i++) delete queues_[i];
+    threads_.resize(0);
   }
 
   void Schedule(std::function<void()> fn) {
@@ -92,13 +89,13 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     PerThread* pt = GetPerThread();
     if (pt->pool == this) {
       // Worker thread of this pool, push onto the thread's queue.
-      Queue* q = queues_[pt->thread_id];
-      t = q->PushFront(std::move(t));
+      Queue& q = queues_[pt->thread_id];
+      t = q.PushFront(std::move(t));
     } else {
       // A free-standing thread (or worker of another pool), push onto a random
       // queue.
-      Queue* q = queues_[Rand(&pt->rand) % queues_.size()];
-      t = q->PushBack(std::move(t));
+      Queue& q = queues_[Rand(&pt->rand) % queues_.size()];
+      t = q.PushBack(std::move(t));
     }
     // Note: below we touch this after making w available to worker threads.
     // Strictly speaking, this can lead to a racy-use-after-free. Consider that
@@ -157,8 +154,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   Environment env_;
   const int num_threads_;
   const bool allow_spinning_;
-  MaxSizeVector<Thread*> threads_;
-  MaxSizeVector<Queue*> queues_;
+  MaxSizeVector<std::unique_ptr<Thread> > threads_;
+  MaxSizeVector<Queue> queues_;
   MaxSizeVector<unsigned> coprimes_;
   MaxSizeVector<EventCount::Waiter> waiters_;
   std::atomic<unsigned> blocked_;
@@ -173,7 +170,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     pt->pool = this;
     pt->rand = std::hash<std::thread::id>()(std::this_thread::get_id());
     pt->thread_id = thread_id;
-    Queue* q = queues_[thread_id];
+    Queue& q = queues_[thread_id];
     EventCount::Waiter* waiter = &waiters_[thread_id];
     // TODO(dvyukov,rmlarsen): The time spent in Steal() is proportional
     // to num_threads_ and we assume that new work is scheduled at a
@@ -189,10 +186,10 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       // counter-productive for the types of I/O workloads the single thread
       // pools tend to be used for.
       while (!cancelled_) {
-        Task t = q->PopFront();
+        Task t = q.PopFront();
         for (int i = 0; i < spin_count && !t.f; i++) {
           if (!cancelled_.load(std::memory_order_relaxed)) {
-            t = q->PopFront();
+            t = q.PopFront();
           }
         }
         if (!t.f) {
@@ -206,7 +203,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       }
     } else {
       while (!cancelled_) {
-        Task t = q->PopFront();
+        Task t = q.PopFront();
         if (!t.f) {
           t = Steal();
           if (!t.f) {
@@ -243,7 +240,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     unsigned inc = coprimes_[r % coprimes_.size()];
     unsigned victim = r % size;
     for (unsigned i = 0; i < size; i++) {
-      Task t = queues_[victim]->PopBack();
+      Task t = queues_[victim].PopBack();
       if (t.f) {
         return t;
       }
@@ -270,7 +267,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       if (cancelled_) {
         return false;
       } else {
-        *t = queues_[victim]->PopBack();
+        *t = queues_[victim].PopBack();
         return true;
       }
     }
@@ -278,7 +275,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // If we are shutting down and all worker threads blocked without work,
     // that's we are done.
     blocked_++;
-    if (done_ && blocked_ == num_threads_) {
+    // TODO is blocked_ required to be unsigned?
+    if (done_ && blocked_ == static_cast<unsigned>(num_threads_)) {
       ec_.CancelWait(waiter);
       // Almost done, but need to re-check queues.
       // Consider that all queues are empty and all worker threads are preempted
@@ -311,7 +309,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     unsigned inc = coprimes_[r % coprimes_.size()];
     unsigned victim = r % size;
     for (unsigned i = 0; i < size; i++) {
-      if (!queues_[victim]->Empty()) {
+      if (!queues_[victim].Empty()) {
         return victim;
       }
       victim += inc;

From e51d9e473aa1f882d3b3106ec2427a44d2a76ceb Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 11:42:05 -0700
Subject: [PATCH 589/680] Protect #undef max with #ifdef max.

---
 unsupported/Eigen/CXX11/ThreadPool | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/unsupported/Eigen/CXX11/ThreadPool b/unsupported/Eigen/CXX11/ThreadPool
index 12aa07c7f..64ea83b7e 100644
--- a/unsupported/Eigen/CXX11/ThreadPool
+++ b/unsupported/Eigen/CXX11/ThreadPool
@@ -49,7 +49,9 @@
 // which trigger a check in test/main.h causing compilation to fail.
 // We work around the check here by removing the check for max in
 // the case where we have to emulate thread_local.
+#ifdef max
 #undef max
+#endif
 #include <unordered_map>
 #endif
 

From 6e0464004a7bcd666d3b5962c3c999ff78f416f1 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 12:10:08 -0700
Subject: [PATCH 590/680] Store std::unique_ptr instead of raw pointers in
 per_thread_map_.

---
 .../CXX11/src/ThreadPool/NonBlockingThreadPool.h | 16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index a800e827f..1ac4de3b5 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -58,9 +58,6 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       }
     }
     queues_.resize(num_threads_);
-    for (int i = 0; i < num_threads_; i++) {
-      queues_.push_back(new Queue());
-    }
 #ifndef EIGEN_THREAD_LOCAL
     init_barrier_.reset(new Barrier(num_threads_));
 #endif
@@ -93,9 +90,6 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     // Join threads explicitly to avoid destruction order issues.
     threads_.resize(0);
     queues_.resize(0);
-#ifndef EIGEN_THREAD_LOCAL
-    for (auto it : per_thread_map_) delete it.second;
-#endif
   }
 
   void Schedule(std::function<void()> fn) {
@@ -176,21 +170,19 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
 #ifndef EIGEN_THREAD_LOCAL
   std::unique_ptr<Barrier> init_barrier_;
   std::mutex mu;  // Protects per_thread_map_.
-  std::unordered_map<uint64_t, PerThread*> per_thread_map_;
+  std::unordered_map<uint64_t, std::unique_ptr<PerThread>> per_thread_map_;
 #endif
 
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
 #ifndef EIGEN_THREAD_LOCAL
-    PerThread* pt = new PerThread();
     mu.lock();
-    per_thread_map_[GlobalThreadIdHash()] = pt;
+    eigen_assert(per_thread_map_.emplace(GlobalThreadIdHash(), new PerThread()).second);
     mu.unlock();
     init_barrier_->Notify();
     init_barrier_->Wait();
-#else
-    PerThread* pt = GetPerThread();
 #endif
+    PerThread* pt = GetPerThread();
     pt->pool = this;
     pt->rand = GlobalThreadIdHash();
     pt->thread_id = thread_id;
@@ -355,7 +347,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     if (it == per_thread_map_.end()) {
       return &dummy;
     } else {
-      return it->second;
+      return it->second.get();
     }
 #else
     EIGEN_THREAD_LOCAL PerThread per_thread_;

From 6cedc5a9b38d6ddda69d532b28dff9ee5c2d1c04 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 12:11:58 -0700
Subject: [PATCH 591/680] rename mu.

---
 .../Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h      | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 1ac4de3b5..d710faa94 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -169,16 +169,16 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   EventCount ec_;
 #ifndef EIGEN_THREAD_LOCAL
   std::unique_ptr<Barrier> init_barrier_;
-  std::mutex mu;  // Protects per_thread_map_.
+  std::mutex per_thread_map_mutex_;  // Protects per_thread_map_.
   std::unordered_map<uint64_t, std::unique_ptr<PerThread>> per_thread_map_;
 #endif
 
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
 #ifndef EIGEN_THREAD_LOCAL
-    mu.lock();
+    per_thread_map_mutex_.lock();
     eigen_assert(per_thread_map_.emplace(GlobalThreadIdHash(), new PerThread()).second);
-    mu.unlock();
+    per_thread_map_mutex_.unlock();
     init_barrier_->Notify();
     init_barrier_->Wait();
 #endif

From 668690978ff66151b2a495767c7daf33d06be4a5 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 12:54:33 -0700
Subject: [PATCH 592/680] Pad PerThread when we emulate thread_local to prevent
 false sharing.

---
 .../Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h        | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index d710faa94..1cb63bcfa 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -153,6 +153,10 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     ThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
     uint64_t rand;          // Random generator state.
     int thread_id;          // Worker thread index in pool.
+#ifndef EIGEN_THREAD_LOCAL
+    // Prevent false sharing.
+    char pad_[128];
+#endif
   };
 
   Environment env_;

From e9f9d70611d0a66751c34b1430ed7649aff6e2bf Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 12:59:46 -0700
Subject: [PATCH 593/680] Don't rely on __had_feature for g++. Don't use
 __thread. Only use thread_local for gcc 4.8 or newer.

---
 unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
index f33759ba9..89ed6e5e5 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
@@ -12,10 +12,10 @@
 
 #undef EIGEN_THREAD_LOCAL
 
-#if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_thread_local))
-  #define EIGEN_THREAD_LOCAL static thread_local
-#elif (EIGEN_COMP_GNUC && EIGEN_GNUC_AT_MOST(4, 7)) || EIGEN_COMP_CLANG
-  #define EIGEN_THREAD_LOCAL static __thread
+#if EIGEN_MAX_CPP_VER >= 11 &&                         \
+    ((EIGEN_COMP_GNUC && EIGEN_GNUC_AT_LEAST(4, 8)) || \
+     __has_feature(cxx_thread_local))
+#define EIGEN_THREAD_LOCAL static thread_local
 #endif
 
 // Disable TLS for Apple and Android builds with older toolchains.

From 8d9bc5cc022bee4a06201c7c5a1dec2b73697f5f Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 23 Aug 2018 13:06:39 -0700
Subject: [PATCH 594/680] Fix g++ compilation.

---
 unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 1cb63bcfa..60a0c9fb6 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -180,8 +180,9 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   // Main worker thread loop.
   void WorkerLoop(int thread_id) {
 #ifndef EIGEN_THREAD_LOCAL
+    std::unique_ptr<PerThread> new_pt(new PerThread());
     per_thread_map_mutex_.lock();
-    eigen_assert(per_thread_map_.emplace(GlobalThreadIdHash(), new PerThread()).second);
+    eigen_assert(per_thread_map_.emplace(GlobalThreadIdHash(), std::move(new_pt)).second);
     per_thread_map_mutex_.unlock();
     init_barrier_->Notify();
     init_barrier_->Wait();

From ad4a08fb68ecd3afde93e4714687e96164db704b Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 24 Aug 2018 19:04:33 +0200
Subject: [PATCH 595/680] Use Intel cast intrinsics, since MSVC does not allow
 direct casting. Reported by David Winkler.

---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 4e2e916de..9e66575a9 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -634,13 +634,13 @@ template<> EIGEN_STRONG_INLINE Packet8d preverse(const Packet8d& a)
 template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a)
 {
   // _mm512_abs_ps intrinsic not found, so hack around it
-  return (__m512)_mm512_and_si512((__m512i)a, _mm512_set1_epi32(0x7fffffff));
+  return _mm512_castsi512_ps(_mm512_and_si512(_mm512_castps_si512(a), _mm512_set1_epi32(0x7fffffff)));
 }
 template <>
 EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
   // _mm512_abs_ps intrinsic not found, so hack around it
-  return (__m512d)_mm512_and_si512((__m512i)a,
-                                   _mm512_set1_epi64(0x7fffffffffffffff));
+  return _mm512_castsi512_pd(_mm512_and_si512(_mm512_castpd_si512(a),
+                                   _mm512_set1_epi64(0x7fffffffffffffff)));
 }
 
 #ifdef EIGEN_VECTORIZE_AVX512DQ

From 744e2fe0dedb697a8802a3e633e37a4f844da372 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 24 Aug 2018 10:24:54 -0700
Subject: [PATCH 596/680] Address comments about EIGEN_THREAD_LOCAL.

---
 unsupported/Eigen/CXX11/ThreadPool                   | 9 ++++-----
 unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h | 2 --
 2 files changed, 4 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/ThreadPool b/unsupported/Eigen/CXX11/ThreadPool
index 64ea83b7e..1dcc4eb6c 100644
--- a/unsupported/Eigen/CXX11/ThreadPool
+++ b/unsupported/Eigen/CXX11/ThreadPool
@@ -44,6 +44,10 @@
 #include <thread>
 #include <functional>
 #include <memory>
+#include "src/util/CXX11Meta.h"
+#include "src/util/MaxSizeVector.h"
+
+#include "src/ThreadPool/ThreadLocal.h"
 #ifndef EIGEN_THREAD_LOCAL
 // There are non-parenthesized calls to "max" in the  <unordered_map> header,
 // which trigger a check in test/main.h causing compilation to fail.
@@ -54,11 +58,6 @@
 #endif
 #include <unordered_map>
 #endif
-
-#include "src/util/CXX11Meta.h"
-#include "src/util/MaxSizeVector.h"
-
-#include "src/ThreadPool/ThreadLocal.h"
 #include "src/ThreadPool/ThreadYield.h"
 #include "src/ThreadPool/ThreadCancel.h"
 #include "src/ThreadPool/EventCount.h"
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
index 89ed6e5e5..a41731c34 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
@@ -10,8 +10,6 @@
 #ifndef EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
 #define EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
 
-#undef EIGEN_THREAD_LOCAL
-
 #if EIGEN_MAX_CPP_VER >= 11 &&                         \
     ((EIGEN_COMP_GNUC && EIGEN_GNUC_AT_LEAST(4, 8)) || \
      __has_feature(cxx_thread_local))

From f7675b826b2f0a33b09b97342e855a8ef059927e Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 24 Aug 2018 22:58:55 +0200
Subject: [PATCH 597/680] Fix several integer conversion and sign-compare
 warnings

---
 .../Eigen/CXX11/src/Tensor/TensorBlock.h      | 27 ++++---
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 72 +++++++++----------
 .../Eigen/CXX11/src/Tensor/TensorExecutor.h   |  4 +-
 .../Eigen/CXX11/src/Tensor/TensorIndexList.h  |  2 +-
 .../test/cxx11_tensor_block_access.cpp        | 54 +++++++-------
 unsupported/test/cxx11_tensor_thread_pool.cpp |  2 +-
 6 files changed, 80 insertions(+), 81 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 24a6343e8..aa500eb70 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -212,11 +212,11 @@ class TensorBlockIO {
         num_size_one_inner_dims, NumDims - num_size_one_inner_dims - 1);
     const StorageIndex block_dim_for_tensor_stride1_dim =
         NumDims == 0 ? 1 : tensor_to_block_dim_map[tensor_stride1_dim];
-    size_t block_inner_dim_size =
+    StorageIndex block_inner_dim_size =
         NumDims == 0 ? 1
                      : block.block_sizes()[block_dim_for_tensor_stride1_dim];
-    for (int i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
-      const int dim = cond<Layout>()(i, NumDims - i - 1);
+    for (Index i = num_size_one_inner_dims + 1; i < NumDims; ++i) {
+      const Index dim = cond<Layout>()(i, NumDims - i - 1);
       const StorageIndex block_stride =
           block.block_strides()[tensor_to_block_dim_map[dim]];
       if (block_inner_dim_size == block_stride &&
@@ -258,8 +258,8 @@ class TensorBlockIO {
 
     // Initialize block iterator state. Squeeze away any dimension of size 1.
     int num_squeezed_dims = 0;
-    for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
-      const int dim = cond<Layout>()(i + 1, NumDims - i - 2);
+    for (Index i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
+      const Index dim = cond<Layout>()(i + 1, NumDims - i - 2);
       const StorageIndex size = block.block_sizes()[tensor_to_block_dim_map[dim]];
       if (size == 1) {
         continue;
@@ -626,7 +626,7 @@ class TensorBlockMapper {
                     const TensorBlockShapeType block_shape,
                     Index min_target_size)
       : m_dimensions(dims),
-        m_block_dim_sizes(BlockDimensions(dims, block_shape, min_target_size)) {
+        m_block_dim_sizes(BlockDimensions(dims, block_shape, internal::convert_index<StorageIndex>(min_target_size))) {
     // Calculate block counts by dimension and total block count.
     DSizes<StorageIndex, NumDims> block_count;
     for (Index i = 0; i < block_count.rank(); ++i) {
@@ -717,8 +717,8 @@ class TensorBlockMapper {
  private:
   static Dimensions BlockDimensions(const Dimensions& tensor_dims,
                                     const TensorBlockShapeType block_shape,
-                                    Index min_target_size) {
-    min_target_size = numext::maxi<Index>(1, min_target_size);
+                                    StorageIndex min_target_size) {
+    min_target_size = numext::maxi<StorageIndex>(1, min_target_size);
 
     // If tensor fully fits into the target size, we'll treat it a single block.
     Dimensions block_dim_sizes = tensor_dims;
@@ -735,16 +735,15 @@ class TensorBlockMapper {
       if (block_shape == kUniformAllDims) {
         // Tensor will not fit within 'min_target_size' budget: calculate tensor
         // block dimension sizes based on "square" dimension size target.
-        const size_t dim_size_target = static_cast<const size_t>(
+        const StorageIndex dim_size_target = internal::convert_index<StorageIndex>(
             std::pow(static_cast<float>(min_target_size),
                      1.0f / static_cast<float>(block_dim_sizes.rank())));
-        for (size_t i = 0; i < block_dim_sizes.rank(); ++i) {
+        for (Index i = 0; i < block_dim_sizes.rank(); ++i) {
           // TODO(andydavis) Adjust the inner most 'block_dim_size' to make it
           // a multiple of the packet size. Note that reducing
           // 'block_dim_size' in this manner can increase the number of
           // blocks, and so will amplify any per-block overhead.
-          block_dim_sizes[i] = numext::mini(
-              dim_size_target, static_cast<size_t>(tensor_dims[i]));
+          block_dim_sizes[i] = numext::mini(dim_size_target, tensor_dims[i]);
         }
         // Add any un-allocated coefficients to inner dimension(s).
         StorageIndex total_size = block_dim_sizes.TotalSize();
@@ -781,7 +780,7 @@ class TensorBlockMapper {
 
     eigen_assert(
         block_dim_sizes.TotalSize() >=
-        numext::mini<size_t>(min_target_size, tensor_dims.TotalSize()));
+        numext::mini<Index>(min_target_size, tensor_dims.TotalSize()));
 
     return block_dim_sizes;
   }
@@ -824,7 +823,7 @@ class TensorSliceBlockMapper {
         m_total_block_count(1) {
     // Calculate block counts by dimension and total block count.
     DSizes<StorageIndex, NumDims> block_count;
-    for (size_t i = 0; i < block_count.rank(); ++i) {
+    for (Index i = 0; i < block_count.rank(); ++i) {
       block_count[i] = divup(m_tensor_slice_extents[i], m_block_dim_sizes[i]);
     }
     m_total_block_count = array_prod(block_count);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 4f973a5b7..ce91bc2a6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -32,12 +32,12 @@ namespace Eigen {
 // Boilerplate code
 namespace internal {
 
-template<std::size_t n, typename Dimension> struct dget {
+template<std::ptrdiff_t n, typename Dimension> struct dget {
   static const std::ptrdiff_t value = get<n, Dimension>::value;
 };
 
 
-template<typename Index, std::size_t NumIndices, std::size_t n, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
 struct fixed_size_tensor_index_linearization_helper
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
@@ -50,7 +50,7 @@ struct fixed_size_tensor_index_linearization_helper
   }
 };
 
-template<typename Index, std::size_t NumIndices, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
 struct fixed_size_tensor_index_linearization_helper<Index, NumIndices, 0, RowMajor>
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
@@ -60,7 +60,7 @@ struct fixed_size_tensor_index_linearization_helper<Index, NumIndices, 0, RowMaj
   }
 };
 
-template<typename Index, std::size_t n>
+template<typename Index, std::ptrdiff_t n>
 struct fixed_size_tensor_index_extraction_helper
 {
   template <typename Dimensions> EIGEN_DEVICE_FUNC
@@ -94,7 +94,7 @@ struct Sizes {
   typedef internal::numeric_list<std::ptrdiff_t, Indices...> Base;
   const Base t = Base();
   static const std::ptrdiff_t total_size = internal::arg_prod(Indices...);
-  static const size_t count = Base::count;
+  static const ptrdiff_t count = Base::count;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t rank() const {
     return Base::count;
@@ -121,16 +121,16 @@ struct Sizes {
     return *this;
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t operator[] (const std::size_t index) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t operator[] (const std::ptrdiff_t index) const {
     return internal::fixed_size_tensor_index_extraction_helper<std::ptrdiff_t, Base::count>::run(index, t);
   }
 
   template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  size_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
+  ptrdiff_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
     return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, false>::run(indices, t);
   }
   template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  size_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
+  ptrdiff_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
     return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, true>::run(indices, t);
   }
 };
@@ -144,25 +144,25 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_prod(const Sizes<Indi
 
 #else
 
-template <std::size_t n>
+template <std::ptrdiff_t n>
 struct non_zero_size {
-  typedef internal::type2val<std::size_t, n> type;
+  typedef internal::type2val<std::ptrdiff_t, n> type;
 };
 template <>
 struct non_zero_size<0> {
   typedef internal::null_type type;
 };
 
-template <std::size_t V1=0, std::size_t V2=0, std::size_t V3=0, std::size_t V4=0, std::size_t V5=0> struct Sizes {
+template <std::ptrdiff_t V1=0, std::ptrdiff_t V2=0, std::ptrdiff_t V3=0, std::ptrdiff_t V4=0, std::ptrdiff_t V5=0> struct Sizes {
   typedef typename internal::make_type_list<typename non_zero_size<V1>::type, typename non_zero_size<V2>::type, typename non_zero_size<V3>::type, typename non_zero_size<V4>::type, typename non_zero_size<V5>::type >::type Base;
-  static const size_t count = Base::count;
-  static const std::size_t total_size = internal::arg_prod<Base>::value;
+  static const std::ptrdiff_t count = Base::count;
+  static const std::ptrdiff_t total_size = internal::arg_prod<Base>::value;
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t rank() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ptrdiff_t rank() const {
     return count;
   }
 
-  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ptrdiff_t TotalSize() {
     return internal::arg_prod<Base>::value;
   }
 
@@ -178,7 +178,7 @@ template <std::size_t V1=0, std::size_t V2=0, std::size_t V3=0, std::size_t V4=0
 
 #if EIGEN_HAS_VARIADIC_TEMPLATES
   template <typename... DenseIndex> Sizes(DenseIndex... /*indices*/) { }
-  explicit Sizes(std::initializer_list<std::size_t>) {
+  explicit Sizes(std::initializer_list<std::ptrdiff_t>) {
     // todo: add assertion
   }
 #else
@@ -213,18 +213,18 @@ template <std::size_t V1=0, std::size_t V2=0, std::size_t V3=0, std::size_t V4=0
   }
 
   template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  size_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
+  ptrdiff_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
     return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, false>::run(indices, *reinterpret_cast<const Base*>(this));
   }
   template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  size_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
+  ptrdiff_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
     return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, true>::run(indices, *reinterpret_cast<const Base*>(this));
   }
 };
 
 namespace internal {
-template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5>
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_prod(const Sizes<V1, V2, V3, V4, V5>&) {
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_prod(const Sizes<V1, V2, V3, V4, V5>&) {
   return Sizes<V1, V2, V3, V4, V5>::total_size;
 }
 }
@@ -233,7 +233,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_prod(const Sizes<V1, V2,
 
 // Boilerplate
 namespace internal {
-template<typename Index, std::size_t NumIndices, std::size_t n, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
 struct tensor_index_linearization_helper
 {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -245,7 +245,7 @@ struct tensor_index_linearization_helper
   }
 };
 
-template<typename Index, std::size_t NumIndices, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
 struct tensor_index_linearization_helper<Index, NumIndices, 0, RowMajor>
 {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -264,7 +264,7 @@ struct DSizes : array<DenseIndex, NumDims> {
   typedef array<DenseIndex, NumDims> Base;
   static const int count = NumDims;
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t rank() const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rank() const {
     return NumDims;
   }
 
@@ -298,7 +298,7 @@ struct DSizes : array<DenseIndex, NumDims> {
     }
   }
 #else
-  template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5>
+  template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5>
   EIGEN_DEVICE_FUNC DSizes(const Sizes<V1, V2, V3, V4, V5>& a) {
     for (int i = 0 ; i < NumDims; ++i) {
       (*this)[i] = a[i];
@@ -359,7 +359,7 @@ struct DSizes : array<DenseIndex, NumDims> {
 
 // Boilerplate
 namespace internal {
-template<typename Index, std::size_t NumIndices, std::size_t n, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
 struct tensor_vsize_index_linearization_helper
 {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -371,7 +371,7 @@ struct tensor_vsize_index_linearization_helper
   }
 };
 
-template<typename Index, std::size_t NumIndices, bool RowMajor>
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
 struct tensor_vsize_index_linearization_helper<Index, NumIndices, 0, RowMajor>
 {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@@ -386,10 +386,10 @@ struct tensor_vsize_index_linearization_helper<Index, NumIndices, 0, RowMajor>
 namespace internal {
 
 template <typename DenseIndex, int NumDims> struct array_size<const DSizes<DenseIndex, NumDims> > {
-  static const size_t value = NumDims;
+  static const ptrdiff_t value = NumDims;
 };
 template <typename DenseIndex, int NumDims> struct array_size<DSizes<DenseIndex, NumDims> > {
-  static const size_t value = NumDims;
+  static const ptrdiff_t value = NumDims;
 };
 #ifndef EIGEN_EMULATE_CXX11_META_H
 template <typename std::ptrdiff_t... Indices> struct array_size<const Sizes<Indices...> > {
@@ -399,33 +399,33 @@ template <typename std::ptrdiff_t... Indices> struct array_size<Sizes<Indices...
 static const std::ptrdiff_t value = Sizes<Indices...>::count;
 };
 template <std::ptrdiff_t n, typename std::ptrdiff_t... Indices> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<Indices...>&) {
-  return get<n, internal::numeric_list<std::size_t, Indices...> >::value;
+  return get<n, internal::numeric_list<std::ptrdiff_t, Indices...> >::value;
 }
 template <std::ptrdiff_t n> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<>&) {
   eigen_assert(false && "should never be called");
   return -1;
 }
 #else
-template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5> struct array_size<const Sizes<V1,V2,V3,V4,V5> > {
-  static const size_t value = Sizes<V1,V2,V3,V4,V5>::count;
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> struct array_size<const Sizes<V1,V2,V3,V4,V5> > {
+  static const ptrdiff_t value = Sizes<V1,V2,V3,V4,V5>::count;
 };
-template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5> struct array_size<Sizes<V1,V2,V3,V4,V5> > {
-  static const size_t value = Sizes<V1,V2,V3,V4,V5>::count;
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> struct array_size<Sizes<V1,V2,V3,V4,V5> > {
+  static const ptrdiff_t value = Sizes<V1,V2,V3,V4,V5>::count;
 };
-template <std::size_t n, std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes<V1,V2,V3,V4,V5>&) {
+template <std::ptrdiff_t n, std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<V1,V2,V3,V4,V5>&) {
   return get<n, typename Sizes<V1,V2,V3,V4,V5>::Base>::value;
 }
 
 #endif
 
 
-template <typename Dims1, typename Dims2, size_t n, size_t m>
+template <typename Dims1, typename Dims2, ptrdiff_t n, ptrdiff_t m>
 struct sizes_match_below_dim {
   static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1&, Dims2&) {
     return false;
   }
 };
-template <typename Dims1, typename Dims2, size_t n>
+template <typename Dims1, typename Dims2, ptrdiff_t n>
 struct sizes_match_below_dim<Dims1, Dims2, n, n> {
   static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1& dims1, Dims2& dims2) {
     return (array_get<n-1>(dims1) == array_get<n-1>(dims2)) &
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 9b9587de5..b756be3b3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -256,8 +256,8 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
     typedef TensorBlockMapper<ScalarNoConst, StorageIndex, NumDims, Evaluator::Layout> TensorBlockMapper;
 
     Evaluator evaluator(expr, device);
-    StorageIndex total_size = array_prod(evaluator.dimensions());
-    StorageIndex cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
+    Index total_size = array_prod(evaluator.dimensions());
+    Index cache_size = device.firstLevelCacheSize() / sizeof(Scalar);
     if (total_size < cache_size) {
       // TODO(andydavis) Reduce block management overhead for small tensors.
       internal::TensorExecutor<Expression, ThreadPoolDevice, Vectorizable,
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index 98ad661ca..3f7d26b18 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -84,7 +84,7 @@ template<DenseIndex n> struct NumTraits<type2index<n> >
 namespace internal {
 template <typename T>
 EIGEN_DEVICE_FUNC void update_value(T& val, DenseIndex new_val) {
-  val = new_val;
+  val = internal::convert_index<T>(new_val);
 }
 template <DenseIndex n>
 EIGEN_DEVICE_FUNC void update_value(type2index<n>& val, DenseIndex new_val) {
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index f572e496d..24a95ab12 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -33,8 +33,8 @@ static internal::TensorBlockShapeType RandomShape() {
 }
 
 template <int NumDims>
-static std::size_t RandomTargetSize(const DSizes<Index, NumDims>& dims) {
-  return internal::random<int>(1, dims.TotalSize());
+static Index RandomTargetSize(const DSizes<Index, NumDims>& dims) {
+  return internal::random<Index>(1, dims.TotalSize());
 }
 
 template <int NumDims>
@@ -178,7 +178,7 @@ static void test_block_mapper_maps_every_element() {
   // Verify that every coefficient in the original Tensor is accessible through
   // TensorBlock only once.
   Index total_coeffs = dims.TotalSize();
-  VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
+  VERIFY_IS_EQUAL(Index(coeff_set.size()), total_coeffs);
   VERIFY_IS_EQUAL(*coeff_set.begin(), 0);
   VERIFY_IS_EQUAL(*coeff_set.rbegin(), total_coeffs - 1);
 }
@@ -208,7 +208,7 @@ static void test_slice_block_mapper_maps_every_element() {
   // Pick a random dimension sizes for the tensor blocks.
   DSizes<Index, NumDims> block_sizes;
   for (int i = 0; i < NumDims; ++i) {
-    block_sizes[i] = internal::random<int>(1, tensor_slice_extents[i]);
+    block_sizes[i] = internal::random<Index>(1, tensor_slice_extents[i]);
   }
 
   TensorSliceBlockMapper block_mapper(tensor_dims, tensor_slice_offsets,
@@ -222,7 +222,7 @@ static void test_slice_block_mapper_maps_every_element() {
                                        &coeff_set);
   }
 
-  VERIFY_IS_EQUAL(coeff_set.size(), total_coeffs);
+  VERIFY_IS_EQUAL(Index(coeff_set.size()), total_coeffs);
 }
 
 template <typename T, int NumDims, int Layout>
@@ -262,14 +262,14 @@ static void test_block_io_copy_data_from_source_to_target() {
 }
 
 template <int Layout, int NumDims>
-static int GetInputIndex(Index output_index,
+static Index GetInputIndex(Index output_index,
                          const array<Index, NumDims>& output_to_input_dim_map,
                          const array<Index, NumDims>& input_strides,
                          const array<Index, NumDims>& output_strides) {
   int input_index = 0;
   if (Layout == ColMajor) {
     for (int i = NumDims - 1; i > 0; --i) {
-      const int idx = output_index / output_strides[i];
+      const Index idx = output_index / output_strides[i];
       input_index += idx * input_strides[output_to_input_dim_map[i]];
       output_index -= idx * output_strides[i];
     }
@@ -277,7 +277,7 @@ static int GetInputIndex(Index output_index,
            output_index * input_strides[output_to_input_dim_map[0]];
   } else {
     for (int i = 0; i < NumDims - 1; ++i) {
-      const int idx = output_index / output_strides[i];
+      const Index idx = output_index / output_strides[i];
       input_index += idx * input_strides[output_to_input_dim_map[i]];
       output_index -= idx * output_strides[i];
     }
@@ -650,7 +650,7 @@ static void test_uniform_block_shape()
   {
     // Test shape 'UniformAllDims' with uniform 'max_coeff count'.
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 5;
+    const Index max_coeff_count = 5 * 5 * 5 * 5 * 5;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -664,7 +664,7 @@ static void test_uniform_block_shape()
   // partially into first inner-most dimension.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 7 * 5 * 5 * 5 * 5;
+    const Index max_coeff_count = 7 * 5 * 5 * 5 * 5;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -675,7 +675,7 @@ static void test_uniform_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 6;
+    const Index max_coeff_count = 5 * 5 * 5 * 5 * 6;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -690,7 +690,7 @@ static void test_uniform_block_shape()
   // fully into first inner-most dimension.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 5 * 5 * 5 * 5;
+    const Index max_coeff_count = 11 * 5 * 5 * 5 * 5;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -701,7 +701,7 @@ static void test_uniform_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 5 * 5 * 5 * 5 * 7;
+    const Index max_coeff_count = 5 * 5 * 5 * 5 * 7;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -716,7 +716,7 @@ static void test_uniform_block_shape()
   // fully into first few inner-most dimensions.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(7, 5, 6, 17, 7);
-    const size_t max_coeff_count = 7 * 5 * 6 * 7 * 5;
+    const Index max_coeff_count = 7 * 5 * 6 * 7 * 5;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -728,7 +728,7 @@ static void test_uniform_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(7, 5, 6, 9, 7);
-    const size_t max_coeff_count = 5 * 5 * 5 * 6 * 7;
+    const Index max_coeff_count = 5 * 5 * 5 * 6 * 7;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -743,7 +743,7 @@ static void test_uniform_block_shape()
   // Test shape 'UniformAllDims' with full allocation to all dims.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(7, 5, 6, 17, 7);
-    const size_t max_coeff_count = 7 * 5 * 6 * 17 * 7;
+    const Index max_coeff_count = 7 * 5 * 6 * 17 * 7;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -755,7 +755,7 @@ static void test_uniform_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(7, 5, 6, 9, 7);
-    const size_t max_coeff_count = 7 * 5 * 6 * 9 * 7;
+    const Index max_coeff_count = 7 * 5 * 6 * 9 * 7;
     TensorBlockMapper block_mapper(dims, internal::kUniformAllDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -777,7 +777,7 @@ static void test_skewed_inner_dim_block_shape()
   // Test shape 'SkewedInnerDims' with partial allocation to inner-most dim.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 10 * 1 * 1 * 1 * 1;
+    const Index max_coeff_count = 10 * 1 * 1 * 1 * 1;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -788,7 +788,7 @@ static void test_skewed_inner_dim_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 1 * 1 * 1 * 1 * 6;
+    const Index max_coeff_count = 1 * 1 * 1 * 1 * 6;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -802,7 +802,7 @@ static void test_skewed_inner_dim_block_shape()
   // Test shape 'SkewedInnerDims' with full allocation to inner-most dim.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 1 * 1 * 1 * 1;
+    const Index max_coeff_count = 11 * 1 * 1 * 1 * 1;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -813,7 +813,7 @@ static void test_skewed_inner_dim_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 1 * 1 * 1 * 1 * 7;
+    const Index max_coeff_count = 1 * 1 * 1 * 1 * 7;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -828,7 +828,7 @@ static void test_skewed_inner_dim_block_shape()
   // and partial allocation to second inner-dim.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 3 * 1 * 1 * 1;
+    const Index max_coeff_count = 11 * 3 * 1 * 1 * 1;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -840,7 +840,7 @@ static void test_skewed_inner_dim_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 1 * 1 * 1 * 15 * 7;
+    const Index max_coeff_count = 1 * 1 * 1 * 15 * 7;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -856,7 +856,7 @@ static void test_skewed_inner_dim_block_shape()
   // and partial allocation to third inner-dim.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 5 * 5 * 1 * 1;
+    const Index max_coeff_count = 11 * 5 * 5 * 1 * 1;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -869,7 +869,7 @@ static void test_skewed_inner_dim_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 1 * 1 * 5 * 17 * 7;
+    const Index max_coeff_count = 1 * 1 * 5 * 17 * 7;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -885,7 +885,7 @@ static void test_skewed_inner_dim_block_shape()
   // Test shape 'SkewedInnerDims' with full allocation to all dims.
   if (Layout == ColMajor) {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
+    const Index max_coeff_count = 11 * 5 * 6 * 17 * 7;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
@@ -897,7 +897,7 @@ static void test_skewed_inner_dim_block_shape()
     VERIFY(block.block_sizes().TotalSize() <= max_coeff_count);
   } else {
     DSizes<Index, 5> dims(11, 5, 6, 17, 7);
-    const size_t max_coeff_count = 11 * 5 * 6 * 17 * 7;
+    const Index max_coeff_count = 11 * 5 * 6 * 17 * 7;
     TensorBlockMapper block_mapper(dims, internal::kSkewedInnerDims,
                                    max_coeff_count);
     TensorBlock block = block_mapper.GetBlockForIndex(0, NULL);
diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp
index 7606b0abf..6d8e58214 100644
--- a/unsupported/test/cxx11_tensor_thread_pool.cpp
+++ b/unsupported/test/cxx11_tensor_thread_pool.cpp
@@ -300,7 +300,7 @@ static void test_multithread_contraction_with_output_kernel() {
 
   m_result = m_left * m_right;
 
-  for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
+  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
     VERIFY(&t_result.data()[i] != &m_result.data()[i]);
     VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
   }

From 8295f02b36bc62584b71550cd48a5c56b173ee4c Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 24 Aug 2018 23:22:20 +0200
Subject: [PATCH 598/680] Hide "maybe uninitialized" warning on gcc

---
 unsupported/test/cxx11_tensor_block_access.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 24a95ab12..74ce2cc80 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -138,8 +138,8 @@ template <typename T, int Layout, int NumDims>
 static void UpdateCoeffSet(
     const internal::TensorBlock<T, Index, NumDims, Layout>& block,
     Index first_coeff_index, int dim_index, std::set<Index>* visited_coeffs) {
-  const DSizes<Index, NumDims> block_sizes = block.block_sizes();
-  const DSizes<Index, NumDims> tensor_strides = block.tensor_strides();
+  const DSizes<Index, NumDims>& block_sizes = block.block_sizes();
+  const DSizes<Index, NumDims>& tensor_strides = block.tensor_strides();
 
   for (int i = 0; i < block_sizes[dim_index]; ++i) {
     if (tensor_strides[dim_index] == 1) {

From 5aaedbecedb6702c5102e13a54d5908b8964174b Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 24 Aug 2018 23:54:12 +0200
Subject: [PATCH 599/680] Fixed more sign-compare and type-limits warnings

---
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +-
 unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h    | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 2ca91709f..4b24e5fc1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -347,7 +347,7 @@ struct TensorContractionEvaluatorBase
     // dimensions and right non-contracting dimensions.
     m_lhs_inner_dim_contiguous = true;
     int dim_idx = 0;
-    unsigned int nocontract_idx = 0;
+    Index nocontract_idx = 0;
 
     for (int i = 0; i < LDims; i++) {
       // find if we are contracting on index i of left tensor
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 9a6431f29..4dd2e7c86 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -336,7 +336,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices())
   {
-    for (std::size_t i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+    for (Index i = 0; i < internal::array_size<Dimensions>::value; ++i) {
       eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]);
     }
 
@@ -739,7 +739,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
   {
     // Handle degenerate intervals by gracefully clamping and allowing m_dimensions to be zero
     DSizes<Index,NumDims> startIndicesClamped, stopIndicesClamped;
-    for (size_t i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+    for (Index i = 0; i < internal::array_size<Dimensions>::value; ++i) {
       eigen_assert(m_strides[i] != 0 && "0 stride is invalid");
       if(m_strides[i]>0){
         startIndicesClamped[i] = clamp(op.startIndices()[i], 0, m_impl.dimensions()[i]);

From 495f6c3c3a93767832f962bf5426b25125917f78 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 24 Aug 2018 23:56:13 +0200
Subject: [PATCH 600/680] Fix missing-braces warnings

