[Trejser]

It's MLAA, AWESOME!

AA on the cpu is MLAA Morphological Antialising. We saved 5-6 miliseconds by moving it off the cpu's. Many props to our coder Cedric for making this happen and it looks way better!

http://forums.godofwar.com/t5/God-of.../30061/page/14

[patsu]

Nice find ! Great to see more people using the technique. Will definitely pay attention to the AA once I have the game.

This sounds useful too:

Number of dynamic lights - ALOT! that is one of the big features of our engine. We built it around being able to use up to 20 dynamic lights per game object. The light can be big or small, it doesn't matter. In the end, i believe we support up to 50 dynamic lights per game object. We are not using a differed lighting scheme. Our lead programmer Vassily came up with this amazing system during pre-production, us artists love it!!! We can place lights in Maya and have them update in realtime in the game on the PS3, its like being able to paint with lights. Lighting is fast and a very enjoyable artistic process.

[Oninotsume]

Hi,

Excellent find. I'm glad to see the Gow III devs are so open to
answering user inquiries.

Great to see more people using the technique.

Yes. If the method proves applicable across various game designs,
it would be nice to see it rolled into the SDK.

I'm even more hyped for this game now. Can't wait to feel my eyes
pop out

Oninotsume

[marcus_rocks]

Originally Posted by Trejser

It's MLAA, AWESOME!



http://forums.godofwar.com/t5/God-of.../30061/page/14

Problem solved. Next problem: solve the transparency problem: fire, smoke, water... move everything to SPUs and we have a PS4

[DonaldDuck]

In fact, if someone achieves that goal (a fast and cheap way of handling transparencies in PS3) that would help programmers of multiplatform games a lot. Nowadays, that thing seems to be the main problem when porting a game from XTS to PS3.

In some ways, the SPU´s are allowing programmers to make things in ways no one really expected. It could be considered a fault in GPU hardware design (RSX being a bit rigid) or an achievement in system design. May be both.

Well, to my taste is complex but interesting. As we say in Spain "el hambre agudiza el ingenio", being hungry makes us smart... I suppose some programers prefer the other way. I probably would. Nevertheless, it´s interesting.

[corduroygt]

If PS3 transparencies are an issue, why does frostbite engine have the exact opposite issue, with the snow covered trees looking noticably worse on the 360? I also think quarter resolution is a satisfactory method of dealing with transparencies in the PS3.

[nightshade]

I think a quater reso with MSAA applied & slight blur or haze to transparency (KZ2) is just perfect for most games.

[_phil_]

From Tim Moss twitter:


The #gow3 AA technique saved 5ms from the GPU, costs ~20ms on 5 SPU's (~4ms Latency), its very pretty and only on #ps3 ;-P

[Shifty Geezer]

Thanks _phil_. That actually makes sense! Although 20ms to add AA is damned expensive.

Edit : Actually it doesn't make sense. 60fps afford ~17ms per frame. 20ms of AA processing just isn't possible! 4 frames would take 80ms, which would be 5 frames at 60fps. The only way I can see it working is if he really means that it adds 4ms of latency due to 4ms processing time. That amounts to a kinda weird total of 20 ms as 16ms to generate the frame plus 4ms to apply AA, but during that 4ms the next frame is being generated, such that the game runs at 60fps with a 4ms lag added to the time taken from updating the game to creating the current frame image.

[patsu]

Can you stuff even more "ms of work" to the SPUs ? Or are they all overwhelmed in GoW3 now ? What other improvements/problems/regrets would you overcome if you were to take another stab at it ?

[T.B.]

Originally Posted by Shifty Geezer

Actually it doesn't make sense.

4ms * 5 SPUs = 20ms of SPU time.

[Shifty Geezer]

Ahh, right. Still sounds expensive (20ms figure is a bit dumb. Are we going to count current GPU times in the hundreds of ms because they have so many shaders!) compared to hardware MSAA, especially at 60fps, but the quality seems worth it.

[T.B.]

Originally Posted by Shifty Geezer

Ahh, right. Still sounds expensive (20ms figure is a bit dumb. Are we going to count current GPU times in the hundreds of ms because they have so many shaders!) compared to hardware MSAA, especially at 60fps, but the quality seems worth it.

That's just how we do it on Cell, TBH. You have 6 cores and while a GPU always runs the same program on all "cores", that's just not true for the SPUs. So if you have a properly parallelisable problem, it makes sense to measure performance in "1 SPU time".

"Example": I have 100ms of SPU time at 60Hz and I budget up to 20ms for a piece of code. Maybe I just run it on 2 SPUs and get 10ms latency. Or I put it on 5 and get 4ms. That decision will depend on scheduling needs, but I still know how much SPU time I've committed.

Still sounds dumb?

[corduroygt]

Considering the ps3 architecture, I'd say 20ms total spu time is less valuable than 5ms rsx time, not to mention the much better quality.

[Shifty Geezer]

Originally Posted by T.B.

Still sounds dumb?

