12:27 pmoreau: imirkin, karolherbst: Some feedback from daniels regarding the pages: no preview in the browser apparently (if you edit locally, you can easily build the pages and view them locally, and some generators support rebuilding automatically on file save, giving some easy previewing).
12:28 pmoreau: Regarding permissions: once someone create an account, we can decide to have them join the nouveau team and give them whatever rights we deem useful.
12:29 pmoreau: Also, as it is a regular git repo, people can clone the wiki and send pull request, that way we don’t need to give them special rights and have them join the team (probably best for one off contributions).
13:14 karolherbst: pmoreau: without online preview it is useless
13:15 karolherbst: there has to be something, no?
13:16 karolherbst: thing is, if it requires users to install random software just to edit a wiki, they won't do it
13:16 karolherbst: then we don't need a wiki at all
13:20 pmoreau: You don’t always need to know how it will look like, if you are just fixing a typo for example, or adding a chipset to the table, so in those cases you just edit the file in the browser or locally and be done with it.
13:20 pmoreau: But for bigger modifications, you would want a preview for sure.
15:54 imirkin: skeggsb: btw, we need those firmware calls to do shader trap handling
15:59 pendingchaos: lachs0r: why does https://0x0.st/sBfK.txt repeat subtest results?
17:12 lachs0r: pendingchaos: the tests print their results every time something happens to the window (input, resize, whatever)
17:12 lachs0r: wasn’t intended
17:12 imirkin: should be able to run with -fbo -auto
17:16 pendingchaos: ah
19:01 karolherbst: imirkin: what would we have to do if we want to access sys memory through gmem access prior to Pascal?
19:28 HdkR: karolherbst: Does it act any differently? :P
19:29 karolherbst: HdkR: for one, you would have to sync the CPU and GPU MMU
19:29 karolherbst: but I am not quite sure if any shader can just access system memory through ld.g
19:30 karolherbst: HdkR: starting with Pascal the GPU can fault on memory access and report it to the host, so that we can copy memory into VRAM and continue
19:30 karolherbst: before that it doesn't work that way
19:31 HdkR: Ah right, you have to ensure the memory is visible to the GPU
19:34 HdkR: I see, the cuda docs also explicitly state that SM 6.x is required for on-demand page migration and GPU memory oversubscription
19:35 karolherbst: HdkR: yeah, I am talking here with openmp in mind though
19:35 karolherbst: I am sure we could do a subset of features, but...
19:35 HdkR: Right, cuda docs just give insight :P
19:36 HdkR: (Also have been working with cuda exclusively for the past two weeks, so it is on the brain)
19:36 karolherbst: imirkin: uhh, it just hit me, with Pascal+ we could in theory report VRAM + sys RAM as the total VRAM, because we can just move memory around as we please. Would be interesting to figure out how that works for non compute work loads
19:36 karolherbst: and VRAM is just a faster region of memory
21:09 skeggsb: karolherbst: it doesn't work at all for non-compute stuff, there's only a couple of GPU units that support replayable faults
21:09 karolherbst: skeggsb: ahh, okay
21:09 karolherbst: skeggsb: but at lost for those units we could swap to system memory without much troubles, right?
21:09 skeggsb: so, unfortunately, we can't use it to do nicer stuff than manual usage tracking / validation like we do in ttm
21:09 karolherbst: *least
21:10 karolherbst: but yeah, it might be not working well with non compute workloads anyway
21:11 skeggsb: that's kinda what we're going to get with hmm, so, yes?
21:12 karolherbst: I was just wondering if we want to also use it for graphics work loads
21:14 skeggsb: i'm not even 100% sure that the replayable faults work outside of the compute class.. it's gcc and tex on gr that support replayable faults, so perhaps, but all the other gr units don't, and i don't know if they've put some check in hw to disable them for !compute class even for gcc/tex
21:14 skeggsb: there's even some minor stuff on the compute class that can't fault, on pascal at least
21:15 karolherbst: ohh, I see
21:16 skeggsb: so yeah, we're basically limited to the narrow scope that hmm will enable for now
21:16 skeggsb: which is a little disappointing, but, yeah