From 59a11bf9011729ad5cf8f30f3ef21adf616bf0b9 Mon Sep 17 00:00:00 2001 From: Brian Popow Date: Wed, 17 Nov 2021 21:45:32 +0100 Subject: [PATCH] Add SSE41 version of UpSample --- .../Formats/Webp/Lossy/YuvConversion.cs | 332 ++++++++++++++++++ 1 file changed, 332 insertions(+) diff --git a/src/ImageSharp/Formats/Webp/Lossy/YuvConversion.cs b/src/ImageSharp/Formats/Webp/Lossy/YuvConversion.cs index 182437e54f..0f5c56c746 100644 --- a/src/ImageSharp/Formats/Webp/Lossy/YuvConversion.cs +++ b/src/ImageSharp/Formats/Webp/Lossy/YuvConversion.cs @@ -4,6 +4,11 @@ using System; using System.Buffers; using System.Runtime.CompilerServices; +using System.Runtime.InteropServices; +#if SUPPORTS_RUNTIME_INTRINSICS +using System.Runtime.Intrinsics; +using System.Runtime.Intrinsics.X86; +#endif using SixLabors.ImageSharp.Memory; using SixLabors.ImageSharp.PixelFormats; @@ -18,8 +23,66 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossy private const int YuvHalf = 1 << (YuvFix - 1); +#if SUPPORTS_RUNTIME_INTRINSICS + private static readonly Vector128 One = Vector128.Create((byte)1); + + // These constants are 14b fixed-point version of ITU-R BT.601 constants. + // R = (19077 * y + 26149 * v - 14234) >> 6 + // G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6 + // B = (19077 * y + 33050 * u - 17685) >> 6 + private static readonly Vector128 K19077 = Vector128.Create((short)19077).AsByte(); + + private static readonly Vector128 K26149 = Vector128.Create((short)26149).AsByte(); + + private static readonly Vector128 K14234 = Vector128.Create((short)14234).AsByte(); + + // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic + private static readonly Vector128 K33050 = Vector128.Create(26, 129, 26, 129, 26, 129, 26, 129, 26, 129, 26, 129, 26, 129, 26, 129); + + private static readonly Vector128 K17685 = Vector128.Create((short)17685).AsByte(); + + private static readonly Vector128 K6419 = Vector128.Create((short)6419).AsByte(); + + private static readonly Vector128 K13320 = Vector128.Create((short)13320).AsByte(); + + private static readonly Vector128 K8708 = Vector128.Create((short)8708).AsByte(); + + private static readonly Vector128 PlanarTo24Shuffle0 = Vector128.Create(0, 255, 255, 1, 255, 255, 2, 255, 255, 3, 255, 255, 4, 255, 255, 5); + + private static readonly Vector128 PlanarTo24Shuffle1 = Vector128.Create(255, 255, 6, 255, 255, 7, 255, 255, 8, 255, 255, 9, 255, 255, 10, 255); + + private static readonly Vector128 PlanarTo24Shuffle2 = Vector128.Create(255, 11, 255, 255, 12, 255, 255, 13, 255, 255, 14, 255, 255, 15, 255, 255); + + private static readonly Vector128 PlanarTo24Shuffle3 = Vector128.Create(255, 0, 255, 255, 1, 255, 255, 2, 255, 255, 3, 255, 255, 4, 255, 255); + + private static readonly Vector128 PlanarTo24Shuffle4 = Vector128.Create(5, 255, 255, 6, 255, 255, 7, 255, 255, 8, 255, 255, 9, 255, 255, 10); + + private static readonly Vector128 PlanarTo24Shuffle5 = Vector128.Create(255, 255, 11, 255, 255, 12, 255, 255, 13, 255, 255, 24, 255, 255, 15, 255); + + private static readonly Vector128 PlanarTo24Shuffle6 = Vector128.Create(255, 255, 0, 255, 255, 1, 255, 255, 2, 255, 255, 3, 255, 255, 4, 255); + + private static readonly Vector128 PlanarTo24Shuffle7 = Vector128.Create(255, 5, 255, 255, 6, 255, 255, 7, 255, 255, 8, 255, 255, 9, 255, 255); + + private static readonly Vector128 PlanarTo24Shuffle8 = Vector128.Create(10, 255, 255, 11, 255, 255, 12, 255, 255, 13, 255, 255, 24, 255, 255, 15); +#endif + + // UpSample from YUV to RGB. public static void UpSample(Span topY, Span bottomY, Span topU, Span topV, Span curU, Span curV, Span