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

Better L4 and array storage design

Browse Source
This commit is contained in:
Youmu Liu
2026-03-04 18:12:58 +08:00
parent 7e970352ab
commit dfa6f4ba92
103 changed files with 853 additions and 6837 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.fma.x86.c
+8 -24
View File
@@ -10,10 +10,8 @@
// float
SN_MK_PROC_DECL_STD(float, sse, , fma)
SN_SK_PROC_DECL_STD(float, sse, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
const __m128 a0 = _mm_load_ps((float *)&left->f32_x[0]);
const __m128 a1 = _mm_load_ps((float *)&left->f32_x[1]);
const __m128 a2 = _mm_load_ps((float *)&left->f32_x[2]);
@@ -60,10 +58,8 @@ SN_MK_PROC_DECL_STD(float, sse, , fma)
_mm_store_ps((float *)&ans->f32_x[3], c3);
}
SN_MK_PROC_DECL_STD(float, avx, , fma)
SN_SK_PROC_DECL_STD(float, avx, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
static const __m256i b01_shift = { .m256i_i32 = { 0, 1, 2, 3, 5, 6, 7, 4 } };
static const __m256i b23_shift = { .m256i_i32 = { 2, 3, 0, 1, 7, 4, 5, 6 } };
@@ -108,10 +104,8 @@ SN_MK_PROC_DECL_STD(float, avx, , fma)
ans->f32_y[1] = c23;
}
SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
SN_SK_PROC_DECL_STD(float, avx, _fma, fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
static const __m256i b01_shift = { .m256i_i32 = { 0, 1, 2, 3, 5, 6, 7, 4 } };
static const __m256i b23_shift = { .m256i_i32 = { 2, 3, 0, 1, 7, 4, 5, 6 } };
@@ -156,10 +150,8 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
ans->f32_y[1] = c23;
}
SN_MK_PROC_DECL_STD(float, 512, , fma)
SN_SK_PROC_DECL_STD(float, 512, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
static const __m512i b_shift = { .m512i_i32 = {
// shuff order
0, 1, 2, 3, // col0
@@ -194,10 +186,8 @@ SN_MK_PROC_DECL_STD(float, 512, , fma)
// double
SN_MK_PROC_DECL_STD(double, sse, , fma)
SN_SK_PROC_DECL_STD(double, sse, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
for (int col = 0; col < 4; ++col) {
/* current answer column (two chunks: rows 0-1 and 2-3) */
register __m128d c_lo = _mm_load_pd(ans->f64_x + (col * 2 + 0));
@@ -240,10 +230,8 @@ SN_MK_PROC_DECL_STD(double, sse, , fma)
}
}
SN_MK_PROC_DECL_STD(double, avx, , fma)
SN_SK_PROC_DECL_STD(double, avx, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
const __m256d a0 = _mm256_load_pd((double *)&left->f64_y[0]);
const __m256d a1 = _mm256_load_pd((double *)&left->f64_y[1]);
const __m256d a2 = _mm256_load_pd((double *)&left->f64_y[2]);
@@ -286,10 +274,8 @@ SN_MK_PROC_DECL_STD(double, avx, , fma)
_mm256_store_pd((double *)&ans->f64_y[3], ccol);
}
SN_MK_PROC_DECL_STD(double, avx, _fma, fma)
SN_SK_PROC_DECL_STD(double, avx, _fma, fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
const __m256d a0 = _mm256_load_pd((double *)&left->f64_y[0]);
const __m256d a1 = _mm256_load_pd((double *)&left->f64_y[1]);
const __m256d a2 = _mm256_load_pd((double *)&left->f64_y[2]);
@@ -332,10 +318,8 @@ SN_MK_PROC_DECL_STD(double, avx, _fma, fma)
_mm256_store_pd((double *)&ans->f64_y[3], ccol);
}
SN_MK_PROC_DECL_STD(double, 512, , fma)
SN_SK_PROC_DECL_STD(double, 512, , fma)
{
__mk_param_std(KERNEL, LEFT_RIGHT_ANS, float);
static const __m512i b01_shift = { .m512i_i64 = { 0, 1, 2, 3, 5, 6, 7, 4 } };
static const __m512i b23_shift = { .m512i_i64 = { 2, 3, 0, 1, 7, 4, 5, 6 } };