| 19 Jul 2025 |
|  Seong Yong-ju joined the room. | 04:56:14 |
 K900 | In reply to @colemickens:matrix.org
I feel so lost trying to keep track of FSR3/3.5/4, the new lossless scaling stuff, what exists in Proton/Gamescope/drivers. Ugh. Drivers are basically out of the picture for this | 05:46:44 |
| K900 | On Linux at least, on Windows there's some driver side upscaling injection nonsense that could just be a Vulkan layer but D3D doesn't have those so lol | 05:47:52 |
 woobilicious | You definitely need hugepages, that gives a 30% boost in performance on native factorio with a custom malloc lol, and find the correct CPU topo hints (try Epyc ones?), passthrough didn't work for me either. look at mem bloon driver too. | 06:51:55 |
| woobilicious | I haven't used this config in like 3 years, but I got good performance with this on my 5600X: https://gist.github.com/YellowOnion/132f85ca3f81a7594fd9bcd1e424cc18 | 06:56:16 |
 olivia | this is for the case where an application isn't built with FSR support? | 07:17:01 |
| K900 | Also yes | 07:17:12 |
| K900 | The AMD driver can inject FSR2 into basically anything | 07:17:22 |
| K900 | And they also have "AMD Fluid Motion Frames" which is kinda FSR3 frame gen retrofitted to be purely spatial | 07:17:42 |
| K900 | That can be injected into uh | 07:17:54 |
| olivia | my impression reading through the vkd3d-proton stuff earlier was that there was a bunch of driver interaction that existed for effectively no reason | 07:17:58 |
| K900 | Some things, but not all of them? | 07:17:58 |
| K900 | I don't know why | 07:18:02 |
| K900 | Oh do you mean for FSR4? | 07:18:13 |
| olivia | is fsr4 different? | 07:18:21 |
| K900 | Yes | 07:18:25 |
| olivia | structurally | 07:18:26 |
| olivia | fun | 07:18:28 |
| K900 | I think what's going on is they basically want to move the actual scaling algorithms into the driver | 07:18:45 |
| olivia | this stuff is so fucked | 07:18:46 |
| K900 | So that the game doesn't have to constantly update to get better upscaling | 07:18:59 |
| K900 | Instead it just asks the driver to do the thing, and the driver does the thing however it considers optimal | 07:19:14 |
| olivia | hmm | 07:19:25 |
| olivia | as an application dev I would be terrified of weird artifacts there | 07:19:33 |
| olivia | with future versions that I haven't tested with | 07:19:40 |
| K900 | There's also the whole "DXIL has no vendor extensions" thing where they have to basically smuggle FP8 coopmat into the shader compiler | 07:20:05 |
| K900 | But that's just DXIL being DXIL | 07:20:10 |
| olivia | I guess this bit is also surprising because I'd expect that you don't want to do the upscaling stuff on UI? | 07:20:21 |
| K900 | And they'll probably switch to native instructions for this once those actually exist (...right?) | 07:20:34 |
| K900 | Yes | 07:20:39 |
