00:53anholt: robclark: looks like wine's actual process don't show up as a child of the wine invocation, so fdinfo is a bust in the case I really wanted it for.
00:55robclark: idk if this helps, but /sys/kernel/debug/dri/0/clients gives you a list of the processes that have drm device file open.. I guess that only helps if you can chase up the process tree to find whatever you need to
01:18karolherbst: I'm having some weird issues with 2D image reads.. no idea why, but the other dimensions work properly, and only with 2D ones, if the pixel size is below 4 bytes I get invalid data... I'm wondering if there are some restrictions on the GL side what are valid formats for images?
01:19karolherbst: e.g. using the PIPE_FORMAT_R16_SNORM format doesn't give me the expected results, but e.g. PIPE_FORMAT_R16A16_SNORM does
01:20karolherbst: though the format isn't known at compile time, but I do require image_load_formatted
01:22karolherbst: though it also just works as expected with llvmpipe.. maybe it's a driver bug
01:26karolherbst: robclark: you do support EXT_shader_image_load_formatted on freedreno, right?
01:27karolherbst: Mind running the "test_conformance/images/kernel_read_write/test_image_streams read_write CL_FILTER_NEAREST" CL CTS test?
01:28robclark: yes we do.. it was passing last time I ran cl cts
01:30robclark: for array formats, like 2d_array, we need to configure hw a bit differently for cl vs gl for array coord rounding mode
01:30robclark: fwiw
01:31karolherbst: mhhhhhh
01:31karolherbst: right.. I'm seeing it failing on radeonsi and only for 2D
01:31karolherbst: and only for 2 and 1 byte formats
01:32karolherbst: let me test on my macbook just to be sure
01:33robclark: for sure cl is a bit more picky.. so could be driver bug ig
01:33karolherbst: yeah.. or maybe just missing test coverage in the GL CTS :)
01:34karolherbst: I was considering disabling read_write images, but maybe I just add driver workarounds and make that all more targeted...
01:35zmike: https://registry.khronos.org/OpenGL/extensions/ARB/ARB_shader_image_load_store.txt R16_SNORM is on the list
01:35karolherbst: okay.. it passes on asahi as well
01:35karolherbst: soo.. works on llvmpipe, asahi and freedreno, doesn't on radeonsi...
01:36karolherbst: "zink Vulkan 1.4(Apple M2 (G14G B0)" amazing
01:37karolherbst: naturally it also runs perfectly on zink + honeykrisp
01:40karolherbst: guess it's a driver bug indeed then..
01:42karolherbst: mhhh zink + radv also works
01:51karolherbst: AMD_DEBUG=notiling fixes this...
01:52karolherbst: AMD_DEBUG=dccstore and AMD_DEBUG=nodcc as well
02:01karolherbst: zmike: but the ARB extension requires format declarations in the shader, right?
02:03karolherbst: anyway, filed an issue: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15218 02:08karolherbst: robclark: oh right I think you've mentioned this in the past.. anything we'd need to change on the rusticl side for that?
02:14robclark: not really.. just a KERNEL vs COMPUTE shader type thing
02:15karolherbst: mhhh
02:15karolherbst: at set_shader_images time?
02:15karolherbst: or when would you ened it
02:16karolherbst: guess sampler views as well
02:16karolherbst: I _guess_ once could fix up all the drivers and make it use mesa_shader_stage_is_compute instead of a direct compare, like radeonsi seems to do for shader images for some stuff
02:16robclark: karolherbst: 005ce4057c067fff262594f9ee06746875b4830f
02:17robclark: part of the program state
02:17karolherbst: ohhh
02:17karolherbst: okay, that makes it simple for me
02:17karolherbst: I won't have to change anything :)
02:17robclark: right :-)