Gamma Corrected FSAA (9700) vs. No Gamma Correction (Ti4600)

Mize

Mize

3dfx Fan
Legend
Thanks to Chalnoth for the proggy. I think gamma corrected FSAA is clearly advantageous even esp. at 2x. You can't see the gradients in these compressed jpgs too well, but the 2x has them if you zoom in on the original capture. I'll post a close-up after I put the kids to bed.

2xfsaa9700.jpg

2xFSAA on the 9700


GF4FSAA2x.jpg

2xFSAA on the GF4

4xfsaa9700.jpg

4xFSAA on the 9700

GF4FSAA4xS.jpg

4xFSAA on the GF4
 
How odd. On these images, also, the 9700 exhibits wierd jpeg pixel "blurring" that aren't apparent on the GF shot.

While I'm sure they're not showing up in the BMPs, I wonder what the hell is causing them.
 
I still for the life of me don't get why png isn't acceptable, especially for such small pictures. Then jpeg compression artifacts are no concern whatsoever. I've been familiar and fond of png ever since CompuServe's .GIF patenting/licensing fiasco, and that was quite a long time ago....how is it such a well kept secret?
 
RussSchultz said:
How odd. On these images, also, the 9700 exhibits wierd jpeg pixel "blurring" that aren't apparent on the GF shot.

While I'm sure they're not showing up in the BMPs, I wonder what the hell is causing them.
Click to expand...

That's because I took the GeForce4 shots, and used maximum-quality compression on them.

As a side note, I'd just like to say that when it comes to situations like this one, jpeg compression is pretty darned stupid. At most, these pics only use 5 colors...there's little reason for the file sizes, even at max-quality compression, these things have in jpeg format...
 
If the NV30 doesn't include gamma correct downsampling, it could easily be done via fragment shader with a small bandwidth/fillrate hit in a 2d image processing post-process.

However, I highly doubt that the NV30 will lack this feature. First, this has been discussed at several Meltdowns with regard to DirectX9 and was added to the DirectX9 spec.

Second, it wasn't included in the past because previous cards (old 8 to 10-bit per component pipeline) lacked the internal precision. Doing x^2.2 and x^0.45 on 8-bit texel/pixel read write is pointless. The error would quickly build up and swamp any gain. The 128-bit FP formats now make gamma correct math useful.


The benefits go way beyond just FSAA correct downsampling.
 
I hope so too, DemoCoder :)

Now that I fully understand the issue at hand, this is another must-have feature on the next video card I purchase...
 
Oh, and one other thing.

The GeForce4 shot is actually 4xS FSAA, in an attempt to try to keep the sampling pattern as close as possible to what the Radeon uses.
 
Chalnoth sent me a copy of his program but swapped the colors on me so I get to show the black on white. I put everything in a big .jpg file and used a quality of 75 via Photoshop 7.

http://www.thoroughbred-data.com/nudies/chtest/chalnothtestcompare.jpg

http://www.thoroughbred-data.com/nudies/chtest/chalnothtestcompare.png

Didn't want to put in directly in the thread because it would bork everyone's readability of the thread and that really irks me ;)

Edit #2 added a link to a .png file. Didn't realize the .png actually turns out smaller, will start using those more in the future.
 
Oh, whoops, that's a big brain fart :p

Just realized that 4xS was a Direct3D-only mode. The shot was using the driver setting for 4xS, however, so I suspect it was actually 4x. Sorry about that...
 
Thanks for posting those shots, Donald.

It seems obvious that the Radeon 9700 looks a whole lot better with the black background, but the 4x shots on the GeForce4 look a bit better, at least on my monitor, with the white background.

What this tells me is that the Radeon 9700 settings could stand to be tweaked a little bit.

It would be really nice if ATI produced a program that displayed scenarios similar to both of these, and allowed user selection of the proper gamma setting. I hope nVidia does the same if they release a card that also does a similar gamma correction during FSAA. Unfortunately, I'm not overly-hopeful for such a thing soon after NV30's launch, as nVidia's never been the best at producing good driver control panels. However, I should suspect that it could be tweaked via a registry setting, but that isn't nearly as good, as you can't see the response in realtime.
 
Chalnoth said:
It seems obvious that the Radeon 9700 looks a whole lot better with the black background, but the 4x shots on the GeForce4 look a bit better, at least on my monitor, with the white background.
Click to expand...
The AA area is much too large on the GeForce. Also, some of the lines in the GeForce 4x case look jaggy, even with the white background.
What this tells me is that the Radeon 9700 settings could stand to be tweaked a little bit.
Click to expand...
Or the GeForce is blurring too much. AA lines should be thin, not fat.

Edit: To clarify: The GeForce lines look too fat. A line at 45 degrees should hit only two adjacent pixels, unless the line happens to cross directly through the center of the pixel. Looking at the GeForce 4x picture, the line is far too thick. Plus, two adjacent pixels should not both be full intensity on a single pixel wide line.
 
What's up with 9700's 6x FSAA? All lines look equal or better than 4x except 45-degree line in 1. and 4. quarter - gamma correction is not working for this particluar line or what?

Gotta love this kind of synthetical tests ;)
 
Sabastian said:
I noticed the Radeon 9700 seems at the center have more depth or something. While the Geforce lines at the center appears to meld into a large spot. What is the significance of that? BTW neat little test.
Click to expand...

I think the main reason is the gamma correction. As OpenGLguy pointed out, the GeForce lines appear too thick, which compounds itself at the center of the image (this is looking at the black lines on white pics, of course).
 
OpenGL guy said:
Or the GeForce is blurring too much. AA lines should be thin, not fat.

Edit: To clarify: The GeForce lines look too fat. A line at 45 degrees should hit only two adjacent pixels, unless the line happens to cross directly through the center of the pixel. Looking at the GeForce 4x picture, the line is far too thick. Plus, two adjacent pixels should not both be full intensity on a single pixel wide line.
Click to expand...

I think that conversely, the AA lines on the Radeon shot look too narrow.

It seems apparent to me that, at least on my monitor, the black background shots look significantly better for the Radeon 9700 than the white background shots do. I think this alone is reason enough to have user-selectable corrected-gamma FSAA whose results are visible in realtime.
 
Is this an OpenGL app? Is Line Anti-Aliasing on? I know I have a registry setting for that for my Radeon drivers...I'd expect that could be a possible explanation for the "fatness" problem.
 
You must log in or register to reply here.

Similar threads

D
Gamma-corrected FSAA - what is it?
2
Replies
24
Views
6K
Xmas
X
M
Unimpressed by Antialiasing (continued)
2
Replies
21
Views
4K
KimB
K
Back
Top