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.Mathematics/Numeric/MatrixCompute.Additional.cs
+65 -25
View File
@@ -30,6 +30,7 @@ using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Numerics;
using System.Runtime.CompilerServices;
using System.Text;
using System.Threading.Tasks;
@@ -39,45 +40,84 @@ namespace StgSharp.Mathematics.Numeric
public partial class MatrixCompute
{
internal static unsafe void ComputeFmaPanelF32(MatrixParallelTaskPackage* p)
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 ansWidth = p-> AnsPrimOffset;
int ansWidth = p->AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
if (NativeIntrinsic.Context.mat_f32_build_panel == null)
{
int i = (int)(c / blocksPerRow) * size;
int j = (int)(c % blocksPerRow) * size;
NativeIntrinsic.Context.f32_clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size)
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
NativeIntrinsic.Context
.f32_build_panel((MatrixPanel*)leftPanel, leftBuffer, commonK,
ansHeight, k, i);
NativeIntrinsic.Context
.f32_build_panel((MatrixPanel*)rightPanel, rightBuffer, ansWidth,
commonK, j, k);
NativeIntrinsic.Context
.f32_panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (MatrixPanel*)ansPanel);
int i = (int)(c / blocksPerRow);
int j = (int)(c % blocksPerRow);
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) {
NativeIntrinsic.Context
.mat_f32_panel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)ansPanel);
}
}
NativeIntrinsic.Context
.f32_store_panel((MatrixPanel*)ansPanel, ansBuffer, ansWidth,
ansHeight, j, i);
} else
{
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
int i = (int)(c / blocksPerRow) * size;
int j = (int)(c % blocksPerRow) * size;
NativeIntrinsic.Context.mat_f32_clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size)
{
NativeIntrinsic.Context
.mat_f32_build_panel((MatrixPanel*)leftPanel, leftBuffer,
commonK, ansHeight, k, i);
NativeIntrinsic.Context
.mat_f32_build_panel((MatrixPanel*)rightPanel, rightBuffer,
ansWidth, commonK, j, k);
NativeIntrinsic.Context
.mat_f32_panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (MatrixPanel*)ansPanel);
}
NativeIntrinsic.Context
.mat_f32_store_panel((MatrixPanel*)ansPanel, ansBuffer, ansWidth,
ansHeight, j, i);
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static unsafe MatrixKernel<T>* GetKernelAddressUnsafe<T>(
MatrixKernel<T>* buffer,
int colLength,
int x,
int y
) where T : unmanaged, INumber<T>
{
// Caller guarantees: indices in range, layout is dense square.
int k_c = y >> 2;
int k_r = x >> 2;
return buffer + colLength * x + y;
}
}