#
#asahi-gpu · 2022-11-28 17:32
marcanand function-ane_data_param_set = 329:dSet()
#asahi-gpu · 2022-11-28 17:32
marcanisp has function-ane_ep_control = 329:eCtl()
#asahi-gpu · 2022-11-28 17:32
marcanI mean t8103 is in desktops (mac mini)
#asahi-gpu · 2022-11-28 17:32
eiln_there's a ane_isp.afpp file somewhere
#asahi-gpu · 2022-11-28 17:32
eiln_no idea, i only have a t8103
#asahi-gpu · 2022-11-28 17:32
marcani.e. desktops
#asahi-gpu · 2022-11-28 17:31
marcanwhat happens on devices without ISP?
#asahi-gpu · 2022-11-28 17:31
eiln_i don't know if isp depends on ane though
#asahi-gpu · 2022-11-28 17:31
marcansince we kind of need ISP to work...
#asahi-gpu · 2022-11-28 17:31
marcanthat sounds like we eventually need to figure out how to make it work then, unless ISP can be told not to talk to ANE
#asahi-gpu · 2022-11-28 17:31
eiln_i believe isp has an ane node?
#asahi-gpu · 2022-11-28 17:30
eiln_unless explicitly turned off with boot-args, the first ane init is called by isp
#asahi-gpu · 2022-11-28 17:30
marcanso isp and ane talk to each other for the ML stuff in the camera processing?
#asahi-gpu · 2022-11-28 17:30
eiln_yes
#asahi-gpu · 2022-11-28 17:29
marcanoh. isp as in the camera thing?
#asahi-gpu · 2022-11-28 17:29
eiln_there's a "program" abstraction that's really just BAR (in ane_context.py)
#asahi-gpu · 2022-11-28 17:28
eiln_i'm pretty certain that my asc init is stuck @ stage 8 because isp isn't
#asahi-gpu · 2022-11-28 17:28
eiln_props up an entire server with isp
#asahi-gpu · 2022-11-28 17:28
marcan(I'm just trying to understand this)
#asahi-gpu · 2022-11-28 17:28
marcanso what does the rest of the 3MB of firmware do? :p
#asahi-gpu · 2022-11-28 17:27
eiln_given i have rvbar & the asc init code technically i can write my own fw to do just that
#asahi-gpu · 2022-11-28 17:26
eiln_are asc duties
#asahi-gpu · 2022-11-28 17:26
eiln_and the 3 reg writes in execute_tq
#asahi-gpu · 2022-11-28 17:25
eiln_specifically enqueue_tq
#asahi-gpu · 2022-11-28 17:25
eiln_yes the asc isn't even turned on
#asahi-gpu · 2022-11-28 17:25
eiln_so i copied the reg diffs to "fill in"
#asahi-gpu · 2022-11-28 17:25
eiln_sigh, the ane-asc init process is a 9-step process, and i've been stuck at stage 8
#asahi-gpu · 2022-11-28 17:24
marcanso you don't really use the asc, just talk to the hardware directly?
#asahi-gpu · 2022-11-28 17:24
eiln_asc is only responsible for queue stacking, which i replicate in TaskManager
#asahi-gpu · 2022-11-28 17:20
marcanusually everything goes through the ASC for hardware that has an ASC
#asahi-gpu · 2022-11-28 17:20
marcaneiln_: that's typical yes, the ASCs sit on the side. but does macOS talk to the engine directly, or to the ASC?
#asahi-gpu · 2022-11-28 17:19
marcanyou probably just need to call p.dapf_init("/arm-io/dart-ane")
#asahi-gpu · 2022-11-28 17:19
eiln_asc is completely separate from the engine circuitry. the dma they refer to isn't asc-emulated-dma
#asahi-gpu · 2022-11-28 17:18
svenpmgr often is apple‘s „let’s stuff things that don’t fit anywhere else“
#asahi-gpu · 2022-11-28 17:18
marcansince they probably map the firmware
#asahi-gpu · 2022-11-28 17:18
marcanthey are mostly likely dynamic and depend on the firmware version and RAM config of the device
#asahi-gpu · 2022-11-28 17:18
marcanand your "dart tunables" are DAPF, see dapf.c in m1n1 (this is already implemented)
#asahi-gpu · 2022-11-28 17:18
eiln_regarding asc, sigh
#asahi-gpu · 2022-11-28 17:17
eiln_ ah hv said pmgr node so it's what i assumed. i've added a patch under tunables_static.c too
#asahi-gpu · 2022-11-28 17:15
marcaneiln_: also I think your "pmgr" tunables are actually fabric tunables, see the analogous https://github.com/AsahiLinux/m1n1/blob/lina/gpu-wip/src/tunables_static.c#L22 from lina's stuff
#asahi-gpu · 2022-11-28 17:12
marcaneiln_: I'm a bit confused about the ASC though; I know it has a big firmware blob but it doesn't look like it uses the usual mailbox RTKit interface? Does it have some bespoke interface?
#asahi-gpu · 2022-11-28 17:00
marcaneiln_: you should probably look at https://lore.kernel.org/lkml/20221022214622.18042-1-ogabbay@kernel.org/
#asahi-gpu · 2022-11-28 16:31
eiln_Neural processor circuit 218 may receive the input data from sensor interface 212 , the image signal processor 206 , persistent storage 228 , system memory 230 or other sources such as network interface 210 or GPU 220"
#asahi-gpu · 2022-11-28 16:31
eiln_"Neural processor circuit 218 is a circuit that performs various machine learning operations based on computation including multiplication , addition , and accumulation . Neural processor circuit 218 is a configurable circuit that performs these operations in a fast and power -efficient manner while relieving CPU 208 of resource - intensive operations associated with neural network operations.
#asahi-gpu · 2022-11-28 16:31
eiln_figured this introduction from their patents may be helpful:
#asahi-gpu · 2022-11-28 16:19
eiln_Being a non-essential component and without much precedent, I feel pretty lost on what direction to head with development. I would rather asahi leverage the hardware to its own needs, not replicate apple's model (which i find very questionable anyway). Hence I currently have no idea of what a final product would even look like (e.g., support what? does it go under drm?). I'm looking for any form of advice / guidance.
#asahi-gpu · 2022-11-28 16:19
eiln_I've got the kernel side / work interface / tiling pretty much down, so the only barrier to all those exotic ops are the compiler output ("regfile" that gets dma'd to configuration registers). I do have prototypes for elementwise add/mul/min/max, 2d convolution, and batched matrix multiplications, but it's still largely a mystery.
#asahi-gpu · 2022-11-28 16:18
eiln_TDLR: ane is a low precision (fp16) configurable mac circuit really good for chained/fifo stream number crunching
#asahi-gpu · 2022-11-28 16:18
eiln_I develop on github but am currently having issues it's a mirror
#asahi-gpu · 2022-11-28 16:18
eiln_https://gitlab.com/eiln/m1n1.git

plus ›