More optimizations

• A lot of misc. optimization work, notably parallelization of Rodinia CPU schedules.
• Support >32 tiles in GPU backend.
• Tweak some transformations.

juno_samples/rodinia/backprop/src/cpu.sch

@@ -12,7 +12,7 @@ simpl!(*);
 inline(layer_forward);
 delete-uncalled(*);
 
-no-memset(layer_forward@res);
+no-memset(layer_forward@res, output_error@res, hidden_error@res);
 lift-dc-math(*);
 loop-bound-canon(*);
 simpl!(*);
@@ -25,7 +25,42 @@ fixpoint {
 }
 reduce-slf(*);
 simpl!(*);
+fork-interchange[0, 1](adjust_weights);
+simpl!(*);
+
+infer-schedules(*);
+
+// The first call to layer_forward can be parallelized by 16 (the size of the
+// hidden layer) and the second can't be parallelized at all (the size of the
+// output layer is 1)
+inline(backprop@forward_input, backprop@forward_hidden);
+let forward_input = outline(backprop@forward_input);
+let forward_hidden = outline(backprop@forward_hidden);
+
+fork-tile[16, 0, false, true](forward_input@outer_loop \ forward_input@inner_loop);
+let (outer, inner) = fork-reshape[[1], [0]](forward_input@outer_loop \ forward_input@inner_loop);
+let forward_input = outline(inner);
+inline(backprop@forward_input);
+
+// The first call to adjust_weights has total loop dimensions of 1 * 17, so not
+// worth parallelizing (given that the body is trivial)
+// The second call to adjust_weights has a total dimension of 16 * (input + 1)
+// which is worth parallelizing, we'll do it by 16
+inline(backprop@adjust_hidden, backprop@adjust_input);
+let adjust_hidden = outline(backprop@adjust_hidden);
+let adjust_input = outline(backprop@adjust_input);
+
+fork-tile[16, 0, false, true](adjust_input);
+let (outer, inner) = fork-reshape[[1], [0, 2]](adjust_input);
+let adjust_input = outline(inner);
+inline(backprop@adjust_input);
 
+delete-uncalled(*);
+const-inline(*);
+
+simpl!(*);
 fork-split(*);
-unforkify(*);
+unforkify(output_error, hidden_error, adjust_hidden, adjust_input, forward_hidden, forward_input);
+simpl!(*);
+
 gcm(*);