Fixes for the CUDA API: first tests pass and the client runs

2 files changed, 3 insertions, 2 deletions

diff --git a/src/kernels/opencl_to_cuda.h b/src/kernels/opencl_to_cuda.h
index fac30dfc..7602b539 100644
--- a/src/kernels/opencl_to_cuda.h
+++ b/src/kernels/opencl_to_cuda.h
@@ -32,8 +32,8 @@ __device__ int get_group_id(const int x) {
   return blockIdx.z;
 }
 __device__ int get_global_size(const int x) {
-  if (x == 0) { return gridDim.x; }
-  if (x == 1) { return gridDim.y; }
+  if (x == 0) { return gridDim.x * blockDim.x; }
+  if (x == 1) { return gridDim.y * blockDim.y; }
   return gridDim.z;
 }
 __device__ int get_global_id(const int x) {
diff --git a/test/routines/level1/xaxpy.hpp b/test/routines/level1/xaxpy.hpp
index cdceb4c7..7491a9e8 100644
--- a/test/routines/level1/xaxpy.hpp
+++ b/test/routines/level1/xaxpy.hpp
@@ -83,6 +83,7 @@ class TestXaxpy {
                          buffers.x_vec(), args.x_offset, args.x_inc,
                          buffers.y_vec(), args.y_offset, args.y_inc,
                          queue.GetContext()(), queue.GetDevice()());
+      cuStreamSynchronize(queue());
     #endif
     return status;
   }