topDst, Span bottomDst, int len) { +#if SUPPORTS_RUNTIME_INTRINSICS + if (Sse41.IsSupported) + { + UpSampleSse41(topY, bottomY, topU, topV, curU, curV, topDst, bottomDst, len); + } + else +#endif + { + UpSampleScalar(topY, bottomY, topU, topV, curU, curV, topDst, bottomDst, len); + } + } + + public static void UpSampleScalar(Span topY, Span bottomY, Span topU, Span topV, Span curU, Span curV, Span topDst, Span bottomDst, int len) + { int xStep = 3; int lastPixelPair = (len - 1) >> 1; uint tluv = LoadUv(topU[0], topV[0]); // top-left sample @@ -72,6 +135,106 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossy } } +#if SUPPORTS_RUNTIME_INTRINSICS + // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows + // u = (9*a + 3*b + 3*c + d + 8) / 16 + // = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2 + // = (a + m + 1) / 2 + // where m = (a + 3*b + 3*c + d) / 8 + // = ((a + b + c + d) / 2 + b + c) / 4 + // + // Let's say k = (a + b + c + d) / 4. + // We can compute k as + // k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1 + // where s = (a + d + 1) / 2 and t = (b + c + 1) / 2 + // + // Then m can be written as + // m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1 + public static void UpSampleSse41(Span topY, Span bottomY, Span topU, Span topV, Span curU, Span curV, Span topDst, Span bottomDst, int len) + { + const int xStep = 3; + byte[] uvBuffer = new byte[(14 * 32) + 15]; + Span ru = uvBuffer.AsSpan(15); + Span rv = ru.Slice(32); + + // Treat the first pixel in regular way. + int uDiag = ((topU[0] + curU[0]) >> 1) + 1; + int vDiag = ((topV[0] + curV[0]) >> 1) + 1; + int u0t = (topU[0] + uDiag) >> 1; + int v0t = (topV[0] + vDiag) >> 1; + YuvToBgr(topY[0], u0t, v0t, topDst); + if (bottomY != null) + { + int u0b = (curU[0] + uDiag) >> 1; + int v0b = (curV[0] + vDiag) >> 1; + YuvToBgr(bottomY[0], u0b, v0b, bottomDst); + } + + // For UpSample32Pixels, 17 u/v values must be read-able for each block. + for (int pos = 1, uvPos = 0; pos + 32 + 1 <= len; pos += 32, uvPos += 16) + { + UpSample32Pixels(topU.Slice(uvPos), curU.Slice(uvPos), ru); + UpSample32Pixels(topV.Slice(uvPos), curV.Slice(uvPos), rv); + ConvertToBgrSse41(topY, bottomY, topDst, bottomDst, ru, rv, pos, xStep); + } + } + + // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels. + public static void UpSample32Pixels(Span r1, Span r2, Span output) + { + // Load inputs. + Vector128 a = Unsafe.As>(ref MemoryMarshal.GetReference(r1)); + Vector128 b = Unsafe.As>(ref Unsafe.Add(ref MemoryMarshal.GetReference(r1), 1)); + Vector128 c = Unsafe.As>(ref MemoryMarshal.GetReference(r2)); + Vector128 d = Unsafe.As>(ref Unsafe.Add(ref MemoryMarshal.GetReference(r2), 1)); + + Vector128 s = Sse2.Average(a, d); // s = (a + d + 1) / 2 + Vector128 t = Sse2.Average(b, c); // t = (b + c + 1) / 2 + Vector128 st = Sse2.Xor(s, t); // st = s^t + + Vector128 ad = Sse2.Xor(a, d); // ad = a^d + Vector128 bc = Sse2.Xor(b, c); // bc = b^c + + Vector128 t1 = Sse2.Or(ad, bc); // (a^d) | (b^c) + Vector128 t2 = Sse2.Or(t1, st); // (a^d) | (b^c) | (s^t) + Vector128 t3 = Sse2.And(t2, One); // (a^d) | (b^c) | (s^t) & 1 + Vector128 t4 = Sse2.Average(s, t); + Vector128 k = Sse2.Subtract(t4, t3); // k = (a + b + c + d) / 4 + + Vector128 diag1 = GetM(k, st, bc, t); + Vector128 diag2 = GetM(k, st, ad, s); + + // Pack the alternate pixels. + PackAndStore(a, b, diag1, diag2, output); // store top. + PackAndStore(c, d, diag2, diag1, output.Slice(2 * 32)); + } + + // Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1 + private static Vector128 GetM(Vector128 k, Vector128 st, Vector128 ij, Vector128 input) + { + Vector128 tmp0 = Sse2.Average(k, input); // (k + in + 1) / 2 + Vector128 tmp1 = Sse2.And(ij, st); // (ij) & (s^t) + Vector128 tmp2 = Sse2.Xor(k, input); // (k^in) + Vector128 tmp3 = Sse2.Or(tmp1, tmp2); // ((ij) & (s^t)) | (k^in) + Vector128 tmp4 = Sse2.And(tmp3, One); // & 1 -> lsb_correction + + return Sse2.Subtract(tmp0, tmp4); // (k + in + 1) / 2 - lsb_correction + } + + private static void PackAndStore(Vector128 a, Vector128 b, Vector128 da, Vector128 db, Span output) + { + Vector128 ta = Sse2.Average(a, da); // (9a + 3b + 3c + d + 8) / 16 + Vector128 tb = Sse2.Average(b, db); // (3a + 9b + c + 3d + 8) / 16 + Vector128 t1 = Sse2.UnpackLow(ta, tb); + Vector128 t2 = Sse2.UnpackHigh(ta, tb); + + ref byte output0Ref = ref MemoryMarshal.GetReference(output); + ref byte output1Ref = ref Unsafe.Add(ref output0Ref, 16); + Unsafe.As>(ref output0Ref) = t1; + Unsafe.As>(ref output1Ref) = t2; + } +#endif + /// /// Converts the RGB values of the image to YUV. /// @@ -366,6 +529,175 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossy bgr[0] = (byte)YuvToB(y, u); } +#if SUPPORTS_RUNTIME_INTRINSICS + + private static void ConvertToBgrSse41(Span topY, Span bottomY, Span topDst, Span bottomDst, Span ru, Span rv, int curX, int step) + { + YuvToBgrSse41(topY.Slice(curX), ru, rv, topDst.Slice(curX * step)); + + if (bottomY != null) + { + YuvToBgrSse41(bottomY.Slice(curX), ru.Slice(64), rv.Slice(64), bottomDst.Slice(curX * step)); + } + } + + public static void YuvToBgrSse41(Span y, Span u, Span v, Span dst) + { + ConvertYuv444ToRgbSse41(y, u, v, out Vector128 r0, out Vector128 g0, out Vector128 b0); + ConvertYuv444ToRgbSse41(y.Slice(8), u.Slice(8), v.Slice(8), out Vector128 r1, out Vector128 g1, out Vector128 b1); + ConvertYuv444ToRgbSse41(y.Slice(16), u.Slice(16), v.Slice(16), out Vector128 r2, out Vector128 g2, out Vector128 b2); + ConvertYuv444ToRgbSse41(y.Slice(24), u.Slice(24), v.Slice(24), out Vector128 r3, out Vector128 g3, out Vector128 b3); + + // Cast to 8b and store as BBBBGGGGRRRR. + Vector128 bgr0 = Sse2.PackUnsignedSaturate(b0, b1); + Vector128 bgr1 = Sse2.PackUnsignedSaturate(b2, b3); + Vector128 bgr2 = Sse2.PackUnsignedSaturate(g0, g1); + Vector128 bgr3 = Sse2.PackUnsignedSaturate(g2, g3); + Vector128 bgr4 = Sse2.PackUnsignedSaturate(r0, r1); + Vector128 bgr5 = Sse2.PackUnsignedSaturate(r2, r3); + + // Pack as BGRBGRBGRBGR. + PlanarTo24bSse41(bgr0, bgr1, bgr2, bgr3, bgr4, bgr5, dst); + } + + // Pack the planar buffers + // rrrr... rrrr... gggg... gggg... bbbb... bbbb.... + // triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ... + private static void PlanarTo24bSse41(Vector128 input0, Vector128 input1, Vector128 input2, Vector128 input3, Vector128 input4, Vector128 input5, Span rgb) + { + // The input is 6 registers of sixteen 8b but for the sake of explanation, + // let's take 6 registers of four 8b values. + // To pack, we will keep taking one every two 8b integer and move it + // around as follows: + // Input: + // r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7 + // Split the 6 registers in two sets of 3 registers: the first set as the even + // 8b bytes, the second the odd ones: + // r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7 + // Repeat the same permutations twice more: + // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7 + // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7 + + // Process R. + ChannelMixing( + input0, + input1, + PlanarTo24Shuffle0, + PlanarTo24Shuffle1, + PlanarTo24Shuffle2, + out Vector128 r0, + out Vector128 r1, + out Vector128 r2, + out Vector128 r3, + out Vector128 