Blender is not middleware, blender is an application, ready to give results, and Blender on AltiVec (and even more on Cell) is a sure winner, there is a lot of demand for Blender artists/modellers, and there are commercial movies that use Blender. Of very high quality too.



Optimizing OpenGL will be nice, but I don't know what else effect it will have, it will still not benefit the games industry for PowerPC not even the rendering markets. In any case, I'll put it in my list, which so far includes Blender, Ogre and perhaps later ODE and now OpenGL. I also liked the idea of a complete AltiVec fork for MySQL, but probably later and if there is enough interest.

Konstantinos

Don't get me wrong.

I love MySQL and I would love an optimized MySQL for Power.
But speaking with my experience as former MySQL employee,
you need to get some commitment of MySQL that they will include (merge back) your work into the main MySQL tree.

Without such a commitment putting any work into Altivec or PPC optimization is lost time!

The option to fork MySQL does only make sense if you are able to keep the fork up2date.
If you create your Altivec fork today and tomorrow MySQL brings out a new release with new storage engines etc - you will have to have the time and man power to keep your fork update or its useless immediatly.
The MySQL source is in development a lot.

I think forking MySQL for Altivec is such a big task as creating a Altivec Linux kernel fork.


If you are interested in doing this work on optimizing MySQL then I would propose you talk to MySQL first to get an commitment. Otherwise it can easely be that your patches will stay in their pipe for some years.

Getting a patch into MySQL can sometimes be more difficult than getting a patch into Linux.

Just my 2 cents
Gunnar

the poll still works here, just voted (xorg)

Well I was thinking of OGRE and the other game-engine and physics-framework kind of things, but Blender is technically middleware - it sits between your model and the render.

However if you are not rendering models, it doesn't give you a great deal of benefit.



Most movie studios use Maya.. we are not talking about a high value market here for people that use Blender to create movies and want to pay for PowerPC render farms (WETA is in league with AMD, for example).

Just who are you targeting that would really appreciate the ability to be locked into a specific modelling application, and a specific rendering technology with a specific processor?



Considering ALL games and ALL rendering engines on Linux and PowerPC require OpenGL, I would say it would benefit them all.

Blender also runs entirely inside an OpenGL context for the user interface. If you speed up OpenGL, then Blender will get faster to use (it may not raytrace faster, but general operation will be improved).



I do think X, OpenSSL, MySQL and OpenGL are all much better and much more broad targets and focussing on very, very specific user needs is not going to bring in a great deal of business (for a start, they are far too specific, but also, there aren't any specific things that PowerPC and AltiVec will do better).

In that view, a compiler is middleware, bash is middleware, everything is middleware (it takes input and produces some output and it sits inbetween :-)



If you're not rendering what do you need Blender at all? :-)



Of course, I was not thinking that WETA for example will drop AMD in favour of powerpc, but I was after small rendering studios, like the guys that did:

http://www.plumiferos.com

or

http://www.elephantsdream.org/

Smaller studios that produce 3D commercials for example, not bound to marketing decisions, might appreciate more the performance gain one would get using a couple of PS3 (since no other cheap PowerPC system exists at this moment).



True



True again, but partly. If I choose Blender I will not go for optimizing the modeller but the renderer. That's the one taking eons to calculate, not the modeller. The renderer can be used standalone in a renderfarm -which is the usual way in commercial setups anyway, with any similar software.



In another post, you mentioned that:



In that case and with this strategy, I don't see why you insist so much in optimizing X, it will benefit very few, and in so minor ways it will not even be noticable (do you think everyone runs compiz-fusion on powerpc?). OpenGL yes, probably, OpenSSL again probably yes, and with MySQL, it will probably mean that PowerPC/POWER-based servers with AltiVec will gain a significant advantage over other arches. But that will have to wait for someone (IBM? Freescale) to ask for it and fund it, as it requires for much more support than we will be able to give over the next 8 months of this "experiment".

Konstantinos

Okay :)




Okay :D



On the other hand optimizing the generic underlying subsystems - Render is heavily used as is OpenGL - benefits any graphics adapter you want to plug into the system that doesn't have it's own onboard acceleration. The number of adapters which actually support full EXA and Render acceleration are actually rather few, so the software fallbacks are important.



