04:09 alyssa: gfxstrand: Is the texture_index/texture_offset split actually doing anything for any driver?
04:09 alyssa: Every backend I've looked at is just lowering to an add if texture_index != 0 and texture_offset present
04:09 alyssa: ditto for samplers
04:10 alyssa: I wrote out a nir_lower_tex option but honestly tempted to just torch the whole thing and replace with a texture_index source or something
05:04 alyssa: jenatali: Need how/why?
05:05 jenatali: I need a constant to identify which texture/sampler array is being referenced, and then an offset within that array
05:18 alyssa: Right..
05:43 gfxstrand: alyssa: IDK. Maybe not these days.
05:44 gfxstrand: alyssa: For drivers with old enough GLES versions where non-constant indexing isn't allowed, I guess it makes a few things easier. :shrug: Also, I think nir_lower_tex uses it a bunch.
05:44 gfxstrand: (Of course, every time nir_lower_tex uses it, it's probably broken so there is that...
06:06 alyssa: The paranthetical is what gets me, yeah
06:06 alyssa: I feel like the lower_tex_packing stuff should probably be moved into v3d
10:57 Lynne: one bad side of vulkan is that it forces you into the low-level details of hardware operation, but when it comes time to actually do something useful with it and write shaders, it lifts its hands up and tells you to read spirv in text form
10:57 Lynne: or read glsl docs which are polluted by decades of opengl and don't tell you much either
11:00 Lynne: I want to know how much local memory I can have at most for each invocation in a workgroup of a compute shader, but none of the device properties expose it
12:55 DavidHeidelberg[m]: lygstate: ping
13:31 pixelcluster: Lynne: local memory = space for variables etc.? The issue with that is that it isn't just one number but depends on a lot of things
13:37 Lynne: yup, space for variables, I have to store around 64k of coeffs unless I want to redo results
13:37 Lynne: *64KiB
13:43 pixelcluster: yeah this sort of thing really depends on what the compiler does with the shader, but 64KiB in variables sounds like a really bad time
13:50 Lynne: yeah, it's a denoiser, but it's the best denoiser, so it's not very fast (~1fps on an intel 770)
13:51 Lynne: I think I can make it go at least a few times that while I'm porting it to vulkan, since it's doing around 80 opencl dispatches
15:28 DavidHeidelberg[m]: lygstate: you added c_cpp_args, that's great, but sadly it's not printed in the environment() output of meson
15:28 DavidHeidelberg[m]: we need add custom summary() line for it
15:29 DavidHeidelberg[m]: even better, we can use environment.append(c_cpp_args)
15:38 DavidHeidelberg[m]: nope, I was too optimistic, this probably wont work
15:45 dcbaker: David Heidelberg: are trying to get the ones set by add_project_args or the ones passed in by the user?
15:47 DavidHeidelberg[m]: dcbaker: in general this https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21539
15:48 DavidHeidelberg[m]: so these passed from the meson
15:48 dcbaker: Yeah, I think that’s the only way to do it
15:53 DavidHeidelberg[m]: dcbaker: ack on that? :P
15:54 dcbaker: Ack
15:55 DavidHeidelberg[m]: dcbaker: thx
16:00 dcbaker: Np
16:03 DavidHeidelberg[m]:uploaded an image: (14KiB) < https://matrix.org/_matrix/media/v3/download/matrix.org/FvsPAvuBrGsbTCTOMcmelrjc/image.png >
16:03 DavidHeidelberg[m]: for x86-64 not extra useful, but I guess for 32-bit it's informative :P
21:04 nielsdg: samueldr: depending on how high level you want it to be, there was a talk about it on FOSDEM: https://fosdem.org/2023/schedule/event/graphics_a_frames_journey/
21:20 nielsdg: Given your problem, you might want to read up on atomic mode setting, eg https://lwn.net/Articles/653071/