00:25 alyssa: karolherbst: how is that defined? fmul isn't associative
00:26 alyssa: "(a*b)*c" apparently.
00:26 karolherbst: alyssa: same for that one, but you can order the sources as you like
00:26 karolherbst: like infinite precision internally
00:26 karolherbst: kinda like fma, but as fmul and three operands
00:27 karolherbst: we only really want to use it in nak to set the constant operand on our fmul instruction (0.125, 0.25, 0.5, 1.0, 2.0, 4.0 or 8.0) but the instruction calculates the result with infinite precision
00:27 karolherbst: so kinda want to make use of all the fp ctrl stuff
00:28 alyssa: meh
00:28 alyssa: fmul3_nv i guess
00:28 karolherbst: yeah.. at which point I thought just calling it fmul3 might make sense in case others want it...
00:28 karolherbst: but can also call it fmul3_nv
00:28 karolherbst: people can rename it if they want to use it
00:28 alyssa: yeah
00:29 alyssa: adding things for 1 driver and then generalizing if/when there's 2 is often a good strategy anyway
00:29 karolherbst: but yeah.. need to honor exact and stuff and using nir algebraic is kinda the obvious choice for it
01:22 alyssa: yeah
03:21 airlied: glehmann, karolherbst : got llama.cpp cm2 paths to parity with cm1 on radv
03:33 airlied: but we definitely need shared memory aliasing
07:49 glehmann: airlied: if it's just parity, is it worth the effort?
10:30 karolherbst: glehmann: apparently nvidia manages to get +30% perf
11:28 karolherbst: alyssa or jenatali: mind reviewing this patch? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/43140/diffs?commit_id=66fc28d1cd0c140128b532b4db727ed7b5629747 It's a weird corner case with libclc using fp16 builtins that libclc does not provide, so the current fp16 lowering kinda falls apart...
15:57 bcheng: Anyone able to help with a a618-android CI failure? I highly doubt it's related to my MR. https://gitlab.freedesktop.org/mesa/mesa/-/jobs/105703126
16:02 glehmann: I think that's a collabora job, so valentine maybe
16:05 valentine: bcheng: looks like an intermittent issue, restarted it
16:15 bcheng: thanks
16:55 bcheng: :( it failed in a different way this time
17:30 rpavlik: One last packaging message: ah, found it. looks like one of those "the preferences of the community are in the tools, not the written policy" things - there is a "lintian" tag that fires on files containing "autogenerated"-like messages. https://udd.debian.org/lintian-tag/generated-file So the requirement is to ensure there is source to clear the
17:30 rpavlik: NEW queue if ever needed (there is some degree of trust involved in updating an existing package), probably noting it somewhere (maybe README.source in the pkging) to remain clear and above reproach, and if it's in a file that looks like text rather than binary possibly adding a lintian suppression. but Mesa has so many lintian warnings already, no
17:30 rpavlik: big deal. 🎉 (yes I was working on openxr packages this morning, how could you tell?)
17:45 sghuge: valentine: Thanks for the pointer on running piglit with rusticl..
19:16 Lyude: oh no. ahahahaha i figured out why drm_vblank_init() takes a manual count of CRTCs it's so cursed
19:17 Lyude: dev->num_crtcs = num_crtcs
19:17 Lyude: we have two num_crtcs :)