---
 unsupported/test/cxx11_tensor_concatenation.cpp | 2 +-
 unsupported/test/cxx11_tensor_contraction.cpp   | 2 +-
 unsupported/test/cxx11_tensor_convolution.cpp   | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp
index f53515b4e..2e1abdfb1 100644
--- a/unsupported/test/cxx11_tensor_concatenation.cpp
+++ b/unsupported/test/cxx11_tensor_concatenation.cpp
@@ -56,7 +56,7 @@ static void test_static_dimension_failure()
   // either the code should change to 
   //  Tensor<int, 2>::Dimensions{{2, 3}}
   // or Tensor<int, 2>::Dimensions{Tensor<int, 2>::Dimensions{{2, 3}}}
-      .concatenate(right.reshape(Tensor<int, 2>::Dimensions{{2, 3}}), 0);
+      .concatenate(right.reshape(Tensor<int, 2>::Dimensions{{{2, 3}}}), 0);
 }
 
 template<int DataLayout>
diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp
index 928d20f6e..4e5922440 100644
--- a/unsupported/test/cxx11_tensor_contraction.cpp
+++ b/unsupported/test/cxx11_tensor_contraction.cpp
@@ -471,7 +471,7 @@ static void test_tensor_product()
   mat1.setRandom();
   mat2.setRandom();
 
-  Tensor<float, 4, DataLayout> result = mat1.contract(mat2, Eigen::array<DimPair, 0>{});
+  Tensor<float, 4, DataLayout> result = mat1.contract(mat2, Eigen::array<DimPair, 0>{{}});
 
   VERIFY_IS_EQUAL(result.dimension(0), 2);
   VERIFY_IS_EQUAL(result.dimension(1), 3);
diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp
index 9fe980648..01bc77bc1 100644
--- a/unsupported/test/cxx11_tensor_convolution.cpp
+++ b/unsupported/test/cxx11_tensor_convolution.cpp
@@ -25,7 +25,7 @@ static void test_evals()
 
   Tensor<float, 2, DataLayout> result(2,3);
   result.setZero();
-  Eigen::array<Tensor<float, 2>::Index, 1> dims3{0};
+  Eigen::array<Tensor<float, 2>::Index, 1> dims3{{0}};
 
   typedef TensorEvaluator<decltype(input.convolve(kernel, dims3)), DefaultDevice> Evaluator;
   Evaluator eval(input.convolve(kernel, dims3), DefaultDevice());

From 209b4972ece60c9193b492745b5bfac2bfd462d5 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 00:02:46 +0200
Subject: [PATCH 601/680] Fix conversion warning

---
 unsupported/test/EulerAngles.cpp | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/EulerAngles.cpp b/unsupported/test/EulerAngles.cpp
index 572fc08a3..67533e364 100644
--- a/unsupported/test/EulerAngles.cpp
+++ b/unsupported/test/EulerAngles.cpp
@@ -197,6 +197,7 @@ template<typename Scalar> void check_singular_cases(const Scalar& singularBeta)
 template<typename Scalar> void eulerangles_manual()
 {
   typedef Matrix<Scalar,3,1> Vector3;
+  typedef Matrix<Scalar,Dynamic,1> VectorX;
   const Vector3 Zero = Vector3::Zero();
   const Scalar PI = Scalar(EIGEN_PI);
   
@@ -213,13 +214,13 @@ template<typename Scalar> void eulerangles_manual()
   check_singular_cases(-PI);
   
   // non-singular cases
-  VectorXd alpha = VectorXd::LinSpaced(Eigen::Sequential, 20, Scalar(-0.99) * PI, PI);
-  VectorXd beta = VectorXd::LinSpaced(Eigen::Sequential, 20, Scalar(-0.49) * PI, Scalar(0.49) * PI);
-  VectorXd gamma = VectorXd::LinSpaced(Eigen::Sequential, 20, Scalar(-0.99) * PI, PI);
+  VectorX alpha = VectorX::LinSpaced(Eigen::Sequential, 20, Scalar(-0.99) * PI, PI);
+  VectorX beta =  VectorX::LinSpaced(Eigen::Sequential, 20, Scalar(-0.49) * PI, Scalar(0.49) * PI);
+  VectorX gamma = VectorX::LinSpaced(Eigen::Sequential, 20, Scalar(-0.99) * PI, PI);
   for (int i = 0; i < alpha.size(); ++i) {
     for (int j = 0; j < beta.size(); ++j) {
       for (int k = 0; k < gamma.size(); ++k) {
-        check_all_var(Vector3d(alpha(i), beta(j), gamma(k)));
+        check_all_var(Vector3(alpha(i), beta(j), gamma(k)));
       }
     }
   }

From f155e97adb45dac0fd0f5e457d0300f0a5e3bada Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 00:10:46 +0200
Subject: [PATCH 602/680] Previous fix broke compilation for clang

---
 unsupported/test/cxx11_tensor_concatenation.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp
index 2e1abdfb1..e223d9ffd 100644
--- a/unsupported/test/cxx11_tensor_concatenation.cpp
+++ b/unsupported/test/cxx11_tensor_concatenation.cpp
@@ -56,7 +56,7 @@ static void test_static_dimension_failure()
   // either the code should change to 
   //  Tensor<int, 2>::Dimensions{{2, 3}}
   // or Tensor<int, 2>::Dimensions{Tensor<int, 2>::Dimensions{{2, 3}}}
-      .concatenate(right.reshape(Tensor<int, 2>::Dimensions{{{2, 3}}}), 0);
+      .concatenate(right.reshape(Tensor<int, 2>::Dimensions(2, 3)), 0);
 }
 
 template<int DataLayout>

From 117bc5d50587d13c80fbe11b9ae7f86873cf5d2d Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 09:06:08 +0200
Subject: [PATCH 603/680] Fix some shadow warnings

---
 unsupported/Eigen/src/Splines/Spline.h |  7 +------
 unsupported/test/matrix_function.cpp   | 12 ++++--------
 unsupported/test/openglsupport.cpp     |  4 ----
 unsupported/test/polynomialsolver.cpp  |  1 -
 4 files changed, 5 insertions(+), 19 deletions(-)

diff --git a/unsupported/Eigen/src/Splines/Spline.h b/unsupported/Eigen/src/Splines/Spline.h
index 627f6e482..c1cf5b7e4 100644
--- a/unsupported/Eigen/src/Splines/Spline.h
+++ b/unsupported/Eigen/src/Splines/Spline.h
@@ -249,15 +249,13 @@ namespace Eigen
     DenseIndex degree,
     const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& knots)
   {
-    typedef typename Spline<_Scalar, _Dim, _Degree>::BasisVectorType BasisVectorType;
-
     const DenseIndex p = degree;
     const DenseIndex i = Spline::Span(u, degree, knots);
 
     const KnotVectorType& U = knots;
 
     BasisVectorType left(p+1); left(0) = Scalar(0);
-    BasisVectorType right(p+1); right(0) = Scalar(0);        
+    BasisVectorType right(p+1); right(0) = Scalar(0);
 
     VectorBlock<BasisVectorType,Degree>(left,1,p) = u - VectorBlock<const KnotVectorType,Degree>(U,i+1-p,p).reverse();
     VectorBlock<BasisVectorType,Degree>(right,1,p) = VectorBlock<const KnotVectorType,Degree>(U,i+1,p) - u;
@@ -380,9 +378,6 @@ namespace Eigen
     typedef Spline<_Scalar, _Dim, _Degree> SplineType;
     enum { Order = SplineTraits<SplineType>::OrderAtCompileTime };
 
-    typedef typename SplineTraits<SplineType>::Scalar Scalar;
-    typedef typename SplineTraits<SplineType>::BasisVectorType BasisVectorType;
-  
     const DenseIndex span = SplineType::Span(u, p, U);
 
     const DenseIndex n = (std::min)(p, order);
diff --git a/unsupported/test/matrix_function.cpp b/unsupported/test/matrix_function.cpp
index 93fb71430..2049b8ba0 100644
--- a/unsupported/test/matrix_function.cpp
+++ b/unsupported/test/matrix_function.cpp
@@ -23,9 +23,8 @@ inline bool test_isApprox_abs(const Type1& a, const Type2& b)
 
 // Returns a matrix with eigenvalues clustered around 0, 1 and 2.
 template<typename MatrixType>
-MatrixType randomMatrixWithRealEivals(const typename MatrixType::Index size)
+MatrixType randomMatrixWithRealEivals(const Index size)
 {
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
   MatrixType diag = MatrixType::Zero(size, size);
@@ -42,16 +41,15 @@ template <typename MatrixType, int IsComplex = NumTraits<typename internal::trai
 struct randomMatrixWithImagEivals
 {
   // Returns a matrix with eigenvalues clustered around 0 and +/- i.
-  static MatrixType run(const typename MatrixType::Index size);
+  static MatrixType run(const Index size);
 };
 
 // Partial specialization for real matrices
 template<typename MatrixType>
 struct randomMatrixWithImagEivals<MatrixType, 0>
 {
-  static MatrixType run(const typename MatrixType::Index size)
+  static MatrixType run(const Index size)
   {
-    typedef typename MatrixType::Index Index;
     typedef typename MatrixType::Scalar Scalar;
     MatrixType diag = MatrixType::Zero(size, size);
     Index i = 0;
@@ -77,9 +75,8 @@ struct randomMatrixWithImagEivals<MatrixType, 0>
 template<typename MatrixType>
 struct randomMatrixWithImagEivals<MatrixType, 1>
 {
-  static MatrixType run(const typename MatrixType::Index size)
+  static MatrixType run(const Index size)
   {
-    typedef typename MatrixType::Index Index;
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
     const Scalar imagUnit(0, 1);
@@ -171,7 +168,6 @@ void testMatrixType(const MatrixType& m)
 {
   // Matrices with clustered eigenvalue lead to different code paths
   // in MatrixFunction.h and are thus useful for testing.
-  typedef typename MatrixType::Index Index;
 
   const Index size = m.rows();
   for (int i = 0; i < g_repeat; i++) {
diff --git a/unsupported/test/openglsupport.cpp b/unsupported/test/openglsupport.cpp
index 460830086..eadd7f985 100644
--- a/unsupported/test/openglsupport.cpp
+++ b/unsupported/test/openglsupport.cpp
@@ -318,10 +318,6 @@ EIGEN_DECLARE_TEST(openglsupport)
         
       GLint prg_id = createShader(vtx,frg);
       
-      typedef Vector2d Vector2d;
-      typedef Vector3d Vector3d;
-      typedef Vector4d Vector4d;
-      
       VERIFY_UNIFORM(dv,v2d, Vector2d);
       VERIFY_UNIFORM(dv,v3d, Vector3d);
       VERIFY_UNIFORM(dv,v4d, Vector4d);
diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp
index 65efea0cb..50c74f797 100644
--- a/unsupported/test/polynomialsolver.cpp
+++ b/unsupported/test/polynomialsolver.cpp
@@ -30,7 +30,6 @@ struct increment_if_fixed_size
 template<int Deg, typename POLYNOMIAL, typename SOLVER>
 bool aux_evalSolver( const POLYNOMIAL& pols, SOLVER& psolve )
 {
-  typedef typename POLYNOMIAL::Index Index;
   typedef typename POLYNOMIAL::Scalar Scalar;
   typedef typename POLYNOMIAL::RealScalar RealScalar;
 

From 4b1ad086b59a8efdc9a4971265868b3cc7a3d290 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 10:07:17 +0200
Subject: [PATCH 604/680] Fix shadow warnings in doc-snippets

---
 doc/snippets/DirectionWise_hnormalized.cpp | 3 +--
 doc/snippets/VectorwiseOp_homogeneous.cpp  | 3 +--
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/doc/snippets/DirectionWise_hnormalized.cpp b/doc/snippets/DirectionWise_hnormalized.cpp
index 3410790a8..2451f6e7b 100644
--- a/doc/snippets/DirectionWise_hnormalized.cpp
+++ b/doc/snippets/DirectionWise_hnormalized.cpp
@@ -1,7 +1,6 @@
-typedef Matrix<double,4,Dynamic> Matrix4Xd;
 Matrix4Xd M = Matrix4Xd::Random(4,5);
 Projective3d P(Matrix4d::Random());
 cout << "The matrix M is:" << endl << M << endl << endl;
 cout << "M.colwise().hnormalized():" << endl << M.colwise().hnormalized() << endl << endl;
 cout << "P*M:" << endl << P*M << endl << endl;
-cout << "(P*M).colwise().hnormalized():" << endl << (P*M).colwise().hnormalized() << endl << endl;
\ No newline at end of file
+cout << "(P*M).colwise().hnormalized():" << endl << (P*M).colwise().hnormalized() << endl << endl;
diff --git a/doc/snippets/VectorwiseOp_homogeneous.cpp b/doc/snippets/VectorwiseOp_homogeneous.cpp
index aba4fed0e..67cf5737d 100644
--- a/doc/snippets/VectorwiseOp_homogeneous.cpp
+++ b/doc/snippets/VectorwiseOp_homogeneous.cpp
@@ -1,7 +1,6 @@
-typedef Matrix<double,3,Dynamic> Matrix3Xd;
 Matrix3Xd M = Matrix3Xd::Random(3,5);
 Projective3d P(Matrix4d::Random());
 cout << "The matrix M is:" << endl << M << endl << endl;
 cout << "M.colwise().homogeneous():" << endl << M.colwise().homogeneous() << endl << endl;
 cout << "P * M.colwise().homogeneous():" << endl << P * M.colwise().homogeneous() << endl << endl;
-cout << "P * M.colwise().homogeneous().hnormalized(): " << endl << (P * M.colwise().homogeneous()).colwise().hnormalized() << endl << endl;
\ No newline at end of file
+cout << "P * M.colwise().homogeneous().hnormalized(): " << endl << (P * M.colwise().homogeneous()).colwise().hnormalized() << endl << endl;

From 42123ff38bb061a25861d3ec0f98c059d07ec4c1 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 11:53:28 +0200
Subject: [PATCH 605/680] Make unit test C++03 compatible

---
 .../test/cxx11_tensor_block_access.cpp        | 59 +++++++++++--------
 1 file changed, 35 insertions(+), 24 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index 74ce2cc80..40e6227e5 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -367,6 +367,40 @@ static void test_block_io_copy_using_reordered_dimensions() {
   delete[] output_data;
 }
 
+template<typename Scalar, typename StorageIndex, int Dim>
+class EqualityChecker
+{
+    const Scalar* input_data;
+    const DSizes<StorageIndex, Dim> &input_dims, &input_strides, &output_dims, &output_strides;
+    void check_recursive(const Scalar* input, const Scalar* output, int depth=0) const
+    {
+        if(depth==Dim)
+        {
+            VERIFY_IS_EQUAL(*input, *output);
+            return;
+        }
+
+        for(int i=0; i<output_dims[depth]; ++i)
+        {
+            check_recursive(input + i % input_dims[depth] * input_strides[depth], output + i*output_strides[depth], depth+1);
+        }
+    }
+public:
+    EqualityChecker(const Scalar* input_data_,
+            const DSizes<StorageIndex, Dim> &input_dims_, const DSizes<StorageIndex, Dim> &input_strides_,
+            const DSizes<StorageIndex, Dim> &output_dims_, const DSizes<StorageIndex, Dim> &output_strides_)
+        : input_data(input_data_)
+        , input_dims(input_dims_), input_strides(input_strides_)
+        , output_dims(output_dims_), output_strides(output_strides_)
+        {}
+
+    void operator()(const Scalar* output_data) const
+    {
+        check_recursive(input_data, output_data);
+    }
+};
+
+
 template <int Layout>
 static void test_block_io_zero_stride()
 {
@@ -398,30 +432,7 @@ static void test_block_io_zero_stride()
   input_tensor_strides_with_zeros[4] = 0;
 
   // Verify that data was correctly read/written from/into the block.
-  const auto verify_is_equal = [&](const float* output_data) {
-    for (int i = 0; i < output_tensor_dims[0]; ++i) {
-      for (int j = 0; j < output_tensor_dims[1]; ++j) {
-        for (int k = 0; k < output_tensor_dims[2]; ++k) {
-          for (int l = 0; l < output_tensor_dims[3]; ++l) {
-            for (int m = 0; m < output_tensor_dims[4]; ++m) {
-              const Index output_offset =
-                  i * output_tensor_strides[0] + j * output_tensor_strides[1] +
-                  k * output_tensor_strides[2] + l * output_tensor_strides[3] +
-                  m * output_tensor_strides[4];
-              const Index input_offset =
-                  i % input_tensor_dims[0] * input_tensor_strides[0] +
-                  j % input_tensor_dims[1] * input_tensor_strides[1] +
-                  k % input_tensor_dims[2] * input_tensor_strides[2] +
-                  l % input_tensor_dims[3] * input_tensor_strides[3] +
-                  m % input_tensor_dims[4] * input_tensor_strides[4];
-              VERIFY_IS_EQUAL(output_data[output_offset],
-                              input_data[input_offset]);
-            }
-          }
-        }
-      }
-    }
-  };
+  const EqualityChecker<float, Index, 5> verify_is_equal(input_data, input_tensor_dims, input_tensor_strides, output_tensor_dims, output_tensor_strides);
 
   {
     float* output_data = new float[output_tensor_dims.TotalSize()];

From b1653d15996b844852e2cefdd4d63e55dbc771f5 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 25 Aug 2018 12:21:00 +0200
Subject: [PATCH 606/680] Fix some trivial C++11 vs C++03 compatibility
 warnings

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h        | 2 +-
 unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h | 4 ++--
 unsupported/Eigen/CXX11/src/util/EmulateArray.h         | 8 ++++----
 3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index aa500eb70..4cb49e9b8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -62,7 +62,7 @@ struct cond<RowMajor> {
  */
 enum TensorBlockShapeType {
   kUniformAllDims,
-  kSkewedInnerDims,
+  kSkewedInnerDims
 };
 
 struct TensorOpResourceRequirements {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index b4a77b022..560e3ec22 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -105,7 +105,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
-  bool isCopy= false, nByOne = false, oneByN = false;
+  bool isCopy, nByOne, oneByN;
 
   enum {
     IsAligned = true,
@@ -116,7 +116,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   };
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
-    : m_broadcast(op.broadcast()),m_impl(op.expression(), device)
+    : isCopy(false), nByOne(false), oneByN(false), m_broadcast(op.broadcast()),m_impl(op.expression(), device)
   {
     // The broadcasting op doesn't change the rank of the tensor. One can't broadcast a scalar
     // and store the result in a scalar. Instead one should reshape the scalar into a a N-D
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index 32db51592..d5c000e08 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -207,16 +207,16 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& array_get(const array<T,N>& a) {
 }
 
 template<class T, std::size_t N> struct array_size<array<T,N> > {
-  static const size_t value = N;
+  enum { value = N };
 };
 template<class T, std::size_t N> struct array_size<array<T,N>& > {
-  static const size_t value = N;
+  enum { value = N };
 };
 template<class T, std::size_t N> struct array_size<const array<T,N> > {
-  static const size_t value = N;
+  enum { value = N };
 };
 template<class T, std::size_t N> struct array_size<const array<T,N>& > {
-  static const size_t value = N;
+  enum { value = N };
 };
 
 }  // end namespace internal

From d5ed64512fc79dff800c90acd73e3e6a08a3d2c3 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 27 Aug 2018 10:38:20 +0200
Subject: [PATCH 607/680] bug #1573: workaround gcc 4.7 and 4.8 bug

---
 test/geo_quaternion.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp
index ed801c71b..27219db10 100644
--- a/test/geo_quaternion.cpp
+++ b/test/geo_quaternion.cpp
@@ -290,6 +290,8 @@ template<typename PlainObjectType> void check_const_correctness(const PlainObjec
 
 // Regression for bug 1573
 struct MovableClass {
+  // The following line is a workaround for gcc 4.7 and 4.8 (see bug 1573 comments).
+  static_assert(std::is_nothrow_move_constructible<Quaternionf>::value,"");
   MovableClass() = default;
   MovableClass(const MovableClass&) = default;
   MovableClass(MovableClass&&) noexcept = default;

From 57472886764ff71ad45338c6538649f7a8fa3d0e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 27 Aug 2018 13:07:34 +0200
Subject: [PATCH 608/680] Disable a bonus unit-test which is broken with gcc
 4.7

---
 test/meta.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/test/meta.cpp b/test/meta.cpp
index a6a67b85c..ea9607fe7 100644
--- a/test/meta.cpp
+++ b/test/meta.cpp
@@ -102,7 +102,13 @@ EIGEN_DECLARE_TEST(meta)
   }
 
   STATIC_CHECK(( !internal::is_convertible<MyInterface, MyImpl>::value ));
+  #if (!EIGEN_COMP_GNUC_STRICT) || (EIGEN_GNUC_AT_LEAST(4,8))
+  // GCC prior to 4.8 fails to compile this test:
+  // error: cannot allocate an object of abstract type 'MyInterface'
+  // In other word, it does not obey SFINAE.
+  // Nevertheless, we don't really care about supporting abstract type as scalar type!
   STATIC_CHECK(( !internal::is_convertible<MyImpl, MyInterface>::value ));
+  #endif
   STATIC_CHECK((  internal::is_convertible<MyImpl, const MyInterface&>::value ));
   {
     int i;

From 42f3ee4fb878b3e16a8cbc99ec90ed330f59899c Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 28 Aug 2018 11:44:15 +0200
Subject: [PATCH 609/680] Old gcc versions have problems with recursive #pragma
 GCC diagnostic push/pop Workaround: Don't include "DisableStupidWarnings.h"
 before including other main-headers

---
 Eigen/Eigenvalues               | 4 ++--
 Eigen/Geometry                  | 4 ++--
 Eigen/QR                        | 4 ++--
 Eigen/SparseQR                  | 1 -
 unsupported/Eigen/ArpackSupport | 4 ++--
 unsupported/Eigen/Polynomials   | 4 ++--
 6 files changed, 10 insertions(+), 11 deletions(-)

diff --git a/Eigen/Eigenvalues b/Eigen/Eigenvalues
index f3f661b07..7d6ac787b 100644
--- a/Eigen/Eigenvalues
+++ b/Eigen/Eigenvalues
@@ -10,14 +10,14 @@
 
 #include "Core"
 
-#include "src/Core/util/DisableStupidWarnings.h"
-
 #include "Cholesky"
 #include "Jacobi"
 #include "Householder"
 #include "LU"
 #include "Geometry"
 
+#include "src/Core/util/DisableStupidWarnings.h"
+
 /** \defgroup Eigenvalues_Module Eigenvalues module
   *
   *
diff --git a/Eigen/Geometry b/Eigen/Geometry
index 131a4edfc..04aa316cb 100644
--- a/Eigen/Geometry
+++ b/Eigen/Geometry
@@ -10,12 +10,12 @@
 
 #include "Core"
 
-#include "src/Core/util/DisableStupidWarnings.h"
-
 #include "SVD"
 #include "LU"
 #include <limits>
 
+#include "src/Core/util/DisableStupidWarnings.h"
+
 /** \defgroup Geometry_Module Geometry module
   *
   * This module provides support for:
diff --git a/Eigen/QR b/Eigen/QR
index c7e914469..1be1863a1 100644
--- a/Eigen/QR
+++ b/Eigen/QR
@@ -10,12 +10,12 @@
 
 #include "Core"
 
-#include "src/Core/util/DisableStupidWarnings.h"
-
 #include "Cholesky"
 #include "Jacobi"
 #include "Householder"
 
+#include "src/Core/util/DisableStupidWarnings.h"
+
 /** \defgroup QR_Module QR module
   *
   *
diff --git a/Eigen/SparseQR b/Eigen/SparseQR
index a6f3b7f7d..f5fc5fa7f 100644
--- a/Eigen/SparseQR
+++ b/Eigen/SparseQR
@@ -28,7 +28,6 @@
   * 
   */
 
-#include "OrderingMethods"
 #include "src/SparseCore/SparseColEtree.h"
 #include "src/SparseQR/SparseQR.h"
 
diff --git a/unsupported/Eigen/ArpackSupport b/unsupported/Eigen/ArpackSupport
index 37a2799ef..a0d4820e1 100644
--- a/unsupported/Eigen/ArpackSupport
+++ b/unsupported/Eigen/ArpackSupport
@@ -11,8 +11,6 @@
 
 #include <Eigen/Core>
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
-
 /** \defgroup ArpackSupport_Module Arpack support module
   *
   * This module provides a wrapper to Arpack, a library for sparse eigenvalue decomposition.
@@ -23,6 +21,8 @@
   */
 
 #include <Eigen/SparseCholesky>
+
+#include <Eigen/src/Core/util/DisableStupidWarnings.h>
 #include "src/Eigenvalues/ArpackSelfAdjointEigenSolver.h"
 
 #include <Eigen/src/Core/util/ReenableStupidWarnings.h>
diff --git a/unsupported/Eigen/Polynomials b/unsupported/Eigen/Polynomials
index cece56337..334b03142 100644
--- a/unsupported/Eigen/Polynomials
+++ b/unsupported/Eigen/Polynomials
@@ -11,10 +11,10 @@
 
 #include <Eigen/Core>
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
-
 #include <Eigen/Eigenvalues>
 
+#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+
 // Note that EIGEN_HIDE_HEAVY_CODE has to be defined per module
 #if (defined EIGEN_EXTERN_INSTANTIATIONS) && (EIGEN_EXTERN_INSTANTIATIONS>=2)
   #ifndef EIGEN_HIDE_HEAVY_CODE

From befaf83f5f4cfafc7b17c8e384e0457c8493393f Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 28 Aug 2018 13:21:28 +0200
Subject: [PATCH 610/680] bug #1590: fix collision with some system headers
 defining the macro FP32

---
 Eigen/src/Core/arch/GPU/Half.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index aca56fa72..b44a4356a 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -399,7 +399,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned sho
   return h;
 }
 
-union FP32 {
+union float32_bits {
   unsigned int u;
   float f;
 };
@@ -416,11 +416,11 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
   return h;
 
 #else
-  FP32 f; f.f = ff;
+  float32_bits f; f.f = ff;
 
-  const FP32 f32infty = { 255 << 23 };
-  const FP32 f16max = { (127 + 16) << 23 };
-  const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
+  const float32_bits f32infty = { 255 << 23 };
+  const float32_bits f16max = { (127 + 16) << 23 };
+  const float32_bits denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
   unsigned int sign_mask = 0x80000000u;
   __half_raw o;
   o.x = static_cast<unsigned short>(0x0u);
@@ -470,9 +470,9 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
   return _cvtsh_ss(h.x);
 
 #else
-  const FP32 magic = { 113 << 23 };
+  const float32_bits magic = { 113 << 23 };
   const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
-  FP32 o;
+  float32_bits o;
 
   o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
   unsigned int exp = shifted_exp & o.u;   // just the exponent

From ef4d79fed8b84f03fe3c7ace98442c822cf3472a Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 28 Aug 2018 18:26:22 +0200
Subject: [PATCH 611/680] Disable/ReenableStupidWarnings did not work properly,
 when included recursively

---
 Eigen/src/Core/util/DisableStupidWarnings.h  | 8 ++++++++
 Eigen/src/Core/util/ReenableStupidWarnings.h | 6 +++++-
 2 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index af832ae77..d04b52649 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -80,4 +80,12 @@
   #pragma diag_suppress 2739
 #endif
 
+#else
+// warnings already disabled:
+# ifndef EIGEN_WARNINGS_DISABLED_2
+#  define EIGEN_WARNINGS_DISABLED_2
+# elif defined(EIGEN_INTERNAL_DEBUGGING)
+#  error "Do not include \"DisableStupidWarnings.h\" recursively more than twice!"
+# endif
+
 #endif // not EIGEN_WARNINGS_DISABLED
diff --git a/Eigen/src/Core/util/ReenableStupidWarnings.h b/Eigen/src/Core/util/ReenableStupidWarnings.h
index e9353a623..e23a128d1 100644
--- a/Eigen/src/Core/util/ReenableStupidWarnings.h
+++ b/Eigen/src/Core/util/ReenableStupidWarnings.h
@@ -1,4 +1,8 @@
-#ifdef EIGEN_WARNINGS_DISABLED
+#ifdef EIGEN_WARNINGS_DISABLED_2
+// "DisableStupidWarnings.h" was included twice recursively: Do not reenable warnings yet!
+#  undef EIGEN_WARNINGS_DISABLED_2
+
+#elif defined(EIGEN_WARNINGS_DISABLED)
 #undef EIGEN_WARNINGS_DISABLED
 
 #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS

From 73ca600bca42f5ccf507ca4be1f53921973b44fe Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 28 Aug 2018 18:32:39 +0200
Subject: [PATCH 612/680] Fix numerous shadow-warnings for GCC<=4.8

---
 Eigen/SparseLU                       | 4 ++++
 Eigen/src/StlSupport/StdDeque.h      | 2 +-
 Eigen/src/StlSupport/StdList.h       | 2 +-
 Eigen/src/StlSupport/StdVector.h     | 2 +-
 blas/common.h                        | 8 ++++++++
 test/mapstaticmethods.cpp            | 4 ++++
 unsupported/Eigen/AlignedVector3     | 4 ++++
 unsupported/Eigen/AutoDiff           | 6 ++++++
 unsupported/Eigen/CXX11/Tensor       | 4 ++--
 unsupported/Eigen/FFT                | 5 +++++
 unsupported/Eigen/IterativeSolvers   | 8 ++++++--
 unsupported/Eigen/LevenbergMarquardt | 4 ++++
 unsupported/Eigen/MatrixFunctions    | 4 ++++
 unsupported/Eigen/Splines            | 4 ++++
 unsupported/test/NumericalDiff.cpp   | 2 +-
 unsupported/test/autodiff.cpp        | 2 +-
 unsupported/test/dgmres.cpp          | 2 +-
 unsupported/test/forward_adolc.cpp   | 2 +-
 18 files changed, 58 insertions(+), 11 deletions(-)

diff --git a/Eigen/SparseLU b/Eigen/SparseLU
index 38b38b531..37c4a5c5a 100644
--- a/Eigen/SparseLU
+++ b/Eigen/SparseLU
@@ -23,6 +23,8 @@
 // Ordering interface
 #include "OrderingMethods"
 
+#include "src/Core/util/DisableStupidWarnings.h"
+
 #include "src/SparseLU/SparseLU_gemm_kernel.h"
 
 #include "src/SparseLU/SparseLU_Structs.h"
@@ -43,4 +45,6 @@
 #include "src/SparseLU/SparseLU_Utils.h"
 #include "src/SparseLU/SparseLU.h"
 
+#include "src/Core/util/ReenableStupidWarnings.h"
+
 #endif // EIGEN_SPARSELU_MODULE_H
diff --git a/Eigen/src/StlSupport/StdDeque.h b/Eigen/src/StlSupport/StdDeque.h
index cf1fedf92..8adfbb5c1 100644
--- a/Eigen/src/StlSupport/StdDeque.h
+++ b/Eigen/src/StlSupport/StdDeque.h
@@ -36,7 +36,7 @@ namespace std \
     deque(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : deque_base(first, last, a) {} \
     deque(const deque& c) : deque_base(c) {}  \
     explicit deque(size_type num, const value_type& val = value_type()) : deque_base(num, val) {} \
-    deque(iterator start, iterator end) : deque_base(start, end) {}  \
+    deque(iterator start_, iterator end_) : deque_base(start_, end_) {}  \
     deque& operator=(const deque& x) {  \
       deque_base::operator=(x);  \
       return *this;  \
diff --git a/Eigen/src/StlSupport/StdList.h b/Eigen/src/StlSupport/StdList.h
index e1eba4985..1b6d44061 100644
--- a/Eigen/src/StlSupport/StdList.h
+++ b/Eigen/src/StlSupport/StdList.h
@@ -35,7 +35,7 @@ namespace std \
     list(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : list_base(first, last, a) {} \
     list(const list& c) : list_base(c) {}  \
     explicit list(size_type num, const value_type& val = value_type()) : list_base(num, val) {} \
-    list(iterator start, iterator end) : list_base(start, end) {}  \
+    list(iterator start_, iterator end_) : list_base(start_, end_) {}  \
     list& operator=(const list& x) {  \
       list_base::operator=(x);  \
       return *this;  \
diff --git a/Eigen/src/StlSupport/StdVector.h b/Eigen/src/StlSupport/StdVector.h
index ec22821d2..7787edd72 100644
--- a/Eigen/src/StlSupport/StdVector.h
+++ b/Eigen/src/StlSupport/StdVector.h
@@ -36,7 +36,7 @@ namespace std \
     vector(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : vector_base(first, last, a) {} \
     vector(const vector& c) : vector_base(c) {}  \
     explicit vector(size_type num, const value_type& val = value_type()) : vector_base(num, val) {} \
-    vector(iterator start, iterator end) : vector_base(start, end) {}  \
+    vector(iterator start_, iterator end_) : vector_base(start_, end_) {}  \
     vector& operator=(const vector& x) {  \
       vector_base::operator=(x);  \
       return *this;  \
diff --git a/blas/common.h b/blas/common.h
index 61d8344d9..960c09cc6 100644
--- a/blas/common.h
+++ b/blas/common.h
@@ -10,6 +10,14 @@
 #ifndef EIGEN_BLAS_COMMON_H
 #define EIGEN_BLAS_COMMON_H
 
+#ifdef __GNUC__
+# if __GNUC__<5
+// GCC < 5.0 does not like the global Scalar typedef
+// we just keep shadow-warnings disabled permanently
+#  define EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+# endif
+#endif
+
 #include "../Eigen/Core"
 #include "../Eigen/Jacobi"
 
diff --git a/test/mapstaticmethods.cpp b/test/mapstaticmethods.cpp
index f32c0beec..d0128ba94 100644
--- a/test/mapstaticmethods.cpp
+++ b/test/mapstaticmethods.cpp
@@ -9,8 +9,12 @@
 
 #include "main.h"
 
+// GCC<=4.8 has spurious shadow warnings, because `ptr` re-appears inside template instantiations
+// workaround: put these in an anonymous namespace
+namespace {
 float *ptr;
 const float *const_ptr;
+}
 
 template<typename PlainObjectType,
          bool IsDynamicSize = PlainObjectType::SizeAtCompileTime == Dynamic,
diff --git a/unsupported/Eigen/AlignedVector3 b/unsupported/Eigen/AlignedVector3
index 47a86d4c0..c16f16a97 100644
--- a/unsupported/Eigen/AlignedVector3
+++ b/unsupported/Eigen/AlignedVector3
@@ -12,6 +12,8 @@
 
 #include <Eigen/Geometry>
 
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 namespace Eigen {
 
 /**
@@ -221,4 +223,6 @@ struct evaluator<AlignedVector3<Scalar> >
 
 }
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
 #endif // EIGEN_ALIGNED_VECTOR3
diff --git a/unsupported/Eigen/AutoDiff b/unsupported/Eigen/AutoDiff
index abf5b7d67..7a4ff460c 100644
--- a/unsupported/Eigen/AutoDiff
+++ b/unsupported/Eigen/AutoDiff
@@ -28,11 +28,17 @@ namespace Eigen {
 //@{
 
 }
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 
 #include "src/AutoDiff/AutoDiffScalar.h"
 // #include "src/AutoDiff/AutoDiffVector.h"
 #include "src/AutoDiff/AutoDiffJacobian.h"
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+
+
 namespace Eigen {
 //@}
 }
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 47514703a..0165d0144 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -26,9 +26,9 @@
 #include <utility>
 #endif
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
-
 #include "../SpecialFunctions"
+
+#include <Eigen/src/Core/util/DisableStupidWarnings.h>
 #include "src/util/CXX11Meta.h"
 #include "src/util/MaxSizeVector.h"
 
diff --git a/unsupported/Eigen/FFT b/unsupported/Eigen/FFT
index d8cf3e642..8ae053677 100644
--- a/unsupported/Eigen/FFT
+++ b/unsupported/Eigen/FFT
@@ -68,6 +68,8 @@
   */
  
 
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 #ifdef EIGEN_FFTW_DEFAULT
 // FFTW: faster, GPL -- incompatible with Eigen in LGPL form, bigger code size
 #  include <fftw3.h>
@@ -415,5 +417,8 @@ void fft_inv_proxy<T_SrcMat,T_FftIfc>::evalTo(T_DestMat& dst) const
 }
 
 }
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
 #endif
 /* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/unsupported/Eigen/IterativeSolvers b/unsupported/Eigen/IterativeSolvers
index 31e880bdc..fe8a6d00d 100644
--- a/unsupported/Eigen/IterativeSolvers
+++ b/unsupported/Eigen/IterativeSolvers
@@ -11,6 +11,8 @@
 #define EIGEN_ITERATIVE_SOLVERS_MODULE_H
 
 #include <Eigen/Sparse>
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/Householder"
 
 /**
   * \defgroup IterativeSolvers_Module Iterative solvers module
@@ -24,19 +26,21 @@
   */
 //@{
 
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 #ifndef EIGEN_MPL2_ONLY
 #include "src/IterativeSolvers/IterationController.h"
 #include "src/IterativeSolvers/ConstrainedConjGrad.h"
 #endif
 
 #include "src/IterativeSolvers/IncompleteLU.h"
-#include "../../Eigen/Jacobi"
-#include "../../Eigen/Householder"
 #include "src/IterativeSolvers/GMRES.h"
 #include "src/IterativeSolvers/DGMRES.h"
 //#include "src/IterativeSolvers/SSORPreconditioner.h"
 #include "src/IterativeSolvers/MINRES.h"
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
 //@}
 
 #endif // EIGEN_ITERATIVE_SOLVERS_MODULE_H
diff --git a/unsupported/Eigen/LevenbergMarquardt b/unsupported/Eigen/LevenbergMarquardt
index 0fe2680ba..e77b746c9 100644
--- a/unsupported/Eigen/LevenbergMarquardt
+++ b/unsupported/Eigen/LevenbergMarquardt
@@ -30,6 +30,9 @@
   */
 
 #include "Eigen/SparseCore"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 #ifndef EIGEN_PARSED_BY_DOXYGEN
 
 #include "src/LevenbergMarquardt/LMqrsolv.h"
@@ -41,5 +44,6 @@
 #include "src/LevenbergMarquardt/LevenbergMarquardt.h"
 #include "src/LevenbergMarquardt/LMonestep.h"
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_LEVENBERGMARQUARDT_MODULE
diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions
index 0320606c1..17eef8813 100644
--- a/unsupported/Eigen/MatrixFunctions
+++ b/unsupported/Eigen/MatrixFunctions
@@ -53,12 +53,16 @@
   *
   */
 
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 #include "src/MatrixFunctions/MatrixExponential.h"
 #include "src/MatrixFunctions/MatrixFunction.h"
 #include "src/MatrixFunctions/MatrixSquareRoot.h"
 #include "src/MatrixFunctions/MatrixLogarithm.h"
 #include "src/MatrixFunctions/MatrixPower.h"
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
 
 /** 
 \page matrixbaseextra_page
diff --git a/unsupported/Eigen/Splines b/unsupported/Eigen/Splines
index 322e6b9f5..2ca581364 100644
--- a/unsupported/Eigen/Splines
+++ b/unsupported/Eigen/Splines
@@ -24,8 +24,12 @@ namespace Eigen
   */
 }
 
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
 #include "src/Splines/SplineFwd.h"
 #include "src/Splines/Spline.h"
 #include "src/Splines/SplineFitting.h"
 
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
 #endif // EIGEN_SPLINES_MODULE_H
diff --git a/unsupported/test/NumericalDiff.cpp b/unsupported/test/NumericalDiff.cpp
index 35f2f6d7c..6d836413b 100644
--- a/unsupported/test/NumericalDiff.cpp
+++ b/unsupported/test/NumericalDiff.cpp
@@ -24,7 +24,7 @@ struct Functor
   int m_inputs, m_values;
   
   Functor() : m_inputs(InputsAtCompileTime), m_values(ValuesAtCompileTime) {}
-  Functor(int inputs, int values) : m_inputs(inputs), m_values(values) {}
+  Functor(int inputs_, int values_) : m_inputs(inputs_), m_values(values_) {}
   
   int inputs() const { return m_inputs; }
   int values() const { return m_values; }
diff --git a/unsupported/test/autodiff.cpp b/unsupported/test/autodiff.cpp
index 495cee03e..bafea6ae9 100644
--- a/unsupported/test/autodiff.cpp
+++ b/unsupported/test/autodiff.cpp
@@ -44,7 +44,7 @@ struct TestFunc1
   int m_inputs, m_values;
 
   TestFunc1() : m_inputs(InputsAtCompileTime), m_values(ValuesAtCompileTime) {}
-  TestFunc1(int inputs, int values) : m_inputs(inputs), m_values(values) {}
+  TestFunc1(int inputs_, int values_) : m_inputs(inputs_), m_values(values_) {}
 
   int inputs() const { return m_inputs; }
   int values() const { return m_values; }
diff --git a/unsupported/test/dgmres.cpp b/unsupported/test/dgmres.cpp
index 04f5ad670..5f63161b2 100644
--- a/unsupported/test/dgmres.cpp
+++ b/unsupported/test/dgmres.cpp
@@ -9,7 +9,7 @@
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #include "../../test/sparse_solver.h"
-#include <Eigen/src/IterativeSolvers/DGMRES.h>
+#include <unsupported/Eigen/IterativeSolvers>
 
 template<typename T> void test_dgmres_T()
 {
diff --git a/unsupported/test/forward_adolc.cpp b/unsupported/test/forward_adolc.cpp
index 688594eff..14a909d3b 100644
--- a/unsupported/test/forward_adolc.cpp
+++ b/unsupported/test/forward_adolc.cpp
@@ -35,7 +35,7 @@ struct TestFunc1
   int m_inputs, m_values;
 
   TestFunc1() : m_inputs(InputsAtCompileTime), m_values(ValuesAtCompileTime) {}
-  TestFunc1(int inputs, int values) : m_inputs(inputs), m_values(values) {}
+  TestFunc1(int inputs_, int values_) : m_inputs(inputs_), m_values(values_) {}
 
   int inputs() const { return m_inputs; }
   int values() const { return m_values; }

From 20ba2eee6d4bc75fe2fbcd2af7218e47abcb76df Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Tue, 28 Aug 2018 18:33:24 +0200
Subject: [PATCH 613/680] gcc thinks this may not be initialized

---
 unsupported/test/NonLinearOptimization.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp
index cc95cea29..2770c3c49 100644
--- a/unsupported/test/NonLinearOptimization.cpp
+++ b/unsupported/test/NonLinearOptimization.cpp
@@ -565,7 +565,7 @@ void testLmdif1()
 
   // do the computation
   lmdif_functor functor;
-  DenseIndex nfev;
+  DenseIndex nfev = -1; // initialize to avoid maybe-uninitialized warning
   info = LevenbergMarquardt<lmdif_functor>::lmdif1(functor, x, &nfev);
 
   // check return value

From 7ec8b40ad90bcc63d36fc5a36b517d4a07c60b4b Mon Sep 17 00:00:00 2001
From: Mehdi Goli <mehdi.goli@codeplay.com>
Date: Tue, 28 Aug 2018 14:20:48 +0100
Subject: [PATCH 614/680] Collapsed revision * Separating SYCL math function. *
 Converting function overload to function specialisation. * Applying the
 suggested design.

