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

remove branch conflict

Browse Source
This commit is contained in:
Nitload
2025-12-21 23:20:56 +08:00
parent cffe44772e
commit ca5fc9a3f5
13 changed files with 17 additions and 1058 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Common/HighPerformance/MatrixKernel.cs
-23
View File
@@ -1,12 +1,4 @@
<<<<<<< HEAD
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
//-----------------------------------------------------------------------
>>>>>>> stgsharp-dev/giga
// -----------------------------------------------------------------------
// file="MatrixKernel"
// Project: StgSharp
@@ -40,10 +32,6 @@ using System.Runtime.InteropServices;
namespace StgSharp.HighPerformance
{
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
public unsafe struct MatrixKernel
{
@@ -51,20 +39,9 @@ namespace StgSharp.HighPerformance
}
<<<<<<< HEAD
[StructLayout(LayoutKind.Sequential)]
[InlineArray(16)]
public unsafe struct MatrixKernel<T> where T : unmanaged, INumber<T>
=======
[StructLayout(LayoutKind.Sequential)]
[InlineArray(16)]
public unsafe struct MatrixKernel<T> where T: unmanaged, INumber<T>
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
[StructLayout(LayoutKind.Sequential)]
[InlineArray(16)]
public unsafe struct MatrixKernel<T> where T : unmanaged, INumber<T>
>>>>>>> stgsharp-dev/giga
{
private T c0r0;
StgSharp.Common/HighPerformance/MatrixPanel.cs
-322
View File
@@ -1,12 +1,4 @@
<<<<<<< HEAD
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
//-----------------------------------------------------------------------
>>>>>>> stgsharp-dev/giga
// -----------------------------------------------------------------------
// file="MatrixPanel"
// Project: StgSharp
@@ -33,14 +25,7 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
<<<<<<< HEAD
<<<<<<< HEAD
using StgSharp.Internal;
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
using StgSharp.Internal;
>>>>>>> stgsharp-dev/giga
using System;
using System.Numerics;
using System.Runtime.CompilerServices;
@@ -48,8 +33,6 @@ using System.Runtime.InteropServices;
namespace StgSharp.HighPerformance
{
<<<<<<< HEAD
<<<<<<< HEAD
public unsafe struct MatrixPanel
{
@@ -178,308 +161,3 @@ namespace StgSharp.HighPerformance
}
}
=======
#pragma warning disable CA1000
public interface IMatrixPanel<T> : IDisposable where T : unmanaged,INumber<T>
=======
public unsafe struct MatrixPanel
>>>>>>> stgsharp-dev/giga
{
public const int AlignmentBytes = 64;
private byte Head;
// Allocate aligned memory for one panel; caller must free via Destroy.
public static unsafe MatrixPanel<T>* Create<T>() where T : unmanaged, INumber<T>
{
return (MatrixPanel<T>*)NativeMemory.AlignedAlloc(Size<T>(), AlignmentBytes);
}
public static unsafe void Destroy<T>(MatrixPanel<T>* panel) where T : unmanaged, INumber<T>
{
NativeMemory.AlignedFree(panel);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static unsafe Span<T> Elements<T>(ref T head) where T : unmanaged, INumber<T>
{
void* p = Unsafe.AsPointer(ref head);
return new Span<T>(p, MaxElementCount<T>());
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int ElementSideLength<T>() where T : unmanaged, INumber<T>
{
return Math.Max(8, AlignmentBytes / sizeof(T));
}
// View an existing pointer; caller guarantees lifetime/alignment.
public static unsafe MatrixPanel<T>* FromPointer<T>(void* ptr)
where T : unmanaged, INumber<T>
{
return (MatrixPanel<T>*)Align64(ptr);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int MaxElementCount<T>() where T : unmanaged, INumber<T>
{
int side = ElementSideLength<T>();
return side * side;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static nuint Size<T>() where T : unmanaged, INumber<T>
{
return (nuint)(MaxElementCount<T>() * sizeof(T));
}
/// <summary>
/// Number of 4x4 kernels that fit in a panel for the given ISA and element T.
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static int KernelCapacity<T>() where T : unmanaged, INumber<T>
{
int perSide = KernelSideCount<T>();
return perSide * perSide;
}
/// <summary>
/// How many 4x4 kernels fit along a single panel dimension for the given ISA.
