Changeset 4093 for proto

Timestamp:

Aug 29, 2014, 2:56:16 AM (4 years ago)

Author:

ksherdy

Message:

Steady-state s2k code generation for 8-4-2-1 reductions.

File:

• proto/s2k/trunk/demo/transpose/transpose.cpp (modified) (8 diffs)

proto/s2k/trunk/demo/transpose/transpose.cpp

@@ -52 +52 @@
 #define LE
 #define BE
-#define LE_S2K
+#define S2K
 
 #define s2p_ideal_le(s_7, s_6, s_5, s_4, s_3, s_2, s_1, s_0, r_0, r_1, r_2, r_3, r_4, r_5, r_6, r_7) \
@@ -86 +86 @@
   } while(0)
 
-static void s2p_le(BytePack b76543210_7, BytePack b76543210_6, BytePack b76543210_5, BytePack b76543210_4, 
+static void s2p_be(BytePack b76543210_7, BytePack b76543210_6, BytePack b76543210_5, BytePack b76543210_4,
                       BytePack b76543210_3, BytePack b76543210_2, BytePack b76543210_1, BytePack b76543210_0, 
                       BitBlock & b0, BitBlock & b1, BitBlock & b2, BitBlock & b3, 
@@ -124 +124 @@
 }
 
-
-static void s2p_be(BytePack b76543210_7, BytePack b76543210_6, BytePack b76543210_5, BytePack b76543210_4, 
+static void s2p_le(BytePack b76543210_7, BytePack b76543210_6, BytePack b76543210_5, BytePack b76543210_4,
                       BytePack b76543210_3, BytePack b76543210_2, BytePack b76543210_1, BytePack b76543210_0, 
-                      BitBlock & b0, BitBlock & b1, BitBlock & b2, BitBlock & b3, 
-                      BitBlock & b4, BitBlock & b5, BitBlock & b6, BitBlock & b7) 
+                      BitBlock & b7, BitBlock & b6, BitBlock & b5, BitBlock & b4,
+                      BitBlock & b3, BitBlock & b2, BitBlock & b1, BitBlock & b0)
 {
         BitBlock b7654_0, b7654_1, b7654_2, b7654_3;
@@ -153 +152 @@
         b10_0 = hsimd<4>::packl(b3210_1, b3210_0);
 
-        b0 = hsimd<2>::packh(b76_1, b76_0);
-        b1 = hsimd<2>::packl(b76_1, b76_0);
-        b2 = hsimd<2>::packh(b54_1, b54_0);
-        b3 = hsimd<2>::packl(b54_1, b54_0);
-        b4 = hsimd<2>::packh(b32_1, b32_0);
-        b5 = hsimd<2>::packl(b32_1, b32_0);
-        b6 = hsimd<2>::packh(b10_1, b10_0);
-        b7 = hsimd<2>::packl(b10_1, b10_0);
+    b7 = hsimd<2>::packh(b76_1, b76_0);
+    b6 = hsimd<2>::packl(b76_1, b76_0);
+    b5 = hsimd<2>::packh(b54_1, b54_0);
+    b4 = hsimd<2>::packl(b54_1, b54_0);
+    b3 = hsimd<2>::packh(b32_1, b32_0);
+    b2 = hsimd<2>::packl(b32_1, b32_0);
+    b1 = hsimd<2>::packh(b10_1, b10_0);
+    b0 = hsimd<2>::packl(b10_1, b10_0);
 }
 
@@ -250 +249 @@
     BitBlock b3, b2, b1, b0;
 
-
     int k = 0;
     int bytes_avail = infile_size;
+
 
 //////////////////////////////////////////////////////////////////////////////////////////
@@ -258 +257 @@
 //////////////////////////////////////////////////////////////////////////////////////////
 