---
 Eigen/src/Core/MathFunctions.h | 253 ++++++++++++++++-----------------
 1 file changed, 123 insertions(+), 130 deletions(-)

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index 58cf7021a..72116e144 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -876,7 +876,7 @@ template<typename T> T generic_fast_tanh_float(const T& a_x);
 
 namespace numext {
 
-#if (!defined(EIGEN_GPUCC) || defined(EIGEN_CONSTEXPR_ARE_DEVICE_FUNC)) && !defined(__SYCL_DEVICE_ONLY__)
+#if (!defined(EIGEN_GPUCC) || defined(EIGEN_CONSTEXPR_ARE_DEVICE_FUNC)) 
 template<typename T>
 EIGEN_DEVICE_FUNC
 EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
@@ -892,80 +892,6 @@ EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
   EIGEN_USING_STD_MATH(max);
   return max EIGEN_NOT_A_MACRO (x,y);
 }
-
-#elif defined(__SYCL_DEVICE_ONLY__)
-template<typename T>
-EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
-{
-  return y < x ? y : x;
-}
-
-template<typename T>
-EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
-{
-  return x < y ? y : x;
-}
-
-EIGEN_ALWAYS_INLINE int mini(const int& x, const int& y)
-{
-  return cl::sycl::min(x,y);
-}
-
-EIGEN_ALWAYS_INLINE int maxi(const int& x, const int& y)
-{
-  return cl::sycl::max(x,y);
-}
-
-EIGEN_ALWAYS_INLINE unsigned int mini(const unsigned int& x, const unsigned int& y)
-{
-  return cl::sycl::min(x,y);
-}
-
-EIGEN_ALWAYS_INLINE unsigned int maxi(const unsigned int& x, const unsigned int& y)
-{
-  return cl::sycl::max(x,y);
-}
-
-EIGEN_ALWAYS_INLINE  long mini(const long & x, const long & y)
-{
-  return cl::sycl::min(x,y);
-}
-
-EIGEN_ALWAYS_INLINE  long maxi(const long & x, const long & y)
-{
-  return cl::sycl::max(x,y);
-}
-
-EIGEN_ALWAYS_INLINE unsigned long mini(const unsigned long& x, const unsigned long& y)
-{
-  return cl::sycl::min(x,y);
-}
-
-EIGEN_ALWAYS_INLINE unsigned long maxi(const unsigned long& x, const unsigned long& y)
-{
-  return cl::sycl::max(x,y);
-}
-
-EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
-{
-  return cl::sycl::fmin(x,y);
-}
-
-EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
-{
-  return cl::sycl::fmax(x,y);
-}
-
-EIGEN_ALWAYS_INLINE double mini(const double& x, const double& y)
-{
-  return cl::sycl::fmin(x,y);
-}
-
-EIGEN_ALWAYS_INLINE double maxi(const double& x, const double& y)
-{
-  return cl::sycl::fmax(x,y);
-}
-
 #else
 template<typename T>
 EIGEN_DEVICE_FUNC
@@ -1028,6 +954,75 @@ EIGEN_ALWAYS_INLINE long double maxi(const long double& x, const long double& y)
 }
 #endif
 
+#if defined(__SYCL_DEVICE_ONLY__)
+
+
+#define SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_BINARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_char)   \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_short)  \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_int)    \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_long)
+#define SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_UNARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_char)   \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_short)  \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_int)    \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_long)
+#define SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_BINARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_uchar)  \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_ushort) \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_uint)   \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_ulong)
+#define SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_UNARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_uchar)  \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_ushort) \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_uint)   \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_ulong)
+#define SYCL_SPECIALIZE_INTEGER_TYPES_BINARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_BINARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_BINARY(NAME, FUNC)
+#define SYCL_SPECIALIZE_INTEGER_TYPES_UNARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_UNARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_UNARY(NAME, FUNC)
+#define SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, cl::sycl::cl_float) \
+  SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC,cl::sycl::cl_double)
+#define SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(NAME, FUNC) \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, cl::sycl::cl_float) \
+  SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC,cl::sycl::cl_double)
+#define SYCL_SPECIALIZE_FLOATING_TYPES_UNARY_FUNC_RET_TYPE(NAME, FUNC, RET_TYPE) \
+  SYCL_SPECIALIZE_GEN_UNARY_FUNC(NAME, FUNC, RET_TYPE, cl::sycl::cl_float) \
+  SYCL_SPECIALIZE_GEN_UNARY_FUNC(NAME, FUNC, RET_TYPE, cl::sycl::cl_double) 
+
+#define SYCL_SPECIALIZE_GEN_UNARY_FUNC(NAME, FUNC, RET_TYPE, ARG_TYPE) \
+template<>                                               \
+  EIGEN_DEVICE_FUNC                                      \
+  EIGEN_ALWAYS_INLINE RET_TYPE NAME(const ARG_TYPE& x) { \
+    return cl::sycl::FUNC(x);                            \
+  }
+
+#define SYCL_SPECIALIZE_UNARY_FUNC(NAME, FUNC, TYPE) \
+  SYCL_SPECIALIZE_GEN_UNARY_FUNC(NAME, FUNC, TYPE, TYPE)
+
+#define SYCL_SPECIALIZE_GEN1_BINARY_FUNC(NAME, FUNC, RET_TYPE, ARG_TYPE1, ARG_TYPE2) \
+  template<>                                                                  \
+  EIGEN_DEVICE_FUNC                                                           \
+  EIGEN_ALWAYS_INLINE RET_TYPE NAME(const ARG_TYPE1& x, const ARG_TYPE2& y) { \
+    return cl::sycl::FUNC(x, y);                                              \
+  }
+
+#define SYCL_SPECIALIZE_GEN2_BINARY_FUNC(NAME, FUNC, RET_TYPE, ARG_TYPE) \
+  SYCL_SPECIALIZE_GEN1_BINARY_FUNC(NAME, FUNC, RET_TYPE, ARG_TYPE, ARG_TYPE)
+
+#define SYCL_SPECIALIZE_BINARY_FUNC(NAME, FUNC, TYPE) \
+  SYCL_SPECIALIZE_GEN2_BINARY_FUNC(NAME, FUNC, TYPE, TYPE)
+
+SYCL_SPECIALIZE_INTEGER_TYPES_BINARY(mini, min)
+SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(mini, fmin)
+SYCL_SPECIALIZE_INTEGER_TYPES_BINARY(maxi, max)
+SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(maxi, fmax)
+
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
 
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
@@ -1109,6 +1104,10 @@ inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar&
   return EIGEN_MATHFUNC_IMPL(hypot, Scalar)::run(x, y);
 }
 
+#if defined(__SYCL_DEVICE_ONLY__)
+  SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(hypot, hypot)
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
@@ -1117,9 +1116,8 @@ inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   log1p(float x) { return cl::sycl::log1p(x); }
-EIGEN_ALWAYS_INLINE double  log1p(double x) { return cl::sycl::log1p(x); }
-#endif // defined(__SYCL_DEVICE_ONLY__)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(log1p, log1p)
+#endif //defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
@@ -1137,8 +1135,7 @@ inline typename internal::pow_impl<ScalarX,ScalarY>::result_type pow(const Scala
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   pow(float x, float y) { return cl::sycl::pow(x, y); }
-EIGEN_ALWAYS_INLINE double  pow(double x, double y) { return cl::sycl::pow(x, y); }
+SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(pow, pow)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 template<typename T> EIGEN_DEVICE_FUNC bool (isnan)   (const T &x) { return internal::isnan_impl(x); }
@@ -1146,12 +1143,9 @@ template<typename T> EIGEN_DEVICE_FUNC bool (isinf)   (const T &x) { return inte
 template<typename T> EIGEN_DEVICE_FUNC bool (isfinite)(const T &x) { return internal::isfinite_impl(x); }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   isnan(float x) { return cl::sycl::isnan(x); }
-EIGEN_ALWAYS_INLINE double  isnan(double x) { return cl::sycl::isnan(x); }
-EIGEN_ALWAYS_INLINE float   isinf(float x) { return cl::sycl::isinf(x); }
-EIGEN_ALWAYS_INLINE double  isinf(double x) { return cl::sycl::isinf(x); }
-EIGEN_ALWAYS_INLINE float   isfinite(float x) { return cl::sycl::isfinite(x); }
-EIGEN_ALWAYS_INLINE double  isfinite(double x) { return cl::sycl::isfinite(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY_FUNC_RET_TYPE(isnan, isnan, bool)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY_FUNC_RET_TYPE(isinf, isinf, bool)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY_FUNC_RET_TYPE(isfinite, isfinite, bool)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 template<typename Scalar>
@@ -1162,8 +1156,7 @@ inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   round(float x) { return cl::sycl::round(x); }
-EIGEN_ALWAYS_INLINE double  round(double x) { return cl::sycl::round(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(round, round)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 template<typename T>
@@ -1175,8 +1168,7 @@ T (floor)(const T& x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   floor(float x) { return cl::sycl::floor(x); }
-EIGEN_ALWAYS_INLINE double  floor(double x) { return cl::sycl::floor(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(floor, floor)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1196,8 +1188,7 @@ T (ceil)(const T& x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   ceil(float x) { return cl::sycl::ceil(x); }
-EIGEN_ALWAYS_INLINE double  ceil(double x) { return cl::sycl::ceil(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(ceil, ceil)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1241,8 +1232,7 @@ T sqrt(const T &x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   sqrt(float x) { return cl::sycl::sqrt(x); }
-EIGEN_ALWAYS_INLINE double  sqrt(double x) { return cl::sycl::sqrt(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(sqrt, sqrt)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 template<typename T>
@@ -1253,8 +1243,7 @@ T log(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   log(float x) { return cl::sycl::log(x); }
-EIGEN_ALWAYS_INLINE double  log(double x) { return cl::sycl::log(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(log, log)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 
@@ -1282,8 +1271,8 @@ abs(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
-EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
+SYCL_SPECIALIZE_INTEGER_TYPES_UNARY(abs, abs)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(abs, fabs)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1312,8 +1301,7 @@ T exp(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   exp(float x) { return cl::sycl::exp(x); }
-EIGEN_ALWAYS_INLINE double  exp(double x) { return cl::sycl::exp(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(exp, exp)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1348,8 +1336,7 @@ inline EIGEN_MATHFUNC_RETVAL(expm1, Scalar) expm1(const Scalar& x)
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   expm1(float x) { return cl::sycl::expm1(x); }
-EIGEN_ALWAYS_INLINE double  expm1(double x) { return cl::sycl::expm1(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(expm1, expm1)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1368,8 +1355,7 @@ T cos(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   cos(float x) { return cl::sycl::cos(x); }
-EIGEN_ALWAYS_INLINE double  cos(double x) { return cl::sycl::cos(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(cos,cos)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1388,8 +1374,7 @@ T sin(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   sin(float x) { return cl::sycl::sin(x); }
-EIGEN_ALWAYS_INLINE double  sin(double x) { return cl::sycl::sin(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(sin, sin)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1408,8 +1393,7 @@ T tan(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   tan(float x) { return cl::sycl::tan(x); }
-EIGEN_ALWAYS_INLINE double  tan(double x) { return cl::sycl::tan(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(tan, tan)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1437,10 +1421,8 @@ T acosh(const T &x) {
 #endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   acos(float x) { return cl::sycl::acos(x); }
-EIGEN_ALWAYS_INLINE double  acos(double x) { return cl::sycl::acos(x); }
-EIGEN_ALWAYS_INLINE float   acosh(float x) { return cl::sycl::acosh(x); }
-EIGEN_ALWAYS_INLINE double  acosh(double x) { return cl::sycl::acosh(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(acos, acos)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(acosh, acosh)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1468,10 +1450,8 @@ T asinh(const T &x) {
 #endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   asin(float x) { return cl::sycl::asin(x); }
-EIGEN_ALWAYS_INLINE double  asin(double x) { return cl::sycl::asin(x); }
-EIGEN_ALWAYS_INLINE float   asinh(float x) { return cl::sycl::asinh(x); }
-EIGEN_ALWAYS_INLINE double  asinh(double x) { return cl::sycl::asinh(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(asin, asin)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(asinh, asinh)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1499,10 +1479,8 @@ T atanh(const T &x) {
 #endif
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   atan(float x) { return cl::sycl::atan(x); }
-EIGEN_ALWAYS_INLINE double  atan(double x) { return cl::sycl::atan(x); }
-EIGEN_ALWAYS_INLINE float   atanh(float x) { return cl::sycl::atanh(x); }
-EIGEN_ALWAYS_INLINE double  atanh(double x) { return cl::sycl::atanh(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(atan, atan)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(atanh, atanh)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1522,8 +1500,7 @@ T cosh(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   cosh(float x) { return cl::sycl::cosh(x); }
-EIGEN_ALWAYS_INLINE double  cosh(double x) { return cl::sycl::cosh(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(cosh, cosh)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1542,8 +1519,7 @@ T sinh(const T &x) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   sinh(float x) { return cl::sycl::sinh(x); }
-EIGEN_ALWAYS_INLINE double  sinh(double x) { return cl::sycl::sinh(x); }
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(sinh, sinh)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1561,14 +1537,15 @@ T tanh(const T &x) {
   return tanh(x);
 }
 
-#if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   tanh(float x) { return cl::sycl::tanh(x); }
-EIGEN_ALWAYS_INLINE double  tanh(double x) { return cl::sycl::tanh(x); }
-#elif (!defined(EIGEN_GPUCC)) && EIGEN_FAST_MATH
+#if (!defined(EIGEN_GPUCC)) && EIGEN_FAST_MATH && (!defined(__SYCL_DEVICE_ONLY__))
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(float x) { return internal::generic_fast_tanh_float(x); }
 #endif
 
+#if defined(__SYCL_DEVICE_ONLY__)
+SYCL_SPECIALIZE_FLOATING_TYPES_UNARY(tanh, tanh)
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
 #if defined(EIGEN_GPUCC)
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(const float &x) { return ::tanhf(x); }
@@ -1585,8 +1562,7 @@ T fmod(const T& a, const T& b) {
 }
 
 #if defined(__SYCL_DEVICE_ONLY__)
-EIGEN_ALWAYS_INLINE float   fmod(float x, float y) { return cl::sycl::fmod(x, y); }
-EIGEN_ALWAYS_INLINE double  fmod(double x, double y) { return cl::sycl::fmod(x, y); }
+SYCL_SPECIALIZE_FLOATING_TYPES_BINARY(fmod, fmod)
 #endif // defined(__SYCL_DEVICE_ONLY__)
 
 #if defined(EIGEN_GPUCC)
@@ -1603,6 +1579,23 @@ double fmod(const double& a, const double& b) {
 }
 #endif
 
+#if defined(__SYCL_DEVICE_ONLY__)
+#undef SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_BINARY
+#undef SYCL_SPECIALIZE_SIGNED_INTEGER_TYPES_UNARY
+#undef SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_BINARY
+#undef SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_UNARY
+#undef SYCL_SPECIALIZE_INTEGER_TYPES_BINARY
+#undef SYCL_SPECIALIZE_UNSIGNED_INTEGER_TYPES_UNARY
+#undef SYCL_SPECIALIZE_FLOATING_TYPES_BINARY
+#undef SYCL_SPECIALIZE_FLOATING_TYPES_UNARY
+#undef SYCL_SPECIALIZE_FLOATING_TYPES_UNARY_FUNC_RET_TYPE
+#undef SYCL_SPECIALIZE_GEN_UNARY_FUNC
+#undef SYCL_SPECIALIZE_UNARY_FUNC
+#undef SYCL_SPECIALIZE_GEN1_BINARY_FUNC
+#undef SYCL_SPECIALIZE_GEN2_BINARY_FUNC
+#undef SYCL_SPECIALIZE_BINARY_FUNC
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
 } // end namespace numext
 
 namespace internal {

From 946c3e2544da35ed0539e433cc906c93c9bda3ed Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Mon, 27 Aug 2018 23:04:08 +0000
Subject: [PATCH 615/680] adding EIGEN_DEVICE_FUNC attribute to fix some GPU
 unit tests that are broken in HIP mode

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 4cb49e9b8..9b92b6179 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -79,7 +79,7 @@ struct TensorOpResourceRequirements {
 };
 
 // Tries to merge multiple resource requirements.
-EIGEN_STRONG_INLINE void MergeResourceRequirements(
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void MergeResourceRequirements(
     const std::vector<TensorOpResourceRequirements>& resources,
     TensorBlockShapeType* block_shape, Index* block_total_size) {
   if (resources.empty()) {

From ddbc5643860d6033fdca88145d2427ae52716a9d Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 30 Aug 2018 16:33:03 +0200
Subject: [PATCH 616/680] Fixed a few more shadowing warnings when compiling
 with g++ (and c++03)

---
 Eigen/src/StlSupport/StdDeque.h  | 2 +-
 Eigen/src/StlSupport/StdList.h   | 2 +-
 Eigen/src/StlSupport/StdVector.h | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/StlSupport/StdDeque.h b/Eigen/src/StlSupport/StdDeque.h
index 8adfbb5c1..045da7b4d 100644
--- a/Eigen/src/StlSupport/StdDeque.h
+++ b/Eigen/src/StlSupport/StdDeque.h
@@ -62,7 +62,7 @@ namespace std {
     : deque_base(first, last, a) {} \
     deque(const deque& c) : deque_base(c) {}  \
     explicit deque(size_type num, const value_type& val = value_type()) : deque_base(num, val) {} \
-    deque(iterator start, iterator end) : deque_base(start, end) {}  \
+    deque(iterator start_, iterator end_) : deque_base(start_, end_) {}  \
     deque& operator=(const deque& x) {  \
       deque_base::operator=(x);  \
       return *this;  \
diff --git a/Eigen/src/StlSupport/StdList.h b/Eigen/src/StlSupport/StdList.h
index 1b6d44061..8ba3fada0 100644
--- a/Eigen/src/StlSupport/StdList.h
+++ b/Eigen/src/StlSupport/StdList.h
@@ -62,7 +62,7 @@ namespace std
     : list_base(first, last, a) {} \
     list(const list& c) : list_base(c) {}  \
     explicit list(size_type num, const value_type& val = value_type()) : list_base(num, val) {} \
-    list(iterator start, iterator end) : list_base(start, end) {}  \
+    list(iterator start_, iterator end_) : list_base(start_, end_) {}  \
     list& operator=(const list& x) {  \
     list_base::operator=(x);  \
     return *this; \
diff --git a/Eigen/src/StlSupport/StdVector.h b/Eigen/src/StlSupport/StdVector.h
index 7787edd72..9fcf19bce 100644
--- a/Eigen/src/StlSupport/StdVector.h
+++ b/Eigen/src/StlSupport/StdVector.h
@@ -62,7 +62,7 @@ namespace std {
     : vector_base(first, last, a) {} \
     vector(const vector& c) : vector_base(c) {}  \
     explicit vector(size_type num, const value_type& val = value_type()) : vector_base(num, val) {} \
-    vector(iterator start, iterator end) : vector_base(start, end) {}  \
+    vector(iterator start_, iterator end_) : vector_base(start_, end_) {}  \
     vector& operator=(const vector& x) {  \
       vector_base::operator=(x);  \
       return *this;  \

From c2f4e8c08e1fc74f426f21ba4e05decaa7dbf53b Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 30 Aug 2018 17:12:53 +0200
Subject: [PATCH 617/680] Fix integer conversion warning

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index ce91bc2a6..7c26b1682 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -194,7 +194,7 @@ template <std::ptrdiff_t V1=0, std::ptrdiff_t V2=0, std::ptrdiff_t V3=0, std::pt
   }
 #endif
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex operator[] (const int index) const {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index operator[] (const Index index) const {
     switch (index) {
       case 0:
         return internal::get<0, Base>::value;
@@ -208,7 +208,7 @@ template <std::ptrdiff_t V1=0, std::ptrdiff_t V2=0, std::ptrdiff_t V3=0, std::pt
         return internal::get<4, Base>::value;
       default:
         eigen_assert(false && "index overflow");
-        return static_cast<DenseIndex>(-1);
+        return static_cast<Index>(-1);
     }
   }
 

From 023ed6b9a83b89bdba7c4b689bc9e69acd1af79f Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 30 Aug 2018 17:14:52 +0200
Subject: [PATCH 618/680] Product of empty array must be 1 and not 0.

---
 unsupported/Eigen/CXX11/src/util/EmulateCXX11Meta.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/util/EmulateCXX11Meta.h
index f3aa1b144..8a536faf6 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateCXX11Meta.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateCXX11Meta.h
@@ -188,7 +188,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array<t, n>& a) {
 }
 template<typename t>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array<t, 0>& /*a*/) {
-  return 0;
+  return 1;
 }
 
 template<typename t>

From ff4e835d6b974ff70bdf12288e58bc073eefe4fb Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 30 Aug 2018 20:10:11 +0200
Subject: [PATCH 619/680] "sparse_product.cpp" must be included before
 "sparse_basic.cpp", otherwise EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN has no
 effect

---
 unsupported/test/sparse_extra.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/test/sparse_extra.cpp b/unsupported/test/sparse_extra.cpp
index 4ac53a9a7..51ba83b13 100644
--- a/unsupported/test/sparse_extra.cpp
+++ b/unsupported/test/sparse_extra.cpp
@@ -10,8 +10,8 @@
 
 // import basic and product tests for deprecated DynamicSparseMatrix
 #define EIGEN_NO_DEPRECATED_WARNING
-#include "sparse_basic.cpp"
 #include "sparse_product.cpp"
+#include "sparse_basic.cpp"
 #include <Eigen/SparseExtra>
 
 template<typename SetterType,typename DenseType, typename Scalar, int Options>

From ba2c8efdcfe8f3b58797080a412410423d223765 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Wed, 12 Sep 2018 11:49:10 +0200
Subject: [PATCH 620/680] EIGEN_UNUSED is not supported by g++4.7 (and not
 portable)

---
 unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
index 13d959df4..2db914765 100755
--- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
+++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h
@@ -534,7 +534,8 @@ struct ScalarBinaryOpTraits<typename DerType::Scalar,AutoDiffScalar<DerType>, Bi
   EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(typename Eigen::internal::remove_all<DerType>::type, typename Eigen::internal::traits<typename Eigen::internal::remove_all<DerType>::type>::Scalar, product) > \
   FUNC(const Eigen::AutoDiffScalar<DerType>& x) { \
     using namespace Eigen; \
-    EIGEN_UNUSED typedef typename Eigen::internal::traits<typename Eigen::internal::remove_all<DerType>::type>::Scalar Scalar; \
+    typedef typename Eigen::internal::traits<typename Eigen::internal::remove_all<DerType>::type>::Scalar Scalar; \
+    EIGEN_UNUSED_VARIABLE(sizeof(Scalar)); \
     CODE; \
   }
 

From 7e9c9fbb2d9b7e3d3467367913be078e0aaa322a Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Wed, 12 Sep 2018 14:40:39 +0200
Subject: [PATCH 621/680] Disable type-limits warnings for g++ < 4.8

---
 Eigen/src/Core/util/DisableStupidWarnings.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index d04b52649..6e93bbc0f 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -52,6 +52,10 @@
   #endif
   // g++ warns about local variables shadowing member functions, which is too strict
   #pragma GCC diagnostic ignored "-Wshadow"
+  #if __GNUC__ == 4 && __GNUC_MINOR__ < 8
+    // Until g++-4.7 there are warnings when comparing unsigned int vs 0, even in templated functions:
+    #pragma GCC diagnostic ignored "-Wtype-limits"
+  #endif
   #if __GNUC__>=6
     #pragma GCC diagnostic ignored "-Wignored-attributes"
   #endif

From 3adece4827c1a11a6e2c4ba95e6fddc682a145d2 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Wed, 12 Sep 2018 14:41:38 +0200
Subject: [PATCH 622/680] Fix misleading indentation of errorCode and make it
 loop-local

---
 Eigen/src/UmfPackSupport/UmfPackSupport.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h
index ba10a9318..e3a333f80 100644
--- a/Eigen/src/UmfPackSupport/UmfPackSupport.h
+++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -613,7 +613,6 @@ bool UmfPackLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBas
   eigen_assert((XDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major result yet");
   eigen_assert(b.derived().data() != x.derived().data() && " Umfpack does not support inplace solve");
 
-  StorageIndex errorCode;
   Scalar* x_ptr = 0;
   Matrix<Scalar,Dynamic,1> x_tmp;
   if(x.innerStride()!=1)
@@ -625,7 +624,7 @@ bool UmfPackLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBas
   {
     if(x.innerStride()==1)
       x_ptr = &x.col(j).coeffRef(0);
-      errorCode = umfpack_solve(UMFPACK_A,
+    StorageIndex errorCode = umfpack_solve(UMFPACK_A,
                                 mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
                                 x_ptr, &b.const_cast_derived().col(j).coeffRef(0),
                                 m_numeric, m_control.data(), m_umfpackInfo.data());

From 5927eef6127aae89abcc7f98e3a7ba66f6bdf56e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 13 Sep 2018 09:44:46 +0200
Subject: [PATCH 623/680] Enable std::result_of for msvc 2015 and later

---
 Eigen/src/Core/util/Macros.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index bcdede61e..3af6c4e37 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -533,8 +533,11 @@
 #endif
 
 // Does the compiler support result_of?
+// It's likely that MSVC 2013 supports result_of but I couldn't not find a good source for that,
+// so let's be conservative.
 #ifndef EIGEN_HAS_STD_RESULT_OF
-#if EIGEN_MAX_CPP_VER>=11 && ((__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L)))
+#if EIGEN_MAX_CPP_VER>=11 && \
+    (__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900)
 #define EIGEN_HAS_STD_RESULT_OF 1
 #else
 #define EIGEN_HAS_STD_RESULT_OF 0

From 8b3d9ed081fc5d4870290649853b19cb5179546e Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 5 Sep 2018 11:20:06 -0700
Subject: [PATCH 624/680] Use padding instead of alignment attribute, which
 MaxSizeVector does not respect. This leads to undefined behavior and
 hard-to-trace bugs.

---
 unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
index 22c952ae1..2cac2d0f1 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
@@ -169,9 +169,7 @@ class EventCount {
 
   class Waiter {
     friend class EventCount;
-    // Align to 128 byte boundary to prevent false sharing with other Waiter
-    // objects in the same vector.
-    EIGEN_ALIGN_TO_BOUNDARY(128) std::atomic<Waiter*> next;
+    std::atomic<Waiter*> next;
     std::mutex mu;
     std::condition_variable cv;
     uint64_t epoch;
@@ -181,6 +179,9 @@ class EventCount {
       kWaiting,
       kSignaled,
     };
+    // Pad past 128 byte boundary to prevent false sharing with other Waiter
+    // objects in the same vector.
+    char pad_[128];
   };
 
  private:

From c64fe9ea1f9f5943864cd9ca27d3fcca07453f82 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Thu, 30 Aug 2018 20:22:16 +0000
Subject: [PATCH 625/680] Updates to fix HIP-clang specific compile errors.

Compiling the eigen unittests with hip-clang (HIP with clang as the underlying compiler instead of hcc or nvcc), results in compile errors. The changes in this commit fix those compile errors. The main change is to convert a few instances of "__device__" to "EIGEN_DEVICE_FUNC"
---
 Eigen/src/Core/arch/GPU/PacketMathHalf.h      | 74 +++++++++----------
 test/gpu_common.h                             |  6 +-
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  |  2 +-
 3 files changed, 41 insertions(+), 41 deletions(-)

diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index b0a72e1f9..c4feda87d 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -43,7 +43,7 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits
 
 template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
 
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
@@ -58,29 +58,29 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half&
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
   return *reinterpret_cast<const half2*>(from);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
   return __halves2half2(from[0], from[1]);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
   return __halves2half2(from[0], from[0]);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
   *reinterpret_cast<half2*>(to) = from;
 }
 
-template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
   to[0] = __low2half(from);
   to[1] = __high2half(from);
 }
 
 template<>
- __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
 
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
@@ -102,7 +102,7 @@ template<>
 }
 
 template<>
-__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
 
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
@@ -123,20 +123,20 @@ __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::ha
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
   return __halves2half2(from[0*stride], from[1*stride]);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
   to[stride*0] = __low2half(from);
   to[stride*1] = __high2half(from);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
   return __low2half(a);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
   half2 result;
   unsigned temp = *(reinterpret_cast<const unsigned*>(&(a)));
   *(reinterpret_cast<unsigned*>(&(result))) = temp & 0x7FFF7FFF;
@@ -144,7 +144,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
 }
 
 
-__device__ EIGEN_STRONG_INLINE void
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<half2,2>& kernel) {
   __half a1 = __low2half(kernel.packet[0]);
   __half a2 = __high2half(kernel.packet[0]);
@@ -154,7 +154,7 @@ ptranspose(PacketBlock<half2,2>& kernel) {
   kernel.packet[1] = __halves2half2(a2, b2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
   
   return __halves2half2(a, __hadd(a, __float2half(1.0f)));
@@ -171,7 +171,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half&
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hadd2(a, b);
@@ -193,7 +193,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, cons
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hsub2(a, b);
@@ -215,7 +215,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, cons
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hneg2(a);
@@ -233,9 +233,9 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hmul2(a, b);
@@ -257,7 +257,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, cons
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
    return __hfma2(a, b, c);
@@ -281,7 +281,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, con
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
   
 #if defined(EIGEN_HAS_OLD_HIP_FP16)
@@ -303,7 +303,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, cons
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -313,7 +313,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, cons
   return __halves2half2(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -323,7 +323,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, cons
   return __halves2half2(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hadd(__low2half(a), __high2half(a));
@@ -341,7 +341,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   __half first = __low2half(a);
@@ -363,7 +363,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const ha
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   __half first = __low2half(a);
@@ -385,7 +385,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const ha
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
 #if defined(EIGEN_HIP_DEVICE_COMPILE)
 
   return __hmul(__low2half(a), __high2half(a));
@@ -403,7 +403,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const ha
 #endif
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = log1pf(a1);
@@ -411,7 +411,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = expm1f(a1);
@@ -422,29 +422,29 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
 #if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
   defined(EIGEN_HIP_DEVICE_COMPILE)
 
-template<>  __device__ EIGEN_STRONG_INLINE
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 half2 plog<half2>(const half2& a) {
   return h2log(a);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 half2 pexp<half2>(const half2& a) {
   return h2exp(a);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 half2 psqrt<half2>(const half2& a) {
   return h2sqrt(a);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 half2 prsqrt<half2>(const half2& a) {
   return h2rsqrt(a);
 }
 
 #else
 
-template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = logf(a1);
@@ -452,7 +452,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = expf(a1);
@@ -460,7 +460,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = sqrtf(a1);
@@ -468,7 +468,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
   return __floats2half2_rn(r1, r2);
 }
 
-template<> __device__ EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = rsqrtf(a1);
diff --git a/test/gpu_common.h b/test/gpu_common.h
index 3aac49e96..79d4ea694 100644
--- a/test/gpu_common.h
+++ b/test/gpu_common.h
@@ -61,9 +61,9 @@ void run_on_gpu(const Kernel& ker, int n, const Input& in, Output& out)
   gpuDeviceSynchronize();
   
 #ifdef EIGEN_USE_HIP
-  hipLaunchKernelGGL(run_on_gpu_meta_kernel<Kernel,
-		     typename std::decay<decltype(*d_in)>::type,
-		     typename std::decay<decltype(*d_out)>::type>, 
+  hipLaunchKernelGGL(HIP_KERNEL_NAME(run_on_gpu_meta_kernel<Kernel,
+				     typename std::decay<decltype(*d_in)>::type,
+				     typename std::decay<decltype(*d_out)>::type>), 
 		     dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
 #else
   run_on_gpu_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 01d3863da..f3f1640b0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -14,7 +14,7 @@
 // clang is incompatible with the CUDA syntax wrt making a kernel a class friend,
 // so we'll use a macro to make clang happy.
 #ifndef KERNEL_FRIEND
-#if defined(__clang__) && defined(__CUDA__)
+#if defined(__clang__) && (defined(__CUDA__) || defined(__HIP__))
 #define KERNEL_FRIEND friend __global__
 #else
 #define KERNEL_FRIEND friend

From edeee16a16373dbd86ff0a1129e3d94e3c7e93a7 Mon Sep 17 00:00:00 2001
From: Alexey Frunze <afrunze@wavecomp.com>
Date: Fri, 31 Aug 2018 14:11:10 -0700
Subject: [PATCH 626/680] Fix build failures in matrix_power and
 matrix_exponential tests.

This fixes the static assertion complaining about double being
used in place of long double. This happened on MIPS32, where
double and long double have the same type representation.
This can be simulated on x86 as well if we pass -mlong-double-64
to g++.
---
 unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
index 54037d58d..8aebcd67a 100644
--- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
+++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h
@@ -234,12 +234,13 @@ struct matrix_exp_computeUV<MatrixType, float>
 template <typename MatrixType>
 struct matrix_exp_computeUV<MatrixType, double>
 {
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
   template <typename ArgType>
   static void run(const ArgType& arg, MatrixType& U, MatrixType& V, int& squarings)
   {
     using std::frexp;
     using std::pow;
-    const double l1norm = arg.cwiseAbs().colwise().sum().maxCoeff();
+    const RealScalar l1norm = arg.cwiseAbs().colwise().sum().maxCoeff();
     squarings = 0;
     if (l1norm < 1.495585217958292e-002) {
       matrix_exp_pade3(arg, U, V);
@@ -250,10 +251,10 @@ struct matrix_exp_computeUV<MatrixType, double>
     } else if (l1norm < 2.097847961257068e+000) {
       matrix_exp_pade9(arg, U, V);
     } else {
-      const double maxnorm = 5.371920351148152;
+      const RealScalar maxnorm = 5.371920351148152;
       frexp(l1norm / maxnorm, &squarings);
       if (squarings < 0) squarings = 0;
-      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<double>(squarings));
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<RealScalar>(squarings));
       matrix_exp_pade13(A, U, V);
     }
   }

From c6066ac4115cafd00db7e02a179957290978142c Mon Sep 17 00:00:00 2001
From: cgs1019 <cgs1019@gmail.com>
Date: Thu, 6 Sep 2018 11:04:17 -0400
Subject: [PATCH 627/680] Make param name and docs constistent for
 JacobiRotation::makeGivens

Previously the rendered math in the doc string called the optional return value
'r', while the actual parameter and the doc string text referred to the
parameter as 'z'. This changeset renames all the z's to r's to match the math.
---
 Eigen/src/Jacobi/Jacobi.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h
index 176ff9d07..4ccd49a04 100644
--- a/Eigen/src/Jacobi/Jacobi.h
+++ b/Eigen/src/Jacobi/Jacobi.h
@@ -73,13 +73,13 @@ template<typename Scalar> class JacobiRotation
     bool makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z);
 
     EIGEN_DEVICE_FUNC
-    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z=0);
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* r=0);
 
   protected:
     EIGEN_DEVICE_FUNC
-    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::true_type);
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::true_type);
     EIGEN_DEVICE_FUNC
-    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::false_type);
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::false_type);
 
     Scalar m_c, m_s;
 };
@@ -145,7 +145,7 @@ inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, Ind
   * \f$ V = \left ( \begin{array}{c} p \\ q \end{array} \right )\f$ yields:
   * \f$ G^* V = \left ( \begin{array}{c} r \\ 0 \end{array} \right )\f$.
   *
-  * The value of \a z is returned if \a z is not null (the default is null).
+  * The value of \a r is returned if \a r is not null (the default is null).
   * Also note that G is built such that the cosine is always real.
   *
   * Example: \include Jacobi_makeGivens.cpp
@@ -159,9 +159,9 @@ inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, Ind
   */
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
-void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* z)
+void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r)
 {
-  makeGivens(p, q, z, typename internal::conditional<NumTraits<Scalar>::IsComplex, internal::true_type, internal::false_type>::type());
+  makeGivens(p, q, r, typename internal::conditional<NumTraits<Scalar>::IsComplex, internal::true_type, internal::false_type>::type());
 }
 
 

From 3d057e0453d7f76330196730abe45806e7e48b5e Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 6 Sep 2018 12:59:36 -0700
Subject: [PATCH 628/680] Avoid compilation error in C++11 test when
 EIGEN_AVOID_STL_ARRAY is set.

