00:43fdobridge: <airlied> Test case 'dEQP-VK.api.copy_and_blit.core.buffer_to_depthstencil.d32_sfloat_s8_uint_DS'..
00:43fdobridge: <airlied> Pass (CopiesAndBlitting test)
00:44fdobridge: <airlied> leet haxor
01:00fdobridge: <airlied> @gfxstrand care to throw https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/217 at a cts run on your hw and see if anything explodes?
01:00fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Let's try this with Overwatch 2 😅
02:24fdobridge: <karolherbst🐧🦀> I wonder if that's indeed cheaper than using shaders...
02:24fdobridge: <karolherbst🐧🦀> but maybe it is
02:28fdobridge: <airlied> I've no idea if that path is even something apps would care about in practice
02:28fdobridge: <karolherbst🐧🦀> some probably do
02:29fdobridge: <airlied> like reading back depth stencil to the CPU side doesn't seem like a major thing to do
02:29fdobridge: <airlied> maybe uploading preinited stencil textures is a thing though
02:30fdobridge: <karolherbst🐧🦀> yeah.. no idea
02:30fdobridge: <karolherbst🐧🦀> at least the copy engine is real, so I wouldn't be surprised it runs alongside shaders
02:30fdobridge: <karolherbst🐧🦀> *if it
02:31fdobridge: <karolherbst🐧🦀> heck.. there are GPUs with like 10 of those things
02:31fdobridge: <karolherbst🐧🦀> really wondering if we have to do something to get automatic parallelism going or if there is some magic we have to do
02:35HdkR: Titan V with its ten async copy engines is great
02:35karolherbst: probably
02:36karolherbst: mhhhhh
02:36HdkR: As long as you know that the copy can be async, otherwise it'll be worthless
02:36karolherbst: I wonder...
02:36karolherbst: sooo
02:36karolherbst: wild guess
02:36karolherbst: NVC7B5_LAUNCH_DMA_DATA_TRANSFER_TYPE_PIPELINED + setting a sempahore we can wait on is the way to do it
02:36karolherbst: or maybe NVC7B5_LAUNCH_DMA_INTERRUPT_TYPE_NON_BLOCKING
02:36karolherbst: whatever
02:38karolherbst: mhhh we only use NON_PIPELINED
02:38karolherbst: I wonder....
02:39karolherbst: If I'd have time I'd even look into it :)
02:40fdobridge: <karolherbst🐧🦀> @gfxstrand if you feel bored, it might be worth checking if PIPELINED dma copies can be make to work if we set a semaphore we wait on when the result is needed
02:40fdobridge: <gfxstrand> Mind filing a bug?
02:40fdobridge: <karolherbst🐧🦀> sure
02:41fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> After some sleep of course 🐸
02:46fdobridge: <karolherbst🐧🦀> https://gitlab.freedesktop.org/nouveau/mesa/-/issues/69
03:11fdobridge: <gfxstrand> Thanks
03:11fdobridge: <gfxstrand> Yeah, right now we're REALLY WFI-happy.
03:11fdobridge: <gfxstrand> We need to clean that up
03:12fdobridge: <karolherbst🐧🦀> I suspect for that to properly work we kinda have to be able to reorder things
03:12fdobridge: <gfxstrand> I've been more focused on correctness than those details so far. That really needs to be cleaned up.
03:12fdobridge: <karolherbst🐧🦀> yeah, fair
04:32fdobridge: <airlied> @gfxstrand !202 is probably also trivial to land
04:39fdobridge: <gfxstrand> I'm run them tomorrow
05:39fdobridge: <airlied> cool thx
08:22fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> This commit was remade with code styling changes (I think)
10:07fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> `m Files (x86)/Overwatch/_retail_/Overwatch.exe: ../mesa/src/nouveau/nil/nil_image.c:407: nil_image_init: Assertion 'nil_format_supports_color_targets(dev, info->format)' failed.` 🤔
10:10fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> The modifier change strikes again 🐸
10:21fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> After reverting that I now get this classic error: `m Files (x86)/Overwatch/_retail_/Overwatch.exe: ../mesa/src/nouveau/vulkan/nvk_descriptor_set.c:64: write_image_view_desc: Assertion 'view->storage_desc_index < (1 << 11)' failed.`
10:40fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> And after disabling the asserts I get this: `kernel: nouveau 0000:01:00.0: fifo: PBDMA0: 00040000 [PBENTRY] ch 4 [00ffa33000 Overwatch.exe[252519]] subc 0 mthd 1d00 data 00000000`
10:50fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> I think I got a bit further with `NV50_PROG_OPTIMIZE=0`: https://pastebin.com/as80zRPM
13:50fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Apparently older DXVK versions need VK_FORMAT_R4G4B4A4_UNORM_PACK16 🫃
13:50fdobridge: <![NVK Whacker] Echo (she) 🇱🇹>
13:50fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Let's add this if possible
14:00mntirc: hmm, trying blender 3.4 on an arm machine with GeForce NV106 (0x800b) results in: nir_alu_instr only supported with 1 component!
14:00mntirc: (and segfault)
14:02fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Or maybe swizzles are problematic for some reason: https://github.com/doitsujin/dxvk/commit/a6d554615e5f939826db758263f9b59ea2333ecc
14:11mntirc: i also get these errors in chromium
14:12mntirc: is that expected, i.e. is NV106 now unsupported, or could this be arm platform specific?
14:15DodoGTA: mntirc: The current nouveau compiler is just not good (that's why NAK is being worked on)
14:16mntirc: DodoGTA: oh :( is there a workaround? can i switch back to TGSI or something?
14:26mntirc: ah, it works with mesa 22.3.6
14:26mntirc: i think the compiler i used to build mesa on this platform might be broken
14:33mntirc: hmm, does nv106 even require firmware?
14:35mntirc: looks like no (???)
14:36DodoGTA: mntirc: This might be the problematic commit: https://gitlab.freedesktop.org/mesa/mesa/-/commit/47251d28527371aeb82f9db7592cff5e78bfcd55
14:39mntirc: DodoGTA: thanks!
14:41mntirc: did kepler gpus not have required firmware?
14:43karolherbst: I really have to look into this one component thing because that got to be a nir regression or something...
14:43karolherbst: we scalarize everything, so I'm surprised we even hit this
14:44mntirc: other than that, i'm surprised how well this gpu+nouveau works with an arm cpu
19:42fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> First DXVK checkbox has been ticked
21:57fdobridge: <gfxstrand> I ran it. CTS is mostly good. I just haven't had a chance to look at the code. I've been in meetings literally all day.
21:57fdobridge: <gfxstrand> @airlied On the upside, @sima and I may have solved synchronization.