CELL from GDC
- Thread starter one
- Start date
Well no - forcing this type of data rearranging effectively eliminates any real chance of writting readable code, especially if it's in C++.nAo said:
And I don't know how much you've dealt with writting non-graphics code for vector units, but this kind of limitations will make it insanely difficult to write efficient physics code. Heck you even run into issues with VUs with a far more capable ISA.
Well, for C/C++ code, it's not like the compiler could rotate vector memory patterns for me, so what code I want to write in straight C will be all manual labour to optimize anyhow. Though I at least expect/hope for decent loop optimizers.
And I suppose Cg could probably help making more readable code possible, though I sinceredly doubt efficiency if you start mixing and matching many single vector operations with matrix transforms and expect it to be handled transparently without doing the above mentioned manual rearrangements.
Well yeah, but that was when I was 16.Guden Oden said:
Thing is people whine about managing multiple processors and this and that, but according to this, writting fast code for just 1 SPE will be obfuscated mess that would make even raw VU assembly code look prettier in some ways.
I share some of your fears, but if they give us a good compiler and I'll just start to use matrices of 4 vectors instead of a single vectors in my loops.Fafalada said:
You're right about potentially fucking up code readibility but if you want huge flops figures (we're talking about being 2/3x powerful than their direct competing company hw..) you have to sacrifice things. Nothing is free..
..my non graphics code library on the VUs is almost nil, nada, zero.. so you're right on this too
Umh..imho it's just a matter to change your way to look how data are stored.
I'm arleady working that way on the PS2. Too bad I realized too late in dev process I should have taken that route from the start, so I had some headache when data stored in different ways have to work together someway..(see my comments about lighting code I wrote here some weeks ago..)
We don't know SPEs ISA so we can't know if they are supporting this kind of rearrangements trough special instructions, let's hope that though.
Faf..life is wonderful
I must say I understood maybe two or three lines from those presentations, so I won't comment if those things are good for the programmability of cell or not.
Did Microsoft give similar info of the insides of their xbox next? Or are they relying XNA solves everything?
That was a game developer's conference after all, right? Not just some pr session for Live end user features
Did Microsoft give similar info of the insides of their xbox next? Or are they relying XNA solves everything?
That was a game developer's conference after all, right? Not just some pr session for Live end user features
AFAIK they are going to lecture on the usage of OpenMP, so it looks like just another SMP configuration where no such special briefing is required except for the obvious focus on multithread programming.rabidrabbit said:Did Microsoft give similar info of the insides of their xbox next? Or are they relying XNA solves everything?
That was a game developer's conference after all, right? Not just some pr session for Live end user features
Actually that's what I was just wondering about - why bother even giving us vector instructions, just stick a 4x4Matrix multiplier in there and be done with it, it'll be about as usable anyhow. :?nAo said:
Question is, is the performance worth it if ISA might limit you to something far lower in everything but trivial code? If all they wanted was something that can transform vertices fast I am sure they could have gotten comparable performance cheaper.
Well, it still forces using you to use bizarro classes such as vector4x4. Basically it boils down to obfuscating your algorithms into quadruples or whatever the larger atomic unit you will define.
And I doubt anyone would prototype any new code in that manner, so we'll be back to writting nonoptimized and optimized versions of code - feels like 1999 all over again.
I actually wanted to ask, what prompted you to do this eventually? I know writting all code unrolledx4 can help with VCL optimizer among other things, but I doubt that would be the only reason. And I don't suppose you store anything but the normals like that?
Personally I've considered the idea early on but rejected it eventually. I still have all my loops written unrolledx2, but that's cuz of the way UVs are stored (saving space and memory accesses).
Megadrive1988
Veteran
I have a stupid, stupid dumb idiotic question
are these 'Jobs' and 'Job Queue' anything like Cell apulets ?
are these 'Jobs' and 'Job Queue' anything like Cell apulets ?
one said:This is reminiscent of BlueGene/L, as its single compute node has a mini-kernel (CNK) running on it.
Each SPE runs a mini- kernal
As I predicted they would run some sort of nano-kernel ala TAOS,
TAOS: http://www.byte.com/art/9407/sec6/art1.htm
Jaws said:
http://www.beyond3d.com/forum/viewtopic.php?p=438987#438987
Hannibal in the article below mentions that SPEs can have various scaler and vector configurations,
http://arstechnica.com/news.ars/post/20050308-4685.html)
Hmm...So we're still on for a CELL Virtual Machine, i.e. CELL VM ISA?
Panajev2001a
Veteran
Fafalada said:Well there was nothing really conclusive said in that discussion thoughnAo said:
Anyway what do they expect me to do with no broadcast and no dotproduct, replicating scalar value into entire vector for every dotproduct during matrix and vector transforms?
Don't even get me started on compiler trying to work in such ways, it took GCC like a decade just to support madds.
GCC == best compiler in the world... there is ntohign better than GCC 2.95.x noooo sir.
That's true..but we still don't know if there is any other kind of support for this stuff in the ISA. a 4x4 matrix multiplier would have been even a bigger a waste with scalar opsFafalada said:
I don't know where the sweet spots is.
Do we like challenges? Hell..yes! [masochist mode off]
100% agreed. But it's clear SPEs will do much more than transforming vertices. Maybe the GPU is going to do most of the vertex shading stuff, who knows :?
yeah..it like an obfuscating C++ code compo
(I don't know if I should cry or if I should laugh..)The idea prompted when I wrote for the first time some VU code for a classic point light. I realized all the renormalization stuff would have taken tons of clock cycles, then I re-discovered (it happened when I workd about doing per triangle mip mapping and I thought about a fast log2 implementantion on the VUs) the ftoi/itof trick (doing calculation in logartmic space..).
That trick takes 4 cycles and work on a single vector, so I could renormalize 4 vectors at the same time (one pow per clock cycle, that' cool
)!I needed to have data in that way..and the better way to do it, imho, needed to work on 4 vertices a the same time and transpose all my matrices. Moreover this accelerated all my dot3 calculations (0.75 cycles per dot3).
I use lower pipeline to transpose non-transposed stuff (veritces position) and it's free cause I had a lot of free slots. I could have used it to just compose a vector full of scalar to be invsquared but VCL worked better with the transposed configuration, dunno why.
In the end it worked and it's faster than my non transposed implementation, even if I've no doubt someone could come up with something even faster.
ciao,
Marco
Megadrive1988
Veteran
up and down, up and down, up and down. sigh.
it seems like, with every new Cell thread, things shift from good to bad, back to good, then back to bad again. ugh, at least as far as the current thoughts of various people who actually work on games, programming, graphics, etc.
Is Cell looking good or bad at this moment in time ?
it seems like, with every new Cell thread, things shift from good to bad, back to good, then back to bad again. ugh, at least as far as the current thoughts of various people who actually work on games, programming, graphics, etc.
Is Cell looking good or bad at this moment in time ?
Well i think it's easy for the gamer to say "YEAH I LOVE IT!", but they're not the ones having to code the bloody thing...
Oh well...
Oh well...
Jaws said:They're just moaning without knowing all the facts. I beleive Faf has 294.912 moan threads con-currently running in his head.
We don't know the CELL ISA yet...and coding to a consistent CELL VM ISA would make life easier...
Well personally i don't care, i'm gonna buy ZOE4 and go WWWWWWWWOOOOOOOOOOOWWWWWWWWWWW.... Who cares if the devs died trying to make the PS3 output 19032trillion particles all moving realistically. One by one. By hand.
