[long]I've found the GFFX AA IQ bug! ( hopefully )

Arun

Unknown.
Moderator
Legend
EDIT: I did a mistake, and GFFX 4x images *are* 4x - the problem is that they're ordered and that, at the area I looked at, it looked a lot more like 3x. But I can nearly guarantee now that it's 4x.
However, the 2x problem remains.

Hey,

I'm ONLY going to base myself on Anand excellent AA comparaison screenshots. Thanks a lot for taking these, BTW, Anand!

I'm mirroring the images because Anand's server is still very busy as I write this.
http://www.notforidiots.com/Images

Okay, so look at the first image:
http://www.notforidiots.com/images/off.jpg
And compare it with:
http://www.notforidiots.com/images/2x-aa-aggressive.jpg

The first is an aliased result. The second is a 2X Antialiased result on a nVidia card.
Look closely and you'll see there IS a difference: the triangles seem *smaller* with 2x AA. Why?
Well, actually, some parts are also bigger. But I'll explain why later, for now only worry about the smaller parts ( such as the bottom-right of one of the left triangle )
Well, there are more very dark pixels. But... once again, why?

Well, let's look at ATI's 2x AA now, shall we?
http://www.notforidiots.com/images/2x-aa.jpg

Now, as you see, it looks a lot smoother. But let's compre it with nVidia 2X AA ( once again, it's here )

Now, look at the pixels which are not filled in nVidia's 2X AA, and which are filled when there is aliasing ( here for the aliased image again )

Try to memorize ( or use multitasking to see several images at the same time, whatever ) which spots are not filled and were without 2x AA.

Now, look again at the ATI 2x AA image ( here )

Look at the spots that I said you to remember. Guess what? Well, they're filled on the 2x AA screenshot, but they more dark than the remaining of the triangle. And what does that tell us?

That those are the pixels that should have only one of their two subpixels filled with that triangle.

Now, as I said before , some pixels are also filled on nVidia 2x and not when there's no AA. But if you look at the ATI 2x mode, it looks there that those pixels are only filled by the triangle in 1/2 of the subpixels.

So... that means AA IS happening ( there's the performance hit ) , but the two subpixel buffers are not merged. Sometimes that means a part looks bigger, sometimes it looks smaller. But in reality, only ONE subsample if being considered for the final image.
How can I say that? Well, sometimes it's filled when it isn't at 0x AA and sometimes it isn't filled when it isn't filled at 2x AA.
So, it's like if there was ONE sample, but that it wasn't at the center of the pixel: you'd get the exact same result. Sometimes it's filled while it shouldn't be, sometimes it's not filled when it should be ( that is, if you consider there should only be one sample )

-------------------- IGNORE THE FOLLOWING ----------------

Now, that seems like a good explanation. But there's more to it! I can also explain the lower AA quality of every single mode compared to ATI, and show it's a mere bug!

Look at the ATI 4x screenshort ( http://www.notforidiots.com/images/4x-aa.jpg ) and the nVidia 4x screenshot ( http://www.notforidiots.com/images/4x-aa-aggressive.jpg.

Look at the right triangle using a zoom ( such as MS Paint one :) )
Now, if you look REALLY close, you'll see something strange
There's only THREE color variations ( 33%, 66%, 100% ) . On the ATI 4x screenshot, there's FOUR color variations ( 25%, 50%, 75%, 100% )
That would indicate the GFFX 4x AA quality is, well, 3x AA!
So, the pattern is amazingly simple. In both 2x and 4x ( and possibly others, didn't check ) , there's ONE subpixel in every pixel which isn't considered when merging the subpixel buffers

-------------------- STOP IGNORING ----------------

Now, that seems a LOT like a driver bug. It's impossible to say if it's also the case ingame, or if it's just a screenshot problem. But no matter what, that's the type of bug which should ( theorically ) be very easy to fix. And NO conspiracy theories please: if nVidia did this to reduce the performance hit, the 2x AA performance hit wouldn't be so huge!

Conclusion
-------------

-------------------- IGNORE THE FOLLOWING ----------------

The 2x/QC problem is nothing more than a driver bug. It also reduces 4x AA quality. So all AA modes quality is reduced due to this bug. Performance is NOT impacted, because else 2x AA would be free ( so no conspiracy theories here, please! )

So DON'T compare the quality of GFFX/R300 AA modes right now. It's a bug, and one that should be fixed quite easily.

-------------------- STOP IGNORING ----------------

It sounds like the problem is only with 2x AA. I wasn't looking correctly at the 4x AA screenshots, the problem is that it's ordered and that cases where the 4 samples are visible are VERY rare. But they do exist!

Let's hope there's a bug with nVidia's AF algorithm too, because aggressive really looks like crap to me ( but then again, that's really unlikely )


Thanks for reading! :) I hope you enjoyed this.


Uttar
 
The 2x/QC problem is nothing more than a driver bug. It also reduces 4x AA quality.

First of all, if it is a bug, there's no guarantee that it can and will be fixed to any degree of satisfaction.

