Initial processor implementation, code base for tests

src/ImageSharp/Formats/Jpeg/Components/Decoder/ComponentProcessors/DownScalingComponentProcessor2.cs

@@ -0,0 +1,157 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Runtime.CompilerServices;
+using SixLabors.ImageSharp.Memory;
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
+{
+    internal sealed class DownScalingComponentProcessor2 : ComponentProcessor
+    {
+        private readonly IRawJpegData rawJpeg;
+
+        public DownScalingComponentProcessor2(MemoryAllocator memoryAllocator, JpegFrame frame, IRawJpegData rawJpeg, Size postProcessorBufferSize, IJpegComponent component)
+            : base(memoryAllocator, frame, postProcessorBufferSize, component, 4)
+            => this.rawJpeg = rawJpeg;
+
+        public override void CopyBlocksToColorBuffer(int spectralStep)
+        {
+            Buffer2D<Block8x8> spectralBuffer = this.Component.SpectralBlocks;
+
+            float maximumValue = this.Frame.MaxColorChannelValue;
+            float normalizationValue = MathF.Ceiling(maximumValue / 2);
+
+            int destAreaStride = this.ColorBuffer.Width;
+
+            int blocksRowsPerStep = this.Component.SamplingFactors.Height;
+            Size subSamplingDivisors = this.Component.SubSamplingDivisors;
+
+            Block8x8F dequantTable = this.rawJpeg.QuantizationTables[this.Component.QuantizationTableIndex];
+            Block8x8F workspaceBlock = default;
+
+            int yBlockStart = spectralStep * blocksRowsPerStep;
+
+            for (int y = 0; y < blocksRowsPerStep; y++)
+            {
+                int yBuffer = y * this.BlockAreaSize.Height;
+
+                Span<float> colorBufferRow = this.ColorBuffer.DangerousGetRowSpan(yBuffer);
+                Span<Block8x8> blockRow = spectralBuffer.DangerousGetRowSpan(yBlockStart + y);
+
+                for (int xBlock = 0; xBlock < spectralBuffer.Width; xBlock++)
+                {
+                    // Integer to float
+                    workspaceBlock.LoadFrom(ref blockRow[xBlock]);
+
+                    // IDCT/Normalization/Range
+                    TransformIDCT_4x4(ref workspaceBlock, ref dequantTable, normalizationValue, maximumValue);
+
+                    // Save to the intermediate buffer
+                    int xColorBufferStart = xBlock * this.BlockAreaSize.Width;
+                    ScaledCopyTo(
+                        ref workspaceBlock,
+                        ref colorBufferRow[xColorBufferStart],
+                        destAreaStride,
+                        subSamplingDivisors.Width,
+                        subSamplingDivisors.Height);
+                }
+            }
+        }
+
+        public static void TransformIDCT_4x4(ref Block8x8F block, ref Block8x8F dequantTable, float normalizationValue, float maxValue)
+        {
+            const int DCTSIZE = 8;
+            const float FIX_0_541196100 = 0.541196100f;
+            const float FIX_0_765366865 = 0.765366865f;
+            const float FIX_1_847759065 = 1.847759065f;
+
+            // input block is transposed so term indices must be tranposed too
+            float tmp0, tmp2, tmp10, tmp12;
+            float z1, z2, z3;
+
+            for (int ctr = 0; ctr < 4; ctr++)
+            {
+                // Even part
+                tmp0 = block[ctr * DCTSIZE] * dequantTable[ctr * DCTSIZE];
