Merge branch 'main' into sn/nullable/various

src/ImageSharp/Formats/Webp/Lossy/LossyUtils.cs

@@ -14,7 +14,7 @@ internal static class LossyUtils
 {
     // Note: method name in libwebp reference implementation is called VP8SSE16x16.
     [MethodImpl(InliningOptions.ShortMethod)]
-    public static int Vp8_Sse16X16(Span<byte> a, Span<byte> b)
+    public static int Vp8_Sse16x16(Span<byte> a, Span<byte> b)
     {
         if (Avx2.IsSupported)
         {
@@ -31,7 +31,7 @@ internal static class LossyUtils
 
     // Note: method name in libwebp reference implementation is called VP8SSE16x8.
     [MethodImpl(InliningOptions.ShortMethod)]
-    public static int Vp8_Sse16X8(Span<byte> a, Span<byte> b)
+    public static int Vp8_Sse16x8(Span<byte> a, Span<byte> b)
     {
         if (Avx2.IsSupported)
         {
@@ -48,7 +48,7 @@ internal static class LossyUtils
 
     // Note: method name in libwebp reference implementation is called VP8SSE4x4.
     [MethodImpl(InliningOptions.ShortMethod)]
-    public static int Vp8_Sse4X4(Span<byte> a, Span<byte> b)
+    public static int Vp8_Sse4x4(Span<byte> a, Span<byte> b)
     {
         if (Avx2.IsSupported)
         {
@@ -77,8 +77,8 @@ internal static class LossyUtils
             Vector256<byte> b01s = Avx2.UnpackLow(b01.AsByte(), Vector256<byte>.Zero);
 
             // subtract, square and accumulate.
-            Vector256<byte> d0 = Avx2.SubtractSaturate(a01s, b01s);
+            Vector256<short> d0 = Avx2.SubtractSaturate(a01s.AsInt16(), b01s.AsInt16());
-            Vector256<int> e0 = Avx2.MultiplyAddAdjacent(d0.AsInt16(), d0.AsInt16());
+            Vector256<int> e0 = Avx2.MultiplyAddAdjacent(d0, d0);
 
             return Numerics.ReduceSum(e0);
         }
@@ -110,10 +110,10 @@ internal static class LossyUtils
             Vector128<byte> b23s = Sse2.UnpackLow(b23.AsByte(), Vector128<byte>.Zero);
 
             // subtract, square and accumulate.
-            Vector128<byte> d0 = Sse2.SubtractSaturate(a01s, b01s);
+            Vector128<short> d0 = Sse2.SubtractSaturate(a01s.AsInt16(), b01s.AsInt16());
-            Vector128<byte> d1 = Sse2.SubtractSaturate(a23s, b23s);
+            Vector128<short> d1 = Sse2.SubtractSaturate(a23s.AsInt16(), b23s.AsInt16());
-            Vector128<int> e0 = Sse2.MultiplyAddAdjacent(d0.AsInt16(), d0.AsInt16());
+            Vector128<int> e0 = Sse2.MultiplyAddAdjacent(d0, d0);
-            Vector128<int> e1 = Sse2.MultiplyAddAdjacent(d1.AsInt16(), d1.AsInt16());
+            Vector128<int> e1 = Sse2.MultiplyAddAdjacent(d1, d1);
             Vector128<int> sum = Sse2.Add(e0, e1);
 
             return Numerics.ReduceSum(sum);
@@ -126,18 +126,16 @@ internal static class LossyUtils
     public static int Vp8_SseNxN(Span<byte> a, Span<byte> b, int w, int h)
     {
         int count = 0;
-        int aOffset = 0;
+        int offset = 0;
-        int bOffset = 0;
         for (int y = 0; y < h; y++)
         {
             for (int x = 0; x < w; x++)
             {
-                int diff = a[aOffset + x] - b[bOffset + x];
+                int diff = a[offset + x] - b[offset + x];
                 count += diff * diff;
             }
 
