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

Mat mul mat is now on full function

Browse Source
This commit is contained in:
Youmu Liu
2025-12-16 16:15:30 +08:00
parent 2c7277f3f9
commit 88ed5a95c5
14 changed files with 357 additions and 488 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Additional.cs
+28 -5
View File
@@ -26,6 +26,7 @@
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
@@ -40,17 +41,39 @@ namespace StgSharp.Mathematics.Numeric
internal static unsafe void ComputeFmaPanelF32(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize;
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
MatrixKernel* rightBuffer = (MatrixKernel*)p->Right;
MatrixKernel* ansBuffer = (MatrixKernel*)p->Answer;
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
ref long offsetRef = ref Unsafe.As<int, long>(ref p->PrimTileOffset);
ref long lengthRef = ref Unsafe.As<int, long>(ref p->PrimCount);
int i_length = p->PrimCount;
int j_length = p->SecCount;
int k_length = 0;
for (long i = 0; i < lengthRef; i++)
int ansWidth = p-> AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
for (int k = 0; k < k_length; k++) { }
int i = (int)(c / blocksPerRow) * size;
int j = (int)(c % blocksPerRow) * size;
NativeIntrinsic.Intrinsic.f32_clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size)
{
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)leftPanel, leftBuffer, commonK,
ansHeight, k, i);
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)rightPanel, rightBuffer, ansWidth,
commonK, j, k);
NativeIntrinsic.Intrinsic
.f32_panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (MatrixPanel*)ansPanel);
}
NativeIntrinsic.Intrinsic
.f32_store_panel((MatrixPanel*)ansPanel, ansBuffer, ansWidth,
ansHeight, j, i);
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);