diff --git a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs index f252864476..d7511fddac 100644 --- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs +++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs @@ -280,7 +280,7 @@ public void AddInPlace(float value) } /// - /// Quantize input block, apply zig-zag ordering and store result as 16bit integers. + /// Quantize input block, transpose, apply zig-zag ordering and store as . /// /// Source block. /// Destination block. @@ -291,19 +291,19 @@ public static void Quantize(ref Block8x8F block, ref Block8x8 dest, ref Block8x8 if (Avx2.IsSupported) { MultiplyIntoInt16_Avx2(ref block, ref qt, ref dest); - ZigZag.ApplyZigZagOrderingAvx2(ref dest); + ZigZag.ApplyTransposingZigZagOrderingAvx2(ref dest); } else if (Ssse3.IsSupported) { MultiplyIntoInt16_Sse2(ref block, ref qt, ref dest); - ZigZag.ApplyZigZagOrderingSsse3(ref dest); + ZigZag.ApplyTransposingZigZagOrderingSsse3(ref dest); } else #endif { for (int i = 0; i < Size; i++) { - int idx = ZigZag.ZigZagOrder[i]; + int idx = ZigZag.TransposingOrder[i]; float quantizedVal = block[idx] * qt[idx]; quantizedVal += quantizedVal < 0 ? -0.5f : 0.5f; dest[i] = (short)quantizedVal; diff --git a/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.Intrinsic.cs b/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.Intrinsic.cs index 94864005ec..8acc4b6269 100644 --- a/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.Intrinsic.cs +++ b/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.Intrinsic.cs @@ -29,11 +29,10 @@ private static void FDCT8x8_Avx(ref Block8x8F block) { DebugGuard.IsTrue(Avx.IsSupported, "Avx support is required to execute this operation."); - // First pass - process rows - block.TransposeInplace(); + // First pass - process columns FDCT8x8_1D_Avx(ref block); - // Second pass - process columns + // Second pass - process rows block.TransposeInplace(); FDCT8x8_1D_Avx(ref block); diff --git a/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.cs b/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.cs index c27ad5b82b..e1bcff30f3 100644 --- a/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.cs +++ b/src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.cs @@ -92,6 +92,11 @@ public static void AdjustToFDCT(ref Block8x8F quantTable) tableRef = 0.125f / (tableRef * Unsafe.Add(ref multipliersRef, i)); tableRef = ref Unsafe.Add(ref tableRef, 1); } + + // Spectral macroblocks are not transposed before quantization + // Transpose is done after quantization at zig-zag stage + // so we must transpose quantization table + quantTable.TransposeInplace(); } /// @@ -133,14 +138,9 @@ public static void TransformFDCT(ref Block8x8F block) } else #endif - if (Vector.IsHardwareAccelerated) { FDCT_Vector4(ref block); } - else - { - FDCT_Scalar(ref block); - } } /// @@ -217,136 +217,17 @@ static void IDCT8x4_Vector4(ref Vector4 vecRef) } } - /// - /// Apply 2D floating point FDCT inplace using scalar operations. - /// - /// - /// Ported from libjpeg-turbo https://github.com/libjpeg-turbo/libjpeg-turbo/blob/main/jfdctflt.c. - /// - /// Input block. - private static void FDCT_Scalar(ref Block8x8F block) - { - const int dctSize = 8; - - float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; - float tmp10, tmp11, tmp12, tmp13; - float z1, z2, z3, z4, z5, z11, z13; - - // First pass - process rows - ref float blockRef = ref Unsafe.As(ref block); - for (int ctr = 7; ctr >= 0; ctr--) - { - tmp0 = Unsafe.Add(ref blockRef, 0) + Unsafe.Add(ref blockRef, 7); - tmp7 = Unsafe.Add(ref blockRef, 0) - Unsafe.Add(ref blockRef, 7); - tmp1 = Unsafe.Add(ref blockRef, 1) + Unsafe.Add(ref blockRef, 6); - tmp6 = Unsafe.Add(ref blockRef, 1) - Unsafe.Add(ref blockRef, 6); - tmp2 = Unsafe.Add(ref blockRef, 2) + Unsafe.Add(ref blockRef, 5); - tmp5 = Unsafe.Add(ref blockRef, 2) - Unsafe.Add(ref blockRef, 5); - tmp3 = Unsafe.Add(ref blockRef, 3) + Unsafe.Add(ref blockRef, 4); - tmp4 = Unsafe.Add(ref blockRef, 3) - Unsafe.Add(ref blockRef, 4); - - // Even part - tmp10 = tmp0 + tmp3; - tmp13 = tmp0 - tmp3; - tmp11 = tmp1 + tmp2; - tmp12 = tmp1 - tmp2; - - Unsafe.Add(ref blockRef, 0) = tmp10 + tmp11; - Unsafe.Add(ref blockRef, 4) = tmp10 - tmp11; - - z1 = (tmp12 + tmp13) * 0.707106781f; - Unsafe.Add(ref blockRef, 2) = tmp13 + z1; - Unsafe.Add(ref