More optimizations

• Speedup backprop
• Fix gpu backend for different CUDA versions

juno_samples/rodinia/bfs/src/cpu.sch

@@ -46,16 +46,14 @@ let traverse_body = outline(inner);
 let (outer, inner) = fork-reshape[[1], [0]](collect);
 let collect_body = outline(inner);
 
-let init_body = init;
-// Following code seems to generate breaking RT code
-//fork-tile[32, 0, false, true](init);
-//let (outer, inner) = fork-reshape[[1], [0]](init);
-//let init_body = outline(inner);
-//inline(bfs@cost_init);
+fork-tile[32, 0, false, true](init);
+let (outer, inner) = fork-reshape[[1], [0]](init);
+let init_body = outline(inner);
 
 inline(bfs@loop1, bfs@loop2);
 delete-uncalled(*);
 const-inline(*);
+simpl!(*);
 
 unforkify(init_body, traverse_body, collect_body);
 simpl!(*);