+                tmp2 = block[(ctr * DCTSIZE) + 2] * dequantTable[(ctr * DCTSIZE) + 2];
+
+                tmp10 = tmp0 + tmp2;
+                tmp12 = tmp0 - tmp2;
+
+                // Odd part
+                z2 = block[(ctr * DCTSIZE) + 1] * dequantTable[(ctr * DCTSIZE) + 1];
+                z3 = block[(ctr * DCTSIZE) + 3] * dequantTable[(ctr * DCTSIZE) + 3];
+
+                z1 = (z2 + z3) * FIX_0_541196100;
+                tmp0 = z1 + (z2 * FIX_0_765366865);
+                tmp2 = z1 - (z3 * FIX_1_847759065);
+
+                /* Final output stage */
+                block[ctr + 4] = tmp10 + tmp0;
+                block[ctr + 28] = tmp10 - tmp0;
+                block[ctr + 12] = tmp12 + tmp2;
+                block[ctr + 20] = tmp12 - tmp2;
+            }
+
+            for (int ctr = 0; ctr < 4; ctr++)
+            {
+                // Even part
+                tmp0 = block[(ctr * 8) + 0 + 4];
+                tmp2 = block[(ctr * 8) + 2 + 4];
+
+                tmp10 = tmp0 + tmp2;
+                tmp12 = tmp0 - tmp2;
+
+                // Odd part
+                z2 = block[(ctr * 8) + 1 + 4];
+                z3 = block[(ctr * 8) + 3 + 4];
+
+                z1 = (z2 + z3) * FIX_0_541196100;
+                tmp0 = z1 + (z2 * FIX_0_765366865);
+                tmp2 = z1 - (z3 * FIX_1_847759065);
+
+                /* Final output stage */
+                block[(ctr * 8) + 0] = (float)Math.Round(Numerics.Clamp(tmp10 + tmp0 + normalizationValue, 0, maxValue));
+                block[(ctr * 8) + 3] = (float)Math.Round(Numerics.Clamp(tmp10 - tmp0 + normalizationValue, 0, maxValue));
+                block[(ctr * 8) + 1] = (float)Math.Round(Numerics.Clamp(tmp12 + tmp2 + normalizationValue, 0, maxValue));
+                block[(ctr * 8) + 2] = (float)Math.Round(Numerics.Clamp(tmp12 - tmp2 + normalizationValue, 0, maxValue));
+            }
+        }
+
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void ScaledCopyTo(ref Block8x8F block, ref float destRef, int destStrideWidth, int horizontalScale, int verticalScale)
+        {
+            // TODO: Optimize: implement all cases with scale-specific, loopless code!
+            CopyArbitraryScale(ref block, ref destRef, destStrideWidth, horizontalScale, verticalScale);
+
+            [MethodImpl(InliningOptions.ColdPath)]
+            static void CopyArbitraryScale(ref Block8x8F block, ref float areaOrigin, int areaStride, int horizontalScale, int verticalScale)
+            {
+                for (int y = 0; y < 4; y++)
+                {
+                    int yy = y * verticalScale;
+                    int y8 = y * 8;
+
+                    for (int x = 0; x < 4; x++)
+                    {
+                        int xx = x * horizontalScale;
+
+                        float value = block[y8 + x];
+
+                        for (int i = 0; i < verticalScale; i++)
+                        {
+                            int baseIdx = ((yy + i) * areaStride) + xx;
+
+                            for (int j = 0; j < horizontalScale; j++)
+                            {
+                                // area[xx + j, yy + i] = value;
+                                Unsafe.Add(ref areaOrigin, baseIdx + j) = value;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ComponentProcessors/DownScalingComponentProcessor4.cs

