Question for developers... PS3 and framerate

[joker454]

Weird, bf culling alone should give you more than that on common geometry

As unnaturally high as 60% sounds, 17% also sounds unnaturally low... Do you just have a lot of planar and/or two-sided geometry or something such that BF culling has little effect?

Figured I'd answer both together. Being a baseball game gives us some advantages. There is a freeform replay camera that lets people go anywhere...but it's still bound to the field area. So we have a pretty good idea which areas need no geometry at all. Without getting too detailed, the artists can pre-backface cull a lot of areas with some basic guidelines. That probably brings our spu pre-cull % to less than other typical games, although I still think 60% sounds extremely ambitious.

Joker, I find the official EDGE numbers a little confusing. If you don't mind answering, in your case;
Is it backface culling only?
How many SPUs are dedicated to culling?
What is the input triangle rate?

750K triangles/s/SPU seems a little low for backface culling, even with the local memory hurdle.

In our case we're currently only doing backface culling. For us that was the "money shot" as it were, where we got the best gains (at least according to gcm replay) by far. In theory a good lod system should minimize the penalty from zero area triangles anyways.

We're using spurs, so all of our spu related work is broken down into chunks and fed to the entire bank of spu's to chew on. I don't know what our triangle processing rate is off hand, I could get that info from our PS3 lead but honestly I can't post that kinda data anyways. I do know though that we still have spu power to spare.

If I remember right, the 750k/s/spu number was for Edge doing a whole lot more than just bf culling. It did the whole suite of tests to eliminate unnecessary verts. I just had lunch yesterday with a friend from Naughty Dog, if only you had asked me a bit earlier I could have asked him

 

[nAo]

Edge can cull back facing triangles, out of the viewport triangles, zero area triangles and micro triangles that dont have a zero area but don't cover any pixel (or subsample when multisampling is enabled).
All public info released at GDC..

 

[joker454]

All public info released at GDC..

Public info is one thing. Private info? Ahhh, now that's where it all gets really interesting

 

[Geoson]

seems very promising, the Cell is a very different beast in its own right. Even though it is In-order, it still is very powerful in comparison to some of the OOO processors today.

 

[Mintmaster]

betan, they're saying 800,000 per frame per SPU. That's 48M per second per SPU. It still seems a tad low to me, but we don't really know the conditions for testing.

Geoson, this is exactly the type of task for which OOO execution gives no benefit (on tuned code). Streaming data and lots of tiny, independent tasks.

 

[nAo]

betan, they're saying 800,000 per frame per SPU. That's 48M per second per SPU. It still seems a tad low to me, but we don't really know the conditions for testing..

Checking if a non zero area triangle covers any sample on screen (think about multisampling and rotated sampling grids) is not exactly straightforward

 

[Mintmaster]

Checking if a non zero area triangle covers any sample on screen (think about multisampling and rotated sampling grids) is not exactly straightforward

Nope, doesn't look easy at all, though I suspect one only checks them if the area is really small.

I'm surprised that they even get a win there. As a percentage of total triangles on the screen, I can't see these triangles being particularly significant unless you have completely borked LOD. Let RSX weed them out.

I guess it all depends on how much you value RSX and Cell time. If you're willing to give up 10% of Cell's CPU time to reduce RSX's load by 5%, then maybe this sort of thing is worth it.

 

[betan]

betan, they're saying 800,000 per frame per SPU. That's 48M per second per SPU. It still seems a tad low to me, but we don't really know the conditions for testing.

Yeah, that's what I meant. During the GDC demo they said the difficult part was due to multisampling but I didn't know they were doing 0 pixel visibility test, value of which is arguable for most games (given the cost).

 

[jobboy]

I want to take advantage of the opportunity to talk to heavenly sword devs....i'd like to know why HS turned out suffering from framerate stutters and vsync issues??....is it from GPU weakness (as stated in many forums) or a premature know-how of cell power?

thanks in advance

 

[betan]

I want to take advantage of the opportunity to talk to heavenly sword devs....i'd like to know why HS turned out suffering from framerate stutters and vsync issues??....is it from GPU weakness (as stated in many forums) or a premature know-how of cell power?

cell power eh?
I suggest you read post #9.

 

[Cyan]

Two UT3 maps (PC) will be left out of the PS3 version because of the lack of smoothness:

http://ps3.ign.com/articles/821/821817p1.html

Mark Rein said that they are shooting for a November release and did say that the PS3 version would have "most" of the content that would be found in its PC brother. He said that there would be a couple of maps that would be left out of the PS3 version simply because the system didn't have the resources to produce a smooth gameplay experience. But rest assured, the classic Unreal Tournament modding could still be performed and uploaded to your friends. Rein even told us that there were plans to allow for players to release mods for sale on the Playstation Network and that the most popular mods and maps would be compiled and sold in an updated Unreal Tournament 3 package later down the line. We'll have much more on Unreal Tournament 3 as its holiday release date draws closer.

