#
…

2021-10-20

00:05bgb has joined #asahi-dev
00:13bgb has quit [Ping timeout: 480 seconds]
00:22bgb has joined #asahi-dev
00:30bgb has quit [Ping timeout: 480 seconds]
00:38bgb has joined #asahi-dev
00:46bgb has quit [Ping timeout: 480 seconds]
00:51yuyichao has quit [Ping timeout: 480 seconds]
00:53bgb has joined #asahi-dev
01:02bgb has quit [Ping timeout: 480 seconds]
01:09bgb has joined #asahi-dev
01:18bgb has quit [Ping timeout: 480 seconds]
01:26bgb has joined #asahi-dev
01:45bgb has quit [Remote host closed the connection]
01:45bgb has joined #asahi-dev
01:59bgb has quit [Ping timeout: 480 seconds]
02:07bgb has joined #asahi-dev
02:16bgb has quit [Ping timeout: 480 seconds]
02:16PhilippvK has joined #asahi-dev
02:19bgb has joined #asahi-dev
02:19phiologe has quit [Ping timeout: 480 seconds]
02:48yuyichao has joined #asahi-dev
04:01kov has quit [Quit: Coyote finally caught me]
04:01kov has joined #asahi-dev
04:08doggkruse has joined #asahi-dev
04:32bgb_ has joined #asahi-dev
04:37stzsch|2 has joined #asahi-dev
04:38bgb has quit [Ping timeout: 480 seconds]
04:38stzsch|3 has joined #asahi-dev
04:42doggkruse has quit [Ping timeout: 480 seconds]
04:43stzsch has quit [Ping timeout: 480 seconds]
04:45stzsch|2 has quit [Ping timeout: 480 seconds]
05:01rohin has quit [Quit: Konversation terminated!]
05:07stzsch|3 has quit [Ping timeout: 480 seconds]
06:57Mary has quit [Quit: The Lounge - https://thelounge.chat]
07:01Mary has joined #asahi-dev
09:39robinp_ has joined #asahi-dev
09:46robinp has quit [Ping timeout: 480 seconds]
10:43marcanhm, since someone mentioned ANE again... I'm wondering *how* that would even be implemented?
10:43marcanmy hunch is it should be a DRI driver, and I see there's been drama over this already and people aren't doing that because DRI requires open source userspace
10:43marcan(cc arnd)
10:44arndwhat is ANE?
10:44marcanthe neural engine thing
10:45arndah, right. What do we know about the hardware side, and about how applications access it in MacOS?
10:46marcanI haven't looked at it at all, but AIUI it's an ASC with command submission and apparently a DART, so probably less fancy address translation than the GPU
10:46marcannot sure how it does context switching
10:46marcanbut it smells like something GPU-shaped to me
10:47marcanoddly enough though, it isn't described as an RTKit ASC in the ADT, so maybe it isn't?
10:47arndthe main question to me is whether it implements a particular set of high-level matrix operations (GEMM, convolution, ...) that can be abstracted using a kernel interface, or if this isx a fully programmable unit that relies on JIT-compiling your ML model into a custom ISA and run it autonomously
10:49TheLinkperhaps already known
10:51vmcs has joined #asahi-dev
10:51marcanthere's definitely at least some finite state machine concept
10:53marcanseems to have a fairly high level interface with program submission, queues, prioritization
10:54marcannot sure if that's the AP interface but it sounds like it could be
11:02arndright, this does sound like an ioctl-type interface to send compiled code to the engine and run that, which is indeed similar to what GPUs would do, and also a bit like what we did a long time ago with spufs
11:04arndhttps://lwn.net/Articles/870418/ summarizes the discussion we had at the kernel summit about those devices.
11:09arndin this case, I don't think the question of open source user space is the main issue. As we won't be running MacOS user space on Linux for this, someone has to complete the reverse-engineering anyway in order to create a new compiler
11:12vmcs has quit [Ping timeout: 480 seconds]
12:10alyssaarnd: right, if we have any linux support, it'd be an open user space
12:20alyssamarcan: my 2c is to stick it in drivers/gpu/drm and use the standard DRM/GEM interface just like AGX
12:20alyssabut I'm also biased as hell
12:20alyssa"gpu" stands for "generic programming unit" ;-)
12:24alyssamarcan: but do AGX first :>
12:25alyssaand AVD and AVE
12:25alyssathose coprocs actually matter
12:25alyssaI am genuinely unsure why ANE would matter on Linux outside of some niche spaces
12:26arndThe GPU programming model always feels like a layering violation to me, the same way that using DPDK to send network data does, or using a custom interface to a DSP doing video encoding instead of using v4l2. OTOH I have no idea what a good kernel abstraction for machine learning hardware would actually look like, so it's probably the best we can do here
12:27TheLinkemulate the cuda api with ane :B
12:27alyssaarnd: nod.
