tsai
/
ImageSharp
mirror of https://github.com/SixLabors/ImageSharp
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Don't use Vector512 (SLOW)

pull/2793/head
James Jackson-South 4 days ago
parent
1e87e4e8fa
commit
a96c78d60c
3 changed files with 30 additions and 116 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2
      README.md
  2. 34
      src/ImageSharp/Common/Helpers/Numerics.cs
  3. 110
      src/ImageSharp/Processing/Processors/Transforms/Resize/ResizeKernel.cs

2
README.md
View File

@ -8,7 +8,7 @@ SixLabors.ImageSharp
<div align="center">
[![Build Status](https://img.shields.io/github/actions/workflow/status/SixLabors/ImageSharp/build-and-test.yml?branch=main)](https://github.com/SixLabors/ImageSharp/actions)
[![Code coverage](https://codecov.io/gh/SixLabors/ImageSharp/branch/main/graph/badge.svg)](https://codecov.io/gh/SixLabors/ImageSharp)
[![codecov](https://codecov.io/gh/SixLabors/ImageSharp/graph/badge.svg?token=g2WJwz770q)](https://codecov.io/gh/SixLabors/ImageSharp)
[![License: Six Labors Split](https://img.shields.io/badge/license-Six%20Labors%20Split-%23e30183)](https://github.com/SixLabors/ImageSharp/blob/main/LICENSE)
[![Twitter](https://img.shields.io/twitter/url/http/shields.io.svg?style=flat&logo=twitter)](https://twitter.com/intent/tweet?hashtags=imagesharp,dotnet,oss&text=ImageSharp.+A+new+cross-platform+2D+graphics+API+in+C%23&url=https%3a%2f%2fgithub.com%2fSixLabors%2fImageSharp&via=sixlabors)

34
src/ImageSharp/Common/Helpers/Numerics.cs
View File

@ -1089,39 +1089,7 @@ internal static class Numerics
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void Normalize(Span<float> span, float sum)
{
if (Vector512.IsHardwareAccelerated)
{
ref float startRef = ref MemoryMarshal.GetReference(span);
ref float endRef = ref Unsafe.Add(ref startRef, span.Length & ~15);
Vector512<float> sum512 = Vector512.Create(sum);
while (Unsafe.IsAddressLessThan(ref startRef, ref endRef))
{
Unsafe.As<float, Vector512<float>>(ref startRef) /= sum512;
startRef = ref Unsafe.Add(ref startRef, (nuint)16);
}
if ((span.Length & 15) >= 8)
{
Unsafe.As<float, Vector256<float>>(ref startRef) /= sum512.GetLower();
startRef = ref Unsafe.Add(ref startRef, (nuint)8);
}
if ((span.Length & 7) >= 4)
{
Unsafe.As<float, Vector128<float>>(ref startRef) /= sum512.GetLower().GetLower();
startRef = ref Unsafe.Add(ref startRef, (nuint)4);
}
endRef = ref Unsafe.Add(ref startRef, span.Length & 3);
while (Unsafe.IsAddressLessThan(ref startRef, ref endRef))
{
startRef /= sum;
startRef = ref Unsafe.Add(ref startRef, (nuint)1);
}
}
else if (Vector256.IsHardwareAccelerated)
if (Vector256.IsHardwareAccelerated)
{
ref float startRef = ref MemoryMarshal.GetReference(span);
ref float endRef = ref Unsafe.Add(ref startRef, span.Length & ~7);

110
src/ImageSharp/Processing/Processors/Transforms/Resize/ResizeKernel.cs
View File

