Re: Fwd: [RFC][gomp4] Offloading patches (2/3): Add tables generation

[Bernd Schmidt <bernds at codesourcery dot com>]

On 03/05/2014 06:15 PM, Ilya Verbin wrote:
> On 28 Feb 17:21, Bernd Schmidt wrote:
> > I think it won't help that much - I still think this entire scheme
> > is likely to fail on nvptx. I'll try to construct an example at
> > some point.
> >
> > One other thing about the split tables is that we don't have to
> > write a useless size of 1 for functions.
> >
> >
> > The concept of "image" is likely to vary somewhat between
> > accelerators. For ptx, it's just a string and it can't really be
> > generated the same way as for your target where you can manipulate
> > ELF images. So I think it is better to have a call to a gomp
> > registration function for every offload target. That should also
> > give you the ordering you said you wanted between shared
> > libraries.
> >
> >
> > Probably. I think the constructor call to the gomp registration
> > function would contain an opaque pointer to whatever data the
> > target wants, so it can arrange its image/table data in whatever
> > way it likes.
>
> Assuming that we're using the scheme with tables. Every DSO with
> offloading must contain a constructor call to GOMP_offload_register
> (const void *openmp_target); The openmp_target descriptor in every
> DSO will have target-independent entries (addresses of host tables)
> and target-dependent entries for each target. Its format may be like
> this:
>
> void *__OPENMP_TARGET__[] = { _omp_host_func_table;
> _omp_host_funcs_end; _omp_host_var_table; _omp_host_vars_end;
> _omp_num_targets; _omp_target_descs[]; /* array of tgt_desc */ }

I don't see why you want the array of target descriptors - it would take
some effort to construct, and as far as I can tell it's unnecessary. You
can just pass a pointer to the corresponding descriptor to every
GOMP_offload_register call.


> struct tgt_desc { int _omp_tgt_id; void *_omp_tgt_%s_image_start;
> void *_omp_tgt_%s_image_end; void *_omp_tgt_%s_func_mappings; void
> *_omp_tgt_%s_var_mappings; /* some other data if needed */ }

This looks reasonable.

> During the devices initialization libgomp will pass the openmp_target
> pointer to all plugins. Each plugin will scan over tgt_descs and find
> the required entries using the _omp_tgt_id.

Once again, that seems unnecessarily complicated. The plugins can 
register their target ID with libgomp, and when libgomp sees a 
GOMP_offload_register call with the corresponding target ID, it can 
invoke the appropriate plugin immediately.

> BTW, do you have any estimate when you will commit your patches to
> the branch, so that we could merge them with ours, and get something
> working for everybody?

I've been waiting for us to reach agreement on how things should look. 
If there are patches in the series that you're happy with, let me know 
and I can commit them (it may be next week though).


Bernd