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static int KernelSideCount<T>() where T : unmanaged, INumber<T>
{
return NativeIntrinsic.IntrinsicLevel switch
{
IntrinsicLevel.SSE => KernelSideCountSse<T>(),
IntrinsicLevel.AVX2 => KernelSideCountAvx2<T>(),
IntrinsicLevel.AVX512 => KernelSideCountAvx512<T>(),
_ => 0
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe nuint Align64(void* p) => ((nuint)p + (AlignmentBytes - 1)) & ~(nuint)(AlignmentBytes - 1);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static int KernelSideCountAvx2<T>() where T : unmanaged, INumber<T>
{
const int minQ = 8;
int lanes = 32 / sizeof(T);
int side = Math.Max(lanes, minQ);
return side / 4;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static int KernelSideCountAvx512<T>() where T : unmanaged, INumber<T>
{
const int minQ = 8;
int lanes = 64 / sizeof(T);
int side = Math.Max(lanes, minQ);
return side / 4;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static int KernelSideCountSse<T>() where T : unmanaged, INumber<T>
{
const int minQ = 4;
int lanes = 16 / sizeof(T);
int side = Math.Max(lanes, minQ);
return side / 4;
}
}
// Single POD-style panel: ZMM-compatible size. Alignment handled by allocator, not by the struct.
public unsafe struct MatrixPanel<T> where T : unmanaged, INumber<T>
{
internal T Head; // first element of the panel buffer
public ref T this[int column, int row]
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => ref Unsafe.Add(ref Head, row * MatrixPanel.ElementSideLength<T>() + column);
}
public Span<T> Elements
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => MatrixPanel.Elements(ref Head);
}
}
}
<<<<<<< HEAD
/// <summary>
/// Buffer for AVX targeted matrix panel, size of a panel is 4x4 elements, and no smaller than
/// 16 bytes.
/// </summary>
/// <typeparam name="T">
/// Type of matrix element.
/// </typeparam>
public unsafe struct MatrixPanelAVX<T> : IMatrixPanel<T> where T : unmanaged, INumber<T>
{
private T element;
[Obsolete("Compute kernel should not be created from new(), use Create instead", true)]
public MatrixPanelAVX() { }
public ref T this[int column, int row]
{
get
{
int index = row * ElementSideLength + column;
return ref Unsafe.Add(ref element, index);
}
}
public static int ElementSideLength
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get;
} = int.Max(8, 32 / sizeof(T));
public static unsafe int MaxElementCount
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get;
} = ElementSideLength * ElementSideLength;
public static unsafe nuint Size { get; } = (nuint)(MaxElementCount * sizeof(T));
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int ChunkCount() => (ElementSideLength + LanesPerVector() - 1) / LanesPerVector();
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixPanelAVX<T>* Create()
{
MatrixPanelAVX<T>* kernel = (MatrixPanelAVX<T>*)NativeMemory.AlignedAlloc(Size, 32);
return kernel;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void Destroy(MatrixPanelAVX<T>* kernel)
{
NativeMemory.AlignedFree(kernel);
}
public void Dispose()
{
fixed (MatrixPanelAVX<T>* thisPtr = &this) {
NativeMemory.AlignedFree(thisPtr);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ref M256 GetVectorChunk(int column, int chunk)
{
int lanes = LanesPerVector();
int rowStart = chunk * lanes;
ref T colBase = ref this[column, 0];
return ref Unsafe.As<T, M256>(ref Unsafe.Add(ref colBase, rowStart));
}
// SIMD access helpers for AVX (column-major)
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int LanesPerVector() => 32 / sizeof(T);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void SetVectorChunk(int column, int chunk, in M256 value)
{
GetVectorChunk(column, chunk) = value;
}
}
/// <summary>
/// Buffer for AVX512 targeted matrix panel, size of a panel is 4x4 elements, and no smaller
/// than 16 bytes.
/// </summary>
/// <typeparam name="T">
/// Type of matrix element.
