<rellla>
and at least the 12 shaders.indexing tests are caused by a GP bug, that leads to hitting the 512 instructions limit.
<rellla>
many of the rest overlap with mali fails iirc
chewitt has quit [Quit: Adios!]
<daniels>
narmstrong: right - we dealt with that by prioritising Mesa jobs above KernelCI jobs, since KCI is non-time-critical post-merge, and Mesa is time-critical pre-merge
<daniels>
narmstrong: we did quite a lot of work on LAVA kernel/rootfs/image generation to make it more efficient, but I think maximal efficiency would probably want a local cache, so that's probably more infrastructure than you'd want
<daniels>
eglSwapBuffers simply places an 'implicit sync' point
<daniels>
(and explicit sync if you ask for a fence)
<daniels>
when you do that and submit the BO to KMS, KMS will want on that sync point to complete
<daniels>
gbm_bo_map will not
psydread has joined #lima
<rellla>
daniels: ok, so when i don't gbm_bo_map but submit the BO to KMS (which should be my next step) eglSwapBuffers is enough?
<narmstrong>
anarsoul: for lima we have 2xsun50i-h5-libretech-all-h3-cc 3xmeson-gxl-s905x-libretech-cc and 2xmeson-gxm-khadas-vim2
<daniels>
rellla: yep!
<narmstrong>
daniels: seems it's not enough, the lava slaves still gets overflowed, the best would be to have some dedicated MESA lava slaves at some point