@ -94,97 +94,43 @@ internal readonly unsafe struct ResizeKernel
{
if (IsHardwareAccelerated)
{
if (Vector512.IsHardwareAccelerated)
{
float* bufferStart = this.bufferPtr;
ref Vector4 rowEndRef = ref Unsafe.Add(ref rowStartRef, this.Length & ~7);
Vector512<float> result512_0 = Vector512<float>.Zero;
Vector512<float> result512_1 = Vector512<float>.Zero;
while (Unsafe.IsAddressLessThan(ref rowStartRef, ref rowEndRef))
{
Vector512<float> pixels512_0 = Unsafe.As<Vector4, Vector512<float>>(ref rowStartRef);
Vector512<float> pixels512_1 = Unsafe.As<Vector4, Vector512<float>>(ref Unsafe.Add(ref rowStartRef, (nuint)4));
result512_0 = Vector512_.MultiplyAdd(result512_0, Vector512.Load(bufferStart), pixels512_0);
result512_1 = Vector512_.MultiplyAdd(result512_1, Vector512.Load(bufferStart + 16), pixels512_1);
bufferStart += 32;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)8);
}
result512_0 += result512_1;
if ((this.Length & 7) >= 4)
{
Vector512<float> pixels512_0 = Unsafe.As<Vector4, Vector512<float>>(ref rowStartRef);
result512_0 = Vector512_.MultiplyAdd(result512_0, Vector512.Load(bufferStart), pixels512_0);
bufferStart += 16;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)4);
}
Vector256<float> result256 = result512_0.GetLower() + result512_0.GetUpper();
if ((this.Length & 3) >= 2)
{
Vector256<float> pixels256_0 = Unsafe.As<Vector4, Vector256<float>>(ref rowStartRef);
result256 = Vector256_.MultiplyAdd(result256, Vector256.Load(bufferStart), pixels256_0);
bufferStart += 8;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)2);
}
Vector128<float> result128 = result256.GetLower() + result256.GetUpper();
if ((this.Length & 1) != 0)
{
Vector128<float> pixels128 = Unsafe.As<Vector4, Vector128<float>>(ref rowStartRef);
result128 = Vector128_.MultiplyAdd(result128, Vector128.Load(bufferStart), pixels128);
}
float* bufferStart = this.bufferPtr;
ref Vector4 rowEndRef = ref Unsafe.Add(ref rowStartRef, this.Length & ~3);
Vector256<float> result256_0 = Vector256<float>.Zero;
Vector256<float> result256_1 = Vector256<float>.Zero;
return result128.AsVector4();
}
else
while (Unsafe.IsAddressLessThan(ref rowStartRef, ref rowEndRef))
{
float* bufferStart = this.bufferPtr;
ref Vector4 rowEndRef = ref Unsafe.Add(ref rowStartRef, this.Length & ~3);
Vector256<float> result256_0 = Vector256<float>.Zero;
Vector256<float> result256_1 = Vector256<float>.Zero;
Vector256<float> pixels256_0 = Unsafe.As<Vector4, Vector256<float>>(ref rowStartRef);
Vector256<float> pixels256_1 = Unsafe.As<Vector4, Vector256<float>>(ref Unsafe.Add(ref rowStartRef, (nuint)2));
while (Unsafe.IsAddressLessThan(ref rowStartRef, ref rowEndRef))
{
Vector256<float> pixels256_0 = Unsafe.As<Vector4, Vector256<float>>(ref rowStartRef);
Vector256<float> pixels256_1 = Unsafe.As<Vector4, Vector256<float>>(ref Unsafe.Add(ref rowStartRef, (nuint)2));
result256_0 = Vector256_.MultiplyAdd(result256_0, Vector256.Load(bufferStart), pixels256_0);
result256_1 = Vector256_.MultiplyAdd(result256_1, Vector256.Load(bufferStart + 8), pixels256_1);
result256_0 = Vector256_.MultiplyAdd(result256_0, Vector256.Load(bufferStart), pixels256_0);
result256_1 = Vector256_.MultiplyAdd(result256_1, Vector256.Load(bufferStart + 8), pixels256_1);
bufferStart += 16;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)4);
}
result256_0 += result256_1;
bufferStart += 16;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)4);
}
if ((this.Length & 3) >= 2)
{
Vector256<float> pixels256_0 = Unsafe.As<Vector4, Vector256<float>>(ref rowStartRef);
result256_0 = Vector256_.MultiplyAdd(result256_0, Vector256.Load(bufferStart), pixels256_0);
result256_0 += result256_1;
bufferStart += 8;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)2);
}
if ((this.Length & 3) >= 2)
{
Vector256<float> pixels256_0 = Unsafe.As<Vector4, Vector256<float>>(ref rowStartRef);
result256_0 = Vector256_.MultiplyAdd(result256_0, Vector256.Load(bufferStart), pixels256_0);
Vector128<float> result128 = result256_0.GetLower() + result256_0.GetUpper();
bufferStart += 8;
rowStartRef = ref Unsafe.Add(ref rowStartRef, (nuint)2);
}
if ((this.Length & 1) != 0)
{
Vector128<float> pixels128 = Unsafe.As<Vector4, Vector128<float>>(ref rowStartRef);
result128 = Vector128_.MultiplyAdd(result128, Vector128.Load(bufferStart), pixels128);
}
Vector128<float> result128 = result256_0.GetLower() + result256_0.GetUpper();
return result128.AsVector4();
if ((this.Length & 1) != 0)
{
Vector128<float> pixels128 = Unsafe.As<Vector4, Vector128<float>>(ref rowStartRef);
result128 = Vector128_.MultiplyAdd(result128, Vector128.Load(bufferStart), pixels128);
}
return result128.AsVector4();
}
else
{
@ -219,7 +165,7 @@ internal readonly unsafe struct ResizeKernel
{
DebugGuard.IsTrue(values.Length == this.Length, nameof(values), "ResizeKernel.Fill: values.Length != this.Length!");
if (Vector256.IsHardwareAccelerated)
if (IsHardwareAccelerated)
{
Vector4* bufferStart = (Vector4*)this.bufferPtr;
ref float valuesStart = ref MemoryMarshal.GetReference(values);

Write Preview
Loading…
Cancel
Save