[Intel-gfx] [igt-dev] [PATCH i-g-t 2/2] tests/gem_exec_nop: Remove submission batching

[Chris Wilson]

Quoting Janusz Krzysztofik (2020-05-11 09:51:12)
> Hi Chris,
> 
> On Fri, 2020-05-08 at 18:54 +0100, Chris Wilson wrote:
> > Quoting Janusz Krzysztofik (2020-05-08 14:56:31)
> > >  static double nop_on_ring(int fd, uint32_t handle,
> > >                           const struct intel_execution_engine2 *e, int timeout,
> > > -                         unsigned long *out)
> > > +                         unsigned long *count)
> > >  {
> > >         struct drm_i915_gem_execbuffer2 execbuf;
> > >         struct drm_i915_gem_exec_object2 obj;
> > >         struct timespec start, now;
> > > -       unsigned long count;
> > > +       unsigned long total;
> > > +
> > > +       igt_assert(*count);
> > >  
> > >         memset(&obj, 0, sizeof(obj));
> > >         obj.handle = handle;
> > > @@ -93,18 +95,18 @@ static double nop_on_ring(int fd, uint32_t handle,
> > >         }
> > >         intel_detect_and_clear_missed_interrupts(fd);
> > >  
> > > -       count = 0;
> > > +       total = 0;
> > >         clock_gettime(CLOCK_MONOTONIC, &start);
> > >         do {
> > > -               for (int loop = 0; loop < 1024; loop++)
> > > +               for (int loop = 0; loop < *count; loop++)
> > 
> > This unnerves me. I expect to get this wrong when writing new callers.
> > 
> > There's no great reason to even have 1024 here, we can survive with
> > doing clock_gettime() every iteration, and just accept it as part of the
> > systematic cost.
> 
> Thanks for fixing this and merging the change.
> 
> Following our discussion on IRC which I probably didn't understand
> precisely enough but maybe have a better understanding now after having
> it reread, how about still addressing the "headless" requirement for
> better precision with the following hunk?
> 
> @@ -373,10 +372,12 @@ stable_nop_on_ring(int fd, uint32_t handle,
>         while (reps--) {
>                 unsigned long count;
>                 double time;
>  
>                 time = nop_on_ring(fd, handle, e, timeout, &count);
> +               igt_skip_on_f(count < 1000 * timeout,
> +                             "submicrosecond precision of time measurement\n");

Yes. I just can't remember off the top of my head what order of
magnitude the error was -- I think it was around an extra 100us (50x
worse latency).
-Chris