add double compute kernel
This commit is contained in:
Youmu Liu
@@ -30,6 +30,7 @@ using StgSharp.Internal;
|
||||
using System;
|
||||
using System.Collections.Generic;
|
||||
using System.Linq;
|
||||
using System.Numerics;
|
||||
using System.Runtime.CompilerServices;
|
||||
using System.Text;
|
||||
using System.Threading.Tasks;
|
||||
@@ -39,45 +40,84 @@ namespace StgSharp.Mathematics.Numeric
|
||||
public partial class MatrixCompute
|
||||
{
|
||||
|
||||
internal static unsafe void ComputeFmaPanelF32(MatrixParallelTaskPackage* p)
|
||||
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 ansWidth = p->AnsPrimOffset;
|
||||
int ansHeight = p->AnsSecOffset;
|
||||
int commonK = p->LeftPrimOffset;
|
||||
int blocksPerRow = ansWidth / size;
|
||||
|
||||
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
|
||||
if (NativeIntrinsic.Context.mat_f32_build_panel == null)
|
||||
{
|
||||
int i = (int)(c / blocksPerRow) * size;
|
||||
int j = (int)(c % blocksPerRow) * size;
|
||||
NativeIntrinsic.Context.f32_clear_panel((MatrixPanel*)ansPanel);
|
||||
for (int k = 0; k < commonK; k += size)
|
||||
// in this occasion size must be 1
|
||||
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
|
||||
{
|
||||
NativeIntrinsic.Context
|
||||
.f32_build_panel((MatrixPanel*)leftPanel, leftBuffer, commonK,
|
||||
ansHeight, k, i);
|
||||
NativeIntrinsic.Context
|
||||
.f32_build_panel((MatrixPanel*)rightPanel, rightBuffer, ansWidth,
|
||||
commonK, j, k);
|
||||
NativeIntrinsic.Context
|
||||
.f32_panel_fma((MatrixPanel*)leftPanel, (MatrixPanel*)rightPanel, (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);
|
||||
}
|
||||
}
|
||||
NativeIntrinsic.Context
|
||||
.f32_store_panel((MatrixPanel*)ansPanel, ansBuffer, ansWidth,
|
||||
ansHeight, j, i);
|
||||
} 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++)
|
||||
{
|
||||
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.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);
|
||||
}
|
||||
MatrixPanel.Destroy(leftPanel);
|
||||
MatrixPanel.Destroy(rightPanel);
|
||||
MatrixPanel.Destroy(ansPanel);
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
internal static unsafe MatrixKernel<T>* GetKernelAddressUnsafe<T>(
|
||||
MatrixKernel<T>* buffer,
|
||||
int colLength,
|
||||
int x,
|
||||
int y
|
||||
) where T : unmanaged, INumber<T>
|
||||
{
|
||||
// Caller guarantees: indices in range, layout is dense square.
|
||||
int k_c = y >> 2;
|
||||
int k_r = x >> 2;
|
||||
|
||||
return buffer + colLength * x + y;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
@@ -77,9 +77,9 @@ namespace StgSharp.Mathematics.Numeric
|
||||
for (int kBlock = 0; kBlock < n; kBlock += 4)
|
||||
{
|
||||
NativeIntrinsic.Context
|
||||
.f32_pivot(matrix.Buffer, (ulong)(kBlock >> 2),
|
||||
(ulong)(matrix.KernelRowLength - (kBlock >> 2)),
|
||||
(ulong)pivots);
|
||||
.mat_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++)
|
||||
{
|
||||
|
||||
@@ -56,7 +56,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* rightPtr = right.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
|
||||
|
||||
return;
|
||||
} else if (count <= 1024)
|
||||
@@ -66,7 +66,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* rightPtr = right.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
|
||||
|
||||
MatrixParallel.ReturnParallelResources(ref pool);
|
||||
return;
|
||||
@@ -79,7 +79,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
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.f32_buffer_add;
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat_f32_buffer_add;
|
||||
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
|
||||
|
||||
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
|
||||
@@ -94,12 +94,12 @@ namespace StgSharp.Mathematics.Numeric
|
||||
scalar->Data<float>(0) = value;
|
||||
if (count <= 256)
|
||||
{
|
||||
NativeIntrinsic.Context.f32_buffer_fill(ans.Buffer, scalar, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_fill(ans.Buffer, scalar, count);
|
||||
return;
|
||||
} else if (count <= 1024)
|
||||
{
|
||||
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
|
||||
NativeIntrinsic.Context.f32_buffer_fill(ans.Buffer, scalar, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_fill(ans.Buffer, scalar, count);
|
||||
MatrixParallel.ReturnParallelResources(ref pool);
|
||||
return;
|
||||
}
|
||||
