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

add simd id and cpu id

Browse Source
This commit is contained in:
Youmu Liu
2026-01-06 18:04:17 +08:00
parent 88fa653323
commit 3d488abd2e
17 changed files with 372 additions and 615 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Float.cs
+14 -54
View File
@@ -143,72 +143,32 @@ namespace StgSharp.Mathematics.Numeric
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
int size = MatrixPanel.KernelSideCount<float>();
// int size = MatrixPanel.KernelSideCount<float>();
if (count <= 1024)
{
MatrixParallelThread[] pool = null;
if (count > 256) {
pool = MatrixParallel.LeadParallel(1);
}
if (NativeIntrinsic.Context.mat[f32].build_panel == null)
for (int i = 0; i < ans.KernelRowLength; i++)
{
for (int i = 0; i < ans.KernelRowLength; i += size)
for (int j = 0; j < ans.KernelColumnLength; j++)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
MatrixKernel<float>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
MatrixKernel<float>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f32].panel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k++)
{
NativeIntrinsic.Context
.mat[f32].kernel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
(MatrixPanel*)ansPanel);
}
}
}
} else
{
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f32].build_panel((MatrixPanel*)leftPanel,
(MatrixKernel*)left.Buffer,
left.KernelColumnLength,
left.KernelRowLength, i, k);
NativeIntrinsic.Context
.mat[f32].build_panel((MatrixPanel*)rightPanel,
(MatrixKernel*)right.Buffer,
right.KernelColumnLength,
right.KernelRowLength, k, j);
NativeIntrinsic.Context
.mat[f32].panel_fma((MatrixPanel*)leftPanel,
(MatrixPanel*)rightPanel,
(MatrixPanel*)ansPanel);
}
NativeIntrinsic.Context
.mat[f32].store_panel((MatrixPanel*)ansPanel,
(MatrixKernel*)ans.Buffer,
ans.KernelColumnLength,
ans.KernelRowLength, i, j);
}
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
}
if (count > 256) {
MatrixParallel.ReturnParallelResources(ref pool!);
}
@@ -221,8 +181,8 @@ namespace StgSharp.Mathematics.Numeric
package->AnsPrimOffset = ans.KernelRowLength;
package->AnsSecOffset = ans.KernelColumnLength;
package->LeftPrimOffset = left.KernelColumnLength;
package->PrimCount = ans.KernelColumnLength / size;
package->SecCount = ans.KernelRowLength / size;
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ElementSize = (int)MatrixElementType.F32;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationSpecial.SEQ_FMA);
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
StgSharp.Mathematics/Numeric/MatrixCompute.cs
+26 -85
View File
@@ -54,55 +54,26 @@ namespace StgSharp.Mathematics.Numeric
int ansWidth = p->AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
if (NativeIntrinsic.Context.mat[f32].build_panel == null)
// int blocksPerRow = ansWidth / size;
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
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(
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) {
NativeIntrinsic.Context
.mat[f32].kernel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)ansPanel);
}
}
} 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);
}
}
@@ -111,7 +82,8 @@ namespace StgSharp.Mathematics.Numeric
)
{
int f64 = (int)MatrixElementType.F64;
int size = MatrixPanel.KernelSideCount<double>();
// int size = 1;
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
MatrixKernel* rightBuffer = (MatrixKernel*)p->Right;
MatrixKernel* ansBuffer = (MatrixKernel*)p->Answer;
@@ -120,55 +92,24 @@ namespace StgSharp.Mathematics.Numeric
int ansWidth = p->AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
if (NativeIntrinsic.Context.mat[f64].build_panel == null)
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
int i = (int)(c / blocksPerRow);
int j = (int)(c % blocksPerRow);
MatrixKernel<double>* ansPanel = GetKernelAddressUnsafe((MatrixKernel<double>*)ansBuffer,
ansHeight, i, j);
NativeIntrinsic.Context.mat[f64].clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size) {
NativeIntrinsic.Context
.mat[f64].panel_fma(
int i = (int)(c / ansWidth);
int j = (int)(c % ansWidth);
MatrixKernel<double>* ansPanel = GetKernelAddressUnsafe((MatrixKernel<double>*)ansBuffer,
ansHeight, i, j);
NativeIntrinsic.Context.mat[f64].clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += 1) {
NativeIntrinsic.Context
.mat[f64].kernel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)ansPanel);
}
}
} else
{
MatrixPanel<double>* leftPanel = MatrixPanel.Create<double>();
MatrixPanel<double>* rightPanel = MatrixPanel.Create<double>();
MatrixPanel<double>* ansPanel = MatrixPanel.Create<double>();
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
int i = (int)(c / blocksPerRow) * size;
int j = (int)(c % blocksPerRow) * size;
NativeIntrinsic.Context.mat[f64].clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size)
{
NativeIntrinsic.Context
.mat[f64].build_panel((MatrixPanel*)leftPanel, leftBuffer,
commonK, ansHeight, k, i);
NativeIntrinsic.Context
.mat[f64].build_panel((MatrixPanel*)rightPanel, rightBuffer,
ansWidth, commonK, j, k);
NativeIntrinsic.Context
.mat[f64].panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (MatrixPanel*)ansPanel);
}
NativeIntrinsic.Context
.mat[f64].store_panel((MatrixPanel*)ansPanel, ansBuffer,
ansWidth, ansHeight, j, i);
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.double.cs
+2 -2
View File
@@ -159,7 +159,7 @@ namespace StgSharp.Mathematics.Numeric
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f64].panel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
.mat[f64].kernel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
(MatrixPanel*)ansPanel);
}
@@ -190,7 +190,7 @@ namespace StgSharp.Mathematics.Numeric
right.KernelColumnLength,
right.KernelRowLength, k, j);
NativeIntrinsic.Context
.mat[f64].panel_fma((MatrixPanel*)leftPanel,
.mat[f64].kernel_fma((MatrixPanel*)leftPanel,
(MatrixPanel*)rightPanel,
(MatrixPanel*)ansPanel);
}