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

working on dynamic wrapping, a good awaiting is on way

Browse Source
This commit is contained in:
Youmu Liu
2025-11-07 18:22:50 +08:00
parent ec1b8dbd51
commit f959339e18
16 changed files with 447 additions and 355 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
+12 -38
View File
@@ -47,6 +47,16 @@ namespace StgSharp.Mathematics.Numeric
}
Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 256)
{
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
for (int i = 0; i < count; i++) {
NativeIntrinsic.Intrinsic.f32_add(leftPtr + i, rightPtr + i, ansPtr + i);
}
return ans;
}
if (count <= 1024)
{
ReadOnlySpan<MatrixParallelThread> pool = MatrixParallel.LeadParallel(1);
@@ -64,50 +74,14 @@ namespace StgSharp.Mathematics.Numeric
}
return ans;
}
Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
MatrixParallelHandle taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
(delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_add);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
/*
* run parallel
*/
return ans;
}
public static unsafe Matrix<float> Sub(Matrix<float> left, Matrix<float> right)
{
if (left.ColumnLength != right.ColumnLength || left.RowLength != right.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 1024)
{
// TODO: this is very awkward, blocking thread renting, need a better way to handle small tasks
ReadOnlySpan<MatrixParallelThread> pool = MatrixParallel.LeadParallel(1);
if (pool.Length == 0)
{
/*
* wait
*/
}
MatrixParallel.ReleaseParallelLeadership(ref pool);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
for (int i = 0; i < count; i++) {
NativeIntrinsic.Intrinsic.f32_sub(leftPtr + i, rightPtr + i, ansPtr + i);
}
// thread.
return ans;
}
Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
(delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_sub);
/*
* run parallel
*/
taskGroup.Dispose();
return ans;
}