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

Better FMA

Browse Source
This commit is contained in:
Youmu Liu
2026-03-05 18:04:34 +08:00
parent dfa6f4ba92
commit 83997f64d0
21 changed files with 893 additions and 1030 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Float.cs
+30 -30
View File
@@ -49,7 +49,7 @@ namespace StgSharp.Mathematics.Numeric
Matrix<float> ans
)
{
int f32 = (int)MatrixElementType.F32;
MatrixElementType f32 = MatrixElementType.F32;
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
@@ -62,8 +62,8 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_add(leftPtr, rightPtr, ansPtr,
count);
NumericalModule.GlobalContext.mat_mk[f32].buffer_add(leftPtr, rightPtr, ansPtr,
count);
return;
} else if (count <= 1024)
@@ -73,21 +73,21 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_add(leftPtr, rightPtr, ansPtr,
count);
NumericalModule.GlobalContext.mat_mk[f32].buffer_add(leftPtr, rightPtr, ansPtr,
count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTask* package = MatrixParallelFactory.CreateBaseTask();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->Mat1 = (MatrixKernel*)left.Buffer;
package->Mat2 = (MatrixKernel*)right.Buffer;
package->Mat3 = (MatrixKernel*)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)MathematicsModule.GlobalContext.mat_sk[f32].buffer_add;
package->ComputeHandle = (IntPtr)NumericalModule.GlobalContext.mat_sk[f32].buffer_add;
MatrixParallelWrap taskGroup = MatrixParallel.ScheduleTask(package);
taskGroup.LaunchParallel(SleepMode.DeepSleep);
@@ -100,18 +100,18 @@ namespace StgSharp.Mathematics.Numeric
float value
)
{
int f32 = (int)MatrixElementType.F32;
MatrixElementType f32 = MatrixElementType.F32;
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = value;
if (count <= 256)
{
MathematicsModule.GlobalContext.mat_sk[f32].buffer_fill(ans.Buffer, scalar, count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_fill(ans.Buffer, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MathematicsModule.GlobalContext.mat_sk[f32].buffer_fill(ans.Buffer, scalar, count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_fill(ans.Buffer, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
@@ -121,7 +121,7 @@ namespace StgSharp.Mathematics.Numeric
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)MathematicsModule.GlobalContext.mat_sk[f32].buffer_fill;
package->ComputeHandle = (IntPtr)NumericalModule.GlobalContext.mat_sk[f32].buffer_fill;
package->Scalar = scalar;
MatrixParallelWrap taskGroup = MatrixParallel.ScheduleTask(package);
taskGroup.LaunchParallel(SleepMode.DeepSleep);
@@ -135,7 +135,7 @@ namespace StgSharp.Mathematics.Numeric
Matrix<float> ans
)
{
int f32 = (int)MatrixElementType.F32;
MatrixElementType f32 = MatrixElementType.F32;
// For matrix multiplication (left: MxK) * (right: KxN) => ans: MxN
// ColumnLength = number of columns (N), RowLength = number of rows (M)
@@ -168,11 +168,11 @@ namespace StgSharp.Mathematics.Numeric
enumeration->Member<int>(0) = i;
enumeration->Member<int>(1) = j;
MathematicsModule.GlobalContext
NumericalModule.GlobalContext
.mat_sk[f32].kernel_tile_fma((MatrixKernel*)left.Buffer,
(MatrixKernel*)right.Buffer,
(MatrixKernel*)ans.Buffer,
enumeration);
(MatrixKernel*)right.Buffer,
(MatrixKernel*)ans.Buffer,
enumeration);
}
}
@@ -205,7 +205,7 @@ namespace StgSharp.Mathematics.Numeric
Matrix<float> ans
)
{
int f32 = (int)MatrixElementType.F32;
MatrixElementType f32 = MatrixElementType.F32;
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
@@ -220,8 +220,8 @@ namespace StgSharp.Mathematics.Numeric
{
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_scalar_mul(leftPtr, ansPtr, scalar,
count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_scalar_mul(leftPtr, ansPtr, scalar,
count);
return;
} else if (count <= 1024)
@@ -230,8 +230,8 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_scalar_mul(leftPtr, ansPtr, scalar,
count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_scalar_mul(leftPtr, ansPtr, scalar,
count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
@@ -244,7 +244,7 @@ namespace StgSharp.Mathematics.Numeric
package->SecCount = left.KernelRowLength;
package->Scalar = scalar;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)MathematicsModule.GlobalContext.mat_sk[f32].buffer_scalar_mul;
package->ComputeHandle = (IntPtr)NumericalModule.GlobalContext.mat_sk[f32].buffer_scalar_mul;
MatrixParallelWrap taskGroup = MatrixParallel.ScheduleTask(package);
taskGroup.LaunchParallel(SleepMode.DeepSleep);
@@ -259,7 +259,7 @@ namespace StgSharp.Mathematics.Numeric
Matrix<float> ans
)
{
int f32 = (int)MatrixElementType.F32;
MatrixElementType f32 = MatrixElementType.F32;
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
@@ -277,8 +277,8 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_sub(leftPtr, rightPtr, ansPtr,
count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_sub(leftPtr, rightPtr, ansPtr,
count);
return;
} else if (count <= 1024)
@@ -288,8 +288,8 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
MathematicsModule.GlobalContext.mat_sk[f32].buffer_sub(leftPtr, rightPtr, ansPtr,
count);
NumericalModule.GlobalContext.mat_sk[f32].buffer_sub(leftPtr, rightPtr, ansPtr,
count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
@@ -302,7 +302,7 @@ namespace StgSharp.Mathematics.Numeric
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)MathematicsModule.GlobalContext.mat_sk[f32].buffer_sub;
package->ComputeHandle = (IntPtr)NumericalModule.GlobalContext.mat_sk[f32].buffer_sub;
MatrixParallelWrap taskGroup = MatrixParallel.ScheduleTask(package);
taskGroup.LaunchParallel(SleepMode.DeepSleep);