diff mbox series

[5/6,OpenACC,libgomp] Async re-work, C/C++ testsuite changes

Message ID 8086c63b-f729-891b-3d21-76871d360734@mentor.com
State New
Headers show
Series Async re-work | expand

Commit Message

Chung-Lin Tang Sept. 25, 2018, 1:11 p.m. UTC
These are the testsuite/libgomp.oacc-c-c++-common/* changes.

Thanks,
Chung-Lin

	* testsuite/libgomp.oacc-c-c++-common/asyncwait-2.c: New testcase.
	* testsuite/libgomp.oacc-c-c++-common/data-2-lib.c: Adjust testcase.
	* testsuite/libgomp.oacc-c-c++-common/data-2.c: Likewise.
	* testsuite/libgomp.oacc-c-c++-common/data-3.c: Likewise.
	* testsuite/libgomp.oacc-c-c++-common/lib-71.c: Likewise.
	* testsuite/libgomp.oacc-c-c++-common/lib-77.c: Likewise.
	* testsuite/libgomp.oacc-c-c++-common/lib-79.c: Likewise.
	* testsuite/libgomp.oacc-c-c++-common/lib-81.c: Likewise.

Comments

Thomas Schwinge Dec. 7, 2018, 3:30 p.m. UTC | #1
Hi Chung-Lin!

On Tue, 25 Sep 2018 21:11:42 +0800, Chung-Lin Tang <chunglin_tang@mentor.com> wrote:
> These are the testsuite/libgomp.oacc-c-c++-common/* changes.

Please commit the following three hunks to trunk: the code as present
doesn't declare its async/wait dependencies correctly.  To record the
review effort, please include "Reviewed-by: Thomas Schwinge
<thomas@codesourcery.com>" in the commit log, see
<https://gcc.gnu.org/wiki/Reviewed-by>.

> diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
> index 2ddfa7d..f553d3d 100644
> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
> @@ -153,7 +153,7 @@ main (int argc, char **argv)
>      d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
>  
>  #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N], N) \
> -  async (4)
> +  wait (1, 2, 3) async (4)
>    for (int ii = 0; ii < N; ii++)
>      e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
>  
> diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
> index 0c6abe6..81d623a 100644
> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
> @@ -162,7 +162,7 @@ main (int argc, char **argv)
>      d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
>  
>  #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
> -  wait (1) async (4)
> +  wait (1, 2, 3) async (4)
>    for (int ii = 0; ii < N; ii++)
>      e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
>  
> diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
> index 0bf706a..5ec50b8 100644
> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
> @@ -138,7 +138,7 @@ main (int argc, char **argv)
>      d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
>  
>  #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
> -  wait (1,5) async (4)
> +  wait (1, 2, 3, 5) async (4)
>    for (int ii = 0; ii < N; ii++)
>      e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
>  


Grüße
 Thomas
Thomas Schwinge Dec. 7, 2018, 3:56 p.m. UTC | #2
Hi Chung-Lin!

On Tue, 25 Sep 2018 21:11:42 +0800, Chung-Lin Tang <chunglin_tang@mentor.com> wrote:
> These are the testsuite/libgomp.oacc-c-c++-common/* changes.

> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
> @@ -114,6 +114,7 @@ main (int argc, char **argv)
>  
>    for (i = 0; i < N; i++)
>      {
> +      stream = (CUstream) acc_get_cuda_stream (i & 1);
>        r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, stream, kargs, 0);

What's the motivation for this change?

And then:

> @@ -122,11 +123,11 @@ main (int argc, char **argv)
>  	}
>      }
>  
> -  acc_wait_async (0, 1);
> -
>    if (acc_async_test (0) != 0)
>      abort ();
>  
> +  acc_wait_async (0, 1);
> +
>    if (acc_async_test (1) != 0)
>      abort ();

I somehow feel that this change...

> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c
> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c
> @@ -133,7 +133,7 @@ main (int argc, char **argv)
>  
>    for (i = 0; i <= N; i++)
>      {
> -      if (acc_async_test (i) != 0)
> +      if (acc_async_test (i) == 0)
>  	abort ();
>      }

..., and this change are needed because we're now more strictly
synchronizing with the local (host) thread.

