#
#asahi · 2026-01-21 08:39
chadmed_outside of the context of like, having the shader compiler recognise certain patterns in the shader and substituting them for hardware-optimised routines
#asahi · 2026-01-21 08:35
chadmed_regardless i think this is a bad example for the application of linear algebra to driver authorship :P
#asahi · 2026-01-21 08:35
chaos_princessfrom what i understand - yes. Maybe also power as you don't need to light up all the multipliers, but at least for a systolic array, i don't see how sparsity would help with speed
#asahi · 2026-01-21 08:33
chadmed_thats basically a memory optimisation for sparse matrices right, so we dont have to worry allocate memory for all the elements which are zero
#asahi · 2026-01-21 08:27
chaos_princessfrom looking around - i don't think anything does diagonalization, but some tensor cores have support for "sparsity", where you store a bitmask of which elements are zero.
#asahi · 2026-01-21 05:00
kitlithsome amount of linear algebra may be useful for optimizations done in the shader compiler, but the specific optimization you're mentioning, involving diagonalization of a matrix, feels more like the domain of an optimization a shader author might apply. It is a different algorithm, if that makes sense.
#asahi · 2026-01-21 04:57
kitlithbigmac: the rough division i'd expect is that the kernel space driver talks to hardware, the user space driver does compiliation (general purpose optimization), and the shader author does algorithm design (special purpose optimization).
#asahi · 2026-01-20 18:36
bigmacHm, when writting userspace or kernelspace gpu drivers, are concepts of linear algebra used? Or perhaps a better question, does one *benefit* from knowing the topics of university level linear algebra?
#asahi · 2026-01-20 17:52
chaos_princessunless we are talking tensor cores and similar, matrix multiplication does not exist in hw, writing one in a gl shader just emits a pile of fmas
#asahi · 2026-01-20 17:50
bigmacIs this something that'd be touched upon only *on* the hardware, or can it be the work of a driver?
#asahi · 2026-01-20 17:49
bigmac..but for this to work the computer would have to (probably) take track of which matrices are singular. Or it would have to prevent creating them in the first place
#asahi · 2026-01-20 17:45
bigmacSay there would, for some reason, have to happen a multiplication of the same matrix many times in a row. Then the procedure of diagonalisation might be worth doing instead of doing the n^3 (or whatever algorithm is used deep down) matrix multiplication procedure
#asahi · 2026-01-20 17:43
bigmacHmm I got curious about GPU matrix diagonalisation while studying linear algebra today. (slightly offtopic for this channel, apologies). Would anyone be able to point me onto resources where I could read about how the computer decides whether a matrix couldn't be diagonalised first before performing operations with it?
#asahi · 2026-01-20 15:52
bigmacwonders whether someone from Prague is here
#asahi · 2026-01-20 11:07
culmanso I'm here
#asahi · 2026-01-20 11:07
culmanI was told to drop by, by the website
#asahi · 2026-01-20 11:07
culmanhello
#asahi · 2026-01-19 16:35
fl0_idalso depends on your budget, m3 might be fun, but also a bit more expensive than m1 or m2 probably. So if you find sth you would like to wokr on for m1/m2, might be cheaper
#asahi · 2026-01-19 16:33
hbockyeah, that's probably not the best starting point for me.
#asahi · 2026-01-19 16:33
chaos_princessi personally would suggest avoiding m4 for now, unless you want to go right into the deepest end of the pool
#asahi · 2026-01-19 16:32
hbockgreat, thanks for your time! i'll ponder that later today.
#asahi · 2026-01-19 16:31
chaos_princessiboot1 lives in spi flash (probably qspi, but don't remember)
#asahi · 2026-01-19 16:30
hbockspi flash, does it boot from NOR/QSPI or something?
#asahi · 2026-01-19 16:30
chaos_princesswhatever machine you prefer, consult the support tables, find something missing that you are interested in, go for it
#asahi · 2026-01-19 16:30
hbockhappy to start with documentation, "walk before you run" etc.
