Possession and the art of PS3 programming

[Kb-smoker]

I spent much of last December writing a preview of PS3 zombie shocker, Possession, for the Official PlayStation 2 magazine. The game is a mouth-watering mix of survival horror adventure and squad-based shooter, with you commanding a ravenous undead army against a city stuffed with civilians, cops and soldiers.

The magazine is now out on the shelves, so I can finally talk about meeting up with Possession developer, Volatile, and chatting about what it's like to write games for Sony's next-gen hardware. At the time, the team were working with an emulator rather than a full PS3 dev kit, but already seemed comfortable with the peculiarities of the multi-processor behemoth.

Of course every developer is going to have a different take on PS3, but this is what one of them had to say...

First of all, as with Killzone developer Guerrilla in a recent online Q&A sesh, they were happy to point out that PS3 is not as complicated to write for as we've all been led to believe. Apparently, the machine's use of Open GL as its graphics API means that anyone who's ever written games for the PC will be intimately familiar with the set-up. In fact, PS3 employs a cut-down version named Open GL ES, which is even simpler - as Volatile's lead PS3 programmer, Lyndon Homewood explained:

"ES is designed for things like set-top boxes and mobile phones, where you want the fundamental graphics but don't need some of the fringe stuff that Open GL has. Because you've got that on PS3, it's going to be much easier than the PS2 to get something up and running - there are hundreds of books out there for it, so you can do your background reading. All the documentation is there."

We also got onto talking about how PS3 will deal with Cg - a version of the programming language C, which allows developers to code for advanced graphics processing units, specifically in the area of 3D shaders. (You can read more about Cg here and here). You may be completely up to spec on how this works, but Lyndon gives a decent beginners guide if not:

"Cg gives you a standard documented API for programming graphics chips. The main two segregations of Cg programming are the vertex shader and pixel shader. With the vertex shader you can act on 3D models at the vertex level, so for each triangle you can do something on each corner and then everything in-between is interpolated. So if you want to make your whole shape bigger, you can just push all the vertices out a bit. In this way you could, say, morph your character into a giant just by scaling up all the verts. It's a lot easier to get to that point in the graphics pipeline.

"And then you've got the pixel shaders. When you render each triangle on screen the GPU asks whether you want to do something to each individual pixel you render... so at this point you could run some sort of mathematical algorithm on each individual pixel - perhaps a lighting effect like high dynamic range lighting (a rough Wikipedia entry on HDR lighting can be found here). And that wasn't possible on PS2.

"All of this is already available and won't be a massive leap from what you're seeing on PCs with high-end graphics cards. But obviously on PS3, you've got eight chips to spread the processing cost over - the main PowerPC chip and seven SPE chips. In a PC, there's just one CPU, two in a dual processor machine. Having an eight CPU multi-processor system in your living room is pretty flash.

"At the end of the day it's just a multi-processor architecture. If you can get something running on eight threads of a PC CPU, you can get it running on eight processors on a PS3 - it's not massively different. There is a small 'gotcha' in there though. The main processor can access all the machine's video memory, but each of the seven SPE chips has access only to its own 256k of onboard memory - so if you have, say, a big mesh to process, it'll be necessary to stream it through a small amount of memory - you'd have to DMA it up to your cell chip and then process a little chunk, then DMA the next chunk, so you won't be able to jump around the memory as easily, which I guess you will be able to do on the Xbox 360.

"The graphics capabilities of PS3 will, I think, be slightly above the absolutely top-end graphics cards on the PC, but you've got much more processing power in the box so you're going to see a lot more physics, a lot more generated geometry. With water ripples, for example - they're pretty much algorithms, you have a flat plane of triangles and you run some sort of mathematical algorithm over it to generate a surface rippling effect - well, you will have the processing power to do these sorts of generated geometry effects On PS3. You could actually put one chip aside just to do that..."

According to Homewood, the management of the SPE chips is going to be a major consideration. One way is to assign specific roles to each of the chips - get one handling physics, one working on AI, etc. This might sound tidy, but Volatile are not convinced - partly because certain gameplay events, like a massive shoot-out, are going to create spikes in demand for, say, animation, collision detection and rag doll physics. One SPU working alone on each of these elements won't be able to deal with such spikes efficiently. Also, there's the question of co-ordination:

"The way we're thinking of doing it ourselves is via a job queue. We'll stick the jobs we want to do into a queue on the main processor and then we'll get the SPEs to pull off a queue entry and process it whenever they're free. You want to make sure all of your processors are always running. If you give the chips specific jobs, you'll end up with a lot of them being idle - you won't get the maximum out of PS3 doing that unless you time everything perfectly, so that the time it takes to do the animation on the first chip is exactly the same amount of time to do the physics on the next chip, which is exactly the same length of time it takes to do all your AI on the next chip - I think that would be extremely problematic."

