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

Mat mul mat is now on full function

Browse Source
This commit is contained in:
Youmu Liu
2025-12-16 16:15:30 +08:00
parent 2c7277f3f9
commit 88ed5a95c5
14 changed files with 357 additions and 488 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Additional.cs
+28 -5
View File
@@ -26,6 +26,7 @@
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
@@ -40,17 +41,39 @@ namespace StgSharp.Mathematics.Numeric
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 i_length = p->PrimCount;
int j_length = p->SecCount;
int k_length = 0;
for (long i = 0; i < lengthRef; i++)
int ansWidth = p-> AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
for (int k = 0; k < k_length; k++) { }
int i = (int)(c / blocksPerRow) * size;
int j = (int)(c % blocksPerRow) * size;
NativeIntrinsic.Intrinsic.f32_clear_panel((MatrixPanel*)ansPanel);
for (int k = 0; k < commonK; k += size)
{
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)leftPanel, leftBuffer, commonK,
ansHeight, k, i);
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)rightPanel, rightBuffer, ansWidth,
commonK, j, k);
NativeIntrinsic.Intrinsic
.f32_panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (MatrixPanel*)ansPanel);
}
NativeIntrinsic.Intrinsic
.f32_store_panel((MatrixPanel*)ansPanel, ansBuffer, ansWidth,
ansHeight, j, i);
}
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
-281
View File
@@ -1,281 +0,0 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute.Basic"
// Project: StgSharp
// AuthorGroup: Nitload
// Copyright (c) Nitload. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
// SOFTWARE.
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.Collections;
using StgSharp.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.ComponentModel.DataAnnotations.Schema;
using System.Linq;
using System.Numerics;
using System.Security.AccessControl;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static partial class MatrixCompute
{
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
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;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
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 = 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;
}
public static unsafe void Fill(Matrix<float> ans, float value)
{
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = value;
if (count <= 256)
{
NativeIntrinsic.Intrinsic.f32_buffer_fill(ans.Buffer, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
NativeIntrinsic.Intrinsic.f32_buffer_fill(ans.Buffer, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Result = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationParam.ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_fill;
package->Scalar = scalar;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
MatrixParallelFactory.Release(scalar);
taskGroup.Dispose();
}
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)
if (left.ColumnLength != right.RowLength ||
ans.ColumnLength != right.ColumnLength ||
ans.RowLength != left.RowLength)
{
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 1024)
{
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)
{
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;
}
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 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.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = right;
if (count <= 256)
{
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
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->Scalar = scalar;
package->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationSpecial.SEQ_FMA);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
MatrixParallelFactory.Release(scalar);
return;
}
public static unsafe void Sub(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
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;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
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->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_sub;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.cs
+251 -1
View File
@@ -25,13 +25,263 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.Collections;
using StgSharp.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.ComponentModel.DataAnnotations.Schema;
using System.Linq;
using System.Numerics;
using System.Security.AccessControl;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static partial class MatrixCompute { }
public static partial class MatrixCompute
{
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
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;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)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)NativeIntrinsic.Intrinsic.f32_buffer_add;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
public static unsafe void Fill(Matrix<float> ans, float value)
{
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = value;
if (count <= 256)
{
NativeIntrinsic.Intrinsic.f32_buffer_fill(ans.Buffer, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
NativeIntrinsic.Intrinsic.f32_buffer_fill(ans.Buffer, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_fill;
package->Scalar = scalar;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
MatrixParallelFactory.Release(scalar);
taskGroup.Dispose();
}
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)
if (left.ColumnLength != right.RowLength ||
ans.ColumnLength != right.ColumnLength ||
ans.RowLength != left.RowLength)
{
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
int size = MatrixPanel.KernelSideCount<float>();
if (count <= 1024)
{
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>();
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
NativeIntrinsic.Intrinsic.f32_clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)leftPanel,
(MatrixKernel*)left.Buffer,
left.KernelColumnLength,
left.KernelRowLength, i, k);
NativeIntrinsic.Intrinsic
.f32_build_panel((MatrixPanel*)rightPanel,
(MatrixKernel*)right.Buffer,
right.KernelColumnLength,
right.KernelRowLength, k, j);
NativeIntrinsic.Intrinsic
.f32_panel_fma((MatrixPanel*)leftPanel,
(MatrixPanel*)rightPanel,
(MatrixPanel*)ansPanel);
}
NativeIntrinsic.Intrinsic
.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!);
}
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->AnsPrimOffset = ans.KernelRowLength;
package->AnsSecOffset = ans.KernelColumnLength;
package->LeftPrimOffset = left.KernelColumnLength;
package->PrimCount = ans.KernelColumnLength / size;
package->SecCount = ans.KernelRowLength / size;
package->ElementSize = size;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationSpecial.SEQ_FMA);
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
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.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = right;
if (count <= 256)
{
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Right = (nint)left.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->Scalar = scalar;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, 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();
MatrixParallelFactory.Release(scalar);
return;
}
public static unsafe void Sub(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
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;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_sub;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
}
}
StgSharp.Mathematics/Numeric/MatrixComputeModel.cs
+13 -15
View File
@@ -34,7 +34,7 @@ namespace StgSharp.Mathematics.Numeric
internal static class MatrixComputeModel
{
internal static unsafe void KernelComputeBinary(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeBinary(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -42,12 +42,12 @@ namespace StgSharp.Mathematics.Numeric
((delegate* unmanaged[Cdecl]<void*, void*, void*, long, void>)p->ComputeHandle)((void*)(p->Left + offset),
(void*)(p->Right + offset),
(void*)(p->Result + offset),
(void*)(p->Answer + offset),
count);
MatrixParallelFactory.Release(p);
}
internal static unsafe void KernelComputeBinaryScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeBinaryScalar(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -55,57 +55,55 @@ namespace StgSharp.Mathematics.Numeric
ScalarPacket* scalar = p->Scalar;
((delegate* unmanaged[Cdecl]<void*, void*, void*,ScalarPacket*, long, void>)p->ComputeHandle)((void*)(p->Left + offset),
(void*)(p->Right + offset),
(void*)(p->Result + offset),
(void*)(p->Answer + offset),
scalar,
count);
MatrixParallelFactory.Release(p);
}
internal static unsafe void KernelComputeNoOperatorScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeNoOperatorScalar(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
long offset = Unsafe.As<int, long>(ref p->PrimTileOffset) * size;
((delegate* unmanaged[Cdecl]<void*, ScalarPacket*,long, void>)p->ComputeHandle)((void*)(p->Result + offset),
((delegate* unmanaged[Cdecl]<void*, ScalarPacket*,long, void>)p->ComputeHandle)((void*)(p->Answer + offset),
p->Scalar, count);
MatrixParallelFactory.Release(p);
}
internal static unsafe void KernelComputeUnary(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeUnary(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
long offset = Unsafe.As<int, long>(ref p->PrimTileOffset) * size;
((delegate* unmanaged[Cdecl]<void*, void*, long, void>)p->ComputeHandle)((void*)(p->Right + offset),
(void*)(p->Result + offset),
(void*)(p->Answer + offset),
count);
MatrixParallelFactory.Release(p);
}
internal static unsafe void KernelComputeUnaryScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeUnaryScalar(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
long offset = Unsafe.As<int, long>(ref p->PrimTileOffset) * size;
((delegate* unmanaged[Cdecl]< void*, void*, ScalarPacket*, long, void>)p->ComputeHandle)((void*)(p->Right + offset),
(void*)(p->Result + offset),
(void*)(p->Answer + offset),
p->Scalar,
count);
MatrixParallelFactory.Release(p);
}
internal static unsafe void PanelComputeBinary(MatrixParallelTaskPackage* p) { }
internal static unsafe void SpecialCompute(MatrixParallelTaskPackage* p)
internal static unsafe void SpecialBufferCompute(MatrixParallelTaskPackage* p)
{
switch ((MatrixOperationSpecial)p->ComputeMode.OperationStyle)
switch ((MatrixOperationSpecial)p->ComputeMode.ParameterStyle)
{
case MatrixOperationSpecial.SEQ_FMA:
MatrixCompute.ComputeFmaPanelF32(p);
break;
default:
throw new NotSupportedException("The special matrix operation style is not supported.");
break;
}
}
StgSharp.Mathematics/Numeric/MatrixParallel.Publish.cs
+1 -76
View File
@@ -51,7 +51,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelTaskPackage* nonGeneric = package;
switch (nonGeneric->ComputeMode.OperationStyle)
{
case MatrixOperationStyle.BUFFER_INDEX:
case MatrixIndexStyle.BUFFER_INDEX:
return ScheduleBufferTask(nonGeneric);
default:
throw new InvalidOperationException($"Unknown mode {nonGeneric->ComputeMode.OperationStyle}.");
@@ -110,80 +110,5 @@ namespace StgSharp.Mathematics.Numeric
return handle;
}
private static unsafe MatrixParallelHandle ScheduleKernelTask(
MatrixParallelTaskPackage* package)
{
int primCount = package->PrimCount;
int secCount = package->SecCount;
int tileWidth = (1024 / secCount) + 1;
int tileCount = (primCount + tileWidth - 1) / tileWidth;
MatrixParallelThread[] pool = LeadParallel(tileCount);
/*
if (pool.Length == 0)
{
// no need
// LeadParallel has contained capacity check
}
/**/
MatrixParallelHandle handle = new MatrixParallelHandle(pool);
if (handle.Wraps.Length == 1)
{
handle.Wraps[0].WrapTask.Enqueue((nint)package);
return handle;
}
int threadCount = pool.Length;
int tilePerThreadLess = tileCount / threadCount; // tiles per thread without extra slice
int columnAfterBaseTile = tileCount % threadCount * tileWidth; // tile sliced into columns
int extraColumnPerThread = columnAfterBaseTile / threadCount;
int extraColumnCount = columnAfterBaseTile % threadCount; // columns to be sliced
int total = primCount;
MatrixParallelTaskPackage* current = package;
foreach (MatrixParallelWrap wrap in handle.Wraps)
{
bool hasSlice = columnAfterBaseTile > 0;
bool isLargeSlice = extraColumnCount > 0;
// _count tiles for wrap
int wrapWidth = wrap!.WrapTaskCapacity/**/;
int columnFromSlicingCount = (hasSlice ? extraColumnPerThread : 0) * wrapWidth;
int baseColumn = wrapWidth * tilePerThreadLess * tileWidth;
int columnInWrap = baseColumn + columnFromSlicingCount + (isLargeSlice ? 1 : 0);
// set tasks
if (columnInWrap == 0)
{
#if DEBUG
throw new InvalidOperationException("Zero tasks published.");
#else
break;
#endif
}
// update current task
current->PrimColumnCountInTile = columnInWrap;
