01:06karolherbst: skeggsb: did you change something inside the kernel module so that issues like that won't appear in userspace? https://trello.com/c/Kgv3gKBf/26-khr-gl45vertexattribbindingadvanced-iterations-fails-in-full-run
01:07karolherbst: allthough I hardly believe that anything transformfeedback related should be caused by something inside the kernel
01:09skeggsb: not that i'm aware of... unless it was something in the gr magics
01:10karolherbst: mhh, maybe I am just lucky in the cts run
01:10karolherbst: currently doing a full one just to see what issues I migh encounter
01:10karolherbst: but until now everything is fine (except those weird glx issues)
01:15HdkR: How does Nouveau implement blend_equation_advanced? Trying to grep the source gives me blend state object which doesn't seem to directly deal with it
01:20HdkR: Is it actually implemented with a framebuffer fetch?
01:45nyef: HdkR: Looks like it's built on PIPE_CAP_TGSI_FS_FBFETCH, and is implemented as a shader fragment at the state-tracker level, if I'm reading this right?
01:46HdkR: That's what I seem to have learned from grepping as well
01:46HdkR: Which sounds unsafe when you have overlapping primitives
01:47nyef: That I couldn't comment on.
01:47karolherbst: skeggsb: we kind of have to improve how o handle engine recoveries :( it can take minutes until something happens
01:48karolherbst: or sometimes nothing happens at all and the userspace process is simply stuck
08:38tagr: skeggsb: any chance we can get this series applied for v4.19? https://patchwork.freedesktop.org/series/42101/
08:39tagr: I think we may want backports all the way to 4.16 since that's where IOMMU was originally broken depending on kernel configuration
09:02skeggsb: tagr: ack, i can take it in my tree, was somehow expecting the arm guys were dealing with it, so forgot about the series :P
09:18RSpliet: skeggsb: speaking of gr magic, I pushed one of the tools to envytools last week that I used for instrumenting context switch info, one that tells the size of a context (plus how it's split out) of GPUs up to Maxwell. When I have time* I consider looking into whether I can actually extract the registers it swaps - not sure where the list of regs is stored after initialisation. Either way I hope this proves helpful in debugging rela
10:58pendingchaos: imirkin, karolherbst: I think I'll try to create an updated patch that takes into account register bank conflicts and throughput sometime
17:11RSpliet: pendingchaos: https://github.com/RSpliet/mesa/commits/bank-aware-RA
17:12RSpliet: I had played with a bank-aware register allocation scheme, code is up there. I didn't see a perf difference on Kepler or Maxwell, as I suspect bottlenecks are currently elsewhere
17:14RSpliet: But invite you to take the code and evolve it if you think it makes a difference
17:15RSpliet: (in particular, it might benefit from some logic wrt. dual issue... but that's a tougher nut to crack considering we don't know what will be issued together at the point of RA)
17:40nyef: Hrm. The blob driver really likes to reload the context-switch microcode, doesn't it?
17:57pendingchaos: RSpliet: what's mapBankBitsetToDisableMask()?
18:02pendingchaos: never mind, I think I figured it out
18:35RSpliet: nyef: hw bug in certain gens of hw, related to BLCG/ELPG/....
18:57nyef: RSpliet: Interesting... does nouveau account for this as well?
18:58nyef: Or does nouveau just not do the relevant gating, and thus (in theory) isn't affected?
19:41RSpliet: nyef: currently we don't do the relevant gating
19:44nyef: We don't do the relevant gating, thus we hope that we're not affected? Sounds about right.
20:15RSpliet: nyef: lyude is the person to ask further about this
20:16RSpliet: although he too does not have the details on this hw bug ;-)
20:25nyef: Eh, right now I'm more interested in trying to nail this dma_pusher thing than in getting power management up and going.
22:16pendingchaos: imirkin: am I correct in thinking this shouldn't break anything: https://hastebin.com/cecomiquro.diff?
22:41cliff-hm: one char change from 'C' to 'D' ? :)
22:42cliff-hm: (only diff my tired eyes saw)
22:43cliff-hm: CFG -> DFS .. not a great font and small :)
22:50nyef: Control-flow graph to depth-first sort?
22:51nyef: So, basically the same nodes, just now with a specific / different specific order?
22:51pendingchaos: seems to cause Unigine Superposition to crash
22:53pendingchaos: after spending a short while on the load screen
22:55pendingchaos: it doesn't seem to crash during compilation though, so I suspect it's creating incorrect code