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

Define sse and avx masks as static readonly

pull/1799/head
Brian Popow 5 years ago
parent
ed8d2afcb0
commit
15a10126d2
3 changed files with 93 additions and 71 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. 65
      src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
  2. 43
      src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
  3. 56
      src/ImageSharp/Formats/Webp/WebpCommonUtils.cs

65
src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
View File

@ -27,6 +27,30 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
private const double Log2Reciprocal = 1.44269504088896338700465094007086; private const double Log2Reciprocal = 1.44269504088896338700465094007086;
#if SUPPORTS_RUNTIME_INTRINSICS
private static readonly Vector256<byte> AddGreenToBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
private static readonly Vector128<byte> AddGreenToBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
private static readonly byte AddGreenToBlueAndRedShuffleMask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
private static readonly Vector256<byte> SubtractGreenFromBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
private static readonly Vector128<byte> SubtractGreenFromBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
private static readonly byte SubtractGreenFromBlueAndRedShuffleMask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
private static readonly Vector128<byte> TransformColorAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
private static readonly Vector128<byte> TransformColorRedBlueMask = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
private static readonly byte TransformColorShuffleMask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
private static readonly Vector128<byte> TransformColorInverseAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
private static readonly byte TransformColorInverseShuffleMask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
#endif
/// <summary> /// <summary>
/// Returns the exact index where array1 and array2 are different. For an index /// Returns the exact index where array1 and array2 are different. For an index
/// inferior or equal to bestLenMatch, the return value just has to be strictly /// inferior or equal to bestLenMatch, the return value just has to be strictly
@ -97,7 +121,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
#if SUPPORTS_RUNTIME_INTRINSICS #if SUPPORTS_RUNTIME_INTRINSICS
if (Avx2.IsSupported) if (Avx2.IsSupported)
{ {
var mask = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -106,7 +129,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* idx = p + i; uint* idx = p + i;
Vector256<byte> input = Avx.LoadVector256((ushort*)idx).AsByte(); Vector256<byte> input = Avx.LoadVector256((ushort*)idx).AsByte();
Vector256<byte> in0g0g = Avx2.Shuffle(input, mask); Vector256<byte> in0g0g = Avx2.Shuffle(input, AddGreenToBlueAndRedMaskAvx2);
Vector256<byte> output = Avx2.Add(input, in0g0g); Vector256<byte> output = Avx2.Add(input, in0g0g);
Avx.Store((byte*)idx, output); Avx.Store((byte*)idx, output);
} }
@ -119,7 +142,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
} }
else if (Ssse3.IsSupported) else if (Ssse3.IsSupported)
{ {
var mask = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -128,7 +150,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* idx = p + i; uint* idx = p + i;
Vector128<byte> input = Sse2.LoadVector128((ushort*)idx).AsByte(); Vector128<byte> input = Sse2.LoadVector128((ushort*)idx).AsByte();
Vector128<byte> in0g0g = Ssse3.Shuffle(input, mask); Vector128<byte> in0g0g = Ssse3.Shuffle(input, AddGreenToBlueAndRedMaskSsse3);
Vector128<byte> output = Sse2.Add(input, in0g0g); Vector128<byte> output = Sse2.Add(input, in0g0g);
Sse2.Store((byte*)idx, output.AsByte()); Sse2.Store((byte*)idx, output.AsByte());
} }
@ -141,7 +163,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
} }
else if (Sse2.IsSupported) else if (Sse2.IsSupported)
{ {
byte mask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -151,8 +172,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
uint* idx = p + i; uint* idx = p + i;
Vector128<ushort> input = Sse2.LoadVector128((ushort*)idx); Vector128<ushort> input = Sse2.LoadVector128((ushort*)idx);
Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g
Vector128<ushort> b = Sse2.ShuffleLow(a, mask); Vector128<ushort> b = Sse2.ShuffleLow(a, AddGreenToBlueAndRedShuffleMask);
Vector128<ushort> c = Sse2.ShuffleHigh(b, mask); // 0g0g Vector128<ushort> c = Sse2.ShuffleHigh(b, AddGreenToBlueAndRedShuffleMask); // 0g0g
Vector128<byte> output = Sse2.Add(input.AsByte(), c.AsByte()); Vector128<byte> output = Sse2.Add(input.AsByte(), c.AsByte());
Sse2.Store((byte*)idx, output); Sse2.Store((byte*)idx, output);
} }
@ -189,7 +210,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
#if SUPPORTS_RUNTIME_INTRINSICS #if SUPPORTS_RUNTIME_INTRINSICS
if (Avx2.IsSupported) if (Avx2.IsSupported)
{ {
var mask = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -198,7 +218,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* idx = p + i; uint* idx = p + i;
Vector256<byte> input = Avx.LoadVector256((ushort*)idx).AsByte(); Vector256<byte> input = Avx.LoadVector256((ushort*)idx).AsByte();
Vector256<byte> in0g0g = Avx2.Shuffle(input, mask); Vector256<byte> in0g0g = Avx2.Shuffle(input, SubtractGreenFromBlueAndRedMaskAvx2);
Vector256<byte> output = Avx2.Subtract(input, in0g0g); Vector256<byte> output = Avx2.Subtract(input, in0g0g);
Avx.Store((byte*)idx, output); Avx.Store((byte*)idx, output);
} }
@ -211,7 +231,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
} }
else if (Ssse3.IsSupported) else if (Ssse3.IsSupported)
{ {
var mask = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -220,7 +239,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* idx = p + i; uint* idx = p + i;
Vector128<byte> input = Sse2.LoadVector128((ushort*)idx).AsByte(); Vector128<byte> input = Sse2.LoadVector128((ushort*)idx).AsByte();
Vector128<byte> in0g0g = Ssse3.Shuffle(input, mask); Vector128<byte> in0g0g = Ssse3.Shuffle(input, SubtractGreenFromBlueAndRedMaskSsse3);
Vector128<byte> output = Sse2.Subtract(input, in0g0g); Vector128<byte> output = Sse2.Subtract(input, in0g0g);
Sse2.Store((byte*)idx, output.AsByte()); Sse2.Store((byte*)idx, output.AsByte());
} }
@ -233,7 +252,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
} }
else if (Sse2.IsSupported) else if (Sse2.IsSupported)
{ {
byte mask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
int numPixels = pixelData.Length; int numPixels = pixelData.Length;
fixed (uint* p = pixelData) fixed (uint* p = pixelData)
{ {
@ -243,8 +261,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
uint* idx = p + i; uint* idx = p + i;
Vector128<ushort> input = Sse2.LoadVector128((ushort*)idx); Vector128<ushort> input = Sse2.LoadVector128((ushort*)idx);
Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g
Vector128<ushort> b = Sse2.ShuffleLow(a, mask); Vector128<ushort> b = Sse2.ShuffleLow(a, SubtractGreenFromBlueAndRedShuffleMask);
Vector128<ushort> c = Sse2.ShuffleHigh(b, mask); // 0g0g Vector128<ushort> c = Sse2.ShuffleHigh(b, SubtractGreenFromBlueAndRedShuffleMask); // 0g0g