blockRef, 6) = tmp13 - z1; - - // Odd part - tmp10 = tmp4 + tmp5; - tmp11 = tmp5 + tmp6; - tmp12 = tmp6 + tmp7; - - z5 = (tmp10 - tmp12) * 0.382683433f; - z2 = (0.541196100f * tmp10) + z5; - z4 = (1.306562965f * tmp12) + z5; - z3 = tmp11 * 0.707106781f; - - z11 = tmp7 + z3; - z13 = tmp7 - z3; - - Unsafe.Add(ref blockRef, 5) = z13 + z2; - Unsafe.Add(ref blockRef, 3) = z13 - z2; - Unsafe.Add(ref blockRef, 1) = z11 + z4; - Unsafe.Add(ref blockRef, 7) = z11 - z4; - - blockRef = ref Unsafe.Add(ref blockRef, dctSize); - } - - // Second pass - process columns - blockRef = ref Unsafe.As(ref block); - for (int ctr = 7; ctr >= 0; ctr--) - { - tmp0 = Unsafe.Add(ref blockRef, dctSize * 0) + Unsafe.Add(ref blockRef, dctSize * 7); - tmp7 = Unsafe.Add(ref blockRef, dctSize * 0) - Unsafe.Add(ref blockRef, dctSize * 7); - tmp1 = Unsafe.Add(ref blockRef, dctSize * 1) + Unsafe.Add(ref blockRef, dctSize * 6); - tmp6 = Unsafe.Add(ref blockRef, dctSize * 1) - Unsafe.Add(ref blockRef, dctSize * 6); - tmp2 = Unsafe.Add(ref blockRef, dctSize * 2) + Unsafe.Add(ref blockRef, dctSize * 5); - tmp5 = Unsafe.Add(ref blockRef, dctSize * 2) - Unsafe.Add(ref blockRef, dctSize * 5); - tmp3 = Unsafe.Add(ref blockRef, dctSize * 3) + Unsafe.Add(ref blockRef, dctSize * 4); - tmp4 = Unsafe.Add(ref blockRef, dctSize * 3) - Unsafe.Add(ref blockRef, dctSize * 4); - - // Even part - tmp10 = tmp0 + tmp3; - tmp13 = tmp0 - tmp3; - tmp11 = tmp1 + tmp2; - tmp12 = tmp1 - tmp2; - - Unsafe.Add(ref blockRef, dctSize * 0) = tmp10 + tmp11; - Unsafe.Add(ref blockRef, dctSize * 4) = tmp10 - tmp11; - - z1 = (tmp12 + tmp13) * 0.707106781f; - Unsafe.Add(ref blockRef, dctSize * 2) = tmp13 + z1; - Unsafe.Add(ref blockRef, dctSize * 6) = tmp13 - z1; - - // Odd part - tmp10 = tmp4 + tmp5; - tmp11 = tmp5 + tmp6; - tmp12 = tmp6 + tmp7; - - z5 = (tmp10 - tmp12) * 0.382683433f; - z2 = (0.541196100f * tmp10) + z5; - z4 = (1.306562965f * tmp12) + z5; - z3 = tmp11 * 0.707106781f; - - z11 = tmp7 + z3; - z13 = tmp7 - z3; - - Unsafe.Add(ref blockRef, dctSize * 5) = z13 + z2; - Unsafe.Add(ref blockRef, dctSize * 3) = z13 - z2; - Unsafe.Add(ref blockRef, dctSize * 1) = z11 + z4; - Unsafe.Add(ref blockRef, dctSize * 7) = z11 - z4; - - blockRef = ref Unsafe.Add(ref blockRef, 1); - } - } - /// /// Apply floating point FDCT inplace using API. /// - /// - /// This implementation must be called only if hardware supports 4 - /// floating point numbers vector. Otherwise explicit scalar - /// implementation is faster - /// because it does not rely on block transposition. - /// /// Input block. public static void FDCT_Vector4(ref Block8x8F block) { - DebugGuard.IsTrue(Vector.IsHardwareAccelerated, "Scalar implementation should be called for non-accelerated hardware."); - - // First pass - process rows - block.TransposeInplace(); + // First pass - process columns FDCT8x4_Vector4(ref block.V0L); FDCT8x4_Vector4(ref block.V0R); - // Second pass - process columns + // Second pass - process rows block.TransposeInplace(); FDCT8x4_Vector4(ref block.V0L); FDCT8x4_Vector4(ref block.V0R); diff --git a/src/ImageSharp/Formats/Jpeg/Components/ZigZag.Intrinsic.cs b/src/ImageSharp/Formats/Jpeg/Components/ZigZag.Intrinsic.cs index 6577739c1a..850de26c30 100644 --- a/src/ImageSharp/Formats/Jpeg/Components/ZigZag.Intrinsic.cs +++ b/src/ImageSharp/Formats/Jpeg/Components/ZigZag.Intrinsic.cs @@ -3,6 +3,7 @@ #if SUPPORTS_RUNTIME_INTRINSICS using System; +using System.Runtime.InteropServices; using System.Runtime.Intrinsics; using System.Runtime.Intrinsics.X86; @@ -18,120 +19,138 @@ internal static partial class ZigZag #pragma warning restore SA1309 /// - /// Gets shuffle vectors for + /// Gets shuffle vectors for /// zig zag implementation. /// private static ReadOnlySpan SseShuffleMasks => new byte[] { - // row0 - 0, 1, 2, 3, _, _, _, _, _, _, 4, 5, 6, 7, _, _, - _, _, _, _, 0, 1, _, _, 2, 3, _, _, _, _, 4, 5, - _, _, _, _, _, _, 0, 1, _, _, _, _, _, _, _, _, - - // row1 - _, _, _, _, _, _, _, _, _, _, _, _, 8, 9, 10, 11, - 2, 3, _, _, _, _, _, _, 4, 5, _, _, _, _, _, _, - _, _, 0, 1, _, _, 2, 3, _, _, _, _, _, _, _, _, - - // row2 - _, _, _, _, _, _, 2, 3, _, _, _, _, _, _, 4, 5, - _, _, _, _, _, _, _, _, 0, 1, _, _, 2, 3, _, _, - - // row3 - _, _, _, _, _, _, 12, 13, 14, 15, _, _, _, _, _, _, - _, _, _, _, 10, 