Quincunx & FSAA: how is it changed between Geforces
- Thread starter Above
- Start date
Ailuros said:
I dunno, less aliasing by blurring doesnt seem like a positive to me...
I mean, i can get you no aliasing, for almost free!! I will just blur more!
WOW.
OTOH, some may see things differently than me. I view quincunx like aths does, as a useless marketing trick.
The package of the Accuview algorithm DOES not contain ONLY Quincunx nor do any other modes blur as Quincunx does.
Aths' point was that a user can very well skip Quincunx overall and use all the other available options and that's what I replied to.
In the end I haven't seen one algorithm that doesn't show from slight to noticable bluring occassionaly up to date and they don't even use a Gaussian filter (or whatever it is) like Quincunx. On contrary I haven't seen many occassions where 2xRGMS blurs significantly the scenery; in the end thanks to it's rotated nature combined with aniso it does a far better job on edges than garden variety ordered grid anything-sampling does.
Feel better now?
Aths' point was that a user can very well skip Quincunx overall and use all the other available options and that's what I replied to.
In the end I haven't seen one algorithm that doesn't show from slight to noticable bluring occassionaly up to date and they don't even use a Gaussian filter (or whatever it is) like Quincunx. On contrary I haven't seen many occassions where 2xRGMS blurs significantly the scenery; in the end thanks to it's rotated nature combined with aniso it does a far better job on edges than garden variety ordered grid anything-sampling does.
Feel better now?
Hellbinder
Banned
I am a bit confused...
I now own a GF4 and you can spot Quincunx a mile away. I spent about an hour comparing all the different FSAA methods in Anarchy Online yesterday.
The quality of quincunx imo is just poor. There is a general "haze" over everything. The jaggies are still quite noticeable. Especialy verticle.
Talking about changes in Quincunx is moot. It may be fast but the quality is just to poor to use. For the life of me i cant figure out why anyone likes it, uses it, or thinks of it so highly...
Its like mass self deception or something.
I now own a GF4 and you can spot Quincunx a mile away. I spent about an hour comparing all the different FSAA methods in Anarchy Online yesterday.
The quality of quincunx imo is just poor. There is a general "haze" over everything. The jaggies are still quite noticeable. Especialy verticle.
Talking about changes in Quincunx is moot. It may be fast but the quality is just to poor to use. For the life of me i cant figure out why anyone likes it, uses it, or thinks of it so highly...
Its like mass self deception or something.
Hi Mike! Sorry for the long time getting back to ya. You know how that work thing goes... ruins our fun digging into lifes little mysteries sometimes. 
I was very perplexed by what I'm seeing in Tribes2 and have a couple theories about it. Here's what I'm finding/guessing/seeing:
It looks as if there is something happening with the framebuffer contents without postfilter, but I'm not entirely sure it's quincunx that is causing it.
A screenshot of Quincunx and playing the game live of Quincunx is still not what I see as being correct as the Quincunx *real* still has additional filtering not seen in the screenshot, yet the screenshot does have *some* level of filtering visible.
My speculation at this point is there is some level of anisotropic filtering being applied erroneously as the setting is LOW to HIGH and I cannot for the life of me find an "off" setting nor a "enableAnisotropic" cvar in the ClientPrefs. Seems there are cvars for every other stupid thing in the world except anisotropic filtering.
The bizarre thing is, enablement of quincunx makes the framebuffer contents change as opposed to without it, so we are indeed on the outer cusp of some interesting findings that might go a long ways towards explaining what is happening. When I have more time, I'll delve a bit deeper but I can still whole heartedly say screenshots are still not illustrative of Quincunx, albeit in this case they are a tad "more accurate" from a clarity standpoint but still missing much of the filter/blur.
Above-
It's not really scabbling, but instead trying to form an understanding about it. It's different from the GF3 and so far, there is much evidence that screenshots do not portray the quality of Quincunx accurately. It's important that depictions of something be as true and accurate as possible.
Also the first step is understanding- which only leads to better applications and usage of said feature.
Like I said, I love using Quincunx on games like Quake2... and I have no way of illustrating what it does for textures here since it doesnt show up in any of my screenshots. I fire up LFire CTF and play for hours and Quincunx looks great. The only problem is it drops performance down to a measly 420 fps timedemo speed so I'm not sure if I can rail as accurately with that kind of handicap. It's also pretty nice in Descent3, Freespace2 and NFS-PU as well. It has it's perks for some older games with incredibly grainy textures and that can stand a little more edge filtering, but the overall effect just cant be shared since there seems to be no way to capture in screengrabs.
