00:03 jenatali: glehmann: Ok trying to retrofit this into the DXIL backend is going to be a massive pain. I really need a nir instruction that's u32 unpacking to f32
00:03 jenatali: Otherwise I need to store effectively an f16 SSA value but pretend that it's u32 and I don't have a great way to do that
00:08 ngcortes: anholt, is there a way to have deqp-runner generate its own caselist? eg if I want to test gles on glcts, can I just have it generate a caselist based on all the KHR-GLES xml files?
09:02 rcv11x: Hello everyone, I have a question about the commit search filter. How can I filter to see what will be included in Mesa 26.0? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests
09:07 K900: You can't
09:07 K900: You can look at https://gitlab.freedesktop.org/mesa/mesa/-/tree/staging/26.0?ref_type=heads
09:18 glehmann: jenatali: okay, I pushed a version that adds a dxil specific unpack opcode
09:19 glehmann: could share it with r600, but meh
10:22 eric_engestrom: ngcortes: deqp itself (not deqp-runner) has `--deqp-runmode=txt-caselist` which dumps the list of all the tests compiled in
10:28 eric_engestrom: (there's also `--deqp-runmode=stdout-caselist` if you want to use it in a script and don't want to have to keep a mapping of the file name for each deqp binary)
10:39 daniels: mupuf, eric_engestrom: all the Steam Decks are having a bad time
10:40 mupuf: daniels: thanks for letting me know, looking into it
10:40 daniels: https://gitlab.freedesktop.org/mesa/mesa/-/jobs/92030823 on -7, https://gitlab.freedesktop.org/mesa/mesa/-/jobs/92030824 on -5, https://gitlab.freedesktop.org/mesa/mesa/-/jobs/92030825 on -4, https://gitlab.freedesktop.org/mesa/mesa/-/jobs/92030826 on -2
10:40 eric_engestrom: they're all idle
10:40 eric_engestrom: thanks, I was typing to ask for links ^^
10:40 eric_engestrom: this is really weird
10:41 daniels: though possibly transient as they did eventually come back it seems, just too late to not tank the pipeline
10:41 eric_engestrom: mupuf: should I restart the executor?
10:41 eric_engestrom: oh, not touching, if it's back
10:41 eric_engestrom: better to leave as is for possible investigation
10:41 daniels: wfm
10:41 mupuf: seems like a gitlab runner issue, not executor :s
10:42 eric_engestrom: this feels like a network interruption between the farm and internet
10:42 daniels: if it falls down a well again, I'll let you know
10:42 daniels: sounds right yeah
10:42 eric_engestrom: daniels: thanks!
10:42 eric_engestrom: mupuf: agreed, but the executor is the one carrying the messages, that's why I was thinking it might be that
10:43 eric_engestrom: s/messages/job log/
10:43 mupuf: eric_engestrom: what do you mean? No, the gitlab runner is the one communicating with gitlab
10:43 eric_engestrom: oh, directly? I misremembered then
10:43 mupuf: yeah :)
10:43 eric_engestrom: ack, thanks ^^
15:21 zmike: what's the best way to move nir instrs into a different block? like if I want to move some stuff into an if()
15:28 mareko: nir_instr_move
15:29 mareko: nir_opt_sink
15:29 zmike: hm neat
15:29 zmike: thx
15:44 zmike: what's the name of that pass that moves all the loads to the top of the block?
15:46 dj-death: zmike: nir_opt_preamble ?
15:47 zmike: 🤔
15:47 zmike: I don't think that's what I'm looking for exactly
15:48 zmike: or is it?
15:49 zmike: I think not
16:03 mareko: nir_opt_move_to_top
16:03 mareko: not the block, but the shader
16:03 zmike: yeah that's the one
16:03 zmike: thanks
16:10 mareko: it would be nice to have 1) only one NIR pass per file, 2) the beginning of each file describes the pass and how it works
16:11 zmike: it would be nice to have a catalog of all the passes which documents them and describes them
16:11 zmike: so I don't have to ask for help any time I'm trying to find one
16:26 glehmann: what we need is a NIR LLM chatbot
16:27 zmike: yes.
16:33 kisak:winces
16:33 cwabbott: I mean, mesa is in every LLM's training data
16:33 cwabbott: so every LLM will confidently get it incorrect already
16:35 zmike: I think what glehmann is saying is that ours should be correct
16:35 cwabbott: lol, good luck with that
16:36 ccr: perhaps it could be a LLM based "Mesa development psychiatrist"
16:52 dcbaker: glehmann: I had to resolve some conflicts with one of your patches for 25.3, could you look at it and make sure it's okay? https://gitlab.freedesktop.org/mesa/mesa/-/commit/dd0e8a62b8fe5f30811a5ac31160a4d28ed500fb
16:55 glehmann: looks fine, I guess the only conflict was the new p_v_fma_mixlo_f16_rtz instructions?
17:00 mareko: only internet-searching LLMs could provide decent up-to-date information on NIR, and LLMs currently can't open FDO Gitlab because robots are banned there
17:01 dcbaker: glehmann: yeah, in both the removed and inserted hunks
17:01 mareko: is 1) and 2) above are done, we can generate a catalog of all passes programmatically
17:01 mareko: *if
17:07 anholt: ngcortes: nobody's done parsing of the mustpass.xmls. Feel free to write it, I suppose.
17:07 jenatali: Seems an LLM agent should be able to handle that if you ask it to search the nir passes in a local clone of the repo?
17:34 mareko: no idea if LLMs can guess what passes do if they are not explained in code comments
17:40 alyssa: it would be nice if I wrote that "how to write a NIR pass" doc that I put on my todo list in 2023
17:40 alyssa: i'm sure someone could train a markov chain on my gitlab comments and generate most of it
18:24 sghuge: pixelcluster: konstantin by any chance do you know or have list of ray tracing games that are using update BVH? I am trying to save myself some efforts of finding list of games by installing and trying out. Thanks.
19:13 olivial: is there a reason that vk_shader_multizalloc isn't doing the same allocation scope thing that vk_shader_zalloc is?