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

Break out cgbi helper methods

pull/3137/head
Erik White 2 months ago
parent
fdc15d283e
commit
a68921cac3
3 changed files with 501 additions and 303 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 325
      src/ImageSharp/Formats/Png/PngCgbiProcessor.cs
  2. 305
      src/ImageSharp/Formats/Png/PngDecoderCore.cs
  3. 174
      tests/ImageSharp.Tests/Formats/Png/PngCgbiProcessorTests.cs

325
src/ImageSharp/Formats/Png/PngCgbiProcessor.cs
View File

@ -0,0 +1,325 @@
// Copyright (c) Six Labors.
// Licensed under the Six Labors Split License.
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
using SixLabors.ImageSharp.Common.Helpers;
using SixLabors.ImageSharp.PixelFormats;
using static SixLabors.ImageSharp.SimdUtils;
namespace SixLabors.ImageSharp.Formats.Png;
/// <summary>
/// Reverses the pixel mangling applied by Apple's CgBI PNG variant. CgBI files
/// (emitted by <c>pngcrush -iphone</c>) swap channel order from RGB(A) to BGR(A)
/// and premultiply RGB samples by alpha. This converts a defiltered scanline back
/// to standard PNG semantics in place so the existing scanline processors can
/// consume it unchanged. CgBI is only emitted for 8-bit truecolor (with or
/// without alpha); other color types are left alone.
/// </summary>
/// <remarks>
/// See https://theapplewiki.com/wiki/PNG_CgBI_Format
/// </remarks>
internal static class PngCgbiProcessor
{
// Per-pixel byte indices that swap CgBI's BGRA layout to Rgba32's RGBA.
// MMShuffle3012 expands to [2, 1, 0, 3] per 4-byte pixel; the same 64-byte
// sequence seeds all three shuffle masks (V128/V256 ctors take the leading
// 16/32 bytes).
private static readonly Vector128<byte> BgraToRgbaShuffle128 = Vector128.Create(BuildShuffleBytes());
private static readonly Vector256<byte> BgraToRgbaShuffle256 = Vector256.Create(BuildShuffleBytes());
private static readonly Vector512<byte> BgraToRgbaShuffle512 = Vector512.Create(BuildShuffleBytes());
/// <summary>
/// Applies the inverse of Apple's CgBI pixel mangling to a defiltered scanline in place.
/// </summary>
/// <param name="configuration">The configuration used by the Rgb24 R/B swap.</param>
/// <param name="scanline">The defiltered pixel bytes (without the leading filter byte).</param>
/// <param name="colorType">The PNG color type from IHDR.</param>
public static void ApplyTransform(Configuration configuration, Span<byte> scanline, PngColorType colorType)
{
if (colorType == PngColorType.RgbWithAlpha)
{
Span<Rgba32> pixels = MemoryMarshal.Cast<byte, Rgba32>(scanline);
int i = 0;
// Avx512BW is required for the per-byte vpshufb. Without it, ShuffleNative
// falls back to a software cross-lane shuffle that is slower than the V256
// path, so skip V512 entirely on Avx512F-only hosts.
if (Vector512.IsHardwareAccelerated && Avx512BW.IsSupported && pixels.Length >= 16)
{
i = ApplyTransformVector512(scanline, pixels.Length);
}
if (Vector256.IsHardwareAccelerated && Avx2.IsSupported && (pixels.Length - i) >= 8)
{
i = ApplyTransformVector256(scanline, i, pixels.Length);
}
if (Vector128.IsHardwareAccelerated && (pixels.Length - i) >= 4)
{
i = ApplyTransformVector128(scanline, i, pixels.Length);
}
for (; i < pixels.Length; i++)
{
ref Rgba32 pixel = ref pixels[i];
pixel = new Rgba32(pixel.B, pixel.G, pixel.R, pixel.A);
UndoPremultiplicationScalar(ref pixel);
}
}
else if (colorType == PngColorType.Rgb)
{
// No alpha channel, so just swap R and B using built in SIMD-optimized pixel operations.