-            aOffset += WebpConstants.Bps;
+            offset += WebpConstants.Bps;
-            bOffset += WebpConstants.Bps;
         }
 
         return count;

src/ImageSharp/Formats/Webp/Lossy/QuantEnc.cs

@@ -53,7 +53,7 @@ internal static unsafe class QuantEnc
             rdCur.Nz = (uint)ReconstructIntra16(it, dqm, rdCur, tmpDst, mode);
 
             // Measure RD-score.
-            rdCur.D = LossyUtils.Vp8_Sse16X16(src, tmpDst);
+            rdCur.D = LossyUtils.Vp8_Sse16x16(src, tmpDst);
             rdCur.SD = tlambda != 0 ? Mult8B(tlambda, LossyUtils.Vp8Disto16X16(src, tmpDst, WeightY, scratch)) : 0;
             rdCur.H = WebpConstants.Vp8FixedCostsI16[mode];
             rdCur.R = it.GetCostLuma16(rdCur, proba, res);
@@ -145,7 +145,7 @@ internal static unsafe class QuantEnc
                 rdTmp.Nz = (uint)ReconstructIntra4(it, dqm, tmpLevels, src, tmpDst, mode);
 
                 // Compute RD-score.
-                rdTmp.D = LossyUtils.Vp8_Sse4X4(src, tmpDst);
+                rdTmp.D = LossyUtils.Vp8_Sse4x4(src, tmpDst);
                 rdTmp.SD = tlambda != 0 ? Mult8B(tlambda, LossyUtils.Vp8Disto4X4(src, tmpDst, WeightY, scratch)) : 0;
                 rdTmp.H = modeCosts[mode];
 
@@ -235,7 +235,7 @@ internal static unsafe class QuantEnc
             rdUv.Nz = (uint)ReconstructUv(it, dqm, rdUv, tmpDst, mode);
 
             // Compute RD-score
-            rdUv.D = LossyUtils.Vp8_Sse16X8(src, tmpDst);
+            rdUv.D = LossyUtils.Vp8_Sse16x8(src, tmpDst);
             rdUv.SD = 0;    // not calling TDisto here: it tends to flatten areas.
             rdUv.H = WebpConstants.Vp8FixedCostsUv[mode];
             rdUv.R = it.GetCostUv(rdUv, proba, res);
@@ -389,7 +389,7 @@ internal static unsafe class QuantEnc
             for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
             {
                 Span<byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I16ModeOffsets[mode]);
-                long score = (LossyUtils.Vp8_Sse16X16(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsI16[mode] * lambdaDi16);
+                long score = (LossyUtils.Vp8_Sse16x16(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsI16[mode] * lambdaDi16);
 
                 if (mode > 0 && WebpConstants.Vp8FixedCostsI16[mode] > bitLimit)
                 {
@@ -436,7 +436,7 @@ internal static unsafe class QuantEnc
                 for (mode = 0; mode < WebpConstants.NumBModes; ++mode)
                 {
                     Span<byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8I4ModeOffsets[mode]);
-                    long score = (LossyUtils.Vp8_Sse4X4(src, reference) * WebpConstants.RdDistoMult) + (modeCosts[mode] * lambdaDi4);
+                    long score = (LossyUtils.Vp8_Sse4x4(src, reference) * WebpConstants.RdDistoMult) + (modeCosts[mode] * lambdaDi4);
                     if (score < bestI4Score)
                     {
                         bestI4Mode = mode;
@@ -485,7 +485,7 @@ internal static unsafe class QuantEnc
             for (mode = 0; mode < WebpConstants.NumPredModes; ++mode)
             {
                 Span<byte> reference = it.YuvP.AsSpan(Vp8Encoding.Vp8UvModeOffsets[mode]);
-                long score = (LossyUtils.Vp8_Sse16X8(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsUv[mode] * lambdaDuv);
+                long score = (LossyUtils.Vp8_Sse16x8(src, reference) * WebpConstants.RdDistoMult) + (WebpConstants.Vp8FixedCostsUv[mode] * lambdaDuv);
                 if (score < bestUvScore)
                 {
                     bestMode = mode;