Not as a measure of SPU usage as it use it, but in it's use in these quotes, it seems a mixed measurement. "We spend 20ms SPU time to do 4ms work that takes 9ms on GPU." That's all a bit muddled! But in terms of Cell's structure, I can see it makes sense. My language interests would rather see a different measure created though - A Cell unit where there are 6000 units (SPE ms) to a PS3's Cell. Maybe 'clicks'. "We saved 5 ms by shifting the AA from GPU to CPU. Our FSAA process takes 20 clicks on Cell, which we spread across 5 SPEs, 4ms each."

[patsu]

I agree with Shifty. I took 20ms to be the end-to-end duration. So it's 4ms in duration (over 5 SPUs) vs 5ms on RSX ? Same quality output ?

EDIT:
Don't quite understand his follow up:

TDMoss @vanmnguyen no, 5ms gain on gpu. It can be doing other stuff while spu is doing AA #gow3

Ok... I think I understand it a little better. Each of the 5 SPUs took 4ms, but the duration is longer (depending on how the SPUs line up their work). And the overall end-to-end time is 5ms faster than on a GPU ? with same quality output ?

[Weaste]

For better or worse, Sony really like maying funny/bizarre machines.

[Titanio]

I read it as ~4ms from start to finish when doing it on 5 SPUs. Occuping 5 SPUs for 20ms per frame each in a 30fps game (33ms per-frame budget) would be far too much.

They previously said this was 5-6ms faster than doing 'regular' AA on GPU (I guess 2xMSAA given that's what they were using before). So I take that to mean 5 SPUs doing MLAA = less than half the cost of 2xMSAA on RSX.

On a side note, by way of comparison...although it's likely not apples-to-apples, Intel reported performance of ~46ms of processing time on a single 3Ghz Intel core with its MLAA implementation (for a 720p frame - 20m pixels per second). A single SPU with Santa Monica's implementation would be ~20ms.

[marcus_rocks]

Originally Posted by DonaldDuck

In fact, if someone achieves that goal (a fast and cheap way of handling transparencies in PS3) that would help programmers of multiplatform games a lot. Nowadays, that thing seems to be the main problem when porting a game from XTS to PS3.

In some ways, the SPU´s are allowing programmers to make things in ways no one really expected. It could be considered a fault in GPU hardware design (RSX being a bit rigid) or an achievement in system design. May be both.

Well, to my taste is complex but interesting. As we say in Spain "el hambre agudiza el ingenio", being hungry makes us smart... I suppose some programers prefer the other way. I probably would. Nevertheless, it´s interesting.

I don't think transparency is a hard problem. Everything within a frame is predictable. We know which object is transparent. We know which object would have to be merged the color from the transparent object's color. We just need the SPUs to calculate every pixel a new color and let the GPU process it.

[patsu]

Originally Posted by Titanio

I read it as ~4ms from start to finish when doing it on 5 SPUs.

Yes, I understand this to be the best case scenario (All 5 SPUs start and run at the same time with no/trivial dependency between them). In this case, the RSX would take 9ms or so to complete.

EDIT:

Originally Posted by Weaste

For better or worse, Sony really like maying funny/bizarre machines.

Ha ha, the local store is to sidestep slow global memory access. The split memory pool is a little awkward. Other than those 2 features, heterogeneous computing, more cores, the EIB and NUMA are not uncommon in high performance computing today.

[Trejser]

Christer our Director of Technology tweeted this about MLAA, hope it answers your questions: There's a paper that describes the MLAA algorithm, and the Saboteur effect is probably only a subset of the technique described. AFAIK the version used on GoW3 goes beyond the original paper. The #gow3 AA technique saved 5ms from the GPU, costs ~20ms on 5 SPU's (~4ms Latency), its very pretty and only on #ps3.

http://forums.godofwar.com/t5/God-of.../30061/page/17 (post 5)

[liolio]

It's interesting that the "close to mlaa"/"custom AA/AAA(analytical anti-aliasing) used in the in Metro 2033 is done by xenon which backup earlier Joker454's claims on the matter

[Lucid_Dreamer]

Originally Posted by Trejser

http://forums.godofwar.com/t5/God-of.../30061/page/17 (post 5)

I wonder how far beyond the original MLAA paper did they go in GoW3?

Also, I wonder how close Metro 2033 has come to a subset of the MLAA paper that the Saboteur implementation uses.

[corduroygt]

This is from the Metro 2033 article:
The closest explanation of the technique I can imagine would be that the shader internally doubles the resolution of the picture using pattern/shape detection (similar to morphological AA) and then scales it back to original resolution producing the anti-aliased version.

It's a little hard to understand, but what I took away from this is that it only identifies regions where there could be edges that need AA and then renders those regions at double resolution before resizing.

I guess you could call it selective 2xSSAA.

[Billy Idol]

What interests me: it should be also possible to use the GPU to do this MLAA stuff, right?

A GPU should potentially be faster compared to CELL in number crunching stuff?

If CELL needs 20ms...and they save 5ms GPU time:
How many ms would MLAA cost on RSX in comparison...

In other words...is MLAA in general a faster algorithm compared to MSAA (with even better quality), or is its advantage solely that the CPU can be used to do it, freeing up GPU?!