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

new compute kernel design, less mem read delay

Browse Source
This commit is contained in:
Nitload
2026-04-07 01:04:05 +08:00
parent d702b3224c
commit 12a216d404
47 changed files with 225 additions and 145 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/mk.fma.x86.c
+38 -23
View File
@@ -136,40 +136,55 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
_mm256_store_ps((float *)&ans_cur->f32_y[1], c23);
}
#define SN_MK_FMA_F32_512_PHASE(i)\
do { \
ak = _mm512_broadcast_f32x4(left_##i ->f32_x[0]);\
c_0 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i);\
ak = _mm512_broadcast_f32x4(left_##i ->f32_x[0]);\
c_1 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i);\
ak = _mm512_broadcast_f32x4(left_##i ->f32_x[0]);\
c_2 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i);\
ak = _mm512_broadcast_f32x4(left_##i ->f32_x[0]);\
c_3 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i);\
}while(0)
SN_MK_PROC_DECL_STD(float, 512, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, float);
SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y);
register __m512 c = _mm512_load_ps((float *)ans_cur);
register __m512 c_1 = c;
register __m512 c_2 = c;
register __m512 c_3 = c;
register __m512 c_0 = _mm512_load_ps((float *)ans_cur);
register __m512 c_1 = c_0;
register __m512 c_2 = c_0;
register __m512 c_3 = c_0;
for (size_t z = 0; z < common_z; z++) {
SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z);
SN_MAT_KERNEL_LOCAL const *left_cur = left + GET_INDEX(common_y, z, y);
for (size_t z = 0; z < common_z; z+=4) {
SN_MAT_KERNEL_LOCAL const *right = right + z;
SN_MAT_KERNEL_LOCAL const *left_0 = left + z + 0;
SN_MAT_KERNEL_LOCAL const *left_1 = left + z + 1;
SN_MAT_KERNEL_LOCAL const *left_2 = left + z + 2;
SN_MAT_KERNEL_LOCAL const *left_3 = left + z + 3;
// full right load：[kb_col0 | kb_col1 | kb_col2 | kb_col3]
register __m512 b = _mm512_load_ps((float *)right_cur);
register __m512 b_0 = _mm512_load_ps((float *)(right+0));
register __m512 b_1 = _mm512_load_ps((float *)(right+1));
register __m512 b_2 = _mm512_load_ps((float *)(right+2));
register __m512 b_3 = _mm512_load_ps((float *)(right+3));
// k=0: broadcast ka_col0 到4个lane，permute_ps 取 kb(0,j)
register __m512 ak;
ak = _mm512_broadcast_f32x4(left_cur->f32_x[0]);
c = _mm512_fmadd_ps(ak, _mm512_permute_ps(b, _FMA_SELECT(0)), c);
// k=1
ak = _mm512_broadcast_f32x4(left_cur->f32_x[1]);
c = _mm512_fmadd_ps(ak, _mm512_permute_ps(b, _FMA_SELECT(1)), c);
// k=2
ak = _mm512_broadcast_f32x4(left_cur->f32_x[2]);
c = _mm512_fmadd_ps(ak, _mm512_permute_ps(b, _FMA_SELECT(2)), c);
// k=3
ak = _mm512_broadcast_f32x4(left_cur->f32_x[3]);
c = _mm512_fmadd_ps(ak, _mm512_permute_ps(b, _FMA_SELECT(3)), c);
register __m512 ak_1;
register __m512 ak_2;
register __m512 ak_3;
SN_MK_FMA_F32_512_PHASE(0);
SN_MK_FMA_F32_512_PHASE(1);
SN_MK_FMA_F32_512_PHASE(2);
SN_MK_FMA_F32_512_PHASE(3);
}
_mm512_store_ps((float *)ans_cur, c);
c_0 = _mm512_add_ps(c_0, c_1);
c_2= _mm512_add_ps(c_2, c_3);
c_0 = _mm512_add_ps(c_0, c_2);
_mm512_store_ps((float *)ans_cur, c_0);
}
// double