Killzone 3 and MLAA

[HollovVpo1nt]

http://twitter.com/digital_foundry/statuses/16069843537

DigitalFoundry usual knows what they're talking about. But I don't understand it. How can an SPU intensive engine like the Killzone (2) engine be doing MLAA? Correct me if I'm wrong, but it should take 20ms to execute the technique on a single SPU. That's a little too long for a single SPU, so my guess is that they'll run it in parralel with 4/5 SPU's reducing it to 4/5 ms. But how does that fit with all the other post processing beeing done by the SPU's in Killzone?

 

[Delta9]

I hope they can get the MLAA in.
Here is a list of some spu jobs in KZ2.

 

[nightshade]

I remember a chart showing the SPU usage of KZ2 in some sort of graph.
And from what I recall there were a lot of unused/empty cycles....like really a lot.

Too bad I can't find that chart now...but I know I had it saved somewhere in my PC, I just can't remember where.

 

[HollovVpo1nt]

I remember a chart showing the SPU usage of KZ2 in some sort of graph.
And from what I recall there were a lot of unused/empty cycles....like really a lot.

Too bad I can't find that chart now...but I know I had it saved somewhere in my PC, I just can't remember where.

Look one post above you

Anywho, that's cool. I'm really eager to know how they managed that. Maybe I'll ask GG personally

 

[nightshade]

Look one post above you

I wasn't talking about that, the one which I'm referring to is a proper chart with percentage/numbers.

 

[HollovVpo1nt]

I wasn't talking about that, the one which I'm referring to is a proper chart with percentage/numbers.

Well if you find it, be sure to let me know

 

[nightshade]

Okay turns out I was wrong about percentage/numbers....but anyways here's the chart:

And this is the thread where this chart was posted first & discussed on further:
http://forum.beyond3d.com/showthread.php?t=41998&page=10

 

[Laa-Yosh]

There's always room to optimize, especially with art assets using experience gained from a completed title. The daytime setting also means that among other things, they can reduce the number of active lights for most scenes, as a start. That should already free up some resources.

 

[nightshade]

I remember someone mentioned about GG's philosophy being "If it works then no need to optimize it further". Donno if it really was that way, but yea there's a lot of headroom left for KZ3.

Also would reducing lights do any good for a deferred renderer ? Considering how less of an issue lights are on a deferred rendered engine you'd think that reducing lights would mean actually mean under utilization of tech.

 

[mitran]

I wonder if Killzone3 will have "with a dash of photonic voxel warp drive potatoes'" XD like LBP2 has in its engine.
http://www.eurogamer.net/articles/digitalfoundry-lbp2-tech-interview?page=3

 

[N_B]

Okay turns out I was wrong about percentage/numbers....but anyways here's the chart:

And this is the thread where this chart was posted first & discussed on further:
http://forum.beyond3d.com/showthread.php?t=41998&page=10

I'm thinking this one is more recent:

 

[patsu]

I remember someone mentioned about GG's philosophy being "If it works then no need to optimize it further". Donno if it really was that way, but yea there's a lot of headroom left for KZ3.

In their KZ2 GDC slides, they recommended people not to optimize too early, and not to code optimize specifically for SPU (e.g., Don't use assembly). They advocate smart design and rely on basic hardware characteristics (e.g., number of cores, vector math) to achieve good performance.

In some subsystems (e.g., the Depth of Field module), it was done as an experiment. They didn't expect it to work but it turned out to be one of the most successful experiments. They left it "as is" since the result looked good enough.


Naughty Dog had similar recommendations (Use smart design and basic hardware characteristics, don't code optimize using assembly because it's hard to maintain). It seems that they went against their advices since one of the interviews mentioned that they used SPU assembly in Uncharted 2. Then again, they also said they had some development "close calls" because they relied too much on paper design instead of trying it out for real, until it's a little too late.

 

[nightshade]

I'm thinking this one is more recent:
http://i43.tinypic.com/2zoy87r.jpg

Scratch that...this was the chart I was referring to back then, not the one which I posted.

 

[i nerini del buio]

aren't you guys overestimating this MLAA?
there's plenty of people all over the internet asking for MLAA being applied in every game °_° or at least in every first party game.

i mean, it's a pretty tech, but i would rather like to see msaa4x applied in every game.
resident evil 5 is the game with the best IQ so far, it has a beautiful aa4x indeed. sega rally too.

i think there's really something wrong with MLAA; beautiful on still images but very inconsistent while moving.

maybe it would be beautiful to see some msaa+mlaa, but if i had to choose between the two, i would always go for msaa

 

[patsu]

I think they just want to replace no AA, bad use of AA filters, and QAA with some variant of MLAA. If it's cleaner than MSAA like GoW3, then all the better.

It is said that aliasing is more noticeable in stereoscopic 3D gaming. It fuels more expectation for MLAA (or equivalent) in these games.

 

[Delta9]

aren't you guys overestimating this MLAA?
there's plenty of people all over the internet asking for MLAA being applied in every game °_° or at least in every first party game.

i mean, it's a pretty tech, but i would rather like to see msaa4x applied in every game.
resident evil 5 is the game with the best IQ so far, it has a beautiful aa4x indeed. sega rally too.

i think there's really something wrong with MLAA; beautiful on still images but very inconsistent while moving.

maybe it would be beautiful to see some msaa+mlaa, but if i had to choose between the two, i would always go for msaa

I don't know,RE5 looks pretty jaggie compared to GOW3.
http://images.eurogamer.net/assets/articles//a/7/1/3/1/4/1/360_012.bmp.jpg
http://images.eurogamer.net/assets/articles//a/1/0/2/4/8/8/1/AA_003.jpg.jpg

 

[SG79]

*snip*

RE5 had a dynamic use of 4XMSAA on the 360 (QAA on the PS3) that went down to 2X and then completely off when things got hectic. Pretty game, but certainly not the best IQ. Capcom's new revision with LP2 uses consistent 2x MSAA and nothing on the PS3.

Neither MSAA or MLAA win in terms of consistency but based on the GOW3 implementation alone, it's no wonder people wish for more use of MLAA. The game did have an astonishing IQ, especially compared to its previous MSAA implementation.

 

[i nerini del buio]

I don't know,RE5 looks pretty jaggie compared to GOW3.
http://images.eurogamer.net/assets/articles//a/7/1/3/1/4/1/360_012.bmp.jpg
http://images.eurogamer.net/assets/articles//a/1/0/2/4/8/8/1/AA_003.jpg.jpg

only in still frames.
while moving mlaa looks wierd to me. and videogames do use to move

 

[Gitaroo]

specular map or shader aliasing probably? If thats the case even 8XMSAA wont help. Personally, I play RE5 on PC with 8XCSAA in 1360X768, still doesnt look as clean as GOW3.

 

[ultragpu]

MLAA has that 16xmsaa+ effect in certain areas and coupled with good DOF and motion blur GOW3 is indeed godly in iq.

Update from DF

No sign of supersampling on sub-pixel edges. It's the real deal. Oh, and video encode is 59.94fps, actual performance similar to Killzone 2.