replaced _mm_prefetch in GeneralBlockPanelKernel.h, with ei_prefetch() inline function.

Implemented NEON and AltiVec versions, copied SSE version over from GeneralBlockPanelKernel.h.
Also in GCC case (or rather !_MSC_VER) it's implemented using __builtin_prefetch().
NEON managed to give a small but welcome boost, 0.88GFLOPS -> 0.91GFLOPS.

Eigen/src/Core/GenericPacketMath.h

@@ -169,6 +169,14 @@ template<typename Scalar, typename Packet> inline void ei_pstore(Scalar* to, con
 template<typename Scalar, typename Packet> inline void ei_pstoreu(Scalar* to, const Packet& from)
 { (*to) = from; }
 
+/** \internal tries to do cache prefetching of \a addr */
+template<typename Scalar> inline void ei_prefetch(const Scalar* addr)
+{
+#if !defined(_MSC_VER)
+__builtin_prefetch(addr);
+#endif
+}
+
 /** \internal \returns the first element of a packet */
 template<typename Packet> inline typename ei_unpacket_traits<Packet>::type ei_pfirst(const Packet& a)
 { return a; }

Eigen/src/Core/arch/AltiVec/PacketMath.h

@@ -67,6 +67,8 @@ typedef __vector unsigned char  Packet16uc;
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   Packet4i ei_p4i_##NAME = ei_pset1<int>(X)
 
+#define DST_CHAN 1
+#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
 
 // Define global static constants:
 static Packet4f ei_p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
@@ -291,8 +293,8 @@ template<> EIGEN_STRONG_INLINE void ei_pstoreu<float>(float*  to, const Packet4f
   edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
   edges=vec_perm(LSQ,MSQ,edgeAlign);                        // extract the edges
   align = vec_lvsr( 0, to );                                // permute map to misalign data
-  MSQ = vec_perm(edges,(Packet16uc)from,align);   // misalign the data (MSQ)
+  MSQ = vec_perm(edges,(Packet16uc)from,align);             // misalign the data (MSQ)
-  LSQ = vec_perm((Packet16uc)from,edges,align);   // misalign the data (LSQ)
+  LSQ = vec_perm((Packet16uc)from,edges,align);             // misalign the data (LSQ)
   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
 }
@@ -315,6 +317,9 @@ template<> EIGEN_STRONG_INLINE void ei_pstoreu<int>(int*      to, const Packet4i
   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
 }
 
+template<> EIGEN_STRONG_INLINE void ei_prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
+template<> EIGEN_STRONG_INLINE void ei_prefetch<int>(const int*     addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
+
 template<> EIGEN_STRONG_INLINE float  ei_pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
 template<> EIGEN_STRONG_INLINE int    ei_pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
 

Eigen/src/Core/arch/NEON/PacketMath.h

@@ -53,6 +53,10 @@ typedef int32x4_t   Packet4i;
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   const Packet4i ei_p4i_##NAME = ei_pset1<int>(X)
 
+#ifndef __pld
+#define __pld(x) asm volatile ( "   pld [%[addr]]\n" :: [addr] "r" (x) : "cc" );
+#endif
+
 template<> struct ei_packet_traits<float>  : ei_default_packet_traits
 {
   typedef Packet4f type; enum {size=4};
@@ -168,6 +172,9 @@ template<> EIGEN_STRONG_INLINE void ei_pstore<int>(int*       to, const Packet4i
 template<> EIGEN_STRONG_INLINE void ei_pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
 template<> EIGEN_STRONG_INLINE void ei_pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
 
+template<> EIGEN_STRONG_INLINE void ei_prefetch<float>(const float* addr) { __pld(addr); }
+template<> EIGEN_STRONG_INLINE void ei_prefetch<int>(const int*     addr) { __pld(addr); }
+
 // FIXME only store the 2 first elements ?
 template<> EIGEN_STRONG_INLINE float  ei_pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
 template<> EIGEN_STRONG_INLINE int    ei_pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }

Eigen/src/Core/arch/SSE/PacketMath.h

@@ -233,6 +233,10 @@ template<> EIGEN_STRONG_INLINE void ei_pstoreu<double>(double* to, const Packet2
 template<> EIGEN_STRONG_INLINE void ei_pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE ei_pstoreu((double*)to, _mm_castps_pd(from)); }
 template<> EIGEN_STRONG_INLINE void ei_pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE ei_pstoreu((double*)to, _mm_castsi128_pd(from)); }
 
+template<> EIGEN_STRONG_INLINE void ei_prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void ei_prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void ei_prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+
 #if defined(_MSC_VER) && (_MSC_VER <= 1500) && defined(_WIN64)
 // The temporary variable fixes an internal compilation error.
 // Direct of the struct members fixed bug #62.

Eigen/src/Core/products/GeneralBlockPanelKernel.h

@@ -117,9 +117,7 @@ struct ei_gebp_kernel
       for(int i=0; i<peeled_mc; i+=mr)
       {
         const Scalar* blA = &blockA[i*strideA+offsetA*mr];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         // TODO move the res loads to the stores
 
@@ -139,12 +137,10 @@ struct ei_gebp_kernel
         Scalar* r2 = r1 + resStride;
         Scalar* r3 = r2 + resStride;
 
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(r0+16);
-        _mm_prefetch((const char*)(r0+16), _MM_HINT_T0);
+        ei_prefetch(r1+16);
-        _mm_prefetch((const char*)(r1+16), _MM_HINT_T0);
+        ei_prefetch(r2+16);
-        _mm_prefetch((const char*)(r2+16), _MM_HINT_T0);
+        ei_prefetch(r3+16);
-        _mm_prefetch((const char*)(r3+16), _MM_HINT_T0);
-        #endif
 
         // performs "inner" product
         // TODO let's check wether the folowing peeled loop could not be
@@ -334,9 +330,7 @@ struct ei_gebp_kernel
       {
         int i = peeled_mc;
         const Scalar* blA = &blockA[i*strideA+offsetA*PacketSize];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         // gets res block as register
         PacketType C0, C1, C2, C3;
@@ -464,9 +458,7 @@ struct ei_gebp_kernel
       for(int i=peeled_mc2; i<rows; i++)
       {
         const Scalar* blA = &blockA[i*strideA+offsetA];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         // gets a 1 x nr res block as registers
         Scalar C0(0), C1(0), C2(0), C3(0);
@@ -524,9 +516,7 @@ struct ei_gebp_kernel
       for(int i=0; i<peeled_mc; i+=mr)
       {
         const Scalar* blA = &blockA[i*strideA+offsetA*mr];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         // TODO move the res loads to the stores
 
@@ -557,9 +547,7 @@ struct ei_gebp_kernel
       {
         int i = peeled_mc;
         const Scalar* blA = &blockA[i*strideA+offsetA*PacketSize];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         PacketType C0 = ei_ploadu(&res[(j2+0)*resStride + i]);
 
@@ -576,9 +564,7 @@ struct ei_gebp_kernel
       for(int i=peeled_mc2; i<rows; i++)
       {
         const Scalar* blA = &blockA[i*strideA+offsetA];
-        #ifdef EIGEN_VECTORIZE_SSE
+        ei_prefetch(&blA[0]);
-        _mm_prefetch((const char*)(&blA[0]), _MM_HINT_T0);
-        #endif
 
         // gets a 1 x 1 res block as registers
         Scalar C0(0);