[Intel-gfx] [PATCH v8] drm/i915: Squash repeated awaits on the same fence

[Chris Wilson]

On Thu, Apr 27, 2017 at 10:50:28AM +0100, Chris Wilson wrote:
> On Thu, Apr 27, 2017 at 08:06:36AM +0100, Chris Wilson wrote:
> > +int i915_gem_timeline_mock_selftests(void)
> > +{
> > +	static const struct i915_subtest tests[] = {
> > +		SUBTEST(igt_seqmap),
> 
> I should add a few benchmarks here as well.
> 
> random insertion
> random lookup (uses same random set as insertion)
> repeated lookups of neighbouring engines
> 
> So that we can compare in situ given our simple api.

Hmm, I may be biased, but on Braswell:

idr:
bench_sync: 196699 random insertions, 515ns/insert
bench_sync: 196699 random lookups, 376ns/lookup
bench_sync: 2428021 repeated insert/lookups, 41ns/op

1<<3 ht:
bench_sync: 7857 random insertions, 12766ns/insert
bench_sync: 7857 random lookups, 12855ns/lookup
bench_sync: 2164705 repeated insert/lookups, 47ns/op

1<<7 ht:
bench_sync: 17891 random insertions, 5733ns/insert
bench_sync: 17891 random lookups, 5618ns/lookup
bench_sync: 1983086 repeated insert/lookups, 52ns/op

That is better than my expectations! Once again, take with a pinch of
salt as random insetions are totally unrealistic.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre