#asahi-gpu · 2023-09-27 14:28
eric_engestromella-0, alyssa: ^#asahi-gpu · 2023-09-27 14:28
eric_engestromit's official, it's called `VK_DRIVER_ID_MESA_AGXV` (https://github.com/KhronosGroup/Vulkan-Docs/pull/2238)#asahi-gpu · 2023-09-26 14:09
alyssaah#asahi-gpu · 2023-09-26 12:06
austriancoderagx_open_device failed and vma alloc stuff exploded :)#asahi-gpu · 2023-09-26 12:04
alyssaoh.. I haven't needed drm-shim in a year. yeah, ij#asahi-gpu · 2023-09-26 12:03
alyssaaustriancoder: little surprised it was necessary#asahi-gpu · 2023-09-26 07:59
austriancoderalyssa: got shader-db working - https://gitlab.freedesktop.org/asahi/mesa/-/merge_requests/193#asahi-gpu · 2023-09-25 21:59
alyssawhat a mood#asahi-gpu · 2023-09-25 19:46
eric_engestromhttps://github.com/KhronosGroup/Vulkan-Docs/pull/2238#asahi-gpu · 2023-09-25 19:46
eric_engestromit only took me 3 weeks to remember to do this 🙃#asahi-gpu · 2023-09-25 18:40
alyssadefinitely works on asahi/mesa#asahi-gpu · 2023-09-25 18:40
alyssaprobably not#asahi-gpu · 2023-09-25 18:40
alyssadon't remember#asahi-gpu · 2023-09-25 18:20
austriancoderShould drm-shim for shader-db be working in upstream mesa?#asahi-gpu · 2023-09-20 15:40
i509vcbFixing vk_meta there sounds reasonable, I imagine a few nvk specific details will be there anyways#asahi-gpu · 2023-09-20 15:38
alyssaand then no special handling will be needed in agxv, i think#asahi-gpu · 2023-09-20 15:38
alyssashouldn't be too hard to generalize/fix vk_meta to do the right thing#asahi-gpu · 2023-09-20 15:37
alyssaNotably, spirv_to_nir does not generate these sysvals for layers, neither does the glsl compiler. this is just a case of the vk_meta abstraction being a bit leaky and inadvertently hardcoding NVK specific details in common code#asahi-gpu · 2023-09-20 15:36
alyssaI think you want to take the helper in that pass, move it to nir_builder with the appropriate rename, and then convert vk_meta to use it to ensure agxv never sees load_layer_id intrinsics#asahi-gpu · 2023-09-20 15:35
alyssasee for contrast nir_lower_input_attachments which has a load_layer_id helper that takes driver nir options to generate a varying LAYER load instead of a sysval load if applicable#asahi-gpu · 2023-09-20 15:34
alyssarather than conditionally doing so only if the driver wants layer IDs as sysvals#asahi-gpu · 2023-09-20 15:34
alyssavk_meta is hardcoding a load_layer_id intrinsic inappropriately#asahi-gpu · 2023-09-20 15:32
alyssaI see the problem now#asahi-gpu · 2023-09-20 15:32
alyssai509vcb: ahhh#asahi-gpu · 2023-09-20 15:31
alyssahmm#asahi-gpu · 2023-09-20 15:31
alyssaa#asahi-gpu · 2023-09-20 14:23
i509vcbI guess I could do tree wide changes but then I have to figure out what other drivers expect...#asahi-gpu · 2023-09-20 14:22
i509vcbone thing I did notice is that layer has a gl_varying_slot, wouldn't that imply it's already a varying or is the common infrastructure not capable of expressing layer as a sysval currently#asahi-gpu · 2023-09-20 14:07
alyssa:\#asahi-gpu · 2023-09-20 01:01
i509vcbSomething about LAYER not having a sysval value feels weird#asahi-gpu · 2023-09-20 00:01
i509vcbI don't see any layer id enum variant in gl_system_value#asahi-gpu · 2023-09-19 23:55
i509vcbOr I guess I am adding new things #asahi-gpu · 2023-09-19 23:54
i509vcbsysvals_to_varyings has no layer option?#asahi-gpu · 2023-09-19 23:45
alyssafor us LAYER is a varying, not a sysval#asahi-gpu · 2023-09-19 23:45
alyssabefore the predication pass etc#asahi-gpu · 2023-09-19 23:45
alyssai509vcb: run nir_lower_sysvals_to_varyings with the appropriate modification to support layer #asahi-gpu · 2023-09-19 23:42
i509vcbThe meta shader uses an actual raw load_layer_id#asahi-gpu · 2023-09-19 23:42
i509vcbAlthough I am definitely still confused as to why agx_nir_predicate_layer seems to never lower load_layer_id in the blit meta shader #asahi-gpu · 2023-09-19 23:41
alyssa:+1:#asahi-gpu · 2023-09-19 23:37
i509vcbSure I guess that makes sense#asahi-gpu · 2023-09-19 23:22
alyssa5#asahi-gpu · 2023-09-19 23:21
alyssakapeesh? :)#asahi-gpu · 2023-09-19 23:21
alyssaI should probably simplify this in the gl driver but whatever, it's not really hurting anything there..#asahi-gpu · 2023-09-19 23:20
alyssathe layered bool matters a lot more for the background and EOT programs, but those are independent of pipeline and you should have that information available for them just fine, I think#asahi-gpu · 2023-09-19 23:19
alyssabut if you're spilling attachments you're already screwed so that seems like a reasonable tradeoff to me lol #asahi-gpu · 2023-09-19 23:18
alyssabut if you just always use a 2D Array PBE descriptor for that (with layers=1 if necessary) that should still work out fine, at a slight cost to instruction count in the spilled path#asahi-gpu · 2023-09-19 23:17
alyssathe only case where it would matter is for spilled render targets#asahi-gpu · 2023-09-19 23:16
alyssawhen compiling frag shaders, just asssume the framebuffer is always layered#asahi-gpu · 2023-09-19 23:14
alyssai509vcb: it doesn't look like you necesarily have that information but you dont really need it#asahi-gpu · 2023-09-19 20:38
i509vcbthis is going several layers deep in vulkan CreateInfo types...