04:33dviola: imirkin: I just booted kernel 5.7.6 on that laptop and I still see the failing to init pmu message on dmesg
09:12PhilippGuehring: Hi! Is there any distribution that comes with the most recent Nouveau drivers, or is recommended for building Nouveau from source?
09:14ccr: if you have new kernel and new Mesa, then you have new "nouveau drivers". also maybe ddx if you use Xorg (e.g. the Xorg driver part), but that one does not change much.
11:26Slack_lover: hi, i have GeForce GT 720M on laptop and wanted to program ML with GPU support. i installed nvidia driver i am facing some problems with X. does nouveau support GPU support for ML?
11:42RSpliet: Slack_lover: "ML" support is not a thing. I suspect that whatever application you use for ML uses either Cuda or OpenCL under the bonnet
11:42RSpliet: Cuda isn't supported
11:43RSpliet: OpenCL... don't know. I don't *think* it's supported well upstream, but others may know better
13:27dviola: any ideas about the pmu thing?
23:05dviola: imirkin: I found that if I comment this code: https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/nouveau/nvkm/subdev/pmu/base.c#L105-L110
23:05dviola: the timeout issue is gone
23:06imirkin: hmmmm
23:06imirkin: skeggsb: --^
23:06imirkin: unfortunately i know nothing about any of this stuff
23:06imirkin: skeggsb: dviola gets a pmu timeout on nouveau probe on a GP108 (mobile)
23:22dviola: increasing 2000 to 10000 here also solves the problem: https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/nouveau/nvkm/subdev/pmu/base.c#L107
23:27imirkin: wow, 2s isn't enough?
23:28dviola: let's try 5 :P
23:28imirkin: well, i think 2s is just some astronomically high value
23:37imirkin: skeggsb: as an aside ... wtf is nvjpg?
23:37skeggsb: i have nfi of why it exists, but some kind of "JPEG Engine", GA100 has one...
23:38skeggsb: the GA102/4 i have don't, however
23:38imirkin: seems more fitting on nv1x...
23:38skeggsb: haha yeah, i was a bit "wtf" the first time i seen it
23:38skeggsb: i believe some tegra boards have it too
23:38imirkin: any idea if it's for encoding or decoding? (or both)
23:38airlied: https://developer.nvidia.com/nvjpeg
23:39imirkin: heh
23:39airlied: machine learning
23:39imirkin: when will those machines learn?!
23:39RSpliet: if you combine that with that p2p network-to-GPU tech, you'll never need a CPU again!
23:39skeggsb: OFA is some "optical flow acceleration" engine, nfi what FLA is
23:39skeggsb: ioctrl is related to nvlink
23:43dviola: still getting the timeout with 5000 msecs
23:54dviola: it's timeouting even with 8000 msec...
23:56RSpliet: appreciate you're trying to tune the timeout param, but generally when things time out after a handful of ms, something else is broken
23:56dviola: makes sense
23:56dviola: I see 2000 is used all over the place
23:57RSpliet: and that's normally very generous
23:57skeggsb: it's the "we have nfi what a reasonable timeout is, but we also don't want to spin forever in the kernel" number
23:58imirkin: 2s seems so long though -- i have a hard time imagining something taking longer to process.
23:59skeggsb: as for pmu... i have nfi what it could be doing, we barely understand the interactions there. the only thing that should be running on it for us on that chipset is the PreOS firmware from the VBIOS that manages fans