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

macro gymnastics to implement storage kernel, buffer op and complex compute panel

Browse Source
This commit is contained in:
Youmu Liu
2025-11-13 19:09:27 +08:00
parent 2dfb4ae86a
commit 1417785726
18 changed files with 373 additions and 383 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/load_intrinsic.c
+15 -15
View File
@@ -65,11 +65,11 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->f32_add = (MAT_PROC_LRA)f32_add_sse;
context->f32_fma = (MAT_PROC_LRA)f32_fma_sse;
context->f32_scalar_mul = f32_scalar_mul_sse;
context->f32_sub = (MAT_PROC_LRA)f32_sub_sse;
context->f32_transpose = (MAT_PROC_LA)f32_transpose_sse;
context->f32_kernel_add = KER_PROC(float, sse, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, sse, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, sse, sub);
context->f32_kernel_transpose = KER_PROC(float, sse, transpose);
context->f32_normalize_3 = f32_normalize;
}
@@ -78,11 +78,11 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->f32_add = (MAT_PROC_LRA)f32_add_avx;
context->f32_fma = (MAT_PROC_LRA)f32_fma_avx;
context->f32_scalar_mul = f32_scalar_mul_avx;
context->f32_sub = (MAT_PROC_LRA)f32_sub_avx;
context->f32_transpose = (MAT_PROC_LA)f32_transpose_avx;
context->f32_kernel_add = KER_PROC(float, avx, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, avx, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, avx, sub);
context->f32_kernel_transpose = KER_PROC(float, avx, transpose);
context->f32_normalize_3 = f32_normalize;
}
@@ -91,10 +91,10 @@ void load_intrinsic_512(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->f32_add = (MAT_PROC_LRA)f32_add_512;
context->f32_fma = (MAT_PROC_LRA)f32_fma_512;
context->f32_scalar_mul = f32_scalar_mul_512;
context->f32_sub = (MAT_PROC_LRA)f32_sub_512;
context->f32_transpose = (MAT_PROC_LA)f32_transpose_512;
context->f32_kernel_add = KER_PROC(float, 512, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, 512, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, 512, sub);
context->f32_kernel_transpose = KER_PROC(float, 512, transpose);
context->f32_normalize_3 = f32_normalize;
}