Regarding the case of "libgomp.oacc-c-c++-common/lib-81.c", as currently
present:

    [...]
      for (i = 0; i < N; i++)
        {
          r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, streams[i], kargs, 0);
          if (r != CUDA_SUCCESS)
            {
              fprintf (stderr, "cuLaunchKernel failed: %d\n", r);
              abort ();
            }
        }

This launches N kernels on N separate async queues/CUDA streams, [0..N).

      acc_wait_all_async (N);

Then, the "acc_wait_all_async (N)" -- in my understanding! -- should
*not* synchronize with the local (host) thread, but instead just set up
the additional async queue/CUDA stream N to "depend" on [0..N).

      for (i = 0; i <= N; i++)
        {
          if (acc_async_test (i) != 0)
            abort ();
        }

Thus, all [0..N) should then still be "acc_async_test (i) != 0" (still
running).

      acc_wait (N);

Here, the "acc_wait (N)" would synchronize the local (host) thread with
async queue/CUDA stream N and thus recursively with [0..N).

      for (i = 0; i <= N; i++)
        {
          if (acc_async_test (i) != 1)
            abort ();
        }
    [...]

So, then all these async queues/CUDA streams here indeed are
"acc_async_test (i) != 1", thas is, idle.


Now, the more strict synchronization with the local (host) thread is not
wrong in term of correctness, but I suppose it will impact performance of
otherwise asynchronous operations, which now get synchronized too much?

Or, of course, I'm misunderstanding something...

(For avoidance of doubt, I would accept the "async re-work" as is, but we
should eventually clarify this, and restore the behavior we -- apparently
-- had before, where we didn't synchronize so much?  (So, technically,
the "async re-work" would constitute a regression for this kind of
usage?)


Grüße
 Thomas
Chung-Lin Tang Dec. 11, 2018, 1:30 p.m. UTC | #3
On 2018/12/7 11:56 PM, Thomas Schwinge wrote:
>> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
>> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
>> @@ -114,6 +114,7 @@ main (int argc, char **argv)
>>   
>>     for (i = 0; i < N; i++)
>>       {
>> +      stream = (CUstream) acc_get_cuda_stream (i & 1);
>>         r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, stream, kargs, 0);
> What's the motivation for this change?

To place work on both streams 0 and 1.

> ..., and this change are needed because we're now more strictly
> synchronizing with the local (host) thread.
> 
> Regarding the case of "libgomp.oacc-c-c++-common/lib-81.c", as currently
> present:
> 
>      [...]
>        for (i = 0; i < N; i++)
>          {
>            r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, streams[i], kargs, 0);
>            if (r != CUDA_SUCCESS)
>              {
>                fprintf (stderr, "cuLaunchKernel failed: %d\n", r);
>                abort ();
>              }
>          }
> 
> This launches N kernels on N separate async queues/CUDA streams, [0..N).
> 
>        acc_wait_all_async (N);
> 
> Then, the "acc_wait_all_async (N)" -- in my understanding! -- should
> *not*  synchronize with the local (host) thread, but instead just set up
> the additional async queue/CUDA stream N to "depend" on [0..N).
> 
>        for (i = 0; i <= N; i++)
>          {
>            if (acc_async_test (i) != 0)
>              abort ();
>          }
> 
> Thus, all [0..N) should then still be "acc_async_test (i) != 0" (still
> running).
> 
>        acc_wait (N);
> 
> Here, the "acc_wait (N)" would synchronize the local (host) thread with
> async queue/CUDA stream N and thus recursively with [0..N).
> 
>        for (i = 0; i <= N; i++)
>          {
>            if (acc_async_test (i) != 1)
>              abort ();
>          }
>      [...]
> 
> So, then all these async queues/CUDA streams here indeed are
> "acc_async_test (i) != 1", thas is, idle.
> 
> 
> Now, the more strict synchronization with the local (host) thread is not
> wrong in term of correctness, but I suppose it will impact performance of
> otherwise asynchronous operations, which now get synchronized too much?
> 
> Or, of course, I'm misunderstanding something...

IIRC, we encountered many issues where people misunderstood the meaning of "wait+async",
using it as if the local host sync happened, where in our original implementation it does not.

Also some areas of the OpenACC spec were vague on whether the local host synchronization should
or should not happen; basically, the wording treated as if it was only an implementation detail
and didn't matter, and didn't acknowledge that this would be something visible to the user.

At the end, IIRC, I decided that adding a local host synchronization is easier for all of us,
and took the opportunity of the re-org to make this change.

That said, I didn't notice those tests you listed above were meant to test such delicate behavior.

> (For avoidance of doubt, I would accept the "async re-work" as is, but we
> should eventually clarify this, and restore the behavior we -- apparently
> -- had before, where we didn't synchronize so much?  (So, technically,
> the "async re-work" would constitute a regression for this kind of
> usage?)

It's not hard to restore the old behavior, just a few lines to delete. Although as described
above, this change was deliberate.

This might be another issue to raise with the committee. I think I tried on this exact issue
a long time ago, but never got answers.

Thanks,
Chung-Lin
Thomas Schwinge Dec. 14, 2018, 2:50 p.m. UTC | #4
Hi Chung-Lin!

On Tue, 11 Dec 2018 21:30:31 +0800, Chung-Lin Tang <chunglin_tang@mentor.com> wrote:
> On 2018/12/7 11:56 PM, Thomas Schwinge wrote:
> >> --- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
> >> +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
> >> @@ -114,6 +114,7 @@ main (int argc, char **argv)
> >>   
> >>     for (i = 0; i < N; i++)
> >>       {
> >> +      stream = (CUstream) acc_get_cuda_stream (i & 1);
> >>         r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, stream, kargs, 0);
> > What's the motivation for this change?
> 
> To place work on both streams 0 and 1.

That's describing what it doesn, not the motivation behind it.  ;-)


> > ..., and this change are needed because we're now more strictly
> > synchronizing with the local (host) thread.
> > 
> > Regarding the case of "libgomp.oacc-c-c++-common/lib-81.c", as currently
> > present:
> > 
> >      [...]
> >        for (i = 0; i < N; i++)
> >          {
> >            r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, streams[i], kargs, 0);
> >            if (r != CUDA_SUCCESS)
> >              {
> >                fprintf (stderr, "cuLaunchKernel failed: %d\n", r);
> >                abort ();
> >              }
> >          }
> > 
> > This launches N kernels on N separate async queues/CUDA streams, [0..N).
> > 
> >        acc_wait_all_async (N);
> > 
> > Then, the "acc_wait_all_async (N)" -- in my understanding! -- should
> > *not*  synchronize with the local (host) thread, but instead just set up
> > the additional async queue/CUDA stream N to "depend" on [0..N).
> > 
> >        for (i = 0; i <= N; i++)
> >          {
> >            if (acc_async_test (i) != 0)
> >              abort ();
> >          }
> > 
> > Thus, all [0..N) should then still be "acc_async_test (i) != 0" (still
> > running).
> > 
> >        acc_wait (N);
> > 
> > Here, the "acc_wait (N)" would synchronize the local (host) thread with
> > async queue/CUDA stream N and thus recursively with [0..N).
> > 
> >        for (i = 0; i <= N; i++)
> >          {
> >            if (acc_async_test (i) != 1)
> >              abort ();
> >          }
> >      [...]
> > 
> > So, then all these async queues/CUDA streams here indeed are
> > "acc_async_test (i) != 1", thas is, idle.
> > 
> > 
> > Now, the more strict synchronization with the local (host) thread is not
> > wrong in term of correctness, but I suppose it will impact performance of
> > otherwise asynchronous operations, which now get synchronized too much?
> > 
> > Or, of course, I'm misunderstanding something...
> 
> IIRC, we encountered many issues where people misunderstood the meaning of "wait+async",
> using it as if the local host sync happened, where in our original implementation it does not.

..., and that's the right thing, in my opinion.  (Do you disagree?)