One last thing. Volatile reckon PS3 is going to be much better for HD cinematics than Xbox 360, thanks to its Blu-Ray storage medium. 20 minutes of HD-TV footage takes up around 4.7GB, so an Xbox 360 game would quickly run out of space. This is going to matter more in the coming years as movies and games merge and we see more film elements being brought across to games.

So there you go. PS3 is relatively easy to program for if you have experience with high-end PCs - certainly more straightforward than PS2 with its proprietary graphics APIs. We've also heard lots of talk - from various studios - that Sony's developer support will be a lot stronger with PS3 (it's already much, much better with PSP apparently). Time will tell.

For now, we have screenshots, demos and promises, we have enthusiastic developers taking their first glance at the next generation of Sony hardware. We don't have games.

http://blogs.guardian.co.uk/games/ar...ogramming.html

[OmniCloud]

Good find son...Looks like the PS3 isn't perfect, but if you put some effort in, you will be able to do some amazing things.

[Sephiroth_VII]

You beat me to it. Good job.

[shagawi]

Intresting read, thanks man.

Volatile reckon PS3 is going to be much better for HD cinematics than Xbox 360, thanks to its Blu-Ray storage medium. 20 minutes of HD-TV footage takes up around 4.7GB, so an Xbox 360 game would quickly run out of space.

Storage capacity would be major problem for MS.

[xbdestroya]

Wow, that article is totally simplifying the real challenges with the PS3 - basically mastering the SPE's. They do touch on the fact, but they go on to end on a happy note saying "if you've programmed for high-end PC's, you've programmed for PS3..."

What could be further from the truth?

Sure there's Cg and OpenGL ES - but I mean of course there is - we're talking about an NVidia chip here. Some believe that's the very reason they went with NVidia in the first place, to avoid forcing a proprietary API on devs to learn. It's not like Microsoft doesn't have that same path of ease via DirectX though.

Not one developer who's complained about the PS3 has complained that the graphics chip was difficult to master; it's all been talk about Cell. Now, I think we're just basically where we've always been - it'll be the PS2 programmers who had experience with the Emotion Engine and the Vector Units that will pick PS3 up the easiest; the PC boys are going to have a harder time.

This gets rushed through in the article, but yet it's the most stand-out portion of it:

According to Homewood, the management of the SPE chips is going to be a major consideration. One way is to assign specific roles to each of the chips - get one handling physics, one working on AI, etc. This might sound tidy, but Volatile are not convinced - partly because certain gameplay events, like a massive shoot-out, are going to create spikes in demand for, say, animation, collision detection and rag doll physics. One SPU working alone on each of these elements won't be able to deal with such spikes efficiently. Also, there's the question of co-ordination:

"The way we're thinking of doing it ourselves is via a job queue. We'll stick the jobs we want to do into a queue on the main processor and then we'll get the SPEs to pull off a queue entry and process it whenever they're free. You want to make sure all of your processors are always running. If you give the chips specific jobs, you'll end up with a lot of them being idle - you won't get the maximum out of PS3 doing that unless you time everything perfectly, so that the time it takes to do the animation on the first chip is exactly the same amount of time to do the physics on the next chip, which is exactly the same length of time it takes to do all your AI on the next chip - I think that would be extremely problematic."

Good find though Kb.

[gablar16]

From the article:

The graphics capabilities of PS3 will, I think, be slightly above the absolutely top-end graphics cards on the PC, but you've got much more processing power in the box so you're going to see a lot more physics, a lot more generated geometry.

What is the "absolutely top-end graphics cards on the PC"?The 7800GTX?
I'm I reading too much into it or this affirms the RSX being derivative of the G70?

[Raijin]

Originally Posted by gablar16

What is the "absolutely top-end graphics cards on the PC"?The 7800GTX?
I'm I reading too much into it or this affirms the RSX being derivative of the G70?

It seems so... I think it would be logical to me, at least!

[xbdestroya]

I don't think it afirms it as a G70 derivative per se, could also refer to a G71 or G80 derivative. Personally I'm on the G71 bandwagon, but the point is 'better' than the fastest card at the time of PS3's launch. It will be quickly superceded I imagine after that.

[Infernal]

Originally Posted by gablar16

What is the "absolutely top-end graphics cards on the PC"?The 7800GTX?
I'm I reading too much into it or this affirms the RSX being derivative of the G70?

