Re: [Caml-list] Re: Camlp4 optimizations (was: productivity improvement)

[Brian Hurt <brian.hurt@qlogic.com>]

On Thu, 17 Oct 2002, Chris Hecker wrote:

> The biggest