@@ -109,7 +109,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
package->PrimCount = ans.KernelColumnLength;
|
||||
package->SecCount = ans.KernelRowLength;
|
||||
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.ANS_SCALAR_PARAM);
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.f32_buffer_fill;
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat_f32_buffer_fill;
|
||||
package->Scalar = scalar;
|
||||
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
|
||||
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
|
||||
@@ -135,43 +135,65 @@ namespace StgSharp.Mathematics.Numeric
|
||||
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)
|
||||
if (NativeIntrinsic.Context.mat_f32_build_panel == null)
|
||||
{
|
||||
for (int j = 0; j < ans.KernelColumnLength; j += size)
|
||||
for (int i = 0; i < ans.KernelRowLength; i += size)
|
||||
{
|
||||
NativeIntrinsic.Context.f32_clear_panel((MatrixPanel*)ansPanel);
|
||||
|
||||
// Console.WriteLine($"{i}{j}");
|
||||
for (int k = 0; k < left.KernelColumnLength; k += size)
|
||||
for (int j = 0; j < ans.KernelColumnLength; j += size)
|
||||
{
|
||||
NativeIntrinsic.Context
|
||||
.f32_build_panel((MatrixPanel*)leftPanel,
|
||||
(MatrixKernel*)left.Buffer,
|
||||
left.KernelColumnLength,
|
||||
left.KernelRowLength, i, k);
|
||||
NativeIntrinsic.Context
|
||||
.f32_build_panel((MatrixPanel*)rightPanel,
|
||||
(MatrixKernel*)right.Buffer,
|
||||
right.KernelColumnLength,
|
||||
right.KernelRowLength, k, j);
|
||||
NativeIntrinsic.Context
|
||||
.f32_panel_fma((MatrixPanel*)leftPanel,
|
||||
(MatrixPanel*)rightPanel,
|
||||
(MatrixPanel*)ansPanel);
|
||||
MatrixKernel<float>* ansPanel = ans.GetKernelAddressUnsafe(i, j);
|
||||
NativeIntrinsic.Context.mat_f32_clear_panel((MatrixPanel*)ansPanel);
|
||||
|
||||
// Console.WriteLine($"{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);
|
||||
}
|
||||
}
|
||||
NativeIntrinsic.Context
|
||||
.f32_store_panel((MatrixPanel*)ansPanel,
|
||||
(MatrixKernel*)ans.Buffer,
|
||||
ans.KernelColumnLength, ans.KernelRowLength,
|
||||
i, j);
|
||||
}
|
||||
} 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);
|
||||
}
|
||||
MatrixPanel.Destroy(leftPanel);
|
||||
MatrixPanel.Destroy(rightPanel);
|
||||
MatrixPanel.Destroy(ansPanel);
|
||||
if (count > 256) {
|
||||
MatrixParallel.ReturnParallelResources(ref pool!);
|
||||
}
|
||||
@@ -211,7 +233,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
{
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
|
||||
|
||||
return;
|
||||
} else if (count <= 1024)
|
||||
@@ -220,7 +242,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
|
||||
|
||||
MatrixParallel.ReturnParallelResources(ref pool);
|
||||
return;
|
||||
@@ -233,7 +255,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
package->SecCount = left.KernelRowLength;
|
||||
package->Scalar = scalar;
|
||||
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM);
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.f32_buffer_scalar_mul;
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat_f32_buffer_scalar_mul;
|
||||
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
|
||||
|
||||
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
|
||||
@@ -261,7 +283,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* rightPtr = right.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
|
||||
|
||||
return;
|
||||
} else if (count <= 1024)
|
||||
@@ -271,7 +293,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
MatrixKernel<float>* leftPtr = left.Buffer;
|
||||
MatrixKernel<float>* rightPtr = right.Buffer;
|
||||
MatrixKernel<float>* ansPtr = ans.Buffer;
|
||||
NativeIntrinsic.Context.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
|
||||
NativeIntrinsic.Context.mat_f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
|
||||
|
||||
MatrixParallel.ReturnParallelResources(ref pool);
|
||||
return;
|
||||
@@ -284,7 +306,7 @@ namespace StgSharp.Mathematics.Numeric
|
||||
package->PrimCount = left.KernelColumnLength;
|
||||
package->SecCount = left.KernelRowLength;
|
||||
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.f32_buffer_sub;
|
||||
package->ComputeHandle = (IntPtr)NativeIntrinsic.Context.mat_f32_buffer_sub;
|
||||
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
|
||||
|
||||
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
|
||||
|
||||
Reference in New Issue
Block a user