Vector128<byte> output = Sse2.Subtract(input.AsByte(), c.AsByte()); Vector128<byte> output = Sse2.Subtract(input.AsByte(), c.AsByte());
Sse2.Store((byte*)idx, output); Sse2.Store((byte*)idx, output);
} }
@ -394,9 +412,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue)); Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0); Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0);
var maskalphagreen = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
var maskredblue = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
byte shufflemask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
fixed (uint* src = data) fixed (uint* src = data)
{ {
int idx; int idx;
@ -404,15 +419,15 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* pos = src + idx; uint* pos = src + idx;
Vector128<uint> input = Sse2.LoadVector128(pos); Vector128<uint> input = Sse2.LoadVector128(pos);
Vector128<byte> a = Sse2.And(input.AsByte(), maskalphagreen); Vector128<byte> a = Sse2.And(input.AsByte(), TransformColorAlphaGreenMask);
Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), shufflemask); Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), TransformColorShuffleMask);
Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), shufflemask); Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), TransformColorShuffleMask);
Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16()); Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
Vector128<short> e = Sse2.ShiftLeftLogical(input.AsInt16(), 8); Vector128<short> e = Sse2.ShiftLeftLogical(input.AsInt16(), 8);
Vector128<short> f = Sse2.MultiplyHigh(e.AsInt16(), multsb2.AsInt16()); Vector128<short> f = Sse2.MultiplyHigh(e.AsInt16(), multsb2.AsInt16());
Vector128<int> g = Sse2.ShiftRightLogical(f.AsInt32(), 16); Vector128<int> g = Sse2.ShiftRightLogical(f.AsInt32(), 16);
Vector128<byte> h = Sse2.Add(g.AsByte(), d.AsByte()); Vector128<byte> h = Sse2.Add(g.AsByte(), d.AsByte());
Vector128<byte> i = Sse2.And(h, maskredblue); Vector128<byte> i = Sse2.And(h, TransformColorRedBlueMask);
Vector128<byte> output = Sse2.Subtract(input.AsByte(), i); Vector128<byte> output = Sse2.Subtract(input.AsByte(), i);
Sse2.Store((byte*)pos, output); Sse2.Store((byte*)pos, output);
} }
@ -460,8 +475,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue)); Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0); Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0);
var maskalphagreen = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
byte shufflemask = SimdUtils.Shuffle.MmShuffle(2, 2, 0, 0);
fixed (uint* src = pixelData) fixed (uint* src = pixelData)
{ {
int idx; int idx;
@ -469,9 +482,9 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
{ {
uint* pos = src + idx; uint* pos = src + idx;
Vector128<uint> input = Sse2.LoadVector128(pos); Vector128<uint> input = Sse2.LoadVector128(pos);
Vector128<byte> a = Sse2.And(input.AsByte(), maskalphagreen); Vector128<byte> a = Sse2.And(input.AsByte(), TransformColorInverseAlphaGreenMask);
Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), shufflemask); Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), TransformColorInverseShuffleMask);
Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), shufflemask); Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), TransformColorInverseShuffleMask);
Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16()); Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
Vector128<byte> e = Sse2.Add(input.AsByte(), d.AsByte()); Vector128<byte> e = Sse2.Add(input.AsByte(), d.AsByte());
