NV50_PROG_USE_NIR=1 doesn't work for piglit/bin/pbo-teximage ?
I was using mesa-git plus patches [0] recently (trying to get compute running on nv50), and noticed some regression in piglit quick_gl run on nv92.
Turned out recent commits (57effa34, da4c4c0e, 584b1070 - "st/mesa: Drop the TGSI paths for PBOs and use nir-to-tgsi if needed.") introduced those regressions for me. I reverted them and pbo-teximage and pbo-teximage-tiling and few others now work again .... But this is suboptimal.
I also tried to run those tests with NV50_PROG_USE_NIR=1 and sure they failed too:
NV50_PROG_USE_NIR=1 bin/pbo-teximage
Probe color at (12,12)
Expected: 1.000000 0.000000 0.000000
Observed: 0.000000 0.000000 0.000000
Probe color at (18,12)
Expected: 0.000000 1.000000 0.000000
Observed: 1.000000 0.078431 0.000000
Probe color at (12,18)
Expected: 0.000000 0.000000 1.000000
Observed: 0.000000 0.000000 0.000000
Probe color at (18,18)
Expected: 1.000000 0.000000 0.000000
Observed: 1.000000 0.078431 0.000000
[0] - https://gitlab.freedesktop.org/randrianasulu/mesa/-/tree/nv50-hacks
Even if some of those patches to blame (I'll try with clean tree some time soon) - it will be interesting/important for me to try and track this down ......