/// </typeparam>
public unsafe struct MatrixComputeKernelAVX512<T> : IMatrixPanel<T>
where T : unmanaged, INumber<T>
{
private T element;
[Obsolete("Compute kernel should not be created from new(), use Create instead", true)]
public MatrixComputeKernelAVX512() { }
public ref T this[int column, int row]
{
get
{
int index = row * ElementSideLength + column;
return ref Unsafe.Add(ref element, index);
}
}
public static int ElementSideLength
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get;
} = int.Max(8, 64 / sizeof(T));
public static unsafe int MaxElementCount
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get;
} = ElementSideLength * ElementSideLength;
public static unsafe nuint Size { get; } = (nuint)(MaxElementCount * sizeof(T));
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int ChunkCount() => (ElementSideLength + LanesPerVector() - 1) / LanesPerVector();
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixComputeKernelAVX512<T>* Create()
{
MatrixComputeKernelAVX512<T>* kernel = (MatrixComputeKernelAVX512<T>*)NativeMemory.AlignedAlloc(Size, 64);
return kernel;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void Destroy(MatrixComputeKernelAVX512<T>* kernel)
{
NativeMemory.AlignedFree(kernel);
}
public void Dispose()
{
fixed (MatrixComputeKernelAVX512<T>* thisPtr = &this) {
NativeMemory.AlignedFree(thisPtr);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ref M512 GetVectorChunk(int column, int chunk)
{
int lanes = LanesPerVector();
int rowStart = chunk * lanes;
ref T colBase = ref this[column, 0];
return ref Unsafe.As<T, M512>(ref Unsafe.Add(ref colBase, rowStart));
}
// SIMD access helpers for AVX-512 (column-major)
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int LanesPerVector() => 64 / sizeof(T);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void SetVectorChunk(int column, int chunk, in M512 value)
{
GetVectorChunk(column, chunk) = value;
}
#pragma warning restore CA1000
}
}
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
StgSharp.Mathematics/Numeric/MatrixCompute.Additional.cs
-20
View File
@@ -26,17 +26,11 @@
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.HighPerformance;
<<<<<<< HEAD
using System;
using System.Collections.Generic;
using System.Linq;
=======
using StgSharp.Internal;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Runtime.CompilerServices;
>>>>>>> stgsharp-dev/giga
using System.Text;
using System.Threading.Tasks;
@@ -45,19 +39,6 @@ namespace StgSharp.Mathematics.Numeric
public partial class MatrixCompute
{
<<<<<<< HEAD
private static unsafe void ComputeFmaPanelF32(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* ans,
ScalarPacket* scalar)
{
const int eSize = 32;
int colum = scalar->Data<int>(0);
int row = scalar->Data<int>(1);
int k =
=======
internal static unsafe void ComputeFmaPanelF32(MatrixParallelTaskPackage* p)
{
int size = p->ElementSize;
@@ -97,7 +78,6 @@ namespace StgSharp.Mathematics.Numeric
MatrixPanel.Destroy(leftPanel);
MatrixPanel.Destroy(rightPanel);
MatrixPanel.Destroy(ansPanel);
>>>>>>> stgsharp-dev/giga
}
}
StgSharp.Mathematics/Numeric/MatrixCompute.cs
-230
View File
@@ -1,16 +1,4 @@
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
//-----------------------------------------------------------------------
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
>>>>>>> stgsharp-dev/giga
// -----------------------------------------------------------------------
// file="MatrixCompute"
// Project: StgSharp
@@ -37,57 +25,24 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
<<<<<<< HEAD
using System;
using System.Collections.Generic;
using System.Linq;
=======
using StgSharp.Collections;
using StgSharp.HighPerformance;
using StgSharp.Internal;
using System;
using System.Collections.Generic;
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
using System.ComponentModel.DataAnnotations.Schema;
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
using System.ComponentModel.DataAnnotations.Schema;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
using System.Linq;
using System.Numerics;
using System.Security.AccessControl;
>>>>>>> stgsharp-dev/giga
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
<<<<<<< HEAD
public static partial class MatrixCompute { }
=======
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
public static partial class MatrixCompute
{
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
=======
public static class MatrixCompute
{
public static unsafe Matrix<float> Add(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
public static partial class MatrixCompute
{
public static unsafe void Add(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
@@ -102,16 +57,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -122,49 +68,21 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_add(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
package->Right = (nint)right.Buffer;
package->Answer = (nint)ans.Buffer;
package->ElementSize = sizeof(float);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
