RADV Zink Sees Another "Massive" Optimization On The Way

[phoronix]

Phoronix: RADV Zink Sees Another "Massive" Optimization On The Way

In going through my recent RADV-Zink vs. RadeonSI OpenGL benchmarks, Valve's Mike Blumenkrantz has already been landing optimizations/fixes and there is another one on the way as a result...

https://www.phoronix.com/news/RADV-Zink-Epilog-Optimization

[aufkrawall]

I'm also seeing a bit higher fps with amdvlk in SotTR when being CPU bound. At least that was the case a few weeks ago.

[CochainComplex]

Since I'm a total layman concerning gfx drivers and gfx api...what is profiting. Only games like ET ..Only Zink or the whole RADV even when vkd3d-proton is used?

[tildearrow]

epilog or eiplog?

[illwieckz]

Since I'm a total layman concerning gfx drivers and gfx api...what is profiting. Only games like ET ..Only Zink or the whole RADV even when vkd3d-proton is used?

This is not optimizing a game, this is optimizing a driver feature, this would benefit every game using that features.

It just happens that sometime a given game do things in a way it's easier to spot the need for optimization. The most obvious a game makes it to spot the problem, the easier it is do track down the problem and easier it is to check and test how much fixes improves the situation (or not).

Sometime the testbed for a feature is a corner case, but if it makes things easy to work with, it's better to focus on the corner case.

[kiffmet]

CochainComplex AFAIU, the change itself was made in RADV, but it only benefits Zink, since that's the sole thing relying on that functionality atm. Out of all applications running with Zink, it's likely that only the CPU-bound ones will see a significant uplift.

[spheenik]

Make sure to benchmark Zink again soon, Michael. Seems to really rustle Mr. Blumenkrantz' jimmies when he sees it. Two more iterations, and it'll be on par!

[smitty3268]

AFAIU, the change itself was made in RADV, but it only benefits Zink, since that's the sole thing relying on that functionality atm.

I don't think that's the case, prolog/epilog stuff is used for various stuff in the driver and it's not anything particular to zink.

That said, maybe zink hits it harder than a normal vulkan application does. Not sure.

[mbriar]

I don't think that's the case, prolog/epilog stuff is used for various stuff in the driver and it's not anything particular to zink.

That said, maybe zink hits it harder than a normal vulkan application does. Not sure.

I think applications using the usual monolithic pipelines won't get any epi/prologs at all, so it will definitely have no effect on any shipping vulkan game or anything using vkd3d-proton.

[kiffmet]

I don't think that's the case, prolog/epilog stuff is used for various stuff in the driver and it's not anything particular to zink.
That said, maybe zink hits it harder than a normal vulkan application does. Not sure.

It's exactly that. Certain games running through Zink on RADV hit a corner case in the RADV driver. Other stuff won't benefit at all (I assume that's because usually this code usually doesn't get called a plethora of times during a single frame), which Mike (the author of these patches and main Zink developer) made clear

Taking out a magnifying glass, this resolves into some more definite problem areas:

• PS epilog lookup (AKA dynamic blend)
• some other stuff

I banged out some epilog fixes, and then I did the same for VS prologs, and it’s great, and it’ll be up in MRs soon™. Huge perf gains. Absolutely massive. But only for zink, because nothing else uses any of this.

twice

(...)
How Much Faster Is It?
If I had to give a number, probably at least 1000% faster. Just as a rough estimate.

That Seems Like A Lot
It sure does, but only zink uses any of it, so who cares.
(...)