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

add double compute kernel

Browse Source
This commit is contained in:
Youmu Liu
2025-12-26 17:16:12 +08:00
parent e5539bcc6d
commit c6d53d97ea
16 changed files with 821 additions and 300 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+45 -12
View File
@@ -215,10 +215,18 @@ 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(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_KER_PROC_RIGHT_ANS(float, sse, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, avx, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, 512, transpose);
@@ -227,10 +235,18 @@ 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_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);
#pragma endregion
#pragma region fma
@@ -239,6 +255,10 @@ 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_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);
@@ -278,20 +298,33 @@ INTERNAL int SN_DECL index_pair_sse(short const *str, uint32_t target, int lengt
// Mirrors C# IntrinsicContext field order so managed/unmanaged layouts stay in sync.
typedef struct sn_intrinsic {
HASH city_hash_simplify;
BUF_PROC_LRA f32_buffer_add;
BUF_PROC_AS f32_buffer_fill;
BUF_PROC_RAS f32_buffer_scalar_mul;
BUF_PROC_LRA f32_buffer_sub;
KER_PROC_RA f32_buffer_transpose;
PANEL_PROC f32_build_panel;
PANEL_PROC f32_clear_panel;
VECTORNORMALIZEPROC f32_normalize_3;
PNL_PROC_LRA f32_panel_fma;
PIVOTPROC f32_pivot;
PANEL_PROC f32_store_panel;
FACTORIALROC factorial_simd;
INDEX_PAIR index_pair;
BUF_PROC_RAS variant;
BUF_PROC_LRA mat_f32_buffer_add;
BUF_PROC_AS mat_f32_buffer_fill;
BUF_PROC_RAS mat_f32_buffer_scalar_mul;
BUF_PROC_LRA mat_f32_buffer_sub;
KER_PROC_RA mat_f32_buffer_transpose;
PANEL_PROC mat_f32_build_panel;
PANEL_PROC mat_f32_clear_panel;
PNL_PROC_LRA mat_f32_panel_fma;
PIVOTPROC mat_f32_pivot;
PANEL_PROC mat_f32_store_panel;
// f64 slots (currently unimplemented; filled with NULL in loader)
BUF_PROC_LRA mat_f64_buffer_add;
BUF_PROC_AS mat_f64_buffer_fill;
BUF_PROC_RAS mat_f64_buffer_scalar_mul;
BUF_PROC_LRA mat_f64_buffer_sub;
KER_PROC_RA mat_f64_buffer_transpose;
PANEL_PROC mat_f64_build_panel;
PANEL_PROC mat_f64_clear_panel;
PNL_PROC_LRA mat_f64_panel_fma;
PIVOTPROC mat_f64_pivot;
PANEL_PROC mat_f64_store_panel;
VECTORNORMALIZEPROC vec_f32_normalize_3;
} sn_intrinsic;
typedef enum most_advanced_instruction {