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

fma code added

Browse Source
This commit is contained in:
Youmu Liu
2025-12-09 18:07:53 +08:00
parent 6acc8629b8
commit 7bcb460a39
7 changed files with 230 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+22 -17
View File
@@ -141,22 +141,26 @@ typedef void(SN_DECL *PNL_PROC_RAS)(mat_panel *right, mat_panel *ans, scalar_pac
#define PNL_PROC(T_type, T_arch, T_op_name) T_type##_pnl_##T_op_name##_##T_arch
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(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, \
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, 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, \
MAT_PANEL##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#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, \
MAT_PANEL##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(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, \
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, \
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, \
MAT_PANEL_##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#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, \
MAT_PANEL_##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#pragma endregion
@@ -219,7 +223,8 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill);
#pragma region fma
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
#pragma endregion
StgSharp.Native/sn-head/sn_matkernel.h
+10 -6
View File
@@ -67,15 +67,19 @@ typedef MAT_KERNEL(uint64_t) sn_zero_kernel_u64;
/* SSE panel layout: column-major buffer holding 4 logical columns with MINQ=4.
Columns: each column packs up to four scalars so ALIGN equals sizeof(__m128).
Alignment: every column boundary lands on a 16-byte multiple for XMM loads. */
#define MAT_PANEL_SSE(T) SN_ALIGNAS(16) MAT_PANEL_SQ_MIN_NOALIGN(T, __m128, 4)
#define MAT_PANEL_sse(T) SN_ALIGNAS(16) MAT_PANEL_SQ_MIN_NOALIGN(T, __m128, 4)
/* AVX2 panel layout: column-major buffer holding 8 logical columns with MINQ=8.
Columns: each column stores eight elements via two 256-bit chunks when T is 32-bit.
Alignment: column starts are 32-byte aligned so AVX2 loads/stores remain contiguous. */
#define MAT_PANEL_AVX(T) SN_ALIGNAS(32) MAT_PANEL_SQ_MIN_NOALIGN(T, __m256, 8)
#define MAT_PANEL_avx(T) SN_ALIGNAS(32) MAT_PANEL_SQ_MIN_NOALIGN(T, __m256, 8)
/* AVX2 panel layout: column-major buffer holding 8 logical columns with MINQ=8.
Columns: each column stores eight elements via two 256-bit chunks when T is 32-bit.
Alignment: column starts are 32-byte aligned so AVX2 loads/stores remain contiguous. */
#define MAT_PANEL_avx_fma(T) SN_ALIGNAS(32) MAT_PANEL_SQ_MIN_NOALIGN(T, __m256, 8)
/* AVX-512 panel layout: column-major buffer holding 8 or more columns depending on T.
Columns: each column expands to ceil(Q/lanes) ZMM chunks; default MINQ=8 keeps double-wide tiles.
Columns: each column expands to ceil(Q/lanes) ZMM chunks; default MINQ=8 keeps doublse-wide tiles.
Alignment: column starts sit on 64-byte boundaries to match AVX-512 streaming access. */
#define MAT_PANEL_512(T) SN_ALIGNAS(64) MAT_PANEL_SQ_MIN_NOALIGN(T, __m512, 8)
#define MAT_PANEL_512(T) SN_ALIGNAS(64) MAT_PANEL_SQ_MIN_NOALIGN(T, __m512, 8)
/* Global zero kernel declaration (defined in dllmain.c), aligned to 64 bytes */
#ifdef __cplusplus
@@ -113,7 +117,7 @@ extern __declspec(align(64)) sn_zero_kernel_u64 SN_ZERO_KERNEL;
int col_index, int row_index)
/* Declarations for AVX float variant */
DECLARE_BUILD_PANEL(AVX, float);
DECLARE_STORE_PANEL(AVX, float);
DECLARE_BUILD_PANEL(avx, float);
DECLARE_STORE_PANEL(avx, float);
#endif /* SN_MATKERNEL */