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

Daily update on 20251230-224423 by Nitload Youmu. Additional Message : Adding compute support for mat<double>. Import all opengl core profile fucntions

Browse Source 
Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
This commit is contained in:
Nitload
2025-12-30 22:44:23 +08:00
parent cf370ca98a 6484307b8a
commit 5decbd390d
83 changed files with 75819 additions and 6024 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.Additional.cs
-84
View File
@@ -1,84 +0,0 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute.Additional"
// 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.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Runtime.CompilerServices;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public partial class MatrixCompute
{
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 ansWidth = p-> AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
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);
MatrixPanel.Destroy(ansPanel);
}
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.Factorize.cs
+80 -28
View File
@@ -25,6 +25,7 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
@@ -36,11 +37,12 @@ namespace StgSharp.Mathematics.Numeric
public static partial class MatrixCompute
{
public static void FactorizePLU(
Matrix<float> matrix,
Span<float> pivot,
Matrix<float> lower,
Matrix<float> upper)
public static unsafe void FactorizePLU(
Matrix<float> matrix,
Span<int> pivot,
Matrix<float> lower,
Matrix<float> upper
)
{
if (matrix.ColumnLength != matrix.RowLength) {
throw new InvalidOperationException("Matrix must be square.");
@@ -54,7 +56,7 @@ namespace StgSharp.Mathematics.Numeric
}
if (lower.RowLength != n ||
lower.ColumnLength != n ||
lower.Layout != MatrixLayout.LowerTriangle) {
lower.Layout != MatrixLayout.UpperTriangle) {
throw new ArgumentException("Lower matrix must be lower-triangular with same dimensions as input.", nameof(lower));
}
if (upper.RowLength != n ||
@@ -66,37 +68,87 @@ namespace StgSharp.Mathematics.Numeric
int size = matrix.KernelColumnLength * matrix.KernelRowLength;
if (size < 256)
{
// Naive Doolittle-style PLU without row swaps (for illustration).
for (int i = 0; i < n; i++)
{
pivot[i] = i; // identity pivot
lower[i, i] = 1.0f; // unit diagonal for L
pivot[i] = i; // identity permutation
}
for (int k = 0; k < n; k++)
// Aggressive: always call native pivot to fetch four-column candidates
ulong* pivots = stackalloc ulong[4];
for (int kBlock = 0; kBlock < n; kBlock += 4)
{
// Compute U[k, j] for j >= k
for (int j = k; j < n; j++)
NativeIntrinsic.Context
.mat[(int)MatrixElementType.F32].pivot(matrix.Buffer,
(ulong)(kBlock >> 2),
(ulong)(matrix.KernelRowLength - (kBlock >> 2)),
(ulong)pivots);
int colsInBlock = Math.Min(4, n - kBlock);
for (int offset = 0; offset < colsInBlock; offset++)
{
float sum = 0f;
for (int s = 0; s < k; s++) {
sum += lower[k, s] * upper[s, j];
}
upper[k, j] = matrix[k, j] - sum;
}
int k = kBlock + offset;
int pivotRow = k;
float maxAbs;
// Compute L[i, k] for i > k
for (int i = k + 1; i < n; i++)
{
float sum = 0f;
for (int s = 0; s < k; s++) {
sum += lower[i, s] * upper[s, k];
int candidateRow = (int)pivots[offset];
int idx = -1;
for (int r = k; r < n; r++)
{
if (pivot[r] == candidateRow)
{
idx = r;
break;
}
}
float u = upper[k, k];
if (u == 0f) {
throw new InvalidOperationException("Zero pivot encountered in demonstration PLU.");
if (idx != -1)
{
pivotRow = idx;
maxAbs = MathF.Abs(matrix.GetElementUnsafe(pivot[idx], k));
} else
{
maxAbs = MathF.Abs(matrix.GetElementUnsafe(pivot[k], k));
for (int r = k + 1; r < n; r++)
{
float val = MathF.Abs(matrix.GetElementUnsafe(pivot[r], k));
if (val > maxAbs)
{
maxAbs = val;
pivotRow = r;
}
}
}
if (maxAbs == 0f) {
throw new InvalidOperationException("Zero pivot encountered in PLU.");
}
if (pivotRow != k) {
(pivot[k], pivot[pivotRow]) = (pivot[pivotRow], pivot[k]);
}
lower[k, k] = 1.0f; // unit diagonal for L
// Compute U[k, j] for j >= k
for (int j = k; j < n; j++)
{
float sum = 0f;
for (int s = 0; s < k; s++) {
sum += lower[k, s] * upper[s, j];
}
upper[k, j] = matrix.GetElementUnsafe(pivot[k], j) - sum;
}
// Compute L[i, k] for i > k
for (int i = k + 1; i < n; i++)
{
float sum = 0f;
for (int s = 0; s < k; s++) {
sum += lower[i, s] * upper[s, k];
}
float u = upper[k, k];
if (u == 0f) {
throw new InvalidOperationException("Zero pivot encountered in PLU.");
}
lower[i, k] = (matrix.GetElementUnsafe(pivot[i], k) - sum) / u;
}
lower[i, k] = (matrix.GetElementUnsafe(i, k) - sum) / u;
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.Float.cs
+343
View File
@@ -0,0 +1,343 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute.Float"
// 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
)
{
int f32 = (int)MatrixElementType.F32;
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.Context.mat[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.Context.mat[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.Context.mat[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
)
{
int f32 = (int)MatrixElementType.F32;
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<float>(0) = value;
if (count <= 256)
{
NativeIntrinsic.Context.mat[f32].buffer_fill(ans.Buffer, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
NativeIntrinsic.Context.mat[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.Context.mat[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
)
{
int f32 = (int)MatrixElementType.F32;
// 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);
}
if (NativeIntrinsic.Context.mat[f32].build_panel == null)
{
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
MatrixKernel<float>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f32].panel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
(MatrixPanel*)ansPanel);
}
}
}
} else
{
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{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!);
}
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
)
{
int f32 = (int)MatrixElementType.F32;
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
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.Context.mat[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.Context.mat[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.Context.mat[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
)
{
int f32 = (int)MatrixElementType.F32;
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
right.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
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.Context.mat[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.Context.mat[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.Context.mat[f32].buffer_sub;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.cs
+124 -228
View File
@@ -25,266 +25,162 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
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.Runtime.CompilerServices;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static partial class MatrixCompute
public partial class MatrixCompute
{
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
internal static unsafe void ComputeFmaPanelF32(
MatrixParallelTaskPackage* p
)
{
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();
}
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 f32 = (int)MatrixElementType.F32;
int size = MatrixPanel.KernelSideCount<float>();
if (count <= 1024)
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
MatrixKernel* rightBuffer = (MatrixKernel*)p->Right;
MatrixKernel* ansBuffer = (MatrixKernel*)p->Answer;
ref long offsetRef = ref Unsafe.As<int, long>(ref p->PrimTileOffset);
ref long lengthRef = ref Unsafe.As<int, long>(ref p->PrimCount);
int ansWidth = p->AnsPrimOffset;
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
int blocksPerRow = ansWidth / size;
if (NativeIntrinsic.Context.mat[f32].build_panel == null)
{
MatrixParallelThread[] pool = null;
if (count > 256) {
pool = MatrixParallel.LeadParallel(1);
// 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(
(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 (int i = 0; i < ans.KernelRowLength; i += size)
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
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.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);
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);
if (count > 256) {
MatrixParallel.ReturnParallelResources(ref pool!);
}
}
internal static unsafe void ComputeFmaPanelF64(
MatrixParallelTaskPackage* p
)
{
int f64 = (int)MatrixElementType.F64;
int size = MatrixPanel.KernelSideCount<double>();
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
MatrixKernel* rightBuffer = (MatrixKernel*)p->Right;
MatrixKernel* ansBuffer = (MatrixKernel*)p->Answer;
ref long offsetRef = ref Unsafe.As<int, long>(ref p->PrimTileOffset);
ref long lengthRef = ref Unsafe.As<int, long>(ref p->PrimCount);
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++)
{
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(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)ansPanel);
}
}
return;
} 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);
}
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)
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static unsafe MatrixKernel<T>* GetKernelAddressUnsafe<T>(
MatrixKernel<T>* buffer,
int colLength,
int x,
int y
) where T : unmanaged, INumber<T>
{
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
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.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
right.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
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 buffer + (colLength * x + y);
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.double.cs
+339
View File
@@ -0,0 +1,339 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute.double"
// 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.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public partial class MatrixCompute
{
public static unsafe void Add(
Matrix<double> left,
Matrix<double> right,
Matrix<double> ans
)
{
int f64 = (int)MatrixElementType.F64;
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<double>* leftPtr = left.Buffer;
MatrixKernel<double>* rightPtr = right.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_add(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<double>* leftPtr = left.Buffer;
MatrixKernel<double>* rightPtr = right.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_add(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<double>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(double);
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat[f64].buffer_add;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
public static unsafe void Fill(
Matrix<double> ans,
double value
)
{
int f64 = (int)MatrixElementType.F64;
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<double>(0) = value;
if (count <= 256)
{
NativeIntrinsic.Context.mat[f64].buffer_fill(ans.Buffer, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
NativeIntrinsic.Context.mat[f64].buffer_fill(ans.Buffer, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<double>();
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(double);
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.ANS_SCALAR_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat[f64].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<double> left,
Matrix<double> right,
Matrix<double> ans
)
{
int f64 = (int)MatrixElementType.F64;
// 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<double>();
if (count <= 1024)
{
MatrixParallelThread[] pool = null;
if (count > 256) {
pool = MatrixParallel.LeadParallel(1);
}
if (NativeIntrinsic.Context.mat[f64].build_panel == null)
{
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
MatrixKernel<double>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
NativeIntrinsic.Context.mat[f64].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f64].panel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
(MatrixPanel*)ansPanel);
}
}
}
} else
{
MatrixPanel<double>* leftPanel = MatrixPanel.Create<double>();
MatrixPanel<double>* rightPanel = MatrixPanel.Create<double>();
MatrixPanel<double>* ansPanel = MatrixPanel.Create<double>();
for (int i = 0; i < ans.KernelRowLength; i += size)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
{
NativeIntrinsic.Context.mat[f64].clear_panel((MatrixPanel*)ansPanel);
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
{
NativeIntrinsic.Context
.mat[f64].build_panel((MatrixPanel*)leftPanel,
(MatrixKernel*)left.Buffer,
left.KernelColumnLength,
left.KernelRowLength, i, k);
NativeIntrinsic.Context
.mat[f64].build_panel((MatrixPanel*)rightPanel,
(MatrixKernel*)right.Buffer,
right.KernelColumnLength,
right.KernelRowLength, k, j);
NativeIntrinsic.Context
.mat[f64].panel_fma((MatrixPanel*)leftPanel,
(MatrixPanel*)rightPanel,
(MatrixPanel*)ansPanel);
}
NativeIntrinsic.Context
.mat[f64].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<double>();
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 = (int)MatrixElementType.F64;
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<double> left,
double right,
Matrix<double> ans
)
{
int f64 = (int)MatrixElementType.F64;
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
ScalarPacket* scalar = MatrixParallelFactory.CreateScalarPacket();
scalar->Data<double>(0) = right;
if (count <= 256)
{
MatrixKernel<double>* leftPtr = left.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<double>* leftPtr = left.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<double>();
package->Right = (nint)left.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(double);
