Refactor forward transformers

src/ImageSharp/Formats/Heif/Av1/Transform/Av1ForwardTransformer.cs

@@ -12,22 +12,22 @@ internal class Av1ForwardTransformer
     private const int NewSqrt = 5793;
     private const int NewSqrtBitCount = 12;
 
-    private static readonly IAv1ForwardTransformer?[] Transformers =
+    private static readonly IAv1Forward1dTransformer?[] Transformers =
         [
-            new Av1Dct4ForwardTransformer(),
-            new Av1Dct8ForwardTransformer(),
-            new Av1Dct16ForwardTransformer(),
-            new Av1Dct32ForwardTransformer(),
-            new Av1Dct64ForwardTransformer(),
-            new Av1Adst4ForwardTransformer(),
-            new Av1Adst8ForwardTransformer(),
-            new Av1Adst16ForwardTransformer(),
-            new Av1Adst32ForwardTransformer(),
-            new Av1Identity4ForwardTransformer(),
-            new Av1Identity8ForwardTransformer(),
-            new Av1Identity16ForwardTransformer(),
-            new Av1Identity32ForwardTransformer(),
-            new Av1Identity64ForwardTransformer(),
+            new Av1Dct4Forward1dTransformer(),
+            new Av1Dct8Forward1dTransformer(),
+            new Av1Dct16Forward1dTransformer(),
+            new Av1Dct32Forward1dTransformer(),
+            new Av1Dct64Forward1dTransformer(),
+            new Av1Adst4Forward1dTransformer(),
+            new Av1Adst8Forward1dTransformer(),
+            new Av1Adst16Forward1dTransformer(),
+            new Av1Adst32Forward1dTransformer(),
+            new Av1Identity4Forward1dTransformer(),
+            new Av1Identity8Forward1dTransformer(),
+            new Av1Identity16Forward1dTransformer(),
+            new Av1Identity32Forward1dTransformer(),
+            new Av1Identity64Forward1dTransformer(),
             null
         ];
 
@@ -37,8 +37,8 @@ internal class Av1ForwardTransformer
     {
         Av1Transform2dFlipConfiguration config = new(transformType, transformSize);
         ref int buffer = ref TemporaryCoefficientsBuffer[0];
-        IAv1ForwardTransformer? columnTransformer = GetTransformer(config.TransformFunctionTypeColumn);
-        IAv1ForwardTransformer? rowTransformer = GetTransformer(config.TransformFunctionTypeRow);
+        IAv1Forward1dTransformer? columnTransformer = GetTransformer(config.TransformFunctionTypeColumn);
+        IAv1Forward1dTransformer? rowTransformer = GetTransformer(config.TransformFunctionTypeRow);
         if (columnTransformer != null && rowTransformer != null)
         {
             Transform2dCore(columnTransformer, rowTransformer, ref input[0], stride, ref coefficients[0], config, ref buffer, bitDepth);
@@ -97,15 +97,15 @@ internal class Av1ForwardTransformer
         }
     }
 
-    private static IAv1ForwardTransformer? GetTransformer(Av1TransformFunctionType transformerType)
+    private static IAv1Forward1dTransformer? GetTransformer(Av1TransformFunctionType transformerType)
         => Transformers[(int)transformerType];
 
     /// <summary>
     /// SVT: av1_tranform_two_d_core_c
     /// </summary>
     private static void Transform2dCore<TColumn, TRow>(TColumn transformFunctionColumn, TRow transformFunctionRow, ref short input, uint inputStride, ref int output, Av1Transform2dFlipConfiguration config, ref int buf, int bitDepth)
-            where TColumn : IAv1ForwardTransformer
-            where TRow : IAv1ForwardTransformer
+            where TColumn : IAv1Forward1dTransformer
+            where TRow : IAv1Forward1dTransformer
     {
         int c, r;
 

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity16ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst16Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Identity16ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Adst16Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst16ForwardTransformer.cs

@@ -1,15 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Adst16ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-        => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity32ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst32Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Identity32ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Adst32Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst32ForwardTransformer.cs

@@ -1,15 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Adst32ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-        => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity4ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst4Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Identity4ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Adst4Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity8ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst8Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Identity8ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Adst8Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct8ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct16Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Dct8ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Dct16Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct16ForwardTransformer.cs

@@ -1,15 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Dct16ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-        => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst8ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct32Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Adst8ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Dct32Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct4Forward1dTransformer.cs

