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

seems tile_fma kernel has index error and indexer of matrix has error as well

Browse Source
This commit is contained in:
Youmu Liu
2026-01-11 16:45:38 +08:00
parent cd8210d9a1
commit b8bf1e9f08
8 changed files with 56 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.fma_tile.x86.c
+19 -11
View File
@@ -55,8 +55,9 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, sse, , fma)
for (size_t k = 0; k < commonK; k++) {
MAT_KERNEL(float)
const *right_cur = right + GET_INDEX(ansHeight, k, j), // GetKernelAddressUnsafe
const *left_cur = left + GET_INDEX(commonK, i, k); // GetKernelAddressUnsafe
const *right_cur = right + GET_INDEX(commonK, i, k), // GetKernelAddressUnsafe
const *left_cur =
left + GET_INDEX(ansHeight, k, j); // GetKernelAddressUnsafe
const __m128 a0 = _mm_load_ps((float *)&left_cur->f32_x[0]);
const __m128 a1 = _mm_load_ps((float *)&left_cur->f32_x[1]);
@@ -182,9 +183,9 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, avx, _fma, fma)
for (size_t k = 0; k < commonK; k++) {
MAT_KERNEL(float)
const *left_cur = left + GET_INDEX(commonK, i, k), // GetKernelAddressUnsafe
const *left_cur = left + GET_INDEX(ansHeight, k, j), // GetKernelAddressUnsafe
const *right_cur =
right + GET_INDEX(ansHeight, k, j); // GetKernelAddressUnsafe
right + GET_INDEX(commonK, i, k); // GetKernelAddressUnsafe
register __m256 a01 = _mm256_load_ps((float *)&left_cur->f32_y[0]);
register __m256 a23 = _mm256_load_ps((float *)&left_cur->f32_y[1]);
@@ -239,8 +240,13 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, 512, , fma)
register __m512i b_shift = _mm512_load_si512(&b_shift_source);
register __m512 c = _mm512_load_ps((float *)(ans + GET_INDEX(ansHeight, i, j)));
for (size_t k = 0; k < commonK; k++) {
register __m512 a = _mm512_load_ps((float *)left);
register __m512 b = _mm512_load_ps((float *)right);
MAT_KERNEL(double) *const left_cur = left + GET_INDEX(ansHeight, k, j),
const *right_cur =
right + GET_INDEX(commonK, i,
k); // GetKernelAddressUnsafe
register __m512 a = _mm512_load_ps((float *)left_cur);
register __m512 b = _mm512_load_ps((float *)right_cur);
b = _mm512_permutexvar_ps(b_shift, b);
register __m512 a_perm = _mm512_shuffle_f32x4(a, a, _MM_SHUFFLE(3, 2, 1, 0));
@@ -260,7 +266,7 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, 512, , fma)
c = _mm512_fmadd_ps(a_perm, b_perm, c);
}
_mm512_store_ps((float *)ans, c);
_mm512_store_ps((float *)(ans + GET_INDEX(ansHeight, i, j)), c);
}
// double
@@ -358,8 +364,9 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, avx, , fma)
for (size_t k = 0; k < commonK; k++) {
MAT_KERNEL(double)
const *right_cur = right + GET_INDEX(ansHeight, k, j), // GetKernelAddressUnsafe
const *left_cur = left + GET_INDEX(commonK, i, k); // GetKernelAddressUnsafe
const *right_cur = right + GET_INDEX(commonK, i, k), // GetKernelAddressUnsafe
const *left_cur =
left + GET_INDEX(ansHeight, k, j); // GetKernelAddressUnsafe
const __m256d a0 = _mm256_load_pd((double *)&left_cur->f64_y[0]);
const __m256d a1 = _mm256_load_pd((double *)&left_cur->f64_y[1]);
@@ -435,8 +442,9 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma)
for (size_t k = 0; k < commonK; k++) {
MAT_KERNEL(double)
const *right_cur = right + GET_INDEX(ansHeight, k, j), // GetKernelAddressUnsafe
const *left_cur = left + GET_INDEX(commonK, i, k); // GetKernelAddressUnsafe
const *right_cur = right + GET_INDEX(commonK, i, k), // GetKernelAddressUnsafe
const *left_cur =
left + GET_INDEX(ansHeight, k, j); // GetKernelAddressUnsafe
const __m256d a0 = _mm256_load_pd((double *)&left_cur->f64_y[0]);
const __m256d a1 = _mm256_load_pd((double *)&left_cur->f64_y[1]);