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.Context.mat[f64].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<double> left,
Matrix<double> right,
Matrix<double> ans
)
{
int f64 = (int)MatrixElementType.F64;
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
left.ColumnLength != ans.ColumnLength ||
left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
right.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 256)
{
MatrixKernel<double>* leftPtr = left.Buffer;
MatrixKernel<double>* rightPtr = right.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_sub(leftPtr, rightPtr, ansPtr, count);
return;
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
MatrixKernel<double>* leftPtr = left.Buffer;
MatrixKernel<double>* rightPtr = right.Buffer;
MatrixKernel<double>* ansPtr = ans.Buffer;
NativeIntrinsic.Context.mat[f64].buffer_sub(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
return;
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<double>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(double);
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat[f64].buffer_sub;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
return;
}
}
}
StgSharp.Mathematics/Numeric/MatrixComputeModel.cs
+29 -7
View File
@@ -34,7 +34,9 @@ namespace StgSharp.Mathematics.Numeric
internal static class MatrixComputeModel
{
internal static unsafe void BufferComputeBinary(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeBinary(
MatrixParallelTaskPackage* p
)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -47,7 +49,9 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelFactory.Release(p);
}
internal static unsafe void BufferComputeBinaryScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeBinaryScalar(
MatrixParallelTaskPackage* p
)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -61,7 +65,9 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelFactory.Release(p);
}
internal static unsafe void BufferComputeNoOperatorScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeNoOperatorScalar(
MatrixParallelTaskPackage* p
)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -71,7 +77,9 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelFactory.Release(p);
}
internal static unsafe void BufferComputeUnary(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeUnary(
MatrixParallelTaskPackage* p
)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -83,7 +91,9 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelFactory.Release(p);
}
internal static unsafe void BufferComputeUnaryScalar(MatrixParallelTaskPackage* p)
internal static unsafe void BufferComputeUnaryScalar(
MatrixParallelTaskPackage* p
)
{
int size = p->ElementSize * 16;
long count = Unsafe.As<int, long>(ref p->PrimCount);
@@ -95,12 +105,24 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelFactory.Release(p);
}
internal static unsafe void SpecialBufferCompute(MatrixParallelTaskPackage* p)
internal static unsafe void SpecialBufferCompute(
MatrixParallelTaskPackage* p
)
{
switch ((MatrixOperationSpecial)p->ComputeMode.ParameterStyle)
{
case MatrixOperationSpecial.SEQ_FMA:
MatrixCompute.ComputeFmaPanelF32(p);
switch ((MatrixElementType)p->ElementSize)
{
case MatrixElementType.F64:
MatrixCompute.ComputeFmaPanelF64(p);
return;
case MatrixElementType.F32:
MatrixCompute.ComputeFmaPanelF32(p);
break;
default:
break;
}
break;
default:
break;
StgSharp.Mathematics/Numeric/MatrixParallelTaskPackage.cs
+20 -3
View File
@@ -78,18 +78,34 @@ namespace StgSharp.Mathematics.Numeric
}
[StructLayout(LayoutKind.Sequential)]
public struct MatrixOpMode(MatrixIndexStyle style, MatrixOperationParam paramLayout)
public struct MatrixOpMode(
MatrixIndexStyle style,
MatrixOperationParam paramLayout
)
{
public MatrixIndexStyle OperationStyle = style;
public MatrixOperationParam ParameterStyle = paramLayout;
public MatrixOpMode(MatrixIndexStyle style, MatrixOperationSpecial paramLayout)
public MatrixOpMode(
MatrixIndexStyle style,
MatrixOperationSpecial paramLayout
)
: this(style, (MatrixOperationParam)paramLayout) { }
}
public enum MatrixElementType : int
{
F32 = 0,
F64 = 1,
I32 = 2,
I64 = 3,
}
[StructLayout(LayoutKind.Explicit, Size = 128)]
public unsafe struct MatrixParallelTaskPackage
{
@@ -97,7 +113,8 @@ namespace StgSharp.Mathematics.Numeric
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static unsafe void Copy(
MatrixParallelTaskPackage* source,
MatrixParallelTaskPackage* target)
MatrixParallelTaskPackage* target
)
{
Buffer.MemoryCopy(source, target, sizeof(MatrixParallelTaskPackage),
sizeof(MatrixParallelTaskPackage));