@@ -0,0 +1,67 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Runtime.CompilerServices;
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Dct4Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => TransformScalar(ref input, ref output, cosBit);
+
+    private static void TransformScalar(ref int input, ref int output, int cosBit)
+    {
+        Span<int> cospi = Av1SinusConstants.CosinusPi(cosBit);
+        ref int bf0 = ref output;
+        ref int bf1 = ref output;
+        Span<int> stepSpan = new int[4];
+        ref int step0 = ref stepSpan[0];
+        ref int step1 = ref Unsafe.Add(ref step0, 1);
+        ref int step2 = ref Unsafe.Add(ref step0, 2);
+        ref int step3 = ref Unsafe.Add(ref step0, 3);
+        ref int output1 = ref Unsafe.Add(ref output, 1);
+        ref int output2 = ref Unsafe.Add(ref output, 2);
+        ref int output3 = ref Unsafe.Add(ref output, 3);
+
+        // stage 0;
+
+        // stage 1;
+        output = input + Unsafe.Add(ref input, 3);
+        output1 = Unsafe.Add(ref input, 1) + Unsafe.Add(ref input, 2);
+        output2 = -Unsafe.Add(ref input, 2) + Unsafe.Add(ref input, 1);
+        output3 = -Unsafe.Add(ref input, 3) + Unsafe.Add(ref input, 0);
+
+        // stage 2
+        step0 = HalfBtf(cospi[32], output, cospi[32], output1, cosBit);
+        step1 = HalfBtf(-cospi[32], output1, cospi[32], output, cosBit);
+        step2 = HalfBtf(cospi[48], output2, cospi[16], output3, cosBit);
+        step3 = HalfBtf(cospi[48], output3, -cospi[16], output2, cosBit);
+
+        // stage 3
+        output = step0;
+        output1 = step2;
+        output2 = step1;
+        output3 = step3;
+    }
+
+    private static int HalfBtf(int w0, int in0, int w1, int in1, int bit)
+    {
+        long result64 = (long)(w0 * in0) + (w1 * in1);
+        long intermediate = result64 + (1L << (bit - 1));
+
+        // NOTE(david.barker): The value 'result_64' may not necessarily fit
+        // into 32 bits. However, the result of this function is nominally
+        // ROUND_POWER_OF_TWO_64(result_64, bit)
+        // and that is required to fit into stage_range[stage] many bits
+        // (checked by range_check_buf()).
+        //
+        // Here we've unpacked that rounding operation, and it can be shown
+        // that the value of 'intermediate' here *does* fit into 32 bits
+        // for any conformant bitstream.
+        // The upshot is that, if you do all this calculation using
+        // wrapping 32-bit arithmetic instead of (non-wrapping) 64-bit arithmetic,
+        // then you'll still get the correct result.
+        return (int)(intermediate >> bit);
+    }
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct4ForwardTransformer.cs

