[Beignet] [PATCH v2] GetEventProfilingInfo could query all event status.

Tue Aug 4 21:48:59 PDT 2015

For this special case, the user application has an serious bug which is forgetting to call clReleaseEvent()
and will leak one event.

For beignet, the real problem is if an application rely on event driven mechanism, then it may never
get a chance to move forward, here is a simple example:

main()
{
  ...
  clEnqueueNDRange(queue,...,event);
  clSetEventCallback(event, .., callbackFn);
  ...
  pthread_create(pFn);//creat a child thread.
  pthread_wait() // wait for the last thread to terminate.
  clReleaseEvent(event);
  clFinish(queue);
}

callbackFn() {
  pthread_cond_signal(); // notify the child thread to do some thing.
  ...
}

pFn() {
  pthread_cond_wait()
  ...//do something on CPU;
  clEnqueueNDRange(queue, ..., event2);
  ...
}

The callbackFn will never be called and the child thread will always wait for the condition variable
even the actual event has been finished soon after the first enqueuer().

So add event status update in clFinish() will not really fix this type of problem. The real fix should be to
rework the whole event handling mechanism and add one daemon thread to maintain event status.
I really can't think of one method without an additional thread to solve this problem.

Thanks,
Zhigang Gong.

> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of
> Yang, Rong R
> Sent: Wednesday, August 5, 2015 11:38 AM
> To: Zhigang Gong; Luo, Xionghu
> Cc: beignet at lists.freedesktop.org
> Subject: Re: [Beignet] [PATCH v2] GetEventProfilingInfo could query all event
> status.
> 
> The root cause of beignet using on-demand manner is no point to update
> event's status. If there have this point, I think there is no side effect to update
> the status.
> 
> I just find an case:
> 
> clEnqueueNDRange(....., event);
> clSetEventCallback(event, ....);
> clFinish(...);
> return;
> 
> In event callback, release all cl resource, include event and command queue.
> Then beignet will never call the callback.
> Of cause, this case is the current beignet event bug, if there is not clFinish, the
> callback should also be called in future, but beignet will not.
> Add the update status in clFinish just a workaround, do you the method to fix
> it?
> 
> > -----Original Message-----
> > From: Zhigang Gong [mailto:zhigang.gong at linux.intel.com]
> > Sent: Tuesday, August 4, 2015 16:48
> > To: Yang, Rong R; Luo, Xionghu
> > Cc: beignet at lists.freedesktop.org
> > Subject: RE: [Beignet] [PATCH v2] GetEventProfilingInfo could query
> > all event status.
> >
> > Beignet current uses an on-demand manner to maintain the events' status.
> > For those APIs which need to get current events' status, we insert a
> > cl_gpgpu_event_update_status(). It seems there is no need for clFinish().
> >
> > If you could find any case that if we don't insert it in clFinish()
> > then it will cause functional error, I will agree that we need both
> > patches. Even though xionghu's patch need to refine the commit log
> > message, as it is not to fix
> > GetEventProfilingInfo() issue.
> >
> > Thanks,
> > Zhigang Gong.
> >
> > > -----Original Message-----
> > > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On
> > > Behalf Of Yang, Rong R
> > > Sent: Tuesday, August 4, 2015 3:31 PM
> > > To: Zhigang Gong; Luo, Xionghu
> > > Cc: beignet at lists.freedesktop.org
> > > Subject: Re: [Beignet] [PATCH v2] GetEventProfilingInfo could query
> > > all event status.
> > >
> > > I think these two patches both need when application
> > GetEventProfilingInfo.
> > >
> > > From application view, it is make sense that all event that don't
> > > depend on other events should have been flushed and finished when
> > clFinish return.
> > > The problem is the event have finished, but beignet don't update the
> > > event's status, so I think it is better to update the event status
> > > before exit
> > clFinish.
> > >
> > > And GetEventProfilingInfo is another point need to update the event'
> > status.
> > > Suppose in the application, don't call clFinish, but maybe event has
> > > completed, should not return CL_PROFILING_INFO_NOT_AVAILABLE.
> > >
> > > > -----Original Message-----
> > > > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On
> > > > Behalf Of Zhigang Gong
> > > > Sent: Tuesday, August 4, 2015 13:28
> > > > To: Luo, Xionghu
> > > > Cc: beignet at lists.freedesktop.org
