Merge pull request #204 from CNugteren/cuda_api

Cuda API to CLBlast

1 files changed, 15 insertions, 7 deletions

diff --git a/test/routines/level1/xcopy.hpp b/test/routines/level1/xcopy.hpp
index 7a5c99b8..58abdbf4 100644
--- a/test/routines/level1/xcopy.hpp
+++ b/test/routines/level1/xcopy.hpp
@@ -69,13 +69,21 @@ class TestXcopy {
 
   // Describes how to run the CLBlast routine
   static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-    auto queue_plain = queue();
-    auto event = cl_event{};
-    auto status = Copy<T>(args.n,
-                          buffers.x_vec(), args.x_offset, args.x_inc,
-                          buffers.y_vec(), args.y_offset, args.y_inc,
-                          &queue_plain, &event);
-    if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
+    #ifdef OPENCL_API
+      auto queue_plain = queue();
+      auto event = cl_event{};
+      auto status = Copy<T>(args.n,
+                            buffers.x_vec(), args.x_offset, args.x_inc,
+                            buffers.y_vec(), args.y_offset, args.y_inc,
+                            &queue_plain, &event);
+      if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
+    #elif CUDA_API
+      auto status = Copy<T>(args.n,
+                            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;
   }