@@ -0,0 +1,127 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Runtime.CompilerServices;
+using SixLabors.ImageSharp.Memory;
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
+{
+    internal sealed class DownScalingComponentProcessor4 : ComponentProcessor
+    {
+        private readonly IRawJpegData rawJpeg;
+
+        public DownScalingComponentProcessor4(MemoryAllocator memoryAllocator, JpegFrame frame, IRawJpegData rawJpeg, Size postProcessorBufferSize, IJpegComponent component)
+            : base(memoryAllocator, frame, postProcessorBufferSize, component, 2)
+            => this.rawJpeg = rawJpeg;
+
+        public override void CopyBlocksToColorBuffer(int spectralStep)
+        {
+            Buffer2D<Block8x8> spectralBuffer = this.Component.SpectralBlocks;
+
+            float maximumValue = this.Frame.MaxColorChannelValue;
+            float normalizationValue = MathF.Ceiling(maximumValue / 2);
+
+            int destAreaStride = this.ColorBuffer.Width;
+
+            int blocksRowsPerStep = this.Component.SamplingFactors.Height;
+            Size subSamplingDivisors = this.Component.SubSamplingDivisors;
+
+            Block8x8F dequantTable = this.rawJpeg.QuantizationTables[this.Component.QuantizationTableIndex];
+            Block8x8F workspaceBlock = default;
+
+            int yBlockStart = spectralStep * blocksRowsPerStep;
+
+            for (int y = 0; y < blocksRowsPerStep; y++)
+            {
+                int yBuffer = y * this.BlockAreaSize.Height;
+
+                Span<float> colorBufferRow = this.ColorBuffer.DangerousGetRowSpan(yBuffer);
+                Span<Block8x8> blockRow = spectralBuffer.DangerousGetRowSpan(yBlockStart + y);
+
+                for (int xBlock = 0; xBlock < spectralBuffer.Width; xBlock++)
+                {
+                    // Integer to float
+                    workspaceBlock.LoadFrom(ref blockRow[xBlock]);
+
+                    // IDCT/Normalization/Range
+                    TransformIDCT_2x2(ref workspaceBlock, ref dequantTable, normalizationValue, maximumValue);
+
+                    // Save to the intermediate buffer
+                    int xColorBufferStart = xBlock * this.BlockAreaSize.Width;
+                    ScaledCopyTo(
+                        ref workspaceBlock,
+                        ref colorBufferRow[xColorBufferStart],
+                        destAreaStride,
+                        subSamplingDivisors.Width,
+                        subSamplingDivisors.Height);
+                }
+            }
+        }
+
+        public static void TransformIDCT_2x2(ref Block8x8F block, ref Block8x8F dequantTable, float normalizationValue, float maxValue)
+        {
+            // input block is transposed so term indices must be tranposed too
+            float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+            // 0
+            //   => 0 1
+            // 8
+            tmp4 = block[0] * dequantTable[0];
+            tmp5 = block[1] * dequantTable[1];
+            tmp0 = tmp4 + tmp5;
+            tmp2 = tmp4 - tmp5;
+
+            // 1
+            //   => 8 9
+            // 9
+            tmp4 = block[8] * dequantTable[8];
+            tmp5 = block[9] * dequantTable[9];
+            tmp1 = tmp4 + tmp5;
+            tmp3 = tmp4 - tmp5;
+
+            // Row 0
+            block[0] = (float)Math.Round(Numerics.Clamp(tmp0 + tmp1 + normalizationValue, 0, maxValue));
+            block[1] = (float)Math.Round(Numerics.Clamp(tmp0 - tmp1 + normalizationValue, 0, maxValue));
+
+            // Row 1
+            block[8] = (float)Math.Round(Numerics.Clamp(tmp2 + tmp3 + normalizationValue, 0, maxValue));
+            block[9] = (float)Math.Round(Numerics.Clamp(tmp2 - tmp3 + normalizationValue, 0, maxValue));
+        }
+
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void ScaledCopyTo(ref Block8x8F block, ref float destRef, int destStrideWidth, int horizontalScale, int verticalScale)
+        {
+            // TODO: Optimize: implement all cases with scale-specific, loopless code!
+            CopyArbitraryScale(ref block, ref destRef, destStrideWidth, horizontalScale, verticalScale);
+
+            [MethodImpl(InliningOptions.ColdPath)]
+            static void CopyArbitraryScale(ref Block8x8F block, ref float areaOrigin, int areaStride, int horizontalScale, int verticalScale)
+            {
+                for (int y = 0; y < 2; y++)
+                {
+                    int yy = y * verticalScale;
+                    int y8 = y * 8;
+
+                    for (int x = 0; x < 2; x++)
+                    {
+                        int xx = x * horizontalScale;
+
+                        float value = block[y8 + x];
+
+                        for (int i = 0; i < verticalScale; i++)
+                        {
+                            int baseIdx = ((yy + i) * areaStride) + xx;
+
+                            for (int j = 0; j < horizontalScale; j++)
+                            {
+                                // area[xx + j, yy + i] = value;
+                                Unsafe.Add(ref areaOrigin, baseIdx + j) = value;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ComponentProcessors/DownScalingComponentProcessor8.cs

