08:27 <tomeu> lynxeye: austriancoder I'm looking again at this problem with compositors trying to render with a NPU-only Vivante GPU, and I'm quite unsure of how etnaviv's architecture could ever handle this scenario

08:28 <tomeu> from what I have seen, a render node will always be created if there is a vivante,gc in the DT, regardless of whether anything can be rendered with it

08:29 <tomeu> and compositors and Mesa will happily try to use that render node to create EGL contexts and then fail spectacularly, of course

08:29 <tomeu> from what I have seen, the hw-independent parts of the stack assume that if there is a render node, then rendering can be done with it

08:30 <tomeu> but the etnaviv kernel driver would create a render node on purpose before even knowing if there is rendering hw from vivante on the machine

08:31 <tomeu> afaics, the exact same problem would happen if there was a SoC with a 2D engine but no 3D engine, do those cases exist and if so, how do people cope with them?

08:32 <tomeu> do you have any ideas on how this should work?

08:34 <austriancoder> what would happen if DRIVER_COMPUTE_ACCEL would be used for NPU-only GPUs instead of DRIVER_RENDER?

08:40 <austriancoder> tomeu: as hwdb landed now I think it would make sense to revert some kernel changes that happend in 6.9 rc cycle like https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/gpu/drm/etnaviv?h=v6.9-rc4&id=1dccdba084897443d116508a8ed71e0ac8a031a4 - what is your opinion?

11:47 <tomeu> <austriancoder> "tomeu: as hwdb landed now I..." <- yes, no problem with that

11:49 <tomeu> <austriancoder> "what would happen if DRIVER_COMP..." <- I think that would be a good thing in itself, but currently the DRM device gets created way before any GPU's register can be read

11:49 <tomeu> but I could look at refactoring if we think that is a good idea

12:12 <tomeu> another possibility would be to add a compatible label for the npu, so we don't need to read from the GPU to figure out what kind of piece of HW it is

12:12 <tomeu> which is I guess how this was supposed to work

13:17 <tomeu> lynxeye: what do you think of adding a new DT compatible for NPU-only GPUs?

13:19 <tomeu> alternatively, we could try instantiating a rendernode, and fail if there is no rendering HW, then retry for an accel node

13:37 <austriancoder> tomeu: hmm?

13:38 <tomeu> was just chuckling at my typo :)

13:43 <tomeu> lynxeye: austriancoder: I drafted a quick PoC for the retrying possibility: https://gitlab.freedesktop.org/tomeu/linux/-/commit/ff65a197d37e8f838b488595c371b2095f2e6700

13:43 <tomeu> seems to work, with minimal changes to the existing code base

13:43 <tomeu> probes of NPUs will be a bit slower, but maybe that is fine

13:48 <austriancoder> tomeu: I'll leave the decision up to Lucas. for me both ways are okay (dt binding, and somthing like your PoC).

13:49 <tomeu> cool, thanks for taking a look