adding generated opengl function binding
This commit is contained in:
Youmu Liu
@@ -25,293 +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.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)
|
||||
{
|
||||
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)
|
||||
{
|
||||
// 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);
|
||||
}
|
||||
if (NativeIntrinsic.Context.mat_f32_build_panel == null)
|
||||
// in this occasion size must be 1
|
||||
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
|
||||
{
|
||||
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($"{i}{j}");
|
||||
for (int k = 0; k < left.KernelColumnLength; k += size)
|
||||
{
|
||||
NativeIntrinsic.Context
|
||||
.mat_f32_panel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
|
||||
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
|
||||
(MatrixPanel*)ansPanel);
|
||||
}
|
||||
}
|
||||
int i = (int)(c / blocksPerRow);
|
||||
int j = (int)(c % blocksPerRow);
|
||||
MatrixKernel<float>* ansPanel = GetKernelAddressUnsafe((MatrixKernel<float>*)ansBuffer,
|
||||
ansHeight, i, j);
|
||||
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
|
||||
for (int k = 0; k < commonK; k += size) {
|
||||
NativeIntrinsic.Context
|
||||
.mat[f32].panel_fma(
|
||||
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer,
|
||||
ansHeight, k, j),
|
||||
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer,
|
||||
commonK, j, k),
|
||||
(MatrixPanel*)ansPanel);
|
||||
}
|
||||
} else
|
||||
}
|
||||
} else
|
||||
{
|
||||
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
|
||||
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
|
||||
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
|
||||
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
|
||||
{
|
||||
MatrixPanel<float>* leftPanel = MatrixPanel.Create<float>();
|
||||
MatrixPanel<float>* rightPanel = MatrixPanel.Create<float>();
|
||||
MatrixPanel<float>* ansPanel = MatrixPanel.Create<float>();
|
||||
for (int i = 0; i < ans.KernelRowLength; i += size)
|
||||
int i = (int)(c / blocksPerRow) * size;
|
||||
int j = (int)(c % blocksPerRow) * size;
|
||||
NativeIntrinsic.Context.mat[f32].clear_panel((MatrixPanel*)ansPanel);
|
||||
for (int k = 0; k < commonK; k += size)
|
||||
{
|
||||
for (int j = 0; j < ans.KernelColumnLength; j += size)
|
||||
{
|
||||
NativeIntrinsic.Context.mat_f32_clear_panel((MatrixPanel*)ansPanel);
|
||||
|
||||
// Console.WriteLine($"{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);
|
||||
}
|
||||
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);
|
||||
}
|
||||
MatrixPanel.Destroy(leftPanel);
|
||||
MatrixPanel.Destroy(rightPanel);
|
||||
MatrixPanel.Destroy(ansPanel);
|
||||
NativeIntrinsic.Context
|
||||
.mat[f32].store_panel((MatrixPanel*)ansPanel, ansBuffer,
|
||||
ansWidth, ansHeight, j, i);
|
||||
}
|
||||
if (count > 256) {
|
||||
MatrixParallel.ReturnParallelResources(ref pool!);
|
||||
}
|
||||
return;
|
||||
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)
|
||||
internal static unsafe void ComputeFmaPanelF64(
|
||||
MatrixParallelTaskPackage* p
|
||||
)
|
||||
{
|
||||
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)
|
||||
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)
|
||||
{
|
||||
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)
|
||||
// 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);
|
||||
}
|
||||
}
|
||||
} else
|
||||
{
|
||||
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;
|
||||
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->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)
|
||||
[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 != 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;
|
||||
return buffer + (colLength * x + y);
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user