06:28 austriancoder: imirkin_: okay - I think I will have a look at the kernel side too to fully understand it. could be the same achieved with a fence attached to the cmd buffer?
08:22 pmoreau: PyroSamurai: No worries, I haven’t had time to do much either. Enjoy the graduation ceremony! :-)
08:48 dboyan_: imirkin, hakzsam: Any idea what that nouveau_pushbuf_space at https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nvc0/nve4_compute.c#n658 means?
08:48 dboyan_:doesn't know a lot about bos
09:47 hakzsam: dboyan_: I don't remember this detail
09:49 hakzsam: but I think it's something related to pushbuffer flushing (ie. it reserves enough space in advance to be sure it won't be flush in between or something)
10:03 dboyan_: yeah, I think I gotta modify that section to get indirect dispatch working, but I'm not really sure about its details
10:12 hakzsam: you only need to update the different offsets and allocate more space if the launch desc is bigger
10:12 karolherbst: you also need to reserve enough space to fences
14:57 dboyan_: imirkin: With https://hastebin.com/hovosibaga.diff, dEQP-GLES31.functional.compute.indirect_dispatch all passed
14:57 dboyan_: imirkin: I'm wondering if there are better ways to organize code
14:58 dboyan_: Also 'if (nvc0->screen->base.device->chipset >= 0x130) {' is occuring 3 times in nve4_launch_grid, which I'm not a fan of.
15:00 hakzsam: dboyan_: put this into a new function and pass the offsets as parameters
15:05 dboyan_: hakzsam: Can I move the nouveau_pushbuf_space and PUSH_REFN forward? It's standing out in between the first upload (nve4 case has 2 uploads).
15:14 hakzsam: dboyan_: you can't do that. But the only thing which changes is the offsets, right?
15:15 hakzsam: so instead of using magic numbers, you can do something like if (pascal) offset = 64; else = 48 or something
15:19 dboyan_: well, the problem is that the starting offset are the same. but nve4 uploads two times to fill in GRIDDIM_YZ the second word while pascal needs only once
15:40 imirkin_: dboyan_: don't check chipset, check the 3d or compute class. e.g. class => GP100_COMPUTE or whatever
15:55 dboyan_: imirkin_: Do you think https://hastebin.com/pecifuxadu.diff is any better than the previous one?
15:56 dboyan_: [I haven't tested this one]
15:58 imirkin_: the PUSH_REFN should be done unconditionally
15:59 imirkin_: also i'd rather you do a class check, not a chipset check
16:00 dboyan_: I will fix the class check later
16:00 dboyan_: but PUSH_REFN is done only once in the previous code
16:00 imirkin_: heh
16:01 imirkin_: well, technically your way will work
16:01 imirkin_: but before it was all one sequence of commands
16:01 imirkin_: whereas now the function is logically a single sequence of commands
16:01 imirkin_: which is invoked twice
16:01 imirkin_: the pushbuf_space/PUSH_REFN are required for that logical sequence of commands
16:02 dboyan_: okay, so you mean only pushbuf_space should be executed conditionally?
16:05 imirkin_: nothing should be conditional
16:05 dboyan_: so just drop the 'refn' flag should be okay?
16:06 imirkin_: yup
16:08 dboyan_: okay, I'll fix it up and do some testing tomorrow because I'm going to sleep soon. I'll send an updated 3/4 after that.
16:10 imirkin_: sgtm
17:23 rubdos: Is there a Nouveau dev that I could make happy with a T61p with a quadro card, for the purpose of rev-eng?
17:23 rubdos: I have one myself, and I could get another one very cheaply. I/my company would pay for it.
17:24 rubdos: `VGA compatible controller: NVIDIA Corporation G84GLM [Quadro FX 570M] (rev a1)` is the thing in question
17:27 leberus: karolherbst: Hi. I'm writing v5 now and I have a question. You told me that we don't want to support writing in temp attributes. This rule also applies to auto_point_temp attrs?
17:39 leberus: or do we want to keep those attrs as they are right now?
