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

Add sse2 version of select

pull/1804/head
Brian Popow 4 years ago
parent
d021222162
commit
35d2afa0bb
2 changed files with 64 additions and 23 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. 60
      src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
  2. 27
      src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs

60
src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
View File

@ -27,6 +27,10 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
private const double Log2Reciprocal = 1.44269504088896338700465094007086;
#if SUPPORTS_RUNTIME_INTRINSICS
private static readonly Vector128<byte> Zero = Vector128.Create(0).AsByte();
#endif
/// <summary>
/// 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
@ -551,6 +555,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
int mask = tileWidth - 1;
int tilesPerRow = SubSampleSize(width, transform.Bits);
int predictorModeIdxBase = (y >> transform.Bits) * tilesPerRow;
Span<short> scratch = stackalloc short[8];
while (y < yEnd)
{
int predictorModeIdx = predictorModeIdxBase;
@ -608,7 +613,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
PredictorAdd10(input + x, output + x - width, xEnd - x, output + x);
break;
case 11:
PredictorAdd11(input + x, output + x - width, xEnd - x, output + x);
PredictorAdd11(input + x, output + x - width, xEnd - x, output + x, scratch);
break;
case 12:
PredictorAdd12(input + x, output + x - width, xEnd - x, output + x);
@ -974,11 +979,11 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
}
[MethodImpl(InliningOptions.ShortMethod)]
private static void PredictorAdd11(uint* input, uint* upper, int numberOfPixels, uint* output)
private static void PredictorAdd11(uint* input, uint* upper, int numberOfPixels, uint* output, Span<short> scratch)
{
for (int x = 0; x < numberOfPixels; x++)
{
uint pred = Predictor11(output[x - 1], upper + x);
uint pred = Predictor11(output[x - 1], upper + x, scratch);
output[x] = AddPixels(input[x], pred);
}
}
@ -1031,7 +1036,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
public static uint Predictor10(uint left, uint* top) => Average4(left, top[-1], top[0], top[1]);
[MethodImpl(InliningOptions.ShortMethod)]
public static uint Predictor11(uint left, uint* top) => Select(top[0], left, top[-1]);
public static uint Predictor11(uint left, uint* top, Span<short> scratch) => Select(top[0], left, top[-1], scratch);
[MethodImpl(InliningOptions.ShortMethod)]
public static uint Predictor12(uint left, uint* top) => ClampedAddSubtractFull(left, top[0], top[-1]);
@ -1148,11 +1153,11 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
}
[MethodImpl(InliningOptions.ShortMethod)]
public static void PredictorSub11(uint* input, uint* upper, int numPixels, uint* output)
public static void PredictorSub11(uint* input, uint* upper, int numPixels, uint* output, Span<short> scratch)
{
for (int x = 0; x < numPixels; x++)
{
uint pred = Predictor11(input[x - 1], upper + x);
uint pred = Predictor11(input[x - 1], upper + x, scratch);
output[x] = SubPixels(input[x], pred);
}
}
@ -1240,14 +1245,43 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
private static Vector128<int> MkCst16(int hi, int lo) => Vector128.Create((hi << 16) | (lo & 0xffff));
#endif
private static uint Select(uint a, uint b, uint c)
private static uint Select(uint a, uint b, uint c, Span<short> scratch)
{
int paMinusPb =
Sub3((int)(a >> 24), (int)(b >> 24), (int)(c >> 24)) +
Sub3((int)((a >> 16) & 0xff), (int)((b >> 16) & 0xff), (int)((c >> 16) & 0xff)) +
Sub3((int)((a >> 8) & 0xff), (int)((b >> 8) & 0xff), (int)((c >> 8) & 0xff)) +
Sub3((int)(a & 0xff), (int)(b & 0xff), (int)(c & 0xff));
return paMinusPb <= 0 ? a : b;
#if SUPPORTS_RUNTIME_INTRINSICS
if (Sse2.IsSupported)
{
Span<short> output = scratch;
fixed (short* p = output)
{
Vector128<byte> a0 = Sse2.ConvertScalarToVector128UInt32(a).AsByte();
Vector128<byte> b0 = Sse2.ConvertScalarToVector128UInt32(b).AsByte();
Vector128<byte> c0 = Sse2.ConvertScalarToVector128UInt32(c).AsByte();
Vector128<byte> ac0 = Sse2.SubtractSaturate(a0, c0);
Vector128<byte> ca0 = Sse2.SubtractSaturate(c0, a0);
Vector128<byte> bc0 = Sse2.SubtractSaturate(b0, c0);
Vector128<byte> cb0 = Sse2.SubtractSaturate(c0, b0);
Vector128<byte> ac = Sse2.Or(ac0, ca0);
Vector128<byte> bc = Sse2.Or(bc0, cb0);
Vector128<byte> pa = Sse2.UnpackLow(ac, Zero); // |a - c|
Vector128<byte> pb = Sse2.UnpackLow(bc, Zero); // |b - c|
Vector128<ushort> diff = Sse2.Subtract(pb.AsUInt16(), pa.AsUInt16());
Sse2.Store((ushort*)p, diff);
}
int paMinusPb = output[0] + output[1] + output[2] + output[3];
return (paMinusPb <= 0) ? a : b;
}
else
#endif
{
int paMinusPb =
Sub3((int)(a >> 24), (int)(b >> 24), (int)(c >> 24)) +
Sub3((int)((a >> 16) & 0xff), (int)((b >> 16) & 0xff), (int)((c >> 16) & 0xff)) +
Sub3((int)((a >> 8) & 0xff), (int)((b >> 8) & 0xff), (int)((c >> 8) & 0xff)) +
Sub3((int)(a & 0xff), (int)(b & 0xff), (int)(c & 0xff));
return paMinusPb <= 0 ? a : b;
}
}
[MethodImpl(InliningOptions.ShortMethod)]