The top end graphics card is the X1900 right now with 48 pixel shaders and 8 vertex shaders, thats a hell of a lot better than the 360 card (which they said wouldn't be surpassed in years) and I wouldnt complain if the RSX was as good as or better than the X1900.

[Domination]

Originally Posted by Kb-smoker

http://blogs.guardian.co.uk/games/ar...ogramming.html

That was a good read.

So, to fully take advantage of the Cell, you're probably going to need a team like those from Sony's elite first party or maybe even Kojima. Sounds like great news overall. Great find, Smoker.

I would give you two reps for that, but it's not allowed.

[Domination]

Originally Posted by gablar16

What is the "absolutely top-end graphics cards on the PC"?The 7800GTX?
I'm I reading too much into it or this affirms the RSX being derivative of the G70?

I read it the very same way when i first saw it. But it's not what you think. Here, I'll repost it and bold what you should have paid more attention to:

"The graphics capabilities of PS3 will, I think, be slightly above the absolutely top-end graphics cards on the PC, but you've got much more processing power in the box so you're going to see a lot more physics, a lot more generated geometry. With water ripples, for example - they're pretty much algorithms, you have a flat plane of triangles and you run some sort of mathematical algorithm over it to generate a surface rippling effect - well, you will have the processing power to do these sorts of generated geometry effects On PS3. You could actually put one chip aside just to do that..."

That's how I comprehended it after the second time. So, in other words, when the PS3 comes out, he believes the RSX will be greater than the highest GPU card in a PC at that time.

[Z]

Looks like the PS3 isn't perfect

hey Omni, let me tell you; of I aint perfect, heck there isn't anything else on this planets that is! take it from me.

The graphics capabilities of PS3 will, I think, be slightly above the absolutely top-end graphics cards on the PC, but you've got much more processing power in the box so you're going to see a lot more physics, a lot more generated geometry.

this is what I am talking about! I am more excited about Cell than I am about RSX. Heavenly Sword, MotorStorm and F1 gives me a glimps on what physics can be like. and don't get me on the Duck demo when they took a glass and 'scooped up' some water. that is one of the most amazing things I have seen!

20 minutes of HD-TV footage takes up around 4.7GB

ouch! that is the advantage of BD. so much for X2 being for the *echo effect* "HD Eraaaaa".

[CreativeWriter]

Possession doesn't look like a great game to me from the previews, but this is interesting to read nonetheless.

About the duck demo. Did an appropriate amount of water volume disappear from the tank when the cups were used to scoop some up, or did the volume stay the same? Volumetric water and displacement would be really impressive, but from what I saw it looked like the water didn't lose volume based on how much was scooped out (which is less impressive, while still being neato, since the water in the cups seemed to be volumetric and you could pour out some of it at a time).

[CrumCon]

I always though that all SPE's can access the main memory, at least the XDR one.
Why didnt they design the SPE's to access the main memory? sounds strange.

but looking at the bright side. with 7 SPE's available Anytime.. there shouldnt be queue problem at all. Even with tons of jobs to do.. with 7 SPE's its more then enough.

anyway i think only high-technical games that require so many SPE's (each character with very smart AI). i dont even see any low-tech games would use so many spe's.
l

[lip2lip]

There is a small 'gotcha' in there though. The main processor can access all the machine's video memory, but each of the seven SPE chips has access only to its own 256k of onboard memory - so if you have, say, a big mesh to process, it'll be necessary to stream it through a small amount of memory - you'd have to DMA it up to your cell chip and then process a little chunk, then DMA the next chunk, so you won't be able to jump around the memory as easily, which I guess you will be able to do on the Xbox 360.

is this a round about way to say mem controller at gpu is more efficent for meshes?

no sale with me. I hope any one can see this is a silly notion.

[VG Aficionado]

I would say that, in terms of graphics and physics:

RSX + Cell >> G70 + Latest PC CPU

RSX + Cell > X1900 or G71 + Latest PC CPU

RSX + Cell ...?... G80 + Latest PC CPU


This is taking into account that G71 is released in March and that PS3 launches a bit later during Spring. RSX + Cell could indeed be more powerful (albeit slightly(?)) than the most powerful and expensive PC components by the time of its launch, as promised by Nvidia. But then, a few months later, G80 is released and who knows what will happen...


But then, in terms of price for end user:

Any recent PC GPU + Any recent PC CPU >> PS3

The choice should be a no-brainer for most people who want something powerful and relatively cheap.

[woundingchaney]

Originally Posted by Infernal

The top end graphics card is the X1900 right now with 48 pixel shaders and 8 vertex shaders, thats a hell of a lot better than the 360 card (which they said wouldn't be surpassed in years) and I wouldnt complain if the RSX was as good as or better than the X1900.

