#
✕ nick:marcan
#asahi · 2023-12-23 17:47
marcanpossible in principle though, particularly with vulkan
#asahi · 2023-12-23 17:47
marcanwould need code in the compositor to proactively garbage-collect unused textures out of actual GPU memory to improve that, but I doubt it would be very useful
#asahi · 2023-12-23 17:46
marcanright, in that case it still counts of course, as long as those textures are mapped
#asahi · 2023-12-23 17:45
marcanthe compositor needs access to the primary framebuffers of everything it's compositing
#asahi · 2023-12-23 17:45
marcanin that case no, but if the compositor is actively using it then you wouldn't be able to compress it anyway
#asahi · 2023-12-23 17:44
marcanthe point is just that the moment there is GPU work for that app/context, all of it has to be swapped back in
#asahi · 2023-12-23 17:44
marcanif there is any shared memory then that wouldn't be swapped out if someone who needs it has it mapped
#asahi · 2023-12-23 17:43
marcaneither way sharing memory is not an issue, you don't actually swap out whole apps. what I mean is that apps are either suspended (some memory swapped out) or not (no memory swapped out).
#asahi · 2023-12-23 17:43
marcanso at that point it's irrelevant and the original shared memory is normally swappable
#asahi · 2023-12-23 17:42
marcanif the shared memory is not GPU memory to begin with, then it needs to be copied into GPU memory
#asahi · 2023-12-23 17:41
marcanwhat do you mean?
#asahi · 2023-12-23 17:39
marcanyeah
#asahi · 2023-12-23 17:39
marcanso it's useful for whole processes/contexts stay idle, not for idle memory within a process
#asahi · 2023-12-23 17:38
marcanif an app has to render anything 100% of its GPU memory needs to be available
#asahi · 2023-12-23 17:38
marcanit is, but it involves essentially (potentially) swapping out whole apps at once
#asahi · 2023-12-23 17:37
marcanwe don't have swapper support for the GPU yet so it all stays pinned in memory
#asahi · 2023-12-23 17:37
marcanno(t yet)
#asahi · 2023-12-23 17:36
marcanswap was added literally the day before release ;)
#asahi · 2023-12-23 17:30
marcanwe do enable proper swap in fedora by default these days, so yeah
#asahi · 2023-12-23 17:29
marcanhowever I think in your case it might be zram backing too, not sure how that works
#asahi · 2023-12-23 17:29
marcankidplayer666: magenta is not cache, it's shared memory which, if you don't have massive stuff in tmpfs, usually means GPU memory.
#asahi · 2023-12-23 17:27
marcanpitust: I'm talking about the routes in alsamixer, but please just use pipewire. the speakers are not designed to be driven directly like that, they need DSP to work properly.
#asahi · 2023-12-22 17:42
marcanit's not useless, you can use it for rendering too
#asahi · 2023-12-22 17:42
marcanbasically any render-only device still gets a cardX device
#asahi · 2023-12-22 17:42
marcanit's complicated(TM)
#asahi · 2023-12-22 16:43
marcanpitust: your ALSA routes are probably wrong. this is handled by UCM if you use pipewire, and you should be using pipewire.
#asahi · 2023-12-19 17:22
marcanit just backends to Accelerate/Metal I think, so not terribly useful
#asahi · 2023-12-19 16:48
marcansgtm
#asahi · 2023-12-19 16:47
marcan(remove = make it identical to non-edge)
#asahi · 2023-12-19 16:47
marcanI'm going to remove -edge from ALARM as soon as the next kernel update round hits, probably in January, so yea
#asahi · 2023-12-19 16:46
marcan(there briefly was but it's been long gone)
#asahi · 2023-12-19 16:46
marcanit's just one build
#asahi · 2023-12-19 16:46
marcanthere is no edge for fedora
#asahi · 2023-12-19 16:45
marcanpretty sure there's some open source stuff for that these days, would be neat to run it on my home Minis
#asahi · 2023-12-19 16:45
marcanI'm actually curious whether I can use ANE to run some kind of voice recog/assistant model to replace Google Home (which I use as a Home Assistant frontend only these days)
#asahi · 2023-12-19 16:44
marcanshould be a nice combo once we have both :)
#asahi · 2023-12-19 16:44
marcanalso I hear OpenCL on the GPU is not far off (which is more broadly useful for ML training and things that use fp32 and whatnot, while ANE is for more limited but efficient use cases)
#asahi · 2023-12-19 16:42
marcanI'm not sure what the long term approach ought to be, but you can already run useful models AIUI
#asahi · 2023-12-19 16:42
marcanydalton: eiln has a basic python toolset for "compiling" for ANE
#asahi · 2023-12-19 16:40
marcansince we need that for ISP integration
#asahi · 2023-12-19 16:40
marcanand moving to using the firmware
#asahi · 2023-12-19 16:39
marcanonly released features count :) (but I'm looking forward to merging that one, there's fun to be had there)
#asahi · 2023-12-19 16:39
marcan(probably lots more I forget)
#asahi · 2023-12-19 16:38
marcanand also lots of general distro integration stuff (like no more -edge so those features by default, the new swapfile setup thing, the fancy widevine-installer hack for netflix/spotify/etc, the openh264 trick aka we're the first Fedora spin to actually have the special lawyer-approved h264 codec out of the box AFAIK)
#asahi · 2023-12-19 16:37
marcaneiln: also you may want to mention this in -dev :)
#asahi · 2023-12-19 16:37
marcantoo many to count lol
#asahi · 2023-12-19 16:37
marcanM2 especially desktop support (which was tied to HDMI)
#asahi · 2023-12-19 16:35
marcan(hwmon is not in)
#asahi · 2023-12-19 16:34
marcanyes
#asahi · 2023-12-19 16:34
marcanI need to keep it in the VM at least another year, it deserves to be of drinking age (in japan) ;)

plus ›