15:15karolherbst: does ldg.s8 sign extend?
15:17karolherbst: mhh, I guess so, otehrwise a signed shift wouldn't be enough
17:19orbea: Hi, with linux 4.19.2 novueau keeps taking my system down with a lot of: nouveau 0000:01:00.0: fifo: CHSW_ERROR 00000002 See this log - http://termbin.com/7hja and there was this trace - http://dpaste.com/2E0W0V2.txt
17:22karolherbst: orbea: is it specific to 4.19.2?
17:25Subv: hey, do dependent shader instructions execute sequentially or is there some sort of crazy result-forwarding in the Maxwell (GM20B) pipeline?
17:25orbea: i been having crashing issues for a while, but I'm not sure if the errors are the same or not
17:26orbea: it happens too infrequently to really bisect
17:26orbea: once every 1-2 days
17:26orbea: maybe more if I am not doing a lot of graphical stuff
17:27karolherbst: Subv: there isn't from an ISA point of view
17:27karolherbst: orbea: mhh, yeah, it's some issue with the context switching which is super annoying to debug
17:27karolherbst: there is some way to test the blob firmware, but I never actually did that
17:29Subv: that means 2 dependent sequential IADDs should be faster than a set of IADD+ISET+LOP that all depend on each other if their latencies were the same right? (any idea where i can get a list of instruction latencies/throughput for this GPU?)
17:33orbea: karolherbst: note that it was around 22,000 lines of that spam in my system logs...
17:33Subv: context: uaddCarry seems to be currently implemented with an IADD plus ISET to check for wrap-around and then a LOP to convert the -1/0 result of ISET into 1/0, i'm wondering if a simpler combination of IADD.C to compute the result + an "IADD.X carry, RZ, RZ" would be better
17:37AndrewR: karolherbst, it seems some of my replies yesterday were eaten by nework interruptions (on my side). No, plain nouveau doesn't work with OpenCL, at least test program prints all zeroes .... But I wonder how much of this spir-llvm I need? Because by default it compiles whole thing.. do I need *.a and *.so , or just tools?
17:38karolherbst: Subv: mhhh
17:39karolherbst: AndrewR: you need a bunch of mesa patches as well
17:40karolherbst: Subv: I have no idea, I usually suspects whatever nvidia is doing is the fastest
17:44Subv: nvidia uses IADD+SET+LOP
17:45AndrewR: karolherbst, one of those? https://github.com/karolherbst/mesa/branches
17:46karolherbst: AndrewR: yeah... but most of that stuff is either not really want you want or seriously incomplete
18:42Lyude: Anyone mind helping out with reviews on some of the MST patch series I submitted for nouveau?
23:11mfulz: does anybody know why I've around 6W more power consumption (1050ti optimus / PRIME) with discrete card in state off compared to bbswitch / nvidia / bumblebee?
23:12mfulz: Are there any tunables for that?
23:26joepublic: You must typicaly stay more than 15 minutes to... oh. he left.
23:30annadane: see i used to yell at people like that but then i started unintentionally closing irc when i had a pending question
23:30annadane: it happens lol
23:31joepublic: Perhaps got a solution from another source, or perhaps off on a different track that required a reboot, many things could have happened.
23:45rhyskidd: karolherbst: i spent some more time on the multi-part vbios issues with GP108+
23:45rhyskidd: nouveau and enyvtools don't parse the PCI Data Structure of the 3.0+ PCI firmware specification
23:46karolherbst: rhyskidd: huh? I kind of thought that's done at least for nouveau
23:46rhyskidd: that's why there's two parts, one is the old Intel x86, PC-AT compatible, the second type present is extensible firmware interface (EFI)
23:46rhyskidd: well -- nouveau doesn't choke on it
23:46rhyskidd: we might be missing some useful info though that's present in the new spec
23:50karolherbst: rhyskidd: is there a nouveau issue where it can't really load gp108 vbios files?
23:51rhyskidd: not that users have reported
23:52karolherbst: I think we shoulud just port nvbios over to use the nouveau code directly
23:53karolherbst: or something