> Also some areas of the OpenACC spec were vague on whether the local host synchronization should
> or should not happen; basically, the wording treated as if it was only an implementation detail
> and didn't matter, and didn't acknowledge that this would be something visible to the user.

I suppose in correct code that correctly uses a different mechanism for
inter-thread synchronization, it shouldn't be visible?  (Well, with the
additional synchronization, it would be visible in terms of performance
degradation.)

For example, OpenACC 2.6, 3.2.11. "acc_wait" explicitly states that "If
two or more threads share the same accelerator, the 'acc_wait' routine
will return only if all matching asynchronous operations initiated by
this thread have completed; there is no guarantee that all matching
asynchronous operations initiated by other threads have completed".

I agree that this could be made more explicit throught the specification,
and also the reading of OpenACC 2.6, 2.16.1. "async clause" is a bit
confusing regarding multiple host threads, but as I understand, the idea
still is that such wait operations do not synchronize at the host thread
level.  (Let's please assume that, and then work with the OpenACC
technical committee to get that clarified in the documentation.)

> At the end, IIRC, I decided that adding a local host synchronization is easier for all of us,

Well...

> and took the opportunity of the re-org to make this change.

Well...  Again, a re-org/re-work should not make such functional
changes...

> That said, I didn't notice those tests you listed above were meant to test such delicate behavior.
> 
> > (For avoidance of doubt, I would accept the "async re-work" as is, but we
> > should eventually clarify this, and restore the behavior we -- apparently
> > -- had before, where we didn't synchronize so much?  (So, technically,
> > the "async re-work" would constitute a regression for this kind of
> > usage?)
> 
> It's not hard to restore the old behavior, just a few lines to delete. Although as described
> above, this change was deliberate.
> 
> This might be another issue to raise with the committee. I think I tried on this exact issue
> a long time ago, but never got answers.

OK, I'll try to find that, or send me a pointer to it, if you still got.

I propose you include the following.  Would you please review the "TODO"
comments, and again also especially review the
"libgomp/oacc-parallel.c:goacc_wait" change, and confirm no corresponding
"libgomp/oacc-parallel.c:GOACC_wait" change to be done, because that code
is structured differently.

commit e44cc6dc8f76e50c6f905cd408475589dee7b3b1
Author: Thomas Schwinge <thomas@codesourcery.com>
Date:   Thu Dec 13 17:54:35 2018 +0100

    into async re-work: don't synchronize with the local thread unless actually necessary
---
 libgomp/oacc-async.c    | 8 ++++++--
 libgomp/oacc-parallel.c | 1 -
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git libgomp/oacc-async.c libgomp/oacc-async.c
index a38e42781aa0..ec5cbc408d4e 100644
--- libgomp/oacc-async.c
+++ libgomp/oacc-async.c
@@ -195,9 +195,11 @@ acc_wait_async (int async1, int async2)
   if (aq1 == aq2)
     gomp_fatal ("identical parameters");
 
-  thr->dev->openacc.async.synchronize_func (aq1);
   if (aq2)
     thr->dev->openacc.async.serialize_func (aq1, aq2);
+  else
+    //TODO Local thread synchronization.  Necessary for the "async2 == acc_async_sync" case, or can just skip?
+    thr->dev->openacc.async.synchronize_func (aq1);
 }
 
 void
@@ -232,9 +234,11 @@ acc_wait_all_async (int async)
   gomp_mutex_lock (&thr->dev->openacc.async.lock);
   for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
     {
-      thr->dev->openacc.async.synchronize_func (l->aq);
       if (waiting_queue)
 	thr->dev->openacc.async.serialize_func (l->aq, waiting_queue);
+      else
+	//TODO Local thread synchronization.  Necessary for the "async == acc_async_sync" case, or can just skip?
+	thr->dev->openacc.async.synchronize_func (l->aq);
     }
   gomp_mutex_unlock (&thr->dev->openacc.async.lock);
 }
diff --git libgomp/oacc-parallel.c libgomp/oacc-parallel.c
index 9519abeccc2c..5a441c9efe38 100644
--- libgomp/oacc-parallel.c
+++ libgomp/oacc-parallel.c
@@ -508,7 +508,6 @@ goacc_wait (int async, int num_waits, va_list *ap)
       else
 	{
 	  goacc_aq aq2 = get_goacc_asyncqueue (async);
-	  acc_dev->openacc.async.synchronize_func (aq);
 	  acc_dev->openacc.async.serialize_func (aq, aq2);
 	}
     }


Grüße
 Thomas
Thomas Schwinge Dec. 14, 2018, 9:08 p.m. UTC | #5
Hi!