Second, I find it more likely that if there is some problem, that's it's because there may be problemes with screen-shots actually capturing the displayed image. I'm interested to find out if the screen-shots accurately represent what's displayed on screen.
 
As I said if you read the whole post ( yeah, I know it's long :D ) , well, I don't know if it's only related to screenshots.

As I said, it could only be related to screenshots ( and thus it could be a driver bug when it merges the buffers in software ) , or it could be a hardware bug.
If it's a hardware bug, there could be a workaround via drivers. Or it could be VERY severe and the bug will have to remain.

I'd doubt such a serious bug ( 3x AA when asking for 4x AA but with Z/Color calculated, giving the performance hit but not the IQ ) would have passed hardware Q&A. Sounds more like a mistake by the driver team to me, but who knows...

There's no real answer to all that. But what we DO know, now, is that the same bug exists in both 4x AA & 2x AA, and that NO AA IQ comparaisons should be done unless nVidia says that bug can't be fixed via drivers.

Simply saying 2x AA is defective is not correct: all GFFX mode are defective :(
I'll try to ask nVidia for an explanation, but I doubt they'll answer me ( they're certainly very busy )


Uttar
 
There's no real answer to all that. But what we DO know, now, is that the same bug exists in both 4x AA & 2x AA, and that NO AA IQ comparaisons should be done unless nVidia says that bug can't be fixed via drivers.

First, you have to get nVidia to agree that there is a bug.

The only "known bug" in FSAA that nVidia has told web-sites about is some obvious corruption in (IIRC) 6xs modes. You would think that if there was some known and obvioust FSAA bug that was definitely fixable, nVdia would have briefed the web sites about it.
 
The screen shots certainly only seem to be taking one sample in 2x and QC. I disagree about your assumptions on 4xmode though. All the screen shots of FX 4xAA look like standard 4x ordered grid to me. A central 50% graduation with single pixel 25% and 75% graduations appearing in places.

You've also got 6xS and 8xS to explain ;)

Assuming it is a bug I would think it is fixable as both 2x and QC show up fine in brents NFSHP2 screenshots.
 
Hmmm,

Well I agree with Joe on this one, as hard as it is to swallow you may have to accept that nVidia still haven't managed to crack the FSAA problem effectively. Sure they may have a missing sample but then correcting the problem may indeed occur another peformance penality. Whose to say they've had to comprimise the quality slightly to ensure decent performance? We all know that the FX is bandwidth limited so any FSAA savings how ever small will be utilised.

I don't doubt that nVidia will be able to tweak FSAA through driver devlopment but I'm with Joe when he says if the Nvidia PR machine could have briefed the sites on a bug that would have given them a get out of jail free card for FSAA comparisons they'd certainly have jumped at the chance.

When analysing quality and speed nVidia have clearly leaned towards speed with each card. Take a look at the cooling solution they've had to adopt. Does that seem to point to a company not desperate to take the performance lead at any cost? I can well believe a tweaked driver version would be utilised for the initial round of benches.

I'd even go so far as to suspect they would have been a little shocked at the depth of some of the reviews considering the time constraints. Until recently you would have been shown a few Q3/3d mark scores with a one liner denoting comparable IQ between the R300 and NV30. The scores and FX FSAA modes would have looked very appealing then.

Thankfully the review sites H and Anand have at least started to hammer IQ settings although I think it was unfair for Anand not to mention the mip map boundaries as they can be extremely off putting, especially for a £300 card!

All in all I think it'll be an interesting few months although I still don't think you'll be seeing a ultra FX being sold with that monstrous fan nor a decent FSAA implementation on the FX!

Just my 2 pennies worth

:)
 
Joe: Well, it does seem like there's a bug to me. If nVidia refuses to say it and don't fix it in the future, it could be because it's a hardware bug and can't be fixed ( that would be really bad, and I certainly wouldn't buy a GFFX if there was a so huge bug in it )

Bambers: As I said, I didn't look at 6xS & 8x yet. It's already late here ( 9:30PM ) , so I won't do so today anyway.
As for 4x... Well...
http://www.notforidiots.com/Images/4xA.png ( 50KB )

That's a HUGE zoomup of a part of a Anand screenshot of nVidia 4x AA.
Quality is rather bad even with the huge size & lossless compression ( to make sure it doesn't get worse ) , but it really only looks like there's 3 color variations on that image ( 33%, 66% and 100% ) ... On ATI 4x AA, I can easily see 4 variations. But there I can't.


Uttar

P.S. : I'd be VERY surprised if this was a way for nVidia to increase performance. 2x AA quality is actually *worse* than no AA ( because it's not aligned correctly anymore ) , and the performance hit remains large
Now, maybe it's a bug nVidia can't fix by drivers. But I'd be very surprised if it was voluntary.
 
Bambers said:
The screen shots certainly only seem to be taking one sample in 2x and QC. I disagree about your assumptions on 4xmode though. All the screen shots of FX 4xAA look like standard 4x ordered grid to me.
Agreed.