@@ -1,138 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.Intrinsics;
-using System.Runtime.Intrinsics.X86;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Dct4ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-    {
-        Span<int> cospi = Av1SinusConstants.CosinusPi(cosBit);
-        ref int bf0 = ref output;
-        ref int bf1 = ref output;
-        Span<int> stepSpan = new int[4];
-        ref int step0 = ref stepSpan[0];
-        ref int step1 = ref Unsafe.Add(ref step0, 1);
-        ref int step2 = ref Unsafe.Add(ref step0, 2);
-        ref int step3 = ref Unsafe.Add(ref step0, 3);
-        ref int output1 = ref Unsafe.Add(ref output, 1);
-        ref int output2 = ref Unsafe.Add(ref output, 2);
-        ref int output3 = ref Unsafe.Add(ref output, 3);
-
-        // stage 0;
-
-        // stage 1;
-        output = input + Unsafe.Add(ref input, 3);
-        output1 = Unsafe.Add(ref input, 1) + Unsafe.Add(ref input, 2);
-        output2 = -Unsafe.Add(ref input, 2) + Unsafe.Add(ref input, 1);
-        output3 = -Unsafe.Add(ref input, 3) + Unsafe.Add(ref input, 0);
-
-        // stage 2
-        step0 = HalfBtf(cospi[32], output, cospi[32], output1, cosBit);
-        step1 = HalfBtf(-cospi[32], output1, cospi[32], output, cosBit);
-        step2 = HalfBtf(cospi[48], output2, cospi[16], output3, cosBit);
-        step3 = HalfBtf(cospi[48], output3, -cospi[16], output2, cosBit);
-
-        // stage 3
-        output = step0;
-        output1 = step2;
-        output2 = step1;
-        output3 = step3;
-    }
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException("Too small block for Vector implementation, use TransformSse() method instead.");
-
-    /// <summary>
-    /// SVT: fdct4x4_sse4_1
-    /// </summary>
-    public static void TransformSse(ref Vector128<int> input, ref Vector128<int> output, byte cosBit, int columnNumber)
-    {
-#pragma warning disable CA1857 // A constant is expected for the parameter
-
-        // We only use stage-2 bit;
-        // shift[0] is used in load_buffer_4x4()
-        // shift[1] is used in txfm_func_col()
-        // shift[2] is used in txfm_func_row()
-        Span<int> cospi = Av1SinusConstants.CosinusPi(cosBit);
-        Vector128<int> cospi32 = Vector128.Create<int>(cospi[32]);
-        Vector128<int> cospi48 = Vector128.Create<int>(cospi[48]);
-        Vector128<int> cospi16 = Vector128.Create<int>(cospi[16]);
-        Vector128<int> rnding = Vector128.Create<int>(1 << (cosBit - 1));
-        Vector128<int> s0, s1, s2, s3;
-        Vector128<int> u0, u1, u2, u3;
-        Vector128<int> v0, v1, v2, v3;
-
-        int endidx = 3 * columnNumber;
-        s0 = Sse2.Add(input, Unsafe.Add(ref input, endidx));
-        s3 = Sse2.Subtract(input, Unsafe.Add(ref input, endidx));
-        endidx -= columnNumber;
