#asahi-gpu · 2024-08-07 22:33
karolherbst(and some flakes I haven't hit this run yet)#asahi-gpu · 2024-08-07 22:33
karolherbstdoing a real run atm and the only thing failing is that image stuff which is fixed by compblit#asahi-gpu · 2024-08-07 22:14
karolherbstseems like my CTS fix for fma already got approved :) yay#asahi-gpu · 2024-08-07 11:49
karolherbst(anyway, will do a proper CTS run today, because I don't think those issues happen when those tests are ran alone)#asahi-gpu · 2024-08-07 09:10
karolherbstmaybe I should try a newer kernel..#asahi-gpu · 2024-08-07 09:08
karolherbstalyssa: are you aware of any bug related to sync objects in a multi-threading world? I have this weird test (test_events userevents_multithreaded) which might end up waiting on a sync obj forever, even though nothing which could block execution forever is submitted#asahi-gpu · 2024-08-06 16:22
karolherbstI wonder if submitting those constant as kernel inputs would be acceptable upstream as well. But I'm mostly worried about "you gotta fix volatile tho" reactions#asahi-gpu · 2024-08-06 15:56
karolherbstand it's indeed fixing the test#asahi-gpu · 2024-08-06 15:56
karolherbstif we can't get the CTS fix, the mesa fix would look like this: https://gitlab.freedesktop.org/karolherbst/mesa/-/commit/e15897c0ad14b15db8cceb2397d3d766d505558c#asahi-gpu · 2024-08-06 13:43
karolherbstbut the spirv volatile thing says "This access cannot be eliminated, duplicated, or combined with other accesses." which we clearly do#asahi-gpu · 2024-08-06 13:41
karolherbstI wonder if us ignoring volatile the way we do can be seen as our bug or not... but yeah, maybe the CTS should check both possibilities here#asahi-gpu · 2024-08-06 13:00
karolherbstor we just deal with volatile correctly, but what a pain#asahi-gpu · 2024-08-06 12:53
karolherbst(and figure out what we need to change)#asahi-gpu · 2024-08-06 12:53
karolherbstI wonder if we should add a nir_option thing for this...#asahi-gpu · 2024-08-06 12:52
karolherbstalyssa: I've checked the fma issue again.. it was this "Tininess is detected before rounding?" issue and constant folding being like that...#asahi-gpu · 2024-08-05 17:09
karolherbstyou'll figure it out :D#asahi-gpu · 2024-08-05 17:09
karolherbstor maybe it's something else#asahi-gpu · 2024-08-05 17:09
karolherbstI hope my hunch about last_block being the broken bit is right..#asahi-gpu · 2024-08-05 17:09
karolherbst🙃#asahi-gpu · 2024-08-05 13:10
karolherbstyeah...#asahi-gpu · 2024-08-05 12:53
karolherbstmhhh.. yeah, so the fma emulation works, but that's kinda expected, but that's a huge pile of lowering code#asahi-gpu · 2024-08-05 12:45
karolherbstpipe_grid_info::last_block I mean here#asahi-gpu · 2024-08-05 12:44
karolherbstso maybe when I'm doing that, I might run into issues there unless you use last_block for anything else#asahi-gpu · 2024-08-05 12:44
karolherbstI wanted to look into non uniform work groups at some point on asahi anyway...#asahi-gpu · 2024-08-05 12:43
karolherbstmaybe something with last_block?#asahi-gpu · 2024-08-05 12:43
karolherbstannoying#asahi-gpu · 2024-08-05 12:41
karolherbstI was hoping for compblit, because that also fixes it :'(#asahi-gpu · 2024-08-05 12:41
karolherbstmhhh#asahi-gpu · 2024-08-05 12:40
karolherbstmeanwhile I've coded up a workaround for that 1Dbuffer failures: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30514#asahi-gpu · 2024-08-05 12:40
karolherbstI see...#asahi-gpu · 2024-08-05 12:40
karolherbstohhh...#asahi-gpu · 2024-08-05 12:39
karolherbstbesides that one, the only other known bug with the OpenCL CTS is fma being a little on the weird side and I haven't figured out what's wrong exactly there#asahi-gpu · 2024-08-05 12:27
karolherbstguess I'll need to figure out what u_blitte is doing then...#asahi-gpu · 2024-08-05 12:23
karolherbst"0xffffffffffffffff vs. 0x7e007e007e007e00" those kind of issues#asahi-gpu · 2024-08-05 12:22
karolherbstno, it's handling NaNs in a way I don't need it to handle#asahi-gpu · 2024-08-05 12:22
karolherbstoh no#asahi-gpu · 2024-08-05 12:21
karolherbstalyssa: sooo.. "ASAHI_MESA_DEBUG=compblit", when will that be the default? It fixes a bug with u_blitter I'm running into 🙃#asahi-gpu · 2024-06-05 15:49
karolherbstthat reminds me.. maybe I should complete the CL stuff and then....#asahi-gpu · 2024-06-05 15:41
karolherbst:3 congrats#asahi-gpu · 2023-10-24 23:32
karolherbstpain#asahi-gpu · 2023-10-24 23:32
karolherbstbut also not much...#asahi-gpu · 2023-10-24 23:32
karolherbstalyssa: the "closed source" CTS has some tests#asahi-gpu · 2023-10-24 23:31
karolherbstexternal/openglcts/modules/gl/gl4cDirectStateAccessQueriesTests.cpp has something#asahi-gpu · 2023-10-24 23:31
karolherbstmhhh#asahi-gpu · 2023-10-24 23:30
karolherbstbut yeah.. seems like the CTS doesn't ... weird#asahi-gpu · 2023-10-24 23:29
karolherbstalyssa: fyi, piglit has a few tests#asahi-gpu · 2023-10-24 23:17
karolherbstjust that I kinda need timestamps on the GPU side so I know when certain commands are actually executed. but probably also totally fine to do it like this, as each command is an ioctl anyway.. kinda... I think?#asahi-gpu · 2023-10-24 23:16
karolherbstbut yeah.. dunno. I can give it a try and see how actually useful it it in the current state#asahi-gpu · 2023-10-24 23:15
karolherbstprobably makes sense to fill the query buffer with the shader clock stuff once that's done#asahi-gpu · 2023-10-24 23:14
karolherbstmhhh, that's all CPU side, right?