Optimize SRAD

juno_samples/rodinia/srad/src/gpu.sch

@@ -66,18 +66,21 @@ ip-sroa(*);
 sroa(*);
 simpl!(*);
 
+fork-interchange[0, 1](main_loops);
 fork-dim-merge(main_loops);
 fork-tile[32, 0, false, true](main_loops);
 dce(main_loops);
 fork-split(main_loops);
 simpl!(main_loops);
 
+fork-interchange[0, 1](extract);
 fork-dim-merge(extract);
 fork-tile[32, 0, false, true](extract);
 dce(extract);
 fork-split(extract);
 simpl!(extract);
 
+fork-interchange[0, 1](compress);
 fork-dim-merge(compress);
 fork-tile[32, 0, false, true](compress);
 dce(compress);