Vector128<short> f = Sse2.ShiftLeftLogical(e.AsInt16(), 8); Vector128<short> f = Sse2.ShiftLeftLogical(e.AsInt16(), 8);

43
src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
View File

@ -36,6 +36,22 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
private const int PredLowEffort = 11; private const int PredLowEffort = 11;
#if SUPPORTS_RUNTIME_INTRINSICS
private static readonly Vector128<byte> CollectColorRedTransformsGreenMask = Vector128.Create(0x00ff00).AsByte();
private static readonly Vector128<byte> CollectColorRedTransformsAndMask = Vector128.Create((short)0xff).AsByte();
private static readonly Vector128<byte> CollectColorBlueTransformsGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
private static readonly Vector128<byte> CollectColorBlueTransformsGreenBlueMask = Vector128.Create(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0);
private static readonly Vector128<byte> CollectColorBlueTransformsBlueMask = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
private static readonly Vector128<byte> CollectColorBlueTransformsShuffleLowMask = Vector128.Create(255, 2, 255, 6, 255, 10, 255, 14, 255, 255, 255, 255, 255, 255, 255, 255);
private static readonly Vector128<byte> CollectColorBlueTransformsShuffleHighMask = Vector128.Create(255, 255, 255, 255, 255, 255, 255, 255, 255, 2, 255, 6, 255, 10, 255, 14);
#endif
/// <summary> /// <summary>
/// Finds the best predictor for each tile, and converts the image to residuals /// Finds the best predictor for each tile, and converts the image to residuals
/// with respect to predictions. If nearLosslessQuality &lt; 100, applies /// with respect to predictions. If nearLosslessQuality &lt; 100, applies
@ -1039,9 +1055,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
if (Sse41.IsSupported) if (Sse41.IsSupported)
{ {
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToRed)); var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToRed));
var maskgreen = Vector128.Create(0x00ff00);
var mask = Vector128.Create((short)0xff);
const int span = 8; const int span = 8;
Span<ushort> values = stackalloc ushort[span]; Span<ushort> values = stackalloc ushort[span];
for (int y = 0; y < tileHeight; y++) for (int y = 0; y < tileHeight; y++)
@ -1057,15 +1070,15 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
uint* input1Idx = src + x + (span / 2); uint* input1Idx = src + x + (span / 2);
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte();
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte();
Vector128<byte> g0 = Sse2.And(input0, maskgreen.AsByte()); // 0 0 | g 0 Vector128<byte> g0 = Sse2.And(input0, CollectColorRedTransformsGreenMask); // 0 0 | g 0
Vector128<byte> g1 = Sse2.And(input1, maskgreen.AsByte()); Vector128<byte> g1 = Sse2.And(input1, CollectColorRedTransformsGreenMask);
Vector128<ushort> g = Sse41.PackUnsignedSaturate(g0.AsInt32(), g1.AsInt32()); // g 0 Vector128<ushort> g = Sse41.PackUnsignedSaturate(g0.AsInt32(), g1.AsInt32()); // g 0
Vector128<int> a0 = Sse2.ShiftRightLogical(input0.AsInt32(), 16); // 0 0 | x r Vector128<int> a0 = Sse2.ShiftRightLogical(input0.AsInt32(), 16); // 0 0 | x r
Vector128<int> a1 = Sse2.ShiftRightLogical(input1.AsInt32(), 16); Vector128<int> a1 = Sse2.ShiftRightLogical(input1.AsInt32(), 16);
Vector128<ushort> a = Sse41.PackUnsignedSaturate(a0, a1); // x r Vector128<ushort> a = Sse41.PackUnsignedSaturate(a0, a1); // x r
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); // x dr Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); // x dr
Vector128<byte> c = Sse2.Subtract(a.AsByte(), b.AsByte()); // x r' Vector128<byte> c = Sse2.Subtract(a.AsByte(), b.AsByte()); // x r'
