Radeon HD5xxx filtering issue

[TLDR]

Guys, take a look at this article.

http://www.computerbase.de/artikel/...5/#abschnitt_anisotrope_filterung_auf_hd_6800

In German but the screenshots speak for themselves (there are videos too but they require registration). They've fixed banding but introduced more moire and shimmering. NVidia's HQ mode still remains untouched. What's even worse - in 10.10 build they've made default quality AF on HD 5800 inferior to previous driver revisions.

 

[BFG10K]

I had to look up SGSSAA.

I thought antialiasing by rendering everything at a higher res then downsampling was dead. That's an outdated 10+ year old technique!!

That isn't how RGSS/SGSS works.

 

[BFG10K]

In German but the screenshots speak for themselves (there are videos too but they require registration). They've fixed banding but introduced more moire and shimmering.

Yep, the texture aliasing is as plain as day in those screenshots. And if it’s that visible in stills without zooming, it’ll be even more visible during in-game movement.

 

[KARpott]

The article also states that the banding is not entirely gone, and they mention a member of 3DCenter as source, well guess who that is

But indeed, those shots show the problem. Without SGSSAA, there's a lot of aliasing, with SGSSAA, there's still a little banding visible.

 

[Throckmorton]

That isn't how RGSS/SGSS works.

I found an explanation here http://www.pcgameshardware.com/aid,...irst-DirectX-11-graphics-card/Reviews/?page=6

It sounds like that's what it is.

 

[BFG10K]

I found an explanation here http://www.pcgameshardware.com/aid,...irst-DirectX-11-graphics-card/Reviews/?page=6

It sounds like that's what it is.

Nowhere does it say that ATi’s method is based on oversampling because if it did, it would be wrong. RGSS/SGSS is not done by oversampling; it’s done by rendering the image multiple times with either a sample or geometry offset.

And you do know that MSAA renders the image bigger, right?

 

[Leadbox]

Guys, take a look at this article.

http://www.computerbase.de/artikel/...5/#abschnitt_anisotrope_filterung_auf_hd_6800

In German but the screenshots speak for themselves (there are videos too but they require registration). They've fixed banding but introduced more moire and shimmering. NVidia's HQ mode still remains untouched. What's even worse - in 10.10 build they've made default quality AF on HD 5800 inferior to previous driver revisions.

Halflife 2, Oblivion and Trackmania, is there any reason why they don't use more recent titles for these comparisons?

 

[Throckmorton]

Nowhere does it say that ATi’s method is based on oversampling because if it did, it would be wrong. RGSS/SGSS is not done by oversampling; it’s done by rendering the image multiple times with either a sample or geometry offset.

And you do know that MSAA renders the image bigger, right?

Do you have a link to a good explanation that I can read? That article doesn't really explain it. Is each polygon with its textures rendered multiple times and the results "blurred" together over time? I saw a post somewhere about how it makes games feel more cinematic because you don't get the texture shimmering etc. Sounds like a pretty big advancement actually.

According to that article, MSAA = edge antialiasing, but I had figured MSAA was rendering the whole image at higher res then downsampling.

 

[Scali]

According to that article, MSAA = edge antialiasing, but I had figured MSAA was rendering the whole image at higher res then downsampling.

Well, MSAA is a bit of both.
It renders the z/stencil buffer at a higher resolution. Then it uses this supersampled information to determine whether the pixels are on a polygon edge or not.
If all subpixels for a given pixel are inside the polygon, then the pixelshader is run only once, and the colour is copied to all subpixels. So technically that pixel will not be supersampled, but there is really no need to, as all subpixels belong to the same polygon, and as such a single sample should suffice (this assumption is obviously based on the assumption that there will not be aliasing inside the polygon, which is not true with more sophisticated pixel shading such as bumpmapping).

If one or more subpixels fall outside the polygon, then all subpixels are shaded, so you are doing 'real' supersampling.

After the image is rendered, you have a complete supersampled colourbuffer, which is then downsampled (resolved).
So in a way it is edge antialiasing.

 

[KARpott]

Halflife 2, Oblivion and Trackmania, is there any reason why they don't use more recent titles for these comparisons?