-    while (bytes_avail >= BLOCK_SIZE) 
+
+    while (bytes_avail >= BLOCK_SIZE)
     {
+{
+        BitBlock b76543210 [8];
+        memcpy(&b76543210, &infile_buffer[k], BLOCK_SIZE);
+
+        BitBlock b7654 [4];
+        BitBlock b3210 [4];
+        BitBlock b76 [2];
+        BitBlock b54 [2];
+        BitBlock b32 [2];
+        BitBlock b10 [2];
+
+        BitBlock b7[1];
+        BitBlock b6[1];
+        BitBlock b5[1];
+        BitBlock b4[1];
+        BitBlock b3[1];
+        BitBlock b2[1];
+        BitBlock b1[1];
+        BitBlock b0[1];
+
+        for(int i=0; i<8/2; i++) {
+            b7654[i] = hsimd<8>::packh(b76543210[i+i+1], b76543210[i+i]);
+        }
+
+        for(int i=0; i<8/2; i++) {
+            b3210[i] = hsimd<8>::packl(b76543210[i+i+1], b76543210[i+i]);
+        }
+
+        for(int i=0; i<4/2; i++) {
+            b76[i] = hsimd<4>::packh(b7654[i+i+1], b7654[i+i]);
+        }
+
+        for(int i=0; i<4/2; i++) {
+            b54[i] = hsimd<4>::packl(b7654[i+i+1], b7654[i+i]);
+        }
+
+        for(int i=0; i<4/2; i++) {
+            b32[i] = hsimd<4>::packh(b3210[i+i+1], b3210[i+i]);
+        }
+
+        for(int i=0; i<4/2; i++) {
+            b10[i] = hsimd<4>::packl(b3210[i+i+1], b3210[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b7[0] = hsimd<2>::packh(b76[i+i+1], b76[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b6[0] = hsimd<2>::packl(b76[i+i+1], b76[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b5[0] = hsimd<2>::packh(b54[i+i+1], b54[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b4[0] = hsimd<2>::packl(b54[i+i+1], b54[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b3[0] = hsimd<2>::packh(b32[i+i+1], b32[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b2[0] = hsimd<2>::packl(b32[i+i+1], b32[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b1[0] = hsimd<2>::packh(b10[i+i+1], b10[i+i]);
+        }
+
+        for(int i=0; i<2/2; i++) {
+            b0[0] = hsimd<2>::packl(b10[i+i+1], b10[i+i]);
+        }
+
+        printf("\n");
+        print_register<BitBlock>("b7", b7[0]);
+        print_register<BitBlock>("b6", b6[0]);
+        print_register<BitBlock>("b5", b5[0]);
+        print_register<BitBlock>("b4", b4[0]);
+        print_register<BitBlock>("b3", b3[0]);
+        print_register<BitBlock>("b2", b2[0]);
+        print_register<BitBlock>("b1", b1[0]);
+        print_register<BitBlock>("b0", b0[0]);
+
+        printf("---");
+}
+      //////////////////////////////////////////////////////////////////////////////////////////
+{
       Byte = (BytePack *) &infile_buffer[k];
 
-      #ifdef BE         
-      s2p_be(Byte[7], Byte[6], Byte[5], Byte[4], 
+      #ifdef BE
+      s2p_le(Byte[7], Byte[6], Byte[5], Byte[4],
              Byte[3], Byte[2], Byte[1], Byte[0],
-             b7, b6, b5, b4, 
+             b7, b6, b5, b4,
              b3, b2, b1, b0);
 
