GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-11 11:19:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Initial VSX commit

Browse Source
This commit is contained in:
Konstantinos Margaritis 2014-08-29 20:03:49 +00:00
parent 01f3ca3e8d
commit 7ff266e3ce
10 changed files with 284 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
CMakeLists.txt
View File

@ -219,6 +219,12 @@ if(NOT MSVC)
message(STATUS "Enabling AltiVec in tests/examples")
endif()
option(EIGEN_TEST_VSX "Enable/Disable VSX in tests/examples" OFF)
if(EIGEN_TEST_VSX)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64 -mvsx")
message(STATUS "Enabling VSX in tests/examples")
endif()
option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF)
if(EIGEN_TEST_NEON)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a8")

17
Eigen/Core
View File

@ -171,6 +171,15 @@
#undef bool
#undef vector
#undef pixel
#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 __ARM_NEON__
#define EIGEN_VECTORIZE
#define EIGEN_VECTORIZE_NEON
@ -235,6 +244,8 @@ inline static const char *SimdInstructionSetsInUse(void) {
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";
#else
@ -286,9 +297,13 @@ using std::ptrdiff_t;
#include "src/Core/arch/SSE/PacketMath.h"
#include "src/Core/arch/SSE/MathFunctions.h"
#include "src/Core/arch/SSE/Complex.h"
#elif defined EIGEN_VECTORIZE_ALTIVEC
#elif defined(EIGEN_VECTORIZE_ALTIVEC)
#include "src/Core/arch/AltiVec/PacketMath.h"
#include "src/Core/arch/AltiVec/Complex.h"
#elif defined(EIGEN_VECTORIZE_ALTIVEC)
#include "src/Core/arch/AltiVec/PacketMath.h"
#include "src/Core/arch/AltiVec/VSX.h"
#include "src/Core/arch/AltiVec/Complex.h"
#elif defined EIGEN_VECTORIZE_NEON
#include "src/Core/arch/NEON/PacketMath.h"
#include "src/Core/arch/NEON/Complex.h"

186
Eigen/src/Core/arch/AltiVec/Complex.h
View File

@ -14,7 +14,13 @@ namespace Eigen {
namespace internal {
#ifdef _BIG_ENDIAN
static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
#else
static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4i_ZERO);//{ 0x80000000, 0x00000000, 0x80000000, 0x00000000 };
static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui)p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
#endif
static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
static Packet16uc p16uc_COMPLEX_IM = vec_sld(p16uc_DUPLICATE, (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_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
@ -237,6 +243,186 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
kernel.packet[0].v = tmp;
}
//---------- double ----------
struct Packet1cd
{
EIGEN_STRONG_INLINE Packet1cd() {}
EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
Packet2d v;
};
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}; typedef Packet1cd half; };
template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2ul)a.v, p2ul_CONJ_XOR)); }
template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
{
Packet2d v1, v2;
// Permute and multiply the real parts of a and b
v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE);
// Get the imaginary parts of a
v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM);
// 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 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR);
// permute back to a proper order
v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV);
return Packet2cf(vec_add(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 pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet4f>((const double*)from)); }
template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet4f>((const double*)from)); }
template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from)
{
return pset1<Packet1cd>(*from);
}
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) { vec_dstt((double *)addr, DST_CTRL(2,2,32), DST_CHAN); }
template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
{
std::complex<double> EIGEN_ALIGN16 res[2];
pstore((float *)&res, a.v);
return res[0];
}
template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a)
{
Packet2d rev_a;
rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2);
return Packet1cd(rev_a);
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
{
Packet2d b;
b = (Packet2d) vec_sld(a.v, a.v, 8);
b = padd(a.v, b);
return pfirst(Packet1cd(b));
}
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
{
Packet2d b1, b2;
b1 = (Packet2d) vec_sld(vecs[0].v, vecs[1].v, 8);
b2 = (Packet2d) vec_sld(vecs[1].v, vecs[0].v, 8);
b2 = (Packet2d) vec_sld(b2, b2, 8);
b2 = padd(b1, b2);
return Packet1cd(b2);
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
{
Packet2d b;
Packet1cd prod;
b = (Packet2d) vec_sld(a.v, a.v, 8);
prod = pmul(a, Packet1cd(b));
return pfirst(prod);
}
template<int Offset>
struct palign_impl<Offset,Packet1cd>
{
static EIGEN_STRONG_INLINE void run(Packet1cd& first, const Packet1cd& second)
{ }
};
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));
}
};
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, p4f_ZERO);
return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV))));
}
template<> EIGEN_STRONG_INLINE Packet1cd pcplxflip<Packet1cd>(const Packet1cd& x)
{
return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX_REV));
}
EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
{
Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0);
kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1);
kernel.packet[0].v = tmp;
}
} // end namespace internal
} // end namespace Eigen