Vector128<byte> d = Sse2.And(c, mask.AsByte()); // 0 r' Vector128<byte> d = Sse2.And(c, CollectColorRedTransformsAndMask); // 0 r'
Sse2.Store(dst, d.AsUInt16()); Sse2.Store(dst, d.AsUInt16());
for (int i = 0; i < span; i++) for (int i = 0; i < span; i++)
{ {
@ -1113,12 +1126,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
Span<ushort> values = stackalloc ushort[span]; Span<ushort> values = stackalloc ushort[span];
var multsr = Vector128.Create(LosslessUtils.Cst5b(redToBlue)); var multsr = Vector128.Create(LosslessUtils.Cst5b(redToBlue));
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToBlue)); var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToBlue));
var maskgreen = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
var maskgreenblue = Vector128.Create(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0);
var maskblue = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
var shufflerLow = Vector128.Create(255, 2, 255, 6, 255, 10, 255, 14, 255, 255, 255, 255, 255, 255, 255, 255);
var shufflerHigh = Vector128.Create(255, 255, 255, 255, 255, 255, 255, 255, 255, 2, 255, 6, 255, 10, 255, 14);
for (int y = 0; y < tileHeight; y++) for (int y = 0; y < tileHeight; y++)
{ {
Span<uint> srcSpan = bgra.Slice(y * stride); Span<uint> srcSpan = bgra.Slice(y * stride);
@ -1132,18 +1139,18 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
uint* input1Idx = src + x + (span / 2); uint* input1Idx = src + x + (span / 2);
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte();
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte();
Vector128<byte> r0 = Ssse3.Shuffle(input0, shufflerLow); Vector128<byte> r0 = Ssse3.Shuffle(input0, CollectColorBlueTransformsShuffleLowMask);
Vector128<byte> r1 = Ssse3.Shuffle(input1, shufflerHigh); Vector128<byte> r1 = Ssse3.Shuffle(input1, CollectColorBlueTransformsShuffleHighMask);
Vector128<byte> r = Sse2.Or(r0, r1); Vector128<byte> r = Sse2.Or(r0, r1);
Vector128<byte> gb0 = Sse2.And(input0, maskgreenblue); Vector128<byte> gb0 = Sse2.And(input0, CollectColorBlueTransformsGreenBlueMask);
Vector128<byte> gb1 = Sse2.And(input1, maskgreenblue); Vector128<byte> gb1 = Sse2.And(input1, CollectColorBlueTransformsGreenBlueMask);
Vector128<ushort> gb = Sse41.PackUnsignedSaturate(gb0.AsInt32(), gb1.AsInt32()); Vector128<ushort> gb = Sse41.PackUnsignedSaturate(gb0.AsInt32(), gb1.AsInt32());
Vector128<byte> g = Sse2.And(gb.AsByte(), maskgreen); Vector128<byte> g = Sse2.And(gb.AsByte(), CollectColorBlueTransformsGreenMask);
Vector128<short> a = Sse2.MultiplyHigh(r.AsInt16(), multsr); Vector128<short> a = Sse2.MultiplyHigh(r.AsInt16(), multsr);
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg);
Vector128<byte> c = Sse2.Subtract(gb.AsByte(), b.AsByte()); Vector128<byte> c = Sse2.Subtract(gb.AsByte(), b.AsByte());
Vector128<byte> d = Sse2.Subtract(c, a.AsByte()); Vector128<byte> d = Sse2.Subtract(c, a.AsByte());
Vector128<byte> e = Sse2.And(d, maskblue); Vector128<byte> e = Sse2.And(d, CollectColorBlueTransformsBlueMask);
Sse2.Store(dst, e.AsUInt16()); Sse2.Store(dst, e.AsUInt16());
for (int i = 0; i < span; i++) for (int i = 0; i < span; i++)
{ {

56
src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
View File

@ -16,6 +16,16 @@ namespace SixLabors.ImageSharp.Formats.Webp
/// </summary> /// </summary>
internal static class WebpCommonUtils internal static class WebpCommonUtils
{ {
#if SUPPORTS_RUNTIME_INTRINSICS
private static readonly Vector256<byte> AlphaMaskVector256 = Vector256.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
private static readonly Vector256<byte> All0x80Vector256 = Vector256.Create((byte)0x80).AsByte();
