Ratchet and Clank not doing LOD on SPUs *

Do you have a link to the original interview?

 

http://threespeech.com/blog/?p=422#more-422

 

Thread's title completly off.They just say they don't use the Spu to do LOD.

 

Absolutely. Cell could be doing culling and whatnot still. And if you found with the graphics targets you've set, you don't need hard LOD as RSX can render it, you wouldn't wate time developing an LOD to eek extra performance from RSX (unless you set new targets).

 

Sliding Window...

 

From R&C thread:

Obviously using SPU's for gfx is associated with a certain cost (hardware resources, human resources, money, time, loss of flexibility ...).

You don't really need an answer from the developers. If someone is not using SPUs it is always because the gain doesn't justify the cost. And majority of the developers seem to be in this boat. Maybe in time gain/cost will go up with better understanding and tool chains but for now how is this jumping the gun?

For the record, there is already an EDGE thread suitable for this discussion.

 

Hahahahaha!

Who pulled that one out their ass and gave it to you?

 

You certainly jumped the gun with your assumption that most developers are in that boat of which you speak. I'd say that the majority of developers haven't even gotten to the point where they've assessed which boat they're in on that topic. Most of the known techniques (among others) are in evaluation at several studios, EDGE itself is still closed beta I think, and the people who are actively working with all this are primarily working on titles we haven't actually seen (or at least seen much of) yet.

That said, they only really speak of heavy vertex processing and complex LOD systems in software, which they say they're just generally not doing, SPUs or otherwise. They're not that specific. Something like backface culling or software skinning is a lot simpler than progressive meshes and edge contraction or other such things to the point that you probably wouldn't refer to it as "heavy" vertex processing.

 

Going back to your original post...

You suggest SPU aid to RSX isn't practical, but the real decision is, as you say above, do the benefits outweigh the gains? That depends entirely on what they're trying to achieve. ie. The choice may be 'do we get more graphics on screen than we have already, or do we use the SPEs to generate better animation for the characters we have?' If they're happy with the graphic look, they may choose the latter. In looking at the game myself, I don't see they've chosen badly! It's not like the graphics look sparse and the game could do with more content!

This choice wasn't, according the dev, whether they could manage it or not. The interview doesn't read 'We wanted to use the SPE's to improve LOD but they weren't very practical, so decided to spend them on of gaming features instead.' And we also don't know what other things they may be using SPE's for to aid RSX. All we have is "We made the decision to steer away from doing heavy amounts of geometry processing on the SPUs, so instead we’ve focused on techniques that can be implemented directly on the graphics chip." We don't have a reason why they chose not to implement geometry processing on SPUs.

So in summary, where you've jumped the gun is both equating not using SPE's on LOD to not using SPEs to aid RSX at all, and in equating the decision not to use SPE's to help RSX as it being too difficult or inefficient, rather than a choice based on other criteria. That may be what happened, but that's quite an assumption without some stronger reasons IMO. And I also question why you think the majority of developers are unable to use SPE's to aid RSX? The majority of developers haven't commented, and those that have give varying stories, from what little I've read. In the main we don't know what SPE's are being used for, or even if they are being used at all!

 

I see no difference, but to be clear they don't seem to be in the SPU gfx processing boat yet. "The decision is yet to be made" part is vague for me. They are not using it yet, they may be using in the future, that much we know (it seems) and I already said. Maybe they made an active decision, I wouldn't know.

Closed still? it was supposed to be out in March. But you think it is closed to first/second party developers as well? I somehow find this very unlikely.

First here is a question regarding heaviness of occlusion culling: How much processing power do you think it takes to do simple backface culling or skinning? Or relatedly what additional complex operations do you think EDGE libraries are doing at 750k*60 triangles/sec/SPU?

Second, most of the people on this thread seem to think that they may be doing "light" vertex processing which is fine, but even if this is the case why would it require SPU at all? The question really is the practical value of SPUs after all.

The guy already says it is inflexible. EDGE animation presentation mentions their complex animation libraries are very very fast. Plus from what we have seen R&C animation system doesn't look complex at all. Nor it has any significant physics or AI. So it is possible they may need SPUs for something we don't know, I see no evidence of heavy SPU usage. Whatever the case SPU usage doesn't have to be constant through out the game. For example, flying through the city may require less adaptive animation or physics but use SPU gfx help.

Let's see what we know: Resistance has significant amount of popins and it is pretty safe to assume its engine is not more complex than R&C's. Genji didn't use SPUs at all. Motorstorm only had basic object occlusion detection which may not even be in _the_ SPU it is using. Heavenly Sword is not doing any vertex processing per triangle level. Warhawk has old school LOD system as far as I can tell (at least for objects and buildings).

No need to talk about third parties I assume.

Again, up until now, it doesn't look practical to me, but the "yet" part is very important and hopefully will change in the future.

 

No, he said he likes the flexibility to the engine that LOD on RSX (and other things on Cell) provides. Which may be flexibility to the game engine rather than graphics engine. By that I think he prefers to have all SPE's available to do whatever game functions he wants, rather than have some dedicated to graphics work and not be moveable between stuff like physics, AI, or what-have-you.

As for your examples of devs, you're talking about launch titles in the main! I'm also not sure what you're talking about regards use of SPEs and graphics. Are you talking about only LOD systems, or any graphics work at all? I think there's lots more graphics stuff SPE's could be doing that wouldn't get a mention. I dunno. I haven't got a nice log of developer quotes to reference! Maybe you can provide suitable links where devs talk about how they're not using SPEs?

 

I cannot provide links as I am tired. I can provide pointers though.

