Mesa RADV Driver Baking Improvements For VKD3D-Proton With AMD FSR3

[phoronix]

Phoronix: Mesa RADV Driver Baking Improvements For VKD3D-Proton With AMD FSR3

Landing in Mesa 24.0-devel this week alongside other exciting changes is some pending work for enhancing VKD3D-Proton and AMD FidelityFX Super Resolution 3 (FSR 3) support for the RADV Vulkan driver...

https://www.phoronix.com/news/Mesa-RADV-Better-FSR3

[rhysperry111]

This Rhys Perry guy sounds awesome

[ms178]

FSR2 also needs some more work in RADV: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7260

[Wielkie G]

https://gitlab.freedesktop.org/mesa/mesa/-/issues/7260#note_1780329

It also doesn't seem to have anything to do with the FSR upscaling itself, the performance discrepancy is just as large when setting it to render in native resolution.

Have you even read the bug report?

[ms178]

https://gitlab.freedesktop.org/mesa/...0#note_1780329



Have you even read the bug report?

Please read my post and the linked issue more carefully before attacking other people. The issue hasn't been root caused properly yet. It occured when using FSR2, I haven't stated anything that contradicts with the findings of the bug reporter, i was merely pointing out that RADV has some issues which showed up when using FSR2. This doesn't imply any form of causation.

[skeevy420]

This Rhys Perry guy sounds awesome

He's a pretty rad feller setting a high mantle due to his fire commits coming at mach speed making him all the rage over at AMD.

[middy]

does anyone know if mesa / vkd3d / dxvk / the likes support the dedicated ai hardware on rdna3 cards? iirc my 7900 xtx has 192 of them and i remember reading awhile ago that fsr3 was post to take advantage of the ai accelerators on rdna3.

[Nocifer]

"Apparently this DXIL instruction is used by FSR 3. The unoptimized sequence is horrible, but I don't know how much of an effect this has on actual FSR 3 shaders."

This strikes me as a bit weird. Given that AMD's hardware is powering the Steam Deck and that both companies have a vested interest in AMD's GPUs increasing their performance on Linux, I'd have expected that Valve developers have direct access to AMD's driver schematics and/or direct communication channels with AMD's driver devs, so this shouldn't be a matter of "apparently" and "I don't know how much", it should be a yes-or-no matter as defined by the AMD devs who designed the driver.

What am I missing here?

[ms178]

What am I missing here?

The cooperation between AMD and Valve on RADV from the outside looks not to be as close as I would have assumed years ago. While certain ISA documents are open sourced, this is not neccessarily the case or on time for upcoming hardware. They rely on peaking at what LLVM/radeonsi does for help to bring up such new hardware or use it as a source for inspiration how certain features are implemented. I would have hoped for a closer relationship getting more help from AMD or having contacts to hw and sw engineers to share knowledge. On the other hand, I don't know what other channels they use to communicate such things not meant for our public eyes.

[smitty3268]

This strikes me as a bit weird. Given that AMD's hardware is powering the Steam Deck and that both companies have a vested interest in AMD's GPUs increasing their performance on Linux, I'd have expected that Valve developers have direct access to AMD's driver schematics and/or direct communication channels with AMD's driver devs, so this shouldn't be a matter of "apparently" and "I don't know how much", it should be a yes-or-no matter as defined by the AMD devs who designed the driver.

What am I missing here?

I think they have some private channels to ask AMD questions, but that's about it. In this case, someone just noticed that instruction in the FSR 3 shaders, so it's probably not worth it to try and ask questions of AMD. They know it's there, and it's easier to just optimize their driver to handle it rather than trying to figure out exactly all the details surrounding it in some other driver that may or may not even directly correlate to radv. It probably wasn't optimized before because they hadn't seen anything use it.