Span<Rgb24> target = MemoryMarshal.Cast<byte, Rgb24>(scanline);
PixelOperations<Rgb24>.Instance.FromBgr24Bytes(configuration, scanline, target, target.Length);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void UndoPremultiplicationScalar(ref Rgba32 pixel)
{
byte a = pixel.A;
if (a is 0 or byte.MaxValue)
{
return;
}
// Reverse: c' = c * a / 255 => c = round(c' * 255 / a)
int half = a >> 1;
byte r = (byte)Math.Min(byte.MaxValue, ((pixel.R * byte.MaxValue) + half) / a);
byte g = (byte)Math.Min(byte.MaxValue, ((pixel.G * byte.MaxValue) + half) / a);
byte b = (byte)Math.Min(byte.MaxValue, ((pixel.B * byte.MaxValue) + half) / a);
pixel = new Rgba32(r, g, b, a);
}
internal static int ApplyTransformVector512(Span<byte> scanline, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = 0;
// The mask only swaps bytes inside each 4-byte pixel, so it is correct
// for the per-lane Avx512BW.Shuffle that ShuffleNative selects here.
Vector512<byte> shuffleMask = BgraToRgbaShuffle512;
Vector512<int> zero = Vector512<int>.Zero;
Vector512<int> one = Vector512<int>.One;
Vector512<int> byteMask = Vector512.Create(0xFF);
Vector512<int> opaque = Vector512.Create(0xFF);
Vector512<int> byteMax = Vector512.Create((int)byte.MaxValue);
for (; i <= pixelCount - 16; i += 16)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector512<byte> bgra = Unsafe.ReadUnaligned<Vector512<byte>>(ref blockRef);
Vector512<byte> rgba = Vector512_.ShuffleNative(bgra, shuffleMask);
Vector512<int> packed = rgba.AsInt32();
Vector512<int> alpha = Vector512.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector512<int> partialMask = ~(Vector512.Equals(alpha, zero) | Vector512.Equals(alpha, opaque));
Vector512<int> r = packed & byteMask;
Vector512<int> g = Vector512.ShiftRightLogical(packed, 8) & byteMask;
Vector512<int> b = Vector512.ShiftRightLogical(packed, 16) & byteMask;
Vector512<int> safeAlpha = Vector512.ConditionalSelect(partialMask, alpha, one);
Vector512<int> halfAlpha = Vector512.ShiftRightLogical(safeAlpha, 1);
Vector512<float> safeAlphaF = Vector512.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector512<int> unpremultipliedR = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector512<int> unpremultipliedG = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector512<int> unpremultipliedB = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector512<int> finalR = Vector512.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector512<int> finalG = Vector512.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector512<int> finalB = Vector512.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector512<int> result =
finalR |
Vector512.ShiftLeft(finalG, 8) |
Vector512.ShiftLeft(finalB, 16) |
Vector512.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
internal static int ApplyTransformVector256(Span<byte> scanline, int startPixel, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = startPixel;
// vpshufb is 128-bit lane-local and uses only the low 4 bits of each
// index, so the same per-pixel [2,1,0,3] pattern in both lanes keeps
// every byte inside its own lane.
