Implement cudaDeviceSynchronize()

src/invert_trm/GPU/CUDA/elpa_invert_trm_cuda.cu

@@ -614,3 +614,7 @@ extern "C" void copy_float_complex_a_tmp1_FromC(float _Complex *a_dev, float _Co
     printf("Error in executing copy_float_complex_a_tmp1_kernel: %s\n",cudaGetErrorString(cuerr));
   }
 }
+
+extern "C" void device_synchronizeFromC() {
+   cudaDeviceSynchronize();
+} 

src/invert_trm/mod_invert_trm_cuda.F90

@@ -50,6 +50,12 @@ module invert_trm_cuda
   implicit none
 
   public
+  interface
+     subroutine device_synchronize_c()  &
+              bind(C, name="device_synchronizeFromC")
+     end subroutine
+  end interface
+
   interface
     subroutine copy_double_a_tmat2_c(a_dev, tmat2_dev, nblk, matrixRows, l_cols, l_colx, l_row1, nb)&
              bind(C, name="copy_double_a_tmat2_FromC")
@@ -492,6 +498,9 @@ module invert_trm_cuda
 #endif
 
     end subroutine
-
+  
+    subroutine device_synchronize() 
+       call device_synchronize_c()
+    end subroutine device_synchronize
 end module
 

[Andreas Marek]

This should not be necessary. We have device_synchronize already implemented in cudaFunctions.cu