It is relatively hard to compare the two primarily due many of Xenos console features and the concept of US. Most of what I am reading leads me to believe that at 720p Xenos wins hands down (primarly because it seems Xenos isnt geared toward higher end resolutions). Where as you may have higher theoretical limits on the x1900 you get better results with the Xenos.

Simply comparing 48 to 48 isnt the story.

[cliffbo]

lets just put technology to one side for a moment and consider the plain and logical truth of any company involved in producing consoles. Yes they will all try and put out technology that beats the opposision (on paper that is) but none of them will want any developers to be able to utilise the full potential of any given console straight out of the blocks. so some inate difficulty will be deliberately built in by the manufacturer to enable a gradual familiarisation to take place. so this initial carping by programmers is going to be inevitable (even on PS5) so get used to it. its good news for publishers such as EA who need a slow incramentation in quality to justify knocking out yet another follow up (but with better graphic, better AI, better physics... bla..bla) i also believe that once the narrow arguments of last generation machines stop being aplied to new generation, the real debate will begin: true AI. so yes PS3 will be difficult to program for! who the hell cares! i'm expecting good things from first gen games, great things from second and fantastic things from third.

[section]

Originally Posted by CrumCon

I always though that all SPE's can access the main memory, at least the XDR one.
Why didnt they design the SPE's to access the main memory? sounds strange.

Answer 1) They can access main memory, although not in a straight manner but via DMA accesses

Answer 2) This design decision was introduced to crush the 'memory wall':


from IBMs 'Introduction to the Cell multiprocessor':

The objectives for the new processor were the following:

-Outstanding performance, especially on game/multimedia applications.
-Real-time responsiveness to the user and the network.
-Applicability to a wide range of platforms.
-Support for introduction in 2005.

The first of these objectives, outstanding performance, especially on game/multimedia applications, was expected to be challenged by limits on performance imposed by memory latency and bandwidth, power (even more than chip size), and diminishing returns from increased processor frequencies achieved by reducing the amount of work per cycle while increasing pipeline depth.

The first major barrier to performance is increased memory latency as measured in cycles, and latency-induced limits on memory bandwidth. Also known as the “memory wall”, the problem is that higher processor frequencies are not met by decreased dynamic random access memory (DRAM) latencies; hence, the effective DRAM latency increases with every generation. ...

The challenge therefore is to find a processor organization that allows for more memory bandwidth to be used effectively by allowing more memory transactions to be in flight simultaneously.

Synergistic processing element

... The SPE operates on a local store memory (256 KB) that stores instructions and data. Data and instructions are transferred between this local memory and system memory by asynchronous coherent DMA commands, executed by the memory flow control unit included in each SPE. Each SPE supports up to 16 outstanding DMA commands. ...

...The DMA unit can be programmed in one of three ways: 1) with instructions on the SPE that insert DMA commands in the queues; 2) by preparing (scatter-gather) lists of commands in the local store and issuing a single “DMA list” of commands; or 3) by inserting commands in the DMA queue from another processor in the system (with the appropriate privilege) by using store or DMA-write commands. For programming convenience, and to allow local-store-to-local-store DMA transactions, the local store is mapped into the memory map of the processor, but this memory (if cached) is not coherent in the system.

The local store organization introduces another level of memory hierarchy beyond the registers that provide local storage of data in most processor architectures. This is to provide a mechanism to combat the “memory wall,” since it allows for a large number of memory transactions to be in flight simultaneously without requiring the deep speculation that drives high degrees of inefficiency on other processors. With main memory latency approaching a thousand cycles, the few cycles it takes to set up a DMA command becomes acceptable overhead to access main memory. Obviously, this organization of the processor can provide good support for streaming, but because the local store is large enough to store more than a simple streaming kernel, a wide variety of programming models can be supported, as discussed later.

Not many processor achitectures allow over 128 memory simultaneous memory accesses So SPE's have local store to tackle the latency of going through loads of coherent system memory and every one of them have specific MFCs (memory flow controller) and DMA engines within them are able to push and pull rather adequate amount of data in and out.

All in all operating with the data must be done on SPE level but transferring the data should be lightning fast both at PPE->SPE(s) and SPE->SPE transfer levels because of the asynchronous DMA transfers.

Yeah and also, you can't do SPE->SPE -transfers with public Cell BE emulator that this 'gotcha' part of Volatile software engineer is partly referring to, it's only possible with the hardware version of the Cell processor.

More Cell DMA engine technobabble here for those interested ones :

www-128.ibm.com

/developerworks/power/library/pa-celldmas/

[RavenFox]

great read fellas.