#asahi-gpu · 2026-08-23 19:55
oliverYea, iirc it did not really make a distinction there, it was "only subtitles" in raster order. So only 16x16px tiles in raster order.#asahi-gpu · 2026-08-23 04:51
oliverchadmed: awesome! #asahi-gpu · 2026-03-10 13:00
oliverSo more like this: https://files.narf.zone/dfb0cf5b3c31fab3/check-stride.patch#asahi-gpu · 2026-03-10 12:56
oliveri feel that it would be better to not have different strides for the different modifiers and to just adjust the check in apple_plane_atomic_check to not check the stride with interchange. As the comment says, it is "fundamentally not meaningful" anyways.#asahi-gpu · 2026-03-10 12:45
oliver<jannau> "it's currently not wrong but I think we want to align the stride to tile width * blocksize in the modifier documentation. Same for the height" You're talking about changing the comment for the existing modifiers, or just interchange?#asahi-gpu · 2026-03-08 09:15
oliverjannau: is this fine for you: https://files.narf.zone/60ab0d2640361b88/mesa.patch#asahi-gpu · 2026-03-08 09:07
oliveryes, confirmed, works as expected.#asahi-gpu · 2026-03-08 09:00
oliverjannau: I imagine that your commit would help, yes. #asahi-gpu · 2026-03-08 07:27
oliverWe can now either "break" userspace and change the comment (and adjust mesa too) or update the atomic_check and only verify stride for linear, because we know tiled will have proper stride. Which one should it be? #asahi-gpu · 2026-03-08 07:03
oliverAnd due to that userspace sets pitch based on the width, not rounded up like jannau suggested. #asahi-gpu · 2026-03-08 06:57
oliverThe pitch is actually 64 bytes aligned, it just does not match the comment. #asahi-gpu · 2026-03-08 06:39
oliverWhat do you guys thing is the best way to go here? Only check the stride for the linear modifier? #asahi-gpu · 2026-03-08 06:37
oliverThis then fails in atomic_check for textured that do not have a stride of a multiple of 64 bytes. #asahi-gpu · 2026-03-08 06:36
oliver * stride is arbitrary here, we pick the simplest convention. #asahi-gpu · 2026-03-08 06:36
oliver * therefore require that stride equals (width * bytes per pixel). Since the#asahi-gpu · 2026-03-08 06:36
oliver * software assumes the stride is at least (width * bytes per pixel). We#asahi-gpu · 2026-03-08 06:36
oliver * convention would be setting the image stride to 0. Unfortunately, some#asahi-gpu · 2026-03-08 06:36
oliver * Userspace aware of the image layouts do not use strides. The most "correct"#asahi-gpu · 2026-03-08 06:36
oliver * will be unable to use correctly the specified stride for any purpose.#asahi-gpu · 2026-03-08 06:36
oliver * specify strides for these layouts, userspace unaware of Apple image layouts#asahi-gpu · 2026-03-08 06:36
oliverAlyssa comments on the tiled modifiers: These layouts fundamentally do not have meaningful strides. No matter how we#asahi-gpu · 2026-02-22 14:10
oliverAlso, unrelated, the value in the metadata/accel buf is the length in bytes of the corresponding 8x4px subtile minus one.#asahi-gpu · 2026-02-22 14:07
oliverI looked into the compression of 16x16 tiles a little. I am confident that the actual compression of the data is exactly the same between gpu tiled and interchange. The only difference is in the order of 8x4px subtiles. with interchange they are ordered left to right, top to bottom. With gpu tiled the subtiles are in morton order.#asahi-gpu · 2026-02-08 19:31
oliverBut that I'll exclude for now. #asahi-gpu · 2026-02-08 19:21
oliverfun fact: rgb with 16 bit per channel is also chunked into 16x16px tiles.#asahi-gpu · 2026-02-08 19:21
oliverI realized I can just tell mesa not reject a modifier for a specific format, did that. now the cts seem to run fine for !39755.#asahi-gpu · 2026-02-08 12:47
oliverhttps://files.narf.zone/42504d42556d0ff9/failures.csv#asahi-gpu · 2026-02-08 12:27
oliver(or all non 32bbp formats with a non one block size)#asahi-gpu · 2026-02-08 12:27
oliverthat one leads to more failures, it will crash for yuv formats. #asahi-gpu · 2026-02-08 12:27
oliverhere is vulkan: https://files.narf.zone/43fb7f449ad6d6a0/0001-hk-add-support-for-interchange_compressed.patch#asahi-gpu · 2026-02-08 12:27
oliverbut, like i said, the changes do not affect vulkan at all, it is just gl.#asahi-gpu · 2026-02-08 12:10
oliver(on mesa main, not rc)#asahi-gpu · 2026-02-08 12:08
oliverhttps://files.narf.zone/27f02c5aa3e17d58/failures.csv#asahi-gpu · 2026-02-08 12:07
olivertest_vk finished with some 42 failures on mesa 26.0.0-rc3. i have a g13s according to glxinfo. also my changes did not affect vulkan at all. i'll run it again with the changes applied to the vk driver too.#asahi-gpu · 2026-02-08 10:54
oliverbest_modifier seems to be only used in the context of the "screen". when checking with ASAHI_MESA_DEBUG i see it only used for the surface, all the other textures allocated are either LINEAR or GPU_TILED ... which is fine imho#asahi-gpu · 2026-02-08 10:37
oliveri have one question: the way i understood mesa, INTERCHANGE should only be used for surface textures with my change, like with the "window" or the "desktop", everything else should still use GPU_TILED as before. is this covered when running those tests?#asahi-gpu · 2026-02-08 10:37
oliverperfect, thank you#asahi-gpu · 2026-02-07 21:54
oliverchanges to the cts?#asahi-gpu · 2026-02-07 21:44
oliverWhat do you mean by the last part, i.e. what kind of changes would that be?#asahi-gpu · 2026-02-07 21:44
oliverNo, sorry, I have no experience with any of this. If you can give me a few pointers, that would be great. #asahi-gpu · 2026-02-07 21:20
oliverI've created pull requests in mesa and asahi kernel for the compressed scanout.#asahi-gpu · 2026-02-05 07:43
oliverwe would need to get a compressed yuv framebuffer from somewhere for it to matter. it is probably possible somehow, i bet macos does that with direct scanout of yuv video, maybe directly from some decoder coprocessor. but if we currently can not do it, we might not need to care right now.#asahi-gpu · 2026-02-05 07:33
olivercan we render to yuv on the gpu side of things?#asahi-gpu · 2026-02-05 07:30
oliverthat is just an alias in the dcp firmware afaik#asahi-gpu · 2026-02-05 07:28
oliverno#asahi-gpu · 2026-02-05 07:15
oliverI think that tile size here is always 16x16px#asahi-gpu · 2026-02-05 07:14
oliverhttps://gist.github.com/oliverbestmann/8b2d24c15450ebd3021f688ae70af47e#asahi-gpu · 2026-02-05 07:13
oliverand as you said earlier, maybe the flip is now faster, due to it needing to transfer less data#asahi-gpu · 2026-02-05 07:12
oliverit is not much more. i can share my hacky kernel diff in a sec#asahi-gpu · 2026-02-05 07:00
oliverin that mode there is no z-curve twiddling, just compressed 16x16 tiles aligned in scanout order, i.e. from left to right, top to bottom. this is exactly the same mode that macos is using to scanout its framebuffers