diff options
Diffstat (limited to 'test/routines/level1/xdot.hpp')
-rw-r--r-- | test/routines/level1/xdot.hpp | 25 |
1 files changed, 17 insertions, 8 deletions
diff --git a/test/routines/level1/xdot.hpp b/test/routines/level1/xdot.hpp index 1ea25994..229d18c9 100644 --- a/test/routines/level1/xdot.hpp +++ b/test/routines/level1/xdot.hpp @@ -73,14 +73,23 @@ class TestXdot { // 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 = Dot<T>(args.n, - buffers.scalar(), args.dot_offset, - 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 = Dot<T>(args.n, + buffers.scalar(), args.dot_offset, + 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 = Dot<T>(args.n, + buffers.scalar(), args.dot_offset, + 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; } |