Isn't it so that, with 2xMSAA, the actual combining of the two multisample buffers happens on the ramdac and not in the framebuffer? I remember never having been able to get a proper 2xMSAA screenshot on my GF4Ti with using the print screen button, as it only grapped one of the two multisample buffers, i.e. "half" the final AA effect.

ta,
-Sascha.rb
 
Joe DeFuria said:
On ATI 4x AA, I can easily see 4 variations. But there I can't.

What makes you think you should see 4 variations given the orderd grid implementation of 4X FSAA on the Geforce?
Excepting horizontal (vertical) and near horizontal (vertical) edges, you should still see four gradiations. However, depending on the angle, the falloff might be very sharp so that you don't get a very good AA effect.
 
I can see the same number of middle shades on each (3)

ati4x.png


fx4x.png
 
Damn, you're right guys :oops:
You people are really good at finding errors. Thanks for finding it, BTW. I think my gamma is a little messed up... I can see the variations you show using PSP and looking at the pixles you show, but it's practically invisible with my eye. Damn, bad gamma. Bad! I'll have to fix it one of those days...

Okay, so the problem only lies with 2x AA, not with 4x AA. I'll try to correct my post to reflect this.


Uttar
 
The 6xS and 8xS to me look like 3x2 and 4x2 arrangements of samples from the pictures on anand and hardocp.

Not sure how nvidia got those though. It could be 4xMSAA mixed with 1.5x1 and 2x1 supersampling but the performance numbers would indicate to me a 4x supersampling with 8x (22% of no AA performance).

It would be interesting to see a shot of a ladder or railing from a game like HL to see what the supersampling component is in these modes.
 
Considering the fact the only numbers nVidia's wanted to show for the product for the last few months are 2x FSAA numbers, I think the current 2xFSAA mode is precisely what nVidia intended it to be. It *might be* .5 or 1x FSAA appearing as 2xFSAA--and I only say that because there is some performance loss associated with the mode (so it's clearly doing more than *nothing*.) But it is clearly not 2xFSAA on the level of ATI's 2x FSAA. Therefore any performance parameters advanced by nVidia thus far to represent 2xFSAA performance on the GF FX are indeed bogus, and I think AnandTech was right in ignoring the modes as if they had not been included (which in reality they have not been.)
 
Bambers said:
That seems a little conspiracy theory based to me ;)

http://www.hardocp.com/image.html?image=MTA0MzYyMDg1OTVjVVNkMzFISXhfNF8xMl9sLmpwZw==

2x works fine in need for speed as that picture from hard ocp shows. QC also works. Question for me is if its only ut2003 and a few games where 2x doesnt work or only nfshp2 and a few games where it does work :-?

*chuckle* The only conspiracy I see is that nVidia will try to entice you to buy this monstrosity-- 8) Heck, this makes the V5 6K look downright elegant by comparison. The hair-dryer whine of the thing would be enough for me even if all of the other facets weren't so poor.

My point with respect to 2xFSAA and the GF FX is that it doesn't appear to work....;) You might be getting some tiny bit of FSAA there, but not a healthy 2x. I imagine it was hoped that web sites around the globe would mindlessly post comparative 2x FSAA speeds between the 9700P and the GF FX and simply note the difference in frame rates without ever paying attention to image quality (Well, I guess Tom's Hardware did exactly that, didn't they?)

But sites like Anandtech, much to his credit, were not taken in by the obvious and transparent ruse. The funny thing about image quality is that when you increase it you pay a performance cost. I guess nVidia wanted to have its 2x FSAA pie and eat it, too. But Anand didn't let 'em have the pie and neither did [H] (recall that in NFS2 on the [H] site the current FX drivers didn't even display the fog)...both sites clearly showing what a terrible 2x FSAA job the GF FX was doing in comparison to the 9700P, and these old hands know that you can get better frame rate performance if you degrade your FSAA sampling routines--so that's the objection in a nutshell. Anand even wrote nVidia about the fact that when selected their 2x FSAA mode didn't seem to do anything....;) They did not reply, he reports.

If nVidia hadn't made such brazen attempts since November to pass around "benchmarks" purporting to show huge differences at 2x FSAA--and no other FSAA modes--I wouldn't be so strident about this. But since the company did, I'm satisifed the current 2x FSAA mode is well known to them and is performing exactly as they intended. Now that it's been called out, they'll probably fix it to provide the modest image quality anti-aliasing improvements one expects to see with 2xFSAA--as well as a corresponding reduction in their current 2x FSAA fps scores, of course. But anyway...the observant sites paid scant attention to the 2xFSAA mode, Anand in particular, for that reason. Not a "conspiracy" just an observation.
 
WaltC said:
My point with respect to 2xFSAA and the GF FX is that it doesn't appear to work....;)
It's possible that the screenshots don't represent what you see on the screen, I have no problem with that. However, nvidia needs to provide a method for capturing the onscreen data for screenshots or else image quality comparisons are useless (and not helping them).

All nvidia has to do is provide an application that can communicate with the driver so that the proper data can be captured. It doesn't even matter if a software pass has to be done during the screenshot as that's not a performance path anyway.

Until then, people will have a tough time comparing image quality... kinda like the problems comparing Quincunx modes on the GeForce 4.
 
Back
Top