On Fri, Jan 15, 2016 at 21:05:47 +0300, Ilya Verbin wrote:
> On Fri, Jan 15, 2016 at 17:45:22 +0100, Jakub Jelinek wrote:
> > On Fri, Jan 15, 2016 at 07:38:14PM +0300, Ilya Verbin wrote:
> > > On Fri, Jan 15, 2016 at 17:09:54 +0100, Jakub Jelinek wrote:
> > > > On Fri, Jan 15, 2016 at 05:02:34PM +0100, Martin Jambor wrote:
> > > > > How do other accelerators cope with the situation when half of the
> > > > > application is compiled with the accelerator disabled?  (Would some of
> > > > > their calls to GOMP_target_ext lead to abort?)
> > > > 
> > > > GOMP_target_ext should never abort (unless internal error), worst case 
> > > > it
> > > > just falls back into the host fallback.
> > > 
> > > Wouldn't that lead to hard-to-find problems in case of nonshared memory?
> > > I mean when someone expects that all target regions are executed on the 
> > > device,
> > > but in fact some of them are silently executed on the host with different 
> > > data
> > > environment.
> > 
> > E.g. for HSA it really shouldn't matter, as it is shared memory accelerator.
> > For XeonPhi we hopefully can offload anything.
> 
> As you said, if compilation of target image fails with ICE or somehow, host
> fallback and offloading to other targets should still work:
> https://gcc.gnu.org/ml/gcc-patches/2015-02/msg00951.html
> That patch was not applied, but it can be simulated by -foffload=disable,

I agree that OpenMP doesn't guarantee that all target regions must be executed
on the device, but in this case a user can't be sure that some library function
always will offload (because the library might be replaced by fallback version),
and he/she will have to write something like:

{
  map_data_to_target ();
  some_library1_fn_with_offload ();
  get_data_from_target ();   /* ! */
  send_data_to_target ();    /* ! */
  some_library2_fn_with_offload ();
  get_data_from_target ();   /* ! */
  send_data_to_target ();    /* ! */
  some_library3_fn_with_offload ();
  unmap_data_from_target ();
}

If you're OK with this, I'll install this patch:


libgomp/
        * target.c (gomp_get_target_fn_addr): Allow host fallback if target
        function wasn't mapped to the device with non-shared memory.

diff --git a/libgomp/target.c b/libgomp/target.c
index f1f5849..96fe3d5 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -1436,12 +1436,7 @@ gomp_get_target_fn_addr (struct gomp_device_descr 
*devicep,
       splay_tree_key tgt_fn = splay_tree_lookup (&devicep->mem_map, &k);
       gomp_mutex_unlock (&devicep->lock);
       if (tgt_fn == NULL)
-       {
-         if (devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
-           return NULL;
-         else
-           gomp_fatal ("Target function wasn't mapped");
-       }
+       return NULL;
 
       return (void *) tgt_fn->tgt_offset;
     }

  -- Ilya

Reply via email to