Greatly optimized PaethFilter

src/ImageSharp/Formats/Png/Filters/PaethFilter.cs

@@ -6,6 +6,11 @@ using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+#endif
+
 namespace SixLabors.ImageSharp.Formats.Png.Filters
 {
     /// <summary>
@@ -84,7 +89,30 @@ namespace SixLabors.ImageSharp.Formats.Png.Filters
             }
 
 #if SUPPORTS_RUNTIME_INTRINSICS
-            if (Vector.IsHardwareAccelerated)
+            if (Avx2.IsSupported)
+            {
+                Vector256<int> sumAccumulator = Vector256<int>.Zero;
+
+                for (int xLeft = x - bytesPerPixel; x + Vector256<byte>.Count <= scanline.Length; xLeft += Vector256<byte>.Count)
+                {
+                    Vector256<byte> scan = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
+                    Vector256<byte> left = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
+                    Vector256<byte> above = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref prevBaseRef, x));
+                    Vector256<byte> upperLeft = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref prevBaseRef, xLeft));
+
+                    Vector256<byte> res = Avx2.Subtract(scan, PaethPredictor(left, above, upperLeft));
+                    Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = res; // +1 to skip filter type
+                    x += Vector256<byte>.Count;
+
+                    sumAccumulator = Avx2.Add(sumAccumulator, Avx2.SumAbsoluteDifferences(Avx2.Abs(res.AsSByte()), Vector256<byte>.Zero).AsInt32());
+                }
+
+                for (int i = 0; i < Vector256<int>.Count; i++)
+                {
+                    sum += sumAccumulator.GetElement(i);
+                }
+            }
+            else if (Vector.IsHardwareAccelerated)
             {
                 Vector<uint> sumAccumulator = Vector<uint>.Zero;
 
@@ -155,6 +183,39 @@ namespace SixLabors.ImageSharp.Formats.Png.Filters
             return upperLeft;
         }
 
+#if SUPPORTS_RUNTIME_INTRINSICS
+        private static Vector256<byte> PaethPredictor(Vector256<byte> left, Vector256<byte> above, Vector256<byte> upleft)
+        {
+            Vector256<byte> zero = Vector256<byte>.Zero;
+
+            // Here, we refactor pa = abs(p - left) = abs(left + above - upleft - left)
+            // to pa = abs(above - upleft).  Same deal for pb.
+            // Using saturated subtraction, if the result is negative, the output is zero.
+            // If we subtract in both directions and `or` the results, only one can be
+            // non-zero, so we end up with the absolute value.
+            Vector256<byte> sac = Avx2.SubtractSaturate(above, upleft);
+            Vector256<byte> sbc = Avx2.SubtractSaturate(left, upleft);
+            Vector256<byte> pa = Avx2.Or(Avx2.SubtractSaturate(upleft, above), sac);
+            Vector256<byte> pb = Avx2.Or(Avx2.SubtractSaturate(upleft, left), sbc);
+
+            // pc = abs(left + above - upleft - upleft), or abs(left - upleft + above - upleft).
+            // We've already calculated left - upleft and above - upleft in `sac` and `sbc`.
+            // If they are both negative or both positive, the absolute value of their
+            // sum can't possibly be less than `pa` or `pb`, so we'll never use the value.
+            // We make a mask that sets the value to 255 if they either both got
+            // saturated to zero or both didn't.  Then we calculate the absolute value
+            // of their difference using saturated subtract and `or`, same as before,
+            // keeping the value only where the mask isn't set.
+            Vector256<byte> pm = Avx2.CompareEqual(Avx2.CompareEqual(sac, zero), Avx2.CompareEqual(sbc, zero));
+            Vector256<byte> pc = Avx2.Or(pm, Avx2.Or(Avx2.SubtractSaturate(pb, pa), Avx2.SubtractSaturate(pa, pb)));
+
+            // Finally, blend the values together.  We start with `upleft` and overwrite on
+            // tied values so that the `left`, `above`, `upleft` precedence is preserved.
+            Vector256<byte> minbc = Avx2.Min(pc, pb);
+            Vector256<byte> resbc = Avx2.BlendVariable(upleft, above, Avx2.CompareEqual(minbc, pb));
+            return Avx2.BlendVariable(resbc, left, Avx2.CompareEqual(Avx2.Min(minbc, pa), pa));
+        }
+
         private static Vector<byte> PaethPredictor(Vector<byte> left, Vector<byte> above, Vector<byte> upperLeft)
         {
             Vector.Widen(left, out Vector<ushort> a1, out Vector<ushort> a2);
@@ -185,5 +246,6 @@ namespace SixLabors.ImageSharp.Formats.Png.Filters
                     left: above,
                     right: upperLeft));
         }
+#endif
     }
 }