---
 unsupported/test/cxx11_tensor_convolution.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp
index 01bc77bc1..78704e5e5 100644
--- a/unsupported/test/cxx11_tensor_convolution.cpp
+++ b/unsupported/test/cxx11_tensor_convolution.cpp
@@ -25,7 +25,7 @@ static void test_evals()
 
   Tensor<float, 2, DataLayout> result(2,3);
   result.setZero();
-  Eigen::array<Tensor<float, 2>::Index, 1> dims3{{0}};
+  Eigen::array<Tensor<float, 2>::Index, 1> dims3({0});
 
   typedef TensorEvaluator<decltype(input.convolve(kernel, dims3)), DefaultDevice> Evaluator;
   Evaluator eval(input.convolve(kernel, dims3), DefaultDevice());

From 4827bec7768e5b58fb2e001d6c1360373d6dd0c2 Mon Sep 17 00:00:00 2001
From: Justin Carpentier <jcarpent@laas.fr>
Date: Tue, 11 Sep 2018 09:33:21 +0200
Subject: [PATCH 629/680] LLT: correct doc and add missing reference for the
 return type of rankUpdate ---  Eigen/src/Cholesky/LLT.h | 2 +-  1 file
 changed, 1 insertion(+), 1 deletion(-)

---
 Eigen/src/Cholesky/LLT.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h
index 22e4be75d..868766365 100644
--- a/Eigen/src/Cholesky/LLT.h
+++ b/Eigen/src/Cholesky/LLT.h
@@ -100,7 +100,7 @@ template<typename _MatrixType, int _UpLo> class LLT
       compute(matrix.derived());
     }
 
-    /** \brief Constructs a LDLT factorization from a given matrix
+    /** \brief Constructs a LLT factorization from a given matrix
       *
       * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when
       * \c MatrixType is a Eigen::Ref.
@@ -200,7 +200,7 @@ template<typename _MatrixType, int _UpLo> class LLT
     inline Index cols() const { return m_matrix.cols(); }
 
     template<typename VectorType>
-    LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
+    LLT & rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
@@ -458,7 +458,7 @@ LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>
   */
 template<typename _MatrixType, int _UpLo>
 template<typename VectorType>
-LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, const RealScalar& sigma)
+LLT<_MatrixType,_UpLo> & LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, const RealScalar& sigma)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorType);
   eigen_assert(v.size()==m_matrix.cols());

From 46f88fc454e78484ebdf9d58990d0489c1103cf4 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Tue, 11 Sep 2018 10:08:10 -0700
Subject: [PATCH 630/680] Use numerically stable tree reduction in
 TensorReduction.

---
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h   | 68 ++++++-------
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 96 ++++++++++++++++---
 .../CXX11/src/Tensor/TensorReductionGpu.h     | 10 +-
 unsupported/test/cxx11_tensor_reduction.cpp   | 28 +++++-
 4 files changed, 141 insertions(+), 61 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index cd666c173..9fd276075 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -58,16 +58,15 @@ template<typename Reducer, typename Device>
 struct reducer_traits {
   enum {
     Cost = 1,
-    PacketAccess = false
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 // Standard reduction functors
 template <typename T> struct SumReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasAdd;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
     internal::scalar_sum_op<T> sum_op;
     *accum = sum_op(*accum, t);
@@ -103,16 +102,14 @@ template <typename T, typename Device>
 struct reducer_traits<SumReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = PacketType<T, Device>::HasAdd
+    PacketAccess = PacketType<T, Device>::HasAdd,
+    IsStateful = false,
+    IsExactlyAssociative = NumTraits<T>::IsInteger
   };
 };
 
-
 template <typename T> struct MeanReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv && !NumTraits<T>::IsInteger;
-  static const bool IsStateful = true;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   MeanReducer() : scalarCount_(0), packetCount_(0) { }
 
@@ -161,7 +158,9 @@ template <typename T, typename Device>
 struct reducer_traits<MeanReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = PacketType<T, Device>::HasAdd
+    PacketAccess = PacketType<T, Device>::HasAdd && !NumTraits<T>::IsInteger,
+    IsStateful = true,
+    IsExactlyAssociative = NumTraits<T>::IsInteger
   };
 };
 
@@ -194,9 +193,6 @@ struct MinMaxBottomValue<T, false, false> {
 
 template <typename T> struct MaxReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasMax;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
     if (t > *accum) { *accum = t; }
   }
@@ -228,16 +224,15 @@ template <typename T, typename Device>
 struct reducer_traits<MaxReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = PacketType<T, Device>::HasMax
+    PacketAccess = PacketType<T, Device>::HasMax,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 
 template <typename T> struct MinReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasMin;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
     if (t < *accum) { *accum = t; }
   }
@@ -269,16 +264,15 @@ template <typename T, typename Device>
 struct reducer_traits<MinReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = PacketType<T, Device>::HasMin
+    PacketAccess = PacketType<T, Device>::HasMin,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 
 template <typename T> struct ProdReducer
 {
-  static const bool PacketAccess = packet_traits<T>::HasMul;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
     internal::scalar_product_op<T> prod_op;
     (*accum) = prod_op(*accum, t);
@@ -314,16 +308,15 @@ template <typename T, typename Device>
 struct reducer_traits<ProdReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::MulCost,
-    PacketAccess = PacketType<T, Device>::HasMul
+    PacketAccess = PacketType<T, Device>::HasMul,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 
 struct AndReducer
 {
-  static const bool PacketAccess = false;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(bool t, bool* accum) const {
     *accum = *accum && t;
   }
@@ -339,15 +332,14 @@ template <typename Device>
 struct reducer_traits<AndReducer, Device> {
   enum {
     Cost = 1,
-    PacketAccess = false
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 
 struct OrReducer {
-  static const bool PacketAccess = false;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(bool t, bool* accum) const {
     *accum = *accum || t;
   }
@@ -363,7 +355,9 @@ template <typename Device>
 struct reducer_traits<OrReducer, Device> {
   enum {
     Cost = 1,
-    PacketAccess = false
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
@@ -371,9 +365,6 @@ struct reducer_traits<OrReducer, Device> {
 // Argmin/Argmax reducers
 template <typename T> struct ArgMaxTupleReducer
 {
-  static const bool PacketAccess = false;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
     if (t.second > accum->second) { *accum = t; }
   }
@@ -389,16 +380,15 @@ template <typename T, typename Device>
 struct reducer_traits<ArgMaxTupleReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = false
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
 
 template <typename T> struct ArgMinTupleReducer
 {
-  static const bool PacketAccess = false;
-  static const bool IsStateful = false;
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T& t, T* accum) const {
     if (t.second < accum->second) { *accum = t; }
   }
@@ -414,7 +404,9 @@ template <typename T, typename Device>
 struct reducer_traits<ArgMinTupleReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = false
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
   };
 };
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 01d3863da..d7a04a525 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -165,7 +165,9 @@ struct GenericDimReducer<-1, Self, Op> {
   }
 };
 
-template <typename Self, typename Op, bool Vectorizable = (Self::InputPacketAccess & Op::PacketAccess)>
+template <typename Self, typename Op, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess),
+          bool UseTreeReduction = (!Self::ReducerTraits::IsStateful &&
+                                   !Self::ReducerTraits::IsExactlyAssociative)>
 struct InnerMostDimReducer {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) {
     typename Self::CoeffReturnType accum = reducer.initialize();
@@ -177,23 +179,88 @@ struct InnerMostDimReducer {
 };
 
 template <typename Self, typename Op>
-struct InnerMostDimReducer<Self, Op, true> {
+struct InnerMostDimReducer<Self, Op, true, false> {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) {
-    const int packetSize = internal::unpacket_traits<typename Self::PacketReturnType>::size;
+    const typename Self::Index packetSize = internal::unpacket_traits<typename Self::PacketReturnType>::size;
     const typename Self::Index VectorizedSize = (numValuesToReduce / packetSize) * packetSize;
-    typename Self::PacketReturnType p = reducer.template initializePacket<typename Self::PacketReturnType>();
+    typename Self::PacketReturnType paccum = reducer.template initializePacket<typename Self::PacketReturnType>();
     for (typename Self::Index j = 0; j < VectorizedSize; j += packetSize) {
-      reducer.reducePacket(self.m_impl.template packet<Unaligned>(firstIndex + j), &p);
+      reducer.reducePacket(self.m_impl.template packet<Unaligned>(firstIndex + j), &paccum);
     }
     typename Self::CoeffReturnType accum = reducer.initialize();
     for (typename Self::Index j = VectorizedSize; j < numValuesToReduce; ++j) {
       reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
     }
-    return reducer.finalizeBoth(accum, p);
+    return reducer.finalizeBoth(accum, paccum);
   }
 };
 
-template <int DimIndex, typename Self, typename Op, bool vectorizable = (Self::InputPacketAccess & Op::PacketAccess)>
+static const int kLeafSize = 1024;
+
+template <typename Self, typename Op>
+struct InnerMostDimReducer<Self, Op, false, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType
+  reduce(const Self& self, typename Self::Index firstIndex,
+         typename Self::Index numValuesToReduce, Op& reducer) {
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    if (numValuesToReduce > kLeafSize) {
+      const typename Self::Index half = numValuesToReduce / 2;
+      reducer.reduce(reduce(self, firstIndex, half, reducer), &accum);
+      reducer.reduce(
+          reduce(self, firstIndex + half, numValuesToReduce - half, reducer),
+          &accum);
+    } else {
+      for (typename Self::Index j = 0; j < numValuesToReduce; ++j) {
+        reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+      }
+    }
+    return reducer.finalize(accum);
+  }
+};
+
+#if !defined(EIGEN_USE_GPU) || !defined(__CUDACC__)
+template <typename Self, typename Op>
+struct InnerMostDimReducer<Self, Op, true, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType
+  reduce(const Self& self, typename Self::Index firstIndex,
+         typename Self::Index numValuesToReduce, Op& reducer) {
+    const typename Self::Index packetSize =
+        internal::unpacket_traits<typename Self::PacketReturnType>::size;
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    if (numValuesToReduce > packetSize * kLeafSize) {
+      // Make sure the split point is aligned on a packet boundary.
+      const typename Self::Index split =
+          packetSize *
+          divup(firstIndex + divup(numValuesToReduce, typename Self::Index(2)),
+                packetSize);
+      const typename Self::Index num_left =
+          numext::mini(split - firstIndex, numValuesToReduce);
+      reducer.reduce(reduce(self, firstIndex, num_left, reducer), &accum);
+      if (num_left < numValuesToReduce) {
+        reducer.reduce(
+            reduce(self, split, numValuesToReduce - num_left, reducer), &accum);
+      }
+      return reducer.finalize(accum);
+    } else {
+      const typename Self::Index VectorizedSize =
+          (numValuesToReduce / packetSize) * packetSize;
+      typename Self::PacketReturnType paccum =
+          reducer.template initializePacket<typename Self::PacketReturnType>();
+      for (typename Self::Index j = 0; j < VectorizedSize; j += packetSize) {
+        reducer.reducePacket(
+            self.m_impl.template packet<Unaligned>(firstIndex + j), &paccum);
+      }
+      for (typename Self::Index j = VectorizedSize; j < numValuesToReduce;
+           ++j) {
+        reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+      }
+      return reducer.finalizeBoth(accum, paccum);
+    }
+  }
+};
+#endif
+
+template <int DimIndex, typename Self, typename Op, bool vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
 struct InnerMostDimPreserver {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self&, typename Self::Index, Op&, typename Self::PacketReturnType*) {
     eigen_assert(false && "should never be called");
@@ -228,7 +295,7 @@ struct InnerMostDimPreserver<-1, Self, Op, true> {
 };
 
 // Default full reducer
-template <typename Self, typename Op, typename Device, bool Vectorizable = (Self::InputPacketAccess & Op::PacketAccess)>
+template <typename Self, typename Op, typename Device, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
 struct FullReducer {
   static const bool HasOptimizedImplementation = false;
 
@@ -242,7 +309,7 @@ struct FullReducer {
 #ifdef EIGEN_USE_THREADS
 // Multithreaded full reducers
 template <typename Self, typename Op,
-          bool Vectorizable = (Self::InputPacketAccess & Op::PacketAccess)>
+          bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
 struct FullReducerShard {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(const Self& self, typename Self::Index firstIndex,
                   typename Self::Index numValuesToReduce, Op& reducer,
@@ -255,8 +322,8 @@ struct FullReducerShard {
 // Multithreaded full reducer
 template <typename Self, typename Op, bool Vectorizable>
 struct FullReducer<Self, Op, ThreadPoolDevice, Vectorizable> {
-  static const bool HasOptimizedImplementation = !Op::IsStateful;
-  static const int PacketSize =
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful;
+  static const Index PacketSize =
       unpacket_traits<typename Self::PacketReturnType>::size;
 
   // launch one reducer per thread and accumulate the result.
@@ -394,6 +461,7 @@ class TensorReductionOp : public TensorBase<TensorReductionOp<Op, Dims, XprType,
 template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>
 struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>
 {
+  typedef internal::reducer_traits<Op, Device> ReducerTraits;
   typedef TensorReductionOp<Op, Dims, ArgType, MakePointer_> XprType;
   typedef typename XprType::Index Index;
   typedef ArgType ChildType;
@@ -407,11 +475,11 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
   static const bool InputPacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess;
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
-  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  static const Index PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
     IsAligned = false,
-    PacketAccess = Self::InputPacketAccess && Op::PacketAccess,
+    PacketAccess = Self::InputPacketAccess && ReducerTraits::PacketAccess,
     BlockAccess = false,
     Layout = TensorEvaluator<ArgType, Device>::Layout,
     CoordAccess = false,  // to be implemented
@@ -696,7 +764,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   private:
   template <int, typename, typename> friend struct internal::GenericDimReducer;
-  template <typename, typename, bool> friend struct internal::InnerMostDimReducer;
+  template <typename, typename, bool, bool> friend struct internal::InnerMostDimReducer;
   template <int, typename, typename, bool> friend struct internal::InnerMostDimPreserver;
   template <typename S, typename O, typename D, bool V> friend struct internal::FullReducer;
 #ifdef EIGEN_USE_THREADS
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
index cd20df505..7504c1598 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
@@ -376,12 +376,12 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
   // so reduce the scope of the optimized version of the code to the simple cases
   // of doubles, floats and half floats
 #ifdef EIGEN_HAS_GPU_FP16
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
 #else // EIGEN_HAS_GPU_FP16
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
 #endif // EIGEN_HAS_GPU_FP16
@@ -697,12 +697,12 @@ struct InnerReducer<Self, Op, GpuDevice> {
   // so reduce the scope of the optimized version of the code to the simple case
   // of floats and half floats.
 #ifdef EIGEN_HAS_GPU_FP16
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
        internal::is_same<typename Self::CoeffReturnType, double>::value ||
        (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
 #else // EIGEN_HAS_GPU_FP16
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
                                                  (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                   internal::is_same<typename Self::CoeffReturnType, double>::value);
 #endif // EIGEN_HAS_GPU_FP16
@@ -759,7 +759,7 @@ struct OuterReducer<Self, Op, GpuDevice> {
   // Unfortunately nvidia doesn't support well exotic types such as complex,
   // so reduce the scope of the optimized version of the code to the simple case
   // of floats.
-  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
                                                  (internal::is_same<typename Self::CoeffReturnType, float>::value ||
                                                   internal::is_same<typename Self::CoeffReturnType, double>::value);
   template <typename Device, typename OutputType>
diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp
index ff8e18c07..9458f4e4e 100644
--- a/unsupported/test/cxx11_tensor_reduction.cpp
+++ b/unsupported/test/cxx11_tensor_reduction.cpp
@@ -386,7 +386,7 @@ static void test_static_dims() {
           expected = (std::max)(expected, in(i, k, j, l));
         }
       }
-      VERIFY_IS_APPROX(out(i, j), expected);
+      VERIFY_IS_EQUAL(out(i, j), expected);
     }
   }
 }
@@ -417,7 +417,7 @@ static void test_innermost_last_dims() {
           expected = (std::max)(expected, in(l, k, i, j));
         }
       }
-      VERIFY_IS_APPROX(out(i, j), expected);
+      VERIFY_IS_EQUAL(out(i, j), expected);
     }
   }
 }
@@ -448,7 +448,7 @@ static void test_innermost_first_dims() {
           expected = (std::max)(expected, in(i, j, k, l));
         }
       }
-      VERIFY_IS_APPROX(out(i, j), expected);
+      VERIFY_IS_EQUAL(out(i, j), expected);
     }
   }
 }
@@ -479,11 +479,30 @@ static void test_reduce_middle_dims() {
           expected = (std::max)(expected, in(i, k, l, j));
         }
       }
-      VERIFY_IS_APPROX(out(i, j), expected);
+      VERIFY_IS_EQUAL(out(i, j), expected);
     }
   }
 }
 
+static void test_sum_accuracy() {
+  Tensor<float, 3> tensor(101, 101, 101);
+  for (float prescribed_mean : {1.0f, 10.0f, 100.0f, 1000.0f, 10000.0f}) {
+    tensor.setRandom();
+    tensor += tensor.constant(prescribed_mean);
+
+    Tensor<float, 0> sum = tensor.sum();
+    double expected_sum = 0.0;
+    for (int i = 0; i < 101; ++i) {
+      for (int j = 0; j < 101; ++j) {
+        for (int k = 0; k < 101; ++k) {
+          expected_sum += static_cast<double>(tensor(i, j, k));
+        }
+      }
+    }
+    VERIFY_IS_APPROX(sum(), static_cast<float>(expected_sum));
+  }
+}
+
 EIGEN_DECLARE_TEST(cxx11_tensor_reduction) {
   CALL_SUBTEST(test_trivial_reductions<ColMajor>());
   CALL_SUBTEST(test_trivial_reductions<RowMajor>());
@@ -506,4 +525,5 @@ EIGEN_DECLARE_TEST(cxx11_tensor_reduction) {
   CALL_SUBTEST(test_innermost_first_dims<RowMajor>());
   CALL_SUBTEST(test_reduce_middle_dims<ColMajor>());
   CALL_SUBTEST(test_reduce_middle_dims<RowMajor>());
+  CALL_SUBTEST(test_sum_accuracy());
 }

From 81b38a155adf5d527bce5c84cf90cd83c28da445 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 11 Sep 2018 13:32:32 -0700
Subject: [PATCH 631/680] Fix compilation of tiled evaluation code with c++03

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h   |  6 +++---
 .../Eigen/CXX11/src/Tensor/TensorBroadcasting.h    | 11 +++++------
 .../Eigen/CXX11/src/Tensor/TensorChipping.h        |  5 ++---
 .../Eigen/CXX11/src/Tensor/TensorImagePatch.h      |  5 ++---
 .../Eigen/CXX11/src/Tensor/TensorMorphing.h        |  5 ++---
 .../Eigen/CXX11/src/Tensor/TensorReduction.h       |  5 ++---
 .../Eigen/CXX11/src/Tensor/TensorShuffling.h       |  4 ++--
 unsupported/test/cxx11_tensor_block_access.cpp     | 14 +++++++-------
 unsupported/test/cxx11_tensor_shuffling.cpp        | 10 +++++-----
 9 files changed, 30 insertions(+), 35 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index 6d90af2d3..13da36257 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -396,8 +396,8 @@ struct TensorBlockCwiseUnaryOp {
     typedef const Eigen::Array<InputScalar, Dynamic, 1> Input;
     typedef Eigen::Array<OutputScalar, Dynamic, 1> Output;
 
-    typedef Eigen::Map<Input, 0, InnerStride<>> InputMap;
-    typedef Eigen::Map<Output, 0, InnerStride<>> OutputMap;
+    typedef Eigen::Map<Input, 0, InnerStride<> > InputMap;
+    typedef Eigen::Map<Output, 0, InnerStride<> > OutputMap;
 
     const InputScalar* input_base = &input_data[input_index];
     OutputScalar* output_base = &output_data[output_index];
@@ -502,7 +502,7 @@ struct TensorBlockCwiseUnaryIO {
                                    input_stride, input_data);
       // Update index.
       for (int j = 0; j < num_squeezed_dims; ++j) {
-        auto& state = block_iter_state[j];
+        BlockIteratorState& state = block_iter_state[j];
         if (++state.count < state.size) {
           output_index += state.output_stride;
           input_index += state.input_stride;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 02d061a9c..e5cf93ab0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -596,12 +596,11 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
     // TODO(wuke): Targeting L1 size is 30% faster than targeting L{-1} on large
     // tensors. But this might need further tuning.
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.firstLevelCacheSize() / sizeof(Scalar));
 
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kSkewedInnerDims,
-        block_total_size_max));
+        internal::kSkewedInnerDims, block_total_size_max));
 
     m_impl.getResourceRequirements(resources);
   }
@@ -617,8 +616,8 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
     // equal to m_dimensions for inner dims, a smaller than m_dimensions[i] size
     // for the first outer dim, and 1 for other outer dims. This is guaranteed
     // by MergeResourceRequirements() in TensorBlock.h.
-    const auto& output_block_sizes = output_block->block_sizes();
-    const auto& output_block_strides = output_block->block_strides();
+    const Dimensions& output_block_sizes = output_block->block_sizes();
+    const Dimensions& output_block_strides = output_block->block_strides();
 
     // Find where outer dims start.
     int outer_dim_start = 0;
@@ -642,7 +641,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       return;
     }
 
-    const auto& input_dims = m_impl.dimensions();
+    const Dimensions& input_dims = m_impl.dimensions();
 
     // Pre-fill input_block_sizes, broadcast_block_sizes,
     // broadcast_block_strides, and broadcast_tensor_strides. Later on we will
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 76fab39e2..b47fa9e8e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -290,11 +290,10 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kSkewedInnerDims,
-        block_total_size_max));
+        internal::kSkewedInnerDims, block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 1826d7022..965bd8f1e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -550,11 +550,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kSkewedInnerDims,
-        block_total_size_max));
+        internal::kSkewedInnerDims, block_total_size_max));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 2f765acb7..16dc74afe 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -677,11 +677,10 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kSkewedInnerDims,
-        block_total_size_max));
+        internal::kSkewedInnerDims, block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 3d534eaa2..eeb2578fd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -771,11 +771,10 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.lastLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kSkewedInnerDims,
-        block_total_size_max));
+        internal::kSkewedInnerDims, block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index a5b541a68..e018d0ab2 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -229,10 +229,10 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
       std::vector<internal::TensorOpResourceRequirements>* resources) const {
-    auto block_total_size_max = numext::maxi<Eigen::Index>(
+    Eigen::Index block_total_size_max = numext::maxi<Eigen::Index>(
         1, m_device.firstLevelCacheSize() / sizeof(Scalar));
     resources->push_back(internal::TensorOpResourceRequirements(
-        internal::TensorBlockShapeType::kUniformAllDims, block_total_size_max));
+        internal::kUniformAllDims, block_total_size_max));
     m_impl.getResourceRequirements(resources);
   }
 
diff --git a/unsupported/test/cxx11_tensor_block_access.cpp b/unsupported/test/cxx11_tensor_block_access.cpp
index eec282ba7..ad12ae557 100644
--- a/unsupported/test/cxx11_tensor_block_access.cpp
+++ b/unsupported/test/cxx11_tensor_block_access.cpp
@@ -535,7 +535,7 @@ static void test_block_cwise_unary_io_basic() {
   DSizes<Index, NumDims> block_sizes = RandomDims<NumDims>();
   DSizes<Index, NumDims> strides(ComputeStrides<Layout, NumDims>(block_sizes));
 
-  const auto total_size = block_sizes.TotalSize();
+  const Index total_size = block_sizes.TotalSize();
 
   // Create a random input tensors.
   T* input_data = GenerateRandomData<T>(total_size);
@@ -562,12 +562,12 @@ static void test_block_cwise_unary_io_squeeze_ones() {
   DSizes<Index, 5> block_sizes(1, 2, 1, 3, 1);
   DSizes<Index, 5> strides(ComputeStrides<Layout, 5>(block_sizes));
 
-  const auto total_size = block_sizes.TotalSize();
+  const Index total_size = block_sizes.TotalSize();
 
   // Create a random input tensors.
-  auto* input_data = GenerateRandomData<float>(total_size);
+  float* input_data = GenerateRandomData<float>(total_size);
 
-  auto* output_data = new float[total_size];
+  float* output_data = new float[total_size];
   UnaryFunctor functor;
   TensorBlockCwiseUnaryIO::Run(functor, block_sizes, strides, output_data,
                                strides, input_data);
@@ -599,13 +599,13 @@ static void test_block_cwise_unary_io_zero_strides() {
   input_strides[4] = 0;
 
   // Generate random data.
-  auto* input_data = GenerateRandomData<float>(input_sizes.TotalSize());
+  float* input_data = GenerateRandomData<float>(input_sizes.TotalSize());
 
   DSizes<Index, 5> output_sizes = rnd_dims;
   DSizes<Index, 5> output_strides(ComputeStrides<Layout, 5>(output_sizes));
 
-  const auto output_total_size = output_sizes.TotalSize();
-  auto* output_data = new float[output_total_size];
+  const Index output_total_size = output_sizes.TotalSize();
+  float* output_data = new float[output_total_size];
 
   UnaryFunctor functor;
   TensorBlockCwiseUnaryIO::Run(functor, output_sizes, output_strides,
diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp
index 467df39c7..062dd1c0f 100644
--- a/unsupported/test/cxx11_tensor_shuffling.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling.cpp
@@ -81,12 +81,12 @@ static void test_expr_shuffling()
   Tensor<float, 4, DataLayout> expected;
   expected = tensor.shuffle(shuffles);
 
-  Tensor<float, 4, DataLayout> result(5,7,3,2);
+  Tensor<float, 4, DataLayout> result(5, 7, 3, 2);
 
-  array<ptrdiff_t, 4> src_slice_dim{{2,3,1,7}};
-  array<ptrdiff_t, 4> src_slice_start{{0,0,0,0}};
-  array<ptrdiff_t, 4> dst_slice_dim{{1,7,3,2}};
-  array<ptrdiff_t, 4> dst_slice_start{{0,0,0,0}};
+  array<ptrdiff_t, 4> src_slice_dim({2, 3, 1, 7});
+  array<ptrdiff_t, 4> src_slice_start({0, 0, 0, 0});
+  array<ptrdiff_t, 4> dst_slice_dim({1, 7, 3, 2});
+  array<ptrdiff_t, 4> dst_slice_start({0, 0, 0, 0});
 
   for (int i = 0; i < 5; ++i) {
     result.slice(dst_slice_start, dst_slice_dim) =

From e289f44c565aa6d12d21a498ef34a515891cf520 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Tue, 11 Sep 2018 14:09:00 -0700
Subject: [PATCH 632/680] Don't vectorize the MeanReducer unless pdiv is
 available.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 9fd276075..51572f9e7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -158,7 +158,8 @@ template <typename T, typename Device>
 struct reducer_traits<MeanReducer<T>, Device> {
   enum {
     Cost = NumTraits<T>::AddCost,
-    PacketAccess = PacketType<T, Device>::HasAdd && !NumTraits<T>::IsInteger,
+    PacketAccess = PacketType<T, Device>::HasAdd &&
+                   PacketType<T, Device>::HasDiv && !NumTraits<T>::IsInteger,
     IsStateful = true,
     IsExactlyAssociative = NumTraits<T>::IsInteger
   };

From d138fe341dab4b7c1a7b7d012a3e13e1796aad8f Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 11 Sep 2018 17:23:18 -0700
Subject: [PATCH 633/680] Fis static_assert in test to conform c++11 standard

---
 unsupported/test/cxx11_tensor_executor.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 1dc18220c..8639e7b38 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -286,7 +286,7 @@ template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
 static void test_execute_reduction(Device d)
 {
-  static_assert(NumDims >= 2);
+  static_assert(NumDims >= 2, "NumDims must be greater or equal than 2");
 
   static constexpr int ReducedDims = NumDims - 2;
   static constexpr int Options = 0 | Layout;
@@ -328,7 +328,7 @@ template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
 static void test_execute_reshape(Device d)
 {
-  static_assert(NumDims >= 2);
+  static_assert(NumDims >= 2, "NumDims must be greater or equal than 2");
 
   static constexpr int ReshapedDims = NumDims - 1;
   static constexpr int Options = 0 | Layout;
@@ -368,7 +368,7 @@ template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
 static void test_execute_slice_rvalue(Device d)
 {
-  static_assert(NumDims >= 2);
+  static_assert(NumDims >= 2, "NumDims must be greater or equal than 2");
   static constexpr int Options = 0 | Layout;
 
   auto dims = RandomDims<NumDims>(5, 10);
@@ -408,7 +408,7 @@ template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
 static void test_execute_slice_lvalue(Device d)
 {
-  static_assert(NumDims >= 2);
+  static_assert(NumDims >= 2, "NumDims must be greater or equal than 2");
   static constexpr int Options = 0 | Layout;
 
   auto dims = RandomDims<NumDims>(5, 10);

From 7f3b17e4031c6b921648906f43432ff728bb772d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 13 Sep 2018 18:15:07 +0200
Subject: [PATCH 634/680] MSVC 2015 supports c++11 thread-local-storage

---
 unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
index a41731c34..7229839ac 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
@@ -12,7 +12,8 @@
 
 #if EIGEN_MAX_CPP_VER >= 11 &&                         \
     ((EIGEN_COMP_GNUC && EIGEN_GNUC_AT_LEAST(4, 8)) || \
-     __has_feature(cxx_thread_local))
+     __has_feature(cxx_thread_local)                || \
+     (EIGEN_COMP_MSVC >= 1900) )
 #define EIGEN_THREAD_LOCAL static thread_local
 #endif
 

From 1141bcf7940cf18974f72fcad3febc41e07bc6ec Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 13 Sep 2018 23:53:28 +0200
Subject: [PATCH 635/680] Fix conjugate-gradient for very small rhs

---
 Eigen/src/IterativeLinearSolvers/ConjugateGradient.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h
index 395daa8e4..f7ce47134 100644
--- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h
+++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h
@@ -50,7 +50,8 @@ void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x,
     tol_error = 0;
     return;
   }
-  RealScalar threshold = tol*tol*rhsNorm2;
+  const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+  RealScalar threshold = numext::maxi(tol*tol*rhsNorm2,considerAsZero);
   RealScalar residualNorm2 = residual.squaredNorm();
   if (residualNorm2 < threshold)
   {
@@ -58,7 +59,7 @@ void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x,
     tol_error = sqrt(residualNorm2 / rhsNorm2);
     return;
   }
-  
+
   VectorType p(n);
   p = precond.solve(residual);      // initial search direction
 

From 01197e44527941c95f9a63e4f60ab3a989f12cbe Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 13 Sep 2018 15:03:36 -0700
Subject: [PATCH 636/680] Fix warnings

---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 2 --
 unsupported/test/cxx11_tensor_executor.cpp           | 8 ++++----
 2 files changed, 4 insertions(+), 6 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index eeb2578fd..261dbffa8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -1140,7 +1140,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       const Index max_coeff_count,
       const DSizes<Index, NumInputDims>& input_slice_sizes,
       DSizes<Index, NumInputDims>* target_input_block_sizes) const {
-    typedef typename internal::packet_traits<Scalar>::type Packet;
     typedef internal::BlockReducer<Self, Op> BlockReducer;
     // TODO(andydavis) Compute reducer overhead correctly for the case where
     // we are preserving the inner most dimension, and a single reducer
@@ -1168,7 +1167,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       } else if (!first_preserved_dim_allocated) {
         // TODO(andydavis) Include output block size in this L1 working set
         // calculation.
-        const Index allocated = max_coeff_count - coeff_to_allocate;
         const Index alloc_size = numext::maxi(
             static_cast<Index>(1), coeff_to_allocate / reducer_overhead);
         (*target_input_block_sizes)[dim] =
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 8639e7b38..1bb99854c 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -27,7 +27,7 @@ static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
     dims[i] = internal::random<int>(min_dim, max_dim);
   }
   return dims;
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -117,7 +117,7 @@ static void test_execute_broadcasting(Device d)
   for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
     VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
   }
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -155,7 +155,7 @@ static void test_execute_chipping_rvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
@@ -207,7 +207,7 @@ static void test_execute_chipping_lvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>

From 2b070181401193f562f38179c1a3bb81496485cf Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 13 Sep 2018 16:12:22 -0700
Subject: [PATCH 637/680] Enable vectorized version on GPUs. The underlying bug
 has been fixed.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index d7a04a525..f5b969ce5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -218,7 +218,6 @@ struct InnerMostDimReducer<Self, Op, false, true> {
   }
 };
 
-#if !defined(EIGEN_USE_GPU) || !defined(__CUDACC__)
 template <typename Self, typename Op>
 struct InnerMostDimReducer<Self, Op, true, true> {
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType
@@ -258,7 +257,6 @@ struct InnerMostDimReducer<Self, Op, true, true> {
     }
   }
 };
-#endif
 
 template <int DimIndex, typename Self, typename Op, bool vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
 struct InnerMostDimPreserver {

From 0db590d22db198154f4302d85fbbd892316d36da Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 13 Sep 2018 16:20:57 -0700
Subject: [PATCH 638/680] Backed out changeset
 01197e44527941c95f9a63e4f60ab3a989f12cbe

---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 2 ++
 unsupported/test/cxx11_tensor_executor.cpp           | 8 ++++----
 2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 8d93aacee..e8c47c5b5 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -1205,6 +1205,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       const Index max_coeff_count,
       const DSizes<Index, NumInputDims>& input_slice_sizes,
       DSizes<Index, NumInputDims>* target_input_block_sizes) const {
+    typedef typename internal::packet_traits<Scalar>::type Packet;
     typedef internal::BlockReducer<Self, Op> BlockReducer;
     // TODO(andydavis) Compute reducer overhead correctly for the case where
     // we are preserving the inner most dimension, and a single reducer
@@ -1232,6 +1233,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       } else if (!first_preserved_dim_allocated) {
         // TODO(andydavis) Include output block size in this L1 working set
         // calculation.
+        const Index allocated = max_coeff_count - coeff_to_allocate;
         const Index alloc_size = numext::maxi(
             static_cast<Index>(1), coeff_to_allocate / reducer_overhead);
         (*target_input_block_sizes)[dim] =
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 1bb99854c..8639e7b38 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -27,7 +27,7 @@ static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
     dims[i] = internal::random<int>(min_dim, max_dim);
   }
   return dims;
-}
+};
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -117,7 +117,7 @@ static void test_execute_broadcasting(Device d)
   for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
     VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
   }
-}
+};
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -155,7 +155,7 @@ static void test_execute_chipping_rvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-}
+};
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
@@ -207,7 +207,7 @@ static void test_execute_chipping_lvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-}
+};
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>

From 6313dde390604cbb05b3317d1dcd13ac75978d29 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 13 Sep 2018 16:42:05 -0700
Subject: [PATCH 639/680] Fix merge error.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 1 -
 unsupported/test/cxx11_tensor_executor.cpp           | 8 ++++----
 2 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index e8c47c5b5..3a405f5e6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -1205,7 +1205,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       const Index max_coeff_count,
       const DSizes<Index, NumInputDims>& input_slice_sizes,
       DSizes<Index, NumInputDims>* target_input_block_sizes) const {
-    typedef typename internal::packet_traits<Scalar>::type Packet;
     typedef internal::BlockReducer<Self, Op> BlockReducer;
     // TODO(andydavis) Compute reducer overhead correctly for the case where
     // we are preserving the inner most dimension, and a single reducer
diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 8639e7b38..1bb99854c 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -27,7 +27,7 @@ static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
     dims[i] = internal::random<int>(min_dim, max_dim);
   }
   return dims;
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -117,7 +117,7 @@ static void test_execute_broadcasting(Device d)
   for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
     VERIFY_IS_EQUAL(dst.coeff(i), golden.coeff(i));
   }
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
@@ -155,7 +155,7 @@ static void test_execute_chipping_rvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
@@ -207,7 +207,7 @@ static void test_execute_chipping_lvalue(Device d) {
   TEST_CHIPPING(5)
 
 #undef TEST_CHIPPING
-};
+}
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>

From d0eef5fe6cb0d3d18bf5f2482a197c44b95235f6 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 13 Sep 2018 17:04:05 -0700
Subject: [PATCH 640/680] Don't use bracket syntax in ctor.

---
 unsupported/test/cxx11_tensor_convolution.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp
index 78704e5e5..c3688f678 100644
--- a/unsupported/test/cxx11_tensor_convolution.cpp
+++ b/unsupported/test/cxx11_tensor_convolution.cpp
@@ -25,7 +25,8 @@ static void test_evals()
 
   Tensor<float, 2, DataLayout> result(2,3);
   result.setZero();
-  Eigen::array<Tensor<float, 2>::Index, 1> dims3({0});
+  Eigen::array<Tensor<float, 2>::Index, 1> dims3;
+  dims3[0] = 0;
 
   typedef TensorEvaluator<decltype(input.convolve(kernel, dims3)), DefaultDevice> Evaluator;
   Evaluator eval(input.convolve(kernel, dims3), DefaultDevice());

From d7378aae8e87f85fcf825bf3a04501a16aca4780 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 14 Sep 2018 20:17:47 +0200
Subject: [PATCH 641/680] Provide EIGEN_ALIGNOF macro, and give
 handmade_aligned_malloc the possibility for alignments larger than the
 standard alignment.

