parallelize sum reduction in srad

juno_samples/rodinia/srad/src/cpu.sch

@@ -28,14 +28,28 @@ fixpoint {
   fork-guard-elim(*);
   fork-coalesce(*);
 }
+fork-dim-merge(loop1);
 simpl!(*);
-fork-interchange[0, 1](loop1);
 reduce-slf(*);
 simpl!(*);
 slf(*);
 simpl!(*);
 
 if !feature("seq") {
+  fork-tile[32, 0, false, false](loop1);
+  simpl!(loop1);
+  let split = fork-split(loop1);
+  simpl!(loop1);
+  clean-monoid-reduces(loop1);
+  let loop1_body = outline(split.srad_0.fj1);
+  simpl!(loop1, loop1_body);
+  unforkify(loop1_body);
+  let fission = fork-fission[split.srad_0.fj0](loop1);
+  simpl!(loop1, loop1_body);
+  unforkify(fission.srad_0.fj_bottom);
+  simpl!(loop1, loop1_body);
+  loop1 = loop1_body;
+
   fork-tile[32, 0, false, false](loop2);
   let split = fork-split(loop2);
   let loop2_body = outline(split.srad_1.fj1);