16:06 anholt: hmm. ssbo stores on nv12b seem unstable. flaky results from tests even as simple as dEQP-GLES31.functional.shaders.builtin_functions.common.abs.ivec3_lowp_compute
19:45 imirkin: anholt: in general, that stuff works. we pass CTS (but not with cts-runner) on fermi+.
19:45 imirkin: technically that's dEQP, but wtvr, should still work
19:45 anholt: interesting. I'm seeing fails on single test, no other activity on the system.
19:46 imirkin: GM20B is obviously slightly special, and things could have regressed without us noticing
19:46 anholt: (jetson nano)
19:46 imirkin: right
19:46 imirkin: otoh, i thought karolherbst_ recently got his up and running, and it was fine
19:46 imirkin: anholt: mind providing dmesg?
19:47 imirkin: want to make sure nouveau init'd properly
19:47 anholt: all of gles2/3 seem to be stable
19:47 imirkin: oh
19:47 imirkin: weird.
19:47 imirkin: so compute is busted?
19:47 imirkin: is this with nir or tgsi?
19:48 anholt: tgsi
19:49 anholt: https://people.freedesktop.org/~anholt/jetson-dmesg
19:49 imirkin: [ 4878.079815] nouveau 57000000.gpu: deqp-gles31[12228]: nv50cal_space: -16
19:49 imirkin: are these in response to running that test?
19:49 imirkin: or were they previous failures?
19:49 anholt: previous fails
19:50 anholt: a flaky fail doesn't print anything in dmesg
19:50 imirkin: hrmph. well it could be we're messing something up wrt the lack of vram
19:50 imirkin: oh. *flaky*. i missed that in your original comment
19:50 anholt: you can see some probe debug stuff I had dropped in from before I found I needed pwm regulator enabled.
19:51 imirkin: that's incredibly weird
19:51 imirkin: when it fails, what's the failure?
19:52 imirkin: oh, could be some fence messed-up-edness
19:52 imirkin: i recently saw patches flying around to fix something which was done in 5.15
19:52 anholt: ERROR: comparison failed for value 0: Expected [0] = 80 inputs: in0 = ivec3(80, -34, -96) outputs: out0 = ivec3(0, 0, 0) ERROR: comparison failed for value 1: Expected [0] = 22 inputs: in0 = ivec3(-22, -92, -120) outputs: out0 = ivec3(0, 0, 0) ERROR: comparison failed for value 2: Expected [0] = 117 inputs: in0 = ivec3(-117, 97, -87) outputs: out0 = ivec3(0, 0, 0) ERROR: comparison failed for value 3: Expected [0] = 4 inputs: in0 = ivec3(4,
19:52 anholt: -28, -47) outputs: out0 = ivec3(0, 0, 0) 96 / 100 values passed 56206 Result comparison failed
19:54 imirkin: ok, so the outputs are all 0's
19:54 anholt: not finding a kernel fix in a quick search
19:54 imirkin: i.e. it hasn't written anything yet
19:54 imirkin: from ckoenig
19:55 anholt: yeah
19:55 imirkin: anholt: "[PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2"
19:55 imirkin: oh hm. that was a "performance regression", not a correctness fix... hrmph
19:56 imirkin: anyways, when in doubt, i try an older kernel. since the mid 4.x's, on average newer kernels are more likely to break things than fix things :)