There are several reasons:
- it depends on the engine, not on the game
- those titles DO NOT HAVE ANY BUGS whatever that reduce image quality. So this argument is simply not valid here. AF is not something that changes from game to game, so the only thing you need is the right content and a shader based rendering model, which is given in each case => no need for newer games
- Newer games might get special treatments concerning AF optimizations, we can be sure that those games reflect most of the games out there in terms of IQ
- HL2 equals source 2007 (it has been ported to orange box engine in case you don't know) which is about the same as source 2009 which will be used in future games
- HL2 (engine) has a considerable user base, results can be verified
- TM is an "old" game, but it's perfect for testing AF, as the developers used really bad content like those metal ground textures => you get an extreme case. And testing AF usually only makes sense when looking at those cases, as in the average case, IQ is about the same on most modern graphics cards from AMD or NV
- TM is for free, everybody can reproduce the results!

On the whole, please forget about this "old games do not represent reality" argument. In general, it's not valid. Granted, old games might have problems with certain driver revisions, but as long as this can be ruled out, they are as good as, sometimes even better than newer games for testing IQ.

 

[BFG10K]

Do you have a link to a good explanation that I can read?

The link in this OP takes you to my latest IQ piece: http://forums.anandtech.com/showthread.php?p=30648648

When you hit the article, there are links to past articles in the introduction that should answer your questions.

Is each polygon with its textures rendered multiple times and the results "blurred" together over time?

The entire scene is rendered multiple times, each with a sample offset. Then the samples are blended for the final pixel resolve. This means it can AA everything.

I saw a post somewhere about how it makes games feel more cinematic because you don't get the texture shimmering etc. Sounds like a pretty big advancement actually.

Yes. The technique is actually ancient (being around on consumer parts since 2000/2001) but it has always come with a steep performance hit. SSAA is the only current universal fix for shader and texture aliasing. A scene with proper SSAA applied to it looks like a pre-rendered CGI movie because not a single pixel is out of place.

According to that article, MSAA = edge antialiasing, but I had figured MSAA was rendering the whole image at higher res then downsampling.

Yes, (pure) MSAA only affects polygon edges, and anti-aliasing is done based on depth coverage.

 

[Xarick]

Man.. this thread is the worst thing that could ever have happened to me. I can't game anymore. I can see this effect on tons of games now. I just tried playing call of duty modern warfare 2.. I can see it there.. most notably on shadows that just seem to all the sudden smooth out as you get closer to them. It is really distracting. I can also occasionally see the two weird lines in front of me. I never noticed this stuff before.
In other games like GTA4 I can also see where it really forks up the shadows. Basically if you run along a shadow like of a building you will see where it jaggy until this filter line passes over it.. then it smooths out.
I am stuck now as I cannot afford to take a loss and buy another card right now.. but basically I find myself unable to play pc games now.

 

[Throckmorton]

Man.. this thread is the worst thing that could ever have happened to me. I can't game anymore. I can see this effect on tons of games now. I just tried playing call of duty modern warfare 2.. I can see it there.. most notably on shadows that just seem to all the sudden smooth out as you get closer to them. It is really distracting. I can also occasionally see the two weird lines in front of me. I never noticed this stuff before.
In other games like GTA4 I can also see where it really forks up the shadows. Basically if you run along a shadow like of a building you will see where it jaggy until this filter line passes over it.. then it smooths out.
I am stuck now as I cannot afford to take a loss and buy another card right now.. but basically I find myself unable to play pc games now.

Thanks for posting this. I've been considering buying an AMD card assuming this wouldn't be noticeable, but if you can notice and be bothered, I'm sure I would too since I'm somewhat OCD! I'll have to go with nVidia

 

[Xarick]

honestly throck.. if nobody ever pointed it out I would never have noticed. Also I am not 100% the shadow issue has anything to do with it. This is the first card I have used AF on so maybe the shadow thing happens on both cards. It just seems so abrupt to me. Jaggie shadows all the sudden become smooth. It is annoying because it happens as an imaginary line passes by..

 

[KARpott]

I'm not the soft shadow expert, but if TMUs come into play, it could be related to the same issue. I just know the regular real time soft-shadowing technique, and it only has to use TMUs for reading screen buffers as far as I know. That shouldn't be affected by such problems. On the other hand, there can be other, maybe texture based solutions that are affected. I have no idea, but the shadows in TrackMania have the same problem, for example.