-        s1 = Sse2.Add(Unsafe.Add(ref input, columnNumber), Unsafe.Add(ref input, endidx));
-        s2 = Sse2.Subtract(Unsafe.Add(ref input, columnNumber), Unsafe.Add(ref input, endidx));
-
-        // btf_32_sse4_1_type0(cospi32, cospi32, s[01], u[02], bit);
-        u0 = Sse41.MultiplyLow(s0, cospi32);
-        u1 = Sse41.MultiplyLow(s1, cospi32);
-        u2 = Sse2.Add(u0, u1);
-        v0 = Sse2.Subtract(u0, u1);
-
-        u3 = Sse2.Add(u2, rnding);
-        v1 = Sse2.Add(v0, rnding);
-
-        u0 = Sse2.ShiftRightArithmetic(u3, cosBit);
-        u2 = Sse2.ShiftRightArithmetic(v1, cosBit);
-
-        // btf_32_sse4_1_type1(cospi48, cospi16, s[23], u[13], bit);
-        v0 = Sse41.MultiplyLow(s2, cospi48);
-        v1 = Sse41.MultiplyLow(s3, cospi16);
-        v2 = Sse2.Add(v0, v1);
-
-        v3 = Sse2.Add(v2, rnding);
-        u1 = Sse2.ShiftRightArithmetic(v3, cosBit);
-
-        v0 = Sse41.MultiplyLow(s2, cospi16);
-        v1 = Sse41.MultiplyLow(s3, cospi48);
-        v2 = Sse2.Subtract(v1, v0);
-
-        v3 = Sse2.Add(v2, rnding);
-        u3 = Sse2.ShiftRightArithmetic(v3, cosBit);
-
-        // Note: shift[1] and shift[2] are zeros
-
-        // Transpose 4x4 32-bit
-        v0 = Sse2.UnpackLow(u0, u1);
-        v1 = Sse2.UnpackHigh(u0, u1);
-        v2 = Sse2.UnpackLow(u2, u3);
-        v3 = Sse2.UnpackHigh(u2, u3);
-
-        output = Sse2.UnpackLow(v0.AsInt64(), v2.AsInt64()).AsInt32();
-        Unsafe.Add(ref output, 1) = Sse2.UnpackHigh(v0.AsInt64(), v2.AsInt64()).AsInt32();
-        Unsafe.Add(ref output, 2) = Sse2.UnpackLow(v1.AsInt64(), v3.AsInt64()).AsInt32();
-        Unsafe.Add(ref output, 3) = Sse2.UnpackHigh(v1.AsInt64(), v3.AsInt64()).AsInt32();
-#pragma warning restore CA1857 // A constant is expected for the parameter
-    }
-
-    private static int HalfBtf(int w0, int in0, int w1, int in1, int bit)
-    {
-        long result64 = (long)(w0 * in0) + (w1 * in1);
-        long intermediate = result64 + (1L << (bit - 1));
-
-        // NOTE(david.barker): The value 'result_64' may not necessarily fit
-        // into 32 bits. However, the result of this function is nominally
-        // ROUND_POWER_OF_TWO_64(result_64, bit)
-        // and that is required to fit into stage_range[stage] many bits
-        // (checked by range_check_buf()).
-        //
-        // Here we've unpacked that rounding operation, and it can be shown
-        // that the value of 'intermediate' here *does* fit into 32 bits
-        // for any conformant bitstream.
-        // The upshot is that, if you do all this calculation using
-        // wrapping 32-bit arithmetic instead of (non-wrapping) 64-bit arithmetic,
-        // then you'll still get the correct result.
-        return (int)(intermediate >> bit);
-    }
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct64Forward1dTransformer.cs