Vector256<byte> shuffleMask = BgraToRgbaShuffle256;
Vector256<int> zero = Vector256<int>.Zero;
Vector256<int> one = Vector256<int>.One;
Vector256<int> byteMask = Vector256.Create(0xFF);
Vector256<int> opaque = Vector256.Create(0xFF);
Vector256<int> byteMax = Vector256.Create((int)byte.MaxValue);
for (; i <= pixelCount - 8; i += 8)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector256<byte> bgra = Unsafe.ReadUnaligned<Vector256<byte>>(ref blockRef);
Vector256<byte> rgba = Vector256_.ShufflePerLane(bgra, shuffleMask);
Vector256<int> packed = rgba.AsInt32();
Vector256<int> alpha = Vector256.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector256<int> partialMask = ~(Vector256.Equals(alpha, zero) | Vector256.Equals(alpha, opaque));
Vector256<int> r = packed & byteMask;
Vector256<int> g = Vector256.ShiftRightLogical(packed, 8) & byteMask;
Vector256<int> b = Vector256.ShiftRightLogical(packed, 16) & byteMask;
Vector256<int> safeAlpha = Vector256.ConditionalSelect(partialMask, alpha, one);
Vector256<int> halfAlpha = Vector256.ShiftRightLogical(safeAlpha, 1);
Vector256<float> safeAlphaF = Vector256.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector256<int> unpremultipliedR = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector256<int> unpremultipliedG = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector256<int> unpremultipliedB = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector256<int> finalR = Vector256.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector256<int> finalG = Vector256.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector256<int> finalB = Vector256.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector256<int> result =
finalR |
Vector256.ShiftLeft(finalG, 8) |
Vector256.ShiftLeft(finalB, 16) |
Vector256.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
internal static int ApplyTransformVector128(Span<byte> scanline, int startPixel, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = startPixel;
Vector128<byte> shuffleMask = BgraToRgbaShuffle128;
Vector128<int> zero = Vector128<int>.Zero;
Vector128<int> one = Vector128<int>.One;
Vector128<int> byteMask = Vector128.Create(0xFF);
Vector128<int> opaque = Vector128.Create(0xFF);
Vector128<int> byteMax = Vector128.Create((int)byte.MaxValue);
for (; i <= pixelCount - 4; i += 4)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector128<byte> bgra = Unsafe.ReadUnaligned<Vector128<byte>>(ref blockRef);
Vector128<byte> rgba = Vector128_.ShuffleNative(bgra, shuffleMask);
Vector128<int> packed = rgba.AsInt32();
Vector128<int> alpha = Vector128.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector128<int> partialMask = ~(Vector128.Equals(alpha, zero) | Vector128.Equals(alpha, opaque));
Vector128<int> r = packed & byteMask;
Vector128<int> g = Vector128.ShiftRightLogical(packed, 8) & byteMask;
Vector128<int> b = Vector128.ShiftRightLogical(packed, 16) & byteMask;
Vector128<int> safeAlpha = Vector128.ConditionalSelect(partialMask, alpha, one);
Vector128<int> halfAlpha = Vector128.ShiftRightLogical(safeAlpha, 1);
Vector128<float> safeAlphaF = Vector128.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector128<int> unpremultipliedR = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector128<int> unpremultipliedG = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector128<int> unpremultipliedB = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector128<int> finalR = Vector128.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector128<int> finalG = Vector128.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector128<int> finalB = Vector128.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector128<int> result =
finalR |
Vector128.ShiftLeft(finalG, 8) |
Vector128.ShiftLeft(finalB, 16) |
Vector128.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
private static byte[] BuildShuffleBytes()
{
byte[] bytes = new byte[64];
Span<byte> span = bytes;
Shuffle.MMShuffleSpan(ref span, Shuffle.MMShuffle3012);
return bytes;
}
}

305
src/ImageSharp/Formats/Png/PngDecoderCore.cs
View File

@ -9,8 +9,6 @@ using System.IO.Compression;
using System.IO.Hashing; using System.IO.Hashing;
using System.Runtime.CompilerServices; using System.Runtime.CompilerServices;
