14:08 alyssa: karolherbst: ..thanks, I hate it
14:08 karolherbst: I have a good reason for it!
14:09 karolherbst: for some reasons non uniform workgroups are supported with reqd_work_group_size kernels and the workgroups on the edges....
14:13 alyssa: karolherbst: re VK_fma, what drivers would even need the emulation?
14:13 karolherbst: turnip because the hw doesn't have it :)
14:13 alyssa: uhhh
14:13 karolherbst: I understand
14:14 karolherbst: I've been there myself
14:15 karolherbst: though I guess "need" is a strong word there, we can just hope it's not going to be needed :)
14:19 glehmann: v3dv doesn't have fma either
15:36 karolherbst: I'm once again annoyed that load_global_constant and load_global CAN_REORDER mean the same thing
15:39 zmike: anholt_: idk if you're still interested in https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24262, but https://docs.vulkan.org/refpages/latest/refpages/source/VK_EXT_multisampled_render_to_swapchain.html just released today
15:43 simon-perretta-img: Does nir already have a way to sort/remap varyings in order of their xfb buffer offsets?
15:45 mareko: yes, search for qsort
15:54 simon-perretta-img: Cheers
16:25 anholt_: zmike: interested but probably not working on it unless we have a demand?
16:26 zmike: anholt: don't think there's any demand atm, just came to mind
17:06 alyssa: daniels: your trolling worked
17:09 daniels: alyssa: I’m every bit as delighted as I was unprepared for this eventuality
17:10 alyssa: lol
17:10 alyssa: is it really that surprising? I like nice UX
17:10 daniels: oh not that, just the successful trolling
17:10 alyssa: ah
17:10 alyssa: I have accumulated piles of one off scripts for r-b tags and drm-shim and such
17:11 daniels: I should probably rebrand that as thought leadership tho
17:11 alyssa: like my ~/bin was littered with script names like `t860` and `m1` as well as `frb`, `brb` (Faith/Boris r-b) etc
17:11 alyssa: deleted the latter set with rb.py
17:11 alyssa: can delete the former set now too
17:12 alyssa: well mostly. I also have scripts for things like "zink on turnip"
17:13 alyssa: I guess I just need a --zink option
17:13 alyssa: implied by zink job names or something
17:13 alyssa: idk I already spent more time on this than I intended :p
17:37 zmike: tfw gitlab has heavy_plus emoji but no heavy_approve button
17:46 MadDogTen: Anyone mind looking at a ggml-org/llama.cpp issue (23995) to verify if it's actually a RADV/Mesa regression?
17:46 MadDogTen: ChatGPT could only get me so far.
18:18 willy: i915_gem_phys.c has some dodgy looking code in it
18:19 willy: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/i915/gem/i915_gem_phys.c#n98 first drew my attention. it looks to me like it's copying an array of struct page pointers into the shmem page, rather than the data in the struct page
18:20 willy: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/i915/gem/i915_gem_phys.c#n142 is also suspect. sg_page() returns a struct page pointer, so adding args->offset means we'll scale the addition by sizeof(struct page)
18:36 willy: ok, so line 61 explains the first. sg_assign_page(sg, (struct page *)vaddr);
18:37 willy: abusing the sg list. nice, totally not confusing, definitely doesn't need to be commented
18:37 willy: but the second one still seems like a bug to me. sizeof(struct page) is not ABI and can change with .config
21:44 karolherbst: do we have an opt pass that folds pack_64_2x32 into load instructions to turn them into 64x1 instead of 32x2 or something?
21:45 karolherbst:goes adding nir_opt_pack...