AA: which one looks better?

[Bigus Dickus]

If 8X really is 2X MX + 4X SS, then the GFFX will probably suffer greatly in performance using this mode.

Why? It starts out with less bandwidth than the R300, 8 samples is going to take more bandwidth than 6 samples, and SSAA isn't anything like as compressible as MSAA is.

In short, take the performance penalty the R300 has for 6X RGMS, and add roughly 50% to it to find the penalty the GFFX will have for enabling this mixed 8X mode.

However,

My feeling is that 8X is actually 4X MS + 2X SS. Why? Because the NV30's 2X is RGMS, and it's 4X is OGMS, and since we know that 8X is ordered grid, I can't see why they would use a different 2X MS sampling pattern for the hybrid mode than they would for the striaght 2X RGMS.

Of course, perhaps there's a perfectly logical reason why they would, and I'm just to ignorant to know any better.

 

[Crusher]

As others have said, the left "looks" better when just viewing the original image and comparing the two sides. The stairstepping isn't nearly as noticable.

I thought just the opposite; the right side looks like a smoother edge to me in the original size picture. Especially if you stare at the purple section in the center, the left edge looks rouger than the right there. Of course, the right side could just have a better fit to the pixels on my monitor at the resolution I'm using, too (1280x1024).

[quick test]

Yep, the resolution change made a difference, the right side doesn't look as good at 1024x768 as it did at 1280x1024.

Looking at the full screen shots, I can see many places where the NV_25 pic looks like it has smoother edges than the R300 pic though. Particularly on far right side of the structure. There appears to be very little anti-aliasing occuring in some areas there.

R300 6x on left / NV25 8x on right:

 

[Simon F]

Right looks much better, but I've got an LCD display with almost linear gamma ramp

Is that a conscious decision or something over which you have no control?
My LCD at home has a gamma of ~2. That's a far more efficient use of 8 bits per channel than a linear setup.

BTW: Guys and Girls, make sure your display is set to 24/32bit when comparing the images (I accidentally had mine set to 16bit when I first looked at the pictures.)

 

[Nick]

Is that a conscious decision or something over which you have no control?

I don't have control over it, but I like it this way. The gamma is around 1.5, which is much lower than the average 2.2 of a CRT. It results in a much more linear gamma ramp.

I tested it with some gamma comparison charts, and with my software renderer by comparing mipmap transitions with bilinear filtering. Both gave the same result.

My LCD at home has a gamma of ~2. That's a far more efficient use of 8 bits per channel than a linear setup.

Huh, isn't a linear ramp the most efficient setup? When I display all shades of gray from 0x000000 to 0xFFFFFF, I see no transitions and it looks 'linear'. On my friend's CRT, I see the transitions on the white shades, and there are too many black shades that look the same.

 

[Ante P]

bandwidth ain't too shabby either

I'd Disagree.
SS AA will eat more bandwidth, and the GFFX already has less bandwidth than the 9700.
I Highly doubt performance of 8xS will be reasonable to use, except in a few cases (very old games), where its ordered grid nature will reduce its possible effectiveness.
IMO, this is the failing of the GFFX - lack of quality AA modes.

look at 9500 Pro
when overclocked to 9700 Pro levels performance with FSAA is very near it even though it has half the bandwidth
FX will have much more (bandwidth) than that plus I'd expect some of their optimizations to work even beter perhaps

of course that's not taking the super sampling nature of the mode into account
but judging by performance with this mode on my Ti4600 and what type of increase I think would be reasonable by upgrading to FX I think it could very well be playable in many games at 1024x768 (perhaps even combined with Aniso too)

 

[Randell]

look at 9500 Pro
when overclocked to 9700 Pro levels performance with FSAA is very near it even though it has half the bandwidth
FX will have much more (bandwidth) than that plus I'd expect some of their optimizations to work even beter perhaps)

To clarify - I take it you mean much more bandwidth than the 9500Pro as oppsoed to the 9700Pro.

 

[Ante P]

look at 9500 Pro
when overclocked to 9700 Pro levels performance with FSAA is very near it even though it has half the bandwidth
FX will have much more (bandwidth) than that plus I'd expect some of their optimizations to work even beter perhaps)

To clarify - I take it you mean much more bandwidth than the 9500Pro as oppsoed to the 9700Pro.

of course

 

[Simon F]

]My LCD at home has a gamma of ~2. That's a far more efficient use of 8 bits per channel than a linear setup.

Huh, isn't a linear ramp the most efficient setup?

Certainly not! The eye/brain is logarithmic in that equal perceived changes in brightness are actually multiplicative changes in intensity. (eg if A->B looks to be about the same change as C->D then A/B ~= C/D). If you use a linear system then you won't get many different colours with only 8 bpc.

When I display all shades of gray from 0x000000 to 0xFFFFFF, I see no transitions and it looks 'linear'.

That, to me, sounds like you have got an overall gamma that is not linear (I have heard that some LCDs have a more "S" shaped curve).

On my friend's CRT, I see the transitions on the white shades, and there are too many black shades that look the same.

I suspect that he hasn't got his monitor set up properly - it's a common fault.

Have a search for Charles Poynton's colour and gamma FAQs. These explain a lot about the eye, gamma, and setting up monitors.
(Edit: Hmm the original home page seems to be down at the moment but I found this mirror)

See the answer to Question 12.

