05:37 rmckeever: jekstrand I'm trying to enable VK_EXT_non_seamless_cube_map and when I run the relevant CTS tests with ./deqp-vk -n *non_seamless*, the tests say:
05:37 rmckeever: NotSupported (Format does not support shadow sampling at vktTextureShadowTests.cpp:1367)
05:37 rmckeever: Do I have to do something additional to enable shadow sampling?
08:40 fdobridge: <D​adSchoorse> rmckeever: there is a todo for VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_DEPTH_COMPARISON_BIT in nvk
08:40 fdobridge: <D​adSchoorse> and the cts checks for that
08:58 rmckeever: I found the todo. Thanks!
18:47 KungFuJesus: Hmm, the endianness issues seemed to have gotten worse on big endian
18:48 KungFuJesus: I'm building with debug symbols now to find the cause of an immediate segfault but it seems like more stuff on the render path is exhibiting byteswapped textures than before :-/
18:49 KungFuJesus: I was hoping imirkin would be on here today since he's fought these issues before, too
18:49 KungFuJesus: (This is on NV47)
18:53 KungFuJesus: Guess it's time for me to step up and be an endian correctness warrior :-p. I may have one of the few remaining working big endian powerpc systems with a functioning nvidia GPU
19:02 KungFuJesus: anything fishy with this backtrace? https://bpa.st/7YHTM
19:10 KungFuJesus: funny, it's in a builtin_memcpy call. My guess is that something writing out a copy size into the attribute is doing something wonky with endianness
19:33 qeeg: hey, does anybody here know how the addresses for where to write notifier objects to are calculated on nv3?
19:34 qeeg: there's little to no documentation of pdma in envytools so :/
22:13 mwk: qeeg: it's notifier_idx * 0x10
22:13 mwk: within the notifier dma object
22:15 mwk: and the notify DMA object is specified by high 16 bits of the second word of the graph object in RAMIN
22:16 mwk: (the hw doesn't provide any method to set it; it's done via a software method emulated by the driver)
22:18 mwk: oh, and notifier_idx is the parameter to the NOTIFY (0x104) method; it has to be in range 0-15, so effectively you have 16 distinct notifiers to use per graph object
22:19 mwk: (and note that it only works that way on NV3; NV1 and NV4+ are both quite different)
23:34 qeeg: mwk, wait, but then what does the pte stuff in pdma do?
23:40 KungFuJesus: surprising the issue doesn't seem to be an issue with sizes but rather one where the src index is misset somewhere. It appears to be whatever the caller to run_elts is