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

small changed

Browse Source
This commit is contained in:
Youmu Liu
2026-04-07 18:08:34 +08:00
parent 7bf2104edb
commit 94a3127c0d
40 changed files with 113 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/mk.fma.x86.c
+6 -49
View File
@@ -2,7 +2,8 @@
/* clang-format off */
#if SN_IS_ARCH(SN_ARCH_X86_64)
#include "sn_intrinsic.h"
// #include "sn_intrinsic.h"
#include "sn_template.fma.x86.h"
#include "sn_intrinsic.std.h"
#include "sn_intrinsic.context.x86.h"
@@ -136,17 +137,6 @@ 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)
{
@@ -187,39 +177,6 @@ SN_MK_PROC_DECL_STD(float, 512, , fma)
_mm512_store_ps((float *)ans_cur, c_0);
}
// double
#define DOUBLE_KERNEL_FMA_CYCLE(i) \
do { \
/* k = 0 */ \
register __m128d b_lo = _mm_load_pd(right_values + (2 * i + 0)); \
register __m128d b_hi = _mm_load_pd(right_values + (2 * i + 1)); \
register __m128d a_lo = _mm_load_pd(left_values + (0 * 2 + 0)); \
register __m128d a_hi = _mm_load_pd(left_values + (0 * 2 + 1)); \
register __m128d b = _mm_shuffle_pd(b_lo, b_lo, 0x0); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* group = 1 */ \
a_lo = _mm_load_pd(left_values + (1 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (1 * 2 + 1)); \
b = _mm_shuffle_pd(b_lo, b_lo, 0x3); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* groupp = 2 */ \
a_lo = _mm_load_pd(left_values + (2 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (2 * 2 + 1)); \
b = _mm_shuffle_pd(b_hi, b_hi, 0x0); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* froup = 3 */ \
a_lo = _mm_load_pd(left_values + (3 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (3 * 2 + 1)); \
b = _mm_shuffle_pd(b_hi, b_hi, 0x3); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
} while (0)
SN_MK_PROC_DECL_STD(double, sse, , fma)
{
@@ -243,13 +200,13 @@ SN_MK_PROC_DECL_STD(double, sse, , fma)
const double *right_values = &right_cur->m[0][0];
const double *left_values = &left_cur->m[0][0];
DOUBLE_KERNEL_FMA_CYCLE(0);
SN_MK_FMA_F64_SSE_PHASE(0);
DOUBLE_KERNEL_FMA_CYCLE(1);
SN_MK_FMA_F64_SSE_PHASE(1);
DOUBLE_KERNEL_FMA_CYCLE(2);
SN_MK_FMA_F64_SSE_PHASE(2);
DOUBLE_KERNEL_FMA_CYCLE(3);
SN_MK_FMA_F64_SSE_PHASE(3);
}
ans_cur->f64_x[0] = c0_lo;
StgSharp.Native/sn-src/thread.windows.c
+3 -3
View File
@@ -48,7 +48,7 @@ SN_API int SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_n
return 1;
}
SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *internal, uint64_t *global)
SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global)
{
ULONGLONG mask = 0;
if (!GetNumaNodeProcessorMask((UCHAR)numa_id, &mask))
@@ -56,9 +56,9 @@ SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *internal, uint64_t *g
int intra = 0;
while (mask) {
unsigned long idx;
uint32_t idx;
_BitScanForward64(&idx, mask);
internal[intra] = intra;
int _ = (internal == NULL) ? intra : (internal[intra] = intra);
global[intra] = (int)idx;
intra++;
mask &= mask - 1;