More misc. rodinia opts

• Make GCM not spill ParallelReduce phis.
• Fix unification for inter-device communication.
• Make notion of parallel indices sets more flexible, and use in infer parallel reduce.
• Misc. schedule tweaks.

juno_samples/edge_detection/src/cpu.sch

@@ -86,7 +86,7 @@ fixpoint {
 simpl!(max_gradient);
 fork-dim-merge(max_gradient);
 simpl!(max_gradient);
-fork-tile[8, 0, false, false](max_gradient);
+fork-tile[16, 0, false, false](max_gradient);
 let split = fork-split(max_gradient);
 clean-monoid-reduces(max_gradient);
 let out = outline(split._4_max_gradient.fj1);
@@ -104,11 +104,18 @@ fixpoint {
 }
 predication(reject_zero_crossings);
 simpl!(reject_zero_crossings);
+fork-tile[4, 1, false, false](reject_zero_crossings);
+fork-tile[4, 0, false, false](reject_zero_crossings);
+fork-interchange[1, 2](reject_zero_crossings);
+let split = fork-split(reject_zero_crossings);
+let reject_zero_crossings_body = outline(split._5_reject_zero_crossings.fj2);
+fork-coalesce(reject_zero_crossings, reject_zero_crossings_body);
+simpl!(reject_zero_crossings, reject_zero_crossings_body);
 
 async-call(edge_detection@le, edge_detection@zc);
 
-fork-split(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings);
-unforkify(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings);
+fork-split(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings_body);
+unforkify(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings_body);
 
 simpl!(*);