Valve Engineer Fixes Massive Performance Issue For RADV Driver With AMD FSR2

[phoronix]

Phoronix: Valve Engineer Fixes Massive Performance Issue For RADV Driver With AMD FSR2

With the upcoming Mesa 24.3 release there is a huge improvement coming for those using the RADV Radeon Vulkan driver with games leveraging AMD FidelityFX Super Resolution 2...

https://www.phoronix.com/news/AMD-FSR2-Mesa-24.3-Fix

[dimko]

cant help but wonder if FSR3 will benefit in any way, i doubt it, but i can dream, right?

[skeevy420]

Will this help in games that have the option to use FSR2 for AA without upscaling, when setting it to Native instead of Quality or Performance?

[varikonniemi]

it's amazing AMD has a driver team that is writing their proprietary driver, that does not communicate with the team that writes the open driver, which has to reverse engineer and find out all the things themselves.

Seems like it can only stem from something like paranoid separation of proprietary information between closed and open drivers, extending so far that not even basic hints at what is the problem don't transfer over.

[blackshard]

it's amazing AMD has a driver team that is writing their proprietary driver, that does not communicate with the team that writes the open driver, which has to reverse engineer and find out all the things themselves.

Seems like it can only stem from something like paranoid separation of proprietary information between closed and open drivers, extending so far that not even basic hints at what is the problem don't transfer over.

It's amazing that this just plain wrong. We're talking of a buggy behaviour in a niche case (RDNA2 + FSR2 on RADV)

[ahrs]

it's amazing AMD has a driver team that is writing their proprietary driver, that does not communicate with the team that writes the open driver, which has to reverse engineer and find out all the things themselves.

Seems like it can only stem from something like paranoid separation of proprietary information between closed and open drivers, extending so far that not even basic hints at what is the problem don't transfer over.

Maybe they just forgot, or didn't think it worthwhile to mention. We're potentially talking about two completely different teams of developers, miscommunication or no communication at all may happen at times, even so they are still leaps and bounds better than Nvidia, there you can be sure there will be little to no communication between them and Nouveau (except the bare minimum to make sure the graphical session can run so you can install "the real driver"). Nvidia's got their driver and Mesa has theirs.

[Danny3]

Fuck you AMD for refusing to contribute directly to RADV and duplicating work!
After you already refuse to give us a control panel like for Windows (of course with all the drivers improvements needed for that) because you don't have the time / resources for that!
But FFS you have time for all this duplicated work!!!

Congratulations and many thanks to Valve!
No wonder why I still buy games from you even if I don't like DRM and centralized stores that can take my access away any time they want from what I have bought!

[mbriar]

This change has no effect on the FSR2 upscale pass itself, that part already worked fast without problems before. It's just happened to be that the FSR2 demo scene exposed this particular optimization opportunity, but it ran slow with or without even enabling the FSR upscale itself.

[varikonniemi]

Maybe they just forgot, or didn't think it worthwhile to mention. We're potentially talking about two completely different teams of developers, miscommunication or no communication at all may happen at times, even so they are still leaps and bounds better than Nvidia, there you can be sure there will be little to no communication between them and Nouveau (except the bare minimum to make sure the graphical session can run so you can install "the real driver"). Nvidia's got their driver and Mesa has theirs.

of course, if your continued employment hinges on the open driver team not reaching your performance, then you don't find it worthwhile to mention...

[varikonniemi]

It's amazing that this just plain wrong. We're talking of a buggy behaviour in a niche case (RDNA2 + FSR2 on RADV)

no we are not. We are talking about a not-so-obvious optimization for a specific hardware design.