No, but that is because the high end graphics card drivers either suck or do not exist. OpenGL is totally unoptimized for PowerPC (something which Apple spent a great deal of time on for OS X) and many basic system libraries do the worst actions possible when running on a PowerPC (and SPARC and MIPS and ARM..) because "it works okay on my AMD64" or simply trust the compiler (gcc is notoriously bad for code generation even on x86 platforms).



I think building a library of vectorisable routines would come in far more handy than vectorising applications themselves. Apple, MacSTL (Intel, even) have focussed on providing low-level, system-level support and callable libraries first rather than going straight for the application. Once the libraries are sound then the compiler is always the next target in order to get code generation up to scratch, before finally painstakingly optimizing single applications.

From a certain point of view, any application which uses GNU libc, OpenSSL (or Netscape SSL), uses OpenGL for matrix calculations, perhaps even uses OpenGL shaders (there should always be a software fallback :) demands more attention for the total system to be responsive and efficient, than taking one application usage case and profiling it for problems.

Most new gfx cards provide full accelaration so the altivec optimizations for XRender will probably be useless in such a case. This is in case a new system (quad-g5 from IBM?) comes along.



Hm, well, the compiler autovectorization covers only a few cases. If it could do better then all the weight would fall on the gcc auto-vectorization effort. This has been going on for what? 3-4 years now? Have you seen any major speed improvements? I haven't. Sure there are some cases where it helps, but not in general and it sure as hell doesn't vectorize the algorithm. It replaces a known-to-be-vectorizable loop with the appropriate code. That's definately NOT the same. Personally, I don't have high hopes for auto-vectorization in the compiler, I trust much more the human vectorization and the right toolkits in case of Cell SPEs (like Xcpu for example).



Exactly. Optimizing STL is far harder than optimizing Ogre. While the benefit might be bigger, I run the risk of running out of time. 8 months is not that long a period and there's always the risk of not finishing the task. Don't misunderstand me, I would *really* like to optimize EVERYTHING in Linux for AltiVec and Cell, but I just cannot afford to do it all for free. I will find a few successfull packages and I will work on them. After that, I will approach companies that would like their software optimized. I will also try IBM/Freescale/whoever to suggest a more complete Linux optimization, but my chances is that they will not easily opt for something like that. Why? I don't know, perhaps because they don't believe enough in their own product, or maybe because they have enough developers inhouse that can do this stuff anyway (if so why haven't they done it already? Dunno...).

Anyway, whatever the case, I'll consider everything said here very seriously

BTW, I also asked Sergei Larin from Freescale about this and asked his opinion. He advised me to go for the 3D stuff, that sells, not SQL (it's x86 domain).

Konstantinos

I started approaching syoyo since lucille render and muda are quite interesting for the point of view of 3d studios/farms

more people involved would be quite nice, I also approached luxrender people but they told me that they aren't interested on altivec nor cell right now.

Whats about RapidMind? What a bummer that RapidMind don't support PowerPC

hm, from what I see on their website they already support Cell. It's just that they don't support powerpc as host platform.

What intrigues me most is that IBM partners with them and could easily provide this support in cooperation with the company but they don't. If IBM doesn't do it, how could we? Not to mention that we're just a 3-person startup (of which only 2 will work on altivec/cell anyway) and too small a fish for these supercomputing guys :-/
Maybe later :)

Konstantinos

Ok, the poll is over, so I guess we'll have to decide soon. The results are visible in the top of the page, and they will certainly affect our decision, though they are not the only factor. Strangely enough, Blender which has marked low, is already decided to be on the list of selected software (along with libfreevec which I'm working on anyway). We'll have 1 or 2 more projects to decide upon (depending on complexity).

Stay tuned...

Konstantinos

Hey Markos, whats about OpenCV:
http://www.ppcnux.de/?q=node/7161 ;-)

what about it? they're already doing it :-)

Hi Folks, please have a look:

MPC8610: Integrated Host Processor

Let us know what you think...;)

R&B :)

We had a significant discussion about this thread today with a *new* -- but only a *potential* -- CELL licensee/buyer.

Any more thoughts on any of this? Please post.

We might end up building a CELL reference design?! :D

We would have never thought that... Thanks!

R&B :)