18:29 <gerddie> Hi all, we are trying to enable etnaviv GPU coded path for model 0x8000, rev 0x6025 and run into a two bigger problems: One one hand there is the flop reset that is required to get the FP hardware run properly. I've found the place in the galcore kernel code where this is emitted, so it seems a compute shader with EXT_shader_image_load_store

18:29 <gerddie> support can do the job. There is an MR open to add basic compute shader support https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18986/, however it is a bit bitrotten, so I wonder whether someone else has already something more up-to-date in the pipeline?

18:30 <gerddie> tomeu: Since you work a lot on etnaviv, do you think it makes sense to rebase above MR in order to get it merged?

18:32 <gerddie> tomeu: You also noted that after running some job FP was working fine, i.e. you somehow got a flop reset through the job that was run. Did this only work for TP, or also for GPU (PPU) work loads?

18:33 <austriancoder> gerddie: that MR will not land - sorry.

18:33 <austriancoder> gerddie: you can give that one a try: https://gitlab.freedesktop.org/austriancoder/mesa/-/tree/ebc-with-opencl?ref_type=heads

18:39 <gerddie> Thanks for the answer and no problem if that MR doesn't get merged, I am basically looking into some way to get compute support to implement the PPU flop reset, so I'll happily try your branch and see where this gets me

18:42 <austriancoder> gerddie: which function in the galore kernel, does the needed flop reset?

18:42 <austriancoder> gerddie: out of interest - what SoC is that?

18:53 <austriancoder> gerddie: if its gckHARDWARE_ResetFlopWithPPU(..), you will have no luck. EVIS instructions are almost not REed, and support in the compiler toolchain is non existing. I started to look into it months ago and got that far: https://gitlab.freedesktop.org/austriancoder/mesa/-/commit/d26c2bae82872accb6d0d7cd6e9fa139d347a5c4

18:55 <austriancoder> gerddie: ohh.. my ebc-with-opencl remote branch is not up to date, but is current state should be okay for experiments

19:12 <gerddie> It is indeed gckHARDWARE_ResetFlopWithPPU that does the needed flop reset, the other two don't help. The SoC is STM32MP2

19:14 <austriancoder> gerddie: the only (hacky) way to get this up and running, is to bypass any compiler and use the raw binary shader to do the job.

19:18 <gerddie> austriancoder: Yeah, that was my thinking too after your comment

19:50 <tomeu> Yeah, that's what I have seen to work. The flop reset is needed for ppu jobs that do floating point operations.

19:51 <tomeu> Guess we don't strictly need a full compiler for this, but at least an assembler.