better kernel fma

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
@@ -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));
}
}
}