Dxvk: Performance tanks in Mass Effect when looking at occluded part of the level

[mirh]

So, I like to benchmark the game in the comm room on the Normandy, since I found out (once the door closes) that I can get the highest framerates of all the game there.
I just walk against the FTL comm link at the bottom of it, and then stare at the empty wall.
In D3D9 I get around 420 fps, while dxvk gives me around 350, not that bad I thought.

I just figured out though, that if I just rotate the camera to "face the rest of the ship", while D3D9 only slows down to about 360fps, dxvk gets almost 3 times slower at 130.

Software information

Mass Effect 1, max in-game settings, no mods.

System information

• GPU: Radeon R9 290
• Driver: 20.11.2
• Wine version: Windows 7 SP1
• DXVK version: 1.7.3

Apitrace file(s)

• Apitrace

Log files

• d3d9.log

[doitsujin]

Are you CPU-bound or GPU-bound? That kind of info is extremely important when dealing with performance issues.

Also, be aware that DXVK is known not to perform well on first- and second-gen GCN cards due to some limitations around how they handle depth buffers.

[mirh]

I think it's actually both in the apitrace you can find.
I lowered the resolution to 640x480, and D3D9 wasn't actually affected at all (even though I could have swore gpu usage at 97% meant it was gpu-bound).

But then this had a big effect on DXVK. When I'm staring at the wall I gain those 70fps that I was missing earlier.
While I'm staring behind my shoulders, I'm still getting the same ~130 fps.

[K0bin]

I can confirm that it's slower when looking at the wall that has geometry behind it and the hit is a bit larger than with Windows D3D9. Windows D3D9 goes from 695 to 560 fps and DXVK drops from 595 fps to 450 for me.

Fully CPU bound of course. I dont really know why though. There's nothing that obviously sticks out.

[K0bin]

Can you try this build?

d3d9.zip

[mirh]

This is enormously better. Performance hit is around 20% now (280fps), which is more or less the same proportional ballpark that you get with native D3D9.

[K0bin]

How about this one:
d3d9.zip

Does it yield similar results to yesterdays build?

[mirh]

Yes.

[unic0rn]

A8-7600, mesa 20.2.3, comm room scenario:
dxvk-master: 60fps (vsync - 88% gpu utilization) vs slightly above 48fps
pull request #1857 : 60fps (vsync - 88% gpu utilization) vs slightly below 48fps

The difference is perhaps 0,5fps or even less, but it does seem to perform worse. Not sure how random it is, rotating the camera seems to change that number a lot, but when looking straight at the door dxvk-master reports 31 render passes, #1857 32.

So no performance gain on radv.

[K0bin]

A performance drop is expected because the game is simply doing a lot more when you look at that other wall.

It was just a special case for mirh where the cpu didnt submit work early enough so it would work for the gpu.

[unic0rn]

I did expect a performance drop, I was just thinking it's some generic performance improvement for that case, and the way you explain it, I guess it's related to way higher framerates he's having. Nevermind, don't mind me, I was just reporting 0,5fps drop in my case, which is acceptable.

[K0bin]

d3d9.zip

How about this one? Still improved? That should hopefully not cause any performance regressions on other systems.

[mirh]

Yup. I also noticed a very marked 40% improvement in my "worst case spot" on the citadel.

[K0bin]

Fixed.

[mirh]

Cool, now, how about tackling performance being "in general" slower?
On my desktop the difference is about 15-20% as I mentioned in the opening, but on my 950M I can get north of 50% faster framerate with native d3d9 (in fact, with a slightly different "not really much to render" camera angle from the same scenario above I can even near twice as much fps)

[K0bin]

There's nothing obviously wrong with DXVK that we could just optimize.
Is performance an issue anyway? As mentioned, on my system it runs at 500 fps with DXVK.
Sure it would be nice to reach the 600 fps that Nvidias D3D9 manages but no one's gonna notice the difference.

[mirh]

I'm not talking about cpu limited situations here though (and in fact these last numbers were taken from W10, which would make even an nvidia suck there).
And yes, it's kind of a big deal when it can make the difference between "playable" and "unplayable" (I have a dedicated gpu in my laptop thankfully, so here if any we are talking about "bearable" vs "great", but still)

[K0bin]

What kinds of numbers are you getting anyway where this is a problem and what settings and resolution?

Mass Effect is pretty much ALWAYS CPU limited on my system.

[mirh]

Increase the resolution then?
I'm using a 6500U+950M laptop now personally, so 1080p is already enough to run around the 60fps mark.

In the most gpu-limited scene I can find score is 52 fps vs 35.
In the most stupid one which is the one in the apitrace (which anyway still isn't cpu-limited without having a big ass gpu) 85 vs 51. Which become 120 vs 60 when I stare to the ceiling.

That doesn't sound really trifles.