Cheers,
-Shark
I was very perplexed by what I'm seeing in Tribes2 and have a couple theories about it. Here's what I'm finding/guessing/seeing:
It looks as if there is something happening with the framebuffer contents without postfilter, but I'm not entirely sure it's quincunx that is causing it.
A screenshot of Quincunx and playing the game live of Quincunx is still not what I see as being correct as the Quincunx *real* still has additional filtering not seen in the screenshot, yet the screenshot does have *some* level of filtering visible.
My speculation at this point is there is some level of anisotropic filtering being applied erroneously as the setting is LOW to HIGH and I cannot for the life of me find an "off" setting nor a "enableAnisotropic" cvar in the ClientPrefs. Seems there are cvars for every other stupid thing in the world except anisotropic filtering.
The bizarre thing is, enablement of quincunx makes the framebuffer contents change as opposed to without it, so we are indeed on the outer cusp of some interesting findings that might go a long ways towards explaining what is happening. When I have more time, I'll delve a bit deeper but I can still whole heartedly say screenshots are still not illustrative of Quincunx, albeit in this case they are a tad "more accurate" from a clarity standpoint but still missing much of the filter/blur.
Above-
It's not really scabbling, but instead trying to form an understanding about it. It's different from the GF3 and so far, there is much evidence that screenshots do not portray the quality of Quincunx accurately. It's important that depictions of something be as true and accurate as possible.
Also the first step is understanding- which only leads to better applications and usage of said feature.
Like I said, I love using Quincunx on games like Quake2... and I have no way of illustrating what it does for textures here since it doesnt show up in any of my screenshots. I fire up LFire CTF and play for hours and Quincunx looks great. The only problem is it drops performance down to a measly 420 fps timedemo speed so I'm not sure if I can rail as accurately with that kind of handicap.
It's also pretty nice in Descent3, Freespace2 and NFS-PU as well. It has it's perks for some older games with incredibly grainy textures and that can stand a little more edge filtering, but the overall effect just cant be shared since there seems to be no way to capture in screengrabs. Cheers,
-Shark
ROFL
Since you meantioned NFS/PU: I haven't cross referenced my results yet, but either something went wrong with nvhardpage or I'm getting almost identical results with either 16tap or 32tap anisotropic in that specific game. As I know you've tested it, how do your results look like?
Do we agree that Quincunx is significantly better on the Geforce4 or even that it is an irrelevance?
Also, are we seeing any settings that can not be hacked on the Geforce 3. As 4xs can be set. Before I had come here I had never read of 2x RGMS before yet it is said to implemented even on the Geforce 3.
So the Geforce 4 strength is in the faster implementation of what the Geforce 3 introduced, mirroring the situation with the earlier Geforce and Geforce 2.
Wishing to avoid specific pricing examples, currently the cheapest Geforce 4 is priced from 160% of the cheapest Geforce 3. Would the forum strongly recommend against the earlier card on such basis? Does the Geforce 4 bring a utopia of light Antialiasing with little cost rendering it practical as opposed to Geforce 3 impractical?
Also, are we seeing any settings that can not be hacked on the Geforce 3. As 4xs can be set. Before I had come here I had never read of 2x RGMS before yet it is said to implemented even on the Geforce 3.
So the Geforce 4 strength is in the faster implementation of what the Geforce 3 introduced, mirroring the situation with the earlier Geforce and Geforce 2.
Wishing to avoid specific pricing examples, currently the cheapest Geforce 4 is priced from 160% of the cheapest Geforce 3. Would the forum strongly recommend against the earlier card on such basis? Does the Geforce 4 bring a utopia of light Antialiasing with little cost rendering it practical as opposed to Geforce 3 impractical?
Hellbinder
Banned
I had only previously seen QC FSAA on a friends GF3... Remembering what that looked like and comparing to my GF4, I would have to say that QC is better now. To me though its still like saying "lucky you, your going to get beheaded with only one swing of the axe instead of 3 or 4."
In fact to take this a step futher I dont like *any* form of FSAA on *any* hardware. After about a week of messing arouond with various games, various combinations of FSAA, Aniso, LOD etc... I simply dont see the benefit. Even the Supersampleing method does not make that big an impact on Texture quality imo. Even 4x aniso leaves very visible jaggies. In my mind if you are going to take a performance hit, it had better be worth it.
