with Kokkos::fence

main.cpp

@@ -36,6 +36,7 @@ void scenario_1(double* array, int nblocks, int nlev, int nproma, bool print=tru
             d_view(jb, jk, jc) = p;
         }});
     
+    Kokkos::fence();
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);
@@ -62,7 +63,8 @@ void scenario_2(double* array, int nblocks, int nlev, int nproma, bool print=tru
 
           //  printf("%f ", d_view(jb, jk, jc));      
         }});
-
+    
+    Kokkos::fence();
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);
@@ -89,7 +91,8 @@ void scenario_2b(double* array, int nblocks, int nlev, int nproma, bool print=tr
 
           //  printf("%f ", d_view(jb, jk, jc));      
         }});
-
+    
+    Kokkos::fence();    
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);
@@ -116,7 +119,8 @@ void scenario_3(double* array, int nblocks, int nlev, int nproma, bool print=tru
 
           //  printf("%f ", d_view(jb, jk, jc));      
         }});
-
+    
+    Kokkos::fence();
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);
@@ -144,6 +148,7 @@ void scenario_4(double* array, int nblocks, int nlev, int nproma, bool print=tru
           //  printf("%f ", d_view(jb, jk, jc));      
         }});
 
+    Kokkos::fence();
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);
@@ -172,6 +177,7 @@ void scenario_5(double* array, int nblocks, int nlev, int nproma, bool print=tru
           //  printf("%f ", d_view(jb, jk, jc));      
         }});
 
+    Kokkos::fence();
     if(print)
         printf("Time = %f ms\n\n", timer.seconds() * 1000);
     Kokkos::deep_copy(view, d_view);

script.sh

@@ -12,7 +12,7 @@ then
     cmake --build build_gpu --parallel 
     ncells=(5000000)
     nlev=(90)
-    nproma=(10000 30000 50000 100000 1000000 5000000)
+    nproma=(5000000)
 else 
     cmake -B build -S . -DMU_ARCH=x86_64 -DCMAKE_CXX_FLAGS="-O3"
     cmake --build build --parallel