cfd is tricky

juno_samples/rodinia/cfd/src/cpu_euler.sch

@@ -24,7 +24,8 @@ fixpoint {
   fork-guard-elim(*);
 }
 simpl!(*);
+no-memset(compute_step_factor@res, compute_flux@res, copy_vars@res);
+parallel-reduce(time_step, copy_vars, compute_flux@outer_loop \ compute_flux@inner_loop);
 
 unforkify(*);
-
 gcm(*);

juno_samples/rodinia/cfd/src/euler.jn

@@ -47,7 +47,7 @@ fn compute_speed_of_sound(density: f32, pressure: f32) -> f32 {
 }
 
 fn compute_step_factor<nelr: usize>(variables: Variables::<nelr>, areas: f32[nelr]) -> f32[nelr] {
-  let step_factors : f32[nelr];
+  @res let step_factors : f32[nelr];
 
   for i in 0..nelr {
     let density = variables.density[i];
@@ -106,9 +106,9 @@ fn compute_flux<nelr: usize>(
   ff_flux_contribution_momentum_z: float3,
 ) -> Variables::<nelr> {
   const smoothing_coefficient : f32 = 0.2;
-  let fluxes: Variables::<nelr>;
+  @res let fluxes: Variables::<nelr>;
 
-  for i in 0..nelr {
+  @outer_loop for i in 0..nelr {
     let density_i = variables.density[i];
 
     let momentum_i = float3 { x: variables.momentum.x[i],
@@ -131,7 +131,7 @@ fn compute_flux<nelr: usize>(
     let flux_i_momentum = float3 { x: 0.0, y: 0.0, z: 0.0 };
     let flux_i_density_energy : f32 = 0.0;
 
-    for j in 0..NNB {
+    @inner_loop for j in 0..NNB {
       let nb = elements_surrounding_elements[j, i];
       let normal = float3 {
         x: normals.x[j, i],
@@ -249,7 +249,7 @@ fn time_step<nelr: usize>(
 }
 
 fn copy_vars<nelr: usize>(variables: Variables::<nelr>) -> Variables::<nelr> {
-  let result : Variables::<nelr>;
+  @res let result : Variables::<nelr>;
 
   for i in 0..nelr {
     result.density[i] = variables.density[i];