ChanServ changed the topic of #zink to: official development channel for the mesa3d zink driver ||
avane_ has quit [Ping timeout: 480 seconds]
avane has joined #zink
phire has quit [Remote host closed the connection]
phire has joined #zink
xperia64 has quit [Remote host closed the connection]
xperia64 has joined #zink
sinanmoh- has joined #zink
sinanmohd has quit [Read error: Connection reset by peer]
<anholt_> zmike: is zink going to be doing cmd buffer annotations that have start/end mixed between inside and outside an rp?
<zmike> anholt_: uh
<zmike> you mean like starting a marker inside a cmdbuf and ending it outside?
<anholt_> yeah
<anholt_> err, inside a renderpass and ending it outsid
<zmike> I don't think I can imagine a case where that could occur?
<zmike> the markers I'm putting in are all around single commands
<zmike> so there's no way for a renderpass to start or stop inside a marker
<anholt_> I've got a tu-marker branch, but it's not going well.
<zmike> :(
<anholt_> ah, image_barrier(VK_IMAGE_LAYOUT_UNDEFINED->VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL) seems to start outside an rp and end inside
<anholt_> (for example)
<zmike> uh
<anholt_> I should be able to sort this out.
<zmike> did I not push all my marker changes? 🤔
<zmike> I think I did
<zmike> yeah I don't get how that's possible
<anholt_> zink_cmd_debug_marker_end(ctx, marker); probably needs to take the cmdbuf too
<zmike> oh wow
<zmike> facepalm
<zmike> yeah that'll do it
<anholt_> I can grab that
<zmike> awesome, thanks
* zmike is so deep into tc mechanics the daylight is but a memory
<airlied> isn't it legal to start them in one cmd buf and end them in another also?
<zmike> good q
<zmike> maybe?
<zmike> but that's definitely not intended
<airlied> yeah I'm pretty sure we had a bug report to allow that, even if zink doesn't use it
<zmike> I meant it's not intended if it's happening now, I'd have to check spec to see if it's allowede
<anholt_> airlied: yeah. it's allowed. but not what we wanted.
<anholt_> zmike: tu-marker is updated with my wip annotation code. I think the annotation is getting use-after-freed, and instead I need to go port intel's "app_event" trick, but I also really want to find a perfetto person and ask if we're just doing it wrong because you can't possibly have to do it like this, right?
<zmike> I'm afraid to look
<zmike> anholt_: grimace
<anholt_> who doesn't love tracing through multiple layers of codegenned c++. with some cpp macros for extra flavor.
<zmike> thanks for your service
<zmike> anholt_: after some work today: supertuxkart zink: 103fps, fd 113fps
<zmike> tiler perf is real.
<anholt_> not too shabby
<zmike> typing up one of the last remaining issues I've seen that you might have some ideas on
<zmike> ugh but somehow this has zero effect on at least the mobile_legends trace, which has been my goto