summaryrefslogtreecommitdiff
path: root/test/routines/level2/xger.hpp
diff options
context:
space:
mode:
Diffstat (limited to 'test/routines/level2/xger.hpp')
-rw-r--r--test/routines/level2/xger.hpp28
1 files changed, 19 insertions, 9 deletions
diff --git a/test/routines/level2/xger.hpp b/test/routines/level2/xger.hpp
index 92a1a2ae..9c5e2e40 100644
--- a/test/routines/level2/xger.hpp
+++ b/test/routines/level2/xger.hpp
@@ -77,15 +77,25 @@ class TestXger {
// 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 = Ger(args.layout,
- args.m, args.n, args.alpha,
- buffers.x_vec(), args.x_offset, args.x_inc,
- buffers.y_vec(), args.y_offset, args.y_inc,
- buffers.a_mat(), args.a_offset, args.a_ld,
- &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 = Ger(args.layout,
+ args.m, args.n, args.alpha,
+ buffers.x_vec(), args.x_offset, args.x_inc,
+ buffers.y_vec(), args.y_offset, args.y_inc,
+ buffers.a_mat(), args.a_offset, args.a_ld,
+ &queue_plain, &event);
+ if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
+ #elif CUDA_API
+ auto status = Ger(args.layout,
+ args.m, args.n, args.alpha,
+ buffers.x_vec(), args.x_offset, args.x_inc,
+ buffers.y_vec(), args.y_offset, args.y_inc,
+ buffers.a_mat(), args.a_offset, args.a_ld,
+ queue.GetContext()(), queue.GetDevice()());
+ cuStreamSynchronize(queue());
+ #endif
return status;
}