GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-19 08:09:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Restrict new AVX512 trsm to AVX512VL, rename files for consistency.

Browse Source
This commit is contained in:
Antonio Sánchez 2022-04-14 16:58:32 +00:00
parent 67eeba6e72
commit 07db964bde
6 changed files with 14 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Eigen/Core
View File

@ -190,7 +190,9 @@ using std::ptrdiff_t;
#include "src/Core/arch/SSE/MathFunctions.h" #include "src/Core/arch/SSE/MathFunctions.h"
#include "src/Core/arch/AVX/MathFunctions.h" #include "src/Core/arch/AVX/MathFunctions.h"
#include "src/Core/arch/AVX512/MathFunctions.h" #include "src/Core/arch/AVX512/MathFunctions.h"
#include "src/Core/arch/AVX512/trsmKernel_impl.hpp" #ifdef __AVX512VL__
#include "src/Core/arch/AVX512/TrsmKernel.h"
#endif
#elif defined EIGEN_VECTORIZE_AVX #elif defined EIGEN_VECTORIZE_AVX
// Use AVX for floats and doubles, SSE for integers // Use AVX for floats and doubles, SSE for integers
#include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/PacketMath.h"

8
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@ -237,7 +237,7 @@ template<> struct unpacket_traits<Packet8f> {
typedef Packet8i integer_packet; typedef Packet8i integer_packet;
typedef uint8_t mask_t; typedef uint8_t mask_t;
enum {size=8, alignment=Aligned32, vectorizable=true, masked_load_available=true, masked_store_available=true enum {size=8, alignment=Aligned32, vectorizable=true, masked_load_available=true, masked_store_available=true
#ifdef EIGEN_VECTORIZE_AVX512 #ifdef __AVX512VL__
, masked_fpops_available=true , masked_fpops_available=true
#endif #endif
}; };
@ -468,7 +468,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float* from) { r
template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); } template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); }
template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); }
#ifdef EIGEN_VECTORIZE_AVX512 #ifdef __AVX512VL__
template <> template <>
EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b, uint8_t umask) { EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b, uint8_t umask) {
__mmask8 mask = static_cast<__mmask8>(umask); __mmask8 mask = static_cast<__mmask8>(umask);
@ -859,7 +859,7 @@ template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { E
template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); } template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); }
template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from, uint8_t umask) { template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from, uint8_t umask) {
#ifdef EIGEN_VECTORIZE_AVX512 #ifdef __AVX512VL__
__mmask8 mask = static_cast<__mmask8>(umask); __mmask8 mask = static_cast<__mmask8>(umask);
EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_maskz_loadu_ps(mask, from); EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_maskz_loadu_ps(mask, from);
#else #else
@ -927,7 +927,7 @@ template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d&
template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f& from, uint8_t umask) { template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f& from, uint8_t umask) {
#ifdef EIGEN_VECTORIZE_AVX512 #ifdef __AVX512VL__
__mmask8 mask = static_cast<__mmask8>(umask); __mmask8 mask = static_cast<__mmask8>(umask);
EIGEN_DEBUG_UNALIGNED_STORE return _mm256_mask_storeu_ps(to, mask, from); EIGEN_DEBUG_UNALIGNED_STORE return _mm256_mask_storeu_ps(to, mask, from);
#else #else

4
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@ -247,7 +247,7 @@ template <>
EIGEN_STRONG_INLINE Packet16f pload1<Packet16f>(const float* from) { EIGEN_STRONG_INLINE Packet16f pload1<Packet16f>(const float* from) {
#if (EIGEN_COMP_GNUC != 0) || (EIGEN_COMP_CLANG != 0) #if (EIGEN_COMP_GNUC != 0) || (EIGEN_COMP_CLANG != 0)
// Inline asm here helps reduce some register spilling in TRSM kernels. // Inline asm here helps reduce some register spilling in TRSM kernels.
// See note in unrolls::gemm::microKernel in trsmKernel_impl.hpp // See note in unrolls::gemm::microKernel in TrsmKernel.h
Packet16f ret; Packet16f ret;
__asm__ ("vbroadcastss %[mem], %[dst]" : [dst] "=v" (ret) : [mem] "m" (*from)); __asm__ ("vbroadcastss %[mem], %[dst]" : [dst] "=v" (ret) : [mem] "m" (*from));
return ret; return ret;
@ -300,6 +300,7 @@ EIGEN_STRONG_INLINE Packet16i padd<Packet16i>(const Packet16i& a,
const Packet16i& b) { const Packet16i& b) {
return _mm512_add_epi32(a, b); return _mm512_add_epi32(a, b);
} }
template <> template <>
EIGEN_STRONG_INLINE Packet16f padd<Packet16f>(const Packet16f& a, EIGEN_STRONG_INLINE Packet16f padd<Packet16f>(const Packet16f& a,
const Packet16f& b, const Packet16f& b,
@ -800,6 +801,7 @@ EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) {
EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_si512( EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_si512(
reinterpret_cast<const __m512i*>(from)); reinterpret_cast<const __m512i*>(from));
} }
template <> template <>
EIGEN_STRONG_INLINE Packet16f ploadu<Packet16f>(const float* from, uint16_t umask) { EIGEN_STRONG_INLINE Packet16f ploadu<Packet16f>(const float* from, uint16_t umask) {
__mmask16 mask = static_cast<__mmask16>(umask); __mmask16 mask = static_cast<__mmask16>(umask);

5
Eigen/src/Core/arch/AVX512/trsmKernel_impl.hpp → Eigen/src/Core/arch/AVX512/TrsmKernel.h
View File

@ -38,8 +38,9 @@ typedef Packet8d vecFullDouble;
typedef Packet8f vecHalfFloat; typedef Packet8f vecHalfFloat;
typedef Packet4d vecHalfDouble; typedef Packet4d vecHalfDouble;
// Compile-time unrolls are implemented here // Compile-time unrolls are implemented here.
#include "unrolls_impl.hpp" // Note: this depends on macros and typedefs above.
#include "TrsmUnrolls.inc"
#if defined(EIGEN_USE_AVX512_TRSM_KERNELS) && (EIGEN_COMP_CLANG != 0) #if defined(EIGEN_USE_AVX512_TRSM_KERNELS) && (EIGEN_COMP_CLANG != 0)

0
Eigen/src/Core/arch/AVX512/unrolls_impl.hpp → Eigen/src/Core/arch/AVX512/TrsmUnrolls.inc
View File

2
test/triangular.cpp
View File

@ -7,7 +7,7 @@
// Public License v. 2.0. If a copy of the MPL was not distributed // 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/. // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifdef EIGEN_TEST_PART_100 #if defined(EIGEN_TEST_PART_100) || defined(EIGEN_TEST_PART_ALL)
# define EIGEN_NO_DEPRECATED_WARNING # define EIGEN_NO_DEPRECATED_WARNING
#endif #endif