Added Unrolled_1: Inner loop unrolled fully

pzemtsov · May 2, 2014 · ba6a1e7 · ba6a1e7
1 parent b2d647e
commit ba6a1e7
Showing 1 changed file with 52 additions and 2 deletions.
diff --git a/E1.java b/E1.java
@@ -1,4 +1,4 @@
-/**  E1 demultiplexer, revision 9
+/**  E1 demultiplexer, revision 10
   *  Created reference implementation
   *  Added test and measurement code
   *  Added correctness test
@@ -13,6 +13,9 @@
   *    Added Dst_First_1: Similar to Src_First_2 but different order of loops
   *    Added Dst_First_2: Dst_First_1 manually optimised
   *    Added Dst_First_3: Dst_First_1 with hard-coded array size
+  *
+  *  Started unrolled family (based on Dst_First_1)
+  *    Added Unrolled_1:      Inner loop unrolled fully
   */
 
 import java.util.Random;
@@ -186,6 +189,52 @@ public void demux (byte[] src, byte[][] dst)
         }
     }
 
+    static final class Unrolled_1 implements Demux
+    {
+        public void demux (byte[] src, byte[][] dst)
+        {
+            assert NUM_TIMESLOTS == 32;
+            assert DST_SIZE == 64;
+            assert src.length == NUM_TIMESLOTS * DST_SIZE;
+
+            for (int j = 0; j < NUM_TIMESLOTS; j++) {
+                final byte[] d = dst[j];
+                d[ 0] = src[j+32* 0]; d[ 1] = src[j+32* 1];
+                d[ 2] = src[j+32* 2]; d[ 3] = src[j+32* 3];
+                d[ 4] = src[j+32* 4]; d[ 5] = src[j+32* 5];
+                d[ 6] = src[j+32* 6]; d[ 7] = src[j+32* 7];
+                d[ 8] = src[j+32* 8]; d[ 9] = src[j+32* 9];
+                d[10] = src[j+32*10]; d[11] = src[j+32*11];
+                d[12] = src[j+32*12]; d[13] = src[j+32*13];
+                d[14] = src[j+32*14]; d[15] = src[j+32*15];
+                d[16] = src[j+32*16]; d[17] = src[j+32*17];
+                d[18] = src[j+32*18]; d[19] = src[j+32*19];
+                d[20] = src[j+32*20]; d[21] = src[j+32*21];
+                d[22] = src[j+32*22]; d[23] = src[j+32*23];
+                d[24] = src[j+32*24]; d[25] = src[j+32*25];
+                d[26] = src[j+32*26]; d[27] = src[j+32*27];
+                d[28] = src[j+32*28]; d[29] = src[j+32*29];
+                d[30] = src[j+32*30]; d[31] = src[j+32*31];
+                d[32] = src[j+32*32]; d[33] = src[j+32*33];
+                d[34] = src[j+32*34]; d[35] = src[j+32*35];
+                d[36] = src[j+32*36]; d[37] = src[j+32*37];
+                d[38] = src[j+32*38]; d[39] = src[j+32*39];
+                d[40] = src[j+32*40]; d[41] = src[j+32*41];
+                d[42] = src[j+32*42]; d[43] = src[j+32*43];
+                d[44] = src[j+32*44]; d[45] = src[j+32*45];
+                d[46] = src[j+32*46]; d[47] = src[j+32*47];
+                d[48] = src[j+32*48]; d[49] = src[j+32*49];
+                d[50] = src[j+32*50]; d[51] = src[j+32*51];
+                d[52] = src[j+32*52]; d[53] = src[j+32*53];
+                d[54] = src[j+32*54]; d[55] = src[j+32*55];
+                d[56] = src[j+32*56]; d[57] = src[j+32*57];
+                d[58] = src[j+32*58]; d[59] = src[j+32*59];
+                d[60] = src[j+32*60]; d[61] = src[j+32*61];
+                d[62] = src[j+32*62]; d[63] = src[j+32*63];
+            }
+        }
+    }
+
     public static void main (String [] args) 
     {
 //      measure (new Reference ());
@@ -194,6 +243,7 @@ public static void main (String [] args)
 //      measure (new Src_First_3 ());
 //      measure (new Dst_First_1 ());
 //      measure (new Dst_First_2 ());
-        measure (new Dst_First_3 ());
+//      measure (new Dst_First_3 ());
+        measure (new Unrolled_1 ());
     }
 }