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
2026-01-06 18:04:17 +08:00
parent 88fa653323
commit 3d488abd2e
17 changed files with 372 additions and 615 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/load_intrinsic.x86.c
+27 -19
View File
@@ -13,10 +13,9 @@
#include <cpuid.h>
#endif // _MSC_VER
PRIVATE void load_intrinsic_sse(sn_intrinsic *context);
PRIVATE void load_intrinsic_avx(sn_intrinsic *context);
PRIVATE void load_intrinsic_avx_fma(sn_intrinsic *context);
PRIVATE void load_intrinsic_512(sn_intrinsic *context, SIMDID simdid);
PRIVATE void load_intrinsic_sse(sn_intrinsic *context, SIMDID id);
PRIVATE void load_intrinsic_avx(sn_intrinsic *context, SIMDID id);
PRIVATE void load_intrinsic_512(sn_intrinsic *context, SIMDID id);
SN_API SIMDID load_intrinsic_function(void *intrinsic_context)
{
@@ -26,18 +25,21 @@ SN_API SIMDID load_intrinsic_function(void *intrinsic_context)
sn_intrinsic *context = (sn_intrinsic *)intrinsic_context;
if (check_isa(AVX512)) {
switch (simdid.make_byte[1]) {
case SIMDID_MAIN_LVL_SSE:
load_intrinsic_sse(context, simdid);
break;
case SIMDID_MAIN_LVL_AVX2:
load_intrinsic_avx(context, simdid);
break;
case SIMDID_MAIN_LVL_AVX512:
load_intrinsic_512(context, simdid);
} else if (check_isa(AVX256)) {
load_intrinsic_avx(context);
} else if (check_isa(SSE)) {
load_intrinsic_sse(context);
break;
}
return simdid;
}
void load_intrinsic_sse(sn_intrinsic *context)
void load_intrinsic_sse(sn_intrinsic *context, SIMDID id)
{
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
@@ -47,10 +49,10 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->mat[F32].buffer_fill = BUF_PROC(float, sse, , fill);
context->mat[F32].buffer_scalar_mul = BUF_PROC(float, sse, , scalar_mul);
context->mat[F32].buffer_sub = BUF_PROC(float, sse, , sub);
context->mat[F32].buffer_transpose = KER_PROC(float, sse, transpose);
context->mat[F32].buffer_transpose = KER_PROC(float, sse, , transpose);
context->mat[F32].build_panel = NULL;
context->mat[F32].clear_panel = PNL_PROC(float, sse, , clear);
context->mat[F32].panel_fma = PNL_PROC(float, sse, , fma);
context->mat[F32].panel_fma = KER_PROC(float, sse, , fma);
context->mat[F32].pivot = pivot_float;
context->mat[F32].store_panel = NULL;
@@ -68,8 +70,12 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->vec_f32_normalize_3 = f32_normalize;
}
void load_intrinsic_avx(sn_intrinsic *context)
void load_intrinsic_avx(sn_intrinsic *context, SIMDID id)
{
char avx_seg = id.make_byte[5];
char is_fma = (avx_seg & 1 << 3) != 0;
char is_f16c = (avx_seg & 1 << 4) != 0;
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
@@ -78,10 +84,11 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->mat[F32].buffer_fill = BUF_PROC(float, avx, , fill);
context->mat[F32].buffer_scalar_mul = BUF_PROC(float, avx, , scalar_mul);
context->mat[F32].buffer_sub = BUF_PROC(float, avx, , sub);
context->mat[F32].buffer_transpose = KER_PROC(float, avx, transpose);
context->mat[F32].buffer_transpose = KER_PROC(float, avx, , transpose);
context->mat[F32].build_panel = BUILD_PANEL(avx, float);
context->mat[F32].clear_panel = PNL_PROC(float, avx, , clear);
context->mat[F32].panel_fma = PNL_PROC(float, avx, , fma);
context->mat[F32].panel_fma = is_fma ? KER_PROC(float, avx, , fma) :
KER_PROC(float, avx, _fma, fma);
context->mat[F32].pivot = pivot_float;
context->mat[F32].store_panel = STORE_PANEL(avx, float);
@@ -92,7 +99,8 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->mat[F64].buffer_transpose = NULL;
context->mat[F64].build_panel = NULL;
context->mat[F64].clear_panel = PNL_PROC(double, avx, , clear);
context->mat[F64].panel_fma = PNL_PROC(double, avx, , fma);
context->mat[F64].panel_fma = is_fma ? PNL_PROC(double, avx, , fma) :
PNL_PROC(double, avx, _fma, fma);
context->mat[F64].pivot = NULL;
context->mat[F64].store_panel = NULL;
@@ -109,10 +117,10 @@ void load_intrinsic_512(sn_intrinsic *context, SIMDID id)
context->mat[F32].buffer_fill = BUF_PROC(float, 512, , fill);
context->mat[F32].buffer_scalar_mul = BUF_PROC(float, 512, , scalar_mul);
context->mat[F32].buffer_sub = BUF_PROC(float, 512, , sub);
context->mat[F32].buffer_transpose = KER_PROC(float, 512, transpose);
context->mat[F32].buffer_transpose = KER_PROC(float, 512, , transpose);
context->mat[F32].build_panel = BUILD_PANEL(512, float);
context->mat[F32].clear_panel = PNL_PROC(float, 512, , clear);
context->mat[F32].panel_fma = PNL_PROC(float, 512, , fma);
context->mat[F32].panel_fma = KER_PROC(float, 512, , fma);
context->mat[F32].pivot = pivot_float;
context->mat[F32].store_panel = STORE_PANEL(512, float);