Tests on GPU stacking

timing.h

@@ -72,7 +72,8 @@ typedef struct {
   double compose;    //
   double phase;      //
   double write;      //
-  double total; } timing_t;
+  double total;
+  double offload;} timing_t;
 
 extern timing_t timing_wt;      // wall-clock process timing, at Task 0
 

w-stacking.cu

@@ -420,8 +420,9 @@ void wstack(
         }
 	
     }
-
+   #ifdef ACCOMP
    #pragma omp target exit data map(delete:uu[0:num_points], vv[0:num_points], ww[0:num_points], vis_real[0:Nvis], vis_img[0:Nvis], weight[0:Nvis/freq_per_chan], grid[0:2*num_w_planes*grid_size_x*grid_size_y])
+   #endif
     // End switch between CUDA and CPU versions
 #endif
     //for (int i=0; i<100000; i++)printf("%f\n",grid[i]);

w-stacking.h

@@ -82,6 +82,9 @@ void phase_correction(
      int);
 
 
+#ifdef ACCOMP
+#pragma omp declare target (gauss_kernel_norm)
+#endif
 
 #ifdef __CUDACC__
 extern "C"