private static readonly Vector128<byte> AlphaMask = Vector128.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
private static readonly Vector128<byte> All0x80 = Vector128.Create((byte)0x80).AsByte();
#endif
/// <summary> /// <summary>
/// Checks if the pixel row is not opaque. /// Checks if the pixel row is not opaque.
/// </summary> /// </summary>
@ -27,11 +37,6 @@ namespace SixLabors.ImageSharp.Formats.Webp
if (Avx2.IsSupported) if (Avx2.IsSupported)
{ {
ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row); ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
var alphaMaskVector256 = Vector256.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
Vector256<byte> all0x80Vector256 = Vector256.Create((byte)0x80).AsByte();
var alphaMask = Vector128.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
Vector128<byte> all0x80 = Vector128.Create((byte)0x80).AsByte();
int i = 0; int i = 0;
int length = (row.Length * 4) - 3; int length = (row.Length * 4) - 3;
fixed (byte* src = rowBytes) fixed (byte* src = rowBytes)
@ -42,14 +47,14 @@ namespace SixLabors.ImageSharp.Formats.Webp
Vector256<byte> a1 = Avx.LoadVector256(src + i + 32).AsByte(); Vector256<byte> a1 = Avx.LoadVector256(src + i + 32).AsByte();
Vector256<byte> a2 = Avx.LoadVector256(src + i + 64).AsByte(); Vector256<byte> a2 = Avx.LoadVector256(src + i + 64).AsByte();
Vector256<byte> a3 = Avx.LoadVector256(src + i + 96).AsByte(); Vector256<byte> a3 = Avx.LoadVector256(src + i + 96).AsByte();
Vector256<int> b0 = Avx2.And(a0, alphaMaskVector256).AsInt32(); Vector256<int> b0 = Avx2.And(a0, AlphaMaskVector256).AsInt32();
Vector256<int> b1 = Avx2.And(a1, alphaMaskVector256).AsInt32(); Vector256<int> b1 = Avx2.And(a1, AlphaMaskVector256).AsInt32();
Vector256<int> b2 = Avx2.And(a2, alphaMaskVector256).AsInt32(); Vector256<int> b2 = Avx2.And(a2, AlphaMaskVector256).AsInt32();
Vector256<int> b3 = Avx2.And(a3, alphaMaskVector256).AsInt32(); Vector256<int> b3 = Avx2.And(a3, AlphaMaskVector256).AsInt32();
Vector256<short> c0 = Avx2.PackSignedSaturate(b0, b1).AsInt16(); Vector256<short> c0 = Avx2.PackSignedSaturate(b0, b1).AsInt16();
Vector256<short> c1 = Avx2.PackSignedSaturate(b2, b3).AsInt16(); Vector256<short> c1 = Avx2.PackSignedSaturate(b2, b3).AsInt16();
Vector256<byte> d = Avx2.PackSignedSaturate(c0, c1).AsByte(); Vector256<byte> d = Avx2.PackSignedSaturate(c0, c1).AsByte();
Vector256<byte> bits = Avx2.CompareEqual(d, all0x80Vector256); Vector256<byte> bits = Avx2.CompareEqual(d, All0x80Vector256);
int mask = Avx2.MoveMask(bits); int mask = Avx2.MoveMask(bits);
if (mask != -1) if (mask != -1)
{ {
@ -59,7 +64,7 @@ namespace SixLabors.ImageSharp.Formats.Webp
for (; i + 64 <= length; i += 64) for (; i + 64 <= length; i += 64)
{ {
if (IsNoneOpaque64Bytes(src, i, alphaMask, all0x80)) if (IsNoneOpaque64Bytes(src, i))
{ {
return true; return true;
} }
@ -67,7 +72,7 @@ namespace SixLabors.ImageSharp.Formats.Webp
for (; i + 32 <= length; i += 32) for (; i + 32 <= length; i += 32)
{ {
if (IsNoneOpaque32Bytes(src, i, alphaMask, all0x80)) if (IsNoneOpaque32Bytes(src, i))
{ {
return true; return true;
} }
@ -85,16 +90,13 @@ namespace SixLabors.ImageSharp.Formats.Webp
else if (Sse2.IsSupported) else if (Sse2.IsSupported)
{ {
ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row); ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
var alphaMask = Vector128.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
Vector128<byte> all0x80 = Vector128.Create((byte)0x80).AsByte();
int i = 0; int i = 0;
int length = (row.Length * 4) - 3; int length = (row.Length * 4) - 3;
