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

working on matrix mul matrix

Browse Source
This commit is contained in:
Youmu Liu
2025-12-11 18:09:35 +08:00
parent 631bc2b6e8
commit 9ff438d3b9
13 changed files with 379 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
+61 -32
View File
@@ -39,13 +39,10 @@ using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static class MatrixCompute
public static partial class MatrixCompute
{
public static unsafe Matrix<float> Add(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
@@ -61,7 +58,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
return ans;
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -72,22 +69,22 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return ans;
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Result = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_add;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return ans;
return;
}
public static unsafe void Fill(Matrix<float> ans, float value)
@@ -120,10 +117,7 @@ namespace StgSharp.Mathematics.Numeric
taskGroup.Dispose();
}
public static unsafe Matrix<float> Mul(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
public static unsafe void Mul(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
// For matrix multiplication (left: MxK) * (right: KxN) => ans: MxN
// ColumnLength = number of columns (N), RowLength = number of rows (M)
@@ -134,30 +128,68 @@ namespace StgSharp.Mathematics.Numeric
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 256)
if (count <= 1024)
{
MatrixPanel<float>* leftPanel = MatrixPanel<float>.Create();
MatrixPanel<float>* rightPanel = MatrixPanel<float>.Create();
MatrixPanel<float>* ansPanel = MatrixPanel<float>.Create();
int size = MatrixPanel<float>.KernelSideCount();
MatrixParallelThread[] pool = null;
if (count > 256) {
pool = MatrixParallel.LeadParallel(1);
}
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
int size = MatrixPanel.KernelSideCount<float>();
for (int i = 0; i < left.RowLength; i += size)
{
for (int j = 0; j < right.ColumnLength; j += size)
{
for (int k = 0; k < left.ColumnLength; k += size) {
NativeIntrinsic.Intrinsic.f32_clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < left.ColumnLength; k += size)
{
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixKernel*)left.Buffer,
(MatrixPanel*)leftPanel,
left.ColumnLength, left.RowLength, i,
k);
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixKernel*)right.Buffer,
(MatrixPanel*)rightPanel,
right.ColumnLength, right.RowLength, k,
j);
NativeIntrinsic.Intrinsic
.f32_panel_fma((MatrixPanel*)leftPanel,
(MatrixPanel*)rightPanel,
(MatrixPanel*)ansPanel);
}
NativeIntrinsic.Intrinsic
.f32_store_panel((MatrixKernel*)ans.Buffer,
(MatrixPanel*)ansPanel, ans.ColumnLength,
ans.RowLength, i, j);
}
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
if (count > 256) {
MatrixParallel.ReturnParallelResources(ref pool!);
}
return;
}
throw new NotImplementedException();
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Right = (nint)left.Buffer;
package->Result = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
public static unsafe Matrix<float> Mul(Matrix<float> left, float right, Matrix<float> ans)
public static unsafe void Mul(Matrix<float> left, float right, Matrix<float> ans)
{
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
@@ -171,7 +203,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
return ans;
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -181,7 +213,7 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return ans;
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Right = (nint)left.Buffer;
@@ -197,13 +229,10 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
MatrixParallelFactory.Release(scalar);
return ans;
return;
}
public static unsafe Matrix<float> Sub(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
public static unsafe void Sub(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
@@ -219,7 +248,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
return ans;
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -230,7 +259,7 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return ans;
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
@@ -245,7 +274,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return ans;
return;
}
}