Merge pull request #1847 from br3aker/jpeg-idct-zigzag-optimization

Jpeg new IDCT & fused zigzag/transpose steps

Author: JimBobSquarePants

shared-infrastructure

@@ -337,6 +337,64 @@ public nint GetLastNonZeroIndex()
             }
         }
 
+        /// <summary>
+        /// Transpose the block inplace.
+        /// </summary>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public void TransposeInplace()
+        {
+            ref short elemRef = ref Unsafe.As<Block8x8, short>(ref this);
+
+            // row #0
+            Swap(ref Unsafe.Add(ref elemRef, 1), ref Unsafe.Add(ref elemRef, 8));
+            Swap(ref Unsafe.Add(ref elemRef, 2), ref Unsafe.Add(ref elemRef, 16));
+            Swap(ref Unsafe.Add(ref elemRef, 3), ref Unsafe.Add(ref elemRef, 24));
+            Swap(ref Unsafe.Add(ref elemRef, 4), ref Unsafe.Add(ref elemRef, 32));
+            Swap(ref Unsafe.Add(ref elemRef, 5), ref Unsafe.Add(ref elemRef, 40));
+            Swap(ref Unsafe.Add(ref elemRef, 6), ref Unsafe.Add(ref elemRef, 48));
+            Swap(ref Unsafe.Add(ref elemRef, 7), ref Unsafe.Add(ref elemRef, 56));
+
+            // row #1
+            Swap(ref Unsafe.Add(ref elemRef, 10), ref Unsafe.Add(ref elemRef, 17));
+            Swap(ref Unsafe.Add(ref elemRef, 11), ref Unsafe.Add(ref elemRef, 25));
+            Swap(ref Unsafe.Add(ref elemRef, 12), ref Unsafe.Add(ref elemRef, 33));
+            Swap(ref Unsafe.Add(ref elemRef, 13), ref Unsafe.Add(ref elemRef, 41));
+            Swap(ref Unsafe.Add(ref elemRef, 14), ref Unsafe.Add(ref elemRef, 49));
+            Swap(ref Unsafe.Add(ref elemRef, 15), ref Unsafe.Add(ref elemRef, 57));
+
+            // row #2
+            Swap(ref Unsafe.Add(ref elemRef, 19), ref Unsafe.Add(ref elemRef, 26));
+            Swap(ref Unsafe.Add(ref elemRef, 20), ref Unsafe.Add(ref elemRef, 34));
+            Swap(ref Unsafe.Add(ref elemRef, 21), ref Unsafe.Add(ref elemRef, 42));
+            Swap(ref Unsafe.Add(ref elemRef, 22), ref Unsafe.Add(ref elemRef, 50));
+            Swap(ref Unsafe.Add(ref elemRef, 23), ref Unsafe.Add(ref elemRef, 58));
+
+            // row #3
+            Swap(ref Unsafe.Add(ref elemRef, 28), ref Unsafe.Add(ref elemRef, 35));
+            Swap(ref Unsafe.Add(ref elemRef, 29), ref Unsafe.Add(ref elemRef, 43));
+            Swap(ref Unsafe.Add(ref elemRef, 30), ref Unsafe.Add(ref elemRef, 51));
+            Swap(ref Unsafe.Add(ref elemRef, 31), ref Unsafe.Add(ref elemRef, 59));
+
+            // row #4
+            Swap(ref Unsafe.Add(ref elemRef, 37), ref Unsafe.Add(ref elemRef, 44));
+            Swap(ref Unsafe.Add(ref elemRef, 38), ref Unsafe.Add(ref elemRef, 52));
+            Swap(ref Unsafe.Add(ref elemRef, 39), ref Unsafe.Add(ref elemRef, 60));
+
+            // row #5
+            Swap(ref Unsafe.Add(ref elemRef, 46), ref Unsafe.Add(ref elemRef, 53));
+            Swap(ref Unsafe.Add(ref elemRef, 47), ref Unsafe.Add(ref elemRef, 61));
+
+            // row #6
+            Swap(ref Unsafe.Add(ref elemRef, 55), ref Unsafe.Add(ref elemRef, 62));
+
+            static void Swap(ref short a, ref short b)
+            {
+                short tmp = a;
+                a = b;
+                b = tmp;
+            }
+        }
+
         /// <summary>
         /// Calculate the total sum of absolute differences of elements in 'a' and 'b'.
         /// </summary>