using System.Runtime.InteropServices; using System.Runtime.InteropServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
using System.Text; using System.Text;
using SixLabors.ImageSharp.Common.Helpers; using SixLabors.ImageSharp.Common.Helpers;
using SixLabors.ImageSharp.Compression.Zlib; using SixLabors.ImageSharp.Compression.Zlib;
@ -922,7 +920,7 @@ internal sealed class PngDecoderCore : ImageDecoderCore
if (this.isCgbi) if (this.isCgbi)
{ {
this.ApplyCgbiTransform(scanSpan[1..], this.pngColorType); PngCgbiProcessor.ApplyTransform(this.configuration, scanSpan[1..], this.pngColorType);
} }
this.ProcessDefilteredScanline(frameControl, currentRow, scanSpan, imageFrame, pngMetadata, blendRowBuffer); this.ProcessDefilteredScanline(frameControl, currentRow, scanSpan, imageFrame, pngMetadata, blendRowBuffer);
@ -1057,7 +1055,7 @@ internal sealed class PngDecoderCore : ImageDecoderCore
if (this.isCgbi) if (this.isCgbi)
{ {
this.ApplyCgbiTransform(scanSpan[1..], this.pngColorType); PngCgbiProcessor.ApplyTransform(this.configuration, scanSpan[1..], this.pngColorType);
} }
Span<TPixel> rowSpan = imageBuffer.DangerousGetRowSpan(currentRow); Span<TPixel> rowSpan = imageBuffer.DangerousGetRowSpan(currentRow);
@ -2529,303 +2527,4 @@ internal sealed class PngDecoderCore : ImageDecoderCore
private void SwapScanlineBuffers() private void SwapScanlineBuffers()
=> (this.scanline, this.previousScanline) = (this.previousScanline, this.scanline); => (this.scanline, this.previousScanline) = (this.previousScanline, this.scanline);
/// <summary>
/// Applies the inverse of Apple's CgBI pixel mangling to a defiltered scanline.
/// CgBI PNGs are emitted by <c>pngcrush -iphone</c> with channel order swapped
/// from RGB(A) to BGR(A) and RGB samples premultiplied by alpha. This converts
/// the bytes back to standard PNG semantics in place so the existing scanline
/// processors can consume them unchanged. CgBI is only emitted for 8-bit
/// truecolor (with or without alpha); other color types are left alone.
/// </summary>
/// <remarks>
/// See https://theapplewiki.com/wiki/PNG_CgBI_Format
/// </remarks>
/// <param name="scanline">The defiltered pixel bytes (without the leading filter byte).</param>
/// <param name="colorType">The PNG color type from IHDR.</param>
private void ApplyCgbiTransform(Span<byte> scanline, PngColorType colorType)
{
if (colorType == PngColorType.RgbWithAlpha)
{
Span<Rgba32> pixels = MemoryMarshal.Cast<byte, Rgba32>(scanline);
int i = 0;
if (Vector512.IsHardwareAccelerated && pixels.Length >= 16)
{
i = ApplyCgbiTransformVector512(scanline, pixels.Length);
}
if (Vector256.IsHardwareAccelerated && Avx2.IsSupported && (pixels.Length - i) >= 8)
{
i = ApplyCgbiTransformVector256(scanline, i, pixels.Length);
}
if (Vector128.IsHardwareAccelerated && (pixels.Length - i) >= 4)
{
i = ApplyCgbiTransformVector128(scanline, i, pixels.Length);
}
for (; i < pixels.Length; i++)
{
ref Rgba32 pixel = ref pixels[i];
pixel = new Rgba32(pixel.B, pixel.G, pixel.R, pixel.A);
UndoCgbiPremultiplicationScalar(ref pixel);
}
}
else if (colorType == PngColorType.Rgb)
{
// No alpha channel, so just swap R and B using built in SIMD-optimized pixel operations.
Span<Rgb24> target = MemoryMarshal.Cast<byte, Rgb24>(scanline);
PixelOperations<Rgb24>.Instance.FromBgr24Bytes(this.configuration, scanline, target, target.Length);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void UndoCgbiPremultiplicationScalar(ref Rgba32 pixel)
{
byte a = pixel.A;
if (a is 0 or byte.MaxValue)
{
return;
}
// Reverse: c' = c * a / 255 => c = round(c' * 255 / a)
int half = a >> 1;
byte r = (byte)Math.Min(byte.MaxValue, ((pixel.R * byte.MaxValue) + half) / a);
byte g = (byte)Math.Min(byte.MaxValue, ((pixel.G * byte.MaxValue) + half) / a);
byte b = (byte)Math.Min(byte.MaxValue, ((pixel.B * byte.MaxValue) + half) / a);
pixel = new Rgba32(r, g, b, a);
}
private static int ApplyCgbiTransformVector512(Span<byte> scanline, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = 0;
Span<byte> temp = stackalloc byte[Vector512<byte>.Count];
SimdUtils.Shuffle.MMShuffleSpan(ref temp, SimdUtils.Shuffle.MMShuffle3012);
// MMShuffle3012 expands to [2, 1, 0, 3] for each 4-byte pixel, converting
// CgBI's BGRA byte order to Rgba32's RGBA layout while keeping alpha in place.
