01:09 DUOLabs[m]: However, when vkCreateInstance is called on the host, QEMU crashes (this is on the guest side).
02:02 sassefa: Kayden: Thanks!
10:26 doras: I'd appreciate it if anyone could assign this to Marge: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23255
10:36 DavidHeidelberg[m]: Dor Askayo: thank you, I think it solves my issue I had and few I seen in the issues, but from missing description of the change it's hard to guess.
10:40 doras: David Heidelberg: I'll reply in the MR.
10:40 DavidHeidelberg[m]: Dor Askayo: edit the description of it ;-) thank you
11:58 doras: David Heidelberg: do you mind sharing what were the issues you're thinking of?
12:08 DavidHeidelberg[m]: Dor Askayo: I found that issue got resolved already (it was missing `opencl-c.h`), I had this too when doing some combination of libclc + microsoft-clc
12:08 doras: Ah.
12:24 jenatali: Dor Askayo: There's a general guideline to wait 24 hours, not including weekends, before merging changes to common code, to give other reviewers a chance to chime in
12:24 jenatali: So I was waiting for that before assigning
15:12 Lynne: I think I found a fundamental issue with how video decoding works
15:13 Lynne: when decoding, you decode to size-aligned images, with the padding being generally at the bottom
15:13 Lynne: the decoder decodes, leaving junk at the bottom in the padding
15:14 Lynne: but because the vkimage is oversized, reading into the padding is valid
15:15 Lynne: so when you do bilinear sampling at the very bottom, the bottom-most line will contain junk
15:15 Lynne: and if you use a multi-tap sampler, it's even worse
15:16 Lynne: you can't alias non-disjoing multiplane images either, so it's not like you could bind the memory to a regular-sized vkimage (the layout will be a mismatch)
15:40 jkqxz_: Lynne: Is there not some way to clamp at the edges to avoid that? I remember hitting exactly the same issue in D3D a while ago (manifesting as greenish borders at right/bottom), but it was easily avoided once you knew about it.
15:43 Lynne: only way I can think of is for users to always use a compute shader with a boundary condition, which is very vulkan indeed
15:44 Lynne: imageviews would be the correct place to put a clamp like that, but I didn't see any way for that
15:47 Lynne: the most correct way is of course to just ask the driver to do the padding, but no such option
15:52 jenatali: Yeah I'm not aware of any way to clamp to an arbitrary coordinate that's not the edge of the image
15:52 jenatali: In D3D or VK for the record
15:57 jkqxz_: Trying to remember the D3D case I had, I think that the answer was actually that it was shared with a different device instance (because asynchronous threading), and then the texture on the second device used the actual size.
16:23 glehmann: do we have a nir pass that does if (inot(cond)) {block2} else {block3} -> if (cond) {block3} else {block2} somewhere?
16:24 glehmann: I only found a pass in nir_opt_if that does the opposite if the if block is empty
17:32 Lynne: any ideas for a workaround?
22:24 benjaminl: hi, I'm currently working on getting rusticl running on my ivy bridge laptop. is this the kind of thing that would be accepted upstream, or is the maintenance effort for supporting the old hardware too much?
22:40 emersion: Lynne: is it possible to create another image aliasing the same VkMemory, same width/stride but smaller height?
22:42 Lynne: nope, layout would change
22:44 Lynne: there's no way to ask for a specific stride for an image either
22:46 Lynne: how did no one think this may be a problem? the extension was first published 2 years ago
22:50 Lynne: it'll be at least 6 months of discussion at TSG pace to fix this, and knowing them, it'll be a yet another capability, because what if some hardware/vendor is unable to?
22:53 Lynne: this is a pretty much a fiasco
22:54 psykose: drama alert in linux graphics
22:58 Lynne: it's a strategy, raise enough stink, and hope on monday they say "we're terrible at this video malerkey, could you tell us what to do?"
23:02 Lynne: but it is pretty much rendering the extension useless without big workarounds that everyone would say "no, I'll use vdpau, interlaced-only output sounds reasonably sane"
23:13 psykose: vdpau was ahead of its time