---
 Eigen/src/Core/util/ConfigureVectorization.h | 6 +++++-
 Eigen/src/Core/util/Memory.h                 | 7 ++++---
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index f30503b33..e75c7d89e 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -39,15 +39,19 @@
  */
 #if (defined EIGEN_CUDACC)
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
+  #define EIGEN_ALIGNOF(x) __alignof(x)
 #elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+  #define EIGEN_ALIGNOF(x) __alignof(x)
 #elif EIGEN_COMP_MSVC
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
+  #define EIGEN_ALIGNOF(x) __alignof(x)
 #elif EIGEN_COMP_SUNCC
   // FIXME not sure about this one:
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+  #define EIGEN_ALIGNOF(x) __alignof(x)
 #else
-  #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
+  #error Please tell me what is the equivalent of __attribute__((aligned(n))) and __alignof(x) for your compiler
 #endif
 
 // If the user explicitly disable vectorization, then we also disable alignment
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 056070159..76472703e 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -96,11 +96,12 @@ inline void throw_std_bad_alloc()
 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
   * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
   */
-inline void* handmade_aligned_malloc(std::size_t size)
+inline void* handmade_aligned_malloc(std::size_t size, std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES)
 {
-  void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES);
+  eigen_assert(alignment >= sizeof(void*) && (alignment & -alignment) == alignment && "Alignment must be at least sizeof(void*) and a power of 2");
+  void *original = std::malloc(size+alignment);
   if (original == 0) return 0;
-  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(alignment-1))) + alignment);
   *(reinterpret_cast<void**>(aligned) - 1) = original;
   return aligned;
 }

From 007f165c69f521af0a215c3739c16822e6e2abca Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Fri, 14 Sep 2018 20:21:56 +0200
Subject: [PATCH 642/680] bug #1598: Let MaxSizeVector respect alignment of
 objects and add a unit test Also revert
 8b3d9ed081fc5d4870290649853b19cb5179546e

---
 .../Eigen/CXX11/src/ThreadPool/EventCount.h   |  7 +-
 .../Eigen/CXX11/src/util/MaxSizeVector.h      |  9 ++-
 unsupported/test/CMakeLists.txt               |  1 +
 unsupported/test/cxx11_maxsizevector.cpp      | 77 +++++++++++++++++++
 4 files changed, 86 insertions(+), 8 deletions(-)
 create mode 100644 unsupported/test/cxx11_maxsizevector.cpp

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
index 2cac2d0f1..22c952ae1 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
@@ -169,7 +169,9 @@ class EventCount {
 
   class Waiter {
     friend class EventCount;
-    std::atomic<Waiter*> next;
+    // Align to 128 byte boundary to prevent false sharing with other Waiter
+    // objects in the same vector.
+    EIGEN_ALIGN_TO_BOUNDARY(128) std::atomic<Waiter*> next;
     std::mutex mu;
     std::condition_variable cv;
     uint64_t epoch;
@@ -179,9 +181,6 @@ class EventCount {
       kWaiting,
       kSignaled,
     };
-    // Pad past 128 byte boundary to prevent false sharing with other Waiter
-    // objects in the same vector.
-    char pad_[128];
   };
 
  private:
diff --git a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
index bc5b3632c..277ab149a 100644
--- a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
+++ b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h
@@ -29,12 +29,13 @@ namespace Eigen {
   */
 template <typename T>
 class MaxSizeVector {
+  static const size_t alignment = EIGEN_PLAIN_ENUM_MAX(EIGEN_ALIGNOF(T), sizeof(void*));
  public:
   // Construct a new MaxSizeVector, reserve n elements.
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   explicit MaxSizeVector(size_t n)
       : reserve_(n), size_(0),
-        data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
+        data_(static_cast<T*>(internal::handmade_aligned_malloc(n * sizeof(T), alignment))) {
   }
 
   // Construct a new MaxSizeVector, reserve and resize to n.
@@ -42,7 +43,7 @@ class MaxSizeVector {
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   MaxSizeVector(size_t n, const T& init)
       : reserve_(n), size_(n),
-        data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
+        data_(static_cast<T*>(internal::handmade_aligned_malloc(n * sizeof(T), alignment))) {
     size_t i = 0;
     EIGEN_TRY
     {
@@ -52,7 +53,7 @@ class MaxSizeVector {
     {
       // Construction failed, destruct in reverse order:
       for(; (i+1) > 0; --i) { data_[i-1].~T(); }
-      internal::aligned_free(data_);
+      internal::handmade_aligned_free(data_);
       EIGEN_THROW;
     }
   }
@@ -62,7 +63,7 @@ class MaxSizeVector {
     for (size_t i = size_; i > 0; --i) {
       data_[i-1].~T();
     }
-    internal::aligned_free(data_);
+    internal::handmade_aligned_free(data_);
   }
 
   void resize(size_t n) {
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 239a80926..1a4a77058 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -142,6 +142,7 @@ ei_add_test(cxx11_tensor_sugar)
 ei_add_test(cxx11_tensor_roundings)
 ei_add_test(cxx11_tensor_layout_swap)
 ei_add_test(cxx11_tensor_io)
+ei_add_test(cxx11_maxsizevector)
 endif()
 
 if(EIGEN_TEST_CXX11)
diff --git a/unsupported/test/cxx11_maxsizevector.cpp b/unsupported/test/cxx11_maxsizevector.cpp
new file mode 100644
index 000000000..46b689a8e
--- /dev/null
+++ b/unsupported/test/cxx11_maxsizevector.cpp
@@ -0,0 +1,77 @@
+#include "main.h"
+
+#include <exception>  // std::exception
+
+#include <unsupported/Eigen/CXX11/Tensor>
+
+struct Foo
+{
+  static Index object_count;
+  static Index object_limit;
+  EIGEN_ALIGN_TO_BOUNDARY(128) int dummy;
+
+  Foo(int x=0) : dummy(x)
+  {
+#ifdef EIGEN_EXCEPTIONS
+    // TODO: Is this the correct way to handle this?
+    if (Foo::object_count > Foo::object_limit) { std::cout << "\nThrow!\n"; throw Foo::Fail(); }
+#endif
+    std::cout << '+';
+    ++Foo::object_count;
+    eigen_assert((internal::UIntPtr(this) & (127)) == 0);
+  }
+  Foo(const Foo&)
+  {
+    std::cout << 'c';
+    ++Foo::object_count;
+    eigen_assert((internal::UIntPtr(this) & (127)) == 0);
+  }
+
+  ~Foo()
+  {
+    std::cout << '~';
+    --Foo::object_count;
+  }
+
+  class Fail : public std::exception {};
+};
+
+Index Foo::object_count = 0;
+Index Foo::object_limit = 0;
+
+
+
+EIGEN_DECLARE_TEST(cxx11_maxsizevector)
+{
+  typedef MaxSizeVector<Foo> VectorX;
+  Foo::object_count = 0;
+  for(int r = 0; r < g_repeat; r++) {
+    Index rows = internal::random<Index>(3,30);
+    Foo::object_limit = internal::random<Index>(0, rows - 2);
+    std::cout << "object_limit = " << Foo::object_limit << std::endl;
+    bool exception_raised = false;
+#ifdef EIGEN_EXCEPTIONS
+    try
+    {
+#endif
+      std::cout <<       "\nVectorX m(" << rows << ");\n";
+      VectorX vect(rows);
+      for(int i=0; i<rows; ++i)
+          vect.push_back(Foo());
+#ifdef EIGEN_EXCEPTIONS
+      VERIFY(false);  // not reached if exceptions are enabled
+    }
+    catch (const Foo::Fail&) { exception_raised = true; }
+    VERIFY(exception_raised);
+#endif
+    VERIFY_IS_EQUAL(Index(0), Foo::object_count);
+
+    {
+      Foo::object_limit = rows+1;
+      VectorX vect2(rows, Foo());
+      VERIFY_IS_EQUAL(Foo::object_count, rows);
+    }
+    VERIFY_IS_EQUAL(Index(0), Foo::object_count);
+    std::cout << '\n';
+  }
+}

From 1b8d70a22b83d63667bbefe3899d9a2e0c2c8b78 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 14 Sep 2018 15:25:27 -0700
Subject: [PATCH 643/680] Support reshaping with static shapes and dimensions
 conversion in tensor broadcasting

---
 .../CXX11/src/Tensor/TensorBroadcasting.h     |  2 +-
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 10 ++++++++
 .../test/cxx11_tensor_broadcasting.cpp        |  2 +-
 unsupported/test/cxx11_tensor_morphing.cpp    | 23 +++++++++++++++++++
 4 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index e5cf93ab0..c102a43fb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -641,7 +641,7 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
       return;
     }
 
-    const Dimensions& input_dims = m_impl.dimensions();
+    const Dimensions& input_dims = Dimensions(m_impl.dimensions());
 
     // Pre-fill input_block_sizes, broadcast_block_sizes,
     // broadcast_block_strides, and broadcast_tensor_strides. Later on we will
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 7c26b1682..fe0d57f31 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -290,6 +290,16 @@ struct DSizes : array<DenseIndex, NumDims> {
     }
   }
 
+#ifdef EIGEN_HAS_INDEX_LIST
+  EIGEN_DEVICE_FUNC
+  template <typename FirstType, typename... OtherTypes>
+  DSizes(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
+    for (int i = 0; i < dimensions.count; ++i) {
+      (*this)[i] = dimensions[i];
+    }
+  }
+#endif
+
 #ifndef EIGEN_EMULATE_CXX11_META_H
   template <typename std::ptrdiff_t... Indices>
   EIGEN_DEVICE_FUNC DSizes(const Sizes<Indices...>& a) {
diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp
index 2f8ab6afd..7df5b53d6 100644
--- a/unsupported/test/cxx11_tensor_broadcasting.cpp
+++ b/unsupported/test/cxx11_tensor_broadcasting.cpp
@@ -115,7 +115,7 @@ static void test_static_broadcasting()
   Tensor<float, 3, DataLayout> tensor(8,3,5);
   tensor.setRandom();
 
-#if EIGEN_HAS_CONSTEXPR
+#if defined(EIGEN_HAS_INDEX_LIST)
   Eigen::IndexList<Eigen::type2index<2>, Eigen::type2index<3>, Eigen::type2index<4>> broadcasts;
 #else
   Eigen::array<int, 3> broadcasts;
diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp
index 6365cd89a..4cbe15b63 100644
--- a/unsupported/test/cxx11_tensor_morphing.cpp
+++ b/unsupported/test/cxx11_tensor_morphing.cpp
@@ -41,6 +41,28 @@ static void test_simple_reshape()
   }
 }
 
+template <typename>
+static void test_static_reshape() {
+#if defined(EIGEN_HAS_INDEX_LIST)
+  using Eigen::type2index;
+
+  Tensor<float, 5> tensor(2, 3, 1, 7, 1);
+  tensor.setRandom();
+
+  // New dimensions: [2, 3, 7]
+  Eigen::IndexList<type2index<2>, type2index<3>, type2index<7>> dim;
+  Tensor<float, 3> reshaped = tensor.reshape(dim);
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 3; ++j) {
+      for (int k = 0; k < 7; ++k) {
+        VERIFY_IS_EQUAL(tensor(i, j, 0, k, 0), reshaped(i, j, k));
+      }
+    }
+  }
+#endif
+}
+
 template<typename>
 static void test_reshape_in_expr() {
   MatrixXf m1(2,3*5*7*11);
@@ -462,6 +484,7 @@ static void test_composition()
 EIGEN_DECLARE_TEST(cxx11_tensor_morphing)
 {
   CALL_SUBTEST_1(test_simple_reshape<void>());
+  CALL_SUBTEST_1(test_static_reshape<void>());
   CALL_SUBTEST_1(test_reshape_in_expr<void>());
   CALL_SUBTEST_1(test_reshape_as_lvalue<void>());
 

From 48633757235b522b37e60022dfbe8813c6cdfdc2 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 14 Sep 2018 16:55:05 -0700
Subject: [PATCH 644/680] Explicitly construct tensor block dimensions from
 evaluator dimensions

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h  | 5 +++--
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 4 ++--
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index ba5ab1396..bfe1f97b8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -282,8 +282,9 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
       TensorOpCost cost = evaluator.costPerCoeff(Vectorizable);
       double taskSize = TensorCostModel<ThreadPoolDevice>::taskSize(1, cost);
       size_t block_size = static_cast<size_t>(1.0 / taskSize);
-      TensorBlockMapper block_mapper(evaluator.dimensions(), block_shape,
-                                     block_size);
+      TensorBlockMapper block_mapper(
+          typename TensorBlockMapper::Dimensions(evaluator.dimensions()),
+          block_shape, block_size);
       block_size = block_mapper.block_dims_total_size();
       const size_t aligned_blocksize =
           EIGEN_MAX_ALIGN_BYTES *
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 3a405f5e6..4263f11a6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -976,7 +976,8 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
         // find that scattered reads are not worth supporting in
         // TensorSliceBlockMapper.
         TensorSliceBlockMapper block_mapper(
-            input_tensor_dims, tensor_slice_offsets, tensor_slice_extents,
+            typename TensorSliceBlockMapper::Dimensions(input_tensor_dims),
+            tensor_slice_offsets, tensor_slice_extents,
             target_input_block_sizes, DimensionList<Index, NumInputDims>());
 
         const Index num_outputs_to_update =
@@ -1232,7 +1233,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
       } else if (!first_preserved_dim_allocated) {
         // TODO(andydavis) Include output block size in this L1 working set
         // calculation.
-        const Index allocated = max_coeff_count - coeff_to_allocate;
         const Index alloc_size = numext::maxi(
             static_cast<Index>(1), coeff_to_allocate / reducer_overhead);
         (*target_input_block_sizes)[dim] =

From 71070a1e84538a91e6ecf364f4a23716eeb780bc Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 14 Sep 2018 17:17:50 -0700
Subject: [PATCH 645/680] Const cast scalar pointer in TensorSlicingOp
 evaluator

---
 unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 16dc74afe..d19aba3b3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -695,7 +695,7 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
-    Scalar* result = m_impl.data();
+    Scalar* result = const_cast<Scalar*>(m_impl.data());
     if (result) {
       Index offset = 0;
       if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {

From f7d0053cf04bcafe2f784aa4ae0cbf3342e2d500 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Fri, 14 Sep 2018 19:19:13 -0700
Subject: [PATCH 646/680] Fix DSizes IndexList constructor

---
 unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index fe0d57f31..94871ef43 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -291,8 +291,8 @@ struct DSizes : array<DenseIndex, NumDims> {
   }
 
 #ifdef EIGEN_HAS_INDEX_LIST
-  EIGEN_DEVICE_FUNC
   template <typename FirstType, typename... OtherTypes>
+  EIGEN_DEVICE_FUNC
   DSizes(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
     for (int i = 0; i < dimensions.count; ++i) {
       (*this)[i] = dimensions[i];

From 42705ba574e8c0a1764ef96e41831ed353b4057e Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Sat, 15 Sep 2018 12:43:41 +0200
Subject: [PATCH 647/680] Fix weird error for building with g++-4.7 in C++03
 mode.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
index e018d0ab2..416948765 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
@@ -222,7 +222,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
   template<int LoadMode>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
-    EIGEN_STATIC_ASSERT(PacketSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
         eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
     return PacketLoader<LoadMode, Self, TensorEvaluator<ArgType, Device>::PacketAccess>::Run(*this, index);
   }

From f313126dab9c6b2eeda0eb8bc7392ea6e664a88f Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Sat, 15 Sep 2018 13:47:54 -0700
Subject: [PATCH 648/680] Fix warnings in IndexList array_prod

---
 unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index 3f7d26b18..c19edd07f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -351,9 +351,9 @@ struct IndexPairList : internal::IndexTuple<FirstType, OtherTypes...> {
 namespace internal {
 
 template<typename FirstType, typename... OtherTypes>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
-  size_t result = 1;
-  for (int i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
+  DenseIndex result = 1;
+  for (size_t i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
     result *= sizes[i];
   }
   return result;

From 66f056776f1220ede1aa5cbfe058b88d6df3e359 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Sat, 15 Sep 2018 15:17:38 -0700
Subject: [PATCH 649/680] Add DSizes index type promotion

---
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 19 ++++++++++++++++++-
 unsupported/test/cxx11_tensor_dimension.cpp   | 18 ++++++++++++++++++
 2 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 94871ef43..5de0d0de7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -290,10 +290,27 @@ struct DSizes : array<DenseIndex, NumDims> {
     }
   }
 
+#ifdef EIGEN_HAS_CXX11
+  // Enable DSizes index type promotion only if we are promoting to the
+  // larger type, e.g. allow to promote dimensions of type int to long.
+  template<typename OtherIndex,
+           typename std::enable_if<
+               std::is_same<
+                 DenseIndex,
+                 typename internal::promote_index_type<DenseIndex, OtherIndex>::type
+               >::value, int>::type = 0>
+  EIGEN_DEVICE_FUNC
+  explicit DSizes(const array<OtherIndex, NumDims>& other) {
+    for (int i = 0; i < NumDims; ++i) {
+      (*this)[i] = static_cast<DenseIndex>(other[i]);
+    }
+  }
+#endif  // EIGEN_HAS_CXX11
+
 #ifdef EIGEN_HAS_INDEX_LIST
   template <typename FirstType, typename... OtherTypes>
   EIGEN_DEVICE_FUNC
-  DSizes(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
+  explicit DSizes(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
     for (int i = 0; i < dimensions.count; ++i) {
       (*this)[i] = dimensions[i];
     }
diff --git a/unsupported/test/cxx11_tensor_dimension.cpp b/unsupported/test/cxx11_tensor_dimension.cpp
index 10364d4b4..26f8edd8a 100644
--- a/unsupported/test/cxx11_tensor_dimension.cpp
+++ b/unsupported/test/cxx11_tensor_dimension.cpp
@@ -60,10 +60,28 @@ static void test_rank_zero()
   VERIFY_IS_EQUAL((int)dscalar.rank(), 0);
 }
 
+static void test_index_type_promotion() {
+#ifdef EIGEN_HAS_CXX11
+  Eigen::DSizes<int, 3> src0(1, 2, 3);
+  Eigen::array<int, 3> src1 = {4, 5, 6};
+
+  Eigen::DSizes<long, 3> dst0(src0);
+  Eigen::DSizes<long, 3> dst1(src1);
+
+  VERIFY_IS_EQUAL(dst0[0], 1L);
+  VERIFY_IS_EQUAL(dst0[1], 2L);
+  VERIFY_IS_EQUAL(dst0[2], 3L);
+  VERIFY_IS_EQUAL(dst1[0], 4L);
+  VERIFY_IS_EQUAL(dst1[1], 5L);
+  VERIFY_IS_EQUAL(dst1[2], 6L);
+#endif  // EIGEN_HAS_CXX11
+}
+
 EIGEN_DECLARE_TEST(cxx11_tensor_dimension)
 {
   CALL_SUBTEST(test_dynamic_size());
   CALL_SUBTEST(test_fixed_size());
   CALL_SUBTEST(test_match());
   CALL_SUBTEST(test_rank_zero());
+  CALL_SUBTEST(test_index_type_promotion());
 }

From 72f19c827a04b01d1ba9ed81a3744d965bdf4e05 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 16 Sep 2018 22:10:34 +0200
Subject: [PATCH 650/680] typo

---
 unsupported/Eigen/CXX11/src/Tensor/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/README.md b/unsupported/Eigen/CXX11/src/Tensor/README.md
index dfd7ab7c7..d6bc6d46f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/README.md
+++ b/unsupported/Eigen/CXX11/src/Tensor/README.md
@@ -1250,7 +1250,7 @@ dd a comment to this line
     =>
     a
     1 2 3
-    6 5 4
+    4 5 6
 
     b
     1  3  6

From b311bfb752c925aab15d9e2148e179577e62afd8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Sep 2018 09:54:29 +0200
Subject: [PATCH 651/680] bug #1596: fix inclusion of Eigen's header within
 unsupported modules.

---
 unsupported/Eigen/AdolcForward                       |  2 +-
 unsupported/Eigen/AlignedVector3                     |  2 +-
 unsupported/Eigen/ArpackSupport                      |  8 ++++----
 unsupported/Eigen/BVH                                |  6 +++---
 unsupported/Eigen/CXX11/Tensor                       |  4 ++--
 unsupported/Eigen/CXX11/TensorSymmetry               |  6 +++---
 unsupported/Eigen/CXX11/ThreadPool                   |  4 ++--
 unsupported/Eigen/EulerAngles                        |  8 ++++----
 unsupported/Eigen/FFT                                |  2 +-
 unsupported/Eigen/IterativeSolvers                   |  2 +-
 unsupported/Eigen/LevenbergMarquardt                 | 12 ++++++------
 unsupported/Eigen/MPRealSupport                      |  2 +-
 unsupported/Eigen/MatrixFunctions                    |  6 +++---
 unsupported/Eigen/MoreVectorization                  |  2 +-
 unsupported/Eigen/NonLinearOptimization              |  8 ++++----
 unsupported/Eigen/NumericalDiff                      |  2 +-
 unsupported/Eigen/OpenGLSupport                      |  2 +-
 unsupported/Eigen/Polynomials                        |  8 ++++----
 unsupported/Eigen/Skyline                            |  6 +++---
 .../src/Eigenvalues/ArpackSelfAdjointEigenSolver.h   |  2 +-
 .../Eigen/src/IterativeSolvers/ConstrainedConjGrad.h |  2 +-
 unsupported/Eigen/src/IterativeSolvers/DGMRES.h      |  2 +-
 unsupported/Eigen/src/Splines/SplineFitting.h        |  4 ++--
 unsupported/Eigen/src/Splines/SplineFwd.h            |  2 +-
 24 files changed, 52 insertions(+), 52 deletions(-)

diff --git a/unsupported/Eigen/AdolcForward b/unsupported/Eigen/AdolcForward
index 15f5f0731..9b8d3cd1a 100644
--- a/unsupported/Eigen/AdolcForward
+++ b/unsupported/Eigen/AdolcForward
@@ -40,7 +40,7 @@
 # undef realloc
 #endif
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
 namespace Eigen {
 
diff --git a/unsupported/Eigen/AlignedVector3 b/unsupported/Eigen/AlignedVector3
index c16f16a97..a67ee6e42 100644
--- a/unsupported/Eigen/AlignedVector3
+++ b/unsupported/Eigen/AlignedVector3
@@ -10,7 +10,7 @@
 #ifndef EIGEN_ALIGNED_VECTOR3
 #define EIGEN_ALIGNED_VECTOR3
 
-#include <Eigen/Geometry>
+#include "../../Eigen/Geometry"
 
 #include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
diff --git a/unsupported/Eigen/ArpackSupport b/unsupported/Eigen/ArpackSupport
index a0d4820e1..28c95ffa2 100644
--- a/unsupported/Eigen/ArpackSupport
+++ b/unsupported/Eigen/ArpackSupport
@@ -9,7 +9,7 @@
 #ifndef EIGEN_ARPACKSUPPORT_MODULE_H
 #define EIGEN_ARPACKSUPPORT_MODULE_H
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
 /** \defgroup ArpackSupport_Module Arpack support module
   *
@@ -20,12 +20,12 @@
   * \endcode
   */
 
-#include <Eigen/SparseCholesky>
+#include "../../Eigen/SparseCholesky"
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 #include "src/Eigenvalues/ArpackSelfAdjointEigenSolver.h"
 
-#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_ARPACKSUPPORT_MODULE_H
 /* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/unsupported/Eigen/BVH b/unsupported/Eigen/BVH
index 0161a5402..666c9835f 100644
--- a/unsupported/Eigen/BVH
+++ b/unsupported/Eigen/BVH
@@ -10,9 +10,9 @@
 #ifndef EIGEN_BVH_MODULE_H
 #define EIGEN_BVH_MODULE_H
 
-#include <Eigen/Core>
-#include <Eigen/Geometry>
-#include <Eigen/StdVector>
+#include "../../Eigen/Core"
+#include "../../Eigen/Geometry"
+#include "../../Eigen/StdVector"
 #include <algorithm>
 #include <queue>
 
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 0165d0144..14a2d0f0d 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -28,7 +28,7 @@
 
 #include "../SpecialFunctions"
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
 #include "src/util/CXX11Meta.h"
 #include "src/util/MaxSizeVector.h"
 
@@ -160,6 +160,6 @@ typedef unsigned __int64 uint64_t;
 
 #include "src/Tensor/TensorIO.h"
 
-#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 //#endif // EIGEN_CXX11_TENSOR_MODULE
diff --git a/unsupported/Eigen/CXX11/TensorSymmetry b/unsupported/Eigen/CXX11/TensorSymmetry
index fb1b0c0fb..b09c5e472 100644
--- a/unsupported/Eigen/CXX11/TensorSymmetry
+++ b/unsupported/Eigen/CXX11/TensorSymmetry
@@ -10,9 +10,9 @@
 #ifndef EIGEN_CXX11_TENSORSYMMETRY_MODULE
 #define EIGEN_CXX11_TENSORSYMMETRY_MODULE
 
-#include <unsupported/Eigen/CXX11/Tensor>
+#include "Tensor"
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
 #include "src/util/CXX11Meta.h"
 
@@ -33,7 +33,7 @@
 #include "src/TensorSymmetry/StaticSymmetry.h"
 #include "src/TensorSymmetry/DynamicSymmetry.h"
 
-#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_CXX11_TENSORSYMMETRY_MODULE
 
diff --git a/unsupported/Eigen/CXX11/ThreadPool b/unsupported/Eigen/CXX11/ThreadPool
index 1dcc4eb6c..d046af9b2 100644
--- a/unsupported/Eigen/CXX11/ThreadPool
+++ b/unsupported/Eigen/CXX11/ThreadPool
@@ -12,7 +12,7 @@
 
 #include "../../../Eigen/Core"
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
 /** \defgroup CXX11_ThreadPool_Module C++11 ThreadPool Module
   *
@@ -69,6 +69,6 @@
 
 #endif
 
-#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_CXX11_THREADPOOL_MODULE
diff --git a/unsupported/Eigen/EulerAngles b/unsupported/Eigen/EulerAngles
index 521fa3f76..f8f1c5d0b 100644
--- a/unsupported/Eigen/EulerAngles
+++ b/unsupported/Eigen/EulerAngles
@@ -11,10 +11,10 @@
 #define EIGEN_EULERANGLES_MODULE_H
 
 
-#include "Eigen/Core"
-#include "Eigen/Geometry"
+#include "../../Eigen/Core"
+#include "../../Eigen/Geometry"
 
-#include "Eigen/src/Core/util/DisableStupidWarnings.h"
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
 namespace Eigen {
 
@@ -38,6 +38,6 @@ namespace Eigen {
 #include "src/EulerAngles/EulerSystem.h"
 #include "src/EulerAngles/EulerAngles.h"
 
-#include "Eigen/src/Core/util/ReenableStupidWarnings.h"
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_EULERANGLES_MODULE_H
diff --git a/unsupported/Eigen/FFT b/unsupported/Eigen/FFT
index 8ae053677..d9ad21a5a 100644
--- a/unsupported/Eigen/FFT
+++ b/unsupported/Eigen/FFT
@@ -13,7 +13,7 @@
 #include <complex>
 #include <vector>
 #include <map>
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
 
 /**
diff --git a/unsupported/Eigen/IterativeSolvers b/unsupported/Eigen/IterativeSolvers
index fe8a6d00d..0fa129a7b 100644
--- a/unsupported/Eigen/IterativeSolvers
+++ b/unsupported/Eigen/IterativeSolvers
@@ -10,7 +10,7 @@
 #ifndef EIGEN_ITERATIVE_SOLVERS_MODULE_H
 #define EIGEN_ITERATIVE_SOLVERS_MODULE_H
 
-#include <Eigen/Sparse>
+#include "../../Eigen/Sparse"
 #include "../../Eigen/Jacobi"
 #include "../../Eigen/Householder"
 
diff --git a/unsupported/Eigen/LevenbergMarquardt b/unsupported/Eigen/LevenbergMarquardt
index e77b746c9..109050501 100644
--- a/unsupported/Eigen/LevenbergMarquardt
+++ b/unsupported/Eigen/LevenbergMarquardt
@@ -12,12 +12,12 @@
 
 // #include <vector>
 
-#include <Eigen/Core>
-#include <Eigen/Jacobi>
-#include <Eigen/QR>
-#include <unsupported/Eigen/NumericalDiff> 
+#include "../../Eigen/Core"
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/QR"
+#include "NumericalDiff"
 
-#include <Eigen/SparseQR>
+#include "../../Eigen/SparseQR"
 
 /**
   * \defgroup LevenbergMarquardt_Module Levenberg-Marquardt module
@@ -29,7 +29,7 @@
   * 
   */
 
-#include "Eigen/SparseCore"
+#include "../../Eigen/SparseCore"
 
 #include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport
index b770abc4a..6392bea91 100644
--- a/unsupported/Eigen/MPRealSupport
+++ b/unsupported/Eigen/MPRealSupport
@@ -12,7 +12,7 @@
 #ifndef EIGEN_MPREALSUPPORT_MODULE_H
 #define EIGEN_MPREALSUPPORT_MODULE_H
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 #include <mpreal.h>
 
 namespace Eigen {
diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions
index 17eef8813..06acde68d 100644
--- a/unsupported/Eigen/MatrixFunctions
+++ b/unsupported/Eigen/MatrixFunctions
@@ -14,9 +14,9 @@
 #include <cfloat>
 #include <list>
 
-#include <Eigen/Core>
-#include <Eigen/LU>
-#include <Eigen/Eigenvalues>
+#include "../../Eigen/Core"
+#include "../../Eigen/LU"
+#include "../../Eigen/Eigenvalues"
 
 /**
   * \defgroup MatrixFunctions_Module Matrix functions module
diff --git a/unsupported/Eigen/MoreVectorization b/unsupported/Eigen/MoreVectorization
index 470e72430..7662b4780 100644
--- a/unsupported/Eigen/MoreVectorization
+++ b/unsupported/Eigen/MoreVectorization
@@ -9,7 +9,7 @@
 #ifndef EIGEN_MOREVECTORIZATION_MODULE_H
 #define EIGEN_MOREVECTORIZATION_MODULE_H
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
 namespace Eigen {
 
diff --git a/unsupported/Eigen/NonLinearOptimization b/unsupported/Eigen/NonLinearOptimization
index 28eaa5f29..961f192b5 100644
--- a/unsupported/Eigen/NonLinearOptimization
+++ b/unsupported/Eigen/NonLinearOptimization
@@ -12,10 +12,10 @@
 
 #include <vector>
 
-#include <Eigen/Core>
-#include <Eigen/Jacobi>
-#include <Eigen/QR>
-#include <unsupported/Eigen/NumericalDiff>
+#include "../../Eigen/Core"
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/QR"
+#include "NumericalDiff"
 
 /**
   * \defgroup NonLinearOptimization_Module Non linear optimization module
diff --git a/unsupported/Eigen/NumericalDiff b/unsupported/Eigen/NumericalDiff
index 433334ca8..0668f960f 100644
--- a/unsupported/Eigen/NumericalDiff
+++ b/unsupported/Eigen/NumericalDiff
@@ -10,7 +10,7 @@
 #ifndef EIGEN_NUMERICALDIFF_MODULE
 #define EIGEN_NUMERICALDIFF_MODULE
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
 namespace Eigen {
 
diff --git a/unsupported/Eigen/OpenGLSupport b/unsupported/Eigen/OpenGLSupport
index 489fd8354..f8c213003 100644
--- a/unsupported/Eigen/OpenGLSupport
+++ b/unsupported/Eigen/OpenGLSupport
@@ -10,7 +10,7 @@
 #ifndef EIGEN_OPENGL_MODULE
 #define EIGEN_OPENGL_MODULE
 
-#include <Eigen/Geometry>
+#include "../../Eigen/Geometry"
 
 #if defined(__APPLE_CC__)
   #include <OpenGL/gl.h>
diff --git a/unsupported/Eigen/Polynomials b/unsupported/Eigen/Polynomials
index 334b03142..146e5c404 100644
--- a/unsupported/Eigen/Polynomials
+++ b/unsupported/Eigen/Polynomials
@@ -9,11 +9,11 @@
 #ifndef EIGEN_POLYNOMIALS_MODULE_H
 #define EIGEN_POLYNOMIALS_MODULE_H
 
-#include <Eigen/Core>
+#include "../../Eigen/Core"
 
-#include <Eigen/Eigenvalues>
+#include "../../Eigen/Eigenvalues"
 
-#include <Eigen/src/Core/util/DisableStupidWarnings.h>
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
 // Note that EIGEN_HIDE_HEAVY_CODE has to be defined per module
 #if (defined EIGEN_EXTERN_INSTANTIATIONS) && (EIGEN_EXTERN_INSTANTIATIONS>=2)
@@ -132,7 +132,7 @@
   Output: \verbinclude PolynomialSolver1.out
 */
 
-#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_POLYNOMIALS_MODULE_H
 /* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/unsupported/Eigen/Skyline b/unsupported/Eigen/Skyline
index 71a68cb42..ebdf143f7 100644
--- a/unsupported/Eigen/Skyline
+++ b/unsupported/Eigen/Skyline
@@ -10,9 +10,9 @@
 #define EIGEN_SKYLINE_MODULE_H
 
 
-#include "Eigen/Core"
+#include "../../Eigen/Core"
 
-#include "Eigen/src/Core/util/DisableStupidWarnings.h"
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
 
 #include <map>
 #include <cstdlib>
@@ -34,6 +34,6 @@
 #include "src/Skyline/SkylineInplaceLU.h"
 #include "src/Skyline/SkylineProduct.h"
 
-#include "Eigen/src/Core/util/ReenableStupidWarnings.h"
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_SKYLINE_MODULE_H
diff --git a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
index 9f7bff764..3c6cfb1e3 100644
--- a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
+++ b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
@@ -25,7 +25,7 @@
 #ifndef EIGEN_ARPACKGENERALIZEDSELFADJOINTEIGENSOLVER_H
 #define EIGEN_ARPACKGENERALIZEDSELFADJOINTEIGENSOLVER_H
 
-#include <Eigen/Dense>
+#include "../../../../Eigen/Dense"
 
 namespace Eigen { 
 
diff --git a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
index 37d5b4c6c..5f7cdf29a 100644
--- a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
+++ b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h
@@ -31,7 +31,7 @@
 #ifndef EIGEN_CONSTRAINEDCG_H
 #define EIGEN_CONSTRAINEDCG_H
 
-#include <Eigen/Core>
+#include "../../../../Eigen/Core"
 
 namespace Eigen { 
 
diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
index 85a4f696c..2aea53dcb 100644
--- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
+++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_DGMRES_H
 #define EIGEN_DGMRES_H
 
-#include <Eigen/Eigenvalues>
+#include "../../../../Eigen/Eigenvalues"
 
 namespace Eigen { 
   
diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h
index 1a4b80a2f..850f78a5e 100644
--- a/unsupported/Eigen/src/Splines/SplineFitting.h
+++ b/unsupported/Eigen/src/Splines/SplineFitting.h
@@ -17,8 +17,8 @@
 
 #include "SplineFwd.h"
 
-#include <Eigen/LU>
-#include <Eigen/QR>
+#include "../../../../Eigen/LU"
+#include "../../../../Eigen/QR"
 
 namespace Eigen
 {
diff --git a/unsupported/Eigen/src/Splines/SplineFwd.h b/unsupported/Eigen/src/Splines/SplineFwd.h
index 0a95fbf3e..00d6b4921 100644
--- a/unsupported/Eigen/src/Splines/SplineFwd.h
+++ b/unsupported/Eigen/src/Splines/SplineFwd.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_SPLINES_FWD_H
 #define EIGEN_SPLINES_FWD_H
 
-#include <Eigen/Core>
+#include "../../../../Eigen/Core"
 
 namespace Eigen
 {

From a5cd4e9ad140362bded78a06065bd41d0fe4ecce Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Mon, 17 Sep 2018 10:58:07 -0700
Subject: [PATCH 652/680] Replace deprecated Eigen::DenseIndex with
 Eigen::Index in TensorIndexList

---
 .../Eigen/CXX11/src/Tensor/TensorIndexList.h  | 176 +++++++++---------
 unsupported/test/cxx11_tensor_index_list.cpp  |  32 ++--
 2 files changed, 104 insertions(+), 104 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
index c19edd07f..32caccf87 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h
@@ -37,36 +37,36 @@ namespace Eigen {
   * \sa Tensor
   */
 
-template <DenseIndex n>
+template <Index n>
 struct type2index {
-  static const DenseIndex value = n;
-  EIGEN_DEVICE_FUNC constexpr operator DenseIndex() const { return n; }
-  EIGEN_DEVICE_FUNC void set(DenseIndex val) {
+  static const Index value = n;
+  EIGEN_DEVICE_FUNC constexpr operator Index() const { return n; }
+  EIGEN_DEVICE_FUNC void set(Index val) {
     eigen_assert(val == n);
   }
 };
 
 // This can be used with IndexPairList to get compile-time constant pairs,
 // such as IndexPairList<type2indexpair<1,2>, type2indexpair<3,4>>().
-template <DenseIndex f, DenseIndex s>
+template <Index f, Index s>
 struct type2indexpair {
-  static const DenseIndex first = f;
-  static const DenseIndex second = s;
+  static const Index first = f;
+  static const Index second = s;
 
-  constexpr EIGEN_DEVICE_FUNC operator IndexPair<DenseIndex>() const {
-    return IndexPair<DenseIndex>(f, s);
+  constexpr EIGEN_DEVICE_FUNC operator IndexPair<Index>() const {
+    return IndexPair<Index>(f, s);
   }
 
-  EIGEN_DEVICE_FUNC void set(const IndexPair<DenseIndex>& val) {
+  EIGEN_DEVICE_FUNC void set(const IndexPair<Index>& val) {
     eigen_assert(val.first == f);
     eigen_assert(val.second == s);
   }
 };
 
 
-template<DenseIndex n> struct NumTraits<type2index<n> >
+template<Index n> struct NumTraits<type2index<n> >
 {
-  typedef DenseIndex Real;
+  typedef Index Real;
   enum {
     IsComplex = 0,
     RequireInitialization = false,
@@ -83,20 +83,20 @@ template<DenseIndex n> struct NumTraits<type2index<n> >
 
 namespace internal {
 template <typename T>
-EIGEN_DEVICE_FUNC void update_value(T& val, DenseIndex new_val) {
+EIGEN_DEVICE_FUNC void update_value(T& val, Index new_val) {
   val = internal::convert_index<T>(new_val);
 }
-template <DenseIndex n>
-EIGEN_DEVICE_FUNC void update_value(type2index<n>& val, DenseIndex new_val) {
+template <Index n>
+EIGEN_DEVICE_FUNC void update_value(type2index<n>& val, Index new_val) {
   val.set(new_val);
 }
 
 template <typename T>
-EIGEN_DEVICE_FUNC void update_value(T& val, IndexPair<DenseIndex> new_val) {
+EIGEN_DEVICE_FUNC void update_value(T& val, IndexPair<Index> new_val) {
   val = new_val;
 }
-template <DenseIndex f, DenseIndex s>
-EIGEN_DEVICE_FUNC void update_value(type2indexpair<f, s>& val, IndexPair<DenseIndex> new_val) {
+template <Index f, Index s>
+EIGEN_DEVICE_FUNC void update_value(type2indexpair<f, s>& val, IndexPair<Index> new_val) {
   val.set(new_val);
 }
 
@@ -106,36 +106,36 @@ struct is_compile_time_constant {
   static constexpr bool value = false;
 };
 
-template <DenseIndex idx>
+template <Index idx>
 struct is_compile_time_constant<type2index<idx> > {
   static constexpr bool value = true;
 };
-template <DenseIndex idx>
+template <Index idx>
 struct is_compile_time_constant<const type2index<idx> > {
   static constexpr bool value = true;
 };
-template <DenseIndex idx>
+template <Index idx>
 struct is_compile_time_constant<type2index<idx>& > {
   static constexpr bool value = true;
 };
-template <DenseIndex idx>
+template <Index idx>
 struct is_compile_time_constant<const type2index<idx>& > {
   static constexpr bool value = true;
 };
 
-template <DenseIndex f, DenseIndex s>
+template <Index f, Index s>
 struct is_compile_time_constant<type2indexpair<f, s> > {
   static constexpr bool value = true;
 };
-template <DenseIndex f, DenseIndex s>
+template <Index f, Index s>
 struct is_compile_time_constant<const type2indexpair<f, s> > {
   static constexpr bool value = true;
 };
-template <DenseIndex f, DenseIndex s>
+template <Index f, Index s>
 struct is_compile_time_constant<type2indexpair<f, s>& > {
   static constexpr bool value = true;
 };
-template <DenseIndex f, DenseIndex s>
+template <Index f, Index s>
 struct is_compile_time_constant<const type2indexpair<f, s>& > {
   static constexpr bool value = true;
 };
@@ -228,15 +228,15 @@ template <typename T, typename... O>
 
 
 
-template <DenseIndex Idx, typename ValueT>
+template <Index Idx, typename ValueT>
 struct tuple_coeff {
   template <typename... T>
-  EIGEN_DEVICE_FUNC static constexpr ValueT get(const DenseIndex i, const IndexTuple<T...>& t) {
+  EIGEN_DEVICE_FUNC static constexpr ValueT get(const Index i, const IndexTuple<T...>& t) {
     //    return array_get<Idx>(t) * (i == Idx) + tuple_coeff<Idx-1>::get(i, t) * (i != Idx);
     return (i == Idx ? array_get<Idx>(t) : tuple_coeff<Idx-1, ValueT>::get(i, t));
   }
   template <typename... T>
-  EIGEN_DEVICE_FUNC static void set(const DenseIndex i, IndexTuple<T...>& t, const ValueT& value) {
+  EIGEN_DEVICE_FUNC static void set(const Index i, IndexTuple<T...>& t, const ValueT& value) {
     if (i == Idx) {
       update_value(array_get<Idx>(t), value);
     } else {
@@ -245,7 +245,7 @@ struct tuple_coeff {
   }
 
   template <typename... T>
-  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const DenseIndex i, const IndexTuple<T...>& t) {
+  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const Index i, const IndexTuple<T...>& t) {
     return ((i == Idx) & is_compile_time_constant<typename IndexTupleExtractor<Idx, T...>::ValType>::value) ||
         tuple_coeff<Idx-1, ValueT>::value_known_statically(i, t);
   }
@@ -268,17 +268,17 @@ struct tuple_coeff {
 template <typename ValueT>
 struct tuple_coeff<0, ValueT> {
   template <typename... T>
-  EIGEN_DEVICE_FUNC static constexpr ValueT get(const DenseIndex /*i*/, const IndexTuple<T...>& t) {
+  EIGEN_DEVICE_FUNC static constexpr ValueT get(const Index /*i*/, const IndexTuple<T...>& t) {
     //  eigen_assert (i == 0);  // gcc fails to compile assertions in constexpr
     return array_get<0>(t)/* * (i == 0)*/;
   }
   template <typename... T>
-  EIGEN_DEVICE_FUNC static void set(const DenseIndex i, IndexTuple<T...>& t, const ValueT value) {
+  EIGEN_DEVICE_FUNC static void set(const Index i, IndexTuple<T...>& t, const ValueT value) {
     eigen_assert (i == 0);
     update_value(array_get<0>(t), value);
   }
   template <typename... T>
-  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const DenseIndex i, const IndexTuple<T...>&) {
+  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const Index i, const IndexTuple<T...>&) {
     return is_compile_time_constant<typename IndexTupleExtractor<0, T...>::ValType>::value & (i == 0);
   }
 
@@ -298,29 +298,29 @@ struct tuple_coeff<0, ValueT> {
 
 template<typename FirstType, typename... OtherTypes>
 struct IndexList : internal::IndexTuple<FirstType, OtherTypes...> {
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr DenseIndex operator[] (const DenseIndex i) const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::get(i, *this);
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr Index operator[] (const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::get(i, *this);
   }
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr DenseIndex get(const DenseIndex i) const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::get(i, *this);
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr Index get(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::get(i, *this);
   }
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const DenseIndex i, const DenseIndex value) {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::set(i, *this, value);
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const Index i, const Index value) {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::set(i, *this, value);
   }
 