 

[Scali]

honestly throck.. if nobody ever pointed it out I would never have noticed. Also I am not 100% the shadow issue has anything to do with it. This is the first card I have used AF on so maybe the shadow thing happens on both cards. It just seems so abrupt to me. Jaggie shadows all the sudden become smooth. It is annoying because it happens as an imaginary line passes by..

That shadow thing sounds more like a side-effect of projected shadow maps to me.
As the projection of the shadowmap on the screen changes (because of movement of the camera and/or lightsource), the mapping between the shadowmap and the pixels of the screen change, which can cause magnification in some places, resulting in jaggies.
Another common effect is 'shadow acne', when pixels will randomly fall in front or behind a polygon because the z-values are very close to eachother, but because of the different projections you get inaccuracies.
For some reason my 5770 had shadow acne issues in code that I had written many years ago, and worked fine on a Radeon 9600 and all other cards I've had since.

 

[Xarick]

Here is the best way I can describe it. You are standing next to a building that has a straight line shadow cast parallel to your direction. It looks very jaggie in a staircase pattern up until it gets close at which point it simple is smooth. as you walk the smooth line follows forward. This doesn't happen on all shadows, but it just looks bad when it does.
Of course.. for most people they are playing the game and not walking parallel along shadows to see the effect. I first noticed in in GTA4 where shadows would do this but they would come from no shadow to being.. others would do the jaggie thing.

 

[Scali]

Well, that's just a side-effect of shadowmapping in general. The shadowmap is projected from the lightsource into screen space, which causes distortion. Shadowmap resolutions may jump up and down, and projections may switch from one method to another, depending on the distortion. In 3DMark05 it was very obvious as well.

 

[Xarick]

Well I just picked up Risen. This is definitely due to ATIs filtering. The effect happens on both Nvidia and ATI cards where the shadows are blurry then come into focus as you get close. Here is the problem. On my 5850 you can see it happen in a line.. this line goes out directly infront of you making all the shadows clear. It is distracting. I can't believe no reviews of this card have ever meantioned it. On Nvidia cards the blurry to clear happens as well. It just doesn't do it in a harsh like. They just sorta clear up.
I assume this is because Nvidia blurs the transition line and ATI fails to.
I took a video of risen.. gonna upload it here in a bit so you guys can evaluate it.

Here you go

http://www.mediafire.com/?ixtl7lasg6sfg4b

Let me know whats going on. Cause this is making it unfun to play for me.

 

[Scali]

Well I just picked up Risen. This is definitely due to ATIs filtering. The effect happens on both Nvidia and ATI cards where the shadows are blurry then come into focus as you get close. Here is the problem. On my 5850 you can see it happen in a line.. this line goes out directly infront of you making all the shadows clear. It is distracting. I can't believe no reviews of this card have ever meantioned it. On Nvidia cards the blurry to clear happens as well. It just doesn't do it in a harsh like. They just sorta clear up.
I assume this is because Nvidia blurs the transition line and ATI fails to.

Could you make some kind of screen shots or perhaps even a small video with Fraps or such, to demonstrate what you mean exactly?

 

[Xarick]

I uploaded on in the last post scali. Take a look a the shadows of the trees on the ground. If it is something else.. or I am wrong.. let me know.. cause I hate this effect.

 

[Scali]

Ah, thanks for the video, Xarick. Now I see what you mean.
I think these lines are caused by the mipmapping done inside the shadowmaps. It appears that the mipmap transitions are not being filtered properly, so you clearly see a band between the larger, more detailed mipmap, and the smaller mipmap further in the distance. There could indeed be a difference between how AMD and nVidia handle the filtering of mipmaps in hardware (or what 'optimizations' their drivers may use).

 

[Xarick]

Okay.. but I am lost how any reviewer would miss this. I understanding missing the filtering on repeating patterns as I never would have noticed that if it hadn't been pointed out.. but the shadows?
EVERY game I play the shadows do this. It doesn't make any sense. Is this the same filtering issue? How could anyone not see this.

 

[Scali]

Well, I have one theory:
In the early days of shadowmapping, all games looked like this, on all hardware.
Perhaps reviewers got so used to it that they took it for granted, and didn't even notice that some hardware now does a better job at filtering.