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

redesign matrix design

Browse Source
This commit is contained in:
Youmu Liu
2025-12-23 18:13:05 +08:00
parent b32c1cb4eb
commit b60ce56f1c
12 changed files with 493 additions and 225 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+27 -46
View File
@@ -96,40 +96,6 @@ static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mis
_Static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size mismatch");
#endif
#pragma region matrix buffer operation function type
typedef void(SN_DECL *BUF_PROC_AS)(mat_kernel *ans, scalar_pack *scalar, int count);
typedef void(SN_DECL *BUF_PROC_RAS)(mat_kernel *right, mat_kernel *ans, scalar_pack *scalar,
int count);
typedef void(SN_DECL *BUF_PROC_LRAS)(mat_kernel *left, mat_kernel *right, mat_kernel *ans,
scalar_pack *scalar, int count);
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 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)( \
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)( \
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, \
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)( \
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)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, size_t count)
#pragma endregion
#pragma region matrix panel function type
typedef void(SN_DECL *PANEL_PROC)(mat_panel *panel, mat_kernel *kernel, int col_length,
@@ -176,6 +142,9 @@ typedef void(SN_DECL *KER_PROC_LAS)(mat_kernel *matrix, mat_kernel *ans, scalar_
#define KER_PROC(T_type, T_arch, T_op_name) T_type##_kernel_##T_op_name##_##T_arch
#define DECLARE_KER_PROC_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL KER_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar)
#define DECLARE_KER_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL KER_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
@@ -216,25 +185,25 @@ typedef uint64_t(SN_DECL *FACTORIALROC)(int n);
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_KER_PROC_LEFT_RIGHT_ANS(float, sse, add);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, add);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 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_KER_PROC_LEFT_RIGHT_ANS(float, sse, sub);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, sub);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 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_KER_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul);
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul);
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, 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_KER_PROC_ANS_SCALAR(float, sse, fill);
DECLARE_KER_PROC_ANS_SCALAR(float, avx, fill);
DECLARE_KER_PROC_ANS_SCALAR(float, 512, fill);
#pragma endregion
@@ -280,6 +249,18 @@ INTERNAL int SN_DECL index_pair_sse(short const *str, uint32_t target, int lengt
#pragma endregion
typedef struct sn_mat_enumerator {
int CurrentColumn;
int CurrentRow;
int KernelColumnLength;
int KernelRowLength;
} sn_mat_enumerator;
typedef struct sn_mat_boarder {
int (*SN_DECL is_in_buffer)(int col, int row, sn_mat_enumerator *enumerator);
long (*SN_DECL get_index_offset)(int col, int row, sn_mat_enumerator *wnumerator);
} sn_mat_boarder;
// Mirrors C# IntrinsicContext field order so managed/unmanaged layouts stay in sync.
typedef struct sn_intrinsic {
HASH city_hash_simplify;