   EIGEN_DEVICE_FUNC constexpr IndexList(const internal::IndexTuple<FirstType, OtherTypes...>& other) : internal::IndexTuple<FirstType, OtherTypes...>(other) { }
   EIGEN_DEVICE_FUNC constexpr IndexList(FirstType& first, OtherTypes... other) : internal::IndexTuple<FirstType, OtherTypes...>(first, other...) { }
   EIGEN_DEVICE_FUNC constexpr IndexList() : internal::IndexTuple<FirstType, OtherTypes...>() { }
 
-  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const DenseIndex i) const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::value_known_statically(i, *this);
+  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::value_known_statically(i, *this);
   }
   EIGEN_DEVICE_FUNC constexpr bool all_values_known_statically() const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::values_up_to_known_statically(*this);
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::values_up_to_known_statically(*this);
   }
 
   EIGEN_DEVICE_FUNC constexpr bool values_statically_known_to_increase() const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::values_up_to_statically_known_to_increase(*this);
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::values_up_to_statically_known_to_increase(*this);
   }
 };
 
@@ -333,26 +333,26 @@ constexpr IndexList<FirstType, OtherTypes...> make_index_list(FirstType val1, Ot
 
 template<typename FirstType, typename... OtherTypes>
 struct IndexPairList : internal::IndexTuple<FirstType, OtherTypes...> {
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr IndexPair<DenseIndex> operator[] (const DenseIndex i) const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, IndexPair<DenseIndex>>::get(i, *this);
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr IndexPair<Index> operator[] (const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, IndexPair<Index>>::get(i, *this);
   }
-  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const DenseIndex i, const IndexPair<DenseIndex> value) {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...>>::value-1, IndexPair<DenseIndex> >::set(i, *this, value);
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const Index i, const IndexPair<Index> value) {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...>>::value-1, IndexPair<Index> >::set(i, *this, value);
   }
 
   EIGEN_DEVICE_FUNC  constexpr IndexPairList(const internal::IndexTuple<FirstType, OtherTypes...>& other) : internal::IndexTuple<FirstType, OtherTypes...>(other) { }
   EIGEN_DEVICE_FUNC  constexpr IndexPairList() : internal::IndexTuple<FirstType, OtherTypes...>() { }
 
-  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const DenseIndex i) const {
-    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, DenseIndex>::value_known_statically(i, *this);
+  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::value_known_statically(i, *this);
   }
 };
 
 namespace internal {
 
 template<typename FirstType, typename... OtherTypes>
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
-  DenseIndex result = 1;
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
+  Index result = 1;
   for (size_t i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
     result *= sizes[i];
   }
@@ -373,30 +373,30 @@ template<typename FirstType, typename... OtherTypes> struct array_size<const Ind
   static const size_t value = std::tuple_size<std::tuple<FirstType, OtherTypes...> >::value;
 };
 
-template<DenseIndex N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr DenseIndex array_get(IndexList<FirstType, OtherTypes...>& a) {
+template<Index N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr Index array_get(IndexList<FirstType, OtherTypes...>& a) {
   return IndexTupleExtractor<N, FirstType, OtherTypes...>::get_val(a);
 }
-template<DenseIndex N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr DenseIndex array_get(const IndexList<FirstType, OtherTypes...>& a) {
+template<Index N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr Index array_get(const IndexList<FirstType, OtherTypes...>& a) {
   return IndexTupleExtractor<N, FirstType, OtherTypes...>::get_val(a);
 }
 
 template <typename T>
 struct index_known_statically_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_known_statically_impl<IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i);
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_known_statically_impl<const IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i);
   }
 };
@@ -448,14 +448,14 @@ template <typename FirstType, typename... OtherTypes>
 
 template <typename Tx>
 struct index_statically_eq_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_eq_impl<IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) == value);
   }
@@ -463,7 +463,7 @@ struct index_statically_eq_impl<IndexList<FirstType, OtherTypes...> > {
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_eq_impl<const IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) == value);
   }
@@ -472,14 +472,14 @@ struct index_statically_eq_impl<const IndexList<FirstType, OtherTypes...> > {
 
 template <typename T>
 struct index_statically_ne_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_ne_impl<IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) != value);
   }
@@ -487,7 +487,7 @@ struct index_statically_ne_impl<IndexList<FirstType, OtherTypes...> > {
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_ne_impl<const IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) != value);
   }
@@ -496,14 +496,14 @@ struct index_statically_ne_impl<const IndexList<FirstType, OtherTypes...> > {
 
 template <typename T>
 struct index_statically_gt_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_gt_impl<IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) > value);
   }
@@ -511,7 +511,7 @@ struct index_statically_gt_impl<IndexList<FirstType, OtherTypes...> > {
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_gt_impl<const IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) > value);
   }
@@ -521,14 +521,14 @@ struct index_statically_gt_impl<const IndexList<FirstType, OtherTypes...> > {
 
 template <typename T>
 struct index_statically_lt_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_lt_impl<IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) < value);
   }
@@ -536,7 +536,7 @@ struct index_statically_lt_impl<IndexList<FirstType, OtherTypes...> > {
 
 template <typename FirstType, typename... OtherTypes>
 struct index_statically_lt_impl<const IndexList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexList<FirstType, OtherTypes...>().get(i) < value);
   }
@@ -546,14 +546,14 @@ struct index_statically_lt_impl<const IndexList<FirstType, OtherTypes...> > {
 
 template <typename Tx>
 struct index_pair_first_statically_eq_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_pair_first_statically_eq_impl<IndexPairList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexPairList<FirstType, OtherTypes...>().operator[](i).first == value);
   }
@@ -561,7 +561,7 @@ struct index_pair_first_statically_eq_impl<IndexPairList<FirstType, OtherTypes..
 
 template <typename FirstType, typename... OtherTypes>
 struct index_pair_first_statically_eq_impl<const IndexPairList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexPairList<FirstType, OtherTypes...>().operator[](i).first == value);
   }
@@ -571,14 +571,14 @@ struct index_pair_first_statically_eq_impl<const IndexPairList<FirstType, OtherT
 
 template <typename Tx>
 struct index_pair_second_statically_eq_impl {
-  EIGEN_DEVICE_FUNC static constexpr bool run(DenseIndex, DenseIndex) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename FirstType, typename... OtherTypes>
 struct index_pair_second_statically_eq_impl<IndexPairList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexPairList<FirstType, OtherTypes...>().operator[](i).second == value);
   }
@@ -586,7 +586,7 @@ struct index_pair_second_statically_eq_impl<IndexPairList<FirstType, OtherTypes.
 
 template <typename FirstType, typename... OtherTypes>
 struct index_pair_second_statically_eq_impl<const IndexPairList<FirstType, OtherTypes...> > {
-  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
     return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
         (IndexPairList<FirstType, OtherTypes...>().operator[](i).second == value);
   }
@@ -603,7 +603,7 @@ namespace internal {
 
 template <typename T>
 struct index_known_statically_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const Index) {
     return false;
   }
 };
@@ -624,42 +624,42 @@ struct indices_statically_known_to_increase_impl {
 
 template <typename T>
 struct index_statically_eq_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename T>
 struct index_statically_ne_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename T>
 struct index_statically_gt_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename T>
 struct index_statically_lt_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename Tx>
 struct index_pair_first_statically_eq_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
 
 template <typename Tx>
 struct index_pair_second_statically_eq_impl {
-  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(DenseIndex, DenseIndex) {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
     return false;
   }
 };
@@ -675,7 +675,7 @@ struct index_pair_second_statically_eq_impl {
 namespace Eigen {
 namespace internal {
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_known_statically(DenseIndex i) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_known_statically(Index i) {
   return index_known_statically_impl<T>::run(i);
 }
 
@@ -690,32 +690,32 @@ static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool indices_statically_known_to_increa
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_eq(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_eq(Index i, Index value) {
   return index_statically_eq_impl<T>::run(i, value);
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_ne(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_ne(Index i, Index value) {
   return index_statically_ne_impl<T>::run(i, value);
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_gt(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_gt(Index i, Index value) {
   return index_statically_gt_impl<T>::run(i, value);
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_lt(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_lt(Index i, Index value) {
   return index_statically_lt_impl<T>::run(i, value);
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_first_statically_eq(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_first_statically_eq(Index i, Index value) {
   return index_pair_first_statically_eq_impl<T>::run(i, value);
 }
 
 template <typename T>
-static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_second_statically_eq(DenseIndex i, DenseIndex value) {
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_second_statically_eq(Index i, Index value) {
   return index_pair_second_statically_eq_impl<T>::run(i, value);
 }
 
diff --git a/unsupported/test/cxx11_tensor_index_list.cpp b/unsupported/test/cxx11_tensor_index_list.cpp
index 294677a4d..2166532c8 100644
--- a/unsupported/test/cxx11_tensor_index_list.cpp
+++ b/unsupported/test/cxx11_tensor_index_list.cpp
@@ -22,9 +22,9 @@ static void test_static_index_list()
   VERIFY_IS_EQUAL(internal::array_get<0>(reduction_axis), 0);
   VERIFY_IS_EQUAL(internal::array_get<1>(reduction_axis), 1);
   VERIFY_IS_EQUAL(internal::array_get<2>(reduction_axis), 2);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[0]), 0);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[1]), 1);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[2]), 2);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[0]), 0);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[1]), 1);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[2]), 2);
 
   EIGEN_STATIC_ASSERT((internal::array_get<0>(reduction_axis) == 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
   EIGEN_STATIC_ASSERT((internal::array_get<1>(reduction_axis) == 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
@@ -167,18 +167,18 @@ static void test_type2indexpair_list()
 
   typedef Eigen::IndexPairList<Eigen::type2indexpair<0,10>> Dims0;
   typedef Eigen::IndexPairList<Eigen::type2indexpair<0,10>, Eigen::type2indexpair<1,11>, Eigen::type2indexpair<2,12>> Dims2_a;
-  typedef Eigen::IndexPairList<Eigen::type2indexpair<0,10>, Eigen::IndexPair<DenseIndex>, Eigen::type2indexpair<2,12>> Dims2_b;
-  typedef Eigen::IndexPairList<Eigen::IndexPair<DenseIndex>, Eigen::type2indexpair<1,11>, Eigen::IndexPair<DenseIndex>> Dims2_c;
+  typedef Eigen::IndexPairList<Eigen::type2indexpair<0,10>, Eigen::IndexPair<Index>, Eigen::type2indexpair<2,12>> Dims2_b;
+  typedef Eigen::IndexPairList<Eigen::IndexPair<Index>, Eigen::type2indexpair<1,11>, Eigen::IndexPair<Index>> Dims2_c;
 
   Dims2_a d2_a;
 
   Dims2_b d2_b;
-  d2_b.set(1, Eigen::IndexPair<DenseIndex>(1,11));
+  d2_b.set(1, Eigen::IndexPair<Index>(1,11));
 
   Dims2_c d2_c;
-  d2_c.set(0, Eigen::IndexPair<DenseIndex>(Eigen::IndexPair<DenseIndex>(0,10)));
-  d2_c.set(1, Eigen::IndexPair<DenseIndex>(1,11));  // setting type2indexpair to correct value.
-  d2_c.set(2, Eigen::IndexPair<DenseIndex>(2,12));
+  d2_c.set(0, Eigen::IndexPair<Index>(Eigen::IndexPair<Index>(0,10)));
+  d2_c.set(1, Eigen::IndexPair<Index>(1,11));  // setting type2indexpair to correct value.
+  d2_c.set(2, Eigen::IndexPair<Index>(2,12));
 
   VERIFY_IS_EQUAL(d2_a[0].first, 0);
   VERIFY_IS_EQUAL(d2_a[0].second, 10);
@@ -277,9 +277,9 @@ static void test_dynamic_index_list()
   VERIFY_IS_EQUAL(internal::array_get<0>(reduction_axis), 2);
   VERIFY_IS_EQUAL(internal::array_get<1>(reduction_axis), 1);
   VERIFY_IS_EQUAL(internal::array_get<2>(reduction_axis), 0);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[0]), 2);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[1]), 1);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[2]), 0);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[0]), 2);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[1]), 1);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[2]), 0);
 
   Tensor<float, 1> result = tensor.sum(reduction_axis);
   for (int i = 0; i < result.size(); ++i) {
@@ -309,10 +309,10 @@ static void test_mixed_index_list()
   VERIFY_IS_EQUAL(internal::array_get<1>(reduction_axis), 1);
   VERIFY_IS_EQUAL(internal::array_get<2>(reduction_axis), 2);
   VERIFY_IS_EQUAL(internal::array_get<3>(reduction_axis), 3);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[0]), 0);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[1]), 1);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[2]), 2);
-  VERIFY_IS_EQUAL(static_cast<DenseIndex>(reduction_axis[3]), 3);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[0]), 0);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[1]), 1);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[2]), 2);
+  VERIFY_IS_EQUAL(static_cast<Index>(reduction_axis[3]), 3);
 
   typedef IndexList<type2index<0>, int, type2index<2>, int> ReductionIndices;
   ReductionIndices reduction_indices;

From 1f0c941c3dee8b30814828c7e11f764ccc4f2a5f Mon Sep 17 00:00:00 2001
From: Ravi Kiran <c.ravikiran123@gmail.com>
Date: Mon, 17 Sep 2018 18:29:12 -0700
Subject: [PATCH 653/680] Collapsed revision * Merged eigen/eigen into default

---
 .../src/ThreadPool/NonBlockingThreadPool.h    | 222 +++++++++++++-----
 .../src/ThreadPool/ThreadPoolInterface.h      |   9 +
 .../test/cxx11_non_blocking_thread_pool.cpp   |  53 +++++
 3 files changed, 222 insertions(+), 62 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 60a0c9fb6..2130625bc 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -26,9 +26,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       : env_(env),
         num_threads_(num_threads),
         allow_spinning_(allow_spinning),
-        threads_(num_threads),
-        queues_(num_threads),
-        coprimes_(num_threads),
+        thread_data_(num_threads),
+        all_coprimes_(num_threads),
         waiters_(num_threads),
         blocked_(0),
         spinning_(0),
@@ -36,33 +35,26 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
         cancelled_(false),
         ec_(waiters_) {
     waiters_.resize(num_threads_);
-
-    // Calculate coprimes of num_threads_.
-    // Coprimes are used for a random walk over all threads in Steal
+    // Calculate coprimes of all numbers [1, num_threads].
+    // Coprimes are used for random walks over all threads in Steal
     // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
-    // a walk starting thread index t and calculate num_threads - 1 subsequent
+    // a random starting thread index t and calculate num_threads - 1 subsequent
     // indices as (t + coprime) % num_threads, we will cover all threads without
     // repetitions (effectively getting a presudo-random permutation of thread
     // indices).
-    for (int i = 1; i <= num_threads_; i++) {
-      unsigned a = i;
-      unsigned b = num_threads_;
-      // If GCD(a, b) == 1, then a and b are coprimes.
-      while (b != 0) {
-        unsigned tmp = a;
-        a = b;
-        b = tmp % b;
-      }
-      if (a == 1) {
-        coprimes_.push_back(i);
-      }
+    eigen_assert(num_threads_ < kMaxThreads);
+    for (int i = 1; i <= num_threads_; ++i) {
+      all_coprimes_.emplace_back(i);
+      ComputeCoprimes(i, &all_coprimes_.back());
     }
-    queues_.resize(num_threads_);
 #ifndef EIGEN_THREAD_LOCAL
     init_barrier_.reset(new Barrier(num_threads_));
 #endif
+    thread_data_.resize(num_threads_);
     for (int i = 0; i < num_threads_; i++) {
-      threads_.emplace_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
+      SetStealPartition(i, EncodePartition(0, num_threads_));
+      thread_data_[i].thread.reset(
+          env_.CreateThread([this, i]() { WorkerLoop(i); }));
     }
 #ifndef EIGEN_THREAD_LOCAL
     // Wait for workers to initialize per_thread_map_. Otherwise we might race
@@ -82,27 +74,51 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     } else {
       // Since we were cancelled, there might be entries in the queues.
       // Empty them to prevent their destructor from asserting.
-      for (size_t i = 0; i < queues_.size(); i++) {
-        queues_[i].Flush();
+      for (size_t i = 0; i < thread_data_.size(); i++) {
+        thread_data_[i].queue.Flush();
       }
     }
+    // Join threads explicitly (by destroying) to avoid destruction order within
+    // this class.
+    for (size_t i = 0; i < thread_data_.size(); ++i)
+      thread_data_[i].thread.reset();
+  }
 
-    // Join threads explicitly to avoid destruction order issues.
-    threads_.resize(0);
-    queues_.resize(0);
+  void SetStealPartitions(const std::vector<std::pair<unsigned, unsigned>>& partitions) {
+    int num_partitions = partitions.size();
+    eigen_assert(num_partitions == num_threads_);
+
+    // Pass this information to each thread queue.
+    for (int i = 0; i < num_threads_; i++) {
+      const auto& pair = partitions[i];
+      unsigned start = pair.first, end = pair.second;
+      AssertBounds(start, end);
+      unsigned val = EncodePartition(start, end);
+      SetStealPartition(i, val);
+    }
   }
 
   void Schedule(std::function<void()> fn) {
+    ScheduleWithHint(std::move(fn), 0, num_threads_);
+  }
+
+  void ScheduleWithHint(std::function<void()> fn, int start,
+                        int limit) override {
     Task t = env_.CreateTask(std::move(fn));
     PerThread* pt = GetPerThread();
     if (pt->pool == this) {
       // Worker thread of this pool, push onto the thread's queue.
-      Queue& q = queues_[pt->thread_id];
+      Queue& q = thread_data_[pt->thread_id].queue;
       t = q.PushFront(std::move(t));
     } else {
       // A free-standing thread (or worker of another pool), push onto a random
       // queue.
-      Queue& q = queues_[Rand(&pt->rand) % queues_.size()];
+      eigen_assert(start < limit);
+      eigen_assert(limit <= num_threads_);
+      int num_queues = limit - start;
+      int rnd = Rand(&pt->rand) % num_queues;
+      eigen_assert(start + rnd < limit);
+      Queue& q = thread_data_[start + rnd].queue;
       t = q.PushBack(std::move(t));
     }
     // Note: below we touch this after making w available to worker threads.
@@ -125,8 +141,8 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
 
     // Let each thread know it's been cancelled.
 #ifdef EIGEN_THREAD_ENV_SUPPORTS_CANCELLATION
-    for (size_t i = 0; i < threads_.size(); i++) {
-      threads_[i]->OnCancel();
+    for (size_t i = 0; i < thread_data_.size(); i++) {
+      thread_data_[i].thread->OnCancel();
     }
 #endif
 
@@ -146,6 +162,56 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   }
 
  private:
+  // Create a single atomic<int> that encodes start and limit information for
+  // each thread.
+  // We expect num_threads_ < 65536, so we can store them in a single
+  // std::atomic<unsigned>.
+  // Exposed publicly as static functions so that external callers can reuse
+  // this encode/decode logic for maintaining their own thread-safe copies of
+  // scheduling and steal domain(s).
+  static const int kMaxPartitionBits = 16;
+  static const int kMaxThreads = 1 << kMaxPartitionBits;
+
+  inline unsigned EncodePartition(unsigned start, unsigned limit) {
+    return (start << kMaxPartitionBits) | limit;
+  }
+
+  inline void DecodePartition(unsigned val, unsigned* start, unsigned* limit) {
+    *limit = val & (kMaxThreads - 1);
+    val >>= kMaxPartitionBits;
+    *start = val;
+  }
+
+  void AssertBounds(int start, int end) {
+    eigen_assert(start >= 0);
+    eigen_assert(start < end);  // non-zero sized partition
+    eigen_assert(end <= num_threads_);
+  }
+
+  inline void SetStealPartition(size_t i, unsigned val) {
+    thread_data_[i].steal_partition.store(val, std::memory_order_relaxed);
+  }
+
+  inline unsigned GetStealPartition(int i) {
+    return thread_data_[i].steal_partition.load(std::memory_order_relaxed);
+  }
+
+  void ComputeCoprimes(int N, MaxSizeVector<unsigned>* coprimes) {
+    for (int i = 1; i <= N; i++) {
+      unsigned a = i;
+      unsigned b = N;
+      // If GCD(a, b) == 1, then a and b are coprimes.
+      while (b != 0) {
+        unsigned tmp = a;
+        a = b;
+        b = tmp % b;
+      }
+      if (a == 1) {
+        coprimes->push_back(i);
+      }
+    }
+  }
+
   typedef typename Environment::EnvThread Thread;
 
   struct PerThread {
@@ -159,12 +225,18 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
 #endif
   };
 
+  struct ThreadData {
+    constexpr ThreadData() : thread(), steal_partition(0), queue() {}
+    std::unique_ptr<Thread> thread;
+    std::atomic<unsigned> steal_partition;
+    Queue queue;
+  };
+
   Environment env_;
   const int num_threads_;
   const bool allow_spinning_;
-  MaxSizeVector<std::unique_ptr<Thread> > threads_;
-  MaxSizeVector<Queue> queues_;
-  MaxSizeVector<unsigned> coprimes_;
+  MaxSizeVector<ThreadData> thread_data_;
+  MaxSizeVector<MaxSizeVector<unsigned>> all_coprimes_;
   MaxSizeVector<EventCount::Waiter> waiters_;
   std::atomic<unsigned> blocked_;
   std::atomic<bool> spinning_;
@@ -191,19 +263,19 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     pt->pool = this;
     pt->rand = GlobalThreadIdHash();
     pt->thread_id = thread_id;
-    Queue& q = queues_[thread_id];
+    Queue& q = thread_data_[thread_id].queue;
     EventCount::Waiter* waiter = &waiters_[thread_id];
-    // TODO(dvyukov,rmlarsen): The time spent in Steal() is proportional
-    // to num_threads_ and we assume that new work is scheduled at a
-    // constant rate, so we set spin_count to 5000 / num_threads_. The
+    // TODO(dvyukov,rmlarsen): The time spent in NonEmptyQueueIndex() is
+    // proportional to num_threads_ and we assume that new work is scheduled at
+    // a constant rate, so we set spin_count to 5000 / num_threads_. The
     // constant was picked based on a fair dice roll, tune it.
     const int spin_count =
         allow_spinning_ && num_threads_ > 0 ? 5000 / num_threads_ : 0;
     if (num_threads_ == 1) {
       // For num_threads_ == 1 there is no point in going through the expensive
-      // steal loop. Moreover, since Steal() calls PopBack() on the victim
-      // queues it might reverse the order in which ops are executed compared to
-      // the order in which they are scheduled, which tends to be
+      // steal loop. Moreover, since NonEmptyQueueIndex() calls PopBack() on the
+      // victim queues it might reverse the order in which ops are executed
+      // compared to the order in which they are scheduled, which tends to be
       // counter-productive for the types of I/O workloads the single thread
       // pools tend to be used for.
       while (!cancelled_) {
@@ -226,23 +298,26 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       while (!cancelled_) {
         Task t = q.PopFront();
         if (!t.f) {
-          t = Steal();
+          t = LocalSteal();
           if (!t.f) {
-            // Leave one thread spinning. This reduces latency.
-            if (allow_spinning_ && !spinning_ && !spinning_.exchange(true)) {
-              for (int i = 0; i < spin_count && !t.f; i++) {
-                if (!cancelled_.load(std::memory_order_relaxed)) {
-                  t = Steal();
-                } else {
+            t = GlobalSteal();
+            if (!t.f) {
+              // Leave one thread spinning. This reduces latency.
+              if (allow_spinning_ && !spinning_ && !spinning_.exchange(true)) {
+                for (int i = 0; i < spin_count && !t.f; i++) {
+                  if (!cancelled_.load(std::memory_order_relaxed)) {
+                    t = GlobalSteal();
+                  } else {
+                    return;
+                  }
+                }
+                spinning_ = false;
+              }
+              if (!t.f) {
+                if (!WaitForWork(waiter, &t)) {
                   return;
                 }
               }
-              spinning_ = false;
-            }
-            if (!t.f) {
-              if (!WaitForWork(waiter, &t)) {
-                return;
-              }
             }
           }
         }
@@ -253,15 +328,18 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     }
   }
 
-  // Steal tries to steal work from other worker threads in best-effort manner.
-  Task Steal() {
+  // Steal tries to steal work from other worker threads in the range [start,
+  // limit) in best-effort manner.
+  Task Steal(unsigned start, unsigned limit) {
     PerThread* pt = GetPerThread();
-    const size_t size = queues_.size();
+    const size_t size = limit - start;
     unsigned r = Rand(&pt->rand);
-    unsigned inc = coprimes_[r % coprimes_.size()];
     unsigned victim = r % size;
+    unsigned inc = all_coprimes_[size - 1][r % all_coprimes_[size - 1].size()];
+
     for (unsigned i = 0; i < size; i++) {
-      Task t = queues_[victim].PopBack();
+      eigen_assert(start + victim < limit);
+      Task t = thread_data_[start + victim].queue.PopBack();
       if (t.f) {
         return t;
       }
@@ -273,6 +351,23 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
     return Task();
   }
 
+  // Steals work within threads belonging to the partition.
+  Task LocalSteal() {
+    PerThread* pt = GetPerThread();
+    unsigned partition = GetStealPartition(pt->thread_id);
+    unsigned start, limit;
+    DecodePartition(partition, &start, &limit);
+    AssertBounds(start, limit);
+
+    return Steal(start, limit);
+  }
+
+  // Steals work from any other thread in the pool.
+  Task GlobalSteal() {
+    return Steal(0, num_threads_);
+  }
+
+
   // WaitForWork blocks until new work is available (returns true), or if it is
   // time to exit (returns false). Can optionally return a task to execute in t
   // (in such case t.f != nullptr on return).
@@ -288,7 +383,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
       if (cancelled_) {
         return false;
       } else {
-        *t = queues_[victim].PopBack();
+        *t = thread_data_[victim].queue.PopBack();
         return true;
       }
     }
@@ -325,12 +420,15 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
 
   int NonEmptyQueueIndex() {
     PerThread* pt = GetPerThread();
-    const size_t size = queues_.size();
+    // We intentionally design NonEmptyQueueIndex to steal work from
+    // anywhere in the queue so threads don't block in WaitForWork() forever
+    // when all threads in their partition go to sleep. Steal is still local.
+    const size_t size = thread_data_.size();
     unsigned r = Rand(&pt->rand);
-    unsigned inc = coprimes_[r % coprimes_.size()];
+    unsigned inc = all_coprimes_[size - 1][r % all_coprimes_[size - 1].size()];
     unsigned victim = r % size;
     for (unsigned i = 0; i < size; i++) {
-      if (!queues_[victim].Empty()) {
+      if (!thread_data_[victim].queue.Empty()) {
         return victim;
       }
       victim += inc;
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
index 84e1e6cc0..d1fa4b23e 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
@@ -19,6 +19,15 @@ class ThreadPoolInterface {
   // Submits a closure to be run by a thread in the pool.
   virtual void Schedule(std::function<void()> fn) = 0;
 
+  // Submits a closure to be run by threads in the range [start, end) in the
+  // pool.
+  virtual void ScheduleWithHint(std::function<void()> fn, int start,
+                                int end) {
+    // Just defer to Schedule in case sub-classes aren't interested in
+    // overriding this functionality.
+    Schedule(fn);
+  }
+
   // If implemented, stop processing the closures that have been enqueued.
   // Currently running closures may still be processed.
   // If not implemented, does nothing.
diff --git a/unsupported/test/cxx11_non_blocking_thread_pool.cpp b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
index e73a034b1..90b330fdc 100644
--- a/unsupported/test/cxx11_non_blocking_thread_pool.cpp
+++ b/unsupported/test/cxx11_non_blocking_thread_pool.cpp
@@ -116,10 +116,63 @@ static void test_cancel()
   tp.Cancel();
 }
 
+static void test_pool_partitions() {
+  const int kThreads = 2;
+  ThreadPool tp(kThreads);
+
+  // Assign each thread to its own partition, so that stealing other work only
+  // occurs globally when a thread is idle.
+  std::vector<std::pair<unsigned, unsigned>> steal_partitions(kThreads);
+  for (int i = 0; i < kThreads; ++i) {
+    steal_partitions[i] = std::make_pair(i, i + 1);
+  }
+  tp.SetStealPartitions(steal_partitions);
+
+  std::atomic<int> running(0);
+  std::atomic<int> done(0);
+  std::atomic<int> phase(0);
+
+  // Schedule kThreads tasks and ensure that they all are running.
+  for (int i = 0; i < kThreads; ++i) {
+    tp.Schedule([&]() {
+      const int thread_id = tp.CurrentThreadId();
+      VERIFY_GE(thread_id, 0);
+      VERIFY_LE(thread_id, kThreads - 1);
+      ++running;
+      while (phase < 1) {
+      }
+      ++done;
+    });
+  }
+  while (running != kThreads) {
+  }
+  // Schedule each closure to only run on thread 'i' and verify that it does.
+  for (int i = 0; i < kThreads; ++i) {
+    tp.ScheduleWithHint(
+        [&, i]() {
+          ++running;
+          const int thread_id = tp.CurrentThreadId();
+          VERIFY_IS_EQUAL(thread_id, i);
+          while (phase < 2) {
+          }
+          ++done;
+        },
+        i, i + 1);
+  }
+  running = 0;
+  phase = 1;
+  while (running != kThreads) {
+  }
+  running = 0;
+  phase = 2;
+}
+
+
 EIGEN_DECLARE_TEST(cxx11_non_blocking_thread_pool)
 {
   CALL_SUBTEST(test_create_destroy_empty_pool());
   CALL_SUBTEST(test_parallelism(true));
   CALL_SUBTEST(test_parallelism(false));
   CALL_SUBTEST(test_cancel());
+  CALL_SUBTEST(test_pool_partitions());
 }

From 218a7b98401aee558b96e21d8ba90641eb3e9371 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 18 Sep 2018 10:57:00 -0700
Subject: [PATCH 654/680] Enable DSizes type promotion with c++03 compilers

---
 .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 20 ++++++++++---------
 unsupported/test/cxx11_tensor_dimension.cpp   |  7 ++++---
 2 files changed, 15 insertions(+), 12 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
index 5de0d0de7..dbf5af094 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h
@@ -290,22 +290,24 @@ struct DSizes : array<DenseIndex, NumDims> {
     }
   }
 
-#ifdef EIGEN_HAS_CXX11
   // Enable DSizes index type promotion only if we are promoting to the
   // larger type, e.g. allow to promote dimensions of type int to long.
-  template<typename OtherIndex,
-           typename std::enable_if<
-               std::is_same<
-                 DenseIndex,
-                 typename internal::promote_index_type<DenseIndex, OtherIndex>::type
-               >::value, int>::type = 0>
+  template<typename OtherIndex>
   EIGEN_DEVICE_FUNC
-  explicit DSizes(const array<OtherIndex, NumDims>& other) {
+  explicit DSizes(const array<OtherIndex, NumDims>& other,
+                  // Default template parameters require c++11.
+                  typename internal::enable_if<
+                     internal::is_same<
+                         DenseIndex,
+                         typename internal::promote_index_type<
+                             DenseIndex,
+                             OtherIndex
+                         >::type
+                     >::value, void*>::type = 0) {
     for (int i = 0; i < NumDims; ++i) {
       (*this)[i] = static_cast<DenseIndex>(other[i]);
     }
   }
-#endif  // EIGEN_HAS_CXX11
 
 #ifdef EIGEN_HAS_INDEX_LIST
   template <typename FirstType, typename... OtherTypes>
diff --git a/unsupported/test/cxx11_tensor_dimension.cpp b/unsupported/test/cxx11_tensor_dimension.cpp
index 26f8edd8a..ee416e14a 100644
--- a/unsupported/test/cxx11_tensor_dimension.cpp
+++ b/unsupported/test/cxx11_tensor_dimension.cpp
@@ -61,9 +61,11 @@ static void test_rank_zero()
 }
 
 static void test_index_type_promotion() {
-#ifdef EIGEN_HAS_CXX11
   Eigen::DSizes<int, 3> src0(1, 2, 3);
-  Eigen::array<int, 3> src1 = {4, 5, 6};
+  Eigen::array<int, 3> src1;
+  src1[0] = 4;
+  src1[1] = 5;
+  src1[2] = 6;
 
   Eigen::DSizes<long, 3> dst0(src0);
   Eigen::DSizes<long, 3> dst1(src1);
@@ -74,7 +76,6 @@ static void test_index_type_promotion() {
   VERIFY_IS_EQUAL(dst1[0], 4L);
   VERIFY_IS_EQUAL(dst1[1], 5L);
   VERIFY_IS_EQUAL(dst1[2], 6L);
-#endif  // EIGEN_HAS_CXX11
 }
 
 EIGEN_DECLARE_TEST(cxx11_tensor_dimension)

From 3e8188fc77970e9bce485309431e6c7f278b15e9 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Tue, 18 Sep 2018 21:24:48 +0200
Subject: [PATCH 655/680] bug #1600: initialize m_info to InvalidInput by
 default, even though m_info is not accessible until it has been initialized
 (assert)

---
 Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
index 2d1a82c57..f95606206 100644
--- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
+++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -121,6 +121,7 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
         : m_eivec(),
           m_eivalues(),
           m_subdiag(),
+          m_info(InvalidInput),
           m_isInitialized(false)
     { }
 

From c4627039ac217ea1415b43fc07cc98c8e90c1ada Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 18 Sep 2018 14:25:21 -0700
Subject: [PATCH 656/680] Support static dimensions (aka IndexList) in
 Tensor::resize(...)

---
 unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
index aed71b265..95349d0ed 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h
@@ -477,6 +477,18 @@ class Tensor : public TensorBase<Tensor<Scalar_, NumIndices_, Options_, IndexTyp
       // Nothing to do: rank 0 tensors have fixed size
     }
 
+#ifdef EIGEN_HAS_INDEX_LIST
+    template <typename FirstType, typename... OtherTypes>
+    EIGEN_DEVICE_FUNC
+    void resize(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
+      array<Index, NumIndices> dims;
+      for (int i = 0; i < NumIndices; ++i) {
+        dims[i] = static_cast<Index>(dimensions[i]);
+      }
+      resize(dims);
+    }
+#endif
+
     /** Custom Dimension */
 #ifdef EIGEN_HAS_SFINAE
     template<typename CustomDimension,

From 82772e8d9de3e04d84bcd3aafaf714e88404f784 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 15 Sep 2018 14:16:20 +0200
Subject: [PATCH 657/680] Rename Symbolic namespace to symbolic to be
 consistent with numext namespace

---
 Eigen/src/Core/ArithmeticSequence.h     | 76 ++++++++++++-------------
 Eigen/src/Core/util/IndexedViewHelper.h | 10 ++--
 Eigen/src/Core/util/SymbolicIndex.h     | 12 ++--
 Eigen/src/plugins/IndexedViewMethods.h  |  4 +-
 test/symbolic_index.cpp                 | 10 ++--
 5 files changed, 56 insertions(+), 56 deletions(-)

diff --git a/Eigen/src/Core/ArithmeticSequence.h b/Eigen/src/Core/ArithmeticSequence.h
index 2083cf178..941028893 100644
--- a/Eigen/src/Core/ArithmeticSequence.h
+++ b/Eigen/src/Core/ArithmeticSequence.h
@@ -29,17 +29,17 @@ template<int N> struct aseq_negate<FixedInt<N> > {
 template<> struct aseq_negate<FixedInt<DynamicIndex> > {};
 
 template<typename FirstType,typename SizeType,typename IncrType,
-         bool FirstIsSymbolic=Symbolic::is_symbolic<FirstType>::value,
-         bool SizeIsSymbolic =Symbolic::is_symbolic<SizeType>::value>
+         bool FirstIsSymbolic=symbolic::is_symbolic<FirstType>::value,
+         bool SizeIsSymbolic =symbolic::is_symbolic<SizeType>::value>
 struct aseq_reverse_first_type {
   typedef Index type;
 };
 
 template<typename FirstType,typename SizeType,typename IncrType>
 struct aseq_reverse_first_type<FirstType,SizeType,IncrType,true,true> {
-  typedef Symbolic::AddExpr<FirstType,
-                            Symbolic::ProductExpr<Symbolic::AddExpr<SizeType,Symbolic::ValueExpr<FixedInt<-1> > >,
-                                                  Symbolic::ValueExpr<IncrType> >
+  typedef symbolic::AddExpr<FirstType,
+                            symbolic::ProductExpr<symbolic::AddExpr<SizeType,symbolic::ValueExpr<FixedInt<-1> > >,
+                                                  symbolic::ValueExpr<IncrType> >
                            > type;
 };
 
@@ -56,14 +56,14 @@ struct aseq_reverse_first_type_aux<SizeType,IncrType,typename internal::enable_i
 template<typename FirstType,typename SizeType,typename IncrType>
 struct aseq_reverse_first_type<FirstType,SizeType,IncrType,true,false> {
   typedef typename aseq_reverse_first_type_aux<SizeType,IncrType>::type Aux;
-  typedef Symbolic::AddExpr<FirstType,Symbolic::ValueExpr<Aux> > type;
+  typedef symbolic::AddExpr<FirstType,symbolic::ValueExpr<Aux> > type;
 };
 
 template<typename FirstType,typename SizeType,typename IncrType>
 struct aseq_reverse_first_type<FirstType,SizeType,IncrType,false,true> {
-  typedef Symbolic::AddExpr<Symbolic::ProductExpr<Symbolic::AddExpr<SizeType,Symbolic::ValueExpr<FixedInt<-1> > >,
-                                                  Symbolic::ValueExpr<IncrType> >,
-                            Symbolic::ValueExpr<> > type;
+  typedef symbolic::AddExpr<symbolic::ProductExpr<symbolic::AddExpr<SizeType,symbolic::ValueExpr<FixedInt<-1> > >,
+                                                  symbolic::ValueExpr<IncrType> >,
+                            symbolic::ValueExpr<> > type;
 };
 #endif
 