src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs

@@ -502,7 +502,7 @@ private void DecodeBlockBaseline(
                 {
                     i += r;
                     s = buffer.Receive(s);
-                    Unsafe.Add(ref blockDataRef, ZigZag.ZigZagOrder[i++]) = (short)s;
+                    Unsafe.Add(ref blockDataRef, ZigZag.TransposingOrder[i++]) = (short)s;
                 }
                 else
                 {
@@ -571,7 +571,7 @@ private void DecodeBlockProgressiveAC(ref Block8x8 block, ref HuffmanTable acTab
                     if (s != 0)
                     {
                         s = buffer.Receive(s);
-                        Unsafe.Add(ref blockDataRef, ZigZag.ZigZagOrder[i]) = (short)(s << low);
+                        Unsafe.Add(ref blockDataRef, ZigZag.TransposingOrder[i]) = (short)(s << low);
                     }
                     else
                     {
@@ -647,7 +647,7 @@ private void DecodeBlockProgressiveACRefined(ref short blockDataRef, ref Huffman
 
                     do
                     {
-                        ref short coef = ref Unsafe.Add(ref blockDataRef, ZigZag.ZigZagOrder[k]);
+                        ref short coef = ref Unsafe.Add(ref blockDataRef, ZigZag.TransposingOrder[k]);
                         if (coef != 0)
                         {
                             buffer.CheckBits();
@@ -673,7 +673,7 @@ private void DecodeBlockProgressiveACRefined(ref short blockDataRef, ref Huffman
 
                     if ((s != 0) && (k < 64))
                     {
-                        Unsafe.Add(ref blockDataRef, ZigZag.ZigZagOrder[k]) = (short)s;
+                        Unsafe.Add(ref blockDataRef, ZigZag.TransposingOrder[k]) = (short)s;
                     }
                 }
             }
@@ -682,7 +682,7 @@ private void DecodeBlockProgressiveACRefined(ref short blockDataRef, ref Huffman
             {
                 for (; k <= end; k++)
                 {
-                    ref short coef = ref Unsafe.Add(ref blockDataRef, ZigZag.ZigZagOrder[k]);
+                    ref short coef = ref Unsafe.Add(ref blockDataRef, ZigZag.TransposingOrder[k]);
 
                     if (coef != 0)
                     {

src/ImageSharp/Formats/Jpeg/Components/Decoder/HuffmanScanDecoder.cs

@@ -18,11 +18,6 @@ internal struct JpegBlockPostProcessor
         /// </summary>
         public Block8x8F SourceBlock;
 
-        /// <summary>
-        /// Temporal block to store intermediate computation results.
-        /// </summary>
-        public Block8x8F WorkspaceBlock;
-
         /// <summary>
         /// The quantization table as <see cref="Block8x8F"/>.
         /// </summary>
@@ -45,7 +40,6 @@ public JpegBlockPostProcessor(IRawJpegData decoder, IJpegComponent component)
             this.subSamplingDivisors = component.SubSamplingDivisors;
 
             this.SourceBlock = default;
-            this.WorkspaceBlock = default;
         }
 
         /// <summary>
@@ -71,7 +65,7 @@ public void ProcessBlockColorsInto(
             // Dequantize:
             block.MultiplyInPlace(ref this.DequantiazationTable);
 
-            FastFloatingPointDCT.TransformIDCT(ref block, ref this.WorkspaceBlock);
+            FastFloatingPointDCT.TransformIDCT(ref block);
 
             // To conform better to libjpeg we actually NEED TO loose precision here.
             // This is because they store blocks as Int16 between all the operations.