Already mentioned here ( https://gcc.gnu.org/ml/gcc-patches/2016-11/msg00972.html ):
...
With OpenMP/PTX offloading there are 5 additional failures in check-target-libgomp:

<SNIP>

Two with 'target nowait' (not implemented)
FAIL: libgomp.c/target-33.c execution test
FAIL: libgomp.c/target-34.c execution test
...

These tests pass on gomp-4_0-branch, because there GOMP_OFFLOAD_get_caps does not return GOMP_OFFLOAD_CAP_OPENMP_400, while on trunk it does.

Created attachment 42484 [details]
Patch that moves async_run from plugin-hsa.c to default_sync_run

I wonder if this will work.

(In reply to Tom de Vries from comment #3)
> Created attachment 42484 [details]
> Patch that moves async_run from plugin-hsa.c to default_sync_run
> 
> I wonder if this will work.

That got me this error for target-33.c:
...
libgomp: 
libgomp: cuMemAlloc error: invalid device context
cuMemAlloc error: invalid device context

libgomp: cuCtxPushCurrent_v2 error: unknown cuda error

libgomp: device finalization failed
...

Using this patch in addition, both target-33.c and target34.c pass:
...
diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c
index d5262639578..4e0009f650e 100644
--- a/libgomp/plugin/plugin-nvptx.c
+++ b/libgomp/plugin/plugin-nvptx.c
@@ -2127,6 +2127,8 @@ GOMP_OFFLOAD_run (int ord, void *tgt_fn, void *tgt_vars, void **args)
   const char *maybe_abort_msg = "(perhaps abort was called)";
   int teams = 0, threads = 0;
 
+  nvptx_attach_host_thread_to_device (ord);
+
   if (!args)
     GOMP_PLUGIN_fatal ("No target arguments provided");
   while (*args)
...

patch submitted: https://gcc.gnu.org/ml/gcc-patches/2017-10/msg02079.html

This no longer fails at current trunk.