@@ -38,14 +38,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
 
                 for (int xBlock = 0; xBlock < spectralBuffer.Width; xBlock++)
                 {
-                    // get direct current term - averaged 8x8 pixel value
+                    float dc = TransformIDCT_1x1(blockRow[xBlock][0], this.dcDequantizer, normalizationValue, maximumValue);
-                    float dc = blockRow[xBlock][0];
-
-                    // dequantization
-                    dc *= this.dcDequantizer;
-
-                    // Normalize & round
-                    dc = (float)Math.Round(Numerics.Clamp(dc + normalizationValue, 0, maximumValue));
 
                     // Save to the intermediate buffer
                     int xColorBufferStart = xBlock * this.BlockAreaSize.Width;
@@ -59,6 +52,13 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
             }
         }
 
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static float TransformIDCT_1x1(float dc, float dequantizer, float normalizationValue, float maxValue)
+        {
+            dc *= dequantizer;
+            return (float)Math.Round(Numerics.Clamp(dc + normalizationValue, 0, maxValue));
+        }
+
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ScaledCopyTo(float value, ref float destRef, int destStrideWidth, int horizontalScale, int verticalScale)
         {

src/ImageSharp/Formats/Jpeg/Components/Decoder/SpectralConverter{TPixel}.cs

@@ -33,6 +33,12 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
             // 8 => 8, no scaling
             8,
 
+            // 8 => 4, 1/2 of the original size
+            4,
+
+            // 8 => 2, 1/4 of the original size
+            2,
+
             // 8 => 1, 1/8 of the original size
             1,
         };
@@ -256,6 +262,20 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
                         this.componentProcessors[i] = new DirectComponentProcessor(allocator, frame, jpegData, postProcessorBufferSize, frame.Components[i]);
                     }
 
+                    break;
+                case 4:
+                    for (int i = 0; i < this.componentProcessors.Length; i++)
+                    {
+                        this.componentProcessors[i] = new DownScalingComponentProcessor2(allocator, frame, jpegData, postProcessorBufferSize, frame.Components[i]);
+                    }
+
+                    break;
+                case 2:
+                    for (int i = 0; i < this.componentProcessors.Length; i++)
+                    {
+                        this.componentProcessors[i] = new DownScalingComponentProcessor4(allocator, frame, jpegData, postProcessorBufferSize, frame.Components[i]);
+                    }
+
                     break;
                 case 1:
                     for (int i = 0; i < this.componentProcessors.Length; i++)

tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs

@@ -3,6 +3,7 @@
 
 using System;
 using SixLabors.ImageSharp.Formats.Jpeg.Components;
+using SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder;
 using SixLabors.ImageSharp.Tests.Formats.Jpg.Utils;
 using SixLabors.ImageSharp.Tests.TestUtilities;
 using Xunit;
@@ -14,8 +15,13 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
     [Trait("Format", "Jpg")]
     public static class DCTTests
     {
-        private const int MaxAllowedValue = short.MaxValue;
+        // size of input values is 10 bit max
-        private const int MinAllowedValue = short.MinValue;
+        private const float MaxInputValue = 1023;
+        private const float MinInputValue = -1024;
+
+        // output value range is 12 bit max
+        private const float MaxOutputValue = 4096;
+        private const float NormalizationValue = MaxOutputValue / 2;
 
         internal static Block8x8F CreateBlockFromScalar(float value)
         {
@@ -41,7 +47,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             [InlineData(3)]
             public void LLM_TransformIDCT_CompareToNonOptimized(int seed)
             {
-                float[] sourceArray = Create8x8RoundedRandomFloatData(MinAllowedValue, MaxAllowedValue, seed);
+                float[] sourceArray = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
 
                 var srcBlock = Block8x8F.Load(sourceArray);
 