#asahi · 2026-01-19 16:30
hbockokay. i'll see what i can find for a reasonable price. if i could be of use, what is something you'd like to prioritize?
#asahi · 2026-01-19 16:29
chaos_princessthere is a low level recovery system in the chip rom that can recover from a disk and spi flash wipe
#asahi · 2026-01-19 16:29
hbockfamous last words :)
#asahi · 2026-01-19 16:28
chaos_princessshould be impossible as long as you have a second mac or a linux machine
#asahi · 2026-01-19 16:28
hbockas long as it's relatively hard to brick... that's fine. i spend most of my real job building yocto and reflashing ARMv7/ARMv8 boards anyway.
#asahi · 2026-01-19 16:28
chaos_princessand m4 is "hard" due to reasons, m3 is easier
#asahi · 2026-01-19 16:26
hbockoooh. that's fun.
#asahi · 2026-01-19 16:25
chaos_princessinstaller is for cases where it is "user facing", for development installs, you install a second copy of macos and overwrite it's kernel
#asahi · 2026-01-19 16:25
hbockit didn't seem like there was any M4/air support for the installer, from what i read this morning. unless i misunderstand the table.
#asahi · 2026-01-19 16:24
hbocki figured. i imagine being able to boot back into mac os is critical for reverse angineering anyway.
#asahi · 2026-01-19 16:22
chaos_princessyour dev machine is still useful for other things, the dev/experimentation/etc environment is just yet another os install that you boot into
#asahi · 2026-01-19 16:19
hbockit would be used for other things, not just a dev toy. but maybe it would be better to buy a nice one for me, and an older mini to "sacrifice" for asahi dev that's more pressing.
#asahi · 2026-01-19 16:13
chaos_princessi would suggest not buying pretty expensive computers just for this unless you can find some other use for them
#asahi · 2026-01-19 16:12
hbockShiz: i'm up for anything, but my skillset is primarily embedded Linux/QNX, board bringup, driver development, etc. I've done some reverse engineering in the past, but not on this level. Familiar with a handful of subsystems (I2C, SPI, USDHC/SDIO, UART, MII) and ARM Cortex-A/M stuff. Basically, a little bit about a lot of stuff, but not really a master of anything ;)
#asahi · 2026-01-19 14:28
Shizhbock: it all depends on what you want to do and what your existing skillst is :)
#asahi · 2026-01-19 14:18
hbockHi, I'm interested in contributing to Asahi development. I don't own a Mac and never have, but have been looking at the Air. It seems like M4 support is not an active focus - should I buy an M1/M2/M3 instead to get started?
#asahi · 2026-01-18 10:28
northstari followed this (https://asahilinux.org/docs/sw/windows-11-vm/), swapped the names in the script, and it worked :DDDD thank you very much chaos-princess!!
#asahi · 2026-01-18 10:21
chadmedmissed the ok
#asahi · 2026-01-18 10:20
chadmedoh nvm "if you are _ok_ using raw qemu..."
#asahi · 2026-01-18 10:20
chaos_princessyes, i was just looking at what it was trying to do, as it is all qemu under the hood and idk how virt-manager works
#asahi · 2026-01-18 10:19
chadmedthat vm was created with libvirt/virt-manager
#asahi · 2026-01-18 10:16
chaos_princessdon't see anything that strange. If you are ok using raw qemu, try following this guide, yes, it says it is for windows 11, just replace the windows iso with linux iso https://asahilinux.org/docs/sw/windows-11-vm/
#asahi · 2026-01-18 10:10
northstartried more based on the output but still failed
#asahi · 2026-01-18 10:09
northstarhttps://dustebin.com/gm0AlWfq.ts
#asahi · 2026-01-18 09:54
chaos_princesslaunch the vm, do `ps ax | grep qemu`, post output to a pastebin

plus ›