64
Eigen/src/Core/arch/AltiVec/PacketMath.h
View File

@ -31,6 +31,12 @@ namespace internal {
#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16
#endif
#ifdef __VSX__
typedef __vector double Packet2d;
typedef __vector unsigned long Packet2ul;
typedef __vector long Packet2l;
#endif // __VSX__
typedef __vector float Packet4f;
typedef __vector int Packet4i;
typedef __vector unsigned int Packet4ui;
@ -50,22 +56,37 @@ typedef __vector unsigned char Packet16uc;
#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
Packet4f p4f_##NAME = pset1<Packet4f>(X)
#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
Packet4i p4i_##NAME = pset1<Packet4i>(X)
#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
Packet2d p2d_##NAME = pset1<Packet2d>(X)
#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
Packet2l p2l_##NAME = pset1<Packet2l>(X)
#define DST_CHAN 1
#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
// Handle endianness properly while loading constants
// Define global static constants:
#ifdef _BIG_ENDIAN
static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
static Packet16uc p16uc_REVERSE = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15}
static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
#else
static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 };
static Packet16uc p16uc_REVERSE = { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 };
static Packet16uc p16uc_FORWARD = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
static Packet16uc p16uc_DUPLICATE = { 4,5,6,7, 4,5,6,7, 0,1,2,3, 0,1,2,3 };
static Packet2d p2d_ZERO_ = (Packet2d) { 0x8000000000000000, 0x8000000000000000 };
static Packet2l p2l_ZERO = (Packet2l) { 0x0, 0x0 };
#endif // _BIG_ENDIAN
// These constants are endian-agnostic
static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 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}
@ -93,6 +114,24 @@ template<> struct packet_traits<float> : default_packet_traits
HasSqrt = 0
};
};
#ifdef __VSX__
template<> struct packet_traits<double> : default_packet_traits
{
typedef Packet2d type;
typedef Packet2d half;
enum {
Vectorizable = 1,
AlignedOnScalar = 1,
size=2,
HasHalfPacket = 0,
HasDiv = 1,
HasExp = 0,
HasSqrt = 0
};
};
#endif // __VSX__
template<> struct packet_traits<int> : default_packet_traits
{
typedef Packet4i type;
@ -105,6 +144,10 @@ template<> struct packet_traits<int> : default_packet_traits
};
};
#ifdef __VSX__
template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; typedef Packet2d half; };
#endif // __VSX__
template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; typedef Packet4f half; };
template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; typedef Packet4i half; };
/*
@ -311,7 +354,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
@ -327,10 +369,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, con
template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
#ifndef __VSX__
template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
{
EIGEN_DEBUG_ALIGNED_LOAD
// Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
Packet16uc MSQ, LSQ;
Packet16uc mask;
MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword
@ -350,6 +392,18 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
mask = vec_lvsl(0, from); // create the permute mask
return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data
}
#else
template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
{
EIGEN_DEBUG_ALIGNED_LOAD
return (Packet4i) vec_vsx_ld((long)from & 15, from); // align the data
}
template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
{
EIGEN_DEBUG_ALIGNED_LOAD
return (Packet4f) vec_vsx_ld((long)from & 15, from); // align the data
}
#endif
template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
{

4
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -182,7 +182,7 @@ public:
nr = 4,
// register block size along the M direction (currently, this one cannot be modified)
#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC)
#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
// we assume 16 registers
mr = 3*LhsPacketSize,
#else
@ -290,7 +290,7 @@ public:
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
nr = 4,
#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC)
#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
// we assume 16 registers
mr = 3*LhsPacketSize,
#else

6
Eigen/src/Core/util/Constants.h
View File

@ -413,10 +413,16 @@ namespace Architecture
Generic = 0x0,
SSE = 0x1,
AltiVec = 0x2,
VSX = 0x3,
NEON = 0x4,
#if defined EIGEN_VECTORIZE_SSE
Target = SSE
#elif defined EIGEN_VECTORIZE_ALTIVEC
Target = AltiVec
#elif defined EIGEN_VECTORIZE_VSX
Target = VSX
#elif defined EIGEN_VECTORIZE_NEON
Target = NEON
#else
Target = Generic
#endif

2
bench/btl/libs/eigen2/eigen2_interface.hh
View File

@ -47,7 +47,7 @@ public :
{
#if defined(EIGEN_VECTORIZE_SSE)
if (SIZE==Dynamic) return "eigen2"; else return "tiny_eigen2";
#elif defined(EIGEN_VECTORIZE_ALTIVEC)
#elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX)
if (SIZE==Dynamic) return "eigen2"; else return "tiny_eigen2";
#else
if (SIZE==Dynamic) return "eigen2_novec"; else return "tiny_eigen2_novec";

6
cmake/EigenTesting.cmake
View File

@ -282,6 +282,12 @@ macro(ei_testing_print_summary)
message(STATUS "Altivec: Using architecture defaults")
endif()
if(EIGEN_TEST_VSX)
message(STATUS "VSX: ON")
else()
message(STATUS "VSX: Using architecture defaults")
endif()
if(EIGEN_TEST_NEON)
message(STATUS "ARM NEON: ON")
else()

2
test/main.h
View File

@ -76,7 +76,7 @@
#endif
// bounds integer values for AltiVec
#ifdef __ALTIVEC__
#if defined(__ALTIVEC__) || defined(__VSX__)
#define EIGEN_MAKING_DOCS
#endif

2
test/packetmath.cpp
View File

@ -156,7 +156,7 @@ template<typename Scalar> void packetmath()
CHECK_CWISE2(REF_ADD, internal::padd);
CHECK_CWISE2(REF_SUB, internal::psub);
CHECK_CWISE2(REF_MUL, internal::pmul);
#ifndef EIGEN_VECTORIZE_ALTIVEC
#if !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
if (!internal::is_same<Scalar,int>::value)
CHECK_CWISE2(REF_DIV, internal::pdiv);
#endif