00:18 <alyssa> Oh, this is odd -- top shows 2710M free, 388M used, 553M buf/cache, and 1024M unused swap

00:18 <alyssa> yet CREATE_DUMB is still failing due to supposed OOM

00:18 <alyssa> bbrezillon: robher: ^^

00:19 <alyssa> I've only ever seen this on Bifrost

00:24 <robher> CREATE_DUMB is on the KMS device, so could be CMA underneath.

01:46 <chewitt_> Amlogic device-tree sets 256MB CMA by default but for 4K things, at least for vdec, the numbers need to be a lot larger

01:47 <chewitt> I have my kernels hacked to 896MB

02:27 <alyssa> chewitt: ahhhh

02:27 <alyssa> so not a panfrost issue?

02:28 <chewitt> narmstrong is the person who understands how it all hangs together

06:04 <icecream95> macc24: Have you fixed all of the bugs with 5.11 yet or will I have to try and backport audio to 5.10?

07:23 <narmstrong> alyssa: yep probably CMA is too small

10:20 <macc24> icecream95: "backport the audio"

10:20 <macc24> ?

10:21 <macc24> are you implying that audio is working in 5.11?

10:22 <icecream95> macc24: Only with the usb-c adapter, the speakers don't work

10:22 <macc24> kernel version doesn't matter for usbc adapter

10:23 <macc24> it's literally usb c audio card

10:24 <icecream95> I never got it working on 5.10, but maybe that was just because of the config or not applying the right patches

10:24 <macc24> config

10:24 <macc24> Works for me™

10:25 <icecream95> So I've been enduring all of those glitches and not being able to check the battery charge for nothing?

10:25 <macc24> yep

10:26 <macc24> should've been using cadmium on duet

13:26 <alyssa> narmstrong: ack, thanks

13:27 <alyssa> does it make sense to bump itupstream?

13:28 <macc24> chewitt: CMA? is it like "max vram size" on intels?

13:29 <urjaman> not really... it just makes sure that it's possible to make physically continuous allocations from that area

13:31 <urjaman> like it's not "vram" in the sense that it's managed by the kernel (not dedicated to the graphics adapter) and kernel can stuff other things there too (like page cache or such i think atleast...) and other non-graphics hardware can also use it if they happen to need such allocations

13:31 <macc24> so it's kinda what i thought it was

13:52 <HdkR> Fosdem is live, I missed some things :O

13:55 <narmstrong> alyssa: probably yes, it was calculated before panfrost and video decoder... but we have some boards with 512m ddr so only the new gen should have more

13:56 <alyssa> libmali doesn't have this issue?

17:41 <KungFuJesus> narmstrong: does the video decode fixed function stuff work yet on amlogic S922x?

17:42 <narmstrong> KungFuJesus: what do you mean ? H264, mpeg2 and vp9 should work, but bugs subsists

17:42 <KungFuJesus> awesome, what userspace APIs exist to access them?

17:42 <KungFuJesus> someone should write a libva driver for it

17:43 <narmstrong> KungFuJesus: V4L2 m2m, with ffmpeg or gstreamer

17:43 <KungFuJesus> ah, that explains perhaps the weird error message I saw when trying to do --hwaccel=gpu with mpv, it was some V4L based memory allocation failure

17:44 <macc24> narmstrong: huh? not v4l2 request api?

17:45 <narmstrong> macc24: nop it’s a statefull decoder

17:45 <macc24> narmstrong: cool

17:45 <narmstrong> KungFuJesus: this may be because of too small cma

17:46 <KungFuJesus> is that something at compile time in the kernel?

17:47 <KungFuJesus> and I guess the question for the holy grail of desktop usability - can the ffmpeg compiled and bundled with chromium on archlinuxarm make use of this API?

17:49 <KungFuJesus> oh awesome, archlinuxarm seems to have gone through the trouble of debundling ffmpeg from chromium so that it uses the system one

18:38 <KungFuJesus> narmstrong: modified the dtb from some googling, thanks for the tip, seems to be working

18:38 <KungFuJesus> except seeking is a little meh, hard to know if that's an NFS problem with the wonky realtek nic or not

20:59 <macc24> lol

20:59 <macc24> retroarch lags on G31 at 854x480

21:00 <HdkR> The UI?

21:00 <macc24> xmb ui

21:00 <macc24> 20fps

21:00 <macc24> with "simplified" ribbon

21:01 <macc24> attempting to change it resulted in 'panfrost ff400000.gpu: gpu sched timeout, js=0, config=0x7300, status=0x8, head=0x4468040, tail=0x4468040, sched_job=0000000032303e37'

21:01 <HdkR> Might be worth trying Ludo instead

21:01 <macc24> i prefer xmb ui

21:02 <HdkR> Sure, but prefer 20fps? :P

21:02 <macc24> well

21:02 <macc24> i am not releasing until panfrost gets optimizations

21:32 <KungFuJesus> hmm, appears the inability to seek is probably v4lm2m meson bug related

21:33 <KungFuJesus> a shame but it's _kinda_ working. CPU utilization goes way down. Needed for VP9, libVPX doesn't do so hot it would seem. But H.264 seems reasonable with the CPU

21:33 <KungFuJesus> eats 20-30% on a few cores, but it's not horrible

21:34 <KungFuJesus> I suspect this will eventually improve, as rpi can leverage this API as well

21:34 <KungFuJesus> though mmal appears to be the preferred method for accelerating video on it

21:41 <KungFuJesus> icecream95: I suspect you or alyssa might know: how much potential performance is left to unlock with bifrost with regard to driver level optimizations? Are we anywhere near peak glx performance or is there still a lot of hacks sitting in place before real optimizations go in?