06:41karolherbst: on main: [110751/110751] Pass: 101855 Fail: 764 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 67 Timeout: 24 Missing: 0 Flake: 38 Duration: 5:46:02 Remaining: 0
06:41karolherbst: mt fixes: [110751/110751] Pass: 101892 Fail: 726 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 69 Timeout: 27 Missing: 0 Flake: 34 Duration: 5:50:46 Remaining: 0
07:54mupuf: karolherbst: yeeaaaahhh!
07:55karolherbst: mupuf: yeah.. finally :D
07:55karolherbst: there are a handful of regressions but I think I figured those out
07:55mupuf: karolherbst: that's your v2, right?
07:55karolherbst: now I have to see if it survives the android emulator
07:55karolherbst: mupuf: well more like v10
07:56karolherbst: I did a lot of experimenting
07:56ccr: v10 in binary is still v2 in decimal!
07:56mupuf: ccr: good one :)
07:56karolherbst: but I finally have a good plan on how to fix this issue for good
07:57karolherbst: fencing was always the most annoying bit to figure out
07:57karolherbst: because.. how to do fencing when you have multiple contexts with their own pushbuffer but just one timeline on the hw
07:58karolherbst: oh well
08:00mupuf: karolherbst: how are others dealing with this?
08:00mupuf: other drivers*
08:01karolherbst: I only checked radeonsi at some point but they don't seem to have a global fence list as we do
08:02mupuf: what do you mean by global? Per GPU context or driver instance?
08:02karolherbst: per gallium screen
08:02karolherbst: I am not sure if we still have to keep it this way though, but I didn't really want to rework all the driver at this point
08:03karolherbst: especially as we have persistent state
08:03karolherbst: and radeon doesn't
08:03karolherbst: so.. from what I know is that radeonsi just builds their command buffer and everything is in there
08:03karolherbst: but nouveau just assumes old state and makes a diff
08:04karolherbst: and then order of submissions between contexts and stuff matters
08:05karolherbst: and I think the bigger issue is, if I start to rework everything I can just write a new driver instead :D
08:05ccr: nouveau^2 ? or would it be 3 now? :P
08:06mupuf: hehe, yeah
08:06karolherbst: we still have a classic driver :D
08:06karolherbst: my secret hope is that zink will be awesomely fast
08:06mupuf: nvc0 vs nv50 was also quite different
08:06karolherbst: and we just have to deal with a vulkan driver
08:06HdkR:puts all the burden on Mike to be as fast as a native driver
08:07karolherbst: so I am not in the mood of really writing a new driver when there is still vulkan
08:07ccr: there's a finnish proverb "kolmas kerta toden sanoo", which roughly translates to the english version "third time is the charm", but more accurately "third time tells the truth"
08:07mupuf: karolherbst: that is indeed the way to go :)
08:08karolherbst: " 64 files changed, 696 insertions(+), 459 deletions(-)" is the change of my mt fixes
08:08karolherbst: and most of it is just refactoring
08:08karolherbst: which isn't all that much in the end :D
08:08karolherbst: and like 30% is inside nouveau_fence
08:09karolherbst: I was running into fun dead locks where we deadlocked between the GPU and CPU :(
08:10karolherbst: because the order of the fences on the CPU and the GPU differed and our code totally can't handle that
08:10mupuf: oh joy
08:10mupuf: hopefully, you didn't end up flushing too much :)
08:10karolherbst: yeah... so we waited on an alreadyed hit fence, but we didn't know on the CPU side :D
08:10karolherbst: ahh no
08:10karolherbst: I didn't add any flushing at all
08:11karolherbst: just have to be careful when submitting fences
08:11mupuf: karolherbst: sooo, will you talk about Nouveau at XDC2021?
08:11karolherbst: uhm.... maybe?
08:11mupuf: deadline is on the 4th of July
08:12karolherbst: I am mildy aware
08:12karolherbst: I guess that could be a fun topic to talk about
08:16mupuf: sure, and maybe say "Vulkan support will come... one day"!
08:17karolherbst: sadly there is still so much to fix..
08:17karolherbst: even our channel recovery code is bonkers
08:18mupuf: seems like a constant among drivers...
08:18mupuf: one day... one day!
08:18karolherbst: yeah.. just other drivers are way more stable on the userspace side
08:19karolherbst: ohh, you meant the genreal fixing
08:19karolherbst: oh well
08:19karolherbst: they have more devs :D
08:20mupuf: no no, i meant hang recovery
08:20karolherbst: the problem isn't the code
08:20karolherbst: the problem is that it isn't race free in nouveau
08:20karolherbst: so if you get multiple channels dieing at the same time you are screwed
08:20karolherbst: which... forces me to run that stupid deqp stuff single threaded :D
08:21mupuf: ah, that explains the terrible runtime :D
08:21karolherbst: and even single threaded the GPU might just crash comepletly and I have to shutdown and wait and boot :p
08:22karolherbst: I even hit an issue which persisted across reboots
08:22mupuf: these are the best...
14:28karolherbst: getting better: [110751/110751] Pass: 101899 Fail: 713 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 69 Timeout: 24 Missing: 0 Flake: 43 Duration: 5:44:51 Remaining: 0
14:29karolherbst: too many flakes though :D
14:29karolherbst: but this run fixed a few stable regressions
14:31karolherbst: ehh.. one regression is still there.. I thought I fixed that one
14:34karolherbst: ehh.. because I broke it twice
15:08Santurysim: What does flake mean?
15:08karolherbst: that it sometimes passes and sometimes fails
19:41pmoreau[m]: karolherbst: I think you pinged me, but it unfortunately took me too long to look at it.
19:41karolherbst: don't remember why :D
19:42imirkin: pmoreau[m]: have you done, or plan to do, any more work on the opencl-for-nv50 project?
19:42pmoreau[m]: Not at the moment
19:42imirkin: i pushed the ES 3.1 for gt21x stuff, so that's nice
19:42pmoreau[m]: So far I have planning to have weekends and evenings again, but still some time off it looks like.
19:42pmoreau[m]: *I am
19:43imirkin: there are some outstanding issues with that impl, esp around graphics/compute interaction
19:48pmoreau[m]: I will be on vacation in 2 weeks, but so far I am looking forward to no longer seeing a computer. I might pick it up again in late August/September depending on how things go, or maybe I'll do some over summer but unlikely.
19:48karolherbst: pmoreau[m]: yeah.. good luck with whatever you are up to, but yeah...
19:49karolherbst: I feel like I need a huge break after fixing MT as well
20:06imirkin: pmoreau[m]: cool, enjoy!
20:08imirkin: karolherbst: btw, i didn't have a chance to review your stuff yet
20:08karolherbst: no worries
21:13karolherbst: [110751/110751] Pass: 101906 Fail: 703 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 69 Timeout: 24 Missing: 0 Flake: 46 Duration: 5:42:51 Remaining: 0 :)
21:14karolherbst: not sure why I pass 50 more tests with my fixes, but maybe they help.. dunno
21:54mrnuke: orbea: You were right about Dell stringing me along for refurb parts. But The machine I got is in much better condition that the one I sent in, and the SSD is practically new -- "Percentage Used 0%"
21:54orbea: at least something improved
21:55mrnuke: I'm curious how nouveau will perform as soon as I replace the Redmond os with something that can run nouvea ;)
22:11mangix: mrnuke: depends on GPU
22:46mrnuke: mangix: It's a GTX 1650