On Fri, 7 Dec 2018 16:30:53 +0100, I wrote:
> On Tue, 25 Sep 2018 21:11:42 +0800, Chung-Lin Tang <chunglin_tang@mentor.com> wrote:
> > These are the testsuite/libgomp.oacc-c-c++-common/* changes.
> 
> Please commit the following three hunks to trunk: the code as present
> doesn't declare its async/wait dependencies correctly.

As I had this queued as a prerequisite for other changes, in r267148 I
have now committed the following to trunk:

commit fef25f06de8e800d2a6ac04b12b6399923d414a9
Author: tschwinge <tschwinge@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Fri Dec 14 20:42:18 2018 +0000

    Correctly describe OpenACC async/wait dependencies
    
            libgomp/
            * testsuite/libgomp.oacc-c-c++-common/data-2-lib.c: Adjust.
            * testsuite/libgomp.oacc-c-c++-common/data-2.c: Likewise.
            * testsuite/libgomp.oacc-c-c++-common/data-3.c: Likewise.
    
    Reviewed-by: Thomas Schwinge <thomas@codesourcery.com>
    
    git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@267148 138bc75d-0d04-0410-961f-82ee72b054a4
---
 libgomp/ChangeLog                                        | 6 ++++++
 libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c | 2 +-
 libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c     | 2 +-
 libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c     | 2 +-
 4 files changed, 9 insertions(+), 3 deletions(-)

diff --git libgomp/ChangeLog libgomp/ChangeLog
index b4ab6b690553..b6cbb34908a2 100644
--- libgomp/ChangeLog
+++ libgomp/ChangeLog
@@ -1,3 +1,9 @@
+2018-12-14  Chung-Lin Tang  <cltang@codesourcery.com>
+
+	* testsuite/libgomp.oacc-c-c++-common/data-2-lib.c: Adjust.
+	* testsuite/libgomp.oacc-c-c++-common/data-2.c: Likewise.
+	* testsuite/libgomp.oacc-c-c++-common/data-3.c: Likewise.
+
 2018-12-14  Thomas Schwinge  <thomas@codesourcery.com>
 
 	PR libgomp/88370
diff --git libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
index 2ddfa7d4a01b..f553d3d839c5 100644
--- libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
+++ libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
@@ -153,7 +153,7 @@ main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N], N) \
-  async (4)
+  wait (1, 2, 3) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 
diff --git libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
index 0c6abe69dc17..81d623afa0ea 100644
--- libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
+++ libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
@@ -162,7 +162,7 @@ main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
-  wait (1) async (4)
+  wait (1, 2, 3) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 
diff --git libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
index 0bf706a1b5d4..5ec50b808a73 100644
--- libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
+++ libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
@@ -138,7 +138,7 @@ main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
-  wait (1,5) async (4)
+  wait (1, 2, 3, 5) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 


Grüße
 Thomas
diff mbox series

Patch

diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/asyncwait-2.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/asyncwait-2.c
new file mode 100644
index 0000000..9420540
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/asyncwait-2.c
@@ -0,0 +1,904 @@ 
+/* { dg-do run { target openacc_nvidia_accel_selected } } */
+/* { dg-additional-options "-lcuda" } */
+
+#include <openacc.h>
+#include <stdlib.h>
+#include <cuda.h>
+
+#include <stdio.h>
+#include <time.h>
+#include <sys/time.h>
+
+int
+main (int argc, char **argv)
+{
+    CUresult r;
+    CUstream stream1;
+    int N = 128; //1024 * 1024;
+    float *a, *b, *c, *d, *e;
+    int i;
+    int nbytes;
+
+    srand (time (NULL));
+    int s = rand () % 100;
+
+    acc_init (acc_device_nvidia);
+
+    nbytes = N * sizeof (float);
+
+    a = (float *) malloc (nbytes);
+    b = (float *) malloc (nbytes);
+    c = (float *) malloc (nbytes);
+    d = (float *) malloc (nbytes);
+    e = (float *) malloc (nbytes);
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+    }
+
+    acc_set_default_async (s);
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 3.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 2.0;
+        b[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 2.0)
+            abort ();
+
+        if (b[i] != 2.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copy (c[0:N]) copy (d[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 9.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 2.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc parallel wait (s) async (s)
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 2.0)
+            abort ();
+
+        if (b[i] != 4.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+
+        if (e[i] != 11.0)
+            abort ();
+    }
+
+
+    r = cuStreamCreate (&stream1, CU_STREAM_NON_BLOCKING);
+    if (r != CUDA_SUCCESS)
+    {
+        fprintf (stderr, "cuStreamCreate failed: %d\n", r);
+        abort ();
+    }
+
+    acc_set_cuda_stream (1, stream1);
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 5.0;
+        b[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 5.0)
+            abort ();
+
+        if (b[i] != 5.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 7.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copy (c[0:N]) copy (d[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 7.0)
+            abort ();
+
+        if (b[i] != 49.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc parallel wait (s) async (s)
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 9.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+
+        if (e[i] != 17.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 4.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copyin (a[0:N], b[0:N], c[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc update host (a[0:N], b[0:N], c[0:N]) wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 4.0)
+            abort ();
+
+        if (b[i] != 16.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+    }
+
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 5.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copyin (a[0:N], b[0:N], c[0:N]) copyin (N)
+    {
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc parallel async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc update host (a[0:N], b[0:N], c[0:N]) async
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 5.0)
+            abort ();
+
+        if (b[i] != 25.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 3.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 2.0;
+        b[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 2.0)
+            abort ();
+
+        if (b[i] != 2.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copy (c[0:N]) copy (d[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 9.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 2.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc kernels wait (s) async (s)
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 2.0)
+            abort ();
+
+        if (b[i] != 4.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+
+        if (e[i] != 11.0)
+            abort ();
+    }
+
+
+    r = cuStreamCreate (&stream1, CU_STREAM_NON_BLOCKING);
+    if (r != CUDA_SUCCESS)
+    {
+        fprintf (stderr, "cuStreamCreate failed: %d\n", r);
+        abort ();
+    }
+
+    acc_set_cuda_stream (1, stream1);
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 5.0;
+        b[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 5.0)
+            abort ();
+
+        if (b[i] != 5.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 7.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N]) copy (b[0:N]) copy (c[0:N]) copy (d[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 7.0)
+            abort ();
+
+        if (b[i] != 49.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 3.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copy (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
+    }
+
+#pragma acc kernels wait (s) async (s)
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
+    }
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 3.0)
+            abort ();
+
+        if (b[i] != 9.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+
+        if (d[i] != 1.0)
+            abort ();
+
+        if (e[i] != 17.0)
+            abort ();
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 4.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copyin (a[0:N], b[0:N], c[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc update host (a[0:N], b[0:N], c[0:N]) wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 4.0)
+            abort ();
+
+        if (b[i] != 16.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+    }
+
+
+    for (i = 0; i < N; i++)
+    {
+        a[i] = 5.0;
+        b[i] = 0.0;
+        c[i] = 0.0;
+        d[i] = 0.0;
+        e[i] = 0.0;
+    }
+
+#pragma acc data copyin (a[0:N], b[0:N], c[0:N]) copyin (N)
+    {
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            b[ii] = (a[ii] * a[ii] * a[ii]) / a[ii];
+    }
+
+#pragma acc kernels async
+    {
+        int ii;
+
+        for (ii = 0; ii < N; ii++)
+            c[ii] = (a[ii] + a[ii] + a[ii] + a[ii]) / a[ii];
+    }
+
+#pragma acc update host (a[0:N], b[0:N], c[0:N]) async
+
+#pragma acc wait (s)
+
+    }
+
+    for (i = 0; i < N; i++)
+    {
+        if (a[i] != 5.0)
+            abort ();
+
+        if (b[i] != 25.0)
+            abort ();
+
+        if (c[i] != 4.0)
+            abort ();
+    }
+
+    acc_shutdown (acc_device_nvidia);
+
+    return 0;
+}
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
index 2ddfa7d..f553d3d 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2-lib.c
@@ -153,7 +153,7 @@  main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N], N) \
-  async (4)
+  wait (1, 2, 3) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
index 0c6abe6..81d623a 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-2.c
@@ -162,7 +162,7 @@  main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
-  wait (1) async (4)
+  wait (1, 2, 3) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
index 0bf706a..5ec50b8 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/data-3.c
@@ -138,7 +138,7 @@  main (int argc, char **argv)
     d[ii] = ((a[ii] * a[ii] + a[ii]) / a[ii]) - a[ii];
 
