Re: [hsa merge 07/10] IPA-HSA pass

[Ilya Verbin <iverbin at gmail dot com>]

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've created a testcase:

$ cat main.c

#pragma omp declare target
int x;
#pragma omp end declare target
extern int foo ();

int main ()
{
  int shared_mem = 0;
  #pragma omp target map (alloc: x, shared_mem)
    {
      x = 10;
      shared_mem = 1;
    }

  x = 20;
  int r = foo ();
  if (!shared_mem && r != 100)
    __builtin_abort ();
  return 0;
}


$ cat liba.c 

#pragma omp declare target
extern int x;
#pragma omp end declare target

int foo ()
{
  int r;
  #pragma omp target map (from: r) map (alloc: x)
    r = x * x;
  return r;
}


$ gcc -fopenmp -fPIC -shared liba.c -o liba.so -foffload=disable
$ gcc -fopenmp -L. -la main.c


Currently it prints "libgomp: Target function wasn't mapped", but after this
change:

--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -1390,7 +1390,7 @@ 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)
-	gomp_fatal ("Target function wasn't mapped");
+	return NULL;

... it will fail at __builtin_abort, but without -foffload=disable it will pass.

  -- Ilya