[Detrius]

I'm looking for a way to use vectors and matrices as defined by linear algebra. I know it's possible to do this with AltiVec, but I'm also concerned about doing this on a G3 processor, not just a G4. I really don't want to have to start from scratch and define these myself.

I've just recently stumbled across the vecLib FrameWork in the System folder. I haven't seen any real documentation on it yet, but I'm reading the header files. Is this library AltiVec accelerated? If so, will it also work properly on non-G4's?


Thanks in advance for any information or advice.

[Detrius]

I found the answer to my own question (mostly).

/Developer/Documentation/CoreTechnologies/vDSP/vDSP_Library.pdf


From what I can tell, through skimming this pdf and skimming through the actual header files, it's up to the developer to determine whether or not the machine has AltiVec available. I wouldn't even know where to begin to look to find this. Also, some of the functions in the vecLib framework don't use altivec at all. So it seems it's entirely possible to do basic linear algebra subroutines (BLAS) without the G4... you just have to use different functions.


hmmm... no fun. I couldn't afford spending the extra $1000 for my laptop to match my desktop. oh well.


Any input is still appreciated.

[DaGuy]

There's some pretty good stuff for numerical computing on OS X at:
http://floyd.liunet.edu/~gkhanna/hpc.html

Enjoy,

[Seb G]

As far as I know, AltiVec would limit you to work in up to four dimensions anyway (unless you use 8bit integers, then you get 16).

If you happen to have big matrices and end up coding your own matrix multiplication (it's not implemented in vDSPlib) , you might want to check out Strassen's algorithm (see Introduction to Algorithms by Cormen, Leiserson, Rivest - or web search).

[larkost]

Actually, a CS vector has nothing to do with a LinAlg vector. The Computer Science concept vector calculation is a row of numbers that you do the same thing to. For example take the vector 4 : 6 : 8 : 3, and do a vector add of 1 to them (implicitly 1 : 1 : 1 : 1) and you get 5 : 7 : 9 : 4. The Math/Geometric concept is very different.

If you were brute forcing the equations, or calculating end results, then AltiVec might come in handy to paralelize the operations, but teh AltiVec unit is not going to make symbolic math programming any easier.

[The Godfather]

Shouldn't vector operations (mathematical or geometric) be hand-optimized in assembly to take advantage of Altivec? How does one know that code is being G4-optimized?

[Detrius]

Originally posted by DaGuy:
<STRONG>There's some pretty good stuff for numerical computing on OS X at:
http://floyd.liunet.edu/~gkhanna/hpc.html

Enjoy,
</STRONG>

Actually, I've seen that site. It came in quite handy for my parallel programming class. However, I didn't find it entirely useful for this other project, as the vectorizing compilers aren't free (best I could tell), and they also won't work with Cocoa (Objective-C). I could, technically, do the work in C++, but then I would have to write an additional objective-c class to interact between the C++ class and the user interface. ...not much fun.

[Detrius]

Originally posted by The Godfather:
<STRONG>Shouldn't vector operations (mathematical or geometric) be hand-optimized in assembly to take advantage of Altivec? How does one know that code is being G4-optimized?</STRONG>

There are certain libraries offered by Apple that have already been written to take advantage of AltiVec and that have been hand optimized. If the developer uses those functions, the developer gets the benefits of AltiVec without writing assembly by hand.

[Wixar]

You want MathArray

ftp://next-ftp.peak.org/pub/openstep/multi-platforms/sourcelibrary/classes/MathArray.0.60.s.tar.gz
ftp://ftp.informatik.uni-muenchen.de/pub/comp/platforms/next/Science/mathematics/MathArray.0.60.s.tar.gz

From the readme:

"MathArray is a general library of classes for performing mathematical operations on arrays (vectors, matrices, etc) of values. It can operate on any standard 'C' number type plus numbers of complex type."

-Peter

[kamprath]

Originally posted by Detrius:
<STRONG>I'm looking for a way to use vectors and matrices as defined by linear algebra. I know it's possible to do this with AltiVec, but I'm also concerned about doing this on a G3 processor, not just a G4. I really don't want to have to start from scratch and define these myself.

I've just recently stumbled across the vecLib FrameWork in the System folder. I haven't seen any real documentation on it yet, but I'm reading the header files. Is this library AltiVec accelerated? If so, will it also work properly on non-G4's?