18:02 karolherbst: leberus: I would say we rather keep them as they are, because we don't do much with them anyway. If somebody get ridiculous low value, they should report a bug and we fix that. Controlling the fans is a different thing and is okay if done by the user
18:03 karolherbst: but changing the temp limits has like no effect anyway afaik
18:28 Lyude: neat, clockgating patch should be ready to go. seems to save a little more power on kepler then it does fermi as well
18:29 karolherbst: Lyude: yeah. nice
18:29 Lyude: will submit in just a little bit
18:46 Lyude: btw, just curious, I am assuming this is turning on the ENG_PWR bit on things that aren't getting used? https://trello.com/c/Jd4FLlqf/43-pm-add-power-gating-support
18:48 karolherbst: Lyude: it isn't merged yet :p
18:48 Lyude: oh, didn't realize I should keep things on the board until then
18:48 Lyude:fixes
18:49 karolherbst: will comment tomorrow evening most likely, maybe today
18:53 hakzsam: imirkin_: any thoughts about skipping the merge registers pass for nv50/nvc0?
19:17 PyroSamurai: rubdos: since you are wanting it RE'ed, what features are you missing/wanting from it?
19:17 rubdos: Well, at least, it shouldn't crash when I'm using pronterface :')
19:17 rubdos: I'll come back tomorrow, gotta spend time with the SO now.
19:18 rubdos: I'm on a broker, so feel free to lecture me while I'm gone
19:18 rubdos: I read backlogs
19:18 PyroSamurai: rubdos: okay, I'll check in tomorrow with you.
21:03 karolherbst: done :)
21:09 karolherbst: I've put up the source code of the presentation here: https://github.com/karolherbst/nouveau_presentations
21:29 mupuf: karolherbst: how did it go?
21:31 karolherbst: very good
21:32 karolherbst: I wanted to do 20 mins, I did nearly 50 :D
21:32 karolherbst: most because of questions
21:32 karolherbst: I am quite sure there will be follow up talks with more specilized topics
21:32 karolherbst: a lot of interest in the crypto falcons :p
22:14 Lyude: imirkin_: working on https://trello.com/c/kuMPh88f/160-gm200-arb-post-depth-coverage again and I am still stuck on that GL test, for the life of me I can't figure out how the multisampling test for this extension in piglit was ever supposed to actually work correctly in the first place :\
22:36 karolherbst: Lyude: reviewing your patches now :p
22:47 Lyude: karolherbst: if you might know anything on why this test passes unconditionally (with or without the GL extension enabled) btw, I'd really appreciate it https://cgit.freedesktop.org/piglit/tree/tests/spec/arb_post_depth_coverage/multisampling.c#n269
22:47 karolherbst: Lyude: I think you can reduce the CC list drastically :P usually it is enough to send to the Nouveau ML and if you think it is mergeable to skeggsb. airlied can't do much with it except reviewing. Same goes for everybody else, but they all read the Nouveau ML as well.
22:48 Lyude: ah, I'll remember that for kernel patches in the future
22:48 karolherbst: well, skeggsb merges those patches inside his tree
22:48 karolherbst: at some point
22:48 karolherbst: so if your patches don't get merge, ping him daily until he does
22:48 karolherbst: *merged
22:49 karolherbst: sent
22:49 karolherbst: mhh, I could have cleared the CC list as well... sry
22:50 Lyude: np
22:50 karolherbst: mhh I think having func pointers inside the public structs is super bad style and we shouldn't do it
22:50 karolherbst: therm is kind of a bad example for how things should be
22:59 karolherbst: mupuf: to get into more details, because I was in a hurry to leave the place there: it was very nice, there were interesting questions, more than I expected. Nice discussions between inside the audience as well. It was nice, and actually covered enough to get a broad overview I think. Showed them a few tools, made a few jokes, it was fun
23:00 karolherbst: And they actually seemed to care about nthe crypto situation as well
23:00 karolherbst: and also tried to discuss ways of getting stuff done