11, _, _, _, _, 12, 13, _, _, _, _, - _, _, 8, 9, _, _, _, _, _, _, _, _, 10, 11, _, _, - 6, 7, _, _, _, _, _, _, _, _, _, _, _, _, 8, 9, - - // row4 - _, _, 4, 5, _, _, _, _, _, _, _, _, 6, 7, _, _, - _, _, _, _, 2, 3, _, _, _, _, 4, 5, _, _, _, _, - _, _, _, _, _, _, 0, 1, 2, 3, _, _, _, _, _, _, - - // row5 - _, _, 12, 13, _, _, 14, 15, _, _, _, _, _, _, _, _, - 10, 11, _, _, _, _, _, _, 12, 13, _, _, _, _, _, _, - - // row6 - _, _, _, _, _, _, _, _, 12, 13, _, _, 14, 15, _, _, - _, _, _, _, _, _, 10, 11, _, _, _, _, _, _, 12, 13, - 4, 5, 6, 7, _, _, _, _, _, _, _, _, _, _, _, _, - - // row7 - 10, 11, _, _, _, _, 12, 13, _, _, 14, 15, _, _, _, _, - _, _, 8, 9, 10, 11, _, _, _, _, _, _, 12, 13, 14, 15 +#pragma warning disable SA1515 + /* row0 - A0 B0 A1 A2 B1 C0 D0 C1 */ + // A + 0, 1, _, _, 2, 3, 4, 5, _, _, _, _, _, _, _, _, + // B + _, _, 0, 1, _, _, _, _, 2, 3, _, _, _, _, _, _, + // C + _, _, _, _, _, _, _, _, _, _, 0, 1, _, _, 2, 3, + + /* row1 - B2 A3 A4 B3 C2 D1 E0 F0 */ + // A + _, _, 6, 7, 8, 9, _, _, _, _, _, _, _, _, _, _, + // B + 4, 5, _, _, _, _, 6, 7, _, _, _, _, _, _, _, _, + + /* row2 - E1 D2 C3 B4 A5 A6 B5 C4 */ + // A + _, _, _, _, _, _, _, _, 10, 11, 12, 13, _, _, _, _, + // B + _, _, _, _, _, _, 8, 9, _, _, _, _, 10, 11, _, _, + // C + _, _, _, _, 6, 7, _, _, _, _, _, _, _, _, 8, 9, + + /* row3 - D3 E2 F1 G0 H0 G1 F2 E3 */ + // E + _, _, 4, 5, _, _, _, _, _, _, _, _, _, _, 6, 7, + // F + _, _, _, _, 2, 3, _, _, _, _, _, _, 4, 5, _, _, + // G + _, _, _, _, _, _, 0, 1, _, _, 2, 3, _, _, _, _, + + /* row4 - D4 C5 B6 A7 B7 C6 D5 E4 */ + // B + _, _, _, _, 12, 13, _, _, 14, 15, _, _, _, _, _, _, + // C + _, _, 10, 11, _, _, _, _, _, _, 12, 13, _, _, _, _, + // D + 8, 9, _, _, _, _, _, _, _, _, _, _, 10, 11, _, _, + + /* row5 - F3 G2 H1 H2 G3 F4 E5 D6 */ + // F + 6, 7, _, _, _, _, _, _, _, _, 8, 9, _, _, _, _, + // G + _, _, 4, 5, _, _, _, _, 6, 7, _, _, _, _, _, _, + // H + _, _, _, _, 2, 3, 4, 5, _, _, _, _, _, _, _, _, + + /* row6 - C7 D7 E6 F5 G4 H3 H4 G5 */ + // G + _, _, _, _, _, _, _, _, 8, 9, _, _, _, _, 10, 11, + // H + _, _, _, _, _, _, _, _, _, _, 6, 7, 8, 9, _, _, + + /* row7 - F6 E7 F7 G6 H5 H6 G7 H7 */ + // F + 12, 13, _, _, 14, 15, _, _, _, _, _, _, _, _, _, _, + // G + _, _, _, _, _, _, 12, 13, _, _, _, _, 14, 15, _, _, + // H + _, _, _, _, _, _, _, _, 10, 11, 12, 13, _, _, 14, 15, +#pragma warning restore SA1515 }; /// - /// Gets shuffle vectors for + /// Gets shuffle vectors for /// zig zag implementation. /// private static ReadOnlySpan AvxShuffleMasks => new byte[] { - // 01_AB/01_EF/23_CD - cross-lane - 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, - - // 01_AB - inner-lane - 0, 1, 2, 3, 8, 9, _, _, 10, 11, 4, 5, 6, 7, 12, 13, _, _, _, _, _, _, _, _, _, _, 10, 11, 4, 5, 6, 7, - - // 01_CD/23_GH - cross-lane - 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, _, _, _, _, 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, _, _, _, _, - - // 01_CD - inner-lane - _, _, _, _, _, _, 0, 1, _, _, _, _, _, _, _, _, 2, 3, 8, 9, _, _, 10, 11, 4, 5, _, _, _, _, _, _, - - // 01_EF - inner-lane - _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 0, 1, _, _, _, _, _, _, _, _, _, _, - - // 23_AB/45_CD/67_EF - cross-lane - 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, _, _, _, _, 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, _, _, _, _, - - // 23_AB - inner-lane - 4, 5, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 6, 7, 0, 1, 2, 3, 8, 9, _, _, _, _, - - // 23_CD - inner-lane - _, _, 6, 7, 12, 13, _, _, _, _, _, _, _, _, _, _, 10, 11, 4, 5, _, _, _, _, _, _, _, _, 6, 7, 12, 13, - - // 23_EF - inner-lane - _, _, _, _, _, _, 2, 3, 8, 9, _, _, 10, 11, 4, 5, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, - - // 23_GH - inner-lane - _, _, _, _, _, _, _, _, _, _, 0, 1, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, - - // 45_AB - inner-lane - _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 10, 11, _, _, _, _, _, _, _, _, _, _, - - // 45_CD - inner-lane - _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 6, 7, 0, 1, _, _, 2, 3, 8, 9, _, _, _, _, _, _, - - // 45_EF - cross-lane - 1, 0, 0, 0, 2, 0, 0, 0, 5, 0, 0, 0, _, _, _, _, 2, 0, 0, 0, 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, - - // 45_EF - inner-lane - 2, 3, 8, 9, _, _, _, _, _, _, _, _, 