tests/ImageSharp.Tests/Formats/WebP/LossyUtilsTests.cs

@@ -77,151 +77,73 @@ public class LossyUtilsTests
     private static void RunVp8Sse16X16Test()
     {
         // arrange
-        byte[] a =
+        Random rand = new(1234);
+        byte[] a = new byte[512 * 10];
+        byte[] b = new byte[512 * 10];
+        for (int i = 0; i < a.Length; i++)
         {
-            154, 154, 151, 151, 149, 148, 151, 157, 163, 163, 154, 132, 102, 98, 104, 108, 107, 104, 104, 103,
+            a[i] = (byte)rand.Next(byte.MaxValue);
-            101, 106, 123, 119, 170, 171, 172, 171, 168, 175, 171, 173, 151, 151, 149, 150, 147, 147, 146, 159,
+            b[i] = (byte)rand.Next(byte.MaxValue);
-            164, 165, 154, 129, 92, 90, 101, 105, 104, 103, 104, 101, 100, 105, 123, 117, 172, 172, 172, 168,
+        }
-            170, 177, 170, 175, 151, 149, 150, 150, 147, 147, 156, 161, 161, 161, 151, 126, 93, 90, 102, 107,
+        int[] expected = { 2533110, 2818581, 2984663, 2891188, 2855134, 2634604, 2466504, 3061747, 2626010, 2640965 };
-            104, 103, 104, 101, 104, 104, 122, 117, 172, 172, 170, 168, 170, 177, 172, 175, 150, 149, 152, 151,
+
-            148, 151, 160, 159, 157, 157, 148, 133, 96, 90, 103, 107, 104, 104, 101, 100, 102, 102, 121, 117,
+        // act + assert
-            170, 170, 169, 171, 171, 179, 173, 175, 149, 151, 152, 151, 148, 154, 162, 157, 154, 154, 151, 132,
+        int offset = 0;
-            92, 89, 101, 108, 104, 102, 101, 101, 103, 103, 123, 118, 171, 168, 177, 173, 171, 178, 172, 176,
+        for (int i = 0; i < expected.Length; i++)
-            152, 152, 152, 151, 154, 162, 161, 155, 149, 157, 156, 129, 92, 87, 101, 107, 102, 100, 107, 100,
-            101, 102, 123, 118, 170, 175, 182, 172, 171, 179, 173, 175, 152, 151, 154, 155, 160, 162, 161, 153,
-            150, 156, 153, 129, 92, 91, 102, 106, 100, 109, 115, 99, 101, 102, 124, 120, 171, 179, 178, 172,
-            171, 181, 171, 173, 154, 154, 154, 162, 160, 158, 156, 152, 153, 157, 151, 128, 86, 86, 102, 105,
-            102, 122, 114, 99, 101, 102, 125, 120, 178, 173, 177, 172, 171, 180, 172, 173, 154, 152, 158, 163,
-            150, 148, 148, 156, 151, 158, 152, 129, 87, 87, 101, 105, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 154, 151, 165, 156, 141, 137, 146, 158, 152, 159, 152, 133,
-            90, 88, 99, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            154, 160, 164, 150, 126, 127, 149, 159, 155, 161, 153, 131, 84, 86, 97, 103, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 157, 167, 157, 137, 102, 128, 155, 161,
-            157, 159, 154, 134, 84, 82, 97, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 163, 163, 150, 113, 78, 132, 156, 162, 159, 160, 154, 132, 83, 78, 91, 97, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 163, 157, 137, 80, 78,
