00:02RSpliet: or... you could use this off-line to generate some of the missing tables (3,5,6...)
09:31karolherbst: pmoreau: what is the status with your clover patches now? Will you have time to work on those or shall I?
10:33RSpliet: mwk: have you by any chance spotted branch prediction hints in the fμc opcode format?
10:35RSpliet: Their RISC-V presentation from last year mentions falcon cores have "static" prediction, which either means opcode-flags or some strange heuristic like "short backward jumps are loops, assume taken, otherwise assume not taken"...
10:44mwk: RSpliet: nope, nothing spotted
10:45mwk: if there is something like that, which I doubt, then it's never been used in official microcode
10:45mwk: I'd guess "static" prediction means the usual loop-or-if heuristic based on direction
10:46RSpliet: mwk: that was my second guess ;-) for the record, source - https://www.youtube.com/watch?v=gg1lISJfJI0&t=820s
10:48mwk: yeah, seen that
10:49mwk: btw, another interesting takeaway from that is that falcon has externally-vectored interrupts in addition to the ones going through its interrupt controller
10:49mwk: which I suppose explains the weirdo unknown bits in flags registers that look like another enable + saved enable pair
10:54RSpliet: Oh I missed that hint saying falcons run at 1.5GHz
10:54RSpliet: That probably means they do actually run in the GPC clock domain
12:47karolherbst: pmoreau: I updated your log2 64 bit patches: https://github.com/karolherbst/mesa/commits/gallium_log64
12:47karolherbst: please ack the changes I made :)
12:47karolherbst: or send them out to the ML
12:50RSpliet: karolherbst: can I ask you to sacrifice 30 seconds to e-mail me my gt640 (NVE7/1) VBIOS? I've got three machines directly accessible, none of which have nvidia-vbios checked out.
12:51karolherbst: RSpliet: yeah
12:52karolherbst: but uhm... which one?
12:52karolherbst: there are like 5 variants of the gt640 :)
12:52karolherbst: gddr5 or ddr3?
12:54RSpliet: there's one in a folder with my ancient (presumably TU Delft, or did I change it to eclipso) e-mail address
12:54RSpliet: the TU Delft e-mail address is defunct by the way ;-) Should change those dir names at some point
12:54karolherbst: it has the eclipso email
12:55RSpliet: Thanks, much appreciated!
16:19juri_: has there been a 'how to use nouveau as a cuda replacement' document written yet?
16:23Natehaxx: good question
16:28HdkR: juri_: Wouldn't it just be a guide to using OpenGL compute? ;)
16:29RSpliet: juri_: that'd be a rather short document I'm afraid.
16:29RSpliet: OpenCL is in the pipeline, nobody looking at CUDA runtime at the moment afaik
16:30RSpliet: but "in the pipeline" means "steps are taken towards making it work in controlled settings", not "ready to be released to the general public"
16:30Natehaxx: if nouveau would take the steps for vulkan api into the maxwell drivers i would be happy
16:31Natehaxx: but that seems not the case for the moment
16:31HdkR: https://github.com/ROCm-Developer-Tools/HIP Might be worthwhile looking in to
16:31RSpliet: Writing the manual unfortunately would currently take more time documenting all the limitations than it takes lifting those limitations one by one, which in an under-staffed project is more useful ;-)
16:32RSpliet: Natehaxx: same applies. All the work you read on Phoronix about a NIR front-end to the nouveau code compiler is in the anticipation of a shared SPIR-V -> NIR translation. SPIR-V is sorta-kinda the intermediate representation for both OpenCL and Vulkan
16:33RSpliet: "Vulkan support" is a monstrous task, and the "baby steps" taken currently are insufficient to make it work straight away but necessary on the path nonetheless
16:34Natehaxx: RSplite: thanks for the replay i will watch the progress good luck with that
16:35RSpliet: I wish I was in a position to contribute myself, but I have to prioritise PhD research at the moment... uni isn't paying me for hacking on nouveau unfortunately!
16:36Natehaxx: yeah if you are doing the uni around here in germany you are paying too much to have less free time
16:37RSpliet: UK, they pay worse here
16:37Natehaxx: the complete college cost 13k here
16:38RSpliet: IIRC in Germany you get a stipend or contract for doing a PhD, around... €1700,=/month?
16:39RSpliet: but the cost of living there is significantly lower than at least the southern half of England
16:40Natehaxx: are you sure about that ?
16:40Natehaxx: i pay 400 euros for my apartment
16:40Natehaxx: and its small
16:40Natehaxx: if you go to a big city with a college on it you will pay around 800 euros
16:41Natehaxx: but with the stipend you are right
16:42RSpliet: I pay the equivalent of 800 euros (£700*€1.15) for a 14m^2 apartment... yes, the cost of living in the UK is ridiculous
16:43RSpliet: with apartment, I mean room btw.
16:43RSpliet: I've been in Saarbrücken for 9 months, it's not just the rent that's lower, also the daily expenses
16:45Natehaxx: what 800 euros for 14m ?
16:45Natehaxx: wtf are they crazy
16:46Natehaxx: its like you are paying to live inside a cage
16:46RSpliet: Told you...
16:49Natehaxx: i mean my apartment is 55m
16:49RSpliet: England is both financially and politically a messed up country, worse than most EU members. People here are rightfully proud of their achievements (scientifically and engineering wise), but seem to overestimate their significance and underestimate the contributions of others in these areas. Brexit won't make any of that better, but at least they won't drag the rest of the EU down when/if things do go tits up.
16:51Natehaxx: never tought the situation of UK was so bad
16:51Natehaxx: this is much more worse
16:52RSpliet: It's all relative. They're still significantly wealthier than the bulk of the countries, but politically and wealth-division-wise they're a mini-US more than a European welfare state
16:53RSpliet: Anyway, digressing. Don't mean to make this channel political (sorry, too late now...)
17:05juri_: I run a hackerspace in washington, DC, so i'm really getting a kick out of your replies.
17:05juri_: I drink a lot.
17:07juri_: I'm now at the point of tieing accelerate-hs to nouveau, hense my question above. I'll take a good whack at it, then follow up with more questions.
17:09annadane:wishes nvidia wasn't stubborn
19:53imirkin_: skeggsb: you should disable issues in your github repo
20:20pmoreau: karolherbst: I was on vacation + work meeting for the last two weeks; I’ll try to send the latest version + ping curro for the missing reviews this week/tomorrow, if I’m not suffering too much from the jetlag.
23:29karolherbst: pmoreau: okay, nice. I have some patches on top of yours I want to get merged soonish: https://github.com/karolherbst/mesa/commits/add_clover_spirv_backend_v2
23:30karolherbst: you might want to squash the OpTypeBool into yours though
23:30karolherbst: pmoreau: that is basically everything we need API wise for system fine grained SVM on Pascal+ using HMM
23:31karolherbst: + that callback fix
23:36karolherbst: imirkin: https://bugzilla.kernel.org/show_bug.cgi?id=199693
23:36karolherbst: wondering if this also fixes runpm bugs on nouveau
23:38karolherbst: or at least some system suspend/resume issues at least