current->SecColumnCountInTile = current->SecCount;
current->SecTileOffset = 0;
// remove tiles _count
extraColumnCount--;
columnAfterBaseTile -= columnFromSlicingCount;
total -= columnInWrap;
// set to wrap and push
wrap.WrapTask.Enqueue((nint)current);
if (total <= 0)
{
break;
}
current = MatrixParallelFactory.FromExistPackage(current);
current->PrimTileOffset += columnInWrap;
}
return handle;
}
}
}
StgSharp.Mathematics/Numeric/MatrixParallel.cs
+1 -1
View File
@@ -34,7 +34,7 @@ using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static class MatrixParallel
public static partial class MatrixParallel
{
private static bool _meaningful;
StgSharp.Mathematics/Numeric/MatrixParallelTaskPackage.cs
+11 -10
View File
@@ -59,7 +59,7 @@ namespace StgSharp.Mathematics.Numeric
}
public enum MatrixOperationStyle : short
public enum MatrixIndexStyle : short
{
BUFFER_INDEX = 1,
@@ -78,13 +78,14 @@ namespace StgSharp.Mathematics.Numeric
}
[StructLayout(LayoutKind.Sequential)]
public struct MatrixOpMode(MatrixOperationStyle style, MatrixOperationParam paramLayout)
public struct MatrixOpMode(MatrixIndexStyle style, MatrixOperationParam paramLayout)
{
public MatrixOperationParam ParameterStyle = paramLayout;
public MatrixOperationStyle OperationStyle = style;
public MatrixIndexStyle OperationStyle = style;
public MatrixOpMode(MatrixOperationStyle style, MatrixOperationSpecial paramLayout)
public MatrixOperationParam ParameterStyle = paramLayout;
public MatrixOpMode(MatrixIndexStyle style, MatrixOperationSpecial paramLayout)
: this(style, (MatrixOperationParam)paramLayout) { }
}
@@ -113,7 +114,7 @@ namespace StgSharp.Mathematics.Numeric
#region ans
// Group 2: Result kernel base address + reserved
[FieldOffset(16)] public IntPtr Result; // 8
[FieldOffset(16)] public IntPtr Answer; // 8
[FieldOffset(24)] public int ElementSize;
[FieldOffset(28)] public int ReservedPtr0; // 4 (extension/reserved)
@@ -139,10 +140,10 @@ namespace StgSharp.Mathematics.Numeric
#region ans enum
[FieldOffset(64)]public int ResultPrimOffset; // 4
[FieldOffset(68)]public int ResultPrimStride; // 4
[FieldOffset(72)]public int ResultSecOffset; // 4
[FieldOffset(76)] public int ResultSecStride; // 4
[FieldOffset(64)]public int AnsPrimOffset; // 4
[FieldOffset(68)]public int AnsPrimStride; // 4
[FieldOffset(72)]public int AnsSecOffset; // 4
[FieldOffset(76)] public int AnsSecStride; // 4
#endregion
StgSharp.Mathematics/Numeric/MatrixParallelThread.cs
+8 -29
View File
@@ -85,49 +85,28 @@ namespace StgSharp.Mathematics.Numeric
while (taskQueue.TryDequeue(out nint handle))
{
MatrixParallelTaskPackage* p = (MatrixParallelTaskPackage*)handle;
if ((int)p->ComputeMode.ParameterStyle < 0) {
MatrixComputeModel.SpecialCompute(p);
if ((short)p->ComputeMode.ParameterStyle < 0) {
MatrixComputeModel.SpecialBufferCompute(p);
}
switch (p->ComputeMode.OperationStyle)
{
case MatrixOperationStyle.PANEL_OP:
case MatrixIndexStyle.BUFFER_INDEX:
switch (p->ComputeMode.ParameterStyle)
{
case MatrixOperationParam.RIGHT_ANS_PARAM:
// MatrixComputeModel.PanelComputeUnary(p);
MatrixComputeModel.BufferComputeUnary(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_PARAM:
MatrixComputeModel.PanelComputeBinary(p);
MatrixComputeModel.BufferComputeBinary(p);
break;
case MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM:
// MatrixComputeModel.PanelComputeUnaryScalar(p);
MatrixComputeModel.BufferComputeUnaryScalar(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_SCALAR_PARAM:
MatrixComputeModel.BufferComputeBinaryScalar(p);
break;
case MatrixOperationParam.ANS_SCALAR_PARAM:
// MatrixComputeModel.PanelComputeNoOperatorScalar(p);
break;
default:
break;
}
break;
case MatrixOperationStyle.KERNEL_OP:
switch (p->ComputeMode.ParameterStyle)
{
case MatrixOperationParam.RIGHT_ANS_PARAM:
MatrixComputeModel.KernelComputeUnary(p);
break;
case MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM:
MatrixComputeModel.KernelComputeUnaryScalar(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_PARAM:
MatrixComputeModel.KernelComputeBinary(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_SCALAR_PARAM:
MatrixComputeModel.KernelComputeBinaryScalar(p);
MatrixComputeModel.BufferComputeNoOperatorScalar(p);
break;
default:
break;