> > > > Subject: Re: [Beignet] [PATCH v2] GetEventProfilingInfo could
> > > > query all event status.
> > > >
> > > >
> > > > The spec only says if event is not in CL_COMPLETE status, it will
> > > > get CL_PROFILING_INFO_NOT_AVAILABLE. It doesn't mean a clFinish()
> > > > should update all events status. According to the spec, clFinish()
> > > > only need to make sure the previously enqueued task has been
> > > > processed and
> > > completed.
> > > >
> > > > Form my point of view, if the application wants to call
> > > > clGetEventProfilingInfo() to get an event's profiling information,
> > > > the application should call
> > > > clWaitForEvents() firstly. But given the fact that the
> > > > specification is not very clear for whether clFinish() should update all
> events'
> > > > status and some applications are really believe all events has
> > > > been also updated after the
> > > > clFinish() call. We may make a work around in Beignet to make
> > > > those applications happy. Your patch is one solution and another
> > > > may be better solution may be to add cl_event_update_status(event,
> > > > 0) when calling to clGetEventProfilingInfo(). The patch is as
> > > > below,
> > > >
> > > > From a5a1b3f372d17f26cc20fba078490b61614f07e5 Mon Sep 17
> 00:00:00
> > > 2001
> > > > From: Zhigang Gong <zhigang.gong at intel.com>
> > > > Date: Tue, 4 Aug 2015 13:21:27 +0800
> > > > Subject: [PATCH] runtime: always try to update event status in
> > > > clGetEventProfilingInfo().
> > > >
> > > > Some applications forgot to call clWaitForEvents() before calling
> > > > to clGetEventProfilingInfo(). Let's update the event's status here.
> > > >
> > > > Signed-off-by: Zhigang Gong <zhigang.gong at intel.com>
> > > > ---
> > > >  src/cl_api.c | 1 +
> > > >  1 file changed, 1 insertion(+)
> > > >
> > > > diff --git a/src/cl_api.c b/src/cl_api.c index 69eb0bc..5c9b250
> > > > 100644
> > > > --- a/src/cl_api.c
> > > > +++ b/src/cl_api.c
> > > > @@ -1468,6 +1468,7 @@ clGetEventProfilingInfo(cl_event
> > > event,
> > > >    cl_ulong ret_val;
> > > >
> > > >    CHECK_EVENT(event);
> > > > +  cl_event_update_status(event, 0);
> > > >
> > > >    if (event->type == CL_COMMAND_USER ||
> > > >        !(event->queue->props & CL_QUEUE_PROFILING_ENABLE) ||
> > > > --
> > > > 1.9.1
> > > >
> > > >
> > > > This way we can avoid unecessary event updating in clFinish() call.
> > > > The overhead is only introduced when the application calling
> > > > clGetEventProfilingInfo() and we do not need busy wait for the
> > > > event updating here.
> > > >
> > > > Thanks,
> > > > Zhigang Gong.
> > > >
> > > > On Tue, Aug 04, 2015 at 01:11:10PM +0800, xionghu.luo at intel.com
> wrote:
> > > > > From: Luo Xionghu <xionghu.luo at intel.com>
> > > > >
> > > > > this could fix the shoc project reported
> > > > > CL_PROFILING_INFO_NOT_AVAILABLE issue.
> > > > > https://github.com/vetter/shoc/issues/47
> > > > > v2: per spec, the GetEventProfilingInfo need return
> > > > > CL_PROFILING_INFO_NOT_AVAILABLE if the event is not
> > CL_COMPLETE,
> > > > so
> > > > > the event should be updated in clFinish.
> > > > >
> > > > > Signed-off-by: Luo Xionghu <xionghu.luo at intel.com>
> > > > > ---
> > > > >  src/cl_command_queue.c | 3 +++
> > > > >  1 file changed, 3 insertions(+)
> > > > >
> > > > > diff --git a/src/cl_command_queue.c b/src/cl_command_queue.c
> > > > > index
> > > > > 4e4ebfb..4b92311 100644
> > > > > --- a/src/cl_command_queue.c
> > > > > +++ b/src/cl_command_queue.c
> > > > > @@ -276,6 +276,9 @@ LOCAL cl_int
> > > > > cl_command_queue_finish(cl_command_queue queue)  {
> > > > >    cl_gpgpu_sync(cl_get_thread_batch_buf(queue));
> > > > > +  cl_event last_event = get_last_event(queue);  if (last_event)
> > > > > +    cl_event_update_status(last_event, 1);
> > > > >    return CL_SUCCESS;
> > > > >  }
> > > > >
> > > > > --
> > > > > 1.9.1
> > > > >
> > > > > _______________________________________________
> > > > > Beignet mailing list
> > > > > Beignet at lists.freedesktop.org
> > > > > http://lists.freedesktop.org/mailman/listinfo/beignet
> > > > _______________________________________________
> > > > Beignet mailing list
> > > > Beignet at lists.freedesktop.org
> > > > http://lists.freedesktop.org/mailman/listinfo/beignet
> > > _______________________________________________
> > > Beignet mailing list
> > > Beignet at lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/beignet
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet