02:11 fdobridge: <g​fxstrand> `Pass: 402427, Fail: 932, Crash: 84, Skip: 1729073, Timeout: 2, Missing: 494, Flake: 358, Duration: 1:39:48`
02:13 fdobridge: <g​fxstrand> Oh, and those runs are using only 16 GPRs
02:14 fdobridge: <g​fxstrand> 12 of the crashes I *think* I just fixed. The rest are missing atomic image cmpxchg and shader clock. I've got a minion doing the former and I'll get to the later one of these days.
02:14 fdobridge: <g​fxstrand> I assume shader clock is just a system value?
02:33 fdobridge: <a​irlied> yeah appears to be 0x50
02:34 fdobridge: <a​irlied> are you using linear instruction scheduling?
02:56 fdobridge: <g​fxstrand> Yeah, no scheduling yet
07:12 fdobridge: <k​arolherbst🐧🦀> one of the few 64 bit ones even
07:13 fdobridge: <k​arolherbst🐧🦀> I think?
08:05 fdobridge: <k​arolherbst🐧🦀> yeah.. 0x50 and 0x51
08:07 fdobridge: <k​arolherbst🐧🦀> thing is... S2R only allows 32 bit loads, so I don't know if it's a problem, but I think the value can get messed up? dunno
08:08 fdobridge: <k​arolherbst🐧🦀> ohh, there is a CS2R variant which can do 64 bit loads
08:08 fdobridge: <k​arolherbst🐧🦀> CS2R only supports a handful of values
08:09 fdobridge: <k​arolherbst🐧🦀> looks like all 64 bit ones are supported bia CS2R
08:09 fdobridge: <k​arolherbst🐧🦀> and only those
08:10 fdobridge: <k​arolherbst🐧🦀> and apparently CS2R is faster
13:27 fdobridge: <g​fxstrand> 🤷🏻‍♀️
13:27 fdobridge: <g​fxstrand> Easy enough to wire up, I suppose.
13:31 fdobridge: <k​arolherbst🐧🦀> yeah.. the encoding is super trivial
13:31 fdobridge: <k​arolherbst🐧🦀> no flags, except CS2R has a .32/.64 flip
13:31 fdobridge: <k​arolherbst🐧🦀> and there are 255 special registers
13:32 fdobridge: <k​arolherbst🐧🦀> and they can only write to a gpr and no indirect access
13:32 fdobridge: <k​arolherbst🐧🦀> and no U variant
13:33 fdobridge: <k​arolherbst🐧🦀> I have the full table of all available special registers btw
13:33 fdobridge: <k​arolherbst🐧🦀> 0-79 are reserved in CS2R (they are all 32 bit values)
13:34 fdobridge: <k​arolherbst🐧🦀> same for 132-254
13:34 fdobridge: <k​arolherbst🐧🦀> huh interesting
13:34 fdobridge: <k​arolherbst🐧🦀> there is another clock special value
13:34 fdobridge: <k​arolherbst🐧🦀> at 82:83
13:35 fdobridge: <k​arolherbst🐧🦀> it's synchronized but has lower precision
13:35 fdobridge: <k​arolherbst🐧🦀> and apparently the update interval can be configured
13:42 fdobridge: <!​[NVK Whacker] Echo (she) 🇱🇹> Why is compute a priority for NAK? Is it because of the CUDA potential? :triangle_nvk:
13:42 fdobridge: <k​arolherbst🐧🦀> it's simpler
13:43 fdobridge: <k​arolherbst🐧🦀> don't have to bother with all that input/output stuff
13:43 fdobridge: <k​arolherbst🐧🦀> also.. less instructions to deal with
13:43 fdobridge: <k​arolherbst🐧🦀> NVK also started bringing up compute first
14:23 fdobridge: <g​fxstrand> `Pass: 402573, Fail: 930, Crash: 75, Skip: 1729463, Timeout: 2, Flake: 327, Duration: 1:38:34`
14:28 fdobridge: <k​arolherbst🐧🦀> mhhhhh.. my current status on investigating those GPU crashes.. submitting with `NOUVEAU_GEM_PUSHBUF_SYNC` makes those errors go away, mostly
14:29 fdobridge: <k​arolherbst🐧🦀> there is a lot of funky stuff going on, but it's clearly related to channel allocation and deallocation
17:53 fdobridge: <k​arolherbst🐧🦀> this issue is soo annoying
21:42 dakr: airlied, gfxstrand: Just pushed a branch containing the scheduler rework: https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/sched-single-entity
21:44 dakr: It contains my GPUVA manager patches (common dma-resv, ext/evicted obj tacking), Matt's DRM_SCHED_POLICY_SINGLE_ENTITY work, some fixes I made on top of that, nouveau making use of DRM_SCHED_POLICY_SINGLE_ENTITY, my attempt of dynamic flow control in the scheduler.
21:45 dakr: The top commit enables dynamic flow control for nouveau. Hence, to see how much of a difference it makes, it can just be reverted.
22:13 fdobridge: <g​fxstrand> What is "dynamic flow control" in this context?
22:15 dakr: It's what we discussed with Christian lately.
22:15 dakr: Considering actual job size vs maximum job size.
22:57 fdobridge: <g​fxstrand> ah
22:57 fdobridge: <g​fxstrand> Sorry. Flow-control means something very different in my brain. 😅