[igt-dev] [PATCH i-g-t 1/2] lib/i915/gem_engine_topology: add iterator for choosing one random engine from each class

[Zbigniew Kempczyński]

On Mon, Mar 21, 2022 at 12:28:58PM +0000, Tvrtko Ursulin wrote:
> 
> On 21/03/2022 09:52, Petri Latvala wrote:
> > On Fri, Mar 18, 2022 at 05:04:44PM +0100, Kamil Konieczny wrote:
> > > In new GPUs there are many engines so tests with fixed timeout
> > > which iterate over all engines can take much longer than on older
> > > gens. Add new iterator for_one_random_cfg_ctx_engine, which will
> > > iterate over each class of engines and will choose one from each
> > > class at random.
> > > 
> > > Cc: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> > > Signed-off-by: Kamil Konieczny <kamil.konieczny at linux.intel.com>
> > > ---
> > >   lib/i915/gem_engine_topology.c | 64 ++++++++++++++++++++++++++++++++++
> > >   lib/i915/gem_engine_topology.h | 19 ++++++++++
> > >   2 files changed, 83 insertions(+)
> > > 
> > > diff --git a/lib/i915/gem_engine_topology.c b/lib/i915/gem_engine_topology.c
> > > index ca3333c2..52a2f3ef 100644
> > > --- a/lib/i915/gem_engine_topology.c
> > > +++ b/lib/i915/gem_engine_topology.c
> > > @@ -208,6 +208,70 @@ void intel_next_engine(struct intel_engine_data *ed)
> > >   	}
> > >   }
> > > +static void __intel_random_init(void)
> > > +{
> > > +	struct timespec start;
> > > +
> > > +	clock_gettime(CLOCK_MONOTONIC, &start);
> > > +	igt_debug("seed %d\n", (int)start.tv_nsec);
> > > +	srand((int)start.tv_nsec);
> > > +}
> > 
> > I don't like randomness for testing, but it sure is sometimes needed.
> > 
> > We have quite a bit of rand()-using tests and some of them have a
> > --seed parameter so their results can be reproduced. Currently there's
> > no overlap in sight with them and this thing, but if testing ever
> > expands that way, mixing RNGs might break reproduceability. Can this
> > instead use another RNG stream that's self-contained, like maybe
> > rand_r() or such? Or the one implemented in lib/igt_rand.
> 
> Agreed, for developer use it is an essential requirement to be able to run a
> test multiple times with same parameters. So any test which would be
> converted to use for_one_random_ctx_cfg_engine would need to have a common
> way of specifying a stable config.
> 
> Hm if we look at the end result:
> 
>  	igt_subtest_with_dynamic("active") {
> -		for_each_ctx_cfg_engine(fd, &cfg, e) {
> +		for_one_random_ctx_cfg_engine(fd, &cfg, e) {
>  			igt_dynamic_f("%s", e->name)
>  				active(fd, &cfg, e, 20, 1);
>  		}
> 
> Would that even work better at the test runner layer? New testlist syntax to
> pick one random out of a list of dynamic subtests?

Getting random engine is prone for two problems:

1. if we have same seed we always choose same engine in single test
   (as igt_runner is executing tests in separate processes we lost randomness
    at all)

2. if we have different seed we will have fluctuations on e->name across
   different runs. I don't know how CI will react if subtest will be rcs0
   then bcs0 in next run. 

I think subtest name crc/hash would be good seed, it would allow us to 
keep same engine across runs for single subtest and selecting different 
engines across subtests.

--
Zbigniew


> 
> Regards,
> 
> Tvrtko