Re: [PATCH 2/2] [OpenACC] Detect pointer updates for attach operations (PR95590)

[Julian Brown <julian@codesourcery.com>]

On Fri, 24 Jul 2020 16:04:02 +0200
Thomas Schwinge <thomas@codesourcery.com> wrote:

> Hi Julian!
> 
> On 2020-06-22T05:14:44-0700, Julian Brown <julian@codesourcery.com>
> wrote:
> > As mentioned in the parent email, this is a fix for PR95590 that
> > detects updates of attached pointers in blocks, and rewrites the
> > attached pointer and resets its attachment counter appropriately. I
> > am however not entirely sure this is desirable or required by the
> > spec: points against are:
> >
> >  - To avoid expensive copies from the device to the host and/or
> > "wrong way" device-to-host splay tree lookups, it requires keeping
> > an extra shadow copy of mapped blocks on the host in order to
> > detect if a host pointer with attachments in the block has been
> > changed between attach operations.  
> 
> I haven't spent too much time trying, but I too have not yet seen a
> way to avoid keeping this state ("shadow copy"), or looking it up on
> demand ("expensive copies from the device to the host").
> 
> I suppose we cannot get the necessary information/state from the
> host-side pointer (value) alone, and/or other state kept in the
> 'splay_tree_key n' etc.?

I don't think so.

A different implementation might keep the attachment counters
associated with the target_mem_desc (on the "target side"), rather than
the splay tree key (the "host side"), in which case the
reset-on-host-pointer-modification might sort-of happen for free. But I
think that would be quite problematic for other reasons with our
current implementation. (Purely speculating, but maybe it "works"
somewhat accidentally for PGI because of the way its host-to-device
pointer mapping is implemented?)

> >    We incur this overhead unconditionally if
> >    attach/detach are in use for what's not likely to be a common
> > use case  
> 
> Is the overhead so bad, though?  As soon as there's an 'attach', we
> have to 'malloc' anyway (can combine the two, as you've done), and the
> checking overhead doesn't seem so bad either?
> 
> Should we reach out to other OpenACC compiler implementors, and ask
> for their understanding/approach to this aspect?

I haven't measured the performance impact (it's probably negligible). It
may be worth trying to get clarification from OpenACC upstream, though.

> >    (it's slightly tricky to write a test case to exercise the
> > behaviour, even -- Thomas's unmodified original for the PR raises
> > an error after the previous patch in this series).  
> 
> Challange accepted!  ;-P (..., but not right now.)

There was a test case attached to the parent email, too :-).

> >  - From a user perspective, I think it's going to be quite easy to
> > get confused wrt. the hidden attachment counter state  
> 
> (Indeed that "hidden" aspect is a bit confusing.  I've even thought
> whether we should add some 'gomp_get_attach_count' function just for
> our own testing purposes.)

Yeah, maybe.

> >    with this kind of
> >    reset-on-host-pointer-modification behaviour.  Mind you,
> > silently *not* doing the update is likewise going to be confusing
> > (the stale device pointer would be updated at present).  Maybe this
> > should be detected as an error instead?  
> 
> I don't understand that, I'm afraid, because as I have quoted in
> <https://gcc.gnu.org/PR95590> "OpenACC 'attach' behavior if already
> attached to different data", OpenACC explicitly mandates the
> "reset-on-host-pointer-modification" behavior, so I don't see a way to
> avoid implementing that?
> 
> >  - The text in "2.6.8. Attachment Counter" *might* contribute to the
> >    argument that this kind of pointer-update detection is not
> > required.  
> 
> Do you think these texts are in conflict in some way (that's not
> obvious to me)?

I'm still not sure that the intended meaning (in OpenACC 2.6, 2.7.2.
"Data Clause Actions", "Attach Action") is what you are reading into
it. See also "2.7.1. Data Specification in Data Clauses", under
Restrictions:

"* In C and C++, modifying pointers in pointer arrays during the data
lifetime, either on the host or on the device, may result in undefined
behavior."

That isn't explicitly about pointers within structs (as we're talking
about here), but is of a similar flavour, I think -- in that
recognizing host pointer modifications in arrays of pointers would
require similar housekeeping in the runtime, but OpenACC 2.6 makes such
modifications undefined behaviour instead.

The text in "2.6.7. Attachment Counter" (in OpenACC 2.6) is
specifically about update operations (acc_update API routines or
equivalent directives), but again, detecting pointer modifications
(on the host side) between successive "attach" operations seems like a
departure from *not* needing to do the same for update operations.

Should we also support modifications of attached pointers (e.g. in
mapped structs) in device-side code? Why or why not? (That wouldn't
be impossible, but the details of how it could work would be ugly
indeed...)

Here's a quick example of "weird" behaviour that would arise with the
pointer-modification detection patch:

#include <assert.h>
#include <stdlib.h>

struct mystr {
  int *ptr;
};

#define N 1024

int 
main (int argc, char *argv[])
{
  int *arr1 = malloc (sizeof (int) * N);
  int *arr2 = malloc (sizeof (int) * N);
  struct mystr s;
 
  for (int i = 0; i < N; i++) 
    { 
      arr1[i] = i;
      arr2[i] = i * 2;
    } 
 
  s.ptr = arr1;

  #pragma acc enter data copyin(s)
 
  #pragma acc data copy(s.ptr[0:N])
  {
    s.ptr = arr2;
    #pragma acc parallel loop copy(s.ptr[0:N])
    for (int i = 0; i < N; i++)
      s.ptr[i] = i * 3;
  }
 
  for (int i = 0; i < N; i++)
    {
      assert (arr1[i] == i);
      assert (arr2[i] == i * 3);
    }
 
  free (arr1);
  free (arr2);
}

With the patch, this gives:

libgomp: attach count underflow

Though of course it doesn't work properly without the
pointer-modification detection patch either.

This example could be made to work, but it would mean *not* resetting
the attachment counter to one on detecting a modified host pointer --
the pointer mapping would be modified but the attachment counter would
be incremented as usual (at the start of the "acc parallel"). That's
arguably the right thing to do perhaps, but it's clearly not what the
spec says, even with your reading.

HTH,

Julian