 #pragma acc parallel present (a[0:N], b[0:N], c[0:N], d[0:N], e[0:N]) \
-  wait (1,5) async (4)
+  wait (1, 2, 3, 5) async (4)
   for (int ii = 0; ii < N; ii++)
     e[ii] = a[ii] + b[ii] + c[ii] + d[ii];
 
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-71.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-71.c
index c85e824..6afe2a0 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-71.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-71.c
@@ -92,16 +92,22 @@  main (int argc, char **argv)
       abort ();
     }
 
-  fprintf (stderr, "CheCKpOInT\n");
-  if (acc_async_test (1) != 0)
+  if (acc_async_test (0) != 0)
     {
       fprintf (stderr, "asynchronous operation not running\n");
       abort ();
     }
 
+  /* Test unseen async number.  */
+  if (acc_async_test (1) != 1)
+    {
+      fprintf (stderr, "acc_async_test failed on unseen number\n");
+      abort ();
+    }
+
   sleep ((int) (dtime / 1000.0f) + 1);
 
-  if (acc_async_test (1) != 1)
+  if (acc_async_test (0) != 1)
     {
       fprintf (stderr, "found asynchronous operation still running\n");
       abort ();
@@ -116,7 +122,3 @@  main (int argc, char **argv)
 
   return 0;
 }
-
-/* { dg-output "CheCKpOInT(\n|\r\n|\r).*" } */
-/* { dg-output "unknown async \[0-9\]+" } */
-/* { dg-shouldfail "" } */
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-77.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-77.c
index f4f196d..2821f88 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-77.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-77.c
@@ -111,7 +111,7 @@  main (int argc, char **argv)
 
   start_timer (0);
 
-  acc_wait (1);
+  acc_wait (0);
 
   atime = stop_timer (0);
 
@@ -132,7 +132,3 @@  main (int argc, char **argv)
 
   return 0;
 }
-
-/* { dg-output "CheCKpOInT(\n|\r\n|\r).*" } */
-/* { dg-output "unknown async \[0-9\]+" } */
-/* { dg-shouldfail "" } */
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
index ef3df13..b22af26 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-79.c
@@ -114,6 +114,7 @@  main (int argc, char **argv)
 
   for (i = 0; i < N; i++)
     {
+      stream = (CUstream) acc_get_cuda_stream (i & 1);
       r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, stream, kargs, 0);
       if (r != CUDA_SUCCESS)
 	{
@@ -122,11 +123,11 @@  main (int argc, char **argv)
 	}
     }
 
-  acc_wait_async (0, 1);
-
   if (acc_async_test (0) != 0)
     abort ();
 
+  acc_wait_async (0, 1);
+
   if (acc_async_test (1) != 0)
     abort ();
 
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c
index d5f18f0..30a4b57 100644
--- a/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/lib-81.c
@@ -133,7 +133,7 @@  main (int argc, char **argv)
 
   for (i = 0; i <= N; i++)
     {
-      if (acc_async_test (i) != 0)
+      if (acc_async_test (i) == 0)
 	abort ();
     }