Thanks in advance for any information or advice.</STRONG>

Just some advice . . . be concerned about doing the linear algebra before you pay too much attention to performance optimization (G4 vs. G3 vs. whatever). "Linear Algebra" and "vectorization of code" are two completely different things. The G4 vectorizes it's object codeprocessing - that's not the same thing as a "vector and matrix" in linear algebra.

Writing some classes that represent the linear math concepts of a vector anda matrix and the mathmetical operations between them is a good excercise for any programmer (I did it myself years ago). Once you do that, then revist the code for performance optimization with G4 AltiVec vectorization. In the mean time, keep your terms straight.

Michael Kamprath

[Detrius]

Originally posted by kamprath:
<STRONG>


Just some advice . . . be concerned about doing the linear algebra before you pay too much attention to performance optimization (G4 vs. G3 vs. whatever). "Linear Algebra" and "vectorization of code" are two completely different things. The G4 vectorizes it's object codeprocessing - that's not the same thing as a "vector and matrix" in linear algebra.

Writing some classes that represent the linear math concepts of a vector anda matrix and the mathmetical operations between them is a good excercise for any programmer (I did it myself years ago). Once you do that, then revist the code for performance optimization with G4 AltiVec vectorization. In the mean time, keep your terms straight.

Michael Kamprath</STRONG>

I didn't ever get the terms mixed up. I've written code for linear algebra vectors and matrices before, but I just don't feel like porting them over to objective-c. I was hoping for pre-existing code that I could use instead of porting it.

[kamprath]

Originally posted by Detrius:
<STRONG>


I didn't ever get the terms mixed up. I've written code for linear algebra vectors and matrices before, but I just don't feel like porting them over to objective-c. I was hoping for pre-existing code that I could use instead of porting it.</STRONG>

You don't need to port them to objective-C, why would you think that? If they are in C++ (which I am assuming), and if you want to write a Cocoa-based front end, you can use Apple's Objective-C++ compiler that came with the Dev-Tools December CD. This compiler allows you to intermix objective-C and C++ code. So, if you want to write a Cocoa based front end, you can still use your suite of C/C++ code. Pretty cool.

If you wrote your stuff in Pascal, or some other language, just create a dynamically linked library and export a C-API to it's methods.

:rant on:
Why does everyone think Objective-C is the only thing you can work in on MacOS X? The only reason Objective-C exists is because C++ wasn't ready for the prime time when NextSTEP was originally invented . . .
What's it been since then . . 13 years? I didn't see Be using Objective-C in their system. Be used C++, and it was a dream to program for . . . too bad they tried to bluff Apple :-/
:rant off:

Michael

[Detrius]

Originally posted by kamprath:
<STRONG>

You don't need to port them to objective-C, why would you think that? If they are in C++ (which I am assuming), and if you want to write a Cocoa-based front end, you can use Apple's Objective-C++ compiler that came with the Dev-Tools December CD. This compiler allows you to intermix objective-C and C++ code. So, if you want to write a Cocoa based front end, you can still use your suite of C/C++ code. Pretty cool.

If you wrote your stuff in Pascal, or some other language, just create a dynamically linked library and export a C-API to it's methods.

:rant on:
Why does everyone think Objective-C is the only thing you can work in on MacOS X? The only reason Objective-C exists is because C++ wasn't ready for the prime time when NextSTEP was originally invented . . .
What's it been since then . . 13 years? I didn't see Be using Objective-C in their system. Be used C++, and it was a dream to program for . . . too bad they tried to bluff Apple :-/
:rant off:

Michael</STRONG>

Actually, I've always read that you could use Objective-C and C++ in the same program with the gcc compiler. I tried it, but couldn't get it to compile. I didn't try very hard, but nevertheless, it didn't work. The C++ header file spat out a really bizarre error: the word "class" was not recognized. Without this, of course, there could be no compiling of the C++ class. If you have any pointers for how to get it to compile, I'd be glad to use my own classes until I find better optimized ones.

[Michel Fortin]

I think if you want a file to compile usign Objective-C++ you need to use the .mm extention.

[Detrius]

Originally posted by Michel Fortin:
<STRONG>I think if you want a file to compile usign Objective-C++ you need to use the .mm extention.</STRONG>

My first mental response to this was wondering what difference this would actually make, as this wouldn't have any effect on the header files. But then, I did a little bit of research on the web and found that this is indeed correct. This would be done for both C++ and Objective-C code--anything that will be using the two languages interchangeably.


Here's the reference document:
http://developer.apple.com/techpubs/...ctive-C++.html