// The generated mask only swaps bytes inside each pixel, so it remains
// correct for the optimized 512-bit byte shuffle helper.
Vector512<byte> shuffleMask = Unsafe.As<byte, Vector512<byte>>(ref MemoryMarshal.GetReference(temp));
Vector512<int> zero = Vector512<int>.Zero;
Vector512<int> one = Vector512<int>.One;
Vector512<int> byteMask = Vector512.Create(0xFF);
Vector512<int> opaque = Vector512.Create(0xFF);
Vector512<int> byteMax = Vector512.Create((int)byte.MaxValue);
for (; i <= pixelCount - 16; i += 16)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector512<byte> bgra = Unsafe.ReadUnaligned<Vector512<byte>>(ref blockRef);
Vector512<byte> rgba = Vector512_.ShuffleNative(bgra, shuffleMask);
Vector512<int> packed = rgba.AsInt32();
Vector512<int> alpha = Vector512.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector512<int> partialMask = ~(Vector512.Equals(alpha, zero) | Vector512.Equals(alpha, opaque));
Vector512<int> r = packed & byteMask;
Vector512<int> g = Vector512.ShiftRightLogical(packed, 8) & byteMask;
Vector512<int> b = Vector512.ShiftRightLogical(packed, 16) & byteMask;
Vector512<int> safeAlpha = Vector512.ConditionalSelect(partialMask, alpha, one);
Vector512<int> halfAlpha = Vector512.ShiftRightLogical(safeAlpha, 1);
Vector512<float> safeAlphaF = Vector512.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector512<int> unpremultipliedR = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector512<int> unpremultipliedG = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector512<int> unpremultipliedB = Vector512.Min(
byteMax,
Vector512.ConvertToInt32(Vector512.Floor(Vector512.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector512<int> finalR = Vector512.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector512<int> finalG = Vector512.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector512<int> finalB = Vector512.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector512<int> result =
finalR |
Vector512.ShiftLeft(finalG, 8) |
Vector512.ShiftLeft(finalB, 16) |
Vector512.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
private static int ApplyCgbiTransformVector256(Span<byte> scanline, int startPixel, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = startPixel;
Span<byte> temp = stackalloc byte[Vector512<byte>.Count];
SimdUtils.Shuffle.MMShuffleSpan(ref temp, SimdUtils.Shuffle.MMShuffle3012);
// MMShuffle3012 expands to [2, 1, 0, 3] for each 4-byte pixel, converting
// CgBI's BGRA byte order to Rgba32's RGBA layout while keeping alpha in place.
// Avx2.Shuffle is 128-bit lane-local, and the generated mask repeats inside
// each lane, so no byte ever needs to cross the lane boundary.
