diff options
Diffstat (limited to 'test/routines/level2/xtbmv.hpp')
-rw-r--r-- | test/routines/level2/xtbmv.hpp | 26 |
1 files changed, 18 insertions, 8 deletions
diff --git a/test/routines/level2/xtbmv.hpp b/test/routines/level2/xtbmv.hpp index 8c7aa381..a80d9e26 100644 --- a/test/routines/level2/xtbmv.hpp +++ b/test/routines/level2/xtbmv.hpp @@ -70,14 +70,24 @@ class TestXtbmv { // 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 = Tbmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, - args.n, args.kl, - buffers.a_mat(), args.a_offset, args.a_ld, - 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 = Tbmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, + args.n, args.kl, + buffers.a_mat(), args.a_offset, args.a_ld, + 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 = Tbmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, + args.n, args.kl, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, + &queue_plain, &event); + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif return status; } |