@@ -0,0 +1,10 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Dct64Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => throw new NotImplementedException();
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct64ForwardTransformer.cs

@@ -1,15 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Dct64ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-        => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct32ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Dct8Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Dct32ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Dct8Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1DctDct4Forward2dTransformer.cs

@@ -0,0 +1,100 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Runtime.CompilerServices;
+using System.Runtime.Intrinsics;
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1DctDct4Forward2dTransformer : Av1Forward2dTransformerBase
+{
+    private readonly Av1Transform2dFlipConfiguration config = new(Av1TransformType.DctDct, Av1TransformSize.Size4x4);
+    private readonly Av1Dct4Forward1dTransformer transformer = new();
+    private readonly int[] temp = new int[Av1Constants.MaxTransformSize * Av1Constants.MaxTransformSize];
+
+    public void Transform(ref short input, ref int output, int cosBit, int columnNumber)
+    {
+        /*if (Vector256.IsHardwareAccelerated)
+        {
+            Span<Vector128<int>> inputVectors = stackalloc Vector128<int>[16];
+            ref Vector128<int> outputAsVector = ref Unsafe.As<int, Vector128<int>>(ref output);
+            TransformVector(ref inputVectors[0], ref outputAsVector, cosBit, columnNumber);
+        }
+        else*/
+        {
+            Transform2dCore(this.transformer, this.transformer, ref input, 4, ref output, this.config, ref this.temp[0], 8);
+        }
+    }
+
+    /// <summary>
+    /// SVT: fdct4x4_sse4_1
+    /// </summary>
+    private static void TransformVector(ref Vector128<int> input, ref Vector128<int> output, int cosBit, int columnNumber)
+    {
+        // We only use stage-2 bit;
+        // shift[0] is used in load_buffer_4x4()
+        // shift[1] is used in txfm_func_col()
+        // shift[2] is used in txfm_func_row()
+        Span<int> cospi = Av1SinusConstants.CosinusPi(cosBit);
+        Vector128<int> cospi32 = Vector128.Create<int>(cospi[32]);
+        Vector128<int> cospi48 = Vector128.Create<int>(cospi[48]);
+        Vector128<int> cospi16 = Vector128.Create<int>(cospi[16]);
+        Vector128<int> rnding = Vector128.Create<int>(1 << (cosBit - 1));
+        Vector128<int> s0, s1, s2, s3;
+        Vector128<int> u0, u1, u2, u3;
+        Vector128<int> v0, v1, v2, v3;
+        Vector256<int> interleave32 = Vector256.Create(0, 4, 1, 5, 2, 6, 3, 7);
+        Vector256<int> reverse64 = Vector256.Create(1, 0, 3, 2, 5, 4, 7, 6);
+        Vector256<int> select64 = Vector256.Create(0, 0, -1, -1, 0, 0, -1, -1);
+
+        int endidx = 3 * columnNumber;
+        s0 = Vector128.Add(input, Unsafe.Add(ref input, endidx));
+        s3 = Vector128.Subtract(input, Unsafe.Add(ref input, endidx));
+        endidx -= columnNumber;
+        s1 = Vector128.Add(Unsafe.Add(ref input, columnNumber), Unsafe.Add(ref input, endidx));
+        s2 = Vector128.Subtract(Unsafe.Add(ref input, columnNumber), Unsafe.Add(ref input, endidx));
+
+        // btf_32_sse4_1_type0(cospi32, cospi32, s[01], u[02], bit);
+        u0 = Vector128.Multiply(s0, cospi32);
+        u1 = Vector128.Multiply(s1, cospi32);
+        u2 = Vector128.Add(u0, u1);
+        v0 = Vector128.Subtract(u0, u1);
+
+        u3 = Vector128.Add(u2, rnding);
+        v1 = Vector128.Add(v0, rnding);
+
+        u0 = Vector128.ShiftRightArithmetic(u3, cosBit);
+        u2 = Vector128.ShiftRightArithmetic(v1, cosBit);
+
+        // btf_32_sse4_1_type1(cospi48, cospi16, s[23], u[13], bit);
+        v0 = Vector128.Multiply(s2, cospi48);
+        v1 = Vector128.Multiply(s3, cospi16);
+        v2 = Vector128.Add(v0, v1);
+
+        v3 = Vector128.Add(v2, rnding);
+        u1 = Vector128.ShiftRightArithmetic(v3, cosBit);
+
+        v0 = Vector128.Multiply(s2, cospi16);
+        v1 = Vector128.Multiply(s3, cospi48);
+        v2 = Vector128.Subtract(v1, v0);
+
+        v3 = Vector128.Add(v2, rnding);
+        u3 = Vector128.ShiftRightArithmetic(v3, cosBit);
+
+        // Note: shift[1] and shift[2] are zeros
+
+        // Transpose 4x4 32-bit
+        Vector256<int> w0 = Vector256.Create(u0, u1);
+        Vector256<int> w1 = Vector256.Create(u2, u3);
+        w0 = Vector256.Shuffle(w0, interleave32);
+        w1 = Vector256.Shuffle(w1, interleave32);
+        Vector256<int> w2 = Vector256.ConditionalSelect(select64, w0, w1);
+        Vector256<int> w3 = Vector256.ConditionalSelect(select64, w1, w0);
+        w3 = Vector256.Shuffle(w3, reverse64);
+
+        output = Vector256.GetLower(w2);
+        Unsafe.Add(ref output, 1) = Vector256.GetLower(w3);
+        Unsafe.Add(ref output, 2) = Vector256.GetUpper(w2);
+        Unsafe.Add(ref output, 3) = Vector256.GetUpper(w3);
+    }
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Forward2dTransformerBase.cs

