14:05 <tomeu> lynxeye: thanks, but I think that should be fine. Each operations gets its buffers added as relocs to the cmdstream, so if we split, the right buffers still get added: https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/gallium/drivers/etnaviv/etnaviv_ml.c#L772

14:11 <lynxeye> tomeu: maybe dumping out if any buffers get released due to etna_bo_del in etna_cmd_stream_flush helps to see if there are any changes to buffer usage due to the cmdstream split?

14:12 <tomeu> I will certainly be adding quite some logging and see if there is any obvious difference between when the faults and happen, and when they don't

14:13 <tomeu> but the buffers get allocated during compilation, then the compiled graph is run a number of times, and only when the program finishes and the graph gets deallocated, we free anything

14:14 <tomeu> something I still don't understand is why this may happen on the first invocation, but after a first reset we don't get any further faults

14:18 <lynxeye> yes, that seems quite odd and I have no explanation why this would be triggered by the cmdstream splits