=======
package->PrimCount = left.KernelColumnLength;
package->SecCount = left.KernelRowLength;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
package->ComputeMode = new MatrixOpMode(MatrixOperationStyle.BUFFER_OP, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
========
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationParam.LEFT_RIGHT_ANS_PARAM);
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
package->ComputeHandle = (IntPtr)NativeIntrinsic.Intrinsic.f32_buffer_add;
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
}
public static unsafe void Fill(Matrix<float> ans, float value)
@@ -197,89 +115,7 @@ namespace StgSharp.Mathematics.Numeric
taskGroup.Dispose();
}
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
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)
=======
public static unsafe Matrix<float> Mul(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
========
public static unsafe void Mul(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
{
// For matrix multiplication (left: MxK) * (right: KxN) => ans: MxN
// ColumnLength = number of columns (N), RowLength = number of rows (M)
@@ -357,12 +193,7 @@ namespace StgSharp.Mathematics.Numeric
return;
}
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
public static unsafe Matrix<float> Mul(Matrix<float> left, float right, Matrix<float> ans)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
public static unsafe void Mul(Matrix<float> left, float right, Matrix<float> ans)
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
{
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
@@ -375,16 +206,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -394,15 +216,7 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_scalar_mul(leftPtr, ansPtr, scalar, count);
MatrixParallel.ReturnParallelResources(ref pool);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Right = (nint)left.Buffer;
@@ -418,27 +232,10 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
MatrixParallelFactory.Release(scalar);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
}
public static unsafe void Sub(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
=======
return ans;
}
public static unsafe Matrix<float> Sub(
Matrix<float> left,
Matrix<float> right,
Matrix<float> ans)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
}
public static unsafe void Sub(Matrix<float> left, Matrix<float> right, Matrix<float> ans)
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
{
if (left.ColumnLength != right.ColumnLength ||
left.RowLength != right.RowLength ||
@@ -453,16 +250,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
} else if (count <= 1024)
{
MatrixParallelThread[] pool = MatrixParallel.LeadParallel(1);
@@ -473,15 +261,7 @@ namespace StgSharp.Mathematics.Numeric
NativeIntrinsic.Intrinsic.f32_buffer_sub(leftPtr, rightPtr, ansPtr, count);
MatrixParallel.ReturnParallelResources(ref pool);
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
}
MatrixParallelTaskPackage* package = MatrixParallelFactory.CreateBaseTask<float>();
package->Left = (nint)left.Buffer;
@@ -496,17 +276,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallel.LaunchParallel(taskGroup, SleepMode.DeepSleep);
taskGroup.Dispose();
<<<<<<<< HEAD:StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
<<<<<<< HEAD
return;
=======
return ans;
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
========
return;
>>>>>>>> stgsharp-dev/giga:StgSharp.Mathematics/Numeric/MatrixCompute.cs
}
}
>>>>>>> stgsharp-dev/giga
}
StgSharp.Mathematics/Numeric/MatrixComputeModel.cs
+11 -200
View File
@@ -1,12 +1,4 @@
<<<<<<< HEAD
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
//-----------------------------------------------------------------------
>>>>>>> stgsharp-dev/giga
// -----------------------------------------------------------------------
// file="MatrixComputeModel"
// Project: StgSharp
@@ -61,11 +53,11 @@ namespace StgSharp.Mathematics.Numeric
long count = Unsafe.As<int, long>(ref p->PrimCount);
long offset = Unsafe.As<int, long>(ref p->PrimTileOffset) * size;
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->Answer + offset),
scalar,
count);
((delegate* unmanaged[Cdecl]<void*, void*, void*, ScalarPacket*, long, void>)p->ComputeHandle)((void*)(p->Left + offset),
(void*)(p->Right + offset),
(void*)(p->Answer + offset),
scalar,
count);
MatrixParallelFactory.Release(p);
}
@@ -74,8 +66,8 @@ namespace StgSharp.Mathematics.Numeric
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->Answer + offset),
p->Scalar, count);
((delegate* unmanaged[Cdecl]<void*, ScalarPacket*, long, void>)p->ComputeHandle)((void*)(p->Answer + offset),
p->Scalar, count);
MatrixParallelFactory.Release(p);
}
@@ -86,8 +78,8 @@ namespace StgSharp.Mathematics.Numeric
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->Answer + offset),
count);
(void*)(p->Answer + offset),
count);
MatrixParallelFactory.Release(p);
}