Perhaps the game is still undergoing fine-tuning in development, or the main RAM is the problem here -very big maps, too many elements on screen I guess- but looks like consoles have been surpassed by PCs....

I love consoles, though, and always will. In fact, without my beloved consoles gaming would be history for me.

p.s. I recently installed Diablo 2 in my computer but, while the game is amazing, there are so many things to do, too many builds, items, etc, that these facts make me feel uncomfortable and I don't see myself playing it anymore.

 

[Neb]

p.s. I recently installed Diablo 2 in my computer but, while the game is amazing, there are so many things to do, too many builds, items, etc, that these facts make me feel uncomfortable and I don't see myself playing it anymore.

I tend to get a bit overwelmed to by games like Diablo and BG series. So much small stuff, spells, combinations and things to keep track of.

 

[novcze]

I suggest you read post #9.

oh well, this is nice, but if game suffer from framerate and vsync issues 90% of all gameplay time is rather unplayable. I know HS looks great, but this doesn`t help)

 

[Betanumerical]

oh well, this is nice, but if game suffer from framerate and vsync issues 90% of all gameplay time is rather unplayable. I know HS looks great, but this doesn`t help)

I wouldn't call HS unplayable, I've had maybe 2 or 3 noticeable frame drops (and I've beaten the game) and I haven't noticed much tearing. I'm interested in what NT are doing on the Cell, or are the graphics all RSX? I guess it'd all be under NDA though.

 

[Shifty Geezer]

Two UT3 maps (PC) will be left out of the PS3 version because of the lack of smoothness:

He said that there would be a couple of maps that would be left out of the PS3 version simply because the system didn't have the resources to produce a smooth gameplay experience.

Does that mean they're cutting out those maps on lower spec (anything less than the very best?) PCs as well?

 

[DeanoC]

I wouldn't call HS unplayable, I've had maybe 2 or 3 noticeable frame drops (and I've beaten the game) and I haven't noticed much tearing. I'm interested in what NT are doing on the Cell, or are the graphics all RSX? I guess it'd all be under NDA though.

In terms of graphics, we use the SPU as a form of object processor. So essentially everything up to and including the production of RSX's command stream probably has a module on SPU to help.
This includes,

1. A module that does a lot of object level clipping and culling both for the view frustum and ths shadow maps. Its job is per frame to calculate how big each shadow map should be in world space and what objects needs rendering in each map.
2. Animations, using ATGs (DeanA team) animation library, every animation is blended and bones updated.
3. Blend shapes, a custom module that handles facial animations
4. Skin matrices, even after animation there some work required to get them into the format used by the GPU vertex shader.
5. Flags, a simple verlet based simulation used for the flags in the game
6. Cloth & Hair, a constrained physics solver used for simple chains that are then rendered as Nariko's cloth and hair
7. Pushbuffer generation. This produces the commands used by RSX to actually render the scene. Has a number of optimisers to reduce redundent state changes.

Probably a few i've missed. Essentially a normal skinned or non-skinned character costs very little PPU time and virtually all processing is done on SPU and RSX. Its this that allows us to render the army scenes for example.

We do no per triangle work on the SPU, we let RSX do that, we however do try and prepare things on the SPU for RSX.

Hope that answers a few things.

 

[jobboy]

oh well, this is nice, but if game suffer from framerate and vsync issues 90% of all gameplay time is rather unplayable. I know HS looks great, but this doesn`t help)

I'm scared framerate performance is the real bottleneck of ps3 architecture instead of the shared ram

 

[patsu]

I wouldn't call HS unplayable, I've had maybe 2 or 3 noticeable frame drops (and I've beaten the game) and I haven't noticed much tearing. I'm interested in what NT are doing on the Cell, or are the graphics all RSX? I guess it'd all be under NDA though.

This is my personal experiences as well. I didn't notice any framerate issue except in a few cutscenes (usually when Nariko is moving from combat A to combat B). During fighting, the animation is smooth. I am at the beginning of Chapter 5 (Died a few times in the massive battle but no other issues as far as I can see).

There is only 1 case in Chapter 3 when Kai's crossing the bridge where it stuttered (I moved her back and forth just to see if it was reproducible. It was).

 

[novcze]

I'm scared framerate performance is the real bottleneck of ps3 architecture instead of the shared ram

no, there is not thing called "framerate performance", only developers need shift focus from shiny efects to rock solid framerate

 

[Cyan]

Does that mean they're cutting out those maps on lower spec (anything less than the very best?) PCs as well?

You have a point, and a very smart one. I'm wondering what the minimum requirements for the PC version are, because every PC with 512 MB of system RAM or less couldn't run those maps.

Actually they where surpassed years ago!

Edited, here is the actual Shifty the Overseer watching over the forum:

Seriously, some paragraphs of this post were a bit crummy.