Quoting Tvrtko Ursulin (2018-10-01 11:38:37)
>
> On 28/09/2018 14:58, Chris Wilson wrote:
> > When submitting chains to each engine, we can do so (mostly) in
> > parallel, so delegate submission to threads on a per-engine basis.
> >
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > ---
> > drivers/gpu/drm/i915/selftests/intel_lrc.c | 73 ++++++++++++++++++----
> > 1 file changed, 61 insertions(+), 12 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/selftests/intel_lrc.c b/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > index 3a474bb64c05..d68a924c530e 100644
> > --- a/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > +++ b/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > @@ -587,8 +587,10 @@ static int random_priority(struct rnd_state *rnd)
> > struct preempt_smoke {
> > struct drm_i915_private *i915;
> > struct i915_gem_context **contexts;
> > + struct intel_engine_cs *engine;
> > unsigned int ncontext;
> > struct rnd_state prng;
> > + unsigned long count;
> > };
> >
> > static struct i915_gem_context *smoke_context(struct preempt_smoke *smoke)
> > @@ -597,31 +599,78 @@ static struct i915_gem_context *smoke_context(struct preempt_smoke *smoke)
> > &smoke->prng)];
> > }
> >
> > +static int smoke_crescendo_thread(void *arg)
> > +{
> > + struct preempt_smoke *smoke = arg;
> > + IGT_TIMEOUT(end_time);
> > + unsigned long count;
> > +
> > + count = 0;
> > + do {
> > + struct i915_gem_context *ctx = smoke_context(smoke);
> > + struct i915_request *rq;
> > +
> > + mutex_lock(&smoke->i915->drm.struct_mutex);
> > +
> > + ctx->sched.priority = count % I915_PRIORITY_MAX;
> > +
> > + rq = i915_request_alloc(smoke->engine, ctx);
> > + if (IS_ERR(rq)) {
> > + mutex_unlock(&smoke->i915->drm.struct_mutex);
> > + return PTR_ERR(rq);
> > + }
> > +
> > + i915_request_add(rq);
> > +
> > + mutex_unlock(&smoke->i915->drm.struct_mutex);
> > +
> > + count++;
>
> Very little outside the mutex so I am not sure if parallelization will
> work that well. Every thread could probably fill the ring in it's
> timeslice?
Very unlikely due to the randomised ring. And we are working on that,
right? :)
> And then it blocks the others until there is space. It will
> heavily rely on scheduler behaviour and mutex fairness I think.
But it does bring the overall subtest time from num_engines to 1s, for
the same pattern.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