@@ -96,25 +88,7 @@ namespace StgSharp.Mathematics.Numeric
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),
<<<<<<< HEAD
(void*)(p->Result + offset),
<<<<<<< HEAD
p->Scalar,
count);
=======
p->Scalar, count);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
MatrixParallelFactory.Release(p);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
<<<<<<< HEAD
internal static unsafe void KernelComputeBinary(MatrixParallelTaskPackage* p)
=======
internal static unsafe void Compute_Binary(MatrixParallelTaskPackage* p)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
((delegate* unmanaged[Cdecl]<void*, void*, ScalarPacket*, long, void>)p->ComputeHandle)((void*)(p->Right + offset),
(void*)(p->Answer + offset),
p->Scalar,
count);
@@ -122,7 +96,6 @@ namespace StgSharp.Mathematics.Numeric
}
internal static unsafe void SpecialBufferCompute(MatrixParallelTaskPackage* p)
>>>>>>> stgsharp-dev/giga
{
switch ((MatrixOperationSpecial)p->ComputeMode.ParameterStyle)
{
@@ -132,170 +105,8 @@ namespace StgSharp.Mathematics.Numeric
default:
break;
}
<<<<<<< HEAD
MatrixParallelFactory.Release(p);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
<<<<<<< HEAD
internal static unsafe void KernelComputeBinaryScalar(MatrixParallelTaskPackage* p)
=======
internal static unsafe void Compute_BinaryScalar(MatrixParallelTaskPackage* p)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
{
int eSize = p->ElementSize;
// base pointer to matrix kernels
nint l = p->Left, r = p->Right, a = p->Result;
long
// set beginning value to counters for primary offsets
lCount = (long)p->LeftPrimOffset * p->LeftPrimStride,
rCount = (long)p->RightPrimOffset * p->RightPrimStride,
aCount = (long)p->ResultPrimOffset * p->ResultPrimStride,
pCount = int.Min(p->PrimCount, p->PrimColumnCountInTile + p->PrimTileOffset) /*_count of tile may be too much*/,
//edge to reset counters
lLimit = pCount - p->LeftPrimOffset,
rLimit = pCount - p->RightPrimOffset,
aLimit = pCount - p->ResultPrimOffset;
delegate* unmanaged[Cdecl]<IntPtr, IntPtr, IntPtr, ScalarPacket*, void> op =
(delegate* unmanaged[Cdecl]<IntPtr, IntPtr, IntPtr, ScalarPacket*, void>)p->ComputeHandle;
for (long i = p->PrimTileOffset % pCount; i < pCount; i++)
{
// update primary counters
// if reaches edge, reset to zero
lCount = i < lLimit ? lCount + p->LeftPrimStride : 0;
rCount = i < rLimit ? rCount + p->RightPrimStride : 0;
aCount = i < rLimit ? aCount + p->ResultPrimStride : 0;
// get kernel column pointers
nint left = l + (nint)(lCount * eSize);
nint right = r + (nint)(rCount * eSize);
nint ans = a + (nint)(aCount * eSize);
long
lc = (long)p->LeftSecOffset * p->LeftSecStride,
rc = (long)p->RightSecOffset * p->RightSecStride,
ac = (long)p->ResultSecOffset * p->ResultSecStride,
sCount = int.Min(p->SecCount, p->PrimColumnCountInTile * p->SecTileOffset);
long
ll = sCount - p->LeftSecOffset,
rl = sCount - p->RightSecOffset,
al = sCount - p->ResultSecOffset;
for (long j = p->SecTileOffset % sCount; j < sCount; j++)
{
lc = j < ll ? lc + p->LeftSecStride : 0;
rc = j < rl ? rc + p->RightSecStride : 0;
ac = j < al ? ac + p->ResultSecStride : 0;
op(left + (nint)lc, right + (nint)rc, ans + (nint)ac, p->Scalar);
}
}
MatrixParallelFactory.Release(p);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
<<<<<<< HEAD
internal static unsafe void KernelComputeUnary(MatrixParallelTaskPackage* p)
=======
internal static unsafe void Compute_Unary(MatrixParallelTaskPackage* p)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
{
int eSize = p->ElementSize;
// base pointer to matrix kernels
nint r = p->Right, a = p->Result;
long
// set beginning value to counters for primary offsets
rCount = (long)p->RightPrimOffset * p->RightPrimStride,
aCount = (long)p->ResultPrimOffset * p->ResultPrimStride,
pCount = int.Min(p->PrimCount, p->PrimColumnCountInTile + p->PrimTileOffset) /*_count of tile may be too much*/,
//edge to reset counters
rLimit = pCount - p->RightPrimOffset,
aLimit = pCount - p->ResultPrimOffset;
delegate* unmanaged[Cdecl]<IntPtr, IntPtr, void> op =
(delegate* unmanaged[Cdecl]<IntPtr, IntPtr, void>)p->ComputeHandle;
for (long i = p->PrimTileOffset % pCount; i < pCount; i++)
{
// update primary counters
// if reaches edge, reset to zero
rCount = i < rLimit ? rCount + p->RightPrimStride : 0;
aCount = i < rLimit ? aCount + p->ResultPrimStride : 0;
// get kernel column pointers
nint right = r + (nint)(rCount * eSize);
nint ans = a + (nint)(aCount * eSize);
long
rc = (long)p->RightSecOffset * p->RightSecStride,
ac = (long)p->ResultSecOffset * p->ResultSecStride,
sCount = int.Min(p->SecCount, p->PrimColumnCountInTile * p->SecTileOffset);
long
rl = sCount - p->RightSecOffset,
al = sCount - p->ResultSecOffset;
for (long j = p->SecTileOffset % sCount; j < sCount; j++)
{
rc = j < rl ? rc + p->RightSecStride : 0;
ac = j < al ? ac + p->ResultSecStride : 0;
op(right + (nint)rc, ans + (nint)ac);
}
}
MatrixParallelFactory.Release(p);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
<<<<<<< HEAD
