diff options
Diffstat (limited to 'test/routines/level2/xhpr2.hpp')
-rw-r--r-- | test/routines/level2/xhpr2.hpp | 28 |
1 files changed, 19 insertions, 9 deletions
diff --git a/test/routines/level2/xhpr2.hpp b/test/routines/level2/xhpr2.hpp index f946ba5c..1aa6cc54 100644 --- a/test/routines/level2/xhpr2.hpp +++ b/test/routines/level2/xhpr2.hpp @@ -75,15 +75,25 @@ class TestXhpr2 { // 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 = Hpr2(args.layout, args.triangle, - args.n, args.alpha, - buffers.x_vec(), args.x_offset, args.x_inc, - buffers.y_vec(), args.y_offset, args.y_inc, - buffers.ap_mat(), args.ap_offset, - &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 = Hpr2(args.layout, args.triangle, + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } + #elif CUDA_API + auto status = Hpr2(args.layout, args.triangle, + args.n, args.alpha, + buffers.x_vec(), args.x_offset, args.x_inc, + buffers.y_vec(), args.y_offset, args.y_inc, + buffers.ap_mat(), args.ap_offset, + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif return status; } |