RadeonSI Patches Boost Deus Ex: MD Performance By ~70%

[phoronix]

Phoronix: RadeonSI Patches Boost Deus Ex: MD Performance By ~70%

Marek Olšák is off to a good start with performance optimizations of the RadeonSI Gallium3D open-source driver stack in 2017...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-DeusEx-MD-70

[ElectricPrism]

Am I to understand this patch makes performance 170% of what it was before? That's pretty amazing

[LinAGKar]

A few more of these and it might become playable on low graphics on a Fury.

[Veerappan]

Am I to understand this patch makes performance 170% of what it was before? That's pretty amazing

That's what I'm reading here, and how it is generally calculated in my experience. Looks like a nice thing, and this seems to be implemented for SI through the current gen, so it should work for every card that uses the radeonsi driver [my 7850 gets some love ].

The other thing to see is if this helps out in other places than just Deus Ex: MD

[artivision]

This doesn't cover the 3x difference with the pro driver. I advise AMD to develop Bindless.

[GreenByte]

So it should put the RX 480/470 at 1050ti performance. Cool.

[smitty3268]

This doesn't cover the 3x difference with the pro driver. I advise AMD to develop Bindless.

We don't know the circumstances of the testing Marek did here, but this is a pretty major improvement.

From Michael's testing, the perf difference is (in % faster that the PRO driver is):


GPU 1080p Low 1080p Ultra Fury 183% 119% 480 102% 56%

Depending on exactly what GPU and settings that 70% applies to, it appears to take out a pretty major chunk of the difference.

Everyone knows that bindless is going to help a bit more, and it will be added eventually. But Marek already explained to you quite thoroughly that it was going to be much more minor improvement than you think, and explained exactly why. It's not going to be another 70% improvement.

Even in the worst case scenario, (Fury @ 1080Low) the 70% perf boost leaves it only 67% slower than the pro driver. And there's certainly going to be other stuff besides bindless required to fully overcome that (assuming that the 70% applies there, and it's not already larger than 70% on the Fury card)

It could also very easily be that the test was done on the 480, in which case the Mesa driver would have already surpassed the pro driver on at least some settings and bindless would just be gravy on top of that.

[dungeon]

So as Feral minumum requirements for the game is GTX 680 and that has 34.50 fps on low here , with these 70% it seems some cards will go on about 40 fps... which is cool

It could also very easily be that the test was done on the 480, in which case the Mesa driver would have already surpassed the pro driver on at least some settings and bindless would just be gravy on top of that.

Not sure how you calculate that as i just looking at other results and even with these 70% added looks to me it would still be slower about 15 to 30% than PRO PRO has bindless texture extension supported, so it might be that make remaining difference or maybe not

But best if we wait for benhmark, to be sure what exactly this made... and of course if not regressing something else

Funniest thing to me while looking in all these results is that while GTX 680 do 34 fps that is supported, but PRO who does 47 fps Feral does not support that, we probably really live in weird ignorant world

[smitty3268]

Not sure how you calculate that as i just looking at other results and even with these 70% added looks to me it would still be slower about 15 to 30% than PRO

I looked here at the 480 results for the 1080p Ultra test: http://www.phoronix.com/scan.php?pag...st-2016&num=11

17.7 fps (for Mesa) * 170% = 30.1fps

The PRO driver comes in at 27.7 fps.

My main point here was just to poke holes in the "it needs to go 3x faster, and they needed to do bindless before this thing which was mostly useless timewasting" argument it sounded to me like artvision was making.

PRO has bindless texture extension supported, so it might be that make remaining difference or maybe not

But best if we wait for benhmark, to be sure what exactly this made... and of course if not regressing something else

Agreed, this needs to go through testing. It's impossible to say what conditions it was tested in or how it will affect other cards/settings until we see that done.

[Xicronic]

Open source AMD driver progress is exciting, as usual :-)