internal static unsafe void KernelComputeUnaryScalar(MatrixParallelTaskPackage* p)
=======
internal static unsafe void Compute_UnaryScalar(MatrixParallelTaskPackage* p)
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
{
int eSize = p->ElementSize;
// base pointer to matrix kernels
nint r = p->Right, a = p->Result;
long
// set beginning value to counters for primary offsets
rCount = (long)p->RightPrimOffset * p->RightPrimStride,
aCount = (long)p->ResultPrimOffset * p->ResultPrimStride,
pCount = int.Min(p->PrimCount, p->PrimColumnCountInTile + p->PrimTileOffset) /*_count of tile may be too much*/,
//edge to reset counters
rLimit = pCount - p->RightPrimOffset,
aLimit = pCount - p->ResultPrimOffset;
delegate* unmanaged[Cdecl]<IntPtr, IntPtr, ScalarPacket*, void> op =
(delegate* unmanaged[Cdecl]<IntPtr, IntPtr, ScalarPacket*, void>)p->ComputeHandle;
for (long i = p->PrimTileOffset % pCount; i < pCount; i++)
{
// update primary counters
// if reaches edge, reset to zero
rCount = i < rLimit ? rCount + p->RightPrimStride : 0;
aCount = i < rLimit ? aCount + p->ResultPrimStride : 0;
// get kernel column pointers
nint right = r + (nint)(rCount * eSize);
nint ans = a + (nint)(aCount * eSize);
long
rc = (long)p->RightSecOffset * p->RightSecStride,
ac = (long)p->ResultSecOffset * p->ResultSecStride,
sCount = int.Min(p->SecCount, p->PrimColumnCountInTile * p->SecTileOffset);
long
rl = sCount - p->RightSecOffset,
al = sCount - p->ResultSecOffset;
for (long j = p->SecTileOffset % sCount; j < sCount; j++)
{
rc = j < rl ? rc + p->RightSecStride : 0;
ac = j < al ? ac + p->ResultSecStride : 0;
op(right + (nint)rc, ans + (nint)ac, p->Scalar);
}
}
MatrixParallelFactory.Release(p);
=======
>>>>>>> stgsharp-dev/giga
}
<<<<<<< HEAD
internal static unsafe void PanelComputeBinary(MatrixParallelTaskPackage* p) { }
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
}
}
}
StgSharp.Mathematics/Numeric/MatrixParallel.cs
+1 -32
View File
@@ -1,12 +1,4 @@
<<<<<<< HEAD
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
//-----------------------------------------------------------------------
>>>>>>> stgsharp-dev/giga
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixParallel"
// Project: StgSharp
@@ -35,20 +27,6 @@
// -----------------------------------------------------------------------
using StgSharp.Collections;
using System;
<<<<<<< HEAD
<<<<<<< HEAD
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static class MatrixParallel
=======
using System.Collections.Concurrent;
=======
>>>>>>> stgsharp-dev/giga
using System.Collections.Generic;
using System.Linq;
using System.Text;
@@ -57,7 +35,6 @@ using System.Threading.Tasks;
namespace StgSharp.Mathematics.Numeric
{
public static partial class MatrixParallel
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
{
private static bool _meaningful;
@@ -119,12 +96,4 @@ namespace StgSharp.Mathematics.Numeric
}
}
<<<<<<< HEAD
<<<<<<< HEAD
}
=======
}
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
}
>>>>>>> stgsharp-dev/giga
StgSharp.Mathematics/Numeric/MatrixParallelThread.cs
-53
View File
@@ -1,12 +1,4 @@
<<<<<<< HEAD
<<<<<<< HEAD
//-----------------------------------------------------------------------
=======
//-----------------------------------------------------------------------
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
//-----------------------------------------------------------------------
>>>>>>> stgsharp-dev/giga
// -----------------------------------------------------------------------
// file="MatrixParallelThread"
// Project: StgSharp
@@ -120,41 +112,6 @@ namespace StgSharp.Mathematics.Numeric
break;
}
break;
<<<<<<< HEAD
case MatrixOperationStyle.KERNEL_OP:
switch (p->ComputeMode.ParameterStyle)
{
case MatrixOperationParam.RIGHT_ANS_PARAM:
<<<<<<< HEAD
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.Compute_Unary(p);
break;
case MatrixOperationParam.RIGHT_ANS_SCALAR_PARAM:
MatrixComputeModel.Compute_UnaryScalar(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_PARAM:
MatrixComputeModel.Compute_Binary(p);
break;
case MatrixOperationParam.LEFT_RIGHT_ANS_SCALAR_PARAM:
MatrixComputeModel.Compute_BinaryScalar(p);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
break;
default:
break;
}
break;
=======
>>>>>>> stgsharp-dev/giga
default:
break;
}
@@ -426,19 +383,9 @@ namespace StgSharp.Mathematics.Numeric
private static extern IntPtr GetCurrentThread();
[DllImport("kernel32.dll")]
<<<<<<< HEAD
<<<<<<< HEAD
private static extern nuint SetThreadAffinityMask(
IntPtr hThread,
nuint dwThreadAffinityMask);
=======
private static extern nuint SetThreadAffinityMask(IntPtr hThread, nuint dwThreadAffinityMask);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
private static extern nuint SetThreadAffinityMask(
IntPtr hThread,
nuint dwThreadAffinityMask);
>>>>>>> stgsharp-dev/giga
[LibraryImport("kernel32.dll")]
[return: MarshalAs(UnmanagedType.Bool)]
StgSharp.Native/sn-head/sn_intrinsic.h
-39
View File
@@ -141,34 +141,14 @@ typedef void(SN_DECL *PNL_PROC_LRAS)(mat_panel *left, mat_panel *right, mat_pane
typedef void(SN_DECL *PNL_PROC_RAS)(mat_panel *right, mat_panel *ans, scalar_pack *scalar);
#define PNL_PROC(T_type, T_arch, T_op_name) T_type##_pnl_##T_op_name##_##T_arch
<<<<<<< HEAD
<<<<<<< HEAD
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) *restrict ans)
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) *restrict ans)
>>>>>>> stgsharp-dev/giga
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
MAT_PANEL_##T_arch##(T_type) *restrict ans)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
MAT_PANEL_##T_arch##(T_type) *restrict ans)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_op_name)( \
MAT_PANEL_##T_arch##(T_type) const *left, \
MAT_PANEL_##T_arch##(T_type) const *right, \
MAT_PANEL_##T_arch##(T_type) *restrict ans, scalar_pack const *scalar)
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, T_op_name)( \
MAT_PANEL_##T_arch##(T_type) const *left, \
@@ -249,11 +229,6 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill);
#pragma endregion
#pragma region fma
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_ANS(float, sse, clear);
DECLARE_PNL_PROC_ANS(float, avx, clear);
DECLARE_PNL_PROC_ANS(float, 512, clear);
@@ -262,13 +237,6 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, fma);
<<<<<<< HEAD
=======
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
#pragma endregion
@@ -310,14 +278,7 @@ typedef struct sn_intrinsic {
BUF_PROC_LRA f32_buffer_sub;
KER_PROC_RA f32_buffer_transpose;
PANEL_PROC f32_build_panel;
<<<<<<< HEAD
<<<<<<< HEAD
PANEL_PROC f32_clear_panel;
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
PANEL_PROC f32_clear_panel;
>>>>>>> stgsharp-dev/giga
VECTORNORMALIZEPROC f32_normalize_3;
PNL_PROC_LRA f32_panel_fma;
PANEL_PROC f32_store_panel;
StgSharp.Native/sn-head/sn_matkernel.h
-14
View File
@@ -117,26 +117,12 @@ extern __declspec(align(64)) sn_zero_kernel_u64 SN_ZERO_KERNEL;
int col_index, int row_index)
/* Declarations for AVX float variant */
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_BUILD_PANEL(sse, float);
DECLARE_BUILD_PANEL(avx, float);
DECLARE_BUILD_PANEL(512, float);
DECLARE_STORE_PANEL(sse, float);
DECLARE_STORE_PANEL(avx, float);
DECLARE_STORE_PANEL(512, float);
=======
=======
DECLARE_BUILD_PANEL(sse, float);
>>>>>>> stgsharp-dev/giga
DECLARE_BUILD_PANEL(avx, float);
DECLARE_BUILD_PANEL(512, float);
DECLARE_STORE_PANEL(sse, float);
DECLARE_STORE_PANEL(avx, float);
<<<<<<< HEAD
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
DECLARE_STORE_PANEL(512, float);
>>>>>>> stgsharp-dev/giga
#endif /* SN_MATKERNEL */
StgSharp.Native/sn-src/k_basicop.c
+4 -27
View File
@@ -1,52 +1,29 @@
#include "sn_intrinsic.h"
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_ANS(float, sse, clear)
{
register __m128 zero = _mm_setzero_ps();
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
for (size_t i = 0; i < 4; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
DECLARE_PNL_PROC_ANS(float, avx, clear)
{
register __m256 zero = _mm256_setzero_ps();
<<<<<<< HEAD
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
=======
for (size_t i = 0; i < 8; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
>>>>>>> stgsharp-dev/giga
}
DECLARE_PNL_PROC_ANS(float, 512, clear)
{
register __m512 zero = _mm512_setzero_ps();
<<<<<<< HEAD
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
}
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
>>>>>>> stgsharp-dev/giga
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose)
{
StgSharp.Native/sn-src/k_fma.c
-9
View File
@@ -54,10 +54,6 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
// ans = left * right + ans
// right is not transposed
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
@@ -101,11 +97,6 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
// ans = left * right + ans
// right is not transposed
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
StgSharp.Native/sn-src/k_panel.c
+1 -53
View File
@@ -41,18 +41,10 @@ DECLARE_BUILD_PANEL(avx, float)
}
}
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_BUILD_PANEL(sse, float)
{
/* Single kernel -> 4 columns, each is one __m128 column vector. */
const MAT_KERNEL(float) *src = matrix + (col_index * row_length + row_index);
=======
DECLARE_BUILD_PANEL(sse, float)
{
/* Single kernel -> 4 columns, each is one __m128 column vector. */
const MAT_KERNEL(float) *src = matrix + (row_index * col_length + col_index);
>>>>>>> stgsharp-dev/giga
for (int pc = 0; pc < 4; ++pc) {
PANEL_COL_VEC((*panel), pc, 0) = src->xmm[pc];
@@ -62,35 +54,22 @@ DECLARE_BUILD_PANEL(sse, float)
DECLARE_BUILD_PANEL(512, float)
{
const MAT_KERNEL(float) *const zero = (const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
<<<<<<< HEAD
const MAT_KERNEL(float) *k[4][4];
=======
const MAT_KERNEL(float) * k[4][4];
>>>>>>> stgsharp-dev/giga
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
const int col = col_index + kc;
const int row = row_index + kr;
k[kc][kr] = (col < col_length && row < row_length) ?