-            131, 154, 163, 157, 159, 149, 131, 82, 77, 94, 100, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 159, 151, 108, 72, 88, 132, 156, 162, 159, 157, 151, 130, 79, 78,
-            95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 151, 130,
-            82, 82, 89, 134, 154, 161, 161, 157, 152, 129, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204
-        };
-
-        byte[] b =
         {
-            150, 150, 150, 150, 146, 149, 152, 154, 164, 166, 154, 132, 99, 92, 106, 112, 204, 204, 204, 204,
+            int actual = LossyUtils.Vp8_Sse16x16(a.AsSpan(offset), b.AsSpan(offset));
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 150, 150, 150, 150, 146, 149, 152, 154,
+            Assert.Equal(expected[i], actual);
-            161, 164, 151, 130, 93, 86, 100, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 150, 150, 150, 150, 146, 149, 152, 154, 158, 161, 148, 127, 93, 86, 100, 106,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 150, 150, 150, 150,
-            146, 149, 152, 154, 156, 159, 146, 125, 99, 92, 106, 112, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 148, 148, 148, 148, 149, 158, 162, 159, 155, 155, 153, 129,
-            94, 87, 101, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            151, 151, 151, 151, 152, 159, 161, 156, 155, 155, 153, 129, 94, 87, 101, 106, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 154, 154, 154, 154, 156, 161, 159, 152,
-            155, 155, 153, 129, 94, 87, 101, 106, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 156, 156, 156, 156, 159, 162, 158, 149, 155, 155, 153, 129, 94, 87, 101, 106,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 152, 153, 157, 162,
-            150, 149, 149, 151, 155, 160, 150, 131, 91, 90, 104, 104, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 152, 156, 158, 157, 140, 137, 145, 159, 155, 160, 150, 131,
-            89, 88, 102, 101, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            153, 161, 160, 149, 118, 128, 147, 162, 155, 160, 150, 131, 86, 85, 99, 98, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 154, 165, 161, 144, 96, 128, 154, 159, 155,
-            160, 150, 131, 83, 82, 97, 96, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 161, 160, 149, 105, 78, 127, 156, 170, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 160, 160, 133, 85, 81, 129, 155,
-            167, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 156, 147, 109, 76, 85, 130, 153, 163, 156, 156, 154, 130, 81, 77, 95, 102,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 152, 128, 87, 83,
-            88, 132, 152, 159, 156, 156, 154, 130, 81, 77, 95, 102, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204
-        };
 
