On Sat, Jan 14, 2017 at 10:08:11PM +0100, Loďc Grenié wrote:
> On 2017-01-14 at 10:37 GMT+01:00 Bill wrote:
>
> > On Sat, Jan 14, 2017 at 10:01:43AM +0100, Loďc Grenié wrote:

> > > > On Fri, Jan 13, 2017 at 11:00:28PM +0100, Loďc Grenié wrote:
> > > > > Package: pari
> > > > > Version: dd75740be
> > > > > Severity: normal
> > > > >
> > > > >      Hi Karim and Bill,
> > > > >
> > > > >      I think you probably understood what I will say. With pthread,
> > > > >
> > > > > parfor(n=10,20,addprimes(nextprime(10^n)))
> > > > >
> > > > >     crashes in the free(old) off addp().
> > > > >
> > > > >   addprimes() should probably be either disabled or protected with
> > > > >   pthread.
> > > >
> > > > We need a way for (MPI or posix) threads to inherit primetab from the
> > > > master thread.
> > >
> > >      Right now posix threads inherit primetab (that's the reason why it
> > >   crashes in the first place).
> >
> > Yes, but inheritance is an important feature to keep.
> >
> > >   A new field primetab in mt_queue could
> > >   allow to copy the master thread primetab to the children (for posix).
> > >   For MPI a send_request_GEN(PMPI_primetab, primetab, i);
> > >   could work (with the obvious modifications to mt/mpi.c). In that
> > >   case addprime() would be local to the child thread (in both cases).
> >
> > For posix thread, we should do it the same way as for  modular_eqn:
> >
> > see
> > void
> > pari_thread_sync(void)
> > {
> >   pari_pthread_init_varstate();
> >   pari_pthread_init_seadata();
> > }
> >
> > static GEN global_modular_eqn;
> > static THREAD GEN modular_eqn;
> > void
> > pari_init_seadata(void)  { global_modular_eqn = NULL; }
> > void
> > pari_thread_init_seadata(void)  { modular_eqn = global_modular_eqn; }
> > void
> > pari_pthread_init_seadata(void)  { global_modular_eqn = modular_eqn; }
> >
>
>      What happens with MPI?

seadata is loaded on demand by the MPI thread, as regular GP do.