@@ -74,7 +80,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             [InlineData(3)]
             public void LLM_TransformIDCT_CompareToAccurate(int seed)
             {
-                float[] sourceArray = Create8x8RoundedRandomFloatData(MinAllowedValue, MaxAllowedValue, seed);
+                float[] sourceArray = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
 
                 var srcBlock = Block8x8F.Load(sourceArray);
 
@@ -113,7 +119,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 {
                     int seed = FeatureTestRunner.Deserialize<int>(serialized);
 
-                    Span<float> src = Create8x8RoundedRandomFloatData(MinAllowedValue, MaxAllowedValue, seed);
+                    Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
                     var srcBlock = default(Block8x8F);
                     srcBlock.LoadFrom(src);
 
@@ -156,6 +162,147 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                     HwIntrinsics.AllowAll | HwIntrinsics.DisableFMA | HwIntrinsics.DisableAVX | HwIntrinsics.DisableSIMD);
             }
 
+            //[Theory]
+            //[InlineData(1)]
+            //[InlineData(2)]
+            //public void TranformIDCT_4x4(int seed)
+            //{
+            //    Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 4, 4);
+            //    var srcBlock = default(Block8x8F);
+            //    srcBlock.LoadFrom(src);
+
+            //    float[] expectedDest = new float[64];
+            //    float[] temp = new float[64];
+
+            //    // reference
+            //    ReferenceImplementations.LLM_FloatingPoint_DCT.IDCT2D_llm(src, expectedDest, temp);
+
+            //    // testee
+            //    // Part of the IDCT calculations is fused into the quantization step
+            //    // We must multiply input block with adjusted no-quantization matrix
+            //    // before applying IDCT
+            //    Block8x8F dequantMatrix = CreateBlockFromScalar(1);
+
+            //    // Dequantization using unit matrix - no values are upscaled
+            //    // as quant matrix is all 1's
+            //    // This step is needed to apply adjusting multipliers to the input block
+            //    FastFloatingPointDCT.AdjustToIDCT(ref dequantMatrix);
+
+            //    // testee
+            //    // IDCT implementation tranforms blocks after transposition
+            //    srcBlock.TransposeInplace();
+            //    DownScalingComponentProcessor2.TransformIDCT_4x4(ref srcBlock, ref dequantMatrix, NormalizationValue, MaxOutputValue);
+
+            //    var comparer = new ApproximateFloatComparer(1f);
+
+            //    Span<float> expectedSpan = expectedDest.AsSpan();
+            //    Span<float> actualSpan = srcBlock.ToArray().AsSpan();
+
+            //    AssertEquality(expectedSpan, actualSpan, comparer);
+            //    AssertEquality(expectedSpan.Slice(8), actualSpan.Slice(8), comparer);
+            //    AssertEquality(expectedSpan.Slice(16), actualSpan.Slice(16), comparer);
+            //    AssertEquality(expectedSpan.Slice(24), actualSpan.Slice(24), comparer);
+
+            //    static void AssertEquality(Span<float> expected, Span<float> actual, ApproximateFloatComparer comparer)
+            //    {
+            //        for (int x = 0; x < 4; x++)
+            //        {
+            //            float expectedValue = (float)Math.Round(Numerics.Clamp(expected[x] + NormalizationValue, 0, MaxOutputValue));
+            //            Assert.Equal(expectedValue, actual[x], comparer);
+            //        }
+            //    }
+            //}
+
+            [Theory]
+            [InlineData(1)]
+            [InlineData(2)]
+            public void TranformIDCT_2x2(int seed)
+            {
+                Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 2, 2);
+                var srcBlock = default(Block8x8F);
+                srcBlock.LoadFrom(src);
+
+                float[] expectedDest = new float[64];
+                float[] temp = new float[64];
+
+                // reference