Vector256<byte> shuffleMask = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(temp));
Vector256<int> zero = Vector256<int>.Zero;
Vector256<int> one = Vector256<int>.One;
Vector256<int> byteMask = Vector256.Create(0xFF);
Vector256<int> opaque = Vector256.Create(0xFF);
Vector256<int> byteMax = Vector256.Create((int)byte.MaxValue);
for (; i <= pixelCount - 8; i += 8)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector256<byte> bgra = Unsafe.ReadUnaligned<Vector256<byte>>(ref blockRef);
Vector256<byte> rgba = Vector256_.ShufflePerLane(bgra, shuffleMask);
Vector256<int> packed = rgba.AsInt32();
Vector256<int> alpha = Vector256.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector256<int> partialMask = ~(Vector256.Equals(alpha, zero) | Vector256.Equals(alpha, opaque));
Vector256<int> r = packed & byteMask;
Vector256<int> g = Vector256.ShiftRightLogical(packed, 8) & byteMask;
Vector256<int> b = Vector256.ShiftRightLogical(packed, 16) & byteMask;
Vector256<int> safeAlpha = Vector256.ConditionalSelect(partialMask, alpha, one);
Vector256<int> halfAlpha = Vector256.ShiftRightLogical(safeAlpha, 1);
Vector256<float> safeAlphaF = Vector256.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector256<int> unpremultipliedR = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector256<int> unpremultipliedG = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector256<int> unpremultipliedB = Vector256.Min(
byteMax,
Vector256.ConvertToInt32(Vector256.Floor(Vector256.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector256<int> finalR = Vector256.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector256<int> finalG = Vector256.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector256<int> finalB = Vector256.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector256<int> result =
finalR |
Vector256.ShiftLeft(finalG, 8) |
Vector256.ShiftLeft(finalB, 16) |
Vector256.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
private static int ApplyCgbiTransformVector128(Span<byte> scanline, int startPixel, int pixelCount)
{
ref byte scanlineRef = ref MemoryMarshal.GetReference(scanline);
int i = startPixel;
Span<byte> temp = stackalloc byte[Vector512<byte>.Count];
SimdUtils.Shuffle.MMShuffleSpan(ref temp, SimdUtils.Shuffle.MMShuffle3012);
// MMShuffle3012 expands to [2, 1, 0, 3] for each 4-byte pixel, converting
// CgBI's BGRA byte order to Rgba32's RGBA layout while keeping alpha in place.
Vector128<byte> shuffleMask = Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(temp));
Vector128<int> zero = Vector128<int>.Zero;
Vector128<int> one = Vector128<int>.One;
Vector128<int> byteMask = Vector128.Create(0xFF);
Vector128<int> opaque = Vector128.Create(0xFF);
Vector128<int> byteMax = Vector128.Create((int)byte.MaxValue);
for (; i <= pixelCount - 4; i += 4)
{
ref byte blockRef = ref Unsafe.Add(ref scanlineRef, i * Unsafe.SizeOf<Rgba32>());
Vector128<byte> bgra = Unsafe.ReadUnaligned<Vector128<byte>>(ref blockRef);
Vector128<byte> rgba = Vector128_.ShuffleNative(bgra, shuffleMask);
Vector128<int> packed = rgba.AsInt32();
Vector128<int> alpha = Vector128.ShiftRightLogical(packed, 24);
// Fully transparent and fully opaque pixels are identity cases for
// unpremultiplication. Masking them keeps the scalar behavior and lets
// safeAlpha avoid dividing by zero for alpha == 0.
Vector128<int> partialMask = ~(Vector128.Equals(alpha, zero) | Vector128.Equals(alpha, opaque));
Vector128<int> r = packed & byteMask;
Vector128<int> g = Vector128.ShiftRightLogical(packed, 8) & byteMask;
Vector128<int> b = Vector128.ShiftRightLogical(packed, 16) & byteMask;
Vector128<int> safeAlpha = Vector128.ConditionalSelect(partialMask, alpha, one);
Vector128<int> halfAlpha = Vector128.ShiftRightLogical(safeAlpha, 1);
Vector128<float> safeAlphaF = Vector128.ConvertToSingle(safeAlpha);
// The scalar path computes ((c * 255) + (a >> 1)) / a with integer
// division. Floor the positive quotient before converting so SIMD does
// not use the default round-to-nearest conversion and drift by one.