@@ -0,0 +1,186 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Runtime.CompilerServices;
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal abstract class Av1Forward2dTransformerBase
+{
+    private const int NewSqrt = 5793;
+    private const int NewSqrtBitCount = 12;
+
+    /// <summary>
+    /// SVT: av1_tranform_two_d_core_c
+    /// </summary>
+    protected static void Transform2dCore<TColumn, TRow>(TColumn transformFunctionColumn, TRow transformFunctionRow, ref short input, uint inputStride, ref int output, Av1Transform2dFlipConfiguration config, ref int buf, int bitDepth)
+            where TColumn : IAv1Forward1dTransformer
+            where TRow : IAv1Forward1dTransformer
+    {
+        int c, r;
+
+        // Note when assigning txfm_size_col, we use the txfm_size from the
+        // row configuration and vice versa. This is intentionally done to
+        // accurately perform rectangular transforms. When the transform is
+        // rectangular, the number of columns will be the same as the
+        // txfm_size stored in the row cfg struct. It will make no difference
+        // for square transforms.
+        int transformColumnCount = config.TransformSize.GetWidth();
+        int transformRowCount = config.TransformSize.GetHeight();
+        int transformCount = transformColumnCount * transformRowCount;
+
+        // Take the shift from the larger dimension in the rectangular case.
+        Span<int> shift = config.Shift;
+        int rectangleType = GetRectangularRatio(transformColumnCount, transformRowCount);
+        Span<byte> stageRangeColumn = stackalloc byte[Av1Transform2dFlipConfiguration.MaxStageNumber];
+        Span<byte> stageRangeRow = stackalloc byte[Av1Transform2dFlipConfiguration.MaxStageNumber];
+
+        // assert(cfg->stage_num_col <= MAX_TXFM_STAGE_NUM);
+        // assert(cfg->stage_num_row <= MAX_TXFM_STAGE_NUM);
+        config.GenerateStageRange(bitDepth);
+
+        int cosBitColumn = config.CosBitColumn;
+        int cosBitRow = config.CosBitRow;
+
+        // ASSERT(txfm_func_col != NULL);
+        // ASSERT(txfm_func_row != NULL);
+        // use output buffer as temp buffer
+        ref int tempIn = ref output;
+        ref int tempOut = ref Unsafe.Add(ref output, transformRowCount);
+
+        // Columns
+        for (c = 0; c < transformColumnCount; ++c)
+        {
+            if (!config.FlipUpsideDown)
+            {
+                uint t = (uint)c;
+                for (r = 0; r < transformRowCount; ++r)
+                {
+                    Unsafe.Add(ref tempIn, r) = Unsafe.Add(ref input, t);
+                    t += inputStride;
+                }
+            }
+            else
+            {
+                uint t = (uint)(c + ((transformRowCount - 1) * (int)inputStride));
+                for (r = 0; r < transformRowCount; ++r)
+                {
+                    // flip upside down
+                    Unsafe.Add(ref tempIn, r) = Unsafe.Add(ref input, t);
+                    t -= inputStride;
+                }
+            }
+
+            RoundShiftArray(ref tempIn, transformRowCount, -shift[0]); // NM svt_av1_round_shift_array_c
+            transformFunctionColumn.Transform(ref tempIn, ref tempOut, cosBitColumn, stageRangeColumn);
+            RoundShiftArray(ref tempOut, transformRowCount, -shift[1]); // NM svt_av1_round_shift_array_c
+            if (!config.FlipLeftToRight)
+            {
+                int t = c;
+                for (r = 0; r < transformRowCount; ++r)
+                {
+                    Unsafe.Add(ref buf, t) = Unsafe.Add(ref tempOut, r);
+                    t += transformColumnCount;
+                }
+            }
+            else
+            {
+                int t = transformColumnCount - c - 1;
+                for (r = 0; r < transformRowCount; ++r)
+                {
+                    // flip from left to right
+                    Unsafe.Add(ref buf, t) = Unsafe.Add(ref tempOut, r);
+                    t += transformColumnCount;
+                }
+            }
+        }
+
+        // Rows
+        for (r = 0; r < transformRowCount; ++r)
+        {
+            transformFunctionRow.Transform(ref Unsafe.Add(ref buf, r * transformColumnCount), ref Unsafe.Add(ref output, r * transformColumnCount), cosBitRow, stageRangeRow);
+            RoundShiftArray(ref Unsafe.Add(ref output, r * transformColumnCount), transformColumnCount, -shift[2]);
+
+            if (Math.Abs(rectangleType) == 1)
+            {
+                // Multiply everything by Sqrt2 if the transform is rectangular and the
+                // size difference is a factor of 2.
+                for (c = 0; c < transformColumnCount; ++c)
+                {
+                    ref int current = ref Unsafe.Add(ref output, (r * transformColumnCount) + c);
+                    current = Av1Math.RoundShift((long)current * NewSqrt, NewSqrtBitCount);
+                }
+            }
+        }
+    }
+
+    private static void RoundShiftArray(ref int arr, int size, int bit)
+    {
+        if (bit == 0)
+        {
+            return;
+        }
+        else
+        {
+            nuint sz = (nuint)size;
+            if (bit > 0)
+            {
+                for (nuint i = 0; i < sz; i++)
+                {
+                    ref int a = ref Unsafe.Add(ref arr, i);
+                    a = Av1Math.RoundShift(a, bit);
+                }
+            }
+            else
+            {
+                for (nuint i = 0; i < sz; i++)
+                {
+                    ref int a = ref Unsafe.Add(ref arr, i);
+                    a *= 1 << (-bit);
+                }
+            }
+        }
+    }
+
+    /// <summary>
+    /// SVT: get_rect_tx_log_ratio
+    /// </summary>
+    public static int GetRectangularRatio(int col, int row)
+    {
+        if (col == row)
+        {
+            return 0;
+        }
+
+        if (col > row)
+        {
+            if (col == row * 2)
+            {
+                return 1;
+            }
+
+            if (col == row * 4)
+            {
+                return 2;
+            }
+
+            Guard.IsTrue(false, nameof(row), "Unsupported transform size");
+        }
+        else
+        {
+            if (row == col * 2)
+            {
+                return -1;
+            }
+
+            if (row == col * 4)
+            {
+                return -2;
+            }
+
+            Guard.IsTrue(false, nameof(row), "Unsupported transform size");
+        }
+
+        return 0; // Invalid
+    }
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Adst4ForwardTransformer.cs → src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity16Forward1dTransformer.cs

