Re: Too strict synchronization with the local (host) thread?

[Thomas Schwinge <thomas@codesourcery.com>]

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