@@ -279 +369 @@
       #endif
 
-      #ifdef LE 
-      s2p(Byte[0], Byte[1], Byte[2], Byte[3], 
-          Byte[4], Byte[5], Byte[6], Byte[7],
-                  b0, b1, b2, b3, b4, b5, b6, b7);
+      k += BLOCK_SIZE;
+      bytes_avail -= BLOCK_SIZE;
+    }
+}
+//////////////////////////////////////////////////////////////////////////////////////////
+// Final Partial Block.
+//////////////////////////////////////////////////////////////////////////////////////////
+    if(bytes_avail > 0) 
+    {
+
+      Byte = (BytePack *) &infile_buffer[k];
+      BitBlock EOF_mask = eof_mask(bytes_avail);
+
+      #ifdef BE
+      s2p_le(Byte[7], Byte[6], Byte[5], Byte[4],
+             Byte[3], Byte[2], Byte[1], Byte[0],
+             b7, b6, b5, b4,
+             b3, b2, b1, b0);
+
+
+
+      b7 = simd_and(b7, EOF_mask);
+      b6 = simd_and(b6, EOF_mask);
+      b5 = simd_and(b5, EOF_mask);
+      b4 = simd_and(b4, EOF_mask);
+      b3 = simd_and(b3, EOF_mask);
+      b2 = simd_and(b2, EOF_mask);
+      b1 = simd_and(b1, EOF_mask);
+      b0 = simd_and(b0, EOF_mask);
 
       printf("\n");
@@ -295 +410 @@
       #endif
 
-      k += BLOCK_SIZE;
-      bytes_avail -= BLOCK_SIZE;
-    }
-
-//////////////////////////////////////////////////////////////////////////////////////////
-// Final Partial Block.
-//////////////////////////////////////////////////////////////////////////////////////////
-    if(bytes_avail > 0) 
-    {
-      
-      Byte = (BytePack *) &infile_buffer[k];
-      BitBlock EOF_mask = eof_mask(bytes_avail);
-
-      #ifdef BE        
-      s2p_be(Byte[7], Byte[6], Byte[5], Byte[4], 
-             Byte[3], Byte[2], Byte[1], Byte[0],
-             b7, b6, b5, b4, 
-             b3, b2, b1, b0);
-
-
-
-      b7 = simd_and(b7, EOF_mask);
-      b6 = simd_and(b6, EOF_mask); 
-      b5 = simd_and(b5, EOF_mask); 
-      b4 = simd_and(b4, EOF_mask); 
-      b3 = simd_and(b3, EOF_mask); 
-      b2 = simd_and(b2, EOF_mask); 
-      b1 = simd_and(b1, EOF_mask); 
-      b0 = simd_and(b0, EOF_mask);
-
-      printf("\n");
-      print_register<BitBlock>("b7", b7);
-      print_register<BitBlock>("b6", b6);
-      print_register<BitBlock>("b5", b5);
-      print_register<BitBlock>("b4", b4);
-      print_register<BitBlock>("b3", b3);
-      print_register<BitBlock>("b2", b2);
-      print_register<BitBlock>("b1", b1);
-      print_register<BitBlock>("b0", b0);
-      #endif
-
-      #ifdef LE
-          s2p(Byte[0], Byte[1], Byte[2], Byte[3], 
-          Byte[4], Byte[5], Byte[6], Byte[7],
-                  b0, b1, b2, b3, 
-          b4, b5, b6, b7);
-
-      b7 = simd_and(b7, EOF_mask);
-      b6 = simd_and(b6, EOF_mask); 
-      b5 = simd_and(b5, EOF_mask); 
-      b4 = simd_and(b4, EOF_mask); 
-      b3 = simd_and(b3, EOF_mask); 
-      b2 = simd_and(b2, EOF_mask); 
-      b1 = simd_and(b1, EOF_mask); 
-      b0 = simd_and(b0, EOF_mask);
-
-      printf("\n");
-      print_register<BitBlock>("b7", b7);
-      print_register<BitBlock>("b6", b6);
-      print_register<BitBlock>("b5", b5);
-      print_register<BitBlock>("b4", b4);
-      print_register<BitBlock>("b3", b3);
-      print_register<BitBlock>("b2", b2);
-      print_register<BitBlock>("b1", b1);
-      print_register<BitBlock>("b0", b0);
-      #endif
-
-    }
-
-}
-
-
-
+
+    }
+
+}
+
+
+