Now that I have first hand experience on both sides of the track (Ati/Nvidia) I just dont like the FSAA road that is being followed. IQ would be much more benefited by very high persicion Z, color, and filtering.
On the Other hand, i have a high quality flat screen 19" monitor. I run every game at 1280x1024x32. Perhaps people with 14" and 15" monitors are much more benefited by FSAA as they have to run at 800x600 or 1024x768. Of course, if we are talking about someone who can afford a GF4, then surley they have a better monitor also.
Hellbinder[CE said:
I personally find that hard to believe. Now that I've been playing with at least 2x FSAA in all of my 3D games, I can tell when FSAA is disabled immediately, and it really bothers me. It's not so much that you notice the difference when you turn it on, but you notice it more when you've been playing with it for a while, and you turn it off.
Quite note, however: My current system is still a bit blurry at 1600x1200, so I can't really tell the difference very well at that resolution, but at lower resolutions it is very obvious.
Jerry Cornelius
Newcomer
Supersampling is a crude solution for texture aliasing. You're basically just oversampling the snot out of an image and then generalizing it to make up for a trade off already made to preserve speed. If you allotted a budget of four times the samples do apply however you wanted, you would likely end up with less aliasing and higher quality filtering with better framerates and a smaller memory footprint.
Jerry Cornelius said:
But you really can't do that in an immediate-mode renderer, since the makeup of the scene won't be known until it's ready for display. Still, anisotropic filtering does a good job at this, and the performance hit will get smaller as image quality increases into the future. Today's anisotropic filtering tends to use either too few samples (Radeon 8500) or too many (GeForce3/4). With more silicon space, chip designers will be able to pack in more sample calculation logic in order to optimize fillrate usage. At the same time, the pixel pipelines themselves will get more complex, meaning that it will not make sense to just increase the overall fillrate instead of increasing the efficiency of anisotropic filtering.
I expect that into the future (probably not more than two years, possibly in the next generation...), we will have greater than 16-degree anisotropic with the angle-independent quality we see in the GF3/4 cards, but at a performance hit around 30%.
Jerry Cornelius
Newcomer
I was refering to texture anti aliasing.
It's hard to please a 3dfx fanboy
I have to say after experience of the Voodoo 5500 how disappointed I am at attempts to refine the image on cards.
I play simulations and removing crawling edges and textures is my priority over features like vertex shaders. I can get away 3dfx 2x and even 16 bit colour where I would have to use 4x supersampling and 32 bit colour on other cards. I kid you not.
I really don't like what I'm seeing here: http://www.beyond3d.com/reviews/visiontek/nvidia_gf4ti4400/index6.php
The Geforce 4 method of FSAA requires 4x and anisotropic to really compare with the 3dfx quality, but this takes a large performance hit. From 135 to 33 fps. All the improvements of two graphics generations seem to become less significant at this point.
At this point I am stuck to know what to do. The improvements relative to my uses don't seem to justify spending a lot on a new videocard. I can buy a GF4MX for 90 euros, a GF3 for 130, or a GF4 for 180. What do I do just to basically improve antialiasing speed as cheaply as possible until some quantum leap improvement comes along?
I have to say after experience of the Voodoo 5500 how disappointed I am at attempts to refine the image on cards.
I play simulations and removing crawling edges and textures is my priority over features like vertex shaders. I can get away 3dfx 2x and even 16 bit colour where I would have to use 4x supersampling and 32 bit colour on other cards. I kid you not.
I really don't like what I'm seeing here: http://www.beyond3d.com/reviews/visiontek/nvidia_gf4ti4400/index6.php
The Geforce 4 method of FSAA requires 4x and anisotropic to really compare with the 3dfx quality, but this takes a large performance hit. From 135 to 33 fps. All the improvements of two graphics generations seem to become less significant at this point.
At this point I am stuck to know what to do. The improvements relative to my uses don't seem to justify spending a lot on a new videocard. I can buy a GF4MX for 90 euros, a GF3 for 130, or a GF4 for 180. What do I do just to basically improve antialiasing speed as cheaply as possible until some quantum leap improvement comes along?
Re: It's hard to please a 3dfx <bleep>
Those graphs seem a bit strange...that game seems to take a much larger hit from enabling FSAA than most other games. I guess that particular game is just much more memory bandwidth-limited than most others. Anyway, with a Ti 4600, most games will be easily playable at 1024x768x32 with 4xs FSAA and 8-degree anisotropic (one thing about the aniso...it will take more a performance hit with 4xs FSAA than the other modes...but 4xs is a hybrid supersampling mode that does increase texture clarity somewhat).