r4, + out Vector128 r5); + + // Process G. + // Same as before, just shifted to the left by one and including the right padding. + ChannelMixing( + input2, + input3, + PlanarTo24Shuffle3, + PlanarTo24Shuffle4, + PlanarTo24Shuffle5, + out Vector128 g0, + out Vector128 g1, + out Vector128 g2, + out Vector128 g3, + out Vector128 g4, + out Vector128 g5); + + // Process B. + ChannelMixing( + input4, + input5, + PlanarTo24Shuffle6, + PlanarTo24Shuffle7, + PlanarTo24Shuffle8, + out Vector128 b0, + out Vector128 b1, + out Vector128 b2, + out Vector128 b3, + out Vector128 b4, + out Vector128 b5); + + // OR the different channels. + Vector128 rg0 = Sse2.Or(r0, g0); + Vector128 rg1 = Sse2.Or(r1, g1); + Vector128 rg2 = Sse2.Or(r2, g2); + Vector128 rg3 = Sse2.Or(r3, g3); + Vector128 rg4 = Sse2.Or(r4, g4); + Vector128 rg5 = Sse2.Or(r5, g5); + + ref byte outputRef = ref MemoryMarshal.GetReference(rgb); + Unsafe.As>(ref outputRef) = Sse2.Or(rg0, b0); + Unsafe.As>(ref Unsafe.Add(ref outputRef, 16)) = Sse2.Or(rg1, b1); + Unsafe.As>(ref Unsafe.Add(ref outputRef, 32)) = Sse2.Or(rg2, b2); + Unsafe.As>(ref Unsafe.Add(ref outputRef, 48)) = Sse2.Or(rg3, b3); + Unsafe.As>(ref Unsafe.Add(ref outputRef, 64)) = Sse2.Or(rg4, b4); + Unsafe.As>(ref Unsafe.Add(ref outputRef, 80)) = Sse2.Or(rg5, b5); + } + + // Shuffles the input buffer as A0 0 0 A1 0 0 A2 + private static void ChannelMixing( + Vector128 input0, + Vector128 input1, + Vector128 shuffle0, + Vector128 shuffle1, + Vector128 shuffle2, + out Vector128 output0, + out Vector128 output1, + out Vector128 output2, + out Vector128 output3, + out Vector128 output4, + out Vector128 output5) + { + output0 = Ssse3.Shuffle(input0, shuffle0); + output1 = Ssse3.Shuffle(input0, shuffle1); + output2 = Ssse3.Shuffle(input0, shuffle2); + output3 = Ssse3.Shuffle(input1, shuffle0); + output4 = Ssse3.Shuffle(input1, shuffle1); + output5 = Ssse3.Shuffle(input1, shuffle2); + } + + // Convert 32 samples of YUV444 to R/G/B + private static void ConvertYuv444ToRgbSse41(Span y, Span u, Span v, out Vector128 r, out Vector128 g, out Vector128 b) + { + Vector128 y0 = LoadHigh(y); + Vector128 u0 = LoadHigh(u); + Vector128 v0 = LoadHigh(v); + + Vector128 y1 = Sse2.MultiplyHigh(y0.AsUInt16(), K19077.AsUInt16()); + + Vector128 r0 = Sse2.MultiplyHigh(v0.AsUInt16(), K26149.AsUInt16()); + Vector128 r1 = Sse2.Subtract(y1.AsUInt16(), K14234.AsUInt16()); + Vector128 r2 = Sse2.Add(r1, r0); + + Vector128 g0 = Sse2.MultiplyHigh(u0.AsUInt16(), K6419.AsUInt16()); + Vector128 g1 = Sse2.MultiplyHigh(v0.AsUInt16(), K13320.AsUInt16()); + Vector128 g2 = Sse2.Add(y1.AsUInt16(), K8708.AsUInt16()); + Vector128 g3 = Sse2.Add(g0, g1); + Vector128 g4 = Sse2.Subtract(g2, g3); + + Vector128 b0 = Sse2.MultiplyHigh(u0.AsUInt16(), K33050.AsUInt16()); + Vector128 b1 = Sse2.AddSaturate(b0, y1); + Vector128 b2 = Sse2.SubtractSaturate(b1, K17685.AsUInt16()); + + // use logical shift for B2, which can be larger than 32767 + r = Sse2.ShiftRightArithmetic(r2.AsInt16(), 6); // range: [-14234, 30815] + g = Sse2.ShiftRightArithmetic(g4.AsInt16(), 6); // range: [-10953, 27710] + b = Sse2.ShiftRightLogical(b2.AsInt16(), 6); // range: [0, 34238] + } + + // Load the bytes into the *upper* part of 16b words. That's "<< 8", basically. + private static Vector128 LoadHigh(Span src) + { + Vector64 tmp = Unsafe.As>(ref MemoryMarshal.GetReference(src)); + Vector128 tmp2 = Unsafe.As, Vector128>(ref tmp); + return Sse2.UnpackLow(Vector128.Zero, tmp2); + } +#endif + [MethodImpl(InliningOptions.ShortMethod)] public static int YuvToB(int y, int u) => Clip8(MultHi(y, 19077) + MultHi(u, 33050) - 17685);