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

better kernel fma

Browse Source
This commit is contained in:
Youmu Liu
2026-01-07 18:06:36 +08:00
parent 3d488abd2e
commit 88af8d8957
11 changed files with 347 additions and 149 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Float.cs
+17 -6
View File
@@ -142,9 +142,13 @@ namespace StgSharp.Mathematics.Numeric
{
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
right.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
// int size = MatrixPanel.KernelSideCount<float>();
if (count <= 1024)
{
MatrixParallelThread[] pool = null;
@@ -155,15 +159,22 @@ namespace StgSharp.Mathematics.Numeric
{
for (int j = 0; j < ans.KernelColumnLength; j++)
{
MatrixKernel<float>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
MatrixKernel<float>* ansPanel = GetKernelAddressUnsafe(ans.Buffer,
ans.KernelColumnLength,
i, j);
// Console.WriteLine($"{i}{j}");
// NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < left.KernelColumnLength; k++)
{
NativeIntrinsic.Context
.mat[f32].kernel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
.mat[f32].kernel_fma((MatrixPanel*)GetKernelAddressUnsafe(left.Buffer,
left.KernelColumnLength,
i,
k),
(MatrixPanel*)GetKernelAddressUnsafe(right.Buffer,
right.KernelColumnLength,
k,
j),
(MatrixPanel*)ansPanel);
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.cs
+3 -6
View File
@@ -45,7 +45,6 @@ namespace StgSharp.Mathematics.Numeric
)
{
int f32 = (int)MatrixElementType.F32;
int size = MatrixPanel.KernelSideCount<float>();
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
MatrixKernel* rightBuffer = (MatrixKernel*)p->Right;
MatrixKernel* ansBuffer = (MatrixKernel*)p->Answer;
@@ -62,17 +61,15 @@ namespace StgSharp.Mathematics.Numeric
{
int i = (int)(c / ansWidth);
int j = (int)(c % ansWidth);
MatrixKernel<float>* ansPanel = GetKernelAddressUnsafe((MatrixKernel<float>*)ansBuffer,
ansHeight, i, j);
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size) {
for (int k = 0; k < commonK; k++) {
NativeIntrinsic.Context
.mat[f32].kernel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)ansPanel);
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)ansBuffer,
ansHeight, i, j));
}
}
}