01:28 <hays> its not an aquarium to eat out of haha

01:33 <alyssa> "Some people are vegetarians because they love animals. I'm a vegetarian because I hate plants."

01:35 <hays> what is v10 referring to?

01:53 <alyssa> the architecture in mali-g610

01:53 <alyssa> i.e. rk3588

01:53 <alyssa> (v10 is marketed as "Mali-Gxy0")

03:41 <hays> Ahh figured it was something like that.

21:59 <mensi> so, I'm still on my (foolish?) quest to try rusticl and I got clinfo as far as actually listing my Mali-T860 but I'm running into the unreachable() at https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/panfrost/midgard/midgard_compile.c#L1188 because dest_size=64 and nr_comps=3 so the bitsize check doesn't match any branches... Is that completely wild or explainable? Any debugging hints?

22:04 <alyssa> it's not "try"

22:04 <alyssa> it's "if you want spend hundreds of hours minimum writing an opencl implementation"

22:04 <alyssa> (I say this having spent hundreds of hours on the newer mali and still don't have something passing CTS)

22:06 <mensi> So I guess my impression that rusticl generates NIR that then goes to the card is naiive and there's far more code in the middle missing?

22:19 <alyssa> well, you need a compiler that ingests NIR and spits out native code

22:19 <alyssa> Midgard assembly in your case

22:20 <alyssa> and by nature of OpenCL being so much more expressive, the NIR that Rusticl emits is much more complicated than the NIR that the GLSL compiler producers

22:20 <alyssa> *produces

22:20 <alyssa> need to teach the compiler (Midgard in this case) to handle all the new OpenCL shiny

22:20 <alyssa> 8-bit integers, 64-bit integers, numerical precision issues, etc

22:21 <mensi> I see

22:22 <alyssa> i'm also not aware of a single linux/arm64/opencl workload that's faster on the GPU than the CPU

22:22 <alyssa> so it's difficult to motivate.

22:28 <mensi> oh? are the NEON and whatnot extensions so good or the GPUs so bad?

22:39 <robclark> usually the cpu can saturate bandwidth.. tilers get a big leg up when it comes to 3d workloads by just avoiding a lot of memory bandwidth. But compute workloads don't benefit from that. So if it comes down to CPU saturating bandwidth vs GPU saturating bandwidth, it comes down to which draws less power doing so

22:40 <mensi> as in geometry / textures / the big data is relatively static for 3D -> you only copy once?

22:42 <mensi> actually do these Mali GPUs even have dedicated memory or are they saturating the main memory bus for everything?

22:42 <robclark> as in all the memory traffic to/from color/depth/stencil doesn't leave the gpu core

22:42 <robclark> no dedicated memory other than caches and tile buffer

22:43 <mensi> I see... are these dedicated NPUs like rock chip sticks in newer chips any better in that regard?

22:44 <mensi> or is it also just a memory contention thing?

22:45 <robclark> I mean, it comes down to what the bottleneck in a given workload.. if it is memory bandwidth, the gpu doesn't have any more than the cpu.. that was the point I was trying to make

22:48 <mensi> true, like what I was thinking is all these inference workloads and whatnot sound quite memory intensive if you don't have any considerable dedicated space to keep intermediates

22:51 <alyssa> correct

22:54 <alyssa> in general don't expect to run serious workloads on not serious hardware :p

22:55 <mensi> hehe of course, but I feel bad if I don't at least try anything with all these SBCs littering my desk ;)

22:55 <alyssa> Where are all these SBCs coming from?!

22:55 <mensi> China? ;)

22:55 <alyssa> hmm seems legit

22:57 <mensi> I was positively surprised how much better the OSS support for these is these days thanks to you people here :)

22:57 <alyssa> I guess that's good :)

22:58 <alyssa> usually i only hear when people are unhappy with it so based on my inbox you'd think support is garbage.

22:58 <alyssa> (who's jaded? not me! /s )

23:00 <mensi> yeah I guess that is one of those traits of humanity, you can share as much goodness as you want, there will always be people with entitlement issues and complaints

23:05 <alyssa> well, that, and drivers are infrastructure

23:05 <alyssa> most people don't think about them unless they break

23:05 <alyssa> which means as a driver developer by the time I talk to a user they're probably quite frustrated :)

23:06 <mensi> well, my driver writing experience so far was only for https://barrelfish.org where we didn't really have users ;)

23:11 <alyssa> ah..