02:06karolherbst: another day goes by where I wished meson would just allow me to set env variables for build targets :')
02:08karolherbst: I want to include a SPIR-V file in rusticl, but because rustc doesn't have something like defines and relies on env variables, and meson doesn't want to support it, I have to move it into a C file and bindgen it..
12:42osdev199: Hi, in linux kernel, could we display the panic logo at the top center of the user's panic screen (when DRM_PANIC_SCREEN = user)?
12:43osdev199: I think it will make things more aligned on this screen unless we are trying to achieve a universal UI for all panic screens.
17:38karolherbst: alyssa: what was the proper way of getting nir intrinsic available in CL files again? Do I have to forward declare them or what?
18:18alyssa: karolherbst: forward declare them and they'll magically resolve by nir_lower_calls_to_builtins.c
18:18alyssa: grep src/asahi/libagx for `nir_` for examples
18:18karolherbst: ohh, I'm not calling nir_lower_calls_to_builtins!
18:18karolherbst: so that's why it didn't work...
18:18alyssa: and that pass has a comment explaining it
18:19karolherbst: I'm mostly interested in using globalinvocation id without adding the base, and in theory we have NoGlobalOffsetINTEL from SPV_INTEL_kernel_attributes for that.. but there is no CL C -> SPIR-V path for it 🙃
18:20karolherbst: mhhh
18:24alyssa: yeah lower_call_to_builtins is the path of least resistance hre
18:25karolherbst: mhhhhhhh okay soo.. that also doesn't work for me. I'm doing it Meta style, so I end up creating a real cl_program executable, which means it's not a library, which means the SPIR-V linker complains about unresolved symbols..
18:25karolherbst: I... need to think about this a bit I think..
18:30alyssa: sorry I don`t have a facebook account, can't help with Meta issues
18:31karolherbst: okay.. so if I have a stub impl returning 0, nir_lower_calls_to_builtins will still replace it, but that sounds like a horrible workaround. And anyway.. lower_compute_system_values still replaces my nir_load_global_invocation_id with itself + a base :')
18:32karolherbst: I guess I need to get back to my other idea: add `-cl-rusticl-no-offsets` internal compiler option and just disable offsets...
18:33karolherbst: in other news, I'm starting to write internal helper kernels in CLC... and it's surprising it took that long for me to do that 🙃
18:38alyssa: at least I don`t need to convince you to use write your internal shaders in opencl c :p
18:40karolherbst: the final reason is, that I need clear_buffer to support 32,64 and 128 byte clear patterns and some drivers just don't support it, and while fixing this, might as well speed up all those that use "u_default_clear_buffer" with a compute shader..
18:41karolherbst: the other things I want to accelerate: u_default_buffer_subdata and u_default_texture_subdata
18:44karolherbst: in other news, a new conformant driver incoming (soon): "Pass 4097 Fails 0 Crashes 0"...
18:45alyssa: which ça
18:47karolherbst: radeonsi
18:53alyssa: cool