<<<<<<< HEAD
(matrix + (col * row_length + row)) :
=======
(matrix + (row * col_length + col)) :
>>>>>>> stgsharp-dev/giga
zero;
}
}
/* Each panel column corresponds to a kernel column (0..3) within a block col (0..3). */
for (int pc = 0; pc < 16; ++pc) {
<<<<<<< HEAD
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
=======
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
>>>>>>> stgsharp-dev/giga
const __m128 r0 = k[blk_col][0]->xmm[kc_col];
const __m128 r1 = k[blk_col][1]->xmm[kc_col];
@@ -107,11 +86,6 @@ DECLARE_BUILD_PANEL(512, float)
}
}
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
DECLARE_STORE_PANEL(avx, float)
{
const int Q = 8; /* fixed for AVX float panel */
@@ -167,18 +141,10 @@ DECLARE_STORE_PANEL(avx, float)
k_lb->xmm[kc_col] = bot128;
}
}
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_STORE_PANEL(sse, float)
{
MAT_KERNEL(float) *dst = matrix + (col_index * row_length + row_index);
=======
DECLARE_STORE_PANEL(sse, float)
{
MAT_KERNEL(float) *dst = matrix + (row_index * col_length + col_index);
>>>>>>> stgsharp-dev/giga
for (int pc = 0; pc < 4; ++pc) {
const __m128 col = PANEL_COL_VEC((*panel), pc, 0);
@@ -189,34 +155,21 @@ DECLARE_STORE_PANEL(sse, float)
DECLARE_STORE_PANEL(512, float)
{
MAT_KERNEL(float) zero;
<<<<<<< HEAD
MAT_KERNEL(float) *k[4][4];
=======
MAT_KERNEL(float) * k[4][4];
>>>>>>> stgsharp-dev/giga
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
const int col = col_index + kc;
const int row = row_index + kr;
k[kc][kr] = (col < col_length && row < row_length) ?
<<<<<<< HEAD
(matrix + (col * row_length + row)) :
=======
(matrix + (row * col_length + col)) :
>>>>>>> stgsharp-dev/giga
&zero;
}
}
for (int pc = 0; pc < 16; ++pc) {
<<<<<<< HEAD
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
=======
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
>>>>>>> stgsharp-dev/giga
const __m512 col = PANEL_COL_VEC((*panel), pc, 0);
@@ -234,8 +187,3 @@ DECLARE_STORE_PANEL(512, float)
k[blk_col][3]->xmm[kc_col] = r3;
}
}
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
StgSharp.Native/sn-src/load_intrinsic.c
-36
View File
@@ -65,23 +65,11 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(sse, float);
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = PNL_PROC(float, sse, fma);
=======
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(sse, float);
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(sse, float);
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = PNL_PROC(float, sse, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, sse, fill);
context->variant = BUF_PROC(float, sse, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, sse, sub);
@@ -95,23 +83,11 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, avx, clear);
context->f32_build_panel = BUILD_PANEL(avx, float);
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = PNL_PROC(float, avx, fma);
=======
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(avx, float);
context->f32_clear_panel = PNL_PROC(float, avx, clear);
context->f32_build_panel = BUILD_PANEL(avx, float);
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = PNL_PROC(float, avx, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, avx, fill);
context->variant = BUF_PROC(float, avx, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, avx, sub);
@@ -125,23 +101,11 @@ void load_intrinsic_512(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(512, float);
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = PNL_PROC(float, 512, fma);
=======
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(512, float);
context->f32_clear_panel = PNL_PROC(float, 512, clear);
context->f32_build_panel = BUILD_PANEL(512, float);
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = PNL_PROC(float, 512, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, 512, fill);
context->variant = BUF_PROC(float, 512, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, 512, sub);