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

working on matrix mul matrix

Browse Source
This commit is contained in:
Youmu Liu
2025-12-10 17:52:17 +08:00
parent 7bcb460a39
commit 631bc2b6e8
8 changed files with 151 additions and 221 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_fma.c
+43
View File
@@ -52,6 +52,49 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
PANEL_COL_VEC((*ans), 3, 0) = c3;
}
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
for (int col = 0; col < 8; ++col) {
register __m256 bcol = _mm256_load_ps((float *)&right->vec[col][0]);
register __m256 ccol = _mm256_load_ps((float *)&ans->vec[col][0]);
register __m256i idx = _mm256_setzero_si256();
register __m256i increament = _mm256_load_si256(&idx0);
register __m256 a0 = _mm256_load_ps((float *)&left->vec[0][0]);
register __m256 a1 = _mm256_load_ps((float *)&left->vec[1][0]);
register __m256 a2 = _mm256_load_ps((float *)&left->vec[2][0]);
register __m256 a3 = _mm256_load_ps((float *)&left->vec[3][0]);
ccol = _mm256_fmadd_ps(a0, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a1, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a2, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a3, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
a0 = _mm256_load_ps((float *)&left->vec[4][0]);
a1 = _mm256_load_ps((float *)&left->vec[5][0]);
a2 = _mm256_load_ps((float *)&left->vec[6][0]);
a3 = _mm256_load_ps((float *)&left->vec[7][0]);
ccol = _mm256_fmadd_ps(a0, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a1, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a2, _mm256_permutevar8x32_ps(bcol, idx), ccol);
idx = _mm256_add_epi32(idx, increament);
ccol = _mm256_fmadd_ps(a3, _mm256_permutevar8x32_ps(bcol, idx), ccol);
_mm256_store_ps((float *)&ans->vec[col][0], ccol);
}
}
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma)