diff options
Diffstat (limited to 'test/routines/level1/xnrm2.hpp')
-rw-r--r-- | test/routines/level1/xnrm2.hpp | 22 |
1 files changed, 15 insertions, 7 deletions
diff --git a/test/routines/level1/xnrm2.hpp b/test/routines/level1/xnrm2.hpp index 1db70537..f3a789b5 100644 --- a/test/routines/level1/xnrm2.hpp +++ b/test/routines/level1/xnrm2.hpp @@ -69,13 +69,21 @@ class TestXnrm2 { // 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 = Nrm2<T>(args.n, - buffers.scalar(), args.nrm2_offset, - buffers.x_vec(), args.x_offset, args.x_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 = Nrm2<T>(args.n, + buffers.scalar(), args.nrm2_offset, + buffers.x_vec(), args.x_offset, args.x_inc, + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } + #elif CUDA_API + auto status = Nrm2<T>(args.n, + buffers.scalar(), args.nrm2_offset, + buffers.x_vec(), args.x_offset, args.x_inc, + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif return status; } |