07:37pmoreau: karolherbst: I just saw curro reviewed the series; I'll go through it tonight. It doesn't look like 19.2 has been branched off yet, was it been delayed?
15:14imirkin: airlied: in case you didn't see: https://patchwork.freedesktop.org/patch/324543/
19:09airlied: imirkin: did see, and then fell out of my brain
19:09airlied: thanks for that, might have to kick mareko to look
19:14imirkin: he normally looks at stuff within 24h or so
19:14imirkin: figured i'd kick him tomorrow if he hadn't reviewed by then
22:18mattst88: imirkin: that iMac G4 I mentioned to you ages ago -- I finally got it installed :)
22:19mattst88: nouveau works just fine on it, after I specify video=... on the kernel command line
22:19imirkin: mattst88: nice! it's one of the ones with a NV17 in it, right?
22:19mattst88: I think nouveau isn't able to figure out the native resolution from the vbios/open firmware/whatever apple crap
22:19mattst88: FX5200, whatever that is
22:19imirkin: oh ok. NV34.
22:19imirkin: that's twice as much as a NV17...
22:20mattst88: and still crappy enough that it can't do 1080p :)
22:20imirkin: that's actually a bit surprising ... it should definitely be able to
22:20imirkin: My TNT2 was able to drive 1920x1200 no problem (over VGA no less)
22:20mattst88: (was looking into replacing the panel, but... no point if the graphics card can't even do the higher res!)
22:21imirkin: iirc i bought a FX5200 (which was not super-old at the time of such purchase) coz i could decode 1080i ATSC with it (whereas my CPU-du-jour could not keep up with such a high rate mpeg stream)
22:21mattst88: well, I haven't tried personally, but googling "nvidia 5200 max resolution" turns up https://www.nvidia.com/page/pg_20040109440047.html which says "resolutions up to and including 1600×1200"
22:22imirkin: yeah, that's coz 1920x1200 wasn't really a thing when those docs were produced
22:22mattst88: huh, it's actually a little confusing. says "including 2048×1536" above, and then the stuff about 1600x1200 below in the context of DVI...
22:22mattst88: the built-in panel is 1440x900 TMDS, so I guess that's DVI really
22:22imirkin: LVDS probably
22:23mattst88: no, it's actually TMDS :(
22:23mattst88: makes panel replacement much harder
22:23imirkin: certainly ain't HDMI, i can tell you that!
22:23mattst88: would max resolution with DVI actually be lower than with VGA?
22:24imirkin: NV34 has a max of 135mhz tmds, iirc
22:24mattst88: huh, weird. googling seems to confirm :(
22:25mattst88: weird. I had no idea this was a thing
22:25imirkin: prior to nv18, it was 112mhz :)
22:26imirkin: (not a ton of dvi boards in that group though)
22:26imirkin: your chipset is 0x34
22:28imirkin: 1600x1200 fits into 135mhz
22:28imirkin: but 1920x1080 is JUST outside ... 138.5mhz
22:28imirkin: you can futz with the limits, it's just hardware
22:28imirkin: but if you get snow, don't complain here
22:29imirkin: whereas if you do VGA, you could definitely do 1920x1080, but depending on the panel and cable, could look worse
22:29imirkin: i am a bit curious why you need the video= override
22:30imirkin: maybe the internal screen is missing an edid, and we don't pick it up from the vbios either (which has an area to indicate such things)
22:36imirkin: mattst88: also please note that GL rendering with BE is pretty fubar... it works enough for glxgears, maybe a few other things, but once you start using buffer objects, it all goes downhill pretty fast
22:37mattst88: yeah, I can get you a dmesg log of the fail-to-detect-edid stuff if you're interested
22:37mattst88: I'm under no illusions about GL on big-endian, don't worry :)
22:37imirkin: well, if you can confirm it fails to detect edid, i'd be more interested in the vbios
22:37imirkin: the BE fail is basically a failure in my ability to reason about what data is in what endianness where
22:37mattst88: okay. easy way to capture that? don't we link it in sysfs somewhere?
22:37imirkin: and where all the various flips happen
22:37imirkin: yea, /sys/kernel/debug/dri/0/vbios.rom
22:38mattst88: okay, I'll grab it tonight and email it to you
22:38imirkin: thanks. no promises of an actual investigation, but the commitment from you is small enough...
22:39imirkin: if you use X with it, make sure you're on xf86-video-noueau. if you're using wayland, you won't get acceleration at all coz no ES2
22:40imirkin: (or GL2)
22:42imirkin: ES2 could be achieved by someone who cared
22:42imirkin: GL2 requires mucho fallbacks
22:43airlied: I also envisage endianness requiring a large whiteboard
22:43airlied: I had to use a large whiteboard for figuring out viewport flips with virgl
22:44imirkin: problem is a few-fold
22:44imirkin: problem 1 is deciding what endianness "GART" memory should be -- LE or BE
22:44imirkin: problem 2 is understanding where flips happen
22:45imirkin: problem 3 is tracking down the correct element width for the endian flips
22:45airlied: yeah cpu maps of vram also are painful :-P
22:45imirkin: problem 4 is once you have that width, to track down the undocumented transfers logic for where to set that
22:46airlied: and all GPU views are BE works until you engage GPU swappers
22:46imirkin: yeah, vram maps have to just be disallowed obviously
22:46airlied: at which point some GPU views are BE
22:46airlied: and some are LE
22:46airlied: and then you can't find the magic swapper bits for one of them
22:46airlied: I think that's where I have up last time I tried
22:46imirkin: yeah, having an inventory of what swapping the gpu is able to do would be super-nice
22:47imirkin: right now we enable it so that we feed it BE words in pushbuf
22:47imirkin: and MMIO gets auto-flipped
22:47airlied: I think the first thing would be to just always treat the GPU as LE and use no swapper hw ever
22:47airlied: and deal with the pain until you understand it
22:47airlied: then slowly add back swapper support
22:48imirkin: of course this would require caring
22:48airlied: yeah I also ran out of that
22:48imirkin: glxgears works, ship it, done!
22:48airlied: hey llvmpipe works, screw hw :-P
22:48imirkin: oh, that's another fun one --
22:49imirkin: some stuff requires gallivm vertex processing
22:49imirkin: have to make sure THAT data gets fed in properly too
22:49imirkin: and then the newer GL variants are just fucked
22:49imirkin: like i don't think you can reasonably do SSBO in a mixed endian situation
22:50imirkin: certainly not with persistent/coherent maps :)
22:50airlied: yeah coherent buffers in mixed endian sounds awaful
22:51imirkin:is waiting for the day where someone does pci passthrough of an intel gpu pci device into a softmmu be cpu vm...
22:51airlied: imirkin: just wait until someone plugs a discrete intel into a g5 :-p
22:52imirkin: the i740? :p
22:52airlied: I think there are plenty of news sites reporting things now
22:53imirkin: i was attempting to be funny. attempt ... unsuccessful.
22:53airlied: imirkin: oh I got it, I just have to tiptoe along my ndas :-P
22:54imirkin: would never (purposefully) try to get you to reveal privileged info
22:57airlied: imirkin: requires beer :-P
22:57airlied:wonders if I could write "except while drinking" into NDAs