Anyway, there's no possible way that a Voodoo5 5500 could possibly do as well at that resolution with 4x FSAA as the Ti 4600 does with 4xs and anisotropic, provided all details were the same (the 5500 may well not be able to run with the same quality settings), and the texture detail on the Ti 4600 will be far in excess of what the V5 can offer. Edge AA quality will be similar.
Above said:
Those graphs seem a bit strange...that game seems to take a much larger hit from enabling FSAA than most other games. I guess that particular game is just much more memory bandwidth-limited than most others. Anyway, with a Ti 4600, most games will be easily playable at 1024x768x32 with 4xs FSAA and 8-degree anisotropic (one thing about the aniso...it will take more a performance hit with 4xs FSAA than the other modes...but 4xs is a hybrid supersampling mode that does increase texture clarity somewhat).
Anyway, there's no possible way that a Voodoo5 5500 could possibly do as well at that resolution with 4x FSAA as the Ti 4600 does with 4xs and anisotropic, provided all details were the same (the 5500 may well not be able to run with the same quality settings), and the texture detail on the Ti 4600 will be far in excess of what the V5 can offer. Edge AA quality will be similar.
Yes for sure I would well expect GF4 4xs with anisotropic (but not 4x MSAA plain or hacked 4x OGSS) to look better and be faster than basic 3dfx 4x. Maybe I didn't signal quite blatantly enough that I have to be more concerned with antialiasing than cutting edge features, which is why I would even consider an MX440 (gasp).
I think nvidia did a good thing on 2x and Quincunx performance but that 4xs 8x anisotropic is the ideal to aim at.
I think nvidia did a good thing on 2x and Quincunx performance but that 4xs 8x anisotropic is the ideal to aim at.
Me, I don't like 4xs in the least.
What I would like to see is a pseudo-random multisampling technique (like people thought ATI's smoothvision was going to be) in combination with higher degrees of anisotropic filtering, and possibly the ability to set a minimum amount of filtering in any direction (Currently the minimum amount of filtering in any direction is two samples...increasing that would decrease performance, but decrease texture aliasing, similar to how SSAA works).
I don't have much hope for increasing the minimum texture filtering method anytime soon, but I do seriously hope that the NV30 has support for pseudo-random multisampling, in conjunction with a higher number of samples than is available currently, and ATI actually gets their Smoothvision to finally work in their next product (but with a multisampling implementation this time...).
What I would like to see is a pseudo-random multisampling technique (like people thought ATI's smoothvision was going to be) in combination with higher degrees of anisotropic filtering, and possibly the ability to set a minimum amount of filtering in any direction (Currently the minimum amount of filtering in any direction is two samples...increasing that would decrease performance, but decrease texture aliasing, similar to how SSAA works).
I don't have much hope for increasing the minimum texture filtering method anytime soon, but I do seriously hope that the NV30 has support for pseudo-random multisampling, in conjunction with a higher number of samples than is available currently, and ATI actually gets their Smoothvision to finally work in their next product (but with a multisampling implementation this time...).
Chalnoth said:
Why would you want MS?
If you want to get rid of Aliasing (both edge and texture) then SS is the way to go.
Look at it this way: 4x Smoothvision + 16xAniso (r8500) looks significantly "cleaner" and with less aliasing than 4xMSAA + 8x aniso (gf3).
Aniso is good, dont get me wrong, but it will never get rid of as much of the high freqency noise as aniso+ SS will.
Therefor, the EVENTUAL end solution will be to SOME form of SuperSampling, once rendering fillrate/bandwidth isnt really a concern.
MS is just an interim measure to provide better performance at the cost of IQ, just like ATI's aniso.
Fillrate/bandwidth will always be a concern with SSAA, apart from a deferred renderer.
I still consider a SSAA/aniso combination computational overkill, textures do not need to get filtered twice.
Why not combine SSAA and MSAA in an algorithm like with 4xS, only with far more MSAA samples as so far as an alternative?
Apart from that I´d still love to see how a Z3 algorithm would look and perform combined with good anisotropic filtering.
I still consider a SSAA/aniso combination computational overkill, textures do not need to get filtered twice.
Why not combine SSAA and MSAA in an algorithm like with 4xS, only with far more MSAA samples as so far as an alternative?
Apart from that I´d still love to see how a Z3 algorithm would look and perform combined with good anisotropic filtering.
Similar threads
