02:00 <robclark> AMD_performance_monitor isn't a _great_ extension if you can't read the counters from cmdstream.. it's ok as long as app doesn't try to begin or end the query at a point that would force a flush and cpu stall on gpu.. probably the more sane thing to use is perfetto/pps so the counters don't have to be coordinated with gpu rendering

03:07 <cphealy> Yea, I was thinking of the perfetto/pps integration (which works today). Is there an easy way to associate the perfetto/pps reporting with specific client frame render?

03:21 <robclark> some of the drivers have renderpass support.. the main thing is you need a way to capture a gpu timestamp from cmdstream

03:21 <robclark> u_trace gives the infrastructure to re-associated those timestamps with render-stage (or other) traces after the fact without having to stall waiting for gpu

03:22 <robclark> but not sure if mali has a way to read gpu perfcntr from cmdstream?

03:23 <robclark> on freedreno side, we have a counter that increments at a fixed rate (19.2MHz) which we use for gpu timestamps

03:23 <robclark> and then perfetto clock-sync traces to teach perfetto how to convert from gpu to cpu time

03:24 <HdkR> Very curious if that 19.2Mhz is going to bump to 1Ghz since ARMv9.x requires it at some point :D

03:26 <cphealy> robclark: Thanks for the hints!

03:26 <robclark> how does armv9 make requirements on gpu timestamp precision? Given that most adreno's run at slower clk than 1GHz, I can't see how that would be useful

03:33 <HdkR> robclark: They make requirements on the CPU cycle counter, which on Snapdragon as far as I'm aware are the same clock or at least linked to be the same. Since ARMv9.1/8.6 mandates 1Ghz it'll be interesting if it is strill intrinsically linked

03:35 <HdkR> If it is required on the CPU side, might be good to just reuse the clock since it is there :D

03:36 <robclark> hmm, I'd have to add some new uapi to query the rate instead of just hard-coding 19.2MHz like it has been since forever ;-)

03:36 <HdkR> haha, yep

03:38 <HdkR> Luckily even the latest thing announced is only ARMv9.0