-        int expected = 2063;
+            offset += 512;
-
+        }
-        // act
-        int actual = LossyUtils.Vp8_Sse16X16(a, b);
-
-        // assert
-        Assert.Equal(expected, actual);
     }
 
     private static void RunVp8Sse16X8Test()
     {
         // arrange
-        byte[] a =
+        Random rand = new(1234);
+        byte[] a = new byte[256 * 10];
+        byte[] b = new byte[256 * 10];
+        for (int i = 0; i < a.Length; i++)
         {
-            107, 104, 104, 103, 101, 106, 123, 119, 170, 171, 172, 171, 168, 175, 171, 173, 151, 151, 149, 150,
+            a[i] = (byte)rand.Next(byte.MaxValue);
-            147, 147, 146, 159, 164, 165, 154, 129, 92, 90, 101, 105, 104, 103, 104, 101, 100, 105, 123, 117,
+            b[i] = (byte)rand.Next(byte.MaxValue);
-            172, 172, 172, 168, 170, 177, 170, 175, 151, 149, 150, 150, 147, 147, 156, 161, 161, 161, 151, 126,
+        }
-            93, 90, 102, 107, 104, 103, 104, 101, 104, 104, 122, 117, 172, 172, 170, 168, 170, 177, 172, 175,
+        int[] expected = { 1298274, 1234836, 1325264, 1493317, 1551995, 1432668, 1407891, 1483297, 1537930, 1317204 };
-            150, 149, 152, 151, 148, 151, 160, 159, 157, 157, 148, 133, 96, 90, 103, 107, 104, 104, 101, 100,
+
-            102, 102, 121, 117, 170, 170, 169, 171, 171, 179, 173, 175, 149, 151, 152, 151, 148, 154, 162, 157,
+        // act + assert
-            154, 154, 151, 132, 92, 89, 101, 108, 104, 102, 101, 101, 103, 103, 123, 118, 171, 168, 177, 173,
+        int offset = 0;
-            171, 178, 172, 176, 152, 152, 152, 151, 154, 162, 161, 155, 149, 157, 156, 129, 92, 87, 101, 107,
+        for (int i = 0; i < expected.Length; i++)
-            102, 100, 107, 100, 101, 102, 123, 118, 170, 175, 182, 172, 171, 179, 173, 175, 152, 151, 154, 155,
-            160, 162, 161, 153, 150, 156, 153, 129, 92, 91, 102, 106, 100, 109, 115, 99, 101, 102, 124, 120,
-            171, 179, 178, 172, 171, 181, 171, 173, 154, 154, 154, 162, 160, 158, 156, 152, 153, 157, 151, 128,
-            86, 86, 102, 105, 102, 122, 114, 99, 101, 102, 125, 120, 178, 173, 177, 172, 171, 180, 172, 173,
-            154, 152, 158, 163, 150, 148, 148, 156, 151, 158, 152, 129, 87, 87, 101, 105
-        };
-
-        byte[] b =
         {
-            103, 103, 103, 103, 101, 106, 122, 114, 171, 171, 171, 171, 171, 177, 169, 175, 150, 150, 150, 150,
+            int actual = LossyUtils.Vp8_Sse16x8(a.AsSpan(offset), b.AsSpan(offset));
-            146, 149, 152, 154, 161, 164, 151, 130, 93, 86, 100, 106, 103, 103, 103, 103, 101, 106, 122, 114,
+            Assert.Equal(expected[i], actual);
-            171, 171, 171, 171, 171, 177, 169, 175, 150, 150, 150, 150, 146, 149, 152, 154, 158, 161, 148, 127,
-            93, 86, 100, 106, 103, 103, 103, 103, 101, 106, 122, 114, 171, 171, 171, 171, 171, 177, 169, 175,
-            150, 150, 150, 150, 146, 149, 152, 154, 156, 159, 146, 125, 99, 92, 106, 112, 103, 103, 103, 103,
-            101, 106, 122, 114, 171, 171, 171, 171, 171, 177, 169, 175, 148, 148, 148, 148, 149, 158, 162, 159,
-            155, 155, 153, 129, 94, 87, 101, 106, 102, 100, 100, 102, 100, 101, 120, 122, 170, 176, 176, 170,
-            174, 180, 171, 177, 151, 151, 151, 151, 152, 159, 161, 156, 155, 155, 153, 129, 94, 87, 101, 106,
-            102, 105, 105, 102, 100, 101, 120, 122, 170, 176, 176, 170, 174, 180, 171, 177, 154, 154, 154, 154,
-            156, 161, 159, 152, 155, 155, 153, 129, 94, 87, 101, 106, 102, 112, 112, 102, 100, 101, 120, 122,
-            170, 176, 176, 170, 174, 180, 171, 177, 156, 156, 156, 156, 159, 162, 158, 149, 155, 155, 153, 129,
-            94, 87, 101, 106, 102, 117, 117, 102, 100, 101, 120, 122, 170, 176, 176, 170, 174, 180, 171, 177,
-            152, 153, 157, 162, 150, 149, 149, 151, 155, 160, 150, 131, 91, 90, 104, 104
-        };
 
-        int expected = 749;
+            offset += 256;
-
+        }
-        // act
-        int actual = LossyUtils.Vp8_Sse16X8(a, b);
-
-        // assert
-        Assert.Equal(expected, actual);
     }
 
