04:18fdobridge: <gfxstrand> We probably want to lower it, but I'm not 100% sure. The uniform cache is crazy-fast. Then again, the looping means lots of stalling unless we're REALLY careful in back-end dependency handling. IDK which will win.
04:18fdobridge: <gfxstrand> We can probably also lower some stuff to bound UBOs. We'll need to do that for pre-Turing anyway if we want to get perf.
05:49HdkR: The uniform cache is as fast as registers if abused correctly :)
05:53fdobridge: <airlied> @gfxstrand what the plan for d32s8 to use the meta stuff to do a copy for the stencil plane?
06:01fdobridge: <airlied> oh I suppose that won't work without stencil exports
14:55fdobridge: <gfxstrand> There are a variety of possible plans.
14:56fdobridge: <gfxstrand> One is to spin up 3D and do a meta copy for stencil. Annoying but possibly workable.
14:56fdobridge: <gfxstrand> Another is to keep either some per-device memory reserved or a shadow copy inside the image itself which is 16bpp and do a two-stage copy.
14:56fdobridge: <gfxstrand> Given that that only increases the image size by 25%, it's probably tractable.
14:57fdobridge: <gfxstrand> And lets us keep using the copy queue for copies.
15:37fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> 🐸
15:37fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> https://cdn.discordapp.com/attachments/1034184951790305330/1097546459840258058/Screenshot_20230417_183647.png
15:38fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> DXVK v2.1 on NVK 🚎
15:38fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> https://cdn.discordapp.com/attachments/1034184951790305330/1097546586285940806/Screenshot_20230417_163828.png
16:24fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> NFS MW2012 with DXVK v2.1 🍩
16:24fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> https://cdn.discordapp.com/attachments/1034184951790305330/1097558099763663019/Screenshot_20230417_191922.png
16:25fdobridge: <karolherbst🐧🦀> I hope NFSU2 works :3
16:28fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> I tested NFSU1 and it works, so U2 should work too
16:29fdobridge: <karolherbst🐧🦀> huh? U2 is quite different tho
16:34fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> https://cdn.discordapp.com/attachments/1034184951790305330/1097560725775462410/Screenshot_20230417_193352.png
16:40fdobridge: <karolherbst🐧🦀> best NFS game :3
16:43fdobridge: <karolherbst🐧🦀> though some disagree and say MW1 was better
16:43fdobridge: <karolherbst🐧🦀> I disagree
18:46fdobridge: <airlied> @gfxstrand can a temporary work in the presence of multiple queues?
19:21fdobridge: <gfxstrand> If it has enough space for the whole image, yes. Data races are the client's fault.
19:24fdobridge: <airlied> Wonder what NVIDIA does
19:27fdobridge: <gfxstrand> Not sure.
19:27fdobridge: <gfxstrand> I think they copy through a temporary. I think I've heard Piers mention that.