need to do the data copying after the data generation. Required for later tests

fftw.c

@@ -75,8 +75,6 @@ run_test
         __COMPLEX8__ * hA;
 	__MALLOC(hA, sizeof(*hA) * N * T);
 
-        //  Copy data to device
-        __ASSERT(__PREFIX(Memcpy)(A, hA, sizeof(*hA) * N, __PREFIX(MemcpyHostToDevice)));
         // Create FFT plan
 	__FFTW_PLAN plan;
         size_t length = N;
@@ -121,20 +119,21 @@ run_test
         );
 #endif
 	prepare_data(hA, N * T);
-	for (int r = 0 ; r < 10; r++)
+        //  Copy data to device
+        __ASSERT(__PREFIX(Memcpy)(A, hA, sizeof(*hA) * N, __PREFIX(MemcpyHostToDevice)));
+	sprintf(mes, "T = %zu n = %zu\t 1nd total %d rounds." ,T, N, rep );
+	timer_start(timer, mes);
+	for (int r = 0 ; r < rep; r++)
 	{
 		// Execute plan
-		sprintf(mes, "T = %zu n = %zu\t round %d." ,T, N , r);
-		timer_start(timer, mes);
 #ifdef ROC
 		rocfft_execute(plan, (void**) &A, NULL, NULL);
 #elif CUDA
 		cufftExecC2C(plan, A, A, CUFFT_FORWARD);
 #endif
 		__PREFIX(DeviceSynchronize)();
-		timer_stop(timer);
-
 	}
+	timer_stop(timer);
         // Destroy plan
         __DESTROY_PLAN(plan);
 
@@ -152,7 +151,7 @@ main
 (
 )
 {
-	int rep = 1;
+	int rep = 1000;
 	int t_min = 8;
 	int t_max = 11;
 	int n_min = 11;