12:27svenjust create a new machine learning subsystem!
12:28alyssaarnd: I would also point out the GPU ioctl interfaces are extremely NIH
12:28alyssabut every mature driver by definition is committed to an existing uAPI so there is a strong incentive not to change it.
12:28alyssakrh did a neat PoC of what a "common" interface could look like but. there's no turning the clock back on anything mainline
12:28jnthe other AI accelerator driver that i'm aware of lives in drivers/misc/habanalabs, probably with a very custom interface
12:29arndTheLink: cuda is a user space interface, not kernel level, and their kernel interface is not an abstraction but hardware specific
12:29jn(userspace interface)
12:29arndone could do something that looks like cudnn or cublas, and build on top of that
12:30arndwhich in turn is what apple's Accelerate framework or Intel's OneAPI do as well, but I have not seen anyone do such an abstraction on the kernel to user boundary
12:33arndalyssa: on a related note, do you think it would be possible to have an OpenCL based BLAS implementation on top of your gallium driver, and use that for machine learning on the GPU instead of the ANE?
12:34arndI'm thinking of applications that use cublas on nvidia GPUs today, not their higher level interfaces or the tensor cores
12:35alyssasure, clover provides OpenCL on top of Gallium drivers
12:35alyssait's not ready for production yet, but it's closer to it than the AGX stack ;-)
12:42alyssaI had a dream about Apple sending me an M1 MBA and M1X MBP respectively. It wasn't a very interesting dream.
13:16TheLinkthat's the way it should be ... apple sending you devices and the process being totally uninteresting :)
13:35jbowen has joined #asahi-dev
13:46yuyichao has quit [Ping timeout: 480 seconds]
13:50stzsch has joined #asahi-dev
14:01yuyichao has joined #asahi-dev
14:10chadmedjust reading geohot's description of its data handling, it seems very GPU-
14:10chadmedGPU-like*
14:11chadmedthe ANE, that is
14:24jkkm_ has joined #asahi-dev
14:25jkkm has quit [Ping timeout: 480 seconds]
14:25jkkm_ is now known as jkkm
14:43phire_ has joined #asahi-dev
14:44aleasto has joined #asahi-dev
14:55aleasto has quit [Quit: Konversation terminated!]
14:57phire_ has quit [Quit: Leaving]
15:20aleasto has joined #asahi-dev
16:08aleasto has quit [Remote host closed the connection]
16:31aleasto has joined #asahi-dev
16:48aleasto has quit [Remote host closed the connection]
16:53aleasto has joined #asahi-dev
17:04___nick___ has joined #asahi-dev
17:16___nick___ has quit [Ping timeout: 480 seconds]
17:30phire_ has joined #asahi-dev
17:30phire has quit [Quit: Leaving]
17:30phire_ is now known as phire
17:58jacoxon has joined #asahi-dev
18:15jacoxon has quit []
18:52jacoxon has joined #asahi-dev
18:58erincandescent has quit [Remote host closed the connection]
18:59erincandescent has joined #asahi-dev
19:14Mary has quit [Quit: The Lounge - https://thelounge.chat]
19:17Mary has joined #asahi-dev
19:18jacoxon has quit []
19:19jacoxon has joined #asahi-dev
19:23jacoxon has quit []
19:24jacoxon has joined #asahi-dev
19:47jbowen has quit [Ping timeout: 480 seconds]
19:49jbowen has joined #asahi-dev
20:01jacoxon has quit []
20:02aleasto has quit [Remote host closed the connection]
20:04aleasto has joined #asahi-dev
21:06jbowen has quit [Quit: leaving]
21:51aleasto has quit [Remote host closed the connection]
21:51WhyNotHugo has quit []
21:53WhyNotHugo has joined #asahi-dev
22:46yuyichao has quit [Ping timeout: 480 seconds]
23:18Gues__________________________ has joined #asahi-dev
23:23amw_ has joined #asahi-dev
23:23amw has quit [Ping timeout: 480 seconds]
…