Genji's SPU usage should be somewhere in this forum. Apparently Sony wanted them to use SPUs but they didn't (for anything). Similarly Motorstorm's is from B3d Interview. There was an article regarding Resistance in IEEE Spectrum just before the lunch (popins are personal observation obviously). The Newsweek game blogger has an F1 interview though I'm not 100% sure they are not doing heavy SPU gfx (but I am fairly certain). Regarding Heavenly Sword, nAo mentioned vertex processing on EDGE thread or something similar.

Overall I am fairly certain there is no single game on this list (including R&C) doing any computationally complex gfx stuff. Warhawk may be doing something but it certainly doesn't show up on its LOD system (again personal observation).

As you said there are other stuffs SPUs can help with, but excluding Lair there is no single game publicly disclosed that information. With all those CELL praising and marketing, I find this interesting.

 

So what? We're a mere 6 months into te lifecycle, the Xbox360 is 18months old and we're JUST starting to see games that really take advantage of Xenos.

Give it some time, we're still on the first generation of titles, games that began development 1-2 years ago.

If you look at comments by Ninja Theory, or Insomniac in regards to the potential of the SPU's, they are extremely positive. DeanoC referred to them as one of the best innovations in gaming in the last 2 or 3 generations, now that's a fairly powerful statement imo.

 

Well, Insomniac is one case, and I don't know what the specifics of their story would be, but they've probably made their decision for the case of R&C, anyway. Your prior post made it sound as if you believed that the majority of developers had already made the same decision as well, which isn't the case.

Well, I wasn't talking about first/second party developers specifically. I think they have open access, but I was speaking about PS3/multiplatform developers in general. And the last update I heard was about 2 weeks ago, and it was still said that it was closed beta. Maybe that's changed as of 1.8 SDK, but I can't say for sure. We've only just gotten to the point where we're testing out EDGE in our render pipe and seeing what it does, which is why the PS3 builds at the moment are very not-runnable.

Backface culling is nothing. It's a dot product to decide whether a triangle is visible or not. Skinning is a few times heavier than that, but it's still one of those things that SPUs are good at. Additionally, skinning on the SPUs means sending less complex verts (and in turn vertex shaders) over to the GPU, which is always a good thing.

LOD for instance, is more complicated mainly because of the larger scope at which it operates, so it may not necessarily be heavy on workload, but logically heavy and hence trouble-prone. Skinning is a single-vertex-level operation... Backface culling is a single-triangle-level operation... making both easy to split up and trivial to make into independent SPU jobs. And the point is to save on what you send down to the GPU.

I think you're approaching this from the wrong angle. It isn't so much that SPUs are *required*, but that they're better at it than anything else in the system. It's not that you couldn't do without any SPU culling, it's that you have to somehow lighten the vertex throughput requirements in other ways, since RSX is otherwise limited. And yeah, you could do it on the PPU if you really wanted, but even a single SPU will do it significantly faster and because simple things like skinning are easy to turn into several independent jobs, it's more of a "why wouldn't you?" question.

Of course, if you have games where RSX as is is good enough for what you're sending down the pipe and it's not getting in the way of whatever else you might be doing on the software side, then there isn't much point even if it does improve performance for you. This is of course, assuming that no more spanners are going to be thrown into the works.

Besides which, while it goes outside the scope of this thread, if you really wanted to know about the value of the SPUs, then you have to step outside of vertex processing alone (audio, physics, etc). Which is partly what Insomniac was talking about when they spoke of having the choice of keeping SPUs available.

 

Who talks about the value of SPU? The question is its current practical value in terms of gfx only.
For the gfx, I see hope but no convincing evidence. There is time indeed and hopefully things will change.

Since I am tired of similar responses let me make it clear. I am not anti-Sony nor disappointed with my CELL/PS3. On the contrary I am particularly happy with the machine, but somethings are taken for granted here. I see comments like "Motorstorm uses only one SPU, imagine how Motorstorm 2 will look like". People are convinced 6 SPUs will make things magically better. We can all list couple of things they can potentially help, but the list of real life examples is very short. That doesn't mean it will stay short, it is just short for now. Why are you people going defensive about this?

 

hmm... for the record, I remember that Lair is actively using progressive mesh as "practical value" in terms of gfx...

I don't think this is anything that every developer is quite familiar with doing heavy geometry on the Cell so I'd say it's fair that they are still exploring the practical value of using the SPEs to aid in gfx. So saying that

is really jumping to a conclusion.... IMHO
It might not justify the cost "in their scenario" but it doesn't mean that the majority of the developers are in the same boat unless more evidence is provided.

The interview says

It does say that not doing heavy geometry processing on the SPU gives more flexibility, but nowhere does it say that it is inflexible....



Pardon my language

Consuming viagra gives you more ability in certain aspects but not consuming viagra means that you have an inability in certain aspects.
Sounds rediculous, aint it?

lets change
"Consuming viagra" to "Not doing heavy geometry processing"
"flexibility" to "ability"
"in certain aspects" to "in our engine"

Not doing heavy geometry processing gives you more flexibility in our engine but [strike]not not[/strike] doing heavy geometry means that you have an inflexibility in our engine.
Still a rediculous statement IMO

 

For what it's worth, a beta release of Edge has been available to all PS3 developers (including 3rd parties) for the last few weeks. Indeed, I already know some guys at multiplatform studios who are looking at it.. some who are directly looking at the code, others who are looking at it as a thorough reference of how to build their own equivalent systems.

Cheers,
Dean

 

You are truly smoking some bad stuff…

R:FOM as far as I can tell doesn’t have an inch of popins or draw distance problems (with nice Multi-sampling to boot) on my TV(s).

 

Ok just to give you some reference point of view here - a 300MHZ VU unit can do BF culling at a rate of ~500k*60 triangles/sec.