#
#asahi-gpu · 2024-01-08 23:43
i509vcbOnly the first layer is correct
#asahi-gpu · 2024-01-08 23:21
i509vcbtime to open up gdb again to compare the memory
#asahi-gpu · 2024-01-08 23:17
i509vcbhttps://usercontent.irccloud-cdn.com/file/1xlF2ywk/image.png
#asahi-gpu · 2024-01-08 23:17
i509vcbsome more basic 3d related copy tests are giving me the funny results again
#asahi-gpu · 2024-01-08 22:30
alyssa2 bugs, why not?
#asahi-gpu · 2024-01-08 22:30
i509vcbhnm no different qpa output
#asahi-gpu · 2024-01-08 22:24
i509vcbI guess the bug for layered rendering is elsewhere, setting viewport_target and render_target based on the vs changes nothing
#asahi-gpu · 2024-01-08 17:51
i509vcbYou might be able to get away with initializing a zeroed struct of that type on the stack and then doing a memcpy
#asahi-gpu · 2024-01-08 17:50
knedlikI'll take a look, thanks
#asahi-gpu · 2024-01-08 17:49
i509vcbYou probably want to zero the formatProperties field
#asahi-gpu · 2024-01-08 17:48
knedlikWould the zero properties be just 0 in memory, or some different default value?
#asahi-gpu · 2024-01-08 17:47
knedlikOh right
#asahi-gpu · 2024-01-08 17:47
i509vcbI think you'll want memset? Or was it called something else in C?
#asahi-gpu · 2024-01-08 17:47
i509vcbknedlik: setting the pointer to zero does not change the data referenced by the pointer
#asahi-gpu · 2024-01-08 16:58
knedlikSo I added a guard clause that checks if the format is equal to the undefined value (I believe it is 0, not sure though), and if it is, set the pointer to a nullptr and return. Unfortunately, that does not solve the error of returning non-zero properties. Any ideas? I'm really learning a lot about mesa as I go, so please excuse me if this is something obvious.
#asahi-gpu · 2024-01-06 21:36
alyssaNVK is the reference drivr
#asahi-gpu · 2024-01-06 20:55
knedlikShould I make a pull request with the change to agxv/main once finished?
#asahi-gpu · 2024-01-06 20:54
knedlikI'm starting to get tired, but tommorow I'll do a quick test if the guard clause clears the test
#asahi-gpu · 2024-01-06 20:42
i509vcbThe properties are returned via one of the pointers
#asahi-gpu · 2024-01-06 20:42
i509vcbI guess we could just do the quick guard clause
#asahi-gpu · 2024-01-06 19:49
knedlikNevermind, it's a pointer... :facepalm:
#asahi-gpu · 2024-01-06 19:49
knedlikI need clarification here - the error is about returning non-zero properties, but the function is declared as void, where exactly is it returning the properties?
#asahi-gpu · 2024-01-06 19:42
knedlikI have a suspicion NVK does it in the image functions, which means we could just do a guard clause assuming it won't be needed elsewhere
#asahi-gpu · 2024-01-06 19:08
knedlikAlright
#asahi-gpu · 2024-01-06 19:08
i509vcbI guess take a peek at nvk
#asahi-gpu · 2024-01-06 19:08
i509vcbUhh I'd wonder what other drivers do here actually
#asahi-gpu · 2024-01-06 19:05
knedlikI'd assume it should be as easy as adding a guard clause to check for the undefined format?
#asahi-gpu · 2024-01-06 19:04
i509vcbMesa is also used for several other vulkan implementations so people do want to yeet some boilerplate
#asahi-gpu · 2024-01-06 19:03
knedlikIt makes sense for a project like this to be honest
#asahi-gpu · 2024-01-06 19:03
i509vcbTLDR: Mesa makes a big function pointer lookup table for us
#asahi-gpu · 2024-01-06 19:03
knedlikThere it is, I found it, thanks for the help
#asahi-gpu · 2024-01-06 19:02
knedlikThat makes sense now
#asahi-gpu · 2024-01-06 19:02
i509vcbPretty much look at agxv_GetPhysicalDeviceFormatProperties2
#asahi-gpu · 2024-01-06 19:01
i509vcbYou just need to do stuff inside the agxv_ prefixed functions
#asahi-gpu · 2024-01-06 19:01
i509vcbIt's all linked for you
#asahi-gpu · 2024-01-06 19:00
knedlikSo I'm supposed to patch some function outside the asahi-specific tree and point it to some file in the asahi tree?
#asahi-gpu · 2024-01-06 18:58
i509vcbSearch for vk_common_GetPhysicalDeviceFormatProperties to see where that forwarding is done
#asahi-gpu · 2024-01-06 18:56
i509vcbVulkan has a few of these vkWhatever2 functions since 1.0 didn't consider adding pointer chains to some types or expected all data to be specified via a single fixed signature
#asahi-gpu · 2024-01-06 18:55
i509vcbThat function is delegated to the FormatProperties2 function by mesa for us
#asahi-gpu · 2024-01-06 18:54
knedlikHmm, I may be blind, but is the function not implemented yet at all? I'd assume "returning non-zero properties" doesn't happen in an unimplemented function
#asahi-gpu · 2024-01-06 18:47
knedlikI'll try taking a look at the 'vkGetPhysicalDeviceFormatProperties' from line 49 as you said
#asahi-gpu · 2024-01-06 18:46
knedlikAlright, I was using the binary directly without deqp-runner, so I guess that was my issue
#asahi-gpu · 2024-01-06 18:38
i509vcb(That "Fail (Fail)" dedup is doing some heavy lifting)
#asahi-gpu · 2024-01-06 18:33
i509vcbdeqp-runner gives you a results and failures.csv with every test name and whether it failed, crashed, flakes or passed
#asahi-gpu · 2024-01-06 18:33
i509vcbI guess the problem you have there is that you know how many failed but not what specific test did
#asahi-gpu · 2024-01-06 18:32
i509vcbAnd the tests below testing for required format feature bits
#asahi-gpu · 2024-01-06 18:32
i509vcbhttps://gist.github.com/KnedlikMCPE/4c4fffa220a668d0bb48851c77bb2062#file-deduped-txt-L49 seems like something you could do
#asahi-gpu · 2024-01-06 18:31
i509vcbhttps://gist.github.com/KnedlikMCPE/4c4fffa220a668d0bb48851c77bb2062#file-deduped-txt-L48 ignore this, that's something to fix when we are 90% done
#asahi-gpu · 2024-01-06 18:30
i509vcbYou can run individual tests with the -n argument on dEQP-vk to avoid running the entire suite every time
#asahi-gpu · 2024-01-06 18:29
i509vcbknedlik: VK_EXT_global_priority(_query) the promoted KHR extension for that is implemented. If you did want to implement the older EXT extension I think you just need to enable those two extensions and test if anything changed during promotion?

plus ›