diff options
Diffstat (limited to 'test/routines/level3/xhemm.hpp')
-rw-r--r-- | test/routines/level3/xhemm.hpp | 28 |
1 files changed, 19 insertions, 9 deletions
diff --git a/test/routines/level3/xhemm.hpp b/test/routines/level3/xhemm.hpp index a89617b5..3b70d3f1 100644 --- a/test/routines/level3/xhemm.hpp +++ b/test/routines/level3/xhemm.hpp @@ -83,15 +83,25 @@ class TestXhemm { // 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 = Hemm(args.layout, args.side, args.triangle, - args.m, args.n, args.alpha, - buffers.a_mat(), args.a_offset, args.a_ld, - buffers.b_mat(), args.b_offset, args.b_ld, args.beta, - buffers.c_mat(), args.c_offset, args.c_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 = Hemm(args.layout, args.side, args.triangle, + args.m, args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_ld, args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } + #elif CUDA_API + auto status = Hemm(args.layout, args.side, args.triangle, + args.m, args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_ld, args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif return status; } |