summaryrefslogtreecommitdiff
path: root/test/routines/level3/xhemm.hpp
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2017-10-28 17:32:37 +0200
committerCedric Nugteren <web@cedricnugteren.nl>2017-10-28 17:32:37 +0200
commit12b08ae49154379f7471a40809ace6418857b387 (patch)
treeef958197db0bb8a67c9a5840f828b3f6c72bd8fc /test/routines/level3/xhemm.hpp
parent2949e156f5bfdd724987e67477da3e3608e4aaf9 (diff)
parentfa6e5e67f585b77d34c3031c176de9a0f7904aa9 (diff)
Merge branch 'master' into android_support
Diffstat (limited to 'test/routines/level3/xhemm.hpp')
-rw-r--r--test/routines/level3/xhemm.hpp28
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;
}