27
src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
View File

@ -50,6 +50,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
int tilesPerRow = LosslessUtils.SubSampleSize(width, bits);
int tilesPerCol = LosslessUtils.SubSampleSize(height, bits);
int maxQuantization = 1 << LosslessUtils.NearLosslessBits(nearLosslessQuality);
Span<short> scratch = stackalloc short[8];
// TODO: Can we optimize this?
int[][] histo = new int[4][];
@ -84,7 +85,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
transparentColorMode,
usedSubtractGreen,
nearLossless,
image);
image,
scratch);
image[(tileY * tilesPerRow) + tileX] = (uint)(WebpConstants.ArgbBlack | (pred << 8));
}
@ -192,7 +194,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
WebpTransparentColorMode transparentColorMode,
bool usedSubtractGreen,
bool nearLossless,
Span<uint> modes)
Span<uint> modes,
Span<short> scratch)
{
const int numPredModes = 14;
int startX = tileX << bits;
@ -272,7 +275,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
}
}
GetResidual(width, height, upperRow, currentRow, maxDiffs, mode, startX, startX + maxX, y, maxQuantization, transparentColorMode, usedSubtractGreen, nearLossless, residuals);
GetResidual(width, height, upperRow, currentRow, maxDiffs, mode, startX, startX + maxX, y, maxQuantization, transparentColorMode, usedSubtractGreen, nearLossless, residuals, scratch);
for (int relativeX = 0; relativeX < maxX; ++relativeX)
{
UpdateHisto(histoArgb, residuals[relativeX]);
@ -333,11 +336,12 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
WebpTransparentColorMode transparentColorMode,
bool usedSubtractGreen,
bool nearLossless,
Span<uint> output)
Span<uint> output,
Span<short> scratch)
{
if (transparentColorMode == WebpTransparentColorMode.Preserve)
{
PredictBatch(mode, xStart, y, xEnd - xStart, currentRowSpan, upperRowSpan, output);
PredictBatch(mode, xStart, y, xEnd - xStart, currentRowSpan, upperRowSpan, output, scratch);
}
else
{
@ -395,7 +399,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
predict = LosslessUtils.Predictor10(currentRow[x - 1], upperRow + x);
break;
case 11:
predict = LosslessUtils.Predictor11(currentRow[x - 1], upperRow + x);
predict = LosslessUtils.Predictor11(currentRow[x - 1], upperRow + x, scratch);
break;
case 12:
predict = LosslessUtils.Predictor12(currentRow[x - 1], upperRow + x);
@ -583,6 +587,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
Span<byte> currentMaxDiffs = MemoryMarshal.Cast<uint, byte>(currentRow.Slice(width + 1));
Span<byte> lowerMaxDiffs = currentMaxDiffs.Slice(width);
Span<short> scratch = stackalloc short[8];
for (int y = 0; y < height; y++)
{
Span<uint> tmp32 = upperRow;
@ -593,7 +598,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
if (lowEffort)
{
PredictBatch(PredLowEffort, 0, y, width, currentRow, upperRow, argb.Slice(y * width));
PredictBatch(PredLowEffort, 0, y, width, currentRow, upperRow, argb.Slice(y * width), scratch);
}
else
{
@ -634,7 +639,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
transparentColorMode,
usedSubtractGreen,
nearLossless,
argb.Slice((y * width) + x));
argb.Slice((y * width) + x),
scratch);
x = xEnd;
}
@ -649,7 +655,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
int numPixels,
Span<uint> currentSpan,
Span<uint> upperSpan,
Span<uint> outputSpan)
Span<uint> outputSpan,
Span<short> scratch)
{
#pragma warning disable SA1503 // Braces should not be omitted
fixed (uint* current = currentSpan)
@ -718,7 +725,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
LosslessUtils.PredictorSub10(current + xStart, upper + xStart, numPixels, output);
break;
case 11:
LosslessUtils.PredictorSub11(current + xStart, upper + xStart, numPixels, output);
LosslessUtils.PredictorSub11(current + xStart, upper + xStart, numPixels, output, scratch);
break;
case 12:
LosslessUtils.PredictorSub12(current + xStart, upper + xStart, numPixels, output);

Write Preview
Loading…
Cancel
Save