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

other gemm kernel use half-vector-machine style

Browse Source
This commit is contained in:
Youmu Liu
2026-06-08 16:18:03 +08:00
parent 1c515ea2ae
commit 9d5c0fd8b4
4 changed files with 129 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-template/sn_template.fma.x86.h
+24 -43
View File
@@ -2,55 +2,36 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
#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))); \
#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); \
} while (0)
#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))); \
#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); \
} while (0)
#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))); \
#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); \
} while (0)
#define SN_FMA_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_F64_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_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); \
#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); \
} while (0)
// double