youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

before merge

Browse Source
This commit is contained in:
Nitload
2026-06-27 00:49:01 +08:00
parent 062ffacf88
commit 2a6fc7b633
17 changed files with 184 additions and 121 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/CMakeLists.txt
+1 -1
View File
@@ -49,7 +49,7 @@ else()
set(STGSHARP_ARCH_TAG ${SN_RESOLVED_ARCH_NAME})
endif()
set(STGSHARP_BUILD_OUTPUT_DIR ${STGSHARP_SOLUTION_ROOT}/build/${STGSHARP_OS_TAG}-${STGSHARP_ARCH_TAG})
set(STGSHARP_BUILD_OUTPUT_DIR ${STGSHARP_SOLUTION_ROOT}/cmake_build/${STGSHARP_OS_TAG}_${STGSHARP_ARCH_TAG})
set(STGSHARP_OBJECT_ROOT_DIR ${STGSHARP_SOLUTION_ROOT}/obj/${PROJECT_NAME})
set(STGSHARP_NATIVE_PROJECT_FILES
${STGSHARP_NATIVE_SOURCES}
StgSharp.Native/sn-src/mk.fma.x86.c
+5 -5
View File
@@ -125,11 +125,11 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
register __m256 a1 = _mm256_load_ps((float *)&a_base->y[1]);
register __m256 a2 = _mm256_load_ps((float *)&a_base->y[2]);
register __m256 a3 = _mm256_load_ps((float *)&a_base->y[3]);
SN_MK_FMA_F32_AVX2_FMA_CYCLE(0);
SN_MK_FMA_F32_AVX2_FMA_CYCLE(1);
SN_MK_FMA_F32_AVX2_FMA_CYCLE(2);
SN_MK_FMA_F32_AVX2_FMA_CYCLE(3);
SN_MK_FMA_F32_AVX2_CYCLE(0);
SN_MK_FMA_F32_AVX2_CYCLE(1);
SN_MK_FMA_F32_AVX2_CYCLE(2);
SN_MK_FMA_F32_AVX2_CYCLE(3);
}
_mm256_store_ps((float *)&(ans_cur->y[0]), c0);
_mm256_store_ps((float *)&(ans_cur->y[1]), c1);
StgSharp.Native/sn-template/sn_template.fma.x86.h
+43 -24
View File
@@ -2,36 +2,55 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
#define SN_MK_FMA_F32_512_CYCLE(i) \
do { \
c0 = _mm512_fmadd_ps(a0, _mm512_permute_ps(b0, _FMA_SELECT(i)), c0); \
c1 = _mm512_fmadd_ps(a1, _mm512_permute_ps(b1, _FMA_SELECT(i)), c1); \
c2 = _mm512_fmadd_ps(a2, _mm512_permute_ps(b2, _FMA_SELECT(i)), c2); \
c3 = _mm512_fmadd_ps(a3, _mm512_permute_ps(b3, _FMA_SELECT(i)), c3); \
#define SN_MK_FMA_F32_512_CYCLE(i) \
do { \
c0 = _mm512_fmadd_ps(c0, a0, _mm512_permute_ps(b0, _FMA_SELECT(i))); \
c1 = _mm512_fmadd_ps(c1, a1, _mm512_permute_ps(b1, _FMA_SELECT(i))); \
c2 = _mm512_fmadd_ps(c2, a2, _mm512_permute_ps(b2, _FMA_SELECT(i))); \
c3 = _mm512_fmadd_ps(c3, a3, _mm512_permute_ps(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_MK_FMA_F32_AVX2_FMA_CYCLE(i) \
do { \
c0 = _mm256_fmadd_ps(a0, _mm256_permute_ps(b0, _FMA_SELECT(i)), c0); \
c1 = _mm256_fmadd_ps(a1, _mm256_permute_ps(b1, _FMA_SELECT(i)), c1); \
c2 = _mm256_fmadd_ps(a2, _mm256_permute_ps(b2, _FMA_SELECT(i)), c2); \
c3 = _mm256_fmadd_ps(a3, _mm256_permute_ps(b3, _FMA_SELECT(i)), c3); \
#define SN_MK_FMA_F32_AVX2_CYCLE(i) \
do { \
c0 = _mm256_fmadd_ps(c0, a0, _mm256_permute_ps(b0, _FMA_SELECT(i))); \
c1 = _mm256_fmadd_ps(c1, a1, _mm256_permute_ps(b1, _FMA_SELECT(i))); \
c2 = _mm256_fmadd_ps(c2, a2, _mm256_permute_ps(b2, _FMA_SELECT(i))); \
c3 = _mm256_fmadd_ps(c3, a3, _mm256_permute_ps(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_MK_FMA_F32_AVX2_CYCLE(i) \
do { \
c0 = _mm256_add_ps(_mm256_mul_ps(a0, _mm256_permute_ps(b0, _FMA_SELECT(i))), c0); \
c1 = _mm256_add_ps(_mm256_mul_ps(a1, _mm256_permute_ps(b1, _FMA_SELECT(i))), c1); \
c2 = _mm256_add_ps(_mm256_mul_ps(a2, _mm256_permute_ps(b2, _FMA_SELECT(i))), c2); \
c3 = _mm256_add_ps(_mm256_mul_ps(a3, _mm256_permute_ps(b3, _FMA_SELECT(i))), c3); \
#define SN_MK_FMA_F64_AVX2_CYCLE(i) \
do { \
c0 = _mm512_fmadd_ps(c0, a0, _mm512_permutex_pd(b0, _FMA_SELECT(i))); \
c1 = _mm512_fmadd_ps(c1, a1, _mm512_permutex_pd(b1, _FMA_SELECT(i))); \
c2 = _mm512_fmadd_ps(c2, a2, _mm512_permutex_pd(b2, _FMA_SELECT(i))); \
c3 = _mm512_fmadd_ps(c3, a3, _mm512_permutex_pd(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_MK_FMA_F64_AVX2_CYCLE(i) \
do { \
c0 = _mm512_fmadd_ps(a0, _mm512_permutex_pd(b0, _FMA_SELECT(i)), c0); \
c1 = _mm512_fmadd_ps(a1, _mm512_permutex_pd(b1, _FMA_SELECT(i)), c1); \
c2 = _mm512_fmadd_ps(a2, _mm512_permutex_pd(b2, _FMA_SELECT(i)), c2); \
c3 = _mm512_fmadd_ps(a3, _mm512_permutex_pd(b3, _FMA_SELECT(i)), c3); \
#define SN_FMA_F32_512_PACK_COLUMN(col, target) \
do { \
register __m256 low = _mm512_castps512_ps256(col); \
register __m256 high = _mm512_extractf32x8_ps(col, 1); \
register __m256 half = _mm256_add_ps(high, low); \
register __m128 high128 = _mm256_extractf32x4_ps(half, 1); \
register __m128 low128 = _mm256_castps256_ps128(half); \
register __m128 half128 = _mm_add_ps(high128, low128); \
_mm_store_ps(target, half128); \
} while (0)
#define SN_FMA_F32_AVX2_PACK_COLUMN(col, target) \
do { \
register __m128 high128 = _mm256_extractf32x4_ps(col, 1); \
register __m128 low128 = _mm256_castps256_ps128(col); \
register __m128 half128 = _mm_add_ps(high128, low128); \
_mm_store_ps(target, half128); \
} while (0)
#define SN_FMA_F64_512_PACK_COLUMN(col, target) \
do { \
register __m128 high128 = _mm256_extractf32x4_ps(col, 1); \
register __m128 low128 = _mm256_castps256_ps128(col); \
register __m128 half128 = _mm_add_ps(high128, low128); \
_mm_store_ps(target, half128); \
} while (0)
// double