#
✕ nick:phire
#asahi-gpu · 2022-09-10 04:12
phireyeah, that was all done a while back. A lot of guess work
#asahi-gpu · 2022-08-31 06:03
phirewill watch a stream of someone else doing gpu stuff instead
#asahi-gpu · 2022-08-31 05:56
phiremaybe later, if someone doesn't beat me to getting the t6000 though initialization
#asahi-gpu · 2022-08-31 05:55
phireturns out I might not have the spare brainpower, :(
#asahi-gpu · 2022-08-30 20:48
phireYeah, makes sense
#asahi-gpu · 2022-08-29 08:35
phiremy understanding is that all the data structures are pretty simular (if not identical) and it should mostly just be slightly diffrent initdata
#asahi-gpu · 2022-08-29 08:30
phireI might do that, I need a side project to distract by brain from some other things (no promisies)
#asahi-gpu · 2022-08-29 08:26
phireand if not, would it be useful for me to try and hack it in? Or should we just wait for the proper rust driver?
#asahi-gpu · 2022-08-29 08:25
phirelina: I wasn't paying attention, does the current version of m1n1 support AGX on the t600x or just t8103?
#asahi-gpu · 2022-08-23 16:31
phireI'm not sure which direction these bandwidth counters are? is DISP reading the new frame into itself, and then deciding nothing is changed? Or is the RD the scanout going to the display?
#asahi-gpu · 2022-08-23 16:29
phirewith promotion off, the per-frame DISP RD stays about the same, but the per frame DISP WR drops to just a few hundred KB (for a roughly idle screen)
#asahi-gpu · 2022-08-23 16:27
phirebut they are roughly equal
#asahi-gpu · 2022-08-23 16:26
phireit's clearly compressed, the amount of data depends on whats on screen
#asahi-gpu · 2022-08-23 16:26
phirethe weird thing is that I'm not even sure its sending unmodifed data to the screen. With promotion off, bandwidth metrics show ~9mb per frame of DISP RD and ~6mb per frame of DISP WR
#asahi-gpu · 2022-08-23 16:20
phirewell, its adaptive. goes as low as 24fps
#asahi-gpu · 2022-08-23 16:07
phireUm... Maybe it can store an entire frame, but only with promotion enabled. The bandwidth numbers are very different with promotion (and DRAM power consumption when nothing is changing on the screen is a lot lower)
#asahi-gpu · 2022-08-23 15:58
phireThough, weird things happen with "promotion"
#asahi-gpu · 2022-08-23 15:52
phireThere is definitely not enough ram in DISP for the internal display, you can see the bandwidth changing on an idle screen when you change the refresh rate
#asahi-gpu · 2022-08-23 15:50
phirePlaying around with /usr/bin/powermetrics on my 14" m1 pro max
#asahi-gpu · 2022-08-23 09:23
phireI wonder if there are preformance counters for the number of sdram wakeups
#asahi-gpu · 2022-08-23 09:22
phireIt almost looks big enough to contain a complete frame. Assuming the ram type is optimised for fast-fill and slow scanout, and frame compression is involved
#asahi-gpu · 2022-08-23 08:44
phireWhich peeked my curiousity, have we learned anything more about it since then?
#asahi-gpu · 2022-08-23 08:44
phirean old tweet of marcan's points out the likely candidate for DISP EXT is very large, with what seems to be a lot of RAM https://twitter.com/marcan42/status/1461390053389729794
#asahi-gpu · 2022-08-18 23:50
phireThough, it's useful to have a pronoun that doesn't dicate count. Allows you to construct a gramatically valid sentance about an entity without knowing anything about them
#asahi-gpu · 2022-08-17 04:42
phirehttps://github.com/dougallj/applegpu
#asahi-gpu · 2022-08-17 04:42
phireTellowKrinkle[m]: I think dougall has some tooling
#asahi-gpu · 2022-08-17 03:47
phireYeah, I agree that proc_macro is the superior solution. Just thinking we should be clear about what it's completing against
#asahi-gpu · 2022-08-17 02:51
phireA small perl script would be way more sane than the preprocessor abuse you hinted at; and wouldn't hinder stepping though with a debugger
#asahi-gpu · 2022-08-17 02:48
phireand then push the discussion to a larger mailing list about which would be better, perl or proc_macro?
#asahi-gpu · 2022-08-17 02:48
phirePerhaps it's better to consider a solution that automatically generates c code with a small perl script (they already do a bit of auto-header generation with perl)
#asahi-gpu · 2022-08-17 02:48
phireReally, your main motivation for going with rust is that it allows you to hide auto-generated code as proc_macros, and the thread on rust-for-linux seems to indicate that they are uneasy about using proc_macros?
#asahi-gpu · 2022-08-17 02:48
phirelina: I've been thinking about rust vs c.
#asahi-gpu · 2022-08-16 08:58
phireI am
#asahi-gpu · 2022-08-16 08:57
phireI used to be a huge fan of c, because it was small enough to know the whole language, and most of the standard library
#asahi-gpu · 2022-08-16 08:56
phireBut I guess everyone doesn't agree
#asahi-gpu · 2022-08-16 08:55
phireHa! I'm kind of the opinion that everyone should know a high-preformance language, and a low-brainpower language. In my case, C++ for high preformance, python or typescript for low brainpower.
#asahi-gpu · 2022-08-16 08:53
phiredo you know other languages? or do you always go for c/c++?
#asahi-gpu · 2022-08-16 08:47
phireuseally I would have been better off waiting until preformance was actually an issue (which almost never happens), and then rewriting it. or the hot parts of it
#asahi-gpu · 2022-08-16 08:45
phireI have a bad habbit of whenever I think performance might be an issue in the future, I write in in c++
#asahi-gpu · 2022-08-16 07:31
phireOr maybe everyone just abandons python as soon as they need to do multithreaded
#asahi-gpu · 2022-08-16 07:30
phireTheoretically good. I have no idea why nobody has implemented it. Maybe there is some major roadblock I'm unaware of, or maybe separate processes is good enough.
#asahi-gpu · 2022-08-15 22:41
phirethe overhead would be much lower than full on seperate python processes.
#asahi-gpu · 2022-08-15 22:40
phiremort_: it's more like how rust's ownership model puts restrictions on transfering objects between threads
#asahi-gpu · 2022-08-13 00:06
phireI'm actually supprised that nobody has added functionality to cpython to allow multiple GILs. That way you could have multiple (functional) threads at the cost of not being able to share python data types between them.
#asahi-gpu · 2022-08-13 00:02
phireit's just not very useful
#asahi-gpu · 2022-08-13 00:02
phireI mean, cpython does support multithreadding
#asahi-gpu · 2022-06-09 06:00
phirebut prehaps it's simpler to do it in kernel
#asahi-gpu · 2022-06-09 06:00
phiremy gut would be to put as many things in userspace as possible. since it's looking like it's going to be a pretty big kernel driver.
#asahi-gpu · 2022-06-09 05:58
phirepotenitally the firmware decides when some threshold is met and sends an event telling the kernel to reconfigure?
#asahi-gpu · 2022-06-09 05:52
phireactually, IMO it would make even more sense to be handled by userspace, but we know it isn't

plus ›