<tomeu>
I think the OPPs weren't merged in mainline
<robmur01>
the default GPU clock may be something utterly rubbish (e.g. that's the case on RK3328, where lima doesn't even have devfreq at all so it stays there)
<robher>
tomeu: maybe a module param or something to set the timeout?
<robher>
tomeu: But how do parallel jobs really change things? The timeout should start when the job is given to the h/w. So it's only a question of 1 to 3 job slots...
<tomeu>
maybe related to gpu resets at the wrong time
<tomeu>
running 8 deqp instances in parallel must be a good way to stress test :)
<tomeu>
robher: if we have another job queued in the hw, what happens to it?
<robmur01>
FWIW, I can tell you T860MP1 at 40MHz is still considerably faster than T720MP1 at 40MHz ;)
<robmur01>
the latter can't even manage 1fps on most of glmark :D
<robher>
tomeu: we only have the *_NEXT registers, but I don't think we use those yet.
<tomeu>
ah, it's true, we didn't merge them because they made mesa/panfrost slower
<tomeu>
ok, that's good info
<tomeu>
which explains why the latest faults only failed one test, and not 8
<alyssa>
tomeu: Typo, whoops.
<alyssa>
tomeu: I don't understand the 0xFF specialness.
<alyssa>
tomeu: We should definitely revisit the _NEXT registers
<alyssa>
IIRC there's some stuff related to bbrezillon's job pipelining series that needs to be finished up and then we should get a good boost from _NEXT
davidlt has quit [Ping timeout: 246 seconds]
pH5 has quit [Quit: bye]
<chewitt>
tomeu: alyssa: t820 all green is awesome news! .. things have come a long way since https://imgur.com/a/HwJ0bwP (12 December 2018)
<chewitt>
next stop bifrost!?
karolherbst has quit [Ping timeout: 265 seconds]
yann has quit [Ping timeout: 276 seconds]
chewitt has quit [Quit: Zzz..]
<bbrezillon>
alyssa: I don't remember what was missing
<bbrezillon>
didn't I address all the job inter-deps issues?