Vector128<int> unpremultipliedR = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((r * byteMax) + halfAlpha) / safeAlphaF)));
Vector128<int> unpremultipliedG = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((g * byteMax) + halfAlpha) / safeAlphaF)));
Vector128<int> unpremultipliedB = Vector128.Min(
byteMax,
Vector128.ConvertToInt32(Vector128.Floor(Vector128.ConvertToSingle((b * byteMax) + halfAlpha) / safeAlphaF)));
// ConditionalSelect applies the expensive unpremultiply only to pixels
// where alpha is between 1 and 254; alpha 0 and 255 lanes keep the
// shuffled channel values exactly as the scalar path does.
Vector128<int> finalR = Vector128.ConditionalSelect(partialMask, unpremultipliedR, r);
Vector128<int> finalG = Vector128.ConditionalSelect(partialMask, unpremultipliedG, g);
Vector128<int> finalB = Vector128.ConditionalSelect(partialMask, unpremultipliedB, b);
// Rgba32 is laid out as little-endian 0xAABBGGRR in an int lane, so
// shifting the unpacked channels back to byte offsets 0, 1, 2, and 3
// recreates the in-memory RGBA bytes for the unaligned store.
Vector128<int> result =
finalR |
Vector128.ShiftLeft(finalG, 8) |
Vector128.ShiftLeft(finalB, 16) |
Vector128.ShiftLeft(alpha, 24);
Unsafe.WriteUnaligned(ref blockRef, result.AsByte());
}
return i;
}
} }

174
tests/ImageSharp.Tests/Formats/Png/PngCgbiProcessorTests.cs
View File

@ -0,0 +1,174 @@
// Copyright (c) Six Labors.
// Licensed under the Six Labors Split License.
using System.Runtime.InteropServices;
using SixLabors.ImageSharp.Formats.Png;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Tests.Formats.Png;
[Trait("Format", "Png")]
public class PngCgbiProcessorTests
{
[Theory]
[InlineData(0)]
[InlineData(1)]
[InlineData(3)]
[InlineData(4)]
[InlineData(7)]
[InlineData(8)]
[InlineData(15)]
[InlineData(16)]
[InlineData(17)]
[InlineData(31)]
[InlineData(32)]
[InlineData(33)]
[InlineData(64)]
public void ApplyTransform_RgbWithAlpha_MatchesScalar(int pixelCount)
{
// Drives the full V512/V256/V128/scalar dispatch, so it covers each
// path that is hardware-accelerated on the host plus the scalar tail.
byte[] input = CreateBgraScanline(pixelCount);
byte[] processorOutput = (byte[])input.Clone();
byte[] scalarOutput = (byte[])input.Clone();
PngCgbiProcessor.ApplyTransform(Configuration.Default, processorOutput, PngColorType.RgbWithAlpha);
ApplyCgbiTransformScalarReference(scalarOutput);
Assert.Equal(scalarOutput, processorOutput);
}
[Theory]
[InlineData(0)]
[InlineData(1)]
[InlineData(3)]
[InlineData(4)]
[InlineData(7)]
[InlineData(8)]
[InlineData(15)]
[InlineData(16)]
[InlineData(17)]
[InlineData(31)]
[InlineData(32)]
[InlineData(33)]
[InlineData(64)]
public void ApplyTransformVector512_MatchesScalar(int pixelCount) =>
// Vector512 uses Vector512_.ShuffleNative which falls back to the software
// Vector512.Shuffle when Avx512BW is unavailable, so the body runs regardless
// of whether Vector512 is hardware-accelerated on the host.