fixed (byte* src = rowBytes) fixed (byte* src = rowBytes)
{ {
for (; i + 64 <= length; i += 64) for (; i + 64 <= length; i += 64)
{ {
if (IsNoneOpaque64Bytes(src, i, alphaMask, all0x80)) if (IsNoneOpaque64Bytes(src, i))
{ {
return true; return true;
} }
@ -102,7 +104,7 @@ namespace SixLabors.ImageSharp.Formats.Webp
for (; i + 32 <= length; i += 32) for (; i + 32 <= length; i += 32)
{ {
if (IsNoneOpaque32Bytes(src, i, alphaMask, all0x80)) if (IsNoneOpaque32Bytes(src, i))
{ {
return true; return true;
} }
@ -133,20 +135,20 @@ namespace SixLabors.ImageSharp.Formats.Webp
} }
#if SUPPORTS_RUNTIME_INTRINSICS #if SUPPORTS_RUNTIME_INTRINSICS
private static unsafe bool IsNoneOpaque64Bytes(byte* src, int i, Vector128<byte> alphaMask, Vector128<byte> all0x80) private static unsafe bool IsNoneOpaque64Bytes(byte* src, int i)
{ {
Vector128<byte> a0 = Sse2.LoadVector128(src + i).AsByte(); Vector128<byte> a0 = Sse2.LoadVector128(src + i).AsByte();
Vector128<byte> a1 = Sse2.LoadVector128(src + i + 16).AsByte(); Vector128<byte> a1 = Sse2.LoadVector128(src + i + 16).AsByte();
Vector128<byte> a2 = Sse2.LoadVector128(src + i + 32).AsByte(); Vector128<byte> a2 = Sse2.LoadVector128(src + i + 32).AsByte();
Vector128<byte> a3 = Sse2.LoadVector128(src + i + 48).AsByte(); Vector128<byte> a3 = Sse2.LoadVector128(src + i + 48).AsByte();
Vector128<int> b0 = Sse2.And(a0, alphaMask).AsInt32(); Vector128<int> b0 = Sse2.And(a0, AlphaMask).AsInt32();
Vector128<int> b1 = Sse2.And(a1, alphaMask).AsInt32(); Vector128<int> b1 = Sse2.And(a1, AlphaMask).AsInt32();
Vector128<int> b2 = Sse2.And(a2, alphaMask).AsInt32(); Vector128<int> b2 = Sse2.And(a2, AlphaMask).AsInt32();
Vector128<int> b3 = Sse2.And(a3, alphaMask).AsInt32(); Vector128<int> b3 = Sse2.And(a3, AlphaMask).AsInt32();
Vector128<short> c0 = Sse2.PackSignedSaturate(b0, b1).AsInt16(); Vector128<short> c0 = Sse2.PackSignedSaturate(b0, b1).AsInt16();
Vector128<short> c1 = Sse2.PackSignedSaturate(b2, b3).AsInt16(); Vector128<short> c1 = Sse2.PackSignedSaturate(b2, b3).AsInt16();
Vector128<byte> d = Sse2.PackSignedSaturate(c0, c1).AsByte(); Vector128<byte> d = Sse2.PackSignedSaturate(c0, c1).AsByte();
Vector128<byte> bits = Sse2.CompareEqual(d, all0x80); Vector128<byte> bits = Sse2.CompareEqual(d, All0x80);
int mask = Sse2.MoveMask(bits); int mask = Sse2.MoveMask(bits);
if (mask != 0xFFFF) if (mask != 0xFFFF)
{ {
@ -156,15 +158,15 @@ namespace SixLabors.ImageSharp.Formats.Webp
return false; return false;
} }
private static unsafe bool IsNoneOpaque32Bytes(byte* src, int i, Vector128<byte> alphaMask, Vector128<byte> all0x80) private static unsafe bool IsNoneOpaque32Bytes(byte* src, int i)
{ {
Vector128<byte> a0 = Sse2.LoadVector128(src + i).AsByte(); Vector128<byte> a0 = Sse2.LoadVector128(src + i).AsByte();
Vector128<byte> a1 = Sse2.LoadVector128(src + i + 16).AsByte(); Vector128<byte> a1 = Sse2.LoadVector128(src + i + 16).AsByte();
Vector128<int> b0 = Sse2.And(a0, alphaMask).AsInt32(); Vector128<int> b0 = Sse2.And(a0, AlphaMask).AsInt32();
Vector128<int> b1 = Sse2.And(a1, alphaMask).AsInt32(); Vector128<int> b1 = Sse2.And(a1, AlphaMask).AsInt32();
Vector128<short> c = Sse2.PackSignedSaturate(b0, b1).AsInt16(); Vector128<short> c = Sse2.PackSignedSaturate(b0, b1).AsInt16();
Vector128<byte> d = Sse2.PackSignedSaturate(c, c).AsByte(); Vector128<byte> d = Sse2.PackSignedSaturate(c, c).AsByte();
Vector128<byte> bits = Sse2.CompareEqual(d, all0x80); Vector128<byte> bits = Sse2.CompareEqual(d, All0x80);
int mask = Sse2.MoveMask(bits); int mask = Sse2.MoveMask(bits);
if (mask != 0xFFFF) if (mask != 0xFFFF)
{ {

Write Preview
Loading…
Cancel
Save