@@ -1,15 +1,10 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
-using System.Runtime.Intrinsics;
-
 namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
 
-internal class Av1Adst4ForwardTransformer : IAv1ForwardTransformer
+internal class Av1Identity16Forward1dTransformer : IAv1Forward1dTransformer
 {
     public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
         => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
 }

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity32Forward1dTransformer.cs

@@ -0,0 +1,10 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Identity32Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => throw new NotImplementedException();
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity4Forward1dTransformer.cs

@@ -0,0 +1,10 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Identity4Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => throw new NotImplementedException();
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity64Forward1dTransformer.cs

@@ -0,0 +1,10 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Identity64Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => throw new NotImplementedException();
+}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity64ForwardTransformer.cs

@@ -1,15 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
-
-internal class Av1Identity64ForwardTransformer : IAv1ForwardTransformer
-{
-    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
-        => throw new NotImplementedException();
-
-    public void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber)
-        => throw new NotImplementedException();
-}

src/ImageSharp/Formats/Heif/Av1/Transform/Forward/Av1Identity8Forward1dTransformer.cs

@@ -0,0 +1,10 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform.Forward;
+
+internal class Av1Identity8Forward1dTransformer : IAv1Forward1dTransformer
+{
+    public void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange)
+        => throw new NotImplementedException();
+}

src/ImageSharp/Formats/Heif/Av1/Transform/IAv1Forward1dTransformer.cs

@@ -0,0 +1,19 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform;
+
+/// <summary>
+/// Implementation of a specific forward 1-dimensional transform function.
+/// </summary>
+internal interface IAv1Forward1dTransformer
+{
+    /// <summary>
+    /// Execute the 1 dimensional transformation.
+    /// </summary>
+    /// <param name="input">Input pixels.</param>
+    /// <param name="output">Output coefficients.</param>
+    /// <param name="cosBit">The cosinus bit.</param>
+    /// <param name="stageRange">Stage ranges.</param>
+    void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange);
+}

src/ImageSharp/Formats/Heif/Av1/Transform/IAv1ForwardTransformer.cs

@@ -1,31 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-using System.Runtime.Intrinsics.X86;
-
-namespace SixLabors.ImageSharp.Formats.Heif.Av1.Transform;
-
-/// <summary>
-/// Implementation of a specific forward transform function.
-/// </summary>
-internal interface IAv1ForwardTransformer
-{
-    /// <summary>
-    /// Execute the transformation.
-    /// </summary>
-    /// <param name="input">Input pixels.</param>
-    /// <param name="output">Output coefficients.</param>
-    /// <param name="cosBit">The cosinus bit.</param>
-    /// <param name="stageRange">Stage ranges.</param>
-    void Transform(ref int input, ref int output, int cosBit, Span<byte> stageRange);
-
-    /// <summary>
-    /// Execute the transformation using <see cref="Avx2"/> instructions.
-    /// </summary>
-    /// <param name="input">Array of input vectors.</param>
-    /// <param name="output">Array of output coefficients vectors.</param>
-    /// <param name="cosBit">The cosinus bit.</param>
-    /// <param name="columnNumber">The column number to process.</param>
-    void TransformAvx2(ref Vector256<int> input, ref Vector256<int> output, int cosBit, int columnNumber);
-}

tests/ImageSharp.Tests/Formats/Heif/Av1/Av1ForwardTransformTests.cs

@@ -46,7 +46,7 @@ public class Av1ForwardTransformTests
         Av1Transform2dFlipConfiguration config = new(Av1TransformType.DctDct, transformSize);
         int width = config.TransformSize.GetWidth();
 
-        int[] inputOfTest = new int[width];
+        short[] inputOfTest = new short[width];
         double[] inputReference = new double[width];
         int[] outputOfTest = new int[width];
         double[] outputReference = new double[width];
@@ -62,11 +62,11 @@ public class Av1ForwardTransformTests
             }
 
             // calculate in forward transform functions
-            new Av1Dct4ForwardTransformer().Transform(
+            new Av1DctDct4Forward2dTransformer().Transform(
                 ref inputOfTest[0],
                 ref outputOfTest[0],
                 config.CosBitColumn,
-                config.StageRangeColumn);
+                config.StageNumberColumn);
 
             // calculate in reference forward transform functions
             Av1ReferenceTransform.ReferenceDct1d(inputReference, outputReference, width);