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/backprop.jn

@@ -91,15 +91,15 @@ fn backprop<input_n, hidden_n, output_n: usize>(
 ) -> f32, f32,
      f32[input_n + 1, hidden_n + 1], f32[input_n + 1, hidden_n + 1],
      f32[hidden_n + 1, output_n + 1], f32[hidden_n + 1, output_n + 1] {
-  let hidden_vals = layer_forward::<input_n, hidden_n>(input_vals, input_weights);
-  let output_vals = layer_forward::<hidden_n, output_n>(hidden_vals, hidden_weights);
+  @forward_input let hidden_vals = layer_forward::<input_n, hidden_n>(input_vals, input_weights);
+  @forward_hidden let output_vals = layer_forward::<hidden_n, output_n>(hidden_vals, hidden_weights);
 
   @output_error let out_err, out_delta = output_error::<output_n>(target, output_vals);
   @hidden_error let hid_err, hid_delta = hidden_error::<hidden_n, output_n>(out_delta, hidden_weights, hidden_vals);
 
-  let hidden_weights, hidden_prev_weights
+  @adjust_hidden let hidden_weights, hidden_prev_weights
     = adjust_weights::<hidden_n, output_n>(out_delta, hidden_vals, hidden_weights, hidden_prev_weights);
-  let input_weights, input_prev_weights
+  @adjust_input let input_weights, input_prev_weights
     = adjust_weights::<input_n, hidden_n>(hid_delta, input_vals, input_weights, input_prev_weights);
 
   return out_err, hid_err, input_weights, input_prev_weights, hidden_weights, hidden_prev_weights;