Optimized vector rgb pixel matrix scaling

src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs

@@ -577,6 +577,24 @@ namespace SixLabors.ImageSharp
                 }
             }
 
+            /// <summary>
+            /// Scales 8x8 matrix to 4x2 using 2x2 average
+            /// </summary>
+            /// <param name="v">Input matrix consisting of 4 256bit vectors, first row: (v[0], v[2]), second row: (v[1], v[3])</param>
+            /// <returns>256bit vector containing upper and lower scaled parts of the input matrix</returns>
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            public static Vector256<float> Scale16x2_8x1(ReadOnlySpan<Vector256<float>> v)
+            {
+                DebugGuard.IsTrue(v.Length == 4, "Input span must consist of 4 elements");
+
+                var f025 = Vector256.Create(0.25f);
+
+                Vector256<float> left = Avx.Add(v[0], v[2]);
+                Vector256<float> right = Avx.Add(v[1], v[3]);
+                Vector256<float> avg2x2 = Avx.Multiply(Avx.HorizontalAdd(left, right), f025);
+
+                return Avx2.Permute4x64(avg2x2.AsDouble(), 0b11_01_10_00).AsSingle();
+            }
 
             /// <summary>
             /// <see cref="ByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).

src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbToYCbCrConverterVectorized.cs

@@ -221,9 +221,9 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
                     bDataLanes[j] = b;
                 }
 
-                r = Scale_8x4_4x2(rDataLanes);
-                g = Scale_8x4_4x2(gDataLanes);
-                b = Scale_8x4_4x2(bDataLanes);
+                r = SimdUtils.HwIntrinsics.Scale16x2_8x1(rDataLanes);
+                g = SimdUtils.HwIntrinsics.Scale16x2_8x1(gDataLanes);
+                b = SimdUtils.HwIntrinsics.Scale16x2_8x1(bDataLanes);
 
                 // 128F + ((-0.168736F * r) - (0.331264F * g) + (0.5F * b))
                 Unsafe.Add(ref destCbRef, i) = Avx.Add(f128, SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f05, b), fn0331264, g), fn0168736, r));
@@ -233,27 +233,5 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
             }
 #endif
         }
-
-#if SUPPORTS_RUNTIME_INTRINSICS
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector256<float> Scale_8x4_4x2(Span<Vector256<float>> v)
-        {
-            Vector256<int> switchInnerDoubleWords = Unsafe.As<byte, Vector256<int>>(ref MemoryMarshal.GetReference(SimdUtils.HwIntrinsics.PermuteMaskSwitchInnerDWords8x32));
-            var f025 = Vector256.Create(0.25f);
-
-            Vector256<float> topPairSum = SumHorizontalPairs(v[0], v[2]);
-            Vector256<float> botPairSum = SumHorizontalPairs(v[1], v[3]);
-
-            return Avx2.PermuteVar8x32(Avx.Multiply(SumVerticalPairs(topPairSum, botPairSum), f025), switchInnerDoubleWords);
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector256<float> SumHorizontalPairs(Vector256<float> v0, Vector256<float> v1)
-            => Avx.Add(Avx.Shuffle(v0, v1, 0b10_00_10_00), Avx.Shuffle(v0, v1, 0b11_01_11_01));
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector256<float> SumVerticalPairs(Vector256<float> v0, Vector256<float> v1)
-            => Avx.Add(Avx.Shuffle(v0, v1, 0b01_00_01_00), Avx.Shuffle(v0, v1, 0b11_10_11_10));
-#endif
     }
 }