AssertVectorMatchesScalar(
pixelCount,
scanline => PngCgbiProcessor.ApplyTransformVector512(scanline, scanline.Length / 4),
blockSize: 16);
[Theory]
[InlineData(0)]
[InlineData(1)]
[InlineData(3)]
[InlineData(4)]
[InlineData(7)]
[InlineData(8)]
[InlineData(15)]
[InlineData(16)]
[InlineData(17)]
[InlineData(31)]
[InlineData(32)]
[InlineData(64)]
public void ApplyTransformVector256_MatchesScalar(int pixelCount) => AssertVectorMatchesScalar(
pixelCount,
scanline => PngCgbiProcessor.ApplyTransformVector256(scanline, 0, scanline.Length / 4),
blockSize: 8);
[Theory]
[InlineData(0)]
[InlineData(1)]
[InlineData(3)]
[InlineData(4)]
[InlineData(7)]
[InlineData(8)]
[InlineData(15)]
[InlineData(16)]
[InlineData(64)]
public void ApplyTransformVector128_MatchesScalar(int pixelCount) => AssertVectorMatchesScalar(
pixelCount,
scanline => PngCgbiProcessor.ApplyTransformVector128(scanline, 0, scanline.Length / 4),
blockSize: 4);
private static void AssertVectorMatchesScalar(int pixelCount, Func<byte[], int> applyVector, int blockSize)
{
byte[] input = CreateBgraScanline(pixelCount);
byte[] vectorOutput = (byte[])input.Clone();
byte[] scalarOutput = (byte[])input.Clone();
int processed = applyVector(vectorOutput);
int expectedProcessed = (pixelCount / blockSize) * blockSize;
Assert.Equal(expectedProcessed, processed);
// The vector path is responsible for whole blocks only; remaining pixels are
// handled by the scalar tail in ApplyTransform. Run the scalar reference
// over every pixel and compare the prefix the vector path actually wrote.
ApplyCgbiTransformScalarReference(scalarOutput);
Span<byte> vectorProcessed = vectorOutput.AsSpan(0, processed * 4);
Span<byte> scalarProcessed = scalarOutput.AsSpan(0, processed * 4);
Assert.True(vectorProcessed.SequenceEqual(scalarProcessed), $"Mismatch at pixelCount={pixelCount}");
// Pixels past the vector's processed prefix must be untouched.
Span<byte> vectorTail = vectorOutput.AsSpan(processed * 4);
Span<byte> inputTail = input.AsSpan(processed * 4);
Assert.True(vectorTail.SequenceEqual(inputTail));
}
private static byte[] CreateBgraScanline(int pixelCount)
{
// Deterministic mix of edge cases (a=0, a=255, partial alpha) and varied channels.
byte[] bytes = new byte[pixelCount * 4];
for (int p = 0; p < pixelCount; p++)
{
byte a = (p % 7) switch
{
0 => byte.MinValue,
1 => byte.MaxValue,
_ => (byte)((((p * 37) + 23) & 0xFF) | 1) // never zero
};
// CgBI premultiplied BGRA: c' = c * a / 255
byte r = (byte)((p * 13) & 0xFF);
byte g = (byte)((p * 29) & 0xFF);
byte b = (byte)((p * 53) & 0xFF);
r = (byte)((r * a) / byte.MaxValue);
g = (byte)((g * a) / byte.MaxValue);
b = (byte)((b * a) / byte.MaxValue);
bytes[(p * 4) + 0] = b;
bytes[(p * 4) + 1] = g;
bytes[(p * 4) + 2] = r;
bytes[(p * 4) + 3] = a;
}
return bytes;
}
private static void ApplyCgbiTransformScalarReference(Span<byte> scanline)
{
Span<Rgba32> pixels = MemoryMarshal.Cast<byte, Rgba32>(scanline);
for (int i = 0; i < pixels.Length; i++)
{
ref Rgba32 pixel = ref pixels[i];
pixel = new Rgba32(pixel.B, pixel.G, pixel.R, pixel.A);
byte a = pixel.A;
if (a is 0 or byte.MaxValue)
{
continue;
}
int half = a >> 1;
byte r = (byte)Math.Min(byte.MaxValue, ((pixel.R * byte.MaxValue) + half) / a);
byte g = (byte)Math.Min(byte.MaxValue, ((pixel.G * byte.MaxValue) + half) / a);
byte b = (byte)Math.Min(byte.MaxValue, ((pixel.B * byte.MaxValue) + half) / a);
pixel = new Rgba32(r, g, b, a);
}
}
}
Write Preview
Loading…
Cancel
Save