10, 11, 4, 5, _, _, _, _, _, _, _, _, _, _, 2, 3, 8, 9, _, _, - - // 45_GH - inner-lane - _, _, _, _, 2, 3, 8, 9, 10, 11, 4, 5, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 6, 7, - - // 67_CD - inner-lane - _, _, _, _, _, _, _, _, _, _, 10, 11, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, - - // 67_EF - inner-lane - _, _, _, _, _, _, 6, 7, 0, 1, _, _, 2, 3, 8, 9, _, _, _, _, _, _, _, _, 10, 11, _, _, _, _, _, _, - - // 67_GH - inner-lane - 8, 9, 10, 11, 4, 5, _, _, _, _, _, _, _, _, _, _, 2, 3, 8, 9, 10, 11, 4, 5, _, _, 6, 7, 12, 13, 14, 15 +#pragma warning disable SA1515 + /* 01 */ + // [cr] crln_01_AB_CD + 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, _, _, _, _, 1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, + // (in) AB + 0, 1, 8, 9, 2, 3, 4, 5, 10, 11, _, _, _, _, _, _, 12, 13, 2, 3, 4, 5, 14, 15, _, _, _, _, _, _, _, _, + // (in) CD + _, _, _, _, _, _, _, _, _, _, 0, 1, 8, 9, 2, 3, _, _, _, _, _, _, _, _, 0, 1, 10, 11, _, _, _, _, + // [cr] crln_01_23_EF_23_CD + 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 5, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, + // (in) EF + _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 0, 1, 8, 9, + + /* 23 */ + // [cr] crln_23_AB_23_45_GH + 2, 0, 0, 0, 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, + // (in) AB + _, _, _, _, _, _, 8, 9, 2, 3, 4, 5, 10, 11, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, + // (in) CDe + _, _, 12, 13, 6, 7, _, _, _, _, _, _, _, _, 8, 9, 14, 15, _, _, _, _, _, _, _, _, _, _, _, _, _, _, + // (in) EF + 2, 3, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 4, 5, 10, 11, _, _, _, _, _, _, 12, 13, 6, 7, + // (in) GH + _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 0, 1, 8, 9, 2, 3, _, _, _, _, + + /* 45 */ + // (in) AB + _, _, _, _, 12, 13, 6, 7, 14, 15, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, + // [cr] crln_45_67_CD_45_EF + 2, 0, 0, 0, 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, 2, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, + // (in) CD + 8, 9, 2, 3, _, _, _, _, _, _, 4, 5, 10, 11, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 12, 13, + // (in) EF + _, _, _, _, _, _, _, _, _, _, _, _, _, _, 0, 1, 6, 7, _, _, _, _, _, _, _, _, 8, 9, 2, 3, _, _, + // (in) GH + _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, 4, 5, 10, 11, 12, 13, 6, 7, _, _, _, _, _, _, + + /* 67 */ + // (in) CD + 6, 7, 14, 15, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, _, + // [cr] crln_67_EF_67_GH + 2, 0, 0, 0, 3, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 3, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0, _, _, _, _, + // (in) EF + _, _, _, _, 4, 5, 14, 15, _, _, _, _, _, _, _, _, 8, 9, 2, 3, 10, 11, _, _, _, _, _, _, _, _, _, _, + // (in) GH + _, _, _, _, _, _, _, _, 0, 1, 10, 11, 12, 13, 2, 3, _, _, _, _, _, _, 0, 1, 6, 7, 8, 9, 2, 3, 10, 11, +#pragma warning restore SA1515 }; /// /// Applies zig zag ordering for given 8x8 matrix using SSE cpu intrinsics. /// /// Input matrix. - public static unsafe void ApplyZigZagOrderingSsse3(ref Block8x8 block) + public static unsafe void ApplyTransposingZigZagOrderingSsse3(ref Block8x8 block) { DebugGuard.IsTrue(Ssse3.IsSupported, "Ssse3 support is required to run this operation!"); - fixed (byte* maskPtr = SseShuffleMasks) + fixed (byte* shuffleVectorsPtr = &MemoryMarshal.GetReference(SseShuffleMasks)) { Vector128 rowA = block.V0.AsByte(); Vector128 rowB = block.V1.AsByte(); @@ -142,73 +161,69 @@ public static unsafe void ApplyZigZagOrderingSsse3(ref Block8x8 block) Vector128 rowG = block.V6.AsByte(); Vector128 rowH = block.V7.AsByte(); - // row0 - A0 A1 B0 C0 B1 A2 A3 B2 - Vector128 rowA0 = Ssse3.Shuffle(rowA, Sse2.LoadVector128(maskPtr + (16 * 0))).AsInt16(); - Vector128 rowB0 = Ssse3.Shuffle(rowB, Sse2.LoadVector128(maskPtr + (16 * 1))).AsInt16(); - Vector128 row0 = Sse2.Or(rowA0, rowB0); - Vector128 rowC0 = Ssse3.Shuffle(rowC, Sse2.LoadVector128(maskPtr + (16 * 2))).AsInt16(); - row0 = Sse2.Or(row0, rowC0); - - // row1 - C1 D0 E0 D1 C2 B3 A4 A5 - Vector128 rowA1 = Ssse3.Shuffle(rowA, Sse2.LoadVector128(maskPtr + (16 * 3))).AsInt16(); - Vector128 rowC1 = Ssse3.Shuffle(rowC, Sse2.LoadVector128(maskPtr + (16 * 4))).AsInt16(); - Vector128 row1 = Sse2.Or(rowA1, rowC1); - Vector128 rowD1 = Ssse3.Shuffle(rowD, Sse2.LoadVector128(maskPtr + (16 * 5))).AsInt16(); - row1 = Sse2.Or(row1, rowD1); - row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowB.AsUInt16(), 3), 5).AsInt16(); - row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 0), 2).AsInt16(); - - // row2 - Vector128 rowE2 = Ssse3.Shuffle(rowE, Sse2.LoadVector128(maskPtr + (16 * 6))).AsInt16(); - Vector128 rowF2 = Ssse3.Shuffle(rowF, Sse2.LoadVector128(maskPtr + (16 * 7))).AsInt16(); - Vector128 row2 = Sse2.Or(rowE2, rowF2); - row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowB.AsUInt16(), 4), 0).AsInt16(); - row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowC.AsUInt16(), 3), 1).AsInt16(); - row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 2), 2).AsInt16(); - row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowG.AsUInt16(), 0), 5).AsInt16(); - - // row3 - Vector128 rowA3 = Ssse3.Shuffle(rowA, Sse2.LoadVector128(maskPtr + (16 * 8))).AsInt16().AsInt16(); - Vector128 rowB3 = Ssse3.Shuffle(rowB, Sse2.LoadVector128(maskPtr + (16 * 9))).AsInt16().AsInt16(); - Vector128 row3 = Sse2.Or(rowA3, rowB3); - Vector128 rowC3 = Ssse3.Shuffle(rowC, Sse2.LoadVector128(maskPtr + (16 * 10))).AsInt16(); - row3 = Sse2.Or(row3, rowC3); - Vector128 shuffleRowD3EF = Sse2.LoadVector128(maskPtr + (16 * 11)); - Vector128 rowD3 = Ssse3.Shuffle(rowD, shuffleRowD3EF).AsInt16(); - row3 = Sse2.Or(row3, rowD3); - - // row4 - Vector128 rowE4 = Ssse3.Shuffle(rowE, shuffleRowD3EF).AsInt16(); - Vector128 rowF4 = Ssse3.Shuffle(rowF, Sse2.LoadVector128(maskPtr + (16 * 12))).AsInt16(); - Vector128 row4 = Sse2.Or(rowE4, rowF4); - Vector128 rowG4 = Ssse3.Shuffle(rowG, Sse2.LoadVector128(maskPtr + (16 * 13))).AsInt16(); - row4 = Sse2.Or(row4, rowG4); - Vector128 rowH4 = Ssse3.Shuffle(rowH, Sse2.LoadVector128(maskPtr + (16 * 14))).AsInt16(); - row4 = Sse2.Or(row4, rowH4); - - // row5 - Vector128 rowC5 = Ssse3.Shuffle(rowC, Sse2.LoadVector128(maskPtr + (16 * 15))).AsInt16(); - Vector128 rowD5 = Ssse3.Shuffle(rowD, Sse2.LoadVector128(maskPtr + (16 * 16))).AsInt16(); - Vector128 row5 = Sse2.Or(rowC5, rowD5); - row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowB.AsUInt16(), 7), 2).AsInt16(); - row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 5), 5).AsInt16(); - row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowF.AsUInt16(), 4), 6).AsInt16(); - row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowG.AsUInt16(), 3), 7).AsInt16(); - - // row6 - Vector128 rowE6 = Ssse3.Shuffle(rowE, Sse2.LoadVector128(maskPtr + (16 * 17))).AsInt16(); - Vector128 rowF6 = Ssse3.Shuffle(rowF, Sse2.LoadVector128(maskPtr + (16 * 18))).AsInt16(); - Vector128 row6 = Sse2.Or(rowE6, rowF6); - Vector128 rowH6 = Ssse3.Shuffle(rowH, Sse2.LoadVector128(maskPtr + (16 * 19))).AsInt16(); - row6 = Sse2.Or(row6, rowH6); - row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 7), 5).AsInt16(); - row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowG.AsUInt16(), 4), 2).AsInt16(); - - // row7 - Vector128 rowG7 = Ssse3.Shuffle(rowG, Sse2.LoadVector128(maskPtr + (16 * 20))).AsInt16(); - Vector128 rowH7 = Ssse3.Shuffle(rowH, Sse2.LoadVector128(maskPtr + (16 * 21))).AsInt16(); - Vector128 row7 = Sse2.Or(rowG7, rowH7); - row7 = Sse2.Insert(row7.AsUInt16(), Sse2.Extract(rowF.AsUInt16(), 7), 4).AsInt16(); + // row0 - A0 B0 A1 A2 B1 C0 D0 C1 + Vector128 row0_A = Ssse3.Shuffle(rowA, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 0))).AsInt16(); + Vector128 row0_B = Ssse3.Shuffle(rowB, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 1))).AsInt16(); + Vector128 row0_C = Ssse3.Shuffle(rowC, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 2))).AsInt16(); + Vector128 row0 = Sse2.Or(Sse2.Or(row0_A, row0_B), row0_C); + row0 = Sse2.Insert(row0.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 0), 6).AsInt16(); + + // row1 - B2 A3 A4 B3 C2 D1 E0 F0 + Vector128 row1_A = Ssse3.Shuffle(rowA, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 3))).AsInt16(); + Vector128 row1_B = Ssse3.Shuffle(rowB, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 4))).AsInt16(); + Vector128 row1 = Sse2.Or(row1_A, row1_B); + row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowC.AsUInt16(), 2), 4).AsInt16(); + row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 1), 5).AsInt16(); + row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 0), 6).AsInt16(); + row1 = Sse2.Insert(row1.AsUInt16(), Sse2.Extract(rowF.AsUInt16(), 0), 7).AsInt16(); + + // row2 - E1 D2 C3 B4 A5 A6 B5 C4 + Vector128 row2_A = Ssse3.Shuffle(rowA, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 5))).AsInt16(); + Vector128 row2_B = Ssse3.Shuffle(rowB, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 6))).AsInt16(); + Vector128 row2_C = Ssse3.Shuffle(rowC, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 7))).AsInt16(); + Vector128 row2 = Sse2.Or(Sse2.Or(row2_A, row2_B), row2_C); + row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 2), 1).AsInt16(); + row2 = Sse2.Insert(row2.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 1), 0).AsInt16(); + + // row3 - D3 E2 F1 G0 H0 G1 F2 E3 + Vector128 row3_E = Ssse3.Shuffle(rowE, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 8))).AsInt16(); + Vector128 row3_F = Ssse3.Shuffle(rowF, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 9))).AsInt16(); + Vector128 row3_G = Ssse3.Shuffle(rowG, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 10))).AsInt16(); + Vector128 row3 = Sse2.Or(Sse2.Or(row3_E, row3_F), row3_G); + row3 = Sse2.Insert(row3.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 3), 0).AsInt16(); + row3 = Sse2.Insert(row3.AsUInt16(), Sse2.Extract(rowH.AsUInt16(), 0), 4).AsInt16(); + + // row4 - D4 C5 B6 A7 B7 C6 D5 E4 + Vector128 row4_B = Ssse3.Shuffle(rowB, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 11))).AsInt16(); + Vector128 row4_C = Ssse3.Shuffle(rowC, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 12))).AsInt16(); + Vector128 row4_D = Ssse3.Shuffle(rowD, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 13))).AsInt16(); + Vector128 row4 = Sse2.Or(Sse2.Or(row4_B, row4_C), row4_D); + row4 = Sse2.Insert(row4.AsUInt16(), Sse2.Extract(rowA.AsUInt16(), 7), 3).AsInt16(); + row4 = Sse2.Insert(row4.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 4), 7).AsInt16(); + + // row5 - F3 G2 H1 H2 G3 F4 E5 D6 + Vector128 row5_F = Ssse3.Shuffle(rowF, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 14))).AsInt16(); + Vector128 row5_G = Ssse3.Shuffle(rowG, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 15))).AsInt16(); + Vector128 row5_H = Ssse3.Shuffle(rowH, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 16))).AsInt16(); + Vector128 row5 = Sse2.Or(Sse2.Or(row5_F, row5_G), row5_H); + row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 6), 7).AsInt16(); + row5 = Sse2.Insert(row5.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 5), 6).AsInt16(); + + // row6 - C7 D7 E6 F5 G4 H3 H4 G5 + Vector128 row6_G = Ssse3.Shuffle(rowG, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 17))).AsInt16(); + Vector128 row6_H = Ssse3.Shuffle(rowH, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 18))).AsInt16(); + Vector128 row6 = Sse2.Or(row6_G, row6_H); + row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowC.AsUInt16(), 7), 0).AsInt16(); + row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowD.AsUInt16(), 7), 1).AsInt16(); + row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 6), 2).AsInt16(); + row6 = Sse2.Insert(row6.AsUInt16(), Sse2.Extract(rowF.AsUInt16(), 5), 3).AsInt16(); + + // row7 - F6 E7 F7 G6 H5 H6 G7 H7 + Vector128 row7_F = Ssse3.Shuffle(rowF, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 19))).AsInt16(); + Vector128 row7_G = Ssse3.Shuffle(rowG, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 20))).AsInt16(); + Vector128 row7_H = Ssse3.Shuffle(rowH, Sse2.LoadVector128(shuffleVectorsPtr + (16 * 21))).AsInt16(); + Vector128 row7 = Sse2.Or(Sse2.Or(row7_F, row7_G), row7_H); + row7 = Sse2.Insert(row7.AsUInt16(), Sse2.Extract(rowE.AsUInt16(), 7), 1).AsInt16(); block.V0 = row0; block.V1 = row1; @@ -225,69 +240,61 @@ public static unsafe void ApplyZigZagOrderingSsse3(ref Block8x8 block) /// Applies zig zag ordering for given 8x8 matrix using AVX cpu intrinsics. /// /// Input matrix. - public static unsafe void ApplyZigZagOrderingAvx2(ref Block8x8 block) + public static unsafe void ApplyTransposingZigZagOrderingAvx2(ref Block8x8 block) { DebugGuard.IsTrue(Avx2.IsSupported, "Avx2 support is required to run this operation!"); - fixed (byte* shuffleVectorsPtr = AvxShuffleMasks) + fixed (byte* shuffleVectorsPtr = &MemoryMarshal.GetReference(AvxShuffleMasks)) { - Vector256 rowsAB = block.V01.AsByte(); - Vector256 rowsCD = block.V23.AsByte(); - Vector256 rowsEF = block.V45.AsByte(); - Vector256 rowsGH = block.V67.AsByte(); - - // rows 0 1 - Vector256 rows_AB01_EF01_CD23_shuffleMask = Avx.LoadVector256(shuffleVectorsPtr + (0 * 32)).AsInt32(); - Vector256 row01_AB = Avx2.PermuteVar8x32(rowsAB.AsInt32(), rows_AB01_EF01_CD23_shuffleMask).AsByte(); + Vector256 rowAB = block.V01.AsByte(); + Vector256 rowCD = block.V23.AsByte(); + Vector256 rowEF = block.V45.AsByte(); + Vector256 rowGH = block.V67.AsByte(); + + /* row01 - A0 B0 A1 A2 B1 C0 D0 C1 | B2 A3 A4 B3 C2 D1 E0 F0 */ + Vector256 crln_01_AB_CD = Avx.LoadVector256(shuffleVectorsPtr + (0 * 32)).AsInt32(); + Vector256 row01_AB = Avx2.PermuteVar8x32(rowAB.AsInt32(), crln_01_AB_CD).AsByte(); row01_AB = Avx2.Shuffle(row01_AB, Avx.LoadVector256(shuffleVectorsPtr + (1 * 32))).AsByte(); - - Vector256 rows_CD01_GH23_shuffleMask = Avx.LoadVector256(shuffleVectorsPtr + (2 * 32)).AsInt32(); - Vector256 row01_CD = Avx2.PermuteVar8x32(rowsCD.AsInt32(), rows_CD01_GH23_shuffleMask).AsByte(); - row01_CD = Avx2.Shuffle(row01_CD, Avx.LoadVector256(shuffleVectorsPtr + (3 * 32))).AsByte(); - - Vector256 row0123_EF = Avx2.PermuteVar8x32(rowsEF.AsInt32(), rows_AB01_EF01_CD23_shuffleMask).AsByte(); - Vector256 row01_EF = Avx2.Shuffle(row0123_EF, Avx.LoadVector256(shuffleVectorsPtr + (4 * 32))).AsByte(); - - Vector256 row01 = Avx2.Or(Avx2.Or(row01_AB, row01_CD), row01_EF); - - // rows 2 3 - Vector256 rows_AB23_CD45_EF67_shuffleMask = Avx.LoadVector256(shuffleVectorsPtr + (5 * 32)).AsInt32(); - Vector256 row2345_AB = Avx2.PermuteVar8x32(rowsAB.AsInt32(), rows_AB23_CD45_EF67_shuffleMask).AsByte(); - Vector256 row23_AB = Avx2.Shuffle(row2345_AB, Avx.LoadVector256(shuffleVectorsPtr + (6 * 32))).AsByte(); - - Vector256 row23_CD = Avx2.PermuteVar8x32(rowsCD.AsInt32(), rows_AB01_EF01_CD23_shuffleMask).AsByte(); + Vector256 row01_CD = Avx2.PermuteVar8x32(rowCD.AsInt32(), crln_01_AB_CD).AsByte(); + row01_CD = Avx2.Shuffle(row01_CD, Avx.LoadVector256(shuffleVectorsPtr + (2 * 32))).AsByte(); + Vector256 crln_01_23_EF_23_CD = Avx.LoadVector256(shuffleVectorsPtr + (3 * 32)).AsInt32(); + Vector256 row01_23_EF = Avx2.PermuteVar8x32(rowEF.AsInt32(), crln_01_23_EF_23_CD).AsByte(); + Vector256 row01_EF = Avx2.Shuffle(row01_23_EF, Avx.LoadVector256(shuffleVectorsPtr + (4 * 32))).AsByte(); + + Vector256 row01 = Avx2.Or(row01_AB, Avx2.Or(row01_CD, row01_EF)); + + /* row23 - E1 D2 C3 B4 A5 A6 B5 C4 | D3 E2 F1 G0 H0 G1 F2 E3 */ + Vector256 crln_23_AB_23_45_GH = Avx.LoadVector256(shuffleVectorsPtr + (5 * 32)).AsInt32(); + Vector256 row23_45_AB = Avx2.PermuteVar8x32(rowAB.AsInt32(), crln_23_AB_23_45_GH).AsByte(); + Vector256 row23_AB = Avx2.Shuffle(row23_45_AB, Avx.LoadVector256(shuffleVectorsPtr + (6 * 32))).AsByte(); + Vector256 row23_CD = Avx2.PermuteVar8x32(rowCD.AsInt32(), crln_01_23_EF_23_CD).AsByte(); row23_CD = Avx2.Shuffle(row23_CD, Avx.LoadVector256(shuffleVectorsPtr + (7 * 32))).AsByte(); - - Vector256 row23_EF = Avx2.Shuffle(row0123_EF, Avx.LoadVector256(shuffleVectorsPtr + (8 * 32))).AsByte(); - - Vector256 row2345_GH = Avx2.PermuteVar8x32(rowsGH.AsInt32(), rows_CD01_GH23_shuffleMask).AsByte(); - Vector256 row23_GH = Avx2.Shuffle(row2345_GH, Avx.LoadVector256(shuffleVectorsPtr + (9 * 32)).AsByte()); + Vector256 row23_EF = Avx2.Shuffle(row01_23_EF, Avx.LoadVector256(shuffleVectorsPtr + (8 * 32))).AsByte(); + Vector256 row23_45_GH = Avx2.PermuteVar8x32(rowGH.AsInt32(), crln_23_AB_23_45_GH).AsByte(); + Vector256 row23_GH = Avx2.Shuffle(row23_45_GH, Avx.LoadVector256(shuffleVectorsPtr + (9 * 32))).AsByte(); Vector256 row23 = Avx2.Or(Avx2.Or(row23_AB, row23_CD), Avx2.Or(row23_EF, row23_GH)); - // rows 4 5 - Vector256 row45_AB = Avx2.Shuffle(row2345_AB, Avx.LoadVector256(shuffleVectorsPtr + (10 * 32)).AsByte()); - Vector256 row4567_CD = Avx2.PermuteVar8x32(rowsCD.AsInt32(), rows_AB23_CD45_EF67_shuffleMask).AsByte(); - Vector256 row45_CD = Avx2.Shuffle(row4567_CD, Avx.LoadVector256(shuffleVectorsPtr + (11 * 32)).AsByte()); - - Vector256 rows_EF45_GH67_shuffleMask = Avx.LoadVector256(shuffleVectorsPtr + (12 * 32)).AsInt32(); - Vector256 row45_EF = Avx2.PermuteVar8x32(rowsEF.AsInt32(), rows_EF45_GH67_shuffleMask).AsByte(); - row45_EF = Avx2.Shuffle(row45_EF, Avx.LoadVector256(shuffleVectorsPtr + (13 * 32)).AsByte()); - - Vector256 row45_GH = Avx2.Shuffle(row2345_GH, Avx.LoadVector256(shuffleVectorsPtr + (14 * 32)).AsByte()); + /* row45 - D4 C5 B6 A7 B7 C6 D5 E4 | F3 G2 H1 H2 G3 F4 E5 D6 */ + Vector256 row45_AB = Avx2.Shuffle(row23_45_AB, Avx.LoadVector256(shuffleVectorsPtr + (10 * 32))).AsByte(); + Vector256 crln_45_67_CD_45_EF = Avx.LoadVector256(shuffleVectorsPtr + (11 * 32)).AsInt32(); + Vector256 row45_67_CD = Avx2.PermuteVar8x32(rowCD.AsInt32(), crln_45_67_CD_45_EF).AsByte(); + Vector256 row45_CD = Avx2.Shuffle(row45_67_CD, Avx.LoadVector256(shuffleVectorsPtr + (12 * 32))).AsByte(); + Vector256 row45_EF = Avx2.PermuteVar8x32(rowEF.AsInt32(), crln_45_67_CD_45_EF).AsByte(); + row45_EF = Avx2.Shuffle(row45_EF, Avx.LoadVector256(shuffleVectorsPtr + (13 * 32))).AsByte(); + Vector256 row45_GH = Avx2.Shuffle(row23_45_GH, Avx.LoadVector256(shuffleVectorsPtr + (14 * 32))).AsByte(); Vector256 row45 = Avx2.Or(Avx2.Or(row45_AB, row45_CD), Avx2.Or(row45_EF, row45_GH)); - // rows 6 7 - Vector256 row67_CD = Avx2.Shuffle(row4567_CD, Avx.LoadVector256(shuffleVectorsPtr + (15 * 32)).AsByte()); - - Vector256 row67_EF = Avx2.PermuteVar8x32(rowsEF.AsInt32(), rows_AB23_CD45_EF67_shuffleMask).AsByte(); - row67_EF = Avx2.Shuffle(row67_EF, Avx.LoadVector256(shuffleVectorsPtr + (16 * 32)).AsByte()); - - Vector256 row67_GH = Avx2.PermuteVar8x32(rowsGH.AsInt32(), rows_EF45_GH67_shuffleMask).AsByte(); - row67_GH = Avx2.Shuffle(row67_GH, Avx.LoadVector256(shuffleVectorsPtr + (17 * 32)).AsByte()); + /* row67 - C7 D7 E6 F5 G4 H3 H4 G5 | F6 E7 F7 G6 H5 H6 G7 H7 */ + Vector256 row67_CD = Avx2.Shuffle(row45_67_CD, Avx.LoadVector256(shuffleVectorsPtr + (15 * 32))).AsByte(); + Vector256 crln_67_EF_67_GH = Avx.LoadVector256(shuffleVectorsPtr + (16 * 32)).AsInt32(); + Vector256 row67_EF = Avx2.PermuteVar8x32(rowEF.AsInt32(), crln_67_EF_67_GH).AsByte(); + row67_EF = Avx2.Shuffle(row67_EF, Avx.LoadVector256(shuffleVectorsPtr + (17 * 32))).AsByte(); + Vector256 row67_GH = Avx2.PermuteVar8x32(rowGH.AsInt32(), crln_67_EF_67_GH).AsByte(); + row67_GH = Avx2.Shuffle(row67_GH, Avx.LoadVector256(shuffleVectorsPtr + (18 * 32))).AsByte(); - Vector256 row67 = Avx2.Or(Avx2.Or(row67_CD, row67_EF), row67_GH); + Vector256 row67 = Avx2.Or(row67_CD, Avx2.Or(row67_EF, row67_GH)); block.V01 = row01.AsInt16(); block.V23 = row23.AsInt16(); diff --git a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs index ae7e81254b..9576cbd3c8 100644 --- a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs +++ b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs @@ -220,7 +220,7 @@ static void RunTest(string srcSeedSerialized, string qtSeedSerialized) // Reference implementation quantizes given block via division Block8x8 expected = default; - ReferenceImplementations.Quantize(ref source, ref expected, ref quant, ZigZag.ZigZagOrder); + ReferenceImplementations.Quantize(ref source, ref expected, ref quant, ZigZag.TransposingOrder); // Actual current implementation quantizes given block via multiplication // With quantization table reciprocal diff --git a/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs b/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs index 36570ce55a..9c467a1cc9 100644 --- a/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs +++ b/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs @@ -135,10 +135,9 @@ static void RunTest(string serialized) FastFloatingPointDCT.AdjustToIDCT(ref dequantMatrix); srcBlock.MultiplyInPlace(ref dequantMatrix); + // testee // IDCT implementation tranforms blocks after transposition srcBlock.TransposeInplace(); - - // IDCT calculation FastFloatingPointDCT.TransformIDCT(ref srcBlock); float[] actualDest = srcBlock.ToArray(); @@ -180,7 +179,10 @@ static void RunTest(string serialized) ReferenceImplementations.LLM_FloatingPoint_DCT.FDCT2D_llm(src, expectedDest, temp1, downscaleBy8: true); // testee + // Second transpose call is done by Quantize step + // Do this manually here just to be complient to the reference implementation FastFloatingPointDCT.TransformFDCT(ref block); + block.TransposeInplace(); // Part of the IDCT calculations is fused into the quantization step // We must multiply input block with adjusted no-quantization matrix