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.Common/Internal/Intrinsic/IntrinsicContext.cs
+21 -12
View File
@@ -39,20 +39,29 @@ namespace StgSharp.Internal.Intrinsic
{
public delegate* unmanaged[Cdecl]<char*, int, int> city_hash_simplify; //city_hash_simplify
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> f32_buffer_add; //f32_buffer_add
public delegate* unmanaged[Cdecl]<void*, ScalarPacket*,long, void> f32_buffer_fill; //f32_buffer_fill
public delegate* unmanaged[Cdecl]<void*, void*, ScalarPacket*, long, void> f32_buffer_scalar_mul; //f32_buffer_s_m
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> f32_buffer_sub; //f32_kernel_sub
public delegate* unmanaged[Cdecl]<void*, void*, void> f32_buffer_transpose; //f32_kernel_transpose
public delegate* unmanaged[Cdecl]<MatrixPanel*, MatrixKernel*, int, int, int, int, void> f32_build_panel; //f32_build_panel
public delegate* unmanaged[Cdecl]<MatrixPanel*, void> f32_clear_panel;
public delegate* unmanaged[Cdecl]<Vector4*, Vector4*, void> f32_normalize_3; //f32_normalize_3
public delegate* unmanaged[Cdecl]<void*, void*, void*, void> f32_panel_fma; //f32_panel_fma
public delegate* unmanaged[Cdecl]<void*, ulong, ulong, ulong, void> f32_pivot; //f32_panel_mul
public delegate* unmanaged[Cdecl]<void*, void*, int, int, int, int, void> f32_store_panel; //f32_store_panel
public delegate* unmanaged[Cdecl]<int, ulong> factorial_simd; //factorial_simd
public delegate* unmanaged[Cdecl]<char*, int, int, int> index_pair; //index_pair
public delegate* unmanaged[Cdecl]<void*, void*, ScalarPacket*, long, void> variant ; //variant
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> mat_f32_buffer_add; //f32_buffer_add
public delegate* unmanaged[Cdecl]<void*, ScalarPacket*, long, void> mat_f32_buffer_fill; //f32_buffer_fill
public delegate* unmanaged[Cdecl]<void*, void*, ScalarPacket*, long, void> mat_f32_buffer_scalar_mul; //f32_buffer_s_m
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> mat_f32_buffer_sub; //f32_kernel_sub
public delegate* unmanaged[Cdecl]<void*, void*, void> mat_f32_buffer_transpose; //f32_kernel_transpose
public delegate* unmanaged[Cdecl]<MatrixPanel*, MatrixKernel*, int, int, int, int, void> mat_f32_build_panel; //f32_build_panel
public delegate* unmanaged[Cdecl]<MatrixPanel*, void> mat_f32_clear_panel; //f32_clear_panel
public delegate* unmanaged[Cdecl]<void*, void*, void*, void> mat_f32_panel_fma; //f32_panel_fma
public delegate* unmanaged[Cdecl]<void*, ulong, ulong, ulong, void> mat_f32_pivot; //f32_panel_mul
public delegate* unmanaged[Cdecl]<void*, void*, int, int, int, int, void> mat_f32_store_panel; //f32_store_panel
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> mat_f64_buffer_add; //f32_buffer_add
public delegate* unmanaged[Cdecl]<void*, ScalarPacket*, long, void> mat_f64_buffer_fill; //f32_buffer_fill
public delegate* unmanaged[Cdecl]<void*, void*, ScalarPacket*, long, void> mat_f64_buffer_scalar_mul; //f32_buffer_s_m
public delegate* unmanaged[Cdecl]<void*, void*, void*, long, void> mat_f64_buffer_sub; //f32_kernel_sub
public delegate* unmanaged[Cdecl]<void*, void*, void> mat_f64_buffer_transpose; //f32_kernel_transpose
public delegate* unmanaged[Cdecl]<MatrixPanel*, MatrixKernel*, int, int, int, int, void> mat_f64_build_panel; //f32_build_panel
public delegate* unmanaged[Cdecl]<MatrixPanel*, void> mat_f64_clear_panel; //f32_clear_panel
public delegate* unmanaged[Cdecl]<void*, void*, void*, void> mat_f64_panel_fma; //f32_panel_fma
public delegate* unmanaged[Cdecl]<void*, ulong, ulong, ulong, void> mat_f64_pivot; //f32_panel_mul
public delegate* unmanaged[Cdecl]<void*, void*, int, int, int, int, void> mat_f64_store_panel; //f32_store_panel
public delegate* unmanaged[Cdecl]<Vector4*, Vector4*, void> vec_f32_normalize_3; //f32_normalize_3
public IntrinsicContext()
{