slow GPU schedule

juno_samples/edge_detection/src/gpu.sch

-gvn(*);
-phi-elim(*);
-dce(*);
+macro simpl!(X) {
+  ccp(X);
+  simplify-cfg(X);
+  lift-dc-math(X);
+  gvn(X);
+  phi-elim(X);
+  dce(X);
+  infer-schedules(X);
+}
 
-gpu(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, max_gradient, reject_zero_crossings);
+simpl!(*);
 
 ip-sroa(*);
 sroa(*);
-dce(*);
-gvn(*);
-phi-elim(*);
-dce(*);
+simpl!(*);
 
-//forkify(*);
-infer-schedules(*);
+no-memset(gaussian_smoothing@res);
+fixpoint {
+  forkify(gaussian_smoothing);
+  fork-guard-elim(gaussian_smoothing);
+  fork-coalesce(gaussian_smoothing);
+}
+predication(gaussian_smoothing);
+simpl!(gaussian_smoothing);
+predication(gaussian_smoothing);
+simpl!(gaussian_smoothing);
 
-gcm(*);
+no-memset(laplacian_estimate@res, laplacian_estimate@shr1, laplacian_estimate@shr2);
+fixpoint {
+  forkify(laplacian_estimate);
+  fork-guard-elim(laplacian_estimate);
+  fork-coalesce(laplacian_estimate);
+}
+simpl!(laplacian_estimate);
+
+no-memset(zero_crossings@res, zero_crossings@shr1, zero_crossings@shr2);
+fixpoint {
+  forkify(zero_crossings);
+  fork-guard-elim(zero_crossings);
+  fork-coalesce(zero_crossings);
+}
+simpl!(zero_crossings);
+
+no-memset(gradient@res);
+fixpoint {
+  forkify(gradient);
+  fork-guard-elim(gradient);
+  fork-coalesce(gradient);
+}
+predication(gradient);
+simpl!(gradient);
+predication(gradient);
+simpl!(gradient);
+
+fixpoint {
+  forkify(max_gradient);
+  fork-guard-elim(max_gradient);
+  fork-coalesce(max_gradient);
+}
+simpl!(max_gradient);
+
+no-memset(reject_zero_crossings@res);
 fixpoint {
-  float-collections(*);
-  dce(*);
-  gcm(*);
+  forkify(reject_zero_crossings);
+  fork-guard-elim(reject_zero_crossings);
+  fork-coalesce(reject_zero_crossings);
 }
+predication(reject_zero_crossings);
+simpl!(reject_zero_crossings);
+
+async-call(edge_detection@le, edge_detection@zc);
+
+gpu(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, max_gradient, reject_zero_crossings);
+
+simpl!(*);
+
+delete-uncalled(*);
+gcm(*);
+