@@ -229,7 +229,7 @@ auto seq(FirstType f, LastType l, IncrType incr)
 #else // EIGEN_HAS_CXX11
 
 template<typename FirstType,typename LastType>
-typename internal::enable_if<!(Symbolic::is_symbolic<FirstType>::value || Symbolic::is_symbolic<LastType>::value),
+typename internal::enable_if<!(symbolic::is_symbolic<FirstType>::value || symbolic::is_symbolic<LastType>::value),
                              ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,Index> >::type
 seq(FirstType f, LastType l)
 {
@@ -238,35 +238,35 @@ seq(FirstType f, LastType l)
 }
 
 template<typename FirstTypeDerived,typename LastType>
-typename internal::enable_if<!Symbolic::is_symbolic<LastType>::value,
-    ArithmeticSequence<FirstTypeDerived, Symbolic::AddExpr<Symbolic::AddExpr<Symbolic::NegateExpr<FirstTypeDerived>,Symbolic::ValueExpr<> >,
-                                                            Symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
-seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, LastType l)
+typename internal::enable_if<!symbolic::is_symbolic<LastType>::value,
+    ArithmeticSequence<FirstTypeDerived, symbolic::AddExpr<symbolic::AddExpr<symbolic::NegateExpr<FirstTypeDerived>,symbolic::ValueExpr<> >,
+                                                            symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
+seq(const symbolic::BaseExpr<FirstTypeDerived> &f, LastType l)
 {
   return seqN(f.derived(),(typename internal::cleanup_index_type<LastType>::type(l)-f.derived()+fix<1>()));
 }
 
 template<typename FirstType,typename LastTypeDerived>
-typename internal::enable_if<!Symbolic::is_symbolic<FirstType>::value,
+typename internal::enable_if<!symbolic::is_symbolic<FirstType>::value,
     ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,
-                        Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::ValueExpr<> >,
-                                          Symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
-seq(FirstType f, const Symbolic::BaseExpr<LastTypeDerived> &l)
+                        symbolic::AddExpr<symbolic::AddExpr<LastTypeDerived,symbolic::ValueExpr<> >,
+                                          symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
+seq(FirstType f, const symbolic::BaseExpr<LastTypeDerived> &l)
 {
   return seqN(typename internal::cleanup_index_type<FirstType>::type(f),(l.derived()-typename internal::cleanup_index_type<FirstType>::type(f)+fix<1>()));
 }
 
 template<typename FirstTypeDerived,typename LastTypeDerived>
 ArithmeticSequence<FirstTypeDerived,
-                    Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::NegateExpr<FirstTypeDerived> >,Symbolic::ValueExpr<internal::FixedInt<1> > > >
-seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, const Symbolic::BaseExpr<LastTypeDerived> &l)
+                    symbolic::AddExpr<symbolic::AddExpr<LastTypeDerived,symbolic::NegateExpr<FirstTypeDerived> >,symbolic::ValueExpr<internal::FixedInt<1> > > >
+seq(const symbolic::BaseExpr<FirstTypeDerived> &f, const symbolic::BaseExpr<LastTypeDerived> &l)
 {
   return seqN(f.derived(),(l.derived()-f.derived()+fix<1>()));
 }
 
 
 template<typename FirstType,typename LastType, typename IncrType>
-typename internal::enable_if<!(Symbolic::is_symbolic<FirstType>::value || Symbolic::is_symbolic<LastType>::value),
+typename internal::enable_if<!(symbolic::is_symbolic<FirstType>::value || symbolic::is_symbolic<LastType>::value),
     ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,Index,typename internal::cleanup_seq_incr<IncrType>::type> >::type
 seq(FirstType f, LastType l, IncrType incr)
 {
@@ -276,27 +276,27 @@ seq(FirstType f, LastType l, IncrType incr)
 }
 
 template<typename FirstTypeDerived,typename LastType, typename IncrType>
-typename internal::enable_if<!Symbolic::is_symbolic<LastType>::value,
+typename internal::enable_if<!symbolic::is_symbolic<LastType>::value,
     ArithmeticSequence<FirstTypeDerived,
-                        Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<Symbolic::NegateExpr<FirstTypeDerived>,
-                                                                                   Symbolic::ValueExpr<> >,
-                                                                 Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
-                                              Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                        symbolic::QuotientExpr<symbolic::AddExpr<symbolic::AddExpr<symbolic::NegateExpr<FirstTypeDerived>,
+                                                                                   symbolic::ValueExpr<> >,
+                                                                 symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                              symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
                         typename internal::cleanup_seq_incr<IncrType>::type> >::type
-seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, LastType l, IncrType incr)
+seq(const symbolic::BaseExpr<FirstTypeDerived> &f, LastType l, IncrType incr)
 {
   typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
   return seqN(f.derived(),(typename internal::cleanup_index_type<LastType>::type(l)-f.derived()+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
 }
 
 template<typename FirstType,typename LastTypeDerived, typename IncrType>
-typename internal::enable_if<!Symbolic::is_symbolic<FirstType>::value,
+typename internal::enable_if<!symbolic::is_symbolic<FirstType>::value,
     ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,
-                        Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::ValueExpr<> >,
-                                                                 Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
-                                               Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                        symbolic::QuotientExpr<symbolic::AddExpr<symbolic::AddExpr<LastTypeDerived,symbolic::ValueExpr<> >,
+                                                                 symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                               symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
                         typename internal::cleanup_seq_incr<IncrType>::type> >::type
-seq(FirstType f, const Symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
+seq(FirstType f, const symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
 {
   typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
   return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
@@ -305,12 +305,12 @@ seq(FirstType f, const Symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
 
 template<typename FirstTypeDerived,typename LastTypeDerived, typename IncrType>
 ArithmeticSequence<FirstTypeDerived,
-                    Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,
-                                                                               Symbolic::NegateExpr<FirstTypeDerived> >,
-                                                             Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
-                                          Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                    symbolic::QuotientExpr<symbolic::AddExpr<symbolic::AddExpr<LastTypeDerived,
+                                                                               symbolic::NegateExpr<FirstTypeDerived> >,
+                                                             symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                          symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
                     typename internal::cleanup_seq_incr<IncrType>::type>
-seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, const Symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
+seq(const symbolic::BaseExpr<FirstTypeDerived> &f, const symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
 {
   typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
   return seqN(f.derived(),(l.derived()-f.derived()+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
@@ -353,7 +353,7 @@ namespace internal {
 // Convert a symbolic span into a usable one (i.e., remove last/end "keywords")
 template<typename T>
 struct make_size_type {
-  typedef typename internal::conditional<Symbolic::is_symbolic<T>::value, Index, T>::type type;
+  typedef typename internal::conditional<symbolic::is_symbolic<T>::value, Index, T>::type type;
 };
 
 template<typename FirstType,typename SizeType,typename IncrType,int XprSize>
diff --git a/Eigen/src/Core/util/IndexedViewHelper.h b/Eigen/src/Core/util/IndexedViewHelper.h
index dcba731f0..5d7cf8e91 100644
--- a/Eigen/src/Core/util/IndexedViewHelper.h
+++ b/Eigen/src/Core/util/IndexedViewHelper.h
@@ -42,7 +42,7 @@ struct symbolic_last_tag {};
   *
   * \sa end
   */
-static const Symbolic::SymbolExpr<internal::symbolic_last_tag> last;
+static const symbolic::SymbolExpr<internal::symbolic_last_tag> last;
 
 /** \var end
   * \ingroup Core_Module
@@ -60,7 +60,7 @@ static const auto end = last+1;
 #else
 // Using a FixedExpr<1> expression is important here to make sure the compiler
 // can fully optimize the computation starting indices with zero overhead.
-static const Symbolic::AddExpr<Symbolic::SymbolExpr<internal::symbolic_last_tag>,Symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > end(last+fix<1>());
+static const symbolic::AddExpr<symbolic::SymbolExpr<internal::symbolic_last_tag>,symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > end(last+fix<1>());
 #endif
 
 } // end namespace placeholders
@@ -74,7 +74,7 @@ template<int N>
 FixedInt<N> eval_expr_given_size(FixedInt<N> x, Index /*size*/)   { return x; }
 
 template<typename Derived>
-Index eval_expr_given_size(const Symbolic::BaseExpr<Derived> &x, Index size)
+Index eval_expr_given_size(const symbolic::BaseExpr<Derived> &x, Index size)
 {
   return x.derived().eval(placeholders::last=size-1);
 }
@@ -127,13 +127,13 @@ struct IndexedViewCompatibleType<T,XprSize,typename internal::enable_if<internal
 };
 
 template<typename T, int XprSize>
-struct IndexedViewCompatibleType<T, XprSize, typename enable_if<Symbolic::is_symbolic<T>::value>::type> {
+struct IndexedViewCompatibleType<T, XprSize, typename enable_if<symbolic::is_symbolic<T>::value>::type> {
   typedef SingleRange type;
 };
 
 
 template<typename T>
-typename enable_if<Symbolic::is_symbolic<T>::value,SingleRange>::type
+typename enable_if<symbolic::is_symbolic<T>::value,SingleRange>::type
 makeIndexedViewCompatible(const T& id, Index size, SpecializedType) {
   return eval_expr_given_size(id,size);
 }
diff --git a/Eigen/src/Core/util/SymbolicIndex.h b/Eigen/src/Core/util/SymbolicIndex.h
index b9fe733b7..41d477056 100644
--- a/Eigen/src/Core/util/SymbolicIndex.h
+++ b/Eigen/src/Core/util/SymbolicIndex.h
@@ -12,7 +12,7 @@
 
 namespace Eigen {
 
-/** \namespace Eigen::Symbolic
+/** \namespace Eigen::symbolic
   * \ingroup Core_Module
   *
   * This namespace defines a set of classes and functions to build and evaluate symbolic expressions of scalar type Index.
@@ -20,9 +20,9 @@ namespace Eigen {
   *
   * \code
   * // First step, defines symbols:
-  * struct x_tag {};  static const Symbolic::SymbolExpr<x_tag> x;
-  * struct y_tag {};  static const Symbolic::SymbolExpr<y_tag> y;
-  * struct z_tag {};  static const Symbolic::SymbolExpr<z_tag> z;
+  * struct x_tag {};  static const symbolic::SymbolExpr<x_tag> x;
+  * struct y_tag {};  static const symbolic::SymbolExpr<y_tag> y;
+  * struct z_tag {};  static const symbolic::SymbolExpr<z_tag> z;
   *
   * // Defines an expression:
   * auto expr = (x+3)/y+z;
@@ -38,7 +38,7 @@ namespace Eigen {
   * It is currently only used internally to define and manipulate the placeholders::last and placeholders::end symbols in Eigen::seq and Eigen::seqN.
   *
   */
-namespace Symbolic {
+namespace symbolic {
 
 template<typename Tag> class Symbol;
 template<typename Arg0> class NegateExpr;
@@ -286,7 +286,7 @@ protected:
   Arg1 m_arg1;
 };
 
-} // end namespace Symbolic
+} // end namespace symbolic
 
 } // end namespace Eigen
 
diff --git a/Eigen/src/plugins/IndexedViewMethods.h b/Eigen/src/plugins/IndexedViewMethods.h
index d435aa597..5bfb19ac6 100644
--- a/Eigen/src/plugins/IndexedViewMethods.h
+++ b/Eigen/src/plugins/IndexedViewMethods.h
@@ -161,7 +161,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 
 template<typename Indices>
 typename internal::enable_if<
-  (internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_valid_index_type<Indices>::value) && (!Symbolic::is_symbolic<Indices>::value),
+  (internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_valid_index_type<Indices>::value) && (!symbolic::is_symbolic<Indices>::value),
   VectorBlock<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,internal::array_size<Indices>::value> >::type
 operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 {
@@ -172,7 +172,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
 }
 
 template<typename IndexType>
-typename internal::enable_if<Symbolic::is_symbolic<IndexType>::value, CoeffReturnType >::type
+typename internal::enable_if<symbolic::is_symbolic<IndexType>::value, CoeffReturnType >::type
 operator()(const IndexType& id) EIGEN_INDEXED_VIEW_METHOD_CONST
 {
   return Base::operator()(internal::eval_expr_given_size(id,size()));
diff --git a/test/symbolic_index.cpp b/test/symbolic_index.cpp
index cf89b9fbe..02db2eb51 100644
--- a/test/symbolic_index.cpp
+++ b/test/symbolic_index.cpp
@@ -60,12 +60,12 @@ bool is_same_symb(const T1& a, const T2& b, Index size)
 
 template<typename T>
 void check_is_symbolic(const T&) {
-  STATIC_CHECK(( Symbolic::is_symbolic<T>::value ))
+  STATIC_CHECK(( symbolic::is_symbolic<T>::value ))
 }
 
 template<typename T>
 void check_isnot_symbolic(const T&) {
-  STATIC_CHECK(( !Symbolic::is_symbolic<T>::value ))
+  STATIC_CHECK(( !symbolic::is_symbolic<T>::value ))
 }
 
 #define VERIFY_EQ_INT(A,B) VERIFY_IS_APPROX(int(A),int(B))
@@ -104,9 +104,9 @@ void check_symbolic_index()
 
 #if EIGEN_HAS_CXX14
   {
-    struct x_tag {};  static const Symbolic::SymbolExpr<x_tag> x;
-    struct y_tag {};  static const Symbolic::SymbolExpr<y_tag> y;
-    struct z_tag {};  static const Symbolic::SymbolExpr<z_tag> z;
+    struct x_tag {};  static const symbolic::SymbolExpr<x_tag> x;
+    struct y_tag {};  static const symbolic::SymbolExpr<y_tag> y;
+    struct z_tag {};  static const symbolic::SymbolExpr<z_tag> z;
 
     VERIFY_IS_APPROX( int(((x+3)/y+z).eval(x=6,y=3,z=-13)), (6+3)/3+(-13) );
   }

From 2014c7ae28c592063f6887632682231ed862cfc5 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 15 Sep 2018 14:35:10 +0200
Subject: [PATCH 658/680] Move all, last, end from Eigen::placeholders
 namespace to Eigen::, and rename end to lastp1 to avoid conflicts with
 std::end.

---
 Eigen/src/Core/ArithmeticSequence.h     |  8 +++----
 Eigen/src/Core/util/IndexedViewHelper.h | 28 +++++++------------------
 test/indexed_view.cpp                   | 10 +++------
 test/symbolic_index.cpp                 | 20 +++++++-----------
 4 files changed, 23 insertions(+), 43 deletions(-)

diff --git a/Eigen/src/Core/ArithmeticSequence.h b/Eigen/src/Core/ArithmeticSequence.h
index 941028893..42fbf47e8 100644
--- a/Eigen/src/Core/ArithmeticSequence.h
+++ b/Eigen/src/Core/ArithmeticSequence.h
@@ -329,9 +329,9 @@ seq(const symbolic::BaseExpr<FirstTypeDerived> &f, const symbolic::BaseExpr<Last
   * \sa lastN(SizeType), seqN(FirstType,SizeType), seq(FirstType,LastType,IncrType) */
 template<typename SizeType,typename IncrType>
 auto lastN(SizeType size, IncrType incr)
--> decltype(seqN(Eigen::placeholders::last-(size-fix<1>())*incr, size, incr))
+-> decltype(seqN(Eigen::last-(size-fix<1>())*incr, size, incr))
 {
-  return seqN(Eigen::placeholders::last-(size-fix<1>())*incr, size, incr);
+  return seqN(Eigen::last-(size-fix<1>())*incr, size, incr);
 }
 
 /** \cpp11
@@ -342,9 +342,9 @@ auto lastN(SizeType size, IncrType incr)
   * \sa lastN(SizeType,IncrType, seqN(FirstType,SizeType), seq(FirstType,LastType) */
 template<typename SizeType>
 auto lastN(SizeType size)
--> decltype(seqN(Eigen::placeholders::last+fix<1>()-size, size))
+-> decltype(seqN(Eigen::last+fix<1>()-size, size))
 {
-  return seqN(Eigen::placeholders::last+fix<1>()-size, size);
+  return seqN(Eigen::last+fix<1>()-size, size);
 }
 #endif
 
diff --git a/Eigen/src/Core/util/IndexedViewHelper.h b/Eigen/src/Core/util/IndexedViewHelper.h
index 5d7cf8e91..3ec3dd7bd 100644
--- a/Eigen/src/Core/util/IndexedViewHelper.h
+++ b/Eigen/src/Core/util/IndexedViewHelper.h
@@ -13,13 +13,6 @@
 
 namespace Eigen {
 
-/** \namespace Eigen::placeholders
-  * \ingroup Core_Module
-  *
-  * Namespace containing symbolic placeholder and identifiers
-  */
-namespace placeholders {
-
 namespace internal {
 struct symbolic_last_tag {};
 }
@@ -44,27 +37,26 @@ struct symbolic_last_tag {};
   */
 static const symbolic::SymbolExpr<internal::symbolic_last_tag> last;
 
-/** \var end
+/** \var lastp1
   * \ingroup Core_Module
   *
-  * Can be used as a parameter to Eigen::seq and Eigen::seqN functions to symbolically reference the last+1 element/row/columns
-  * of the underlying vector or matrix once passed to DenseBase::operator()(const RowIndices&, const ColIndices&).
+  * Can be used as a parameter to Eigen::seq and Eigen::seqN functions to symbolically
+  * reference the last+1 element/row/columns of the underlying vector or matrix once
+  * passed to DenseBase::operator()(const RowIndices&, const ColIndices&).
   *
   * This symbolic placeholder support standard arithmetic operation.
-  * It is essentially an alias to last+1
+  * It is essentially an alias to last+fix<1>.
   *
   * \sa last
   */
 #ifdef EIGEN_PARSED_BY_DOXYGEN
-static const auto end = last+1;
+static const auto lastp1 = last+fix<1>;
 #else
 // Using a FixedExpr<1> expression is important here to make sure the compiler
 // can fully optimize the computation starting indices with zero overhead.
-static const symbolic::AddExpr<symbolic::SymbolExpr<internal::symbolic_last_tag>,symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > end(last+fix<1>());
+static const symbolic::AddExpr<symbolic::SymbolExpr<internal::symbolic_last_tag>,symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > lastp1(last+fix<1>());
 #endif
 
-} // end namespace placeholders
-
 namespace internal {
 
  // Replace symbolic last/end "keywords" by their true runtime value
@@ -76,7 +68,7 @@ FixedInt<N> eval_expr_given_size(FixedInt<N> x, Index /*size*/)   { return x; }
 template<typename Derived>
 Index eval_expr_given_size(const symbolic::BaseExpr<Derived> &x, Index size)
 {
-  return x.derived().eval(placeholders::last=size-1);
+  return x.derived().eval(last=size-1);
 }
 
 // Extract increment/step at compile time
@@ -172,16 +164,12 @@ template<int Size> struct get_compile_time_incr<AllRange<Size> > {
 } // end namespace internal
 
 
-namespace placeholders {
-
 /** \var all
   * \ingroup Core_Module
   * Can be used as a parameter to DenseBase::operator()(const RowIndices&, const ColIndices&) to index all rows or columns
   */
 static const Eigen::internal::all_t all;
 
-}
-
 } // end namespace Eigen
 
 #endif // EIGEN_INDEXED_VIEW_HELPER_H
diff --git a/test/indexed_view.cpp b/test/indexed_view.cpp
index e6ffe5a74..9a284cf0a 100644
--- a/test/indexed_view.cpp
+++ b/test/indexed_view.cpp
@@ -82,10 +82,6 @@ enum DummyEnum { XX=0, YY=1 };
 
 void check_indexed_view()
 {
-  using Eigen::placeholders::all;
-  using Eigen::placeholders::last;
-  using Eigen::placeholders::end;
-
   Index n = 10;
 
   ArrayXd a = ArrayXd::LinSpaced(n,0,n-1);
@@ -239,7 +235,7 @@ void check_indexed_view()
   VERIFY_IS_APPROX( A(seq(n-1,2,-2), seqN(n-1-6,4)), A(seq(last,2,-2), seqN(last-6,4)) );
   VERIFY_IS_APPROX( A(seq(n-1-6,n-1-2), seqN(n-1-6,4)), A(seq(last-6,last-2), seqN(6+last-6-6,4)) );
   VERIFY_IS_APPROX( A(seq((n-1)/2,(n)/2+3), seqN(2,4)), A(seq(last/2,(last+1)/2+3), seqN(last+2-last,4)) );
-  VERIFY_IS_APPROX( A(seq(n-2,2,-2), seqN(n-8,4)), A(seq(end-2,2,-2), seqN(end-8,4)) );
+  VERIFY_IS_APPROX( A(seq(n-2,2,-2), seqN(n-8,4)), A(seq(lastp1-2,2,-2), seqN(lastp1-8,4)) );
 
   // Check all combinations of seq:
   VERIFY_IS_APPROX( A(seq(1,n-1-2,2), seq(1,n-1-2,2)), A(seq(1,last-2,2), seq(1,last-2,fix<2>)) );
@@ -249,7 +245,7 @@ void check_indexed_view()
   VERIFY_IS_APPROX( A(seq(n-1-5,n-1-2), seq(n-1-5,n-1-2)), A(seq(last-5,last-2), seq(last-5,last-2)) );
 
   VERIFY_IS_APPROX( A.col(A.cols()-1), A(all,last) );
-  VERIFY_IS_APPROX( A(A.rows()-2, A.cols()/2), A(last-1, end/2) );
+  VERIFY_IS_APPROX( A(A.rows()-2, A.cols()/2), A(last-1, lastp1/2) );
   VERIFY_IS_APPROX( a(a.size()-2), a(last-1) );
   VERIFY_IS_APPROX( a(a.size()/2), a((last+1)/2) );
 
@@ -272,7 +268,7 @@ void check_indexed_view()
 
     VERIFY( is_same_eq(a.head(4), a(seq(0,3))) );
     VERIFY( is_same_eq(a.tail(4), a(seqN(last-3,4))) );
-    VERIFY( is_same_eq(a.tail(4), a(seq(end-4,last))) );
+    VERIFY( is_same_eq(a.tail(4), a(seq(lastp1-4,last))) );
     VERIFY( is_same_eq(a.segment<4>(3), a(seqN(3,fix<4>))) );
   }
 
diff --git a/test/symbolic_index.cpp b/test/symbolic_index.cpp
index 02db2eb51..bccf1b884 100644
--- a/test/symbolic_index.cpp
+++ b/test/symbolic_index.cpp
@@ -54,7 +54,6 @@ is_same_type(const T1&, const T2&)
 template<typename T1,typename T2>
 bool is_same_symb(const T1& a, const T2& b, Index size)
 {
-  using Eigen::placeholders::last;
   return a.eval(last=size-1) == b.eval(last=size-1);
 }
 
@@ -72,14 +71,11 @@ void check_isnot_symbolic(const T&) {
 
 void check_symbolic_index()
 {
-  using Eigen::placeholders::last;
-  using Eigen::placeholders::end;
-
   check_is_symbolic(last);
-  check_is_symbolic(end);
+  check_is_symbolic(lastp1);
   check_is_symbolic(last+1);
-  check_is_symbolic(last-end);
-  check_is_symbolic(2*last-end/2);
+  check_is_symbolic(last-lastp1);
+  check_is_symbolic(2*last-lastp1/2);
   check_isnot_symbolic(fix<3>());
 
   Index size=100;
@@ -93,14 +89,14 @@ void check_symbolic_index()
   VERIFY( is_same_type( fix<9>()|fix<2>(), fix<9|2>() ) );
   VERIFY( is_same_type( fix<9>()/2, int(9/2) ) );
 
-  VERIFY( is_same_symb( end-1, last, size) );
-  VERIFY( is_same_symb( end-fix<1>, last, size) );
+  VERIFY( is_same_symb( lastp1-1, last, size) );
+  VERIFY( is_same_symb( lastp1-fix<1>, last, size) );
 
   VERIFY_IS_EQUAL( ( (last*5-2)/3 ).eval(last=size-1), ((size-1)*5-2)/3 );
   VERIFY_IS_EQUAL( ( (last*fix<5>-fix<2>)/fix<3> ).eval(last=size-1), ((size-1)*5-2)/3 );
-  VERIFY_IS_EQUAL( ( -last*end  ).eval(last=size-1), -(size-1)*size );
-  VERIFY_IS_EQUAL( ( end-3*last  ).eval(last=size-1), size- 3*(size-1) );
-  VERIFY_IS_EQUAL( ( (end-3*last)/end  ).eval(last=size-1), (size- 3*(size-1))/size );
+  VERIFY_IS_EQUAL( ( -last*lastp1  ).eval(last=size-1), -(size-1)*size );
+  VERIFY_IS_EQUAL( ( lastp1-3*last  ).eval(last=size-1), size- 3*(size-1) );
+  VERIFY_IS_EQUAL( ( (lastp1-3*last)/lastp1  ).eval(last=size-1), (size- 3*(size-1))/size );
 
 #if EIGEN_HAS_CXX14
   {

From 297ca623199017bcc548d72dca0570c02edbeb6d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Mon, 17 Sep 2018 16:24:52 +0200
Subject: [PATCH 659/680] ease transition by adding placeholders::all/last/and
 as deprecated

---
 Eigen/src/Core/util/IndexedViewHelper.h | 17 ++++++++++++++---
 1 file changed, 14 insertions(+), 3 deletions(-)

diff --git a/Eigen/src/Core/util/IndexedViewHelper.h b/Eigen/src/Core/util/IndexedViewHelper.h
index 3ec3dd7bd..40e16fdb4 100644
--- a/Eigen/src/Core/util/IndexedViewHelper.h
+++ b/Eigen/src/Core/util/IndexedViewHelper.h
@@ -28,14 +28,14 @@ struct symbolic_last_tag {};
   * A typical usage example would be:
   * \code
   * using namespace Eigen;
-  * using Eigen::placeholders::last;
+  * using Eigen::last;
   * VectorXd v(n);
   * v(seq(2,last-2)).setOnes();
   * \endcode
   *
   * \sa end
   */
-static const symbolic::SymbolExpr<internal::symbolic_last_tag> last;
+static const symbolic::SymbolExpr<internal::symbolic_last_tag> last; // PLEASE use Eigen::last   instead of Eigen::placeholders::last
 
 /** \var lastp1
   * \ingroup Core_Module
@@ -168,7 +168,18 @@ template<int Size> struct get_compile_time_incr<AllRange<Size> > {
   * \ingroup Core_Module
   * Can be used as a parameter to DenseBase::operator()(const RowIndices&, const ColIndices&) to index all rows or columns
   */
-static const Eigen::internal::all_t all;
+static const Eigen::internal::all_t all; // PLEASE use Eigen::all instead of Eigen::placeholders::all
+
+
+namespace placeholders {
+  typedef symbolic::SymbolExpr<internal::symbolic_last_tag> last_t;
+  typedef symbolic::AddExpr<symbolic::SymbolExpr<internal::symbolic_last_tag>,symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > end_t;
+  typedef Eigen::internal::all_t all_t;
+
+  EIGEN_DEPRECATED static const all_t  all  = Eigen::all;    // PLEASE use Eigen::all    instead of Eigen::placeholders::all
+  EIGEN_DEPRECATED static const last_t last = Eigen::last;   // PLEASE use Eigen::last   instead of Eigen::placeholders::last
+  EIGEN_DEPRECATED static const end_t  end  = Eigen::lastp1; // PLEASE use Eigen::lastp1 instead of Eigen::placeholders::end
+}
 
 } // end namespace Eigen
 

From febd09dcc02c1429cd4abd3ddb3ed5108fcd8339 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 19 Sep 2018 11:11:04 -0700
Subject: [PATCH 660/680] Silence compiler warnings in ThreadPoolInterface.h.

---
 unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
index d1fa4b23e..25030dc0b 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadPoolInterface.h
@@ -21,8 +21,8 @@ class ThreadPoolInterface {
 
   // Submits a closure to be run by threads in the range [start, end) in the
   // pool.
-  virtual void ScheduleWithHint(std::function<void()> fn, int start,
-                                int end) {
+  virtual void ScheduleWithHint(std::function<void()> fn, int /*start*/,
+                                int /*end*/) {
     // Just defer to Schedule in case sub-classes aren't interested in
     // overriding this functionality.
     Schedule(fn);

From db9c9df59aaca890e64c99e63d0147365cc1c8f2 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 19 Sep 2018 11:50:27 -0700
Subject: [PATCH 661/680] Silence more compiler warnings.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h     |  5 ++---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 10 +++++-----
 2 files changed, 7 insertions(+), 8 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index ee70d1d76..aa2e5bcc6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -188,7 +188,7 @@ class TensorBlockIO {
     StorageIndex input_span;
     StorageIndex output_span;
     StorageIndex size;
-    StorageIndex count;
+    StorageIndex count = 0;
   };
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(
@@ -257,7 +257,7 @@ class TensorBlockIO {
     array<BlockIteratorState, at_least_1_dim> block_iter_state;
 
     // Initialize block iterator state. Squeeze away any dimension of size 1.
-    int num_squeezed_dims = 0;
+    Index num_squeezed_dims = 0;
     for (Index i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
       const Index dim = cond<Layout>()(i + 1, NumDims - i - 2);
       const StorageIndex size = block.block_sizes()[tensor_to_block_dim_map[dim]];
@@ -280,7 +280,6 @@ class TensorBlockIO {
       block_iter_state[num_squeezed_dims].output_span =
           block_iter_state[num_squeezed_dims].output_stride *
           (block_iter_state[num_squeezed_dims].size - 1);
-      block_iter_state[num_squeezed_dims].count = 0;
       ++num_squeezed_dims;
     }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 4263f11a6..949764f3a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -884,11 +884,11 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     const Index first_preserved_dim_input_index =
         m_output_to_input_dim_map[first_preserved_dim_output_index];
     const bool inner_most_dim_preserved =
-        first_preserved_dim_input_index ==
-            (static_cast<int>(Layout) == static_cast<int>(ColMajor)
-                 ? 0
-                 : NumInputDims - 1) |
-        PreservingInnerMostDims;
+        PreservingInnerMostDims ||
+        (first_preserved_dim_input_index ==
+         (static_cast<int>(Layout) == static_cast<int>(ColMajor)
+              ? 0
+              : NumInputDims - 1));
 
     // Calculate output block inner/outer dimension sizes.
     const Index output_block_inner_dim_size =

From d638b62dda7c1b15eacda844a8e39791e073b459 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 19 Sep 2018 13:27:55 -0700
Subject: [PATCH 662/680] Silence compiler warning.

---
 unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 2130625bc..9687734ff 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -85,8 +85,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   }
 
   void SetStealPartitions(const std::vector<std::pair<unsigned, unsigned>>& partitions) {
-    int num_partitions = partitions.size();
-    eigen_assert(num_partitions == num_threads_);
+    eigen_assert(static_cast<int>(partitions.size()) == num_threads_);
 
     // Pass this information to each thread queue.
     for (int i = 0; i < num_threads_; i++) {

From 44d8274383d379172bedc7eed4cbc5c6ce4f88a2 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Wed, 19 Sep 2018 13:31:42 -0700
Subject: [PATCH 663/680] Cast to longer type.

---
 unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
index 9687734ff..9bd6a175d 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -85,7 +85,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
   }
 
   void SetStealPartitions(const std::vector<std::pair<unsigned, unsigned>>& partitions) {
-    eigen_assert(static_cast<int>(partitions.size()) == num_threads_);
+    eigen_assert(partitions.size() == static_cast<std::size_t>(num_threads_));
 
     // Pass this information to each thread queue.
     for (int i = 0; i < num_threads_; i++) {

From 2cf6d3050c42d14819160e2642a5c3dfc449e44a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 11:38:19 +0200
Subject: [PATCH 664/680] Disable ignoring attributes warning

---
 test/vectorization_logic.cpp | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp
index 01b55e192..c15f75103 100644
--- a/test/vectorization_logic.cpp
+++ b/test/vectorization_logic.cpp
@@ -22,6 +22,14 @@
 #include "main.h"
 #include <typeinfo>
 
+// Disable "ignoring attributes on template argument"
+// for packet_traits<Packet*>
+// => The only workaround would be to wrap _m128 and the likes
+//    within wrappers.
+#if EIGEN_GNUC_AT_LEAST(6,0)
+    #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
 using internal::demangle_flags;
 using internal::demangle_traversal;
 using internal::demangle_unrolling;

From 5a30eed17e170af4aedca1a3ff0c10a8e65bf47e Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 16:58:51 +0200
Subject: [PATCH 665/680] Fix warnings in AVX512

---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 9e66575a9..9fbb256a1 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -551,7 +551,7 @@ EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet16i& from) {
 template <>
 EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
                                                              Index stride) {
-  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_vector = _mm512_set1_epi32(convert_index<int>(stride));
   Packet16i stride_multiplier =
       _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
   Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
@@ -561,7 +561,7 @@ EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
 template <>
 EIGEN_DEVICE_FUNC inline Packet8d pgather<double, Packet8d>(const double* from,
                                                             Index stride) {
-  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_vector = _mm256_set1_epi32(convert_index<int>(stride));
   Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
   Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
 
@@ -572,7 +572,7 @@ template <>
 EIGEN_DEVICE_FUNC inline void pscatter<float, Packet16f>(float* to,
                                                          const Packet16f& from,
                                                          Index stride) {
-  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_vector = _mm512_set1_epi32(convert_index<int>(stride));
   Packet16i stride_multiplier =
       _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
   Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
@@ -582,7 +582,7 @@ template <>
 EIGEN_DEVICE_FUNC inline void pscatter<double, Packet8d>(double* to,
                                                          const Packet8d& from,
                                                          Index stride) {
-  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_vector = _mm256_set1_epi32(convert_index<int>(stride));
   Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
   Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
   _mm512_i32scatter_pd(to, indices, from, 8);
@@ -660,8 +660,7 @@ EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
 
 #ifdef EIGEN_VECTORIZE_AVX512DQ
 #define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \
-  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTA, 0);        \
-  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTB, 1);
+  OUTPUT = _mm512_insertf32x8(_mm512_castps256_ps512(INPUTA), INPUTB, 1);
 #else
 #define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB)                    \
   OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 0), 0); \
@@ -855,7 +854,7 @@ template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs)
 
   final_1 = _mm256_add_pd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC));
 
-  __m512d final_output = _mm512_insertf64x4(final_output, final_0, 0);
+  __m512d final_output = _mm512_castpd256_pd512(final_0);
 
   return _mm512_insertf64x4(final_output, final_1, 1);
 }

From 71496b0e25158fe6c47dd8c959748b74be4ca94c Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 17:01:24 +0200
Subject: [PATCH 666/680] Fix gebp kernel for real+complex in case only reals
 are vectorized (e.g., AVX512). This commit also removes "half-packet" from
 data-mappers: it was not used and conceptually broken anyways.

---
 .../Core/products/GeneralBlockPanelKernel.h   | 106 +++++++++---------
 Eigen/src/Core/products/GeneralMatrixMatrix.h |   2 +-
 .../products/GeneralMatrixMatrixTriangular.h  |   6 +-
 .../Core/products/SelfadjointMatrixMatrix.h   |   6 +-
 .../Core/products/TriangularMatrixMatrix.h    |   6 +-
 .../Core/products/TriangularSolverMatrix.h    |   4 +-
 Eigen/src/Core/util/BlasUtil.h                |  45 +++-----
 .../CXX11/src/Tensor/TensorContraction.h      |   2 +-
 .../src/Tensor/TensorContractionMapper.h      |  75 ++++---------
 9 files changed, 106 insertions(+), 146 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 3ec8eb082..fa844815d 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -1025,9 +1025,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r0.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C8, alphav, R2);
@@ -1035,9 +1035,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r0.storePacket(1 * Traits::ResPacketSize, R1);
           r0.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r1.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r1.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r1.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r1.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r1.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C1, alphav, R0);
           traits.acc(C5, alphav, R1);
           traits.acc(C9, alphav, R2);
@@ -1045,9 +1045,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r1.storePacket(1 * Traits::ResPacketSize, R1);
           r1.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r2.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r2.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r2.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C2, alphav, R0);
           traits.acc(C6, alphav, R1);
           traits.acc(C10, alphav, R2);
@@ -1055,9 +1055,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r2.storePacket(1 * Traits::ResPacketSize, R1);
           r2.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r3.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r3.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r3.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r3.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r3.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C3, alphav, R0);
           traits.acc(C7, alphav, R1);
           traits.acc(C11, alphav, R2);
@@ -1134,9 +1134,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r0.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C8, alphav, R2);
@@ -1244,10 +1244,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2, R3;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r1.loadPacket(0 * Traits::ResPacketSize);
-          R3 = r1.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R3 = r1.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C1, alphav, R2);
@@ -1257,10 +1257,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r1.storePacket(0 * Traits::ResPacketSize, R2);
           r1.storePacket(1 * Traits::ResPacketSize, R3);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r3.loadPacket(0 * Traits::ResPacketSize);
-          R3 = r3.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r2.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R3 = r3.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C2,  alphav, R0);
           traits.acc(C6,  alphav, R1);
           traits.acc(C3,  alphav, R2);
@@ -1337,8 +1337,8 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
@@ -1431,15 +1431,15 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C1,  alphav, R1);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
           r1.storePacket(0 * Traits::ResPacketSize, R1);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C2,  alphav, R0);
           traits.acc(C3,  alphav, R1);
           r2.storePacket(0 * Traits::ResPacketSize, R0);
@@ -1504,7 +1504,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
 #undef EIGEN_GEBGP_ONESTEP
           ResPacket R0;
           ResPacket alphav = pset1<ResPacket>(alpha);
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
         }
@@ -1685,19 +1685,18 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
 //
 //  32 33 34 35 ...
 //  36 36 38 39 ...
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, ColMajor, Conjugate, PanelMode>
 {
   typedef typename DataMapper::LinearMapper LinearMapper;
   EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, ColMajor, Conjugate, PanelMode>
   ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
 {
-  typedef typename packet_traits<Scalar>::type Packet;
-  enum { PacketSize = packet_traits<Scalar>::size };
+  enum { PacketSize = unpacket_traits<Packet>::size };
 
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
   EIGEN_UNUSED_VARIABLE(stride);
@@ -1725,9 +1724,9 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A, B, C;
-        A = lhs.loadPacket(i+0*PacketSize, k);
-        B = lhs.loadPacket(i+1*PacketSize, k);
-        C = lhs.loadPacket(i+2*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
+        B = lhs.template loadPacket<Packet>(i+1*PacketSize, k);
+        C = lhs.template loadPacket<Packet>(i+2*PacketSize, k);
         pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(C)); count+=PacketSize;
@@ -1745,8 +1744,8 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A, B;
-        A = lhs.loadPacket(i+0*PacketSize, k);
-        B = lhs.loadPacket(i+1*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
+        B = lhs.template loadPacket<Packet>(i+1*PacketSize, k);
         pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
       }
@@ -1763,7 +1762,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A;
-        A = lhs.loadPacket(i+0*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
         pstore(blockA+count, cj.pconj(A));
         count+=PacketSize;
       }
@@ -1793,19 +1792,18 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
   }
 }
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, RowMajor, Conjugate, PanelMode>
 {
   typedef typename DataMapper::LinearMapper LinearMapper;
   EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, RowMajor, Conjugate, PanelMode>
   ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
 {
-  typedef typename packet_traits<Scalar>::type Packet;
-  enum { PacketSize = packet_traits<Scalar>::size };
+  enum { PacketSize = unpacket_traits<Packet>::size };
 
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
   EIGEN_UNUSED_VARIABLE(stride);
@@ -1837,7 +1835,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Ro
           for (Index m = 0; m < pack; m += PacketSize)
           {
             PacketBlock<Packet> kernel;
-            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k);
+            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.template loadPacket<Packet>(i+p+m, k);
             ptranspose(kernel);
             for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p]));
           }
@@ -1971,10 +1969,10 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Co
       {
         for(; k<peeled_k; k+=PacketSize) {
           PacketBlock<Packet,(PacketSize%4)==0?4:PacketSize> kernel;
-          kernel.packet[0] = dm0.loadPacket(k);
-          kernel.packet[1%PacketSize] = dm1.loadPacket(k);
-          kernel.packet[2%PacketSize] = dm2.loadPacket(k);
-          kernel.packet[3%PacketSize] = dm3.loadPacket(k);
+          kernel.packet[0           ] = dm0.template loadPacket<Packet>(k);
+          kernel.packet[1%PacketSize] = dm1.template loadPacket<Packet>(k);
+          kernel.packet[2%PacketSize] = dm2.template loadPacket<Packet>(k);
+          kernel.packet[3%PacketSize] = dm3.template loadPacket<Packet>(k);
           ptranspose(kernel);
           pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0]));
           pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1%PacketSize]));
@@ -2075,7 +2073,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Co
       for(Index k=0; k<depth; k++)
       {
         if (PacketSize==4) {
-          Packet A = rhs.loadPacket(k, j2);
+          Packet A = rhs.template loadPacket<Packet>(k, j2);
           pstoreu(blockB+count, cj.pconj(A));
           count += PacketSize;
         } else {
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index bd7b6ff2a..a010e150f 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -75,7 +75,7 @@ static void run(Index rows, Index cols, Index depth,
   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index e436c50a4..27a7921fa 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -84,7 +84,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
 
-    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
     tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
@@ -110,7 +110,6 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
           gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc,
                (std::min)(size,i2), alpha, -1, -1, 0, 0);
 