 

[Nick]

Have a search for Charles Poynton's colour and gamma FAQs. These explain a lot about the eye, gamma, and setting up monitors.

Thanks! It's really strange though, the tests that I've done gave very good results. Only the gray against interlaced black/white lines test is very difficult because the black/white doesn't 'blend' like with a CRT. I can clearly see every line separately. Another example is this test: http://members.aol.com/xfire905/gammatst.htm. Changing my graphic card's gamma doesn't improve it, I always see the center blocks.

You might be right about the S shaped curve. My midtones look slightly better than the white and black ends.

Else it's because I've stared at monitors since I was 10 and now my eyes are used to only that linear range 8)

 

[Simon F]

Have a search for Charles Poynton's colour and gamma FAQs. These explain a lot about the eye, gamma, and setting up monitors.

Thanks! It's really strange though, the tests that I've done gave very good results. Only the gray against interlaced black/white lines test is very difficult because the black/white doesn't 'blend' like with a CRT. I can clearly see every line separately. Another example is this test: http://members.aol.com/xfire905/gammatst.htm. Changing my graphic card's gamma doesn't improve it, I always see the center blocks.

Argh! That image assumes that you want linear behaviour. Avoid it!

I'll see if I can't find (or make) something which is actually correct!

 

[Nick]

Argh! That image assumes that you want linear behaviour. Avoid it!

Oops, you're totally right, those outer pixels have double (digital) intentsity than the inner pixels. I checked it with Paint Shop to be certain. So if it would match I had a perfect linear gamma. I should have noticed this myself

 

[Simon F]

Nick,
Courtesy of a colleague of mine, here is an image for estimating the gamma :gamma_adjust

Note that you HAVE to step a couple of metres away from your monitor in order to average out the chequer board pattern. (It's risky to use 1x1 pattern s due to 'slow' rise times etc of the monitor). I had to reset my monitor brightness (it was far too high) to get it in the range of 2 to 2.2

As for colour, this image is my version of your link tuned for a gamma of 2.2

 

[Nick]

Thanks a lot Simon F!

With your colleague's image, it was very hard to find the gamma. The upper part always looked different, even from a couple of meters away. It looked more like steel, with more blue.

Your corrected image made this even clearer. There was a small but well noticable difference between the colors.

After correcting the colors separately, I tried your colleague's test again. The upper part looked much more like the lower part now. It appears that I have a gamma around 2.1 I need to fine-tune the brightness and contrast too I guess.

I really think there's an overall improvement of the display quality! Thanks! I'm going to play some games to see the difference 'in action' now...

 

[Nick]

Back on topic now...

I have to change my mind and agree with the majority that the left image looks better. It's as if they've switched places

So to everyone who voted for right: check your gamma(s)!

 

[LeStoffer]

I voted left because of those jagged horizontal lines in the lower-left part of the tower. They were enough the tilt the entire weight in favor for the R300 IMHO.

 

[Basic]

It's risky to use 1x1 pattern s due to 'slow' rise times etc of the monitor.

Yes, it's best to have a lot larger patern than 1x1. That's why gamma adjustment programs use horizontal bars (Nx1) instead. There's no use to make the bars higher than 1 pixel though. I've made some interesting images that uses this effect in a somewhat odd way. I'll post it later. (Well, interesting to me at least, but I'm easily amused.)

 

[Ichneumon]

Very interesting!

I played around with my desktop gamma with the hotkey support in Rage3D Tweak just for ease of transitions... I was astonished to see how setting the gamma higher made the right image look better and better and the left worse and worse... The reverse happened as gamma was reduced.

I have no idea how to gamma correct my monitor (not really sure how to use the images Simon provided), but at what I use as default, the left AA image looks way better.

Now I can understand how there can be such incredible differences in opinions about which of the two images is better... I can make either one look better depending on how drastically I play with my gamma settings. I was at a loss originally to understand how anyone could possibly think the right image looked better... well, from my machine they probably would not have.

I guess that makes a good case for "gamma corrected" FSAA, and that the settings used on one machine to take a screenshot very likely will look different when viewed on another machine with different display settings.

Was very eye-opening for me. Thanks a lot Simon and Nick!

 

[Ante P]

^^ me agrees

At first I though people were blind when they didn't see that "the right pic looked obviously better" guess it was just my gamma

as Ichy I too don't really know how to config it properly, where' the noobie guide?

 

[Simon F]

It's risky to use 1x1 pattern s due to 'slow' rise times etc of the monitor.

Yes, it's best to have a lot larger patern than 1x1. That's why gamma adjustment programs use horizontal bars (Nx1) instead. There's no use to make the bars higher than 1 pixel though.

Hmm that's what I originally had (height 1, width 2) but I was getting some very odd patterns on my monitor (possibly aliasing due to insufficient dot pitch). In the end I decided to just double everything <shrug>.

 

[andypski]

^^ me agrees

At first I though people were blind when they didn't see that "the right pic looked obviously better" guess it was just my gamma

as Ichy I too don't really know how to config it properly, where' the noobie guide?

Essentially for each of the red green and blue areas in the test image adjust the gamma ramp for that component (using your hardware's control panel) until the solid colour appears to be exactly the same brightness as the dithered area around it.

When you have done this for all three components the white regions should also be correct, and your display is correctly adjusted.