17:45 tomeu: jhugo: hi, maybe this could interest you? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35441
17:46 tomeu: @wens: yes, that's my impression as well
17:47 jhugo: That does look interesting
17:48 jhugo: Our work hasn't gotten as far as I would have liked, although basically have someone working full time on it now
17:48 tomeu: oh wow, that sounds great
17:49 tomeu: how does the HW work? there are DSP instructions that copy data to the accelerator, configure it, then copy the results out?
17:51 jhugo: The hardware runs an elf compiled for the DSP. There are commands to pull the elf over and start running it. Then the host can program a DMA to get data in/out
17:51 tomeu: ah, I was referring to the actual instructions in the elf
17:51 tomeu: the ones that program the accelerator hardware
17:52 jhugo: The DSP has its own ISA
17:52 jhugo: There are co-processors that do SIMD/NEON like things, and also other co-processors that do matrix stuff
17:53 tomeu: ok, that sounds similar to TI's DSP with MMA
17:53 tomeu: those NIR intrinsics should be a starting point to lowering to those DSP ISAs
17:56 jhugo: We've developed a shim that we can send commands to, with the idea that we can build up .so send them over, and link to the shim. Then "call" a function that we've compiled to run the graph
17:57 jhugo: likely using hand coded ops as the building blocks