04:35imirkin: w00t! dEQP-GLES31.functional.compute.basic.copy_ssbo_single_invocation passes
04:36imirkin: still need to fix up samplers ... and atomics
04:54imirkin: wow, ld lock/st unlock only available starting gt200. sadness. no shared atomics until then.
06:43imirkin: pmoreau: have you had any trouble with atomics on nv50?
06:43imirkin: it feels like they're not working for some reason
13:48kherbst: mhhh, tagr_ posted nvdec patches for tegra :D
13:48kherbst: wondering if we can make use of the firmwate on desktop chips
13:49kherbst: also wondering if we should have a libfalcon module or something both drivers could depend upon
18:49imirkin: so ... it helps to do the compute ubo setup using the compute channel rather than the 3d channel. oops.
18:49imirkin: and it looks like atomic writes are somehow delayed? really unclear...
18:50imirkin: mwk: are you aware of anything like that? i'm not seeing atomic writes immediately hit memory after the compute job is done. or maybe i'm not waiting for it properly? hmmm
18:50imirkin: yeah, the latter's more likely
18:50imirkin: probably waiting for 3d channel to idle, which it will immediately. sigh
20:25imirkin: ah i see. probably need to do something with the 0x0360 method
20:48imirkin: yeah, looks like we have to do some kind of dance
20:55imirkin: gr. and no z grid support, will have to do some math. fun.
21:01RSpliet: sometimes you do wonder what you actually enjoy about this hobby don't you? :-P
21:02imirkin: best not to think about it too much
21:05ccr: art comes from pain?
21:25kherbst: imirkin: mhh, interesting
21:26kherbst: imirkin: for the z grid part I was addind some compiler support to deal with runtime injected offsets (as CL spec requires it anyway)
21:30kherbst: btw, still searching for somebody taking a look and thinking about making nouveau_mm.c thread safe :D https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8765/diffs?commit_id=fbac2a194b6a373ce21d30701783d89daaa8434f
21:53imirkin: karolherbst: it's a single invocation with z > 1, so there's not really any "injection" per se
21:53imirkin: it's just that you have to fold the z dim into the y dim
21:54imirkin: and then do mod's
21:54imirkin: or something
22:25imirkin: i can just invoke it z times
22:25imirkin: and pass in the z value via the grid id
22:25imirkin: he he he
22:31imirkin: just have to figure out where that grid id comes out in the shared mem
23:38imirkin: ah no, that of course won't work. i need both the dimension and the position, and the gridid is just 16-bit.
23:39imirkin: need to push it via constbuf i guess
23:51imirkin: i guess i could push the values via user parameters
23:51imirkin: seems dodgy though
23:53imirkin: hm, that seems by far the simplest though