Better Block8x8F clamp

9 years ago · d2db3163c4
3 changed files with 120 additions and 53 deletions
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.cs
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.cs
@ -5,53 +5,120 @@
 // ReSharper disable InconsistentNaming
 // <auto-generated />
 #pragma warning disable
-using System;
-using System.Numerics;
-using System.Runtime.CompilerServices;
-
-
 namespace ImageSharp.Formats.Jpg
 {
+	using System.Numerics;
+	using System.Runtime.CompilerServices;
+
 	internal partial struct Block8x8F
    {
-        private static readonly Vector4 CMin4 = new Vector4(-128f);
-        private static readonly Vector4 CMax4 = new Vector4(127f);
-        private static readonly Vector4 COff4 = new Vector4(128f);
+        private static readonly Vector4 CMin4 = new Vector4(0F);
+        private static readonly Vector4 CMax4 = new Vector4(255F);
+        private static readonly Vector4 COff4 = new Vector4(128F);

 		/// <summary>
-        /// Transpose the block into d
+        /// Transpose the block into the destination block.
        /// </summary>
-        /// <param name="d">Destination</param>
+        /// <param name="d">The destination block</param>
 		[MethodImpl(MethodImplOptions.AggressiveInlining)]
        public void TransposeInto(ref Block8x8F d)
        {
-            d.V0L.X = V0L.X; d.V1L.X = V0L.Y; d.V2L.X = V0L.Z; d.V3L.X = V0L.W; d.V4L.X = V0R.X; d.V5L.X = V0R.Y; d.V6L.X = V0R.Z; d.V7L.X = V0R.W; 
-            d.V0L.Y = V1L.X; d.V1L.Y = V1L.Y; d.V2L.Y = V1L.Z; d.V3L.Y = V1L.W; d.V4L.Y = V1R.X; d.V5L.Y = V1R.Y; d.V6L.Y = V1R.Z; d.V7L.Y = V1R.W; 
-            d.V0L.Z = V2L.X; d.V1L.Z = V2L.Y; d.V2L.Z = V2L.Z; d.V3L.Z = V2L.W; d.V4L.Z = V2R.X; d.V5L.Z = V2R.Y; d.V6L.Z = V2R.Z; d.V7L.Z = V2R.W; 
-            d.V0L.W = V3L.X; d.V1L.W = V3L.Y; d.V2L.W = V3L.Z; d.V3L.W = V3L.W; d.V4L.W = V3R.X; d.V5L.W = V3R.Y; d.V6L.W = V3R.Z; d.V7L.W = V3R.W; 
-            d.V0R.X = V4L.X; d.V1R.X = V4L.Y; d.V2R.X = V4L.Z; d.V3R.X = V4L.W; d.V4R.X = V4R.X; d.V5R.X = V4R.Y; d.V6R.X = V4R.Z; d.V7R.X = V4R.W; 
-            d.V0R.Y = V5L.X; d.V1R.Y = V5L.Y; d.V2R.Y = V5L.Z; d.V3R.Y = V5L.W; d.V4R.Y = V5R.X; d.V5R.Y = V5R.Y; d.V6R.Y = V5R.Z; d.V7R.Y = V5R.W; 
-            d.V0R.Z = V6L.X; d.V1R.Z = V6L.Y; d.V2R.Z = V6L.Z; d.V3R.Z = V6L.W; d.V4R.Z = V6R.X; d.V5R.Z = V6R.Y; d.V6R.Z = V6R.Z; d.V7R.Z = V6R.W; 
-            d.V0R.W = V7L.X; d.V1R.W = V7L.Y; d.V2R.W = V7L.Z; d.V3R.W = V7L.W; d.V4R.W = V7R.X; d.V5R.W = V7R.Y; d.V6R.W = V7R.Z; d.V7R.W = V7R.W; 
+            d.V0L.X = V0L.X;
+            d.V1L.X = V0L.Y;
+            d.V2L.X = V0L.Z;
+            d.V3L.X = V0L.W;
+            d.V4L.X = V0R.X;
+            d.V5L.X = V0R.Y;
+            d.V6L.X = V0R.Z;
+            d.V7L.X = V0R.W;
+
+            d.V0L.Y = V1L.X;
+            d.V1L.Y = V1L.Y;
+            d.V2L.Y = V1L.Z;
+            d.V3L.Y = V1L.W;
+            d.V4L.Y = V1R.X;
+            d.V5L.Y = V1R.Y;
+            d.V6L.Y = V1R.Z;
+            d.V7L.Y = V1R.W;
+
+            d.V0L.Z = V2L.X;
+            d.V1L.Z = V2L.Y;
+            d.V2L.Z = V2L.Z;
+            d.V3L.Z = V2L.W;
+            d.V4L.Z = V2R.X;
+            d.V5L.Z = V2R.Y;
+            d.V6L.Z = V2R.Z;
+            d.V7L.Z = V2R.W;
+
+            d.V0L.W = V3L.X;
+            d.V1L.W = V3L.Y;
+            d.V2L.W = V3L.Z;
+            d.V3L.W = V3L.W;
+            d.V4L.W = V3R.X;
+            d.V5L.W = V3R.Y;
+            d.V6L.W = V3R.Z;
+            d.V7L.W = V3R.W;
+
+            d.V0R.X = V4L.X;
+            d.V1R.X = V4L.Y;
+            d.V2R.X = V4L.Z;
+            d.V3R.X = V4L.W;
+            d.V4R.X = V4R.X;
+            d.V5R.X = V4R.Y;
+            d.V6R.X = V4R.Z;
+            d.V7R.X = V4R.W;
+
+            d.V0R.Y = V5L.X;
+            d.V1R.Y = V5L.Y;
+            d.V2R.Y = V5L.Z;
+            d.V3R.Y = V5L.W;
+            d.V4R.Y = V5R.X;
+            d.V5R.Y = V5R.Y;
+            d.V6R.Y = V5R.Z;
+            d.V7R.Y = V5R.W;
+
+            d.V0R.Z = V6L.X;
+            d.V1R.Z = V6L.Y;
+            d.V2R.Z = V6L.Z;
+            d.V3R.Z = V6L.W;
+            d.V4R.Z = V6R.X;
+            d.V5R.Z = V6R.Y;
+            d.V6R.Z = V6R.Z;
+            d.V7R.Z = V6R.W;
+
+            d.V0R.W = V7L.X;
+            d.V1R.W = V7L.Y;
+            d.V2R.W = V7L.Z;
+            d.V3R.W = V7L.W;
+            d.V4R.W = V7R.X;
+            d.V5R.W = V7R.Y;
+            d.V6R.W = V7R.Z;
+            d.V7R.W = V7R.W;
        }

