[Intel-gfx] [RFC i-g-t] GEM features into feat_profile.json

Chris Wilson chris at chris-wilson.co.uk
Wed May 31 14:02:18 UTC 2017


On Wed, May 31, 2017 at 04:45:16PM +0300, Joonas Lahtinen wrote:
> On ke, 2017-05-31 at 13:58 +0100, Chris Wilson wrote:
> > On Wed, May 31, 2017 at 03:23:12PM +0300, Joonas Lahtinen wrote:
> > > 
> > > Hello,
> > > 
> > > I went through the gem_* tests from intel-gpu-tools and categorized
> > > them into roughly categories "X | X robustness | X performance" ready
> > > to be added to the feat_profile.json.
> > > 
> > > Lets open a discussion which ones should go where. I tried to place a
> > > single test to under only one category and I'm kind of hopeful that
> > > we'll have the ability to add "depends_on" to create super features in
> > > the future, instead of placing a single test under multiple categories.
> > > 
> > > I didn't check all the subtests nor wildcard matching with other tests,
> > > this is just all the test names placed under some categories.
> > 
> > You seem to have assigned them exclusively to one category or another,
> > most tests belong to a few of these categories. More when you consider a
> > subtest may be targetting a completely different aspect.
> 
> Yes, that's what I meant to say :) Subtests should probably be matched
> by another pattern like "\btiled\b", "\bflink\b" etc.
> 
> Ultimately there would be a resolver which would re-assign the
> subtests:
> 
> "Global objects" would then get:
> 
> 	"include_subtests": "flink",
> 
> Which would steal subtests with /\bflink\b/ from tests. Do we agree
> that one subtest would be assigned to one category only, or do you
> want to see duplication even at that level?

I see duplication everywhere. It's more a concept of tags as opposed to
categories.

The use of such a system would be as
	"give me all the tests that exercise relocation"
	"give me all the tests that use a context"
	"give me all the tests that exercise contention on $mutex"
	"give me all the tests that exercise file.c:line / this patch set"

The last one especially.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre


More information about the Intel-gfx mailing list