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

new compute kernel design, less mem read delay

Browse Source
This commit is contained in:
Nitload
2026-04-07 01:04:05 +08:00
parent d702b3224c
commit 12a216d404
47 changed files with 225 additions and 145 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+3 -9
View File
@@ -39,11 +39,6 @@ typedef union uint64_u {
uint32_t u32[2];
} uint64_u;
/* Opaque placeholder used by generic single-kernel dispatch.
* It preserves pointer semantics without pretending every branch shares the
* same concrete MAT_KERNEL(T) type. */
typedef struct sn_mat_kernel_opaque sn_mat_kernel_opaque;
typedef struct mat_kernel mat_kernel;
typedef struct mat_panel mat_panel;
@@ -76,10 +71,9 @@ static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mis
static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mismatch");
#endif
typedef void(SN_DECL *UNI_MK_PROC)(matrix_task const *task);
typedef void(SN_DECL *UNI_SK_PROC)(sn_mat_kernel_opaque const *left,
sn_mat_kernel_opaque const *right,
sn_mat_kernel_opaque *restrict ans, uint64_t scalar);
typedef void(SN_DECL *UNI_MK_PROC)(matrix_task *task);
typedef void(SN_DECL *UNI_SK_PROC)(mat_kernel const *left, mat_kernel const *right,
mat_kernel *restrict ans, uint64_t scalar);
typedef void(SN_DECL *VECTORNORMALIZEPROC)(sn_vec_f32 *source, sn_vec_f32 *target);
typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans);