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-head/sn_intrinsic.h
+64 -62
View File
@@ -94,7 +94,7 @@ typedef struct matrix_task matrix_task;
#if defined(__cplusplus)
static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mismatch");
#elif defined(__STDC_VERSION__)
static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size mismatch");
static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size mismatch");
#endif
#pragma region matrix buffer operation function type
@@ -108,24 +108,25 @@ typedef void(SN_DECL *BUF_PROC_RA)(mat_kernel *right, mat_kernel *ans, int count
typedef void(SN_DECL *BUF_PROC_LRA)(mat_kernel *left, mat_kernel *right, mat_kernel *ans,
int count);
#define BUF_PROC(T_type, T_arch, T_op_name) T_type##_buf_##T_op_name##_##T_arch
#define BUF_PROC(T_type, T_arch, T_feature, T_op_name) \
T_type##_buf_##T_op_name##_##T_arch##T_feature
#define DECLARE_BUF_PROC_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
#define DECLARE_BUF_PROC_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar, size_t count)
#define DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
#define DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, \
scalar_pack const *scalar, int count)
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
scalar_pack const *scalar, size_t count)
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar, size_t count)
#define DECLARE_BUF_PROC_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
#define DECLARE_BUF_PROC_RIGHT_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, size_t count)
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, size_t count)
@@ -141,19 +142,20 @@ typedef void(SN_DECL *PNL_PROC_LRAS)(mat_panel *left, mat_panel *right, mat_pane
scalar_pack *scalar);
typedef void(SN_DECL *PNL_PROC_RAS)(mat_panel *right, mat_panel *ans, scalar_pack *scalar);
#define PNL_PROC(T_type, T_arch, T_op_name) T_type##_pnl_##T_op_name##_##T_arch
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
#define PNL_PROC(T_type, T_arch, T_feature, T_op_name) \
T_type##_pnl_##T_op_name##_##T_arch##T_feature
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_feature, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) *restrict ans)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_feature, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
MAT_PANEL_##T_arch##(T_type) *restrict ans)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_op_name)( \
MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_op_name)( \
MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
MAT_PANEL_##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#pragma endregion
@@ -288,7 +290,7 @@ typedef enum most_advanced_instruction {
// MainLevel (ordinal, shared bits 8..15; not bit-flags)
#define SIMDID_MAIN_LVL_NONE 0x00u
#define SIMDID_MAIN_LVL_SSE 0x01u /* SSE1..SSE4.2 */
#define SIMDID_MAIN_LVL_SSE 0x01u /* SSE1..SSE4.2 */
#define SIMDID_MAIN_LVL_AVX2 0x02u
#define SIMDID_MAIN_LVL_AVX2_FMA 0x03u
#define SIMDID_MAIN_LVL_AVX2_FMA_F16C 0x04u /* 16-bit scalar/vector ready */
@@ -310,62 +312,62 @@ typedef enum most_advanced_instruction {
INTERNAL void f32_normalize(VEC(float) * source, VEC(float) * target);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, , add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, , sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, , sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, , sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, , sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, , sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, , sub);
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, avx, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, 512, transpose);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, , scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, , scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, , scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, sse, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, avx, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, 512, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, sse, , scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, avx, , scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, 512, , scalar_mul);
DECLARE_BUF_PROC_ANS_SCALAR(float, sse, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, sse, , fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, , fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, , fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, sse, fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, avx, fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, 512, fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, sse, , fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, avx, , fill);
DECLARE_BUF_PROC_ANS_SCALAR(double, 512, , fill);
#pragma endregion
#pragma region fma
DECLARE_PNL_PROC_ANS(float, sse, clear);
DECLARE_PNL_PROC_ANS(float, avx, clear);
DECLARE_PNL_PROC_ANS(float, 512, clear);
DECLARE_PNL_PROC_ANS(float, sse, , clear);
DECLARE_PNL_PROC_ANS(float, avx, , clear);
DECLARE_PNL_PROC_ANS(float, 512, , clear);
DECLARE_PNL_PROC_ANS(double, sse, clear);
DECLARE_PNL_PROC_ANS(double, avx, clear);
DECLARE_PNL_PROC_ANS(double, 512, clear);
DECLARE_PNL_PROC_ANS(double, sse, , clear);
DECLARE_PNL_PROC_ANS(double, avx, , clear);
DECLARE_PNL_PROC_ANS(double, 512, , clear);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, , fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, , fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, _fma, fma);
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, avx, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, avx_fma, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, 512, fma);
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, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(double, 512, , fma);
#pragma endregion