[Intel-gfx] [PATCH i-g-t] i915/gem_exec_schedule: Exercise implicit ordering between engines

[Chris Wilson]

Quoting Antonio Argenziano (2020-02-18 17:56:10)
> 
> 
> On 18/02/20 09:42, Chris Wilson wrote:
> > Check that reads are serialised after a write, and that a subsequent
> > write is after all reads.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Antonio Argenziano <antonio.argenziano at intel.com>
> > Cc: Sravan Kumar Nedunoori <sravan.kumar.nedunoori at intel.com>
> > ---
> >   tests/i915/gem_exec_schedule.c | 73 ++++++++++++++++++++++++++++++++++
> >   1 file changed, 73 insertions(+)
> > 
> > diff --git a/tests/i915/gem_exec_schedule.c b/tests/i915/gem_exec_schedule.c
> > index a20985864..cfd06aa55 100644
> > --- a/tests/i915/gem_exec_schedule.c
> > +++ b/tests/i915/gem_exec_schedule.c
> > @@ -241,6 +241,61 @@ static void fifo(int fd, unsigned ring)
> >       igt_assert_eq_u32(result, 2);
> >   }
> >   
> > +enum implicit_dir {
> > +     READ_WRITE = 0x1,
> > +     WRITE_READ = 0x2,
> > +};
> > +
> > +static void implicit_rw(int i915, unsigned ring, enum implicit_dir dir)
> > +{
> > +     IGT_CORK_FENCE(cork);
> > +     unsigned int count;
> > +     uint32_t scratch;
> > +     uint32_t result;
> > +     int fence;
> > +
> > +     count = 0;
> > +     for_each_physical_engine(other, i915) {
> > +             if (eb_ring(other) == ring)
> > +                     continue;
> > +
> > +             count++;
> > +     }
> > +     igt_require(count);
> > +
> > +     scratch = gem_create(i915, 4096);
> > +     fence = igt_cork_plug(&cork, i915);
> > +
> > +     if (dir & WRITE_READ)
> > +             store_dword_fenced(i915, 0,
> > +                                ring, scratch, 0, -ring,
> > +                                fence, I915_GEM_DOMAIN_RENDER);
> > +
> > +     for_each_physical_engine(other, i915) {
> > +             if (eb_ring(other) == ring)
> > +                     continue;
> > +
> > +             store_dword_fenced(i915, 0,
> > +                                eb_ring(other), scratch, 0, eb_ring(other),
> > +                                fence, 0);
> > +     }
> > +
> > +     if (dir & READ_WRITE)
> > +             store_dword_fenced(i915, 0,
> > +                                ring, scratch, 0, ring,
> > +                                fence, I915_GEM_DOMAIN_RENDER);
> > +
> > +     unplug_show_queue(i915, &cork, ring);
> > +     close(fence);
> > +
> > +     result =  __sync_read_u32(i915, scratch, 0);
> > +     gem_close(i915, scratch);
> > +
> > +     igt_assert_neq_u32(result, -ring);
> 
> if (dir & WRITE_READ) ?

We should never have -ring in there. If we have 2 billion engines, CI
will timeout first. And ring should always be !0.

Although the symmetry is pleasing.
-Chris