+                ReferenceImplementations.LLM_FloatingPoint_DCT.IDCT2D_llm(src, expectedDest, temp);
+
+                // testee
+                // Part of the IDCT calculations is fused into the quantization step
+                // We must multiply input block with adjusted no-quantization matrix
+                // before applying IDCT
+                Block8x8F dequantMatrix = CreateBlockFromScalar(1);
+
+                // Dequantization using unit matrix - no values are upscaled
+                // as quant matrix is all 1's
+                // This step is needed to apply adjusting multipliers to the input block
+                FastFloatingPointDCT.AdjustToIDCT(ref dequantMatrix);
+
+                // testee
+                // IDCT implementation tranforms blocks after transposition
+                srcBlock.TransposeInplace();
+                DownScalingComponentProcessor4.TransformIDCT_2x2(ref srcBlock, ref dequantMatrix, NormalizationValue, MaxOutputValue);
+
+                var comparer = new ApproximateFloatComparer(0.1f);
+
+                // top-left
+                float topLeftExpected = (float)Math.Round(Numerics.Clamp(expectedDest[0] + NormalizationValue, 0, MaxOutputValue));
+                Assert.Equal(topLeftExpected, srcBlock[0], comparer);
+
+                // top-right
+                float topRightExpected = (float)Math.Round(Numerics.Clamp(expectedDest[7] + NormalizationValue, 0, MaxOutputValue));
+                Assert.Equal(topRightExpected, srcBlock[1], comparer);
+
+                // bot-left
+                float botLeftExpected = (float)Math.Round(Numerics.Clamp(expectedDest[56] + NormalizationValue, 0, MaxOutputValue));
+                Assert.Equal(botLeftExpected, srcBlock[8], comparer);
+
+                // bot-right
+                float botRightExpected = (float)Math.Round(Numerics.Clamp(expectedDest[63] + NormalizationValue, 0, MaxOutputValue));
+                Assert.Equal(botRightExpected, srcBlock[9], comparer);
+            }
+
+            [Theory]
+            [InlineData(1)]
+            [InlineData(2)]
+            public void TranformIDCT_1x1(int seed)
+            {
+                Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 1, 1);
+                var srcBlock = default(Block8x8F);
+                srcBlock.LoadFrom(src);
+
+                float[] expectedDest = new float[64];
+                float[] temp = new float[64];
+
+                // reference
+                ReferenceImplementations.LLM_FloatingPoint_DCT.IDCT2D_llm(src, expectedDest, temp);
+
+                // testee
+                // Part of the IDCT calculations is fused into the quantization step
+                // We must multiply input block with adjusted no-quantization matrix
+                // before applying IDCT
+                Block8x8F dequantMatrix = CreateBlockFromScalar(1);
+
+                // Dequantization using unit matrix - no values are upscaled
+                // as quant matrix is all 1's
+                // This step is needed to apply adjusting multipliers to the input block
+                FastFloatingPointDCT.AdjustToIDCT(ref dequantMatrix);
+
+                // testee
+                // IDCT implementation tranforms blocks after transposition
+                srcBlock.TransposeInplace();
+                float actual = DownScalingComponentProcessor8.TransformIDCT_1x1(
+                    srcBlock[0],
+                    dequantMatrix[0],
+                    NormalizationValue,
+                    MaxOutputValue);
+
+                float expected = (float)Math.Round(Numerics.Clamp(expectedDest[0] + NormalizationValue, 0, MaxOutputValue));
+
+                Assert.Equal(actual, expected, new ApproximateFloatComparer(0.1f));
+            }
+
             // Forward transform
             // This test covers entire FDCT conversion chain
             // This test checks all hardware implementations
@@ -168,7 +315,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 {
                     int seed = FeatureTestRunner.Deserialize<int>(serialized);
 
-                    Span<float> src = Create8x8RoundedRandomFloatData(MinAllowedValue, MaxAllowedValue, seed);
+                    Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
                     var block = default(Block8x8F);
                     block.LoadFrom(src);
 

tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.FastFloatingPointDCT.cs

@@ -53,7 +53,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             [InlineData(2, 200)]
             public void LLM_IDCT_IsEquivalentTo_AccurateImplementation(int seed, int range)
             {
-                float[] sourceArray = Create8x8RoundedRandomFloatData(-range, range, seed);
+                float[] sourceArray = Create8x8RandomFloatData(-range, range, seed);
 
                 var source = Block8x8F.Load(sourceArray);
 
@@ -64,23 +64,6 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 this.CompareBlocks(expected, actual, 0.1f);
             }
 
-            [Theory]
-            [InlineData(42, 1000)]
-            [InlineData(1, 1000)]
-            [InlineData(2, 1000)]
-            public void LLM_IDCT_CompareToIntegerRoundedAccurateImplementation(int seed, int range)
-            {
-                Block8x8F fSource = CreateRoundedRandomFloatBlock(-range, range, seed);
-                Block8x8 iSource = fSource.RoundAsInt16Block();
-
-                Block8x8 iExpected = ReferenceImplementations.AccurateDCT.TransformIDCT(ref iSource);
-                Block8x8F fExpected = iExpected.AsFloatBlock();
-
-                Block8x8F fActual = ReferenceImplementations.LLM_FloatingPoint_DCT.TransformIDCT(ref fSource);
-
-                this.CompareBlocks(fExpected, fActual, 2);
-            }
-
             [Theory]
             [InlineData(42)]
             [InlineData(1)]

tests/ImageSharp.Tests/Formats/Jpg/Utils/JpegFixture.cs

@@ -89,33 +89,28 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg.Utils
             return result;
         }
 
-        internal static float[] Create8x8RoundedRandomFloatData(int minValue, int maxValue, int seed = 42)
+        public static float[] Create8x8RandomFloatData(float minValue, float maxValue, int seed = 42, int xBorder = 8, int yBorder = 8)
-            => Create8x8RandomIntData(minValue, maxValue, seed).ConvertAllToFloat();
-
-        public static float[] Create8x8RandomFloatData(float minValue, float maxValue, int seed = 42)
         {
             var rnd = new Random(seed);
-            var result = new float[64];
+            float[] result = new float[64];
-            for (int i = 0; i < 8; i++)
+            for (int y = 0; y < yBorder; y++)
             {
-                for (int j = 0; j < 8; j++)
+                int y8 = y * 8;
+                for (int x = 0; x < xBorder; x++)
                 {
                     double val = rnd.NextDouble();
                     val *= maxValue - minValue;
                     val += minValue;
 
-                    result[(i * 8) + j] = (float)val;
+                    result[y8 + x] = (float)val;
                 }
             }
 
             return result;
         }
 
-        internal static Block8x8F CreateRandomFloatBlock(float minValue, float maxValue, int seed = 42) =>
+        internal static Block8x8F CreateRandomFloatBlock(float minValue, float maxValue, int seed = 42, int xBorder = 8, int yBorder = 8) =>
-            Block8x8F.Load(Create8x8RandomFloatData(minValue, maxValue, seed));
+            Block8x8F.Load(Create8x8RandomFloatData(minValue, maxValue, seed, xBorder, yBorder));
-
-        internal static Block8x8F CreateRoundedRandomFloatBlock(int minValue, int maxValue, int seed = 42) =>
-            Block8x8F.Load(Create8x8RoundedRandomFloatData(minValue, maxValue, seed));
 
         internal void Print8x8Data<T>(T[] data) => this.Print8x8Data(new Span<T>(data));