Added timings

examples/hasegawa-wakatani-3d/data/BOUT.inp

@@ -2,19 +2,23 @@
 # Hasegawa-Wakatani test case
 #
 
-timestep = 1. # Output timestep
-nout = 100 # Number of output steps
+timestep = 0.000001 # Output timestep
+nout = 1 # Number of output steps
 #nout = 1 # Number of output steps
 
 [mesh]
+Lx = 51.2
+Ly = 16.0
+Lz = 51.2
 
-nx = 68 # Note 4 guard cells in X
-ny = 16 
-nz = 64 # Periodic, so no guard cells in Z
+nx = 36 # Note 4 guard cells in X
+ny = 16
+nz = 32 # Periodic, so no guard cells in Z
+
+dx = Lx/(nx-4)
+dy = Ly/ny
+dz = Lz/nz
 
-dx = 0.2
-dy = 1.
-dz = 0.2
 
 # Set periodic boundary condition in y
 ixseps1 = nx

examples/hasegawa-wakatani-3d/hw.cxx

@@ -512,7 +512,7 @@ class HW3D : public PhysicsModel {
  double deltaz;
  double psimtime = 0.0;
 
- bool profile_StylES = false;
+ bool profile_StylES = true;
  bool implicitStylES = false;
 
  Field3D pPhiVort;
@@ -602,15 +602,36 @@ class HW3D : public PhysicsModel {
 
  double *rLES;
 
+
+ if (profile_StylES){
+ timeStart = high_resolution_clock::now();
+ }
+
  // Solve for potential
  if (pStep>0){
  phi = phiSolver->solve(vort, phi);
  }
 
+ if (profile_StylES)
+ {
+ timeStop = high_resolution_clock::now();
+ auto timeDiff = duration_cast<microseconds>(timeStop - timeStart);
+ printf("BOUT++ timing solve potential %f\n", timeDiff.count()/1.0e6);
+ timeStart = high_resolution_clock::now();
+ }
+
  Field3D phi_minus_n = phi - n;
  // Communicate variables
  mesh->communicate(n, vort, phi, phi_minus_n);
 
+ if (profile_StylES)
+ {
+ timeStop = high_resolution_clock::now();
+ auto timeDiff = duration_cast<microseconds>(timeStop - timeStart);
+ printf("BOUT++ timing communicate %f\n", timeDiff.count()/1.0e6);
+ timeStart = high_resolution_clock::now();
+ }
+
  // Create accessors which enable fast access
  auto n_acc = FieldAccessor<>(n);
  auto vort_acc = FieldAccessor<>(vort);
@@ -631,13 +652,30 @@ class HW3D : public PhysicsModel {
  int LES_it = int(rLES[0]);
  int cont=1;
 
+ if (profile_StylES)
+ {
+ timeStop = high_resolution_clock::now();
+ auto timeDiff = duration_cast<microseconds>(timeStop - timeStart);
+ printf("BOUT++ timing findLESTerms %f\n", timeDiff.count()/1.0e6);
+ timeStart = high_resolution_clock::now();
+ }
+
  for(int i=2; i<n.getNx()-2; i++) // we assume 2 guards cells in x-direction
  for(int j=2; j<n.getNy()-2; j++)
  for(int k=0; k<n.getNz(); k++){
  pPhiVort(i,j,k) = rLES[cont + 0*SIZE];
  pPhiN(i,j,k) = rLES[cont + 1*SIZE];
  cont = cont+1;
  }
+
+ if (profile_StylES)
+ {
+ timeStop = high_resolution_clock::now();
+ auto timeDiff = duration_cast<microseconds>(timeStop - timeStart);
+ printf("BOUT++ timing passing to CPU %f\n", timeDiff.count()/1.0e6);
+ timeStart = high_resolution_clock::now();
+ }
+
  }
  else{
  BOUT_FOR_RAJA(i, n.getRegion("RGN_NOBNDRY"), CAPTURE(alpha, kappa, Dn, Dvort)) {
@@ -656,6 +694,14 @@ class HW3D : public PhysicsModel {
  ddt(vort_acc)[i] = -pPhiVort[i] - div_current + Dvort * Delp2(vort_acc, i);
  }
 
+ if (profile_StylES)
+ {
+ timeStop = high_resolution_clock::now();
+ auto timeDiff = duration_cast<microseconds>(timeStop - timeStart);
+ printf("BOUT++ timing integrate %f\n\n", timeDiff.count()/1.0e6);
+ timeStart = high_resolution_clock::now();
+ }
+
  pStep++;
  return 0;
  }