 		/// <summary>
        /// Level shift by +128, clip to [0, 255]
        /// </summary>
-        /// <param name="d">Destination</param>
+        /// <param name="d">The destination block</param>
        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        internal void TransformByteConvetibleColorValuesInto(ref Block8x8F d)
        {
-            d.V0L = Vector4.Max(Vector4.Min(V0L, CMax4), CMin4) + COff4;d.V0R = Vector4.Max(Vector4.Min(V0R, CMax4), CMin4) + COff4;
-            d.V1L = Vector4.Max(Vector4.Min(V1L, CMax4), CMin4) + COff4;d.V1R = Vector4.Max(Vector4.Min(V1R, CMax4), CMin4) + COff4;
-            d.V2L = Vector4.Max(Vector4.Min(V2L, CMax4), CMin4) + COff4;d.V2R = Vector4.Max(Vector4.Min(V2R, CMax4), CMin4) + COff4;
-            d.V3L = Vector4.Max(Vector4.Min(V3L, CMax4), CMin4) + COff4;d.V3R = Vector4.Max(Vector4.Min(V3R, CMax4), CMin4) + COff4;
-            d.V4L = Vector4.Max(Vector4.Min(V4L, CMax4), CMin4) + COff4;d.V4R = Vector4.Max(Vector4.Min(V4R, CMax4), CMin4) + COff4;
-            d.V5L = Vector4.Max(Vector4.Min(V5L, CMax4), CMin4) + COff4;d.V5R = Vector4.Max(Vector4.Min(V5R, CMax4), CMin4) + COff4;
-            d.V6L = Vector4.Max(Vector4.Min(V6L, CMax4), CMin4) + COff4;d.V6R = Vector4.Max(Vector4.Min(V6R, CMax4), CMin4) + COff4;
-            d.V7L = Vector4.Max(Vector4.Min(V7L, CMax4), CMin4) + COff4;d.V7R = Vector4.Max(Vector4.Min(V7R, CMax4), CMin4) + COff4;
+            d.V0L = Vector4.Clamp(V0L + COff4, CMin4, CMax4);
+            d.V0R = Vector4.Clamp(V0R + COff4, CMin4, CMax4);
+            d.V1L = Vector4.Clamp(V1L + COff4, CMin4, CMax4);
+            d.V1R = Vector4.Clamp(V1R + COff4, CMin4, CMax4);
+            d.V2L = Vector4.Clamp(V2L + COff4, CMin4, CMax4);
+            d.V2R = Vector4.Clamp(V2R + COff4, CMin4, CMax4);
+            d.V3L = Vector4.Clamp(V3L + COff4, CMin4, CMax4);
+            d.V3R = Vector4.Clamp(V3R + COff4, CMin4, CMax4);
+            d.V4L = Vector4.Clamp(V4L + COff4, CMin4, CMax4);
+            d.V4R = Vector4.Clamp(V4R + COff4, CMin4, CMax4);
+            d.V5L = Vector4.Clamp(V5L + COff4, CMin4, CMax4);
+            d.V5R = Vector4.Clamp(V5R + COff4, CMin4, CMax4);
+            d.V6L = Vector4.Clamp(V6L + COff4, CMin4, CMax4);
+            d.V6R = Vector4.Clamp(V6R + COff4, CMin4, CMax4);
+            d.V7L = Vector4.Clamp(V7L + COff4, CMin4, CMax4);
+            d.V7R = Vector4.Clamp(V7R + COff4, CMin4, CMax4);
        }
-
-
 	}
 }
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.tt
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.tt
@ -11,31 +11,29 @@
 <#@ output extension=".cs" #>
 // <auto-generated />
 #pragma warning disable
-using System;
-using System.Numerics;
-using System.Runtime.CompilerServices;
-
 <#
 char[] coordz = {'X', 'Y', 'Z', 'W'};
 #>
-
 namespace ImageSharp.Formats.Jpg
 {
+	using System.Numerics;
+	using System.Runtime.CompilerServices;
+
 	internal partial struct Block8x8F
    {
-        private static readonly Vector4 CMin4 = new Vector4(-128f);
-        private static readonly Vector4 CMax4 = new Vector4(127f);
-        private static readonly Vector4 COff4 = new Vector4(128f);
+        private static readonly Vector4 CMin4 = new Vector4(0F);
+        private static readonly Vector4 CMax4 = new Vector4(255F);
+        private static readonly Vector4 COff4 = new Vector4(128F);

