| 17 Jul 2025 |
 woobilicious | "Georg Lehmann" seems to be responsible for most of the recent work. | 00:48:16 |
 olivia | yeah I'm confused because radv at least is only advertising VKEXTshader_float8 for gfx12+. pretty sure there's no GL fp8 extension | 01:18:36 |
| olivia | yeah I'm confused because radv at least is only advertising VK_EXT_shader_float8 for gfx12+. pretty sure there's no GL fp8 extension | 01:19:38 |
| olivia | oh, VK_KHR_cooperative_matrix has 8-bit int/uints. (!36117)[https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36117] is presumably about that | 01:26:11 |
| olivia | oh, VK_KHR_cooperative_matrix has 8-bit int/uints. !36117 is presumably about that | 01:26:22 |
|  nyanbinary 🏳️⚧️ set a profile picture. | 01:27:34 |
| olivia | and cmats are only enabled on gfx11+ | 01:27:47 |
| woobilicious | the && !pdev->use_llvm portion hints that if you can use llvm, you can get float8 support irrespective of the hardware? | 02:50:34 |
|  mall0c joined the room. | 02:52:28 |
| woobilicious | oh wait I'm getting my logic wrong...But that is more confusing, I'd think llvm wouldn't disable a feature. | 02:52:43 |
| olivia | probably use_llvm is about using the llvm shader compiler middle/backend instead of the nir one | 02:57:20 |
| olivia | I suspect nobody has implemented fp8 or cmats in the llvm backend yet | 02:57:51 |
| woobilicious | Yeah I'd just think it would have feature parity, but I guess not. | 02:57:58 |
| olivia | I wouldn't expect it to, nobody likes it lol | 02:58:25 |
| olivia | I believe the intention is to remove it eventually, but I'm not very aware of what's going on with amd stuff so take that with a grain of salt | 02:58:51 |
| olivia | https://docs.mesa3d.org/drivers/radv.html#shader-compilation yeah
We still maintain an LLVM based compiler backend too, which is these days solely used for testing and hardware bringup. Users are recommended NOT to use the LLVM backend.
| 02:59:54 |
| woobilicious | Oh really? how is that related to the ACO compiler? doesn't that use llvm? | 03:00:08 |
| olivia | no ACO is NIR only | 03:00:44 |
| woobilicious | ahh my memory is wrong then. | 03:01:31 |
| olivia | (I believe the llvm backend in radeonsi predates NIR) | 03:01:33 |
| woobilicious | olivia: I think the FSR4 code just converts to fp16 manually instead of them extending the compiler to do the conversion. I guess the code is writting to use cooperative_matrix, so yeah only RDNA3...too bad I wanted slow as dogshit FSR4 haha. | 03:33:25 |
| olivia | yeah, that seems plausible (I don't know anything about the specifics of FSR4, but almost certainly it will not work without cmats) | 04:10:31 |
| olivia | oh, the fp8->fp16 lowering is happening in vkd3d-proton | 04:28:10 |
| olivia | https://themaister.net/blog/2025/05/09/conquering-fidelityfx-fsr4-enabling-the-pretty-pixels-on-linux-through-maniacal-persistence/ | 04:28:12 |
| olivia | https://github.com/HansKristian-Work/vkd3d-proton/pull/2427 | 04:28:22 |
| olivia | kinda cursed | 04:28:23 |
 K900 | In reply to @woobilicious:matrix.org
K900: your post mentions Intel-RT on non-x86, does that mean you can run Arc graphics on Arm? Always could | 04:54:28 |
| K900 | They're just PCIe | 04:54:28 |
| K900 | The RT code had some handrolled x86_64 SIMD but they got rid of that | 04:54:41 |
|  ribosomerocker joined the room. | 23:58:21 |
