added sync after asynch copies in krnls 3 & 4 & grad & fire, #31

device/kernel1.cl

@@ -92,7 +92,7 @@ gpu_calc_initpop(
 	}
 
 	// Asynchronous copy should be finished by here
-	wait_group_events(1,&ev);
+	wait_group_events(1, &ev);
 
 	// Evaluating initial genotypes
 	barrier(CLK_LOCAL_MEM_FENCE);

device/kernel3.cl

@@ -152,7 +152,7 @@ perform_LS(
 
 
 	// Asynchronous copy should be finished by here
-	wait_group_events(1,&ev);
+	wait_group_events(1, &ev);
 
 	barrier(CLK_LOCAL_MEM_FENCE);
 
@@ -437,7 +437,10 @@ perform_LS(
 	// Updating old offspring in population
 	barrier(CLK_LOCAL_MEM_FENCE);
 
-  	async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
-        	              offspring_genotype,
-        	              dockpars_num_of_genes,0);
+  	event_t ev2 = async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
+        	                            offspring_genotype,
+        	                            dockpars_num_of_genes,0);
+
+	// Asynchronous copy should be finished by here
+	wait_group_events(1, &ev2);
 }

device/kernel4.cl

@@ -219,7 +219,7 @@ gpu_gen_and_eval_newpops(
 					                   dockpars_num_of_genes, 0);
 
 			// Asynchronous copy should be finished by here
-			wait_group_events(1,&ev);
+			wait_group_events(1, &ev);
 		} // End of crossover
 
 		barrier(CLK_LOCAL_MEM_FENCE);
@@ -316,8 +316,11 @@ gpu_gen_and_eval_newpops(
 		}
 
 		// Copying new offspring to next generation
-		async_work_group_copy(dockpars_conformations_next + GENOTYPE_LENGTH_IN_GLOBMEM*get_group_id(0),
-				      offspring_genotype,
-				      dockpars_num_of_genes, 0);
+		event_t ev2 = async_work_group_copy(dockpars_conformations_next + GENOTYPE_LENGTH_IN_GLOBMEM*get_group_id(0),
+				                    offspring_genotype,
+				                    dockpars_num_of_genes, 0);
+
+		// Asynchronous copy should be finished by here
+		wait_group_events(1, &ev2);
   }
 }

device/kernel_fire.cl

@@ -123,7 +123,7 @@ gradient_minFire(
                               		   dockpars_num_of_genes, 0);
 
 	// Asynchronous copy should be finished by here
-	wait_group_events(1,&ev);
+	wait_group_events(1, &ev);
 
   	// -----------------------------------------------------------------------------
   	// Some OpenCL compilers don't allow declaring 
@@ -685,7 +685,10 @@ gradient_minFire(
 	// Updating old offspring in population
 	barrier(CLK_LOCAL_MEM_FENCE);
 
-	async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
-			      genotype,
-			      dockpars_num_of_genes, 0);
+	event_t ev2 = async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
+			                    genotype,
+			                    dockpars_num_of_genes, 0);
+
+	// Asynchronous copy should be finished by here
+	wait_group_events(1, &ev2);
 }

device/kernel_gradient.cl

@@ -139,7 +139,7 @@ gradient_minimizer(
                               		   dockpars_num_of_genes, 0);
 
 	// Asynchronous copy should be finished by here
-	wait_group_events(1,&ev);
+	wait_group_events(1, &ev);
 
   	// -----------------------------------------------------------------------------
 	// -----------------------------------------------------------------------------
@@ -814,7 +814,10 @@ gradient_minimizer(
 	// Updating old offspring in population
 	barrier(CLK_LOCAL_MEM_FENCE);
 
-	async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
-			      genotype,
-			      dockpars_num_of_genes, 0);
+	event_t ev2 = async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
+			                    genotype,
+			                    dockpars_num_of_genes, 0);
+
+	// Asynchronous copy should be finished by here
+	wait_group_events(1, &ev2);
 }