     private static void RunVp8Sse4X4Test()
     {
         // arrange
-        byte[] a =
+        Random rand = new(1234);
+        byte[] a = new byte[128 * 10];
+        byte[] b = new byte[128 * 10];
+        for (int i = 0; i < a.Length; i++)
         {
-            27, 27, 28, 29, 29, 28, 27, 27, 27, 28, 28, 29, 29, 28, 28, 27, 129, 129, 129, 129, 129, 129, 129,
+            a[i] = (byte)rand.Next(byte.MaxValue);
-            129, 128, 128, 128, 128, 128, 128, 128, 128, 27, 27, 27, 27, 27, 27, 27, 27, 27, 28, 28, 29, 29, 28,
+            b[i] = (byte)rand.Next(byte.MaxValue);
-            28, 27, 129, 129, 129, 129, 129, 129, 129, 129, 128, 128, 128, 128, 128, 128, 128, 128, 27, 27, 26,
+        }
-            26, 26, 26, 27, 27, 27, 28, 28, 29, 29, 28, 28, 27, 129, 129, 129, 129, 129, 129, 129, 129, 128,
+        int[] expected = { 194133, 125861, 165966, 195688, 106491, 173015, 266960, 200272, 311224, 122545 };
-            128, 128, 128, 128, 128, 128, 128, 28, 27, 27, 26, 26, 27, 27, 28, 27, 28, 28, 29, 29, 28, 28, 27,
+
-            129, 129, 129, 129, 129, 129, 129, 129, 128, 128, 128, 128, 128, 128, 128, 128
+        // act + assert
-        };
+        int offset = 0;
-
+        for (int i = 0; i < expected.Length; i++)
-        byte[] b =
         {
-            26, 26, 26, 26, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 204, 204, 204, 204, 204, 204, 204,
+            int actual = LossyUtils.Vp8_Sse4x4(a.AsSpan(offset), b.AsSpan(offset));
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 26, 26, 26, 26, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+            Assert.Equal(expected[i], actual);
-            28, 28, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 26, 26, 26,
-            26, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 204, 204, 204, 204, 204, 204, 204, 204, 204,
-            204, 204, 204, 204, 204, 204, 204, 26, 26, 26, 26, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
-            204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204, 204
-        };
-
-        int expected = 27;
 
-        // act
+            offset += 128;
-        int actual = LossyUtils.Vp8_Sse4X4(a, b);
+        }
-
-        // assert
-        Assert.Equal(expected, actual);
     }
 
     private static void RunMean16x4Test()
@@ -311,32 +233,41 @@ public class LossyUtilsTests
     [Fact]
     public void TransformOne_WithoutHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunTransformOneTest, HwIntrinsics.DisableHWIntrinsic);
 
+    // This will test the AVX2 version.
     [Fact]
     public void Vp8Sse16X16_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X16Test, HwIntrinsics.AllowAll);
 
+    // This will test the SSE2 version.
     [Fact]
-    public void Vp8Sse16X16_WithoutSSE2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X16Test, HwIntrinsics.DisableSSE2);
+    public void Vp8Sse16X16_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X16Test, HwIntrinsics.DisableAVX2);
 
+    // This will test the fallback scalar version.
     [Fact]
-    public void Vp8Sse16X16_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X16Test, HwIntrinsics.DisableAVX2);
+    public void Vp8Sse16X16_WithoutSSE2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X16Test, HwIntrinsics.DisableSSE2 | HwIntrinsics.DisableAVX);
 
+    // This will test the AVX2 version.
     [Fact]
     public void Vp8Sse16X8_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X8Test, HwIntrinsics.AllowAll);
 
+    // This will test the SSE2 version.
     [Fact]
-    public void Vp8Sse16X8_WithoutSSE2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X8Test, HwIntrinsics.DisableSSE2);
+    public void Vp8Sse16X8_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X8Test, HwIntrinsics.DisableAVX2);
 
+    // This will test the fallback scalar version.
     [Fact]
-    public void Vp8Sse16X8_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X8Test, HwIntrinsics.DisableAVX2);
+    public void Vp8Sse16X8_WithoutHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse16X8Test, HwIntrinsics.DisableSSE2 | HwIntrinsics.DisableAVX);
 
+    // This will test the AVX2 version.
     [Fact]
     public void Vp8Sse4X4_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse4X4Test, HwIntrinsics.AllowAll);
 
+    // This will test the SSE2 version.
     [Fact]
-    public void Vp8Sse4X4_WithoutSSE2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse4X4Test, HwIntrinsics.DisableSSE2);
+    public void Vp8Sse4X4_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse4X4Test, HwIntrinsics.DisableAVX2);
 
+    // This will test the fallback scalar version.
     [Fact]
-    public void Vp8Sse4X4_WithoutAVX2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse4X4Test, HwIntrinsics.DisableAVX2);
+    public void Vp8Sse4X4_WithoutHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVp8Sse4X4Test, HwIntrinsics.DisableSSE2 | HwIntrinsics.DisableAVX);
 
     [Fact]
     public void Mean16x4_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunMean16x4Test, HwIntrinsics.AllowAll);