 		/// <summary>
-        /// Transpose the block into d
+        /// Transpose the block into the destination block.
        /// </summary>
-        /// <param name="d">Destination</param>
+        /// <param name="d">The destination block</param>
 		[MethodImpl(MethodImplOptions.AggressiveInlining)]
        public void TransposeInto(ref Block8x8F d)
        {
            <#
-			PushIndent("            "); 
+			PushIndent("            ");

            for (int i = 0; i < 8; i++)
            {
@ -44,13 +42,16 @@ namespace ImageSharp.Formats.Jpg

                for (int j = 0; j < 8; j++)
                {
+					if(i > 0 && j == 0){
+					WriteLine("");
+					}
+
                    char srcCoord = coordz[j % 4];
                    char srcSide = (j / 4) % 2 == 0 ? 'L' : 'R';
-                    
-                    string expression = $"d.V{j}{destSide}.{destCoord} = V{i}{srcSide}.{srcCoord}; ";
+
+                    string expression = $"d.V{j}{destSide}.{destCoord} = V{i}{srcSide}.{srcCoord};\r\n";
 					Write(expression);
                }
-				WriteLine("");
            }
 			PopIndent();
 			#>
@ -59,27 +60,24 @@ namespace ImageSharp.Formats.Jpg
 		/// <summary>
        /// Level shift by +128, clip to [0, 255]
        /// </summary>
-        /// <param name="d">Destination</param>
+        /// <param name="d">The destination block</param>
        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        internal void TransformByteConvetibleColorValuesInto(ref Block8x8F d)
        {
            <#

-			PushIndent("            "); 
+			PushIndent("            ");

            for (int i = 0; i < 8; i++)
            {
                for (int j = 0; j < 2; j++)
                {
 					char side = j == 0 ? 'L' : 'R';
-					Write($"d.V{i}{side} = Vector4.Max(Vector4.Min(V{i}{side}, CMax4), CMin4) + COff4;");
+					Write($"d.V{i}{side} = Vector4.Clamp(V{i}{side} + COff4, CMin4, CMax4);\r\n");
                }
-				WriteLine("");
            }
 			PopIndent();
 			#>
        }
-
-
 	}
 }
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
@ -57,6 +57,9 @@ namespace ImageSharp.Formats.Jpg
        public Vector4 V7R;
 #pragma warning restore SA1600 // ElementsMustBeDocumented

+        private static readonly Vector4 NegativeOne = new Vector4(-1);
+        private static readonly Vector4 Offset = new Vector4(.5F);
+
        /// <summary>
        /// Get/Set scalar elements at a given index
        /// </summary>
@ -402,12 +405,11 @@ namespace ImageSharp.Formats.Jpg
        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        private static Vector4 DivideRound(Vector4 dividend, Vector4 divisor)
        {
-            // sign(v) = max(min(v, 1), -1)
-            Vector4 sign = Vector4.Min(dividend, Vector4.One);
-            sign = Vector4.Max(sign, new Vector4(-1));
+            // sign(dividend) = max(min(dividend, 1), -1)
+            var sign = Vector4.Clamp(dividend, NegativeOne, Vector4.One);

            // AlmostRound(dividend/divisor) = dividend/divisior + 0.5*sign(dividend)
-            return (dividend / divisor) + (sign * new Vector4(0.5f));
+            return (dividend / divisor) + (sign * Offset);
        }
    }
 }