[Mesa-dev] [PATCH 00/15] GL_AMD_performance_monitor

[Robert Bragg]

On Tue, Mar 31, 2015 at 6:56 PM, Alex Deucher <alexdeucher at gmail.com> wrote:
> On Thu, Mar 19, 2015 at 6:40 PM, Robert Bragg <robert at sixbynine.org> wrote:
>> Hi Samuel,
>>
>> I thought I should let you and others know of some related changes
>> I've been playing around with to try and improve our support for the
>> INTEL_performance_query extension in the i965 dri driver, since it's
>> quite related to your work.
>>
>> Some of my work-in-progress changes can currently be seen on github here:
>> https://github.com/rib/mesa/commits/wip/rib/i915_oa_perf
>>
>> I've only recently scrubbed up these mesa changes enough for reviewing
>> to make much sense, so sorry I haven't sent out an RFC to the list
>> before now about this.
>>
>> What I was finding is that it's not very helpful in practice to
>> support the AMD_performance_monitor + INTEL_performance query
>> extensions via the same core infrastructure and so since there weren't
>> previously any other drivers implementing AMD_performance_monitor, my
>> series currently just drops the combined infrastructure in favour of
>> core support for just INTEL_performance_query.
>>
>> The main thing here that might be interesting to think about with
>> respect to your work, is whether you might like to do anything
>> differently with the main/performance_monitor.c support if you were
>> the only user and if it strictly only dealt with the
>> AMD_performance_monitor extension?
>>
>> I don't guess it would be controversial for us to look at removing the
>> current i965 backend for both AMD_performance_monitor and
>> INTEL_performance_query for an interim if that might be convenient for
>> you to evolve the backend interface. (Currently the i965 backend
>> doesn't report many usable metrics unfortunately - mainly just the
>> pipeline statistics that are also accessible via query objects.) Ken
>> and Petri who did much of the work here are already aware that I've
>> been taking this approach as I ran it by them some time ago and their
>> work still carries forward in the series I have but instead focused
>> only on INTEL_performance_query for now.
>>
>> Anyway, I just thought I should throw that out there as a possibility
>> to consider in case it might be helpful.
>
> I would prefer to keep support for AMD_performance_monitor in mesa.
> We may implement more extensive support for this extension in our
> radeon open source drivers and it would be nice to be compatible with
> our closed source drivers on both Linux and other OSes.

Right, especially given that Samuel has implemented a backend now, I
assume the core support will be staying.

I just deleted it in my own series where I was also removing the only
backend for it but I'll be rebasing my series without the patch to
drop the AMD_performance_monitor infrastructure.

To clarify; what I was suggesting above was also based on the
assumption that the core AMD_performance_monitor support is being
kept. I was talking about removing the intel backend if that might
make Samuel's life easier in case he has some ideas for changing the
backend interface. Such changes would be less effort if he doesn't
have to worry about updating the intel backend too. I know we'd need
to make some changes to the backend interface if we were to look at
updating the intel AMD_performance_monitor support so I wouldn't
really want the current backend lending weight to keeping the
interface as is.

I hope that provides some reassurance.

Regards,
- Robert

>
> Alex