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

arrange of proj structure

Browse Source
This commit is contained in:
Youmu Liu
2026-02-27 17:09:13 +08:00
parent 15434ed053
commit a798ab07f1
173 changed files with 32304 additions and 33955 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+28 -60
View File
@@ -59,41 +59,22 @@ typedef union scalar_pack {
#define LEFT_RIGHT_ANS_SCALAR_PARAM 4
#define ANS_SCALAR_PARAM 5
typedef struct matrix_op_mode {
int16_t operation_style; /* PANEL_OP, BUFFER_OP, KERNEL_OP */
int16_t parameter_style; /* RIGHT_ANS_PARAM, LEFT_RIGHT_ANS_PARAM, e.g. */
} matrix_op_mode;
/* Mirrors StgSharp.Mathematics.Numeric.MatrixParallelTaskPackage<T> (128 bytes). */
#define MAT_TASK(T) \
struct { \
MAT_KERNEL(T) * left; /* Offset 0: left matrix kernel base */ \
MAT_KERNEL(T) * right; /* Offset 8: right matrix kernel base */ \
MAT_KERNEL(T) * result; /* Offset 16: result matrix kernel base */ \
int32_t element_size; /* Offset 24: sizeof(T) */ \
int32_t reserved_ptr0; /* Offset 28: reserved/extension slot */ \
int32_t left_prim_offset; /* Offset 32 */ \
int32_t left_prim_stride; /* Offset 36 */ \
int32_t left_sec_offset; /* Offset 40 */ \
int32_t left_sec_stride; /* Offset 44 */ \
int32_t right_prim_offset; /* Offset 48 */ \
int32_t right_prim_stride; /* Offset 52 */ \
int32_t right_sec_offset; /* Offset 56 */ \
int32_t right_sec_stride; /* Offset 60 */ \
int32_t result_prim_offset; /* Offset 64 */ \
int32_t result_prim_stride; /* Offset 68 */ \
int32_t result_sec_offset; /* Offset 72 */ \
int32_t result_sec_stride; /* Offset 76 */ \
scalar_pack *scalar; /* Offset 80 */ \
void *compute_handle; /* Offset 88 */ \
int32_t prim_count; /* Offset 96 */ \
matrix_op_mode compute_mode; /* Offset 100 */ \
int32_t sec_count; /* Offset 104 */ \
int32_t reserved_int0; /* Offset 108 */ \
int32_t prim_tile_offset; /* Offset 112 */ \
int32_t sec_tile_offset; /* Offset 116 */ \
int32_t prim_cols_in_tile; /* Offset 120 */ \
int32_t sec_cols_in_tile; /* Offset 124 */ \
#define MAT_TASK(T) \
struct { \
MAT_KERNEL(T) * mat_0; /* Offset 0: 1st matrix kernel base */ \
MAT_KERNEL(T) * mat_1; /* Offset 8: 2nd matrix kernel base */ \
MAT_KERNEL(T) * mat_2; /* Offset 16: 3rd matrix kernel base */ \
void *scalar; /* Offset 24: scalar pack */ \
int32_t i; /* Offset 32: row index of the result tile */ \
int32_t j; /* Offset 36: column index of the result tile */ \
int32_t k; /* Offset 40: column index of the result tile */ \
int32_t common_i; /* Offset 44: row index of the result tile */ \
int32_t common_j; /* Offset 48: column index of the result tile */ \
int32_t common_k; /* Offset 52: column index of the result tile */ \
uint32_t mask; /* Offset 56: mask for task schedualing, not useful in C side */ \
uint32_t reverse; /* Offset 60: reserved for future use */ \
scalar_pack padding; /* Offset 64: padding to make the struct size 128 bytes */ \
}
typedef struct matrix_task matrix_task;
@@ -103,6 +84,9 @@ 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
typedef void(SN_DECL *UNIFIKERNELPROC)(matrix_task *task);
#define UNIFIKERNRLPROC(T_arch, T_feature, T_op) mat_proc_##T_op##_##T_arch##_##T_feature
#pragma region matrix buffer operation function type
typedef void(SN_DECL *BUF_PROC_AS)(mat_kernel *ans, scalar_pack *scalar, int count);
@@ -247,17 +231,15 @@ typedef int(SN_DECL *INDEX_PAIR)(short const *stram, uint32_t target, int length
#pragma endregion
typedef struct mat_intrinsic {
BUF_PROC_LRA buffer_add;
BUF_PROC_AS buffer_fill;
BUF_PROC_RAS buffer_scalar_mul;
BUF_PROC_LRA buffer_sub;
KER_PROC_RA buffer_transpose;
PANEL_PROC build_panel;
PANEL_PROC clear_panel;
PNL_PROC_LRA ker_fma;
KERTILE_PROC_LRA ker_tile_fma;
PIVOTPROC pivot;
PANEL_PROC store_panel;
UNIFIKERNELPROC buffer_add;
UNIFIKERNELPROC buffer_fill;
UNIFIKERNELPROC buffer_scalar_mul;
UNIFIKERNELPROC buffer_sub;
UNIFIKERNELPROC buffer_transpose;
UNIFIKERNELPROC clear_panel;
UNIFIKERNELPROC ker_fma;
UNIFIKERNELPROC ker_tile_fma;
UNIFIKERNELPROC pivot;
} mat_intrinsic;
@@ -270,21 +252,7 @@ typedef struct sn_intrinsic {
mat_intrinsic mat[4];
} sn_intrinsic;
typedef enum elment_type { F32 = 0, F64 = 1, I32 = 2, I64 = 3 } element_type;
typedef enum most_advanced_instruction {
#ifdef _MSC_VER
SSE = __IA_SUPPORT_SSE42,
AVX256 = __IA_SUPPORT_VECTOR256,
AVX512 = __IA_SUPPORT_VECTOR512,
NEON
#else
SSE,
AVX256,
AVX512,
NEON
#endif
} most_advanced_instruction;
typedef enum elment_mask { F32 = 0, F64 = 1, I32 = 2, I64 = 3 } element_mask;
#include "sn_target.h"