Need information regard aa cost and perfomance about RSX gpu

[Shifty Geezer]

Hey, Dean, is there anyway that CELL can run supersampling schemes or some other routine that cuts down on aliasing...Maybe CELL can provide some IQ solutions?

Search the forum. Cell doing AA had quite a lengthy discussion at least once.

 

[Megadrive1988]

I remember to have read time ago nao that he said that the aa am is a big problem with rsx but don't remember where i have read it however many games for ps3 have aa4x+hdr,therefore what's the problem ?

graphics that don't totally look "next-gen". sure PS3 can do 4x AA plus some form of HDR (a hack of some kind) in HD resolutions, but there's always a sacrifice, be it to the actual complexity of the graphics, framerate or both.

 

[luke76]

graphics that don't totally look "next-gen". sure PS3 can do 4x AA plus some form of HDR (a hack of some kind) in HD resolutions, but there's always a sacrifice, be it to the actual complexity of the graphics, framerate or both.

well hs it just does not seem to be "ugly" for how much it concerns the graphics and it has aa 4x+hdr

 

[nonamer]

Does the ability to use XDR RAM relieve some of the bandwidth pressure, or does the fillrate limitation still makes AA on RSX particularly difficult?

 

[Cal]

however many games for ps3 have aa4x+hdr,therefore what's the problem ?

I think I found the answer, luke.
http://psinext.e-mpire.com/index.php?categoryid=3&m_articles_articleid=599

Accoring to Marco Salvi (one of Heavenly Sword's developers), they employed a so called "NAO32" format, a variance from LogLuv format. Thus they are bale to use RGBA8 framebuffer to store CIE Luv values. The tradeoff is no alpha blending. Well, I think we can use this on PC, too.

I never believed any PS3 developer could use Cell to do AA.

 

[Acert93]

I think I found the answer, luke.
http://psinext.e-mpire.com/index.php?categoryid=3&m_articles_articleid=599

Accoring to Marco Salvi (one of Heavenly Sword's developers), they employed a so called "NAO32" format, a variance from LogLuv format. Thus they can use RGBA8 framebuffer to store CIE Luv values.

Cal meet nAo, nAo meet Cal.

nAo is Marco Salvi and he is a frequent contributor to the Beyond3D forums. He was gracious enough to discuss and "break" the NAO32 news here first I believe Just as an aside, I believe Bungie is using a number of methods to do HDR in Halo 3. The recent Gamerfest audio and PPT slides is worth looking into in that regards. Maybe something you can use in your next 360 title :smile:

 

[Cal]

Cal meet nAo, nAo meet Cal.

nAo is Marco Salvi and he is a frequent contributor to the Beyond3D forums. He was gracious enough to discuss and "break" the NAO32 news here first I believe Just as an aside, I believe Bungie is using a number of methods to do HDR in Halo 3. The recent Gamerfest audio and PPT slides is worth looking into in that regards. Maybe something you can use in your next 360 title :smile:

Good time to dig old thread.

EDIT: Just dug and read some old threads. So, is AA + NAO32 - blending a good solution for PC game?
Also, I do think Cell is very good at post-processing (especially on those involving convolution and non-linear filters) other than AA.

 

[Shifty Geezer]

EDIT: Just dug and read some old threads. So, is AA + NAO32 - blending a good solution for PC game?

Depends on the game. If you have minimal translucency and pretty much all opaque geometry, and enough shader ops to spare, it'd be a win. If you're rendering a partical-party with loads of alpha blending, it might not be worth it. I guess the most obvious concern would be foliage with alpha'd textures, which feature in so many games. You'd need to render those to a separate buffer in RGB space and composite I think.

 

[Cal]

Depends on the game. If you have minimal translucency and pretty much all opaque geometry, and enough shader ops to spare, it'd be a win. If you're rendering a partical-party with loads of alpha blending, it might not be worth it. I guess the most obvious concern would be foliage with alpha'd textures, which feature in so many games. You'd need to render those to a separate buffer in RGB space and composite I think.

That's enlightening. Particle system can be handled separately due to the huge demand of fill rate, and it's not wise to render particles directly into main framebuffer with AA on. Foliage is not a problem, because it's usually done with alpha mask not alpha blending. Sky box is the one that most possibly contains lots of transparent objects and needs HDR at same time, so it has to be rendered into a separate FP16 buffer. Things definitely get more complex.

 

[nAo]

AA can be expensive per se but on a closed platform you have much more control on it.
It's almost free while you're rendering primivites using long and complex shaders, and you can just not use it on simple shaders (do you really care about multisampling on partricles? .. )
It can also speed up (a lot) shadow maps rendering since you basicly gets z supersampling (and suddenly a ordered sampling grid makes sense )
EDIT: to render foliage you can avoid blending and just use alphatest (this way you also avoid all those nasty rendering order problems..)

 

[nAo]

Sky box is the one that most possibly contains lots of transparent objects and needs HDR at same time, so it has to be rendered into a separate FP16 buffer. Things definitely get more complex.

Transparent sky boxes are easy cause you can blend everything on chip, you really don't need FP16 for them if you employ some other technique (nao32, rgbe, tone mapping in the pixel shader, etc..)