Optimization for miranda

• Optimize cava and edge detection schedules (CPU schedules are multi-core now).
• Add fork-extend pass to make fork dimensions multiple of some number (closes #28 (closed)).
• Make GCM add extra dimensions to per-thread arrays in fork-joins.
• Make GCM allocate extra memory for calls in parallel, make RT backend use different memories per parallel call.
• Make not finding xdot binary crash the compiler.
• Some misc. fixes.
• Add conv-like fork-join test.

juno_samples/edge_detection/src/cpu.sch

@@ -24,6 +24,14 @@ predication(gaussian_smoothing);
 simpl!(gaussian_smoothing);
 predication(gaussian_smoothing);
 simpl!(gaussian_smoothing);
+let par = gaussian_smoothing@image_loop \ gaussian_smoothing@filter_loop;
+fork-tile[4, 1, false, false](par);
+fork-tile[4, 0, false, false](par);
+fork-interchange[1, 2](par);
+let split = fork-split(par);
+let gaussian_smoothing_body = outline(split._0_gaussian_smoothing.fj2);
+fork-coalesce(gaussian_smoothing, gaussian_smoothing_body);
+simpl!(gaussian_smoothing, gaussian_smoothing_body);
 
 no-memset(laplacian_estimate@res, laplacian_estimate@shr1, laplacian_estimate@shr2);
 fixpoint {
@@ -32,6 +40,15 @@ fixpoint {
   fork-coalesce(laplacian_estimate);
 }
 simpl!(laplacian_estimate);
+let par = laplacian_estimate@image_loop \ laplacian_estimate@filter_loop;
+fork-tile[4, 1, false, false](par);
+fork-tile[4, 0, false, false](par);
+fork-interchange[1, 2](par);
+let split = fork-split(par);
+let body = split._1_laplacian_estimate.fj2 | laplacian_estimate.shr1 | laplacian_estimate.shr2;
+let laplacian_estimate_body = outline(body);
+fork-coalesce(laplacian_estimate, laplacian_estimate_body);
+simpl!(laplacian_estimate, laplacian_estimate_body);
 
 no-memset(zero_crossings@res, zero_crossings@shr1, zero_crossings@shr2);
 fixpoint {
@@ -40,6 +57,15 @@ fixpoint {
   fork-coalesce(zero_crossings);
 }
 simpl!(zero_crossings);
+let par = zero_crossings@image_loop \ zero_crossings@filter_loop;
+fork-tile[4, 1, false, false](par);
+fork-tile[4, 0, false, false](par);
+fork-interchange[1, 2](par);
+let split = fork-split(par);
+let body = split._2_zero_crossings.fj2 | zero_crossings.shr1 | zero_crossings.shr2;
+let zero_crossings_body = outline(body);
+fork-coalesce(zero_crossings, zero_crossings_body);
+simpl!(zero_crossings, zero_crossings_body);
 
 no-memset(gradient@res);
 fixpoint {
@@ -81,8 +107,8 @@ simpl!(reject_zero_crossings);
 
 async-call(edge_detection@le, edge_detection@zc);
 
-fork-split(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, reject_zero_crossings);
-unforkify(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, reject_zero_crossings);
+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);
 
 simpl!(*);