-
         sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
 
         if (UpLo==Upper)
@@ -160,7 +159,7 @@ struct tribb_kernel
       if(UpLo==Upper)
         gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
                     -1, -1, 0, 0);
-
+      
       // selfadjoint micro block
       {
         Index i = j;
@@ -168,6 +167,7 @@ struct tribb_kernel
         // 1 - apply the kernel on the temporary buffer
         gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
                     -1, -1, 0, 0);
+
         // 2 - triangular accumulation
         for(Index j1=0; j1<actualBlockSize; ++j1)
         {
diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
index da6f82abc..c43a53b92 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
@@ -352,7 +352,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
     symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
 
     for(Index k2=0; k2<size; k2+=kc)
     {
@@ -387,7 +387,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
       for(Index i2=k2+kc; i2<size; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,size)-i2;
-        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder,false>()
           (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
         gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
@@ -437,7 +437,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
     symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=0; k2<size; k2+=kc)
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index f784507e7..7fe476be2 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -151,7 +151,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=IsLower ? depth : 0;
@@ -222,7 +222,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         for(Index i2=start; i2<end; i2+=mc)
         {
           const Index actual_mc = (std::min)(i2+mc,end)-i2;
-          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder,false>()
             (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
           gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
@@ -299,7 +299,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
 
diff --git a/Eigen/src/Core/products/TriangularSolverMatrix.h b/Eigen/src/Core/products/TriangularSolverMatrix.h
index 223c38b86..96d1b38a9 100644
--- a/Eigen/src/Core/products/TriangularSolverMatrix.h
+++ b/Eigen/src/Core/products/TriangularSolverMatrix.h
@@ -76,7 +76,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
     conj_if<Conjugate> conj;
     gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, TriStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
 
     // the goal here is to subdivise the Rhs panels such that we keep some cache
@@ -229,7 +229,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, ColMajor, false, true> pack_lhs_panel;
 
     for(Index k2=IsLower ? size : 0;
         IsLower ? k2>0 : k2<size;
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index 705925984..a32630ed7 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -24,7 +24,7 @@ struct gebp_kernel;
 template<typename Scalar, typename Index, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
 struct gemm_pack_rhs;
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
 struct gemm_pack_lhs;
 
 template<
@@ -156,11 +156,9 @@ class BlasVectorMapper {
 };
 
 template<typename Scalar, typename Index, int AlignmentType>
-class BlasLinearMapper {
-  public:
-  typedef typename packet_traits<Scalar>::type Packet;
-  typedef typename packet_traits<Scalar>::half HalfPacket;
-
+class BlasLinearMapper
+{
+public:
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
@@ -171,29 +169,25 @@ class BlasLinearMapper {
     return m_data[i];
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
-    return ploadt<Packet, AlignmentType>(m_data + i);
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketType loadPacket(Index i) const {
+    return ploadt<PacketType, AlignmentType>(m_data + i);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
-    return ploadt<HalfPacket, AlignmentType>(m_data + i);
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const PacketType &p) const {
+    pstoret<Scalar, PacketType, AlignmentType>(m_data + i, p);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet &p) const {
-    pstoret<Scalar, Packet, AlignmentType>(m_data + i, p);
-  }
-
-  protected:
+protected:
   Scalar *m_data;
 };
 
 // Lightweight helper class to access matrix coefficients.
 template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned>
-class blas_data_mapper {
-  public:
-  typedef typename packet_traits<Scalar>::type Packet;
-  typedef typename packet_traits<Scalar>::half HalfPacket;
-
+class blas_data_mapper
+{
+public:
   typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
   typedef BlasVectorMapper<Scalar, Index> VectorMapper;
 
@@ -218,8 +212,9 @@ class blas_data_mapper {
     return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride];
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
-    return ploadt<Packet, AlignmentType>(&operator()(i, j));
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketType loadPacket(Index i, Index j) const {
+    return ploadt<PacketType, AlignmentType>(&operator()(i, j));
   }
 
   template <typename PacketT, int AlignmentT>
@@ -227,10 +222,6 @@ class blas_data_mapper {
     return ploadt<PacketT, AlignmentT>(&operator()(i, j));
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
-    return ploadt<HalfPacket, AlignmentType>(&operator()(i, j));
-  }
-
   template<typename SubPacket>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const {
     pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride);
@@ -251,7 +242,7 @@ class blas_data_mapper {
     return internal::first_default_aligned(m_data, size);
   }
 
-  protected:
+protected:
   Scalar* EIGEN_RESTRICT m_data;
   const Index m_stride;
 };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index f0f61fade..9614aeefe 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -620,7 +620,7 @@ struct TensorContractionEvaluatorBase
     typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
 
     // Declare GEBP packing and kernel structs
-    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, ColMajor> pack_lhs;
+    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, typename Traits::LhsPacket, ColMajor> pack_lhs;
     internal::gemm_pack_rhs<RhsScalar, Index, typename RhsMapper::SubMapper, nr, ColMajor> pack_rhs;
 
     internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper, mr, nr, false, false> gebp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
index ab320a50d..dbb0f76bb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
@@ -238,9 +238,6 @@ class BaseTensorContractionMapper : public SimpleTensorContractionMapper<Scalar,
                               const contract_t& k_strides) :
   ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
 
-  typedef typename Tensor::PacketReturnType Packet;
-  typedef typename unpacket_traits<Packet>::half HalfPacket;
-
   template <typename PacketT,int AlignmentType>
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE PacketT load(Index i, Index j) const {
@@ -284,27 +281,10 @@ class BaseTensorContractionMapper : public SimpleTensorContractionMapper<Scalar,
     return pload<PacketT>(data);
   }
 
-  template <int AlignmentType>
+  template <typename PacketT,int AlignmentType>
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const {
-    return this->load<Packet,AlignmentType>(i,j);
-  }
-
-  template <int AlignmentType>
-  EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
-    // whole method makes column major assumption
-
-    // don't need to add offsets for now (because operator handles that)
-    const Index half_packet_size = unpacket_traits<HalfPacket>::size;
-    if (half_packet_size == packet_size) {
-      return loadPacket<AlignmentType>(i, j);
-    }
-    EIGEN_ALIGN_MAX Scalar data[half_packet_size];
-    for (Index k = 0; k < half_packet_size; k++) {
-      data[k] = operator()(i + k, j);
-    }
-    return pload<HalfPacket>(data);
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
+    return this->load<PacketT,AlignmentType>(i,j);
   }
 };
 
@@ -314,7 +294,8 @@ template<typename Scalar, typename Index, int side,
          typename nocontract_t, typename contract_t,
          bool inner_dim_contiguous,
          bool inner_dim_reordered, int Alignment, template <class> class MakePointer_>
-class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_>
+class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_>
+  : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_>
 {
  public:
   typedef SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_> ParentMapper;
@@ -327,12 +308,11 @@ class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, con
                               const contract_t& k_strides) :
   ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
 
-  typedef typename Tensor::PacketReturnType Packet;
-  template <int> EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const {
+  template <typename PacketT,int> EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
     EIGEN_ALIGN_MAX Scalar data[1];
     data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
-    return pload<typename Tensor::PacketReturnType>(data);
+    return pload<PacketT>(data);
   }
   template <typename PacketT,int> EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE PacketT load(Index i, Index j) const {
@@ -340,10 +320,6 @@ class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, con
     data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
     return pload<PacketT>(data);
   }
-  template <int> EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadHalfPacket(Index i, Index j) const {
-    return loadPacket(i, j);
-  }
 };
 
 
@@ -354,8 +330,6 @@ template<typename Scalar, typename Index, int side,
          bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment, template <class> class MakePointer_=MakePointer>
 class TensorContractionSubMapper {
  public:
-  typedef typename Tensor::PacketReturnType Packet;
-  typedef typename unpacket_traits<Packet>::half HalfPacket;
 
   typedef BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> ParentMapper;
   typedef TensorContractionSubMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> Self;
@@ -390,17 +364,20 @@ class TensorContractionSubMapper {
     return m_base_mapper(i + m_vert_offset, j + m_horiz_offset);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i) const {
     if (UseDirectOffsets) {
-      return m_base_mapper.template loadPacket<Alignment>(i, 0);
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, 0);
     }
-    return m_base_mapper.template loadPacket<Alignment>(i + m_vert_offset, m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, m_horiz_offset);
   }
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i, Index j) const {
     if (UseDirectOffsets) {
-      return m_base_mapper.template loadPacket<Alignment>(i, j);
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, j);
     }
-    return m_base_mapper.template loadPacket<Alignment>(i + m_vert_offset, j + m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, j + m_horiz_offset);
   }
 
   template <typename PacketT, int AlignmentType>
@@ -411,14 +388,8 @@ class TensorContractionSubMapper {
     return m_base_mapper.template loadPacket<PacketT,AlignmentType>(i + m_vert_offset, j + m_horiz_offset);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
-    if (UseDirectOffsets) {
-      return m_base_mapper.template loadHalfPacket<Alignment>(i, 0);
-    }
-    return m_base_mapper.template loadHalfPacket<Alignment>(i + m_vert_offset, m_horiz_offset);
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet& p) const {
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const PacketT& p) const {
     if (UseDirectOffsets) {
       m_base_mapper.storePacket(i, 0, p);
     }
@@ -434,15 +405,15 @@ class TensorContractionSubMapper {
 
   template <typename PacketT, int AlignmentType>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT load(Index i) const {
-    EIGEN_STATIC_ASSERT((internal::is_same<PacketT, Packet>::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((internal::is_same<PacketT, PacketT>::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
     const int ActualAlignment = (AlignmentType == Aligned) && (Alignment == Aligned) ? Aligned : Unaligned;
     if (UseDirectOffsets) {
-     return m_base_mapper.template loadPacket<ActualAlignment>(i, 0);
+     return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i, 0);
     }
-    return m_base_mapper.template loadPacket<ActualAlignment>(i + m_vert_offset, m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i + m_vert_offset, m_horiz_offset);
   }
 
-  template <typename Packet>
+  template <typename PacketT>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool aligned(Index) const {
     return false;
   }

From c50250cb241abb3ea90bac86bc1c27dfadd0862c Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 20 Sep 2018 17:03:42 +0200
Subject: [PATCH 667/680] Avoid warning "suggest braces around initialization
 of subobject". This test is not run in C++03 mode, so no compatibility is
 lost.

---
 unsupported/test/cxx11_tensor_shuffling.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp
index 062dd1c0f..2ec85d2d4 100644
--- a/unsupported/test/cxx11_tensor_shuffling.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling.cpp
@@ -83,10 +83,10 @@ static void test_expr_shuffling()
 
   Tensor<float, 4, DataLayout> result(5, 7, 3, 2);
 
-  array<ptrdiff_t, 4> src_slice_dim({2, 3, 1, 7});
-  array<ptrdiff_t, 4> src_slice_start({0, 0, 0, 0});
-  array<ptrdiff_t, 4> dst_slice_dim({1, 7, 3, 2});
-  array<ptrdiff_t, 4> dst_slice_start({0, 0, 0, 0});
+  array<ptrdiff_t, 4> src_slice_dim{{2, 3, 1, 7}};
+  array<ptrdiff_t, 4> src_slice_start{{0, 0, 0, 0}};
+  array<ptrdiff_t, 4> dst_slice_dim{{1, 7, 3, 2}};
+  array<ptrdiff_t, 4> dst_slice_start{{0, 0, 0, 0}};
 
   for (int i = 0; i < 5; ++i) {
     result.slice(dst_slice_start, dst_slice_dim) =

From e38d1ab4d12720751644ed89ff77c4465be70db8 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 17:07:33 +0200
Subject: [PATCH 668/680] Workaround increases required alignment warning

---
 Eigen/src/Core/arch/GPU/PacketMathHalf.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index c4feda87d..8787adcde 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -542,11 +542,15 @@ template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* fr
 }
 
 template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_store_si256((__m256i*)to, from.x);
+  // (void*) -> workaround clang warning:
+  // cast from 'Eigen::half *' to '__m256i *' increases required alignment from 2 to 32
+  _mm256_store_si256((__m256i*)(void*)to, from.x);
 }
 
 template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_storeu_si256((__m256i*)to, from.x);
+  // (void*) -> workaround clang warning:
+  // cast from 'Eigen::half *' to '__m256i *' increases required alignment from 2 to 32
+  _mm256_storeu_si256((__m256i*)(void*)to, from.x);
 }
 
 template<> EIGEN_STRONG_INLINE Packet16h

From a0166ab6514d47bad8db2502c460953af811ea38 Mon Sep 17 00:00:00 2001
From: Christoph Hertzberg <chtz@informatik.uni-bremen.de>
Date: Thu, 20 Sep 2018 17:08:43 +0200
Subject: [PATCH 669/680] Workaround for spurious "array subscript is above
 array bounds" warnings with g++4.x

---
 unsupported/Eigen/CXX11/src/util/EmulateArray.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index d5c000e08..39c255791 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -21,9 +21,9 @@ namespace Eigen {
 template <typename T, size_t n> class array {
  public:
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE T& operator[] (size_t index) { return values[index]; }
+  EIGEN_STRONG_INLINE T& operator[] (size_t index) { eigen_internal_assert(index < size()); return values[index]; }
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { return values[index]; }
+  EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { eigen_internal_assert(index < size()); return values[index]; }
 
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE T& at(size_t index) { eigen_assert(index < size()); return values[index]; }

From 9419f506d0a01a279454f41906ab27c3fe36fe80 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 17:32:34 +0200
Subject: [PATCH 670/680] Fix regression introduced by the previous fix for
 AVX512. It brokes the complex-complex case on SSE.

---
 Eigen/src/Core/products/GeneralBlockPanelKernel.h           | 4 ++++
 Eigen/src/Core/products/GeneralMatrixMatrix.h               | 2 +-
 Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h     | 2 +-
 Eigen/src/Core/products/SelfadjointMatrixMatrix.h           | 6 +++---
 Eigen/src/Core/products/TriangularMatrixMatrix.h            | 6 +++---
 Eigen/src/Core/products/TriangularSolverMatrix.h            | 4 ++--
 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h      | 2 +-
 .../Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h    | 2 +-
 8 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index fa844815d..b012691c1 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -390,6 +390,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
   
@@ -496,6 +497,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
 
@@ -626,6 +628,7 @@ public:
   typedef typename packet_traits<Scalar>::type     ScalarPacket;
   typedef DoublePacket<RealPacket> DoublePacketType;
 
+  typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type LhsPacket4Packing;
   typedef typename conditional<Vectorizable,RealPacket,  Scalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type ResPacket;
@@ -777,6 +780,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
 
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index a010e150f..f49abcad5 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -75,7 +75,7 @@ static void run(Index rows, Index cols, Index depth,
   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index 27a7921fa..ec2825bf0 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -84,7 +84,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
 
-    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
     tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
index c43a53b92..c84c71609 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
@@ -352,7 +352,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
     symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
 
     for(Index k2=0; k2<size; k2+=kc)
     {
@@ -387,7 +387,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
       for(Index i2=k2+kc; i2<size; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,size)-i2;
-        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder,false>()
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder,false>()
           (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
         gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
@@ -437,7 +437,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=0; k2<size; k2+=kc)
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 7fe476be2..85fd744b9 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -151,7 +151,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=IsLower ? depth : 0;
@@ -222,7 +222,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         for(Index i2=start; i2<end; i2+=mc)
         {
           const Index actual_mc = (std::min)(i2+mc,end)-i2;
-          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder,false>()
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder,false>()
             (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
           gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
@@ -299,7 +299,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
 
diff --git a/Eigen/src/Core/products/TriangularSolverMatrix.h b/Eigen/src/Core/products/TriangularSolverMatrix.h
index 96d1b38a9..8ff2e9d9d 100644
--- a/Eigen/src/Core/products/TriangularSolverMatrix.h
+++ b/Eigen/src/Core/products/TriangularSolverMatrix.h
@@ -76,7 +76,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
     conj_if<Conjugate> conj;
     gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, TriStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, TriStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
 
     // the goal here is to subdivise the Rhs panels such that we keep some cache
@@ -229,7 +229,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket, ColMajor, false, true> pack_lhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor, false, true> pack_lhs_panel;
 
     for(Index k2=IsLower ? size : 0;
         IsLower ? k2>0 : k2<size;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 9614aeefe..3b22e43e7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -620,7 +620,7 @@ struct TensorContractionEvaluatorBase
     typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
 
     // Declare GEBP packing and kernel structs
-    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, typename Traits::LhsPacket, ColMajor> pack_lhs;
+    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor> pack_lhs;
     internal::gemm_pack_rhs<RhsScalar, Index, typename RhsMapper::SubMapper, nr, ColMajor> pack_rhs;
 
     internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper, mr, nr, false, false> gebp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 1d145c4b1..0980854b4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -244,7 +244,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         RhsMapper;
     typedef internal::gemm_pack_lhs<LhsScalar, Index,
                                     typename LhsMapper::SubMapper, Traits::mr,
-                                    Traits::LhsProgress, ColMajor>
+                                    Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor>
         LhsPacker;
     typedef internal::gemm_pack_rhs<
         RhsScalar, Index, typename RhsMapper::SubMapper, Traits::nr, ColMajor>

From eeeb18814feafcc7a5a2de27e0b8b25554bf7685 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 17:48:56 +0200
Subject: [PATCH 671/680] Fix warning

---
 test/packetmath.cpp | 31 ++++++++++++++++---------------
 1 file changed, 16 insertions(+), 15 deletions(-)

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 58a1c60bf..2b0dda573 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -148,24 +148,25 @@ template<typename Scalar> void packetmath()
 
   for (int offset=0; offset<PacketSize; ++offset)
   {
+    #define MIN(A,B) (A<B?A:B)
     packets[0] = internal::pload<Packet>(data1);
     packets[1] = internal::pload<Packet>(data1+PacketSize);
          if (offset==0) internal::palign<0>(packets[0], packets[1]);
-    else if (offset==1) internal::palign<1>(packets[0], packets[1]);
-    else if (offset==2) internal::palign<2>(packets[0], packets[1]);
-    else if (offset==3) internal::palign<3>(packets[0], packets[1]);
-    else if (offset==4) internal::palign<4>(packets[0], packets[1]);
-    else if (offset==5) internal::palign<5>(packets[0], packets[1]);
-    else if (offset==6) internal::palign<6>(packets[0], packets[1]);
-    else if (offset==7) internal::palign<7>(packets[0], packets[1]);
-    else if (offset==8) internal::palign<8>(packets[0], packets[1]);
-    else if (offset==9) internal::palign<9>(packets[0], packets[1]);
-    else if (offset==10) internal::palign<10>(packets[0], packets[1]);
-    else if (offset==11) internal::palign<11>(packets[0], packets[1]);
-    else if (offset==12) internal::palign<12>(packets[0], packets[1]);
-    else if (offset==13) internal::palign<13>(packets[0], packets[1]);
-    else if (offset==14) internal::palign<14>(packets[0], packets[1]);
-    else if (offset==15) internal::palign<15>(packets[0], packets[1]);
+    else if (offset==1) internal::palign<MIN(1,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==2) internal::palign<MIN(2,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==3) internal::palign<MIN(3,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==4) internal::palign<MIN(4,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==5) internal::palign<MIN(5,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==6) internal::palign<MIN(6,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==7) internal::palign<MIN(7,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==8) internal::palign<MIN(8,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==9) internal::palign<MIN(9,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==10) internal::palign<MIN(10,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==11) internal::palign<MIN(11,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==12) internal::palign<MIN(12,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==13) internal::palign<MIN(13,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==14) internal::palign<MIN(14,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==15) internal::palign<MIN(15,PacketSize-1)>(packets[0], packets[1]);
     internal::pstore(data2, packets[0]);
 
     for (int i=0; i<PacketSize; ++i)

From e0f6d352fb24ad823125a20a4d12f97035bcc2b7 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 18:07:32 +0200
Subject: [PATCH 672/680] Rename test/array.cpp to test/array_cwise.cpp to
 avoid conflicts with the array header.

---
 test/CMakeLists.txt                 | 2 +-
 test/{array.cpp => array_cwise.cpp} | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
 rename test/{array.cpp => array_cwise.cpp} (99%)

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index b4730cff0..c215bfbb2 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -198,7 +198,7 @@ ei_add_test(smallvectors)
 ei_add_test(mapped_matrix)
 ei_add_test(mapstride)
 ei_add_test(mapstaticmethods)
-ei_add_test(array)
+ei_add_test(array_cwise)
 ei_add_test(array_for_matrix)
 ei_add_test(array_replicate)
 ei_add_test(array_reverse)
diff --git a/test/array.cpp b/test/array_cwise.cpp
similarity index 99%
rename from test/array.cpp
rename to test/array_cwise.cpp
index d9c4626c0..84e46665b 100644
--- a/test/array.cpp
+++ b/test/array_cwise.cpp
@@ -453,7 +453,7 @@ template<typename ArrayType> void min_max(const ArrayType& m)
 
 }
 
-EIGEN_DECLARE_TEST(array)
+EIGEN_DECLARE_TEST(array_cwise)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( array(Array<float, 1, 1>()) );

From 3c6dc93f998f2efd52754805458b0fe7172ab5ed Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 18:29:21 +0200
Subject: [PATCH 673/680] Fix GPU support.

---
 .../CXX11/src/Tensor/TensorContractionGpu.h   | 58 +++++++++----------
 1 file changed, 29 insertions(+), 29 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index b5e186d21..056665749 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -549,12 +549,12 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
 #define prefetch_lhs(reg, row, col)                            \
     if (!CHECK_LHS_BOUNDARY) {                                 \
       if (col < k_size) {                                      \
-        reg =lhs.template loadPacket<Unaligned>(row, col);     \
+        reg =lhs.template loadPacket<float4,Unaligned>(row, col);     \
       }                                                        \
     } else {                                                   \
       if (col < k_size) {                                      \
         if (row + 3 < m_size) {                                \
-          reg =lhs.template loadPacket<Unaligned>(row, col);   \
+          reg =lhs.template loadPacket<float4,Unaligned>(row, col);   \
         } else if (row + 2 < m_size) {                         \
           reg.x =lhs(row + 0, col);                            \
           reg.y =lhs(row + 1, col);                            \
@@ -584,7 +584,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -599,7 +599,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     } else {
       if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
           rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
@@ -799,37 +799,37 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 
      if (!CHECK_LHS_BOUNDARY) {
       if ((threadIdx.y/4+k+24) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-        lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
       } else if ((threadIdx.y/4+k+16) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
       } else if ((threadIdx.y/4+k+8) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
       } else if ((threadIdx.y/4+k) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
       }
     } else {
       // just CHECK_LHS_BOUNDARY
       if (lhs_vert + 3 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-          lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
         } else if ((threadIdx.y/4+k+16) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
         } else if ((threadIdx.y/4+k+8) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         } else if ((threadIdx.y/4+k) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
         }
       } else if (lhs_vert + 2 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
@@ -918,8 +918,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-        rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -941,8 +941,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       if (rhs_horiz1 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
         } else if (rhs_vert + 2 < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -963,7 +963,7 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       } else if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);

From d37188b9c193aaf087f67e21e654f36855b88ac0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 18:30:10 +0200
Subject: [PATCH 674/680] Fix MPrealSupport

---
 unsupported/Eigen/MPRealSupport | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport
index 6392bea91..c4ea4ec5f 100644
--- a/unsupported/Eigen/MPRealSupport
+++ b/unsupported/Eigen/MPRealSupport
@@ -162,6 +162,7 @@ int main()
       typedef ResScalar LhsPacket;
       typedef ResScalar RhsPacket;
       typedef ResScalar ResPacket;
+      typedef LhsPacket LhsPacket4Packing;
       
     };
 

From 5d2e759329b4deac6b3c357bf9dfc8e9eab75b3f Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Thu, 20 Sep 2018 11:40:43 -0700
Subject: [PATCH 675/680] Initialize BlockIteratorState in a C++03 compatible
 way.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
index aa2e5bcc6..558130300 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
@@ -188,7 +188,14 @@ class TensorBlockIO {
     StorageIndex input_span;
     StorageIndex output_span;
     StorageIndex size;
-    StorageIndex count = 0;
+    StorageIndex count;
+    BlockIteratorState()
+        : input_stride(0),
+          output_stride(0),
+          input_span(0),
+          output_span(0),
+          size(0),
+          count(0) {}
   };
 
   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(

From 0c56d22e2e10a81cb72f74a6751049f25c04ba98 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 22:56:21 +0200
Subject: [PATCH 676/680] Fix shadowing

---
 Eigen/src/SVD/BDCSVD.h              | 10 +++++-----
 Eigen/src/SparseCore/SparseMatrix.h |  4 ++--
 doc/examples/nullary_indexing.cpp   |  8 ++++----
 3 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h
index 11df14918..4daa9dd21 100644
--- a/Eigen/src/SVD/BDCSVD.h
+++ b/Eigen/src/SVD/BDCSVD.h
@@ -945,7 +945,7 @@ void BDCSVD<MatrixType>::perturbCol0
     zhat.setZero();
     return;
   }
-  Index last = perm(m-1);
+  Index lastIdx = perm(m-1);
   // The offset permits to skip deflated entries while computing zhat
   for (Index k = 0; k < n; ++k)
   {
@@ -955,12 +955,12 @@ void BDCSVD<MatrixType>::perturbCol0
     {
       // see equation (3.6)
       RealScalar dk = diag(k);
-      RealScalar prod = (singVals(last) + dk) * (mus(last) + (shifts(last) - dk));
+      RealScalar prod = (singVals(lastIdx) + dk) * (mus(lastIdx) + (shifts(lastIdx) - dk));
 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
       if(prod<0) {
         std::cout << "k = " << k << " ;  z(k)=" << col0(k) << ", diag(k)=" << dk << "\n";
-        std::cout << "prod = " << "(" << singVals(last) << " + " << dk << ") * (" << mus(last) << " + (" << shifts(last) << " - " << dk << "))" << "\n";
-        std::cout << "     = " << singVals(last) + dk << " * " << mus(last) + (shifts(last) - dk) <<  "\n";
+        std::cout << "prod = " << "(" << singVals(lastIdx) << " + " << dk << ") * (" << mus(lastIdx) << " + (" << shifts(lastIdx) << " - " << dk << "))" << "\n";
+        std::cout << "     = " << singVals(lastIdx) + dk << " * " << mus(lastIdx) + (shifts(lastIdx) - dk) <<  "\n";
       }
       assert(prod>=0);
 #endif
@@ -1000,7 +1000,7 @@ void BDCSVD<MatrixType>::perturbCol0
         }
       }
 #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
-      std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n";
+      std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(lastIdx) + dk) << " * " << mus(lastIdx) + shifts(lastIdx) << " - " << dk << "\n";
 #endif
       RealScalar tmp = sqrt(prod);
 #ifdef EIGEN_BDCSVD_SANITY_CHECKS
diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h
index 8f77194b6..8bfa5f6b8 100644
--- a/Eigen/src/SparseCore/SparseMatrix.h
+++ b/Eigen/src/SparseCore/SparseMatrix.h
@@ -604,9 +604,9 @@ class SparseMatrix
       m_outerIndex = newOuterIndex;
       if (outerChange > 0)
       {
-        StorageIndex last = m_outerSize == 0 ? 0 : m_outerIndex[m_outerSize];
+        StorageIndex lastIdx = m_outerSize == 0 ? 0 : m_outerIndex[m_outerSize];
         for(Index i=m_outerSize; i<m_outerSize+outerChange+1; i++)          
-          m_outerIndex[i] = last; 
+          m_outerIndex[i] = lastIdx; 
       }
       m_outerSize += outerChange;
     }
diff --git a/doc/examples/nullary_indexing.cpp b/doc/examples/nullary_indexing.cpp
index e27c3585a..ca1745628 100644
--- a/doc/examples/nullary_indexing.cpp
+++ b/doc/examples/nullary_indexing.cpp
@@ -30,7 +30,7 @@ public:
 // [function]
 template <class ArgType, class RowIndexType, class ColIndexType>
 CwiseNullaryOp<indexing_functor<ArgType,RowIndexType,ColIndexType>, typename indexing_functor<ArgType,RowIndexType,ColIndexType>::MatrixType>
-indexing(const Eigen::MatrixBase<ArgType>& arg, const RowIndexType& row_indices, const ColIndexType& col_indices)
+mat_indexing(const Eigen::MatrixBase<ArgType>& arg, const RowIndexType& row_indices, const ColIndexType& col_indices)
 {
   typedef indexing_functor<ArgType,RowIndexType,ColIndexType> Func;
   typedef typename Func::MatrixType MatrixType;
@@ -45,7 +45,7 @@ int main()
   Eigen::MatrixXi A = Eigen::MatrixXi::Random(4,4);
   Array3i ri(1,2,1);
   ArrayXi ci(6); ci << 3,2,1,0,0,2;
-  Eigen::MatrixXi B = indexing(A, ri, ci);
+  Eigen::MatrixXi B = mat_indexing(A, ri, ci);
   std::cout << "A =" << std::endl;
   std::cout << A << std::endl << std::endl;
   std::cout << "A([" << ri.transpose() << "], [" << ci.transpose() << "]) =" << std::endl;
@@ -53,11 +53,11 @@ int main()
   std::cout << "[main1]\n";
 
   std::cout << "[main2]\n";
-  B =  indexing(A, ri+1, ci);
+  B =  mat_indexing(A, ri+1, ci);
   std::cout << "A(ri+1,ci) =" << std::endl;
   std::cout << B << std::endl << std::endl;
 #if __cplusplus >= 201103L
-  B =  indexing(A, ArrayXi::LinSpaced(13,0,12).unaryExpr([](int x){return x%4;}), ArrayXi::LinSpaced(4,0,3));
+  B =  mat_indexing(A, ArrayXi::LinSpaced(13,0,12).unaryExpr([](int x){return x%4;}), ArrayXi::LinSpaced(4,0,3));
   std::cout << "A(ArrayXi::LinSpaced(13,0,12).unaryExpr([](int x){return x%4;}), ArrayXi::LinSpaced(4,0,3)) =" << std::endl;
   std::cout << B << std::endl << std::endl;
 #endif

From 617f75f1178b89c3ec98bd70644d5754b8339a65 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 22:57:10 +0200
Subject: [PATCH 677/680] Add indexing namespace

---
 Eigen/src/Core/ArithmeticSequence.h | 33 +++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/Eigen/src/Core/ArithmeticSequence.h b/Eigen/src/Core/ArithmeticSequence.h
index 42fbf47e8..db6da0001 100644
--- a/Eigen/src/Core/ArithmeticSequence.h
+++ b/Eigen/src/Core/ArithmeticSequence.h
@@ -375,6 +375,39 @@ struct get_compile_time_incr<ArithmeticSequence<FirstType,SizeType,IncrType> > {
 
 } // end namespace internal
 
+/** \namespace Eigen::indexing
+  * \ingroup Core_Module
+  * 
+  * The sole purpose of this namespace is to be able to import all functions
+  * and symbols that are expected to be used within operator() for indexing
+  * and slicing. If you already imported the whole Eigen namespace:
+  * \code using namespace Eigen; \endcode
+  * then you are already all set. Otherwise, if you don't want/cannot import
+  * the whole Eigen namespace, the following line:
+  * \code using namespace Eigen::indexing; \endcode
+  * is equivalent to:
+  * \code
+  using Eigen::all;
+  using Eigen::seq;
+  using Eigen::seqN;
+  using Eigen::lastN; // c++11 only
+  using Eigen::last;
+  using Eigen::lastp1;
+  using Eigen::fix;
+  \endcode
+  */
+namespace indexing {
+  using Eigen::all;
+  using Eigen::seq;
+  using Eigen::seqN;
+  #if EIGEN_HAS_CXX11
+  using Eigen::lastN;
+  #endif
+  using Eigen::last;
+  using Eigen::lastp1;
+  using Eigen::fix;
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_ARITHMETIC_SEQUENCE_H

From f0ef3467de0b9a726024e7d5efb3aff007ad6f93 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 20 Sep 2018 22:57:28 +0200
Subject: [PATCH 678/680] Fix doc

---
 Eigen/src/Core/util/SymbolicIndex.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Eigen/src/Core/util/SymbolicIndex.h b/Eigen/src/Core/util/SymbolicIndex.h
index 41d477056..17cf46f05 100644
--- a/Eigen/src/Core/util/SymbolicIndex.h
+++ b/Eigen/src/Core/util/SymbolicIndex.h
@@ -35,7 +35,7 @@ namespace Eigen {
   * std::cout << expr98.eval(x=6) << "\n";
   * \endcode
   *
-  * It is currently only used internally to define and manipulate the placeholders::last and placeholders::end symbols in Eigen::seq and Eigen::seqN.
+  * It is currently only used internally to define and manipulate the Eigen::last and Eigen::lastp1 symbols in Eigen::seq and Eigen::seqN.
   *
   */
 namespace symbolic {

From 719e438a208350474f384b53c76ae062e210d856 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Thu, 20 Sep 2018 15:19:12 -0700
Subject: [PATCH 679/680] Collapsed revision * Split cxx11_tensor_executor test
 * Register test parts with EIGEN_SUFFIXES * Fix EIGEN_SUFFIXES in
 cxx11_tensor_executor test

---
 unsupported/test/cxx11_tensor_executor.cpp | 129 +++++++++++----------
 1 file changed, 70 insertions(+), 59 deletions(-)

diff --git a/unsupported/test/cxx11_tensor_executor.cpp b/unsupported/test/cxx11_tensor_executor.cpp
index 1bb99854c..aa789c2e4 100644
--- a/unsupported/test/cxx11_tensor_executor.cpp
+++ b/unsupported/test/cxx11_tensor_executor.cpp
@@ -31,7 +31,8 @@ static array<Index, NumDims> RandomDims(int min_dim = 1, int max_dim = 20) {
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
-static void test_execute_unary_expr(Device d) {
+static void test_execute_unary_expr(Device d)
+{
   static constexpr int Options = 0 | Layout;
 
   // Pick a large enough tensor size to bypass small tensor block evaluation
@@ -121,7 +122,8 @@ static void test_execute_broadcasting(Device d)
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
-static void test_execute_chipping_rvalue(Device d) {
+static void test_execute_chipping_rvalue(Device d)
+{
   auto dims = RandomDims<NumDims>(1, 10);
   Tensor<T, NumDims, Layout, Index> src(dims);
   src.setRandom();
@@ -159,7 +161,8 @@ static void test_execute_chipping_rvalue(Device d) {
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
     bool Tileable, int Layout>
-static void test_execute_chipping_lvalue(Device d) {
+static void test_execute_chipping_lvalue(Device d)
+{
   auto dims = RandomDims<NumDims>(1, 10);
 
 #define TEST_CHIPPING(CHIP_DIM)                                             \
@@ -211,7 +214,8 @@ static void test_execute_chipping_lvalue(Device d) {
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
-static void test_execute_shuffle_rvalue(Device d) {
+static void test_execute_shuffle_rvalue(Device d)
+{
   static constexpr int Options = 0 | Layout;
 
   auto dims = RandomDims<NumDims>(1, 10);
@@ -246,7 +250,8 @@ static void test_execute_shuffle_rvalue(Device d) {
 
 template <typename T, int NumDims, typename Device, bool Vectorizable,
           bool Tileable, int Layout>
-static void test_execute_shuffle_lvalue(Device d) {
+static void test_execute_shuffle_lvalue(Device d)
+{
   static constexpr int Options = 0 | Layout;
 
   auto dims = RandomDims<NumDims>(5, 10);
@@ -447,23 +452,26 @@ static void test_execute_slice_lvalue(Device d)
   }
 }
 
-#define CALL_SUBTEST_COMBINATIONS(NAME, T, NUM_DIMS)                                           \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, false, ColMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  ColMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, ColMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  ColMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, false, RowMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  RowMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, RowMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  RowMajor>(default_device))); \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  ColMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, ColMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  ColMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  RowMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, RowMajor>(tp_device)));      \
-  CALL_SUBTEST((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  RowMajor>(tp_device)))
+#define CALL_SUBTEST_PART(PART) \
+  CALL_SUBTEST_##PART
+
+#define CALL_SUBTEST_COMBINATIONS(PART, NAME, T, NUM_DIMS)                                                \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    false, false, ColMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  ColMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, ColMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  ColMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    false, false, RowMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    false, true,  RowMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    true,  false, RowMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice,    true,  true,  RowMajor>(default_device))); \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  ColMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, ColMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  ColMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true,  RowMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  false, RowMajor>(tp_device)));      \
+  CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, true,  true,  RowMajor>(tp_device)))
 
 EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
   Eigen::DefaultDevice default_device;
@@ -472,53 +480,56 @@ EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
   Eigen::ThreadPool tp(num_threads);
   Eigen::ThreadPoolDevice tp_device(&tp, num_threads);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_unary_expr, float, 5);
+  CALL_SUBTEST_COMBINATIONS(1, test_execute_unary_expr, float, 3);
+  CALL_SUBTEST_COMBINATIONS(1, test_execute_unary_expr, float, 4);
+  CALL_SUBTEST_COMBINATIONS(1, test_execute_unary_expr, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_binary_expr, float, 5);
+  CALL_SUBTEST_COMBINATIONS(2, test_execute_binary_expr, float, 3);
+  CALL_SUBTEST_COMBINATIONS(2, test_execute_binary_expr, float, 4);
+  CALL_SUBTEST_COMBINATIONS(2, test_execute_binary_expr, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_broadcasting, float, 5);
+  CALL_SUBTEST_COMBINATIONS(3, test_execute_broadcasting, float, 3);
+  CALL_SUBTEST_COMBINATIONS(3, test_execute_broadcasting, float, 4);
+  CALL_SUBTEST_COMBINATIONS(3, test_execute_broadcasting, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_rvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(4, test_execute_chipping_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(4, test_execute_chipping_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(4, test_execute_chipping_rvalue, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_chipping_lvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(5, test_execute_chipping_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(5, test_execute_chipping_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(5, test_execute_chipping_lvalue, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_rvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(6, test_execute_shuffle_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(6, test_execute_shuffle_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(6, test_execute_shuffle_rvalue, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_shuffle_lvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(7, test_execute_shuffle_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(7, test_execute_shuffle_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(7, test_execute_shuffle_lvalue, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 2);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reduction, float, 5);
+  CALL_SUBTEST_COMBINATIONS(8, test_execute_reduction, float, 2);
+  CALL_SUBTEST_COMBINATIONS(8, test_execute_reduction, float, 3);
+  CALL_SUBTEST_COMBINATIONS(8, test_execute_reduction, float, 4);
+  CALL_SUBTEST_COMBINATIONS(8, test_execute_reduction, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 2);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_reshape, float, 5);
+  CALL_SUBTEST_COMBINATIONS(9, test_execute_reshape, float, 2);
+  CALL_SUBTEST_COMBINATIONS(9, test_execute_reshape, float, 3);
+  CALL_SUBTEST_COMBINATIONS(9, test_execute_reshape, float, 4);
+  CALL_SUBTEST_COMBINATIONS(9, test_execute_reshape, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 2);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_rvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(10, test_execute_slice_rvalue, float, 2);
+  CALL_SUBTEST_COMBINATIONS(10, test_execute_slice_rvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(10, test_execute_slice_rvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(10, test_execute_slice_rvalue, float, 5);
 
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 2);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 3);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 4);
-  CALL_SUBTEST_COMBINATIONS(test_execute_slice_lvalue, float, 5);
+  CALL_SUBTEST_COMBINATIONS(11, test_execute_slice_lvalue, float, 2);
+  CALL_SUBTEST_COMBINATIONS(11, test_execute_slice_lvalue, float, 3);
+  CALL_SUBTEST_COMBINATIONS(11, test_execute_slice_lvalue, float, 4);
+  CALL_SUBTEST_COMBINATIONS(11, test_execute_slice_lvalue, float, 5);
+
+  // Force CMake to split this test.
+  // EIGEN_SUFFIXES;1;2;3;4;5;6;7;8;9;10;11
 }
 
 #undef CALL_SUBTEST_COMBINATIONS

From 651e5d4866fb82f30e548029c22834b18897c116 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 21 Sep 2018 10:33:22 +0200
Subject: [PATCH 680/680] Fix
 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE for AVX512 or AVX
 with malloc aligned on 8 bytes only. This change also make it future proof
 for AVX1024

---
 Eigen/src/Core/util/Memory.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index 76472703e..9dd2e0252 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -812,8 +812,12 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 #endif
 
 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
-#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
-  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0)))
+#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)                        \
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(                                                             \
+        ((Size)!=Eigen::Dynamic) &&                                                                    \
+        (((EIGEN_MAX_ALIGN_BYTES>=16) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES  )==0)) ||    \
+         ((EIGEN_MAX_ALIGN_BYTES>=32) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/2)==0)) ||    \
+         ((EIGEN_MAX_ALIGN_BYTES>=64) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/4)==0))   )))
 
 /****************************************************************************/