Current Server Time: 23:52 (Central Europe)

#apertus IRC Channel Logs

2021/11/15

Timezone: UTC


00:49
illwieckz
joined the channel
03:35
Bertl
off to bed now ... have a good one everyone!
03:35
Bertl
changed nick to: Bertl_zZ
07:04
elafon
left the channel
07:04
elafon
joined the channel
08:17
se6astian
good day
11:19
Bertl_zZ
changed nick to: Bertl
11:19
Bertl
morning folks!
15:06
aombk3
joined the channel
15:10
aombk2
left the channel
15:10
aombk
joined the channel
15:13
aombk3
left the channel
15:20
aombk
left the channel
17:00
se6astian
MEETING TIME, who is here?
17:00
Bertl
is here ...
17:03
se6astian
anyone else?
17:03
se6astian
vup/anuejn news to share with the gerneal public?
17:04
se6astian
Bertl: do you have news for us?
17:04
vup
well I can talk a bit about the recent changes for axiom-firmware and axiom-recorder
17:04
se6astian
great, please do
17:06
vup
sure, so let start with axiom-recorder. anuejn and I have been developing it almost exclusively on iGPU platforms, which can access the main system RAM with essentially the same performance as any device local memory (because they actually don't have any device local memory and are using the main system RAM as memory)
17:07
vup
We made extensive use of this fact to avoid unnecessary data copies, for example the output of the debayerer (which runs on the gpu) get written to a buffer that is directly accessible from the gpu and thus the debayered frames can then directly without any copy be passed to further processing (for example writing them to a file or encoding them using ffmpeg)
17:08
vup
As it turns out, these cpu accessible buffers are also available on dGPUs, but are incredibly slow, because they are allocated in the main system RAM and make every access by the GPU go over PCIe to the main system RAM
17:09
vup
so on these we actually have to use device local memory and copy the data from cpu accessible buffers to device local buffers
17:09
vup
this unfortunately means there are more data copies involved, however for some reason this is even faster on iGPUs :)
17:10
vup
on a 64 core system with a 1080 Ti this made a basic `read raw12 -> convert to bitdepth 8 -> debayer` flow go from ~20fps to ~750fps
17:11
vup
and on my laptop it made the same flow + a final rendering of the result go from ~160fps to ~200fps
17:11
vup
these changes are currently still wip and only a few of the processing nodes have been converted, but converting the rest should not be too tricky
17:11
vup
(we are working on it here: https://github.com/apertus-open-source-cinema/axiom-recorder/pull/8)
17:13
vup
finally se6astian and I did a session yesterday to get the new raw mode supported on fw 2.x (https://github.com/apertus-open-source-cinema/axiom-firmware)
17:14
vup
the required changes were actually quite minimal and the integration is basically done, but we are still working on a problem that now occured with the normal (not raw) output :)
17:15
vup
(the integration is tracked here: https://github.com/apertus-open-source-cinema/axiom-firmware/pull/188)
17:15
vup
I think thats it
17:15
Bertl
no relevant news from my side this week ...
17:15
se6astian
many thanks vup, incredible work!
17:16
se6astian
I tried the raw12 sequence playback myself and its absolutely stunning to view it in realtime
17:16
se6astian
quick updates from my side:
17:16
se6astian
I benchmarked the m.2 pcie ssd on a fast pc with direct PCIe connection
17:17
se6astian
results are here
17:17
se6astian
https://docs.google.com/document/d/1Y3UVPmJ3kzAAtcH0plMICJ2-Z7EOZl4OhE2_RcXdz2U/edit
17:17
se6astian
writing zeros is fast
17:17
se6astian
writing actual data starts fast
17:17
se6astian
and then continously drops until we go below the required datarate the Beta supplies
17:18
se6astian
so it seems the SSD is simply unsuitable for our application
17:19
se6astian
I adapted the PCB brackets design as there was an issue with top/bottom cap slightly interfering in the CP enclosure
17:19
se6astian
new design was 3d printed
17:19
se6astian
and installed and now everything fits fine
17:19
se6astian
for me this is the perfect solution already, also for production
17:20
se6astian
had a VC today with the folks at the technical university vienna we want to collaborate with for CP enclosure part manufacturing and optimization
17:21
se6astian
I want to see if we can participate in the virtual FOSDEM and CCC online events some way
17:21
se6astian
CCC CFP ends in about a week
17:21
se6astian
FOSDEM CfP isnt open yet
17:21
se6astian
CCC has local events and local stages though (none in Austria) so the concept isnt 100% clear to me yet
17:21
se6astian
how the event will actually take place
17:22
se6astian
we started looking for someone to help us with color science and raw processing here: https://www.magiclantern.fm/forum/index.php?topic=26299.0
17:23
se6astian
if you know someone who could help here please share
17:23
se6astian
or spread the call whereever it might make sense
17:23
se6astian
GSoC mentor summit was also last week
17:23
yasmin
joined the channel
17:23
se6astian
where quite significant changes for the program next year were announced
17:24
se6astian
https://opensource.googleblog.com/2021/11/expanding-google-summer-of-code-in-2022.html
17:24
se6astian
which will also affect us
17:24
se6astian
thats it from my side I think
17:24
se6astian
anyone else who wants to report news?
17:24
se6astian
yasmin: hello, any news from your side?
17:25
vup
se6astian: is this the only m.2 ssd you have tested so far, or did you test others already (lost the overview a bit in the ssd testing saga :)
17:25
yasmin
hello! i'm still working on my task (updating the camera HDL to 2021 version) i faced an error which i'm trying to figure it out (after my midterms this week)
17:26
se6astian
yeah its indeed quite confusing, I tested a few already but few natively connected only on rockpi and now this test machine
17:26
se6astian
in USB3 enclosure results are completely different
17:27
se6astian
ordered a 2017 NUC and lattepanda Alpha 864s
17:27
se6astian
which also have native PCIe NVME SSD support
17:27
vup
right, was just trying to correlate a few data points with the SusWrite benchmark here: https://ssd.userbenchmark.com/
17:28
se6astian
the PNY CS3030 has an advertized 3000MB/s write speed...
17:28
vup
well yes, but under which benchmark
17:28
se6astian
the problem seems to be that every benchmark does things differently
17:28
se6astian
some just write a few GB of zeros
17:29
se6astian
I didnt find any benchmark that does pseudo random data writes over the full SSD capacity
17:29
se6astian
thats what Bertl benchmark tool does
17:29
vup
anandtech has some benchmarks they call whole-drive sequantial write, for example here: https://www.anandtech.com/show/13512/the-crucial-p1-1tb-ssd-review/7
17:30
vup
i guess recording similar graphs like they did there would be interesting aswell
17:31
se6astian
thanks
17:31
se6astian
will see what NUC/lattepanda reach with this SSD but I guess as next step will source another SSD that has very high benchmark results on anandtech
17:32
se6astian
samsung 980 PRO also looks promising
17:33
se6astian
rockpi doesnt support all lanes and not newest PCIe so its possible even high end ssds perform poorly on it
17:33
se6astian
NUC should get close to real performance limit though I hope
17:33
vup
yeah iirc the rockpi benchmarks were a bit weird
17:33
se6astian
thanks for the update yasmin!
17:34
se6astian
anyone else with updates/news?
17:35
se6astian
then many thanks all participants, MEETING CONCLUDED!
17:35
Bertl
thanks for moderating!
17:36
yasmin
Thanks
17:36
yasmin
left the channel
17:49
Bertl
off for now, bbl
17:49
Bertl
changed nick to: Bertl_oO