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

add simd id and cpu id

Browse Source
This commit is contained in:
Youmu Liu
2026-01-05 18:10:24 +08:00
parent d0fe4dde1a
commit 88fa653323
11 changed files with 447 additions and 286 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.fma.x86.c
+9 -8
View File
@@ -3,12 +3,13 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
#include "sn_intrinsic.h"
#include "sn_intrincontext.x86.h"
#define _FMA_SELECT(i) _MM_SHUFFLE(i, i, i, i)
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, , fma)
{
const __m128 a0 = PANEL_COL_VEC((*left), 0, 0);
const __m128 a1 = PANEL_COL_VEC((*left), 1, 0);
@@ -58,7 +59,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, _fma, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
@@ -101,7 +102,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, , fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
@@ -142,7 +143,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma)
}
}
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, , fma)
{
const __m512i ones = _mm512_set1_epi32(1);
@@ -208,7 +209,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, fma)
}
}
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, sse, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, sse, , fma)
{
for (int col = 0; col < 4; ++col) {
/* current answer column (two chunks: rows 0-1 and 2-3) */
@@ -260,7 +261,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, sse, fma)
}
}
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx, , fma)
{
register __m256d a0 = _mm256_load_pd((double *)&left->vec[0][0]);
register __m256d a1 = _mm256_load_pd((double *)&left->vec[1][0]);
@@ -301,7 +302,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx, fma)
_mm256_store_pd((double *)&ans->vec[3][0], c3);
}
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx_fma, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma)
{
register __m256d a0 = _mm256_load_pd((double *)&left->vec[0][0]);
register __m256d a1 = _mm256_load_pd((double *)&left->vec[1][0]);
@@ -344,7 +345,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx_fma, fma)
// ans = left * right + ans
// right is not transposed
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, 512, fma)
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, 512, , fma)
{
const __m512i ones = _mm512_set1_epi32(1);