More optimizations

• A lot of misc. optimization work, notably parallelization of Rodinia CPU schedules.
• Support >32 tiles in GPU backend.
• Tweak some transformations.

juno_samples/rodinia/bfs/src/cpu.sch

@@ -40,19 +40,23 @@ simpl!(collect);
 parallel-fork(traverse, collect);
 parallel-reduce(traverse, collect);
 
-fork-tile[32, 0, false, true](init, traverse, collect);
-let (outer, inner) = fork-reshape[[1], [0]](init);
-let init_body = outline(inner);
+fork-tile[32, 0, false, true](traverse, collect);
 let (outer, inner) = fork-reshape[[1], [0]](traverse);
 let traverse_body = outline(inner);
 let (outer, inner) = fork-reshape[[1], [0]](collect);
 let collect_body = outline(inner);
 
-inline(bfs@cost_init, bfs@loop1, bfs@loop2);
+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);
+
+inline(bfs@loop1, bfs@loop2);
 delete-uncalled(*);
 const-inline(*);
 
 unforkify(init_body, traverse_body, collect_body);
 simpl!(*);
 gcm(*);
-xdot[true](bfs);