Akari has quit [Remote host closed the connection]
jabashque has quit [Read error: Connection reset by peer]
jabashque has joined #zink
<baryluk>
zmike: found way to reproduce glmark2 issue: ./build/src/glmark2 --data-path $(pwd)/data --run-forever --benchmark build:use-vbo=false:duration=100
<baryluk>
It leaks some memory (GTT in radeontop) like crazy.
<baryluk>
what is even weirder. is that even if you reduce duration to 3, but keep --run-forever, which should destroy and create new context, it still leaks (with non-lazy descriptors at least)
<baryluk>
I guess. I should be able to bisect this.
<zmike>
that is weird
<zmike>
I'm deep into fixing tomb raider again, but this is looking like another top 3 worst bugs of all time entry
<baryluk>
Well. Because I can observe the radeontop, and kill the glmark before it reaches close to limit, I thik bisect should be quick to do.
<zmike>
great
<airlied>
zmike: got any idea where it broke?
<zmike>
airlied: hm?
<zmike>
the tomb raider thing uh...sort of but also not really
<zmike>
in benchmark mode when using auto descriptors, the fallback from cache -> lazy on a single ubo set for a compute shader somehow breaks this stupid hair rendering
<baryluk>
hmm. I am not sure how to aproach this bisecting. Bisecting in current zink-wip will not work. I think I need to do testing between zink-wip snapshots
<zmike>
even though as near I can tell the descriptors in radv are completely identical
<zmike>
baryluk: try against mesa main? zink-wip doesn't have much these days
<baryluk>
ok.
<baryluk>
when is 21.3 planned?
<zmike>
branchpoint tomorrow
<baryluk>
ok. then I will try to finish bisecting today