From c174ab42bea366257a410e8335055b9d27d487ff Mon Sep 17 00:00:00 2001
From: Brian Popow <popow@gmx.de>
Date: Wed, 24 Nov 2021 20:13:46 +0100
Subject: [PATCH] Remove SSE2 version of VectorMismatch: Profiling does not
 show any speedup

---
 .../Formats/Webp/Lossless/LosslessUtils.cs    | 63 +------------------
 .../Formats/WebP/LosslessUtilsTests.cs        | 28 +--------
 2 files changed, 4 insertions(+), 87 deletions(-)

diff --git a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
index 319aa8c3d6..0ed180a184 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
@@ -83,69 +83,12 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless
             ref uint array1Ref = ref MemoryMarshal.GetReference(array1);
             ref uint array2Ref = ref MemoryMarshal.GetReference(array2);
 
-#if SUPPORTS_RUNTIME_INTRINSICS
-            if (Sse2.IsSupported)
+            while (matchLen < length && Unsafe.Add(ref array1Ref, matchLen) == Unsafe.Add(ref array2Ref, matchLen))
             {
-                if (length >= 12)
-                {
-                    Vector128<uint> a0 = Unsafe.As<uint, Vector128<uint>>(ref array1Ref);
-                    Vector128<uint> a1 = Unsafe.As<uint, Vector128<uint>>(ref array2Ref);
-
-                    do
-                    {
-                        // Loop unrolling and early load both provide a speedup.
-                        Vector128<uint> cmpA = Sse2.CompareEqual(a0, a1);
-                        Vector128<uint> b0 = Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array1Ref, matchLen + 4));
-                        Vector128<uint> b1 = Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array2Ref, matchLen + 4));
-                        if (Sse2.MoveMask(cmpA.AsByte()) != 0xffff)
-                        {
-                            break;
-                        }
-
-                        matchLen += 4;
-
-                        Vector128<uint> cmpB = Sse2.CompareEqual(b0, b1);
-                        a0 = Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array1Ref, matchLen + 4));
-                        a1 = Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array2Ref, matchLen + 4));
-                        if (Sse2.MoveMask(cmpB.AsByte()) != 0xffff)
-                        {
-                            break;
-                        }
-
-                        matchLen += 4;
-                    }
-                    while (matchLen + 12 < length);
-                }
-                else
-                {
-                    // Unroll the potential first two loops.
-                    if (length >= 4
-                        && Sse2.MoveMask(
-                            Sse2.CompareEqual(
-                                Unsafe.As<uint, Vector128<uint>>(ref array1Ref),
-                                Unsafe.As<uint, Vector128<uint>>(ref array2Ref)).AsByte()) == 0xffff)
-                    {
-                        matchLen = 4;
-                        if (length >= 8
-                            && Sse2.MoveMask(
-                                Sse2.CompareEqual(
-                                    Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array1Ref, 4)),
-                                    Unsafe.As<uint, Vector128<uint>>(ref Unsafe.Add(ref array2Ref, 4))).AsByte()) == 0xffff)
-                        {
-                            matchLen = 8;
-                        }
-                    }
-                }
+                matchLen++;
             }
-#endif
-            {
-                while (matchLen < length && Unsafe.Add(ref array1Ref, matchLen) == Unsafe.Add(ref array2Ref, matchLen))
-                {
-                    matchLen++;
-                }
 
-                return matchLen;
-            }
+            return matchLen;
         }
 
         [MethodImpl(InliningOptions.ShortMethod)]
diff --git a/tests/ImageSharp.Tests/Formats/WebP/LosslessUtilsTests.cs b/tests/ImageSharp.Tests/Formats/WebP/LosslessUtilsTests.cs
index b6f15e5376..62e23c1cdf 100644
--- a/tests/ImageSharp.Tests/Formats/WebP/LosslessUtilsTests.cs
+++ b/tests/ImageSharp.Tests/Formats/WebP/LosslessUtilsTests.cs
@@ -10,24 +10,6 @@ namespace SixLabors.ImageSharp.Tests.Formats.Webp
     [Trait("Format", "Webp")]
     public class LosslessUtilsTests
     {
-        private static void RunVectorMismatchTest()
-        {
-            uint[] array1 = { 4278193152, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896 };
-            uint[] array2 = { 4278193152, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896, 4278192896 };
-            int expected1 = 18;
-
-            // Test unroll first two loops path also.
-            uint[] array3 = { 4279238656, 4278714368, 4279238656, 4279238656, 4279238656, 4279238656, 4279238896, 4279238896, 4279238884 };
-            uint[] array4 = { 4279238656, 4278714368, 4279238656, 4279238656, 4278190080, 4278190080, 4278190080, 4278190080, 4278190080 };
-            int expected2 = 4;
-
-            int actual1 = LosslessUtils.VectorMismatch(array1, array2, 18);
-            int actual2 = LosslessUtils.VectorMismatch(array3, array4, 9);
-
-            Assert.Equal(expected1, actual1);
-            Assert.Equal(expected2, actual2);
-        }
-
         private static void RunSubtractGreenTest()
         {
             uint[] pixelData =
@@ -211,9 +193,6 @@ namespace SixLabors.ImageSharp.Tests.Formats.Webp
             }
         }
 
-        [Fact]
-        public void VectorMismatch_Works() => RunVectorMismatchTest();
-
         [Fact]
         public void Predictor11_Works() => RunPredictor11Test();
 
@@ -236,12 +215,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Webp
         public void TransformColorInverse_Works() => RunTransformColorInverseTest();
 
 #if SUPPORTS_RUNTIME_INTRINSICS
-        [Fact]
-        public void VectorMismatch_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVectorMismatchTest, HwIntrinsics.AllowAll);
-
-        [Fact]
-        public void VectorMismatch_WithoutSSE2_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunVectorMismatchTest, HwIntrinsics.DisableSSE2);
-
+        
         [Fact]
         public void Predictor11_WithHardwareIntrinsics_Works() => FeatureTestRunner.RunWithHwIntrinsicsFeature(RunPredictor11Test, HwIntrinsics.AllowAll);