Conditional schedules

• Add if statements to the scheduling language.
• Add a feature(...) function that can be used to test for whether a feature is enabled
  • For this to work, the scheduler needs to be run from a Cargo build script since it relies on testing environment variables (as described here).
• Modified matmul to use a single schedule testing for the "cuda" feature

juno_samples/matmul/src/matmul.sch

@@ -41,21 +41,41 @@ macro unforkify!(X) {
 
 optimize!(*);
 forkify!(*);
-associative(matmul@outer);
 
-// Parallelize by computing output array as 16 chunks
-let par = matmul@outer \ matmul@inner;
-fork-tile![4](par);
-let (outer, inner, _) = fork-reshape[[1, 3], [0], [2]](par);
-parallelize!(outer \ inner);
+if feature("cuda") {
+  fixpoint {
+    reduce-slf(*);
+    slf(*);
+    infer-schedules(*);
+  }
+  fork-coalesce(*);
+  infer-schedules(*);
+  dce(*);
+  rewrite(*);
+  fixpoint {
+    simplify-cfg(*);
+    dce(*);
+  }
 
-let body = outline(inner);
-cpu(body);
+  optimize!(*);
+  codegen-prep!(*);
+} else {
+  associative(matmul@outer);
 
-// Tile for cache, assuming 64B cache lines
-fork-tile![16](body);
-let (outer, inner) = fork-reshape[[0, 2, 4, 1, 3], [5]](body);
+  // Parallelize by computing output array as 16 chunks
+  let par = matmul@outer \ matmul@inner;
+  fork-tile![4](par);
+  let (outer, inner, _) = fork-reshape[[1, 3], [0], [2]](par);
+  parallelize!(outer \ inner);
 
-reduce-slf(inner);
-unforkify!(body);
-codegen-prep!(*);
+  let body = outline(inner);
+  cpu(body);
+
+  // Tile for cache, assuming 64B cache lines
+  fork-tile![16](body);
+  let (outer, inner) = fork-reshape[[0, 2, 4, 1, 3], [5]](body);
+
+  reduce-slf(inner);
+  unforkify!(body);
+  codegen-prep!(*);
+}