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/backprop/src/cpu.sch

@@ -30,27 +30,11 @@ simpl!(*);
 
 infer-schedules(*);
 
-fork-tile[32, 0, false, true](layer_forward@outer_loop \ layer_forward@inner_loop);
-let (forward_outer, forward_inner) = fork-reshape[[1], [0]](layer_forward@outer_loop \ layer_forward@inner_loop);
-
-fork-tile[32, 0, false, true](adjust_weights);
-let (adjust_outer, adjust_inner) = fork-reshape[[1], [0, 2]](adjust_weights);
-
-let forward_body = outline(forward_inner);
-let adjust_body = outline(adjust_inner);
-
-rename["output_error"](output_error);
-rename["hidden_error"](hidden_error);
-
-let output_error_body = auto-outline(output_error).output_error;
-let hidden_error_body = auto-outline(hidden_error).hidden_error;
-
-inline(backprop);
 delete-uncalled(*);
 const-inline(*);
 
 simpl!(*);
-fork-split(forward_body, adjust_body, output_error_body, hidden_error_body);
-unforkify(forward_body, adjust_body, output_error_body, hidden_error_body);
+fork-split(*);
+unforkify(*);
 
 gcm(*);