[Intel-gfx] [igt-dev] [RFT i-g-t 6/6] test: perf_pmu: use the gem_engine_topology library

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Thu Apr 11 14:37:26 UTC 2019


On 11/04/2019 14:50, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2019-04-11 13:53:24)
>>
>> On 11/04/2019 13:32, Chris Wilson wrote:
>>> Quoting Tvrtko Ursulin (2019-04-11 13:26:54)
>>>> From: Andi Shyti <andi.shyti at intel.com>
>>>>
>>>> Replace the legacy for_each_engine* defines with the ones
>>>> implemented in the gem_engine_topology library.
>>>>
>>>> Use whenever possible gem_engine_can_store_dword() that checks
>>>> class instead of flags.
>>>>
>>>> Now the __for_each_engine_class_instance and
>>>> for_each_engine_class_instance are unused, remove them.
>>>>
>>>> Suggested-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>> Signed-off-by: Andi Shyti <andi.shyti at intel.com>
>>>> Cc: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
>>>> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>
>>> I see a lot of new gem_context_create(), but not gem_require_contexts().
>>
>> Problem is __for_each_physical_engine modifies the default context so
>> tests under that iterator can't keep using I915_EXEC_RENDER.
>>
>> It was perhaps a knee-jerk reaction to make them create their own
>> context as a workaround.
> 
>  From the looks of the tests affected, they all should support contexts. I
> would be more concerned if the basic busyness checking tests grew a new
> dependency.

Basic busyness ones have not, but the ones I fixed up by making them 
create a context have. Which is fixable with below.

Failures are only with gem_wait/basic-*-all, so I suspect something's 
off in igt_dummyload.c. Will see what happens with the debugs added.

> Operating on the default context, should be fine though. May require an
> operation to reset in case a previous subtest failed, which is a
> downside.
>   
>> Should we have a helper after all to get eb flags for given class:instance?
> 
> I think resist/delay as much as possible. In 3 months time we can sweep
> away any evidence of the old ways, and only the explicit legacy ABI
> tests need concern themselves with BSD | BSD2 ;)

perf_pmu tests which I added context_create to open PMU for rcs:0 and 
submit batches to I915_EXEC_RENDER.

Well alternative to the helper could be to change them to submit to 
I915_EXEC_DEFAULT and rely on context map to always have rcs:0 at index 
0. Which could even be passable. If the promise is broken the tests will 
fail to let us know.

Okay with this alternative hack?

Regards,

Tvrtko


More information about the Intel-gfx mailing list