|
| 1 | +// Copyright (c) Six Labors. |
| 2 | +// Licensed under the Apache License, Version 2.0. |
| 3 | + |
| 4 | +using System; |
| 5 | +using System.Buffers.Binary; |
| 6 | +using System.Runtime.CompilerServices; |
| 7 | +using System.Runtime.InteropServices; |
| 8 | + |
| 9 | +namespace SixLabors.ImageSharp |
| 10 | +{ |
| 11 | + /// <summary> |
| 12 | + /// Defines the contract for methods that allow the shuffling of pixel components. |
| 13 | + /// Used for shuffling on platforms that do not support Hardware Intrinsics. |
| 14 | + /// </summary> |
| 15 | + internal interface IComponentShuffle |
| 16 | + { |
| 17 | + /// <summary> |
| 18 | + /// Gets the shuffle control. |
| 19 | + /// </summary> |
| 20 | + byte Control { get; } |
| 21 | + |
| 22 | + /// <summary> |
| 23 | + /// Shuffle 8-bit integers within 128-bit lanes in <paramref name="source"/> |
| 24 | + /// using the control and store the results in <paramref name="dest"/>. |
| 25 | + /// </summary> |
| 26 | + /// <param name="source">The source span of bytes.</param> |
| 27 | + /// <param name="dest">The destination span of bytes.</param> |
| 28 | + void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest); |
| 29 | + } |
| 30 | + |
| 31 | + internal readonly struct DefaultShuffle4 : IComponentShuffle |
| 32 | + { |
| 33 | + public DefaultShuffle4(byte p3, byte p2, byte p1, byte p0) |
| 34 | + : this(SimdUtils.Shuffle.MmShuffle(p3, p2, p1, p0)) |
| 35 | + { |
| 36 | + } |
| 37 | + |
| 38 | + public DefaultShuffle4(byte control) => this.Control = control; |
| 39 | + |
| 40 | + public byte Control { get; } |
| 41 | + |
| 42 | + [MethodImpl(InliningOptions.ShortMethod)] |
| 43 | + public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest) |
| 44 | + { |
| 45 | + ref byte sBase = ref MemoryMarshal.GetReference(source); |
| 46 | + ref byte dBase = ref MemoryMarshal.GetReference(dest); |
| 47 | + SimdUtils.Shuffle.InverseMmShuffle( |
| 48 | + this.Control, |
| 49 | + out int p3, |
| 50 | + out int p2, |
| 51 | + out int p1, |
| 52 | + out int p0); |
| 53 | + |
| 54 | + for (int i = 0; i < source.Length; i += 4) |
| 55 | + { |
| 56 | + Unsafe.Add(ref dBase, i) = Unsafe.Add(ref sBase, p0 + i); |
| 57 | + Unsafe.Add(ref dBase, i + 1) = Unsafe.Add(ref sBase, p1 + i); |
| 58 | + Unsafe.Add(ref dBase, i + 2) = Unsafe.Add(ref sBase, p2 + i); |
| 59 | + Unsafe.Add(ref dBase, i + 3) = Unsafe.Add(ref sBase, p3 + i); |
| 60 | + } |
| 61 | + } |
| 62 | + } |
| 63 | + |
| 64 | + internal readonly struct WXYZShuffle4 : IComponentShuffle |
| 65 | + { |
| 66 | + public byte Control => SimdUtils.Shuffle.MmShuffle(2, 1, 0, 3); |
| 67 | + |
| 68 | + [MethodImpl(InliningOptions.ShortMethod)] |
| 69 | + public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest) |
| 70 | + { |
| 71 | + ReadOnlySpan<uint> s = MemoryMarshal.Cast<byte, uint>(source); |
| 72 | + Span<uint> d = MemoryMarshal.Cast<byte, uint>(dest); |
| 73 | + ref uint sBase = ref MemoryMarshal.GetReference(s); |
| 74 | + ref uint dBase = ref MemoryMarshal.GetReference(d); |
| 75 | + |
| 76 | + // The JIT can detect and optimize rotation idioms ROTL (Rotate Left) |
| 77 | + // and ROTR (Rotate Right) emitting efficient CPU instructions: |
| 78 | + // https://github.com/dotnet/coreclr/pull/1830 |
| 79 | + for (int i = 0; i < s.Length; i++) |
| 80 | + { |
| 81 | + uint packed = Unsafe.Add(ref sBase, i); |
| 82 | + |
| 83 | + // packed = [W Z Y X] |
| 84 | + // ROTL(8, packed) = [Z Y X W] |
| 85 | + Unsafe.Add(ref dBase, i) = (packed << 8) | (packed >> 24); |
| 86 | + } |
| 87 | + } |
| 88 | + } |
| 89 | + |
| 90 | + internal readonly struct WZYXShuffle4 : IComponentShuffle |
| 91 | + { |
| 92 | + public byte Control => SimdUtils.Shuffle.MmShuffle(0, 1, 2, 3); |
| 93 | + |
| 94 | + [MethodImpl(InliningOptions.ShortMethod)] |
| 95 | + public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest) |
| 96 | + { |
| 97 | + ReadOnlySpan<uint> s = MemoryMarshal.Cast<byte, uint>(source); |
| 98 | + Span<uint> d = MemoryMarshal.Cast<byte, uint>(dest); |
| 99 | + ref uint sBase = ref MemoryMarshal.GetReference(s); |
| 100 | + ref uint dBase = ref MemoryMarshal.GetReference(d); |
| 101 | + |
| 102 | + for (int i = 0; i < s.Length; i++) |
| 103 | + { |
| 104 | + uint packed = Unsafe.Add(ref sBase, i); |
| 105 | + |
| 106 | + // packed = [W Z Y X] |
| 107 | + // REVERSE(packedArgb) = [X Y Z W] |
| 108 | + Unsafe.Add(ref dBase, i) = BinaryPrimitives.ReverseEndianness(packed); |
| 109 | + } |
| 110 | + } |
| 111 | + } |
| 112 | + |
| 113 | + internal readonly struct YZWXShuffle4 : IComponentShuffle |
| 114 | + { |
| 115 | + public byte Control => SimdUtils.Shuffle.MmShuffle(0, 3, 2, 1); |
| 116 | + |
| 117 | + [MethodImpl(InliningOptions.ShortMethod)] |
| 118 | + public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest) |
| 119 | + { |
| 120 | + ReadOnlySpan<uint> s = MemoryMarshal.Cast<byte, uint>(source); |
| 121 | + Span<uint> d = MemoryMarshal.Cast<byte, uint>(dest); |
| 122 | + ref uint sBase = ref MemoryMarshal.GetReference(s); |
| 123 | + ref uint dBase = ref MemoryMarshal.GetReference(d); |
| 124 | + |
| 125 | + for (int i = 0; i < s.Length; i++) |
| 126 | + { |
| 127 | + uint packed = Unsafe.Add(ref sBase, i); |
| 128 | + |
| 129 | + // packed = [W Z Y X] |
| 130 | + // ROTR(8, packedArgb) = [Y Z W X] |
| 131 | + Unsafe.Add(ref dBase, i) = (packed >> 8) | (packed << 24); |
| 132 | + } |
| 133 | + } |
| 134 | + } |
| 135 | + |
| 136 | + internal readonly struct ZYXWShuffle4 : IComponentShuffle |
| 137 | + { |
| 138 | + public byte Control => SimdUtils.Shuffle.MmShuffle(3, 0, 1, 2); |
| 139 | + |
| 140 | + [MethodImpl(InliningOptions.ShortMethod)] |
| 141 | + public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest) |
| 142 | + { |
| 143 | + ReadOnlySpan<uint> s = MemoryMarshal.Cast<byte, uint>(source); |
| 144 | + Span<uint> d = MemoryMarshal.Cast<byte, uint>(dest); |
| 145 | + ref uint sBase = ref MemoryMarshal.GetReference(s); |
| 146 | + ref uint dBase = ref MemoryMarshal.GetReference(d); |
| 147 | + |
| 148 | + for (int i = 0; i < s.Length; i++) |
| 149 | + { |
| 150 | + uint packed = Unsafe.Add(ref sBase, i); |
| 151 | + |
| 152 | + // packed = [W Z Y X] |
| 153 | + // tmp1 = [W 0 Y 0] |
| 154 | + // tmp2 = [0 Z 0 X] |
| 155 | + // tmp3=ROTL(16, tmp2) = [0 X 0 Z] |
| 156 | + // tmp1 + tmp3 = [W X Y Z] |
| 157 | + uint tmp1 = packed & 0xFF00FF00; |
| 158 | + uint tmp2 = packed & 0x00FF00FF; |
| 159 | + uint tmp3 = (tmp2 << 16) | (tmp2 >> 16); |
| 160 | + |
| 161 | + Unsafe.Add(ref dBase, i) = tmp1 + tmp3; |
| 162 | + } |
| 163 | + } |
| 164 | + } |
| 165 | +} |
0 commit comments