Ensured that acBoundcondStep is called everywhere in the program before acStore

src/standalone/benchmark.cc

@@ -197,7 +197,6 @@ run_benchmark(void)
                double(results[int(NTH_PERCENTILE * NUM_ITERS)]), int(NTH_PERCENTILE * 100),
                mesh_info.int_params[AC_nx]);
 
-        acStore(mesh);
         acQuit();
         acmesh_destroy(mesh);
     }
@@ -269,7 +268,6 @@ run_benchmark(void)
     write_result_to_file(wallclock * 1e3f / steps);
     #endif
 
-    acStore(mesh);
     acQuit();
     acmesh_destroy(mesh);
 

src/standalone/renderer.cc

@@ -32,13 +32,13 @@
 #include <string.h> // memcpy
 
 #include "config_loader.h"
-#include "src/core/errchk.h"
-#include "src/core/math_utils.h"
 #include "model/host_forcing.h"
 #include "model/host_memory.h"
 #include "model/host_timestep.h"
 #include "model/model_reduce.h"
 #include "model/model_rk3.h"
+#include "src/core/errchk.h"
+#include "src/core/math_utils.h"
 #include "timer_hires.h"
 
 // Window
@@ -410,9 +410,10 @@ run_renderer(void)
         /* Render */
         const float timer_diff_sec = timer_diff_nsec(frame_timer) / 1e9f;
         if (timer_diff_sec >= desired_frame_time) {
-            // acStore(mesh);
             const int num_vertices = mesh->info.int_params[AC_mxy];
             const int3 dst         = (int3){0, 0, k_slice};
+            acBoundcondStep();
+            // acStore(mesh);
             acStoreWithOffset(dst, num_vertices, mesh);
             acSynchronizeStream(STREAM_ALL);
             renderer_draw(*mesh); // Bottleneck is here
@@ -422,7 +423,6 @@ run_renderer(void)
     }
     printf("Wallclock time %f s\n", double(timer_diff_nsec(wallclock) / 1e9f));
 
-    acStore(mesh);
     acQuit();
     acmesh_destroy(mesh);
 

src/standalone/simulation.cc

@@ -27,13 +27,13 @@
 #include "run.h"
 
 #include "config_loader.h"
-#include "src/core/errchk.h"
-#include "src/core/math_utils.h"
 #include "model/host_forcing.h"
 #include "model/host_memory.h"
 #include "model/host_timestep.h"
 #include "model/model_reduce.h"
 #include "model/model_rk3.h"
+#include "src/core/errchk.h"
+#include "src/core/math_utils.h"
 #include "timer_hires.h"
 
 #include <string.h>
@@ -205,6 +205,7 @@ run_simulation(void)
     write_mesh_info(&mesh_info);
     print_diagnostics(0, AcReal(.0), t_step, diag_file);
 
+    acBoundcondStep();
     acStore(mesh);
     save_mesh(*mesh, 0, t_step);
 
@@ -298,9 +299,10 @@ run_simulation(void)
                 assumed to be asynchronous, so the meshes must be also synchronized
                 before transferring the data to the CPU. Like so:
 
-                acSynchronize();
+                acBoundcondStep();
                 acStore(mesh);
             */
+            acBoundcondStep();
             acStore(mesh);
 
             save_mesh(*mesh, i, t_step);