Merge branch 'master' into android_support

1 files changed, 17 insertions, 8 deletions

diff --git a/test/routines/levelx/xaxpybatched.hpp b/test/routines/levelx/xaxpybatched.hpp
index 4a8fc564..e9715f4e 100644
--- a/test/routines/levelx/xaxpybatched.hpp
+++ b/test/routines/levelx/xaxpybatched.hpp
@@ -83,14 +83,23 @@ class TestXaxpyBatched {
 
   // 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 = AxpyBatched(args.n, args.alphas.data(),
-                              buffers.x_vec(), args.x_offsets.data(), args.x_inc,
-                              buffers.y_vec(), args.y_offsets.data(), args.y_inc,
-                              args.batch_count,
-                              &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 = AxpyBatched(args.n, args.alphas.data(),
+                                buffers.x_vec(), args.x_offsets.data(), args.x_inc,
+                                buffers.y_vec(), args.y_offsets.data(), args.y_inc,
+                                args.batch_count,
+                                &queue_plain, &event);
+      if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
+    #elif CUDA_API
+      auto status = AxpyBatched(args.n, args.alphas.data(),
+                                buffers.x_vec(), args.x_offsets.data(), args.x_inc,
+                                buffers.y_vec(), args.y_offsets.data(), args.y_inc,
+                                args.batch_count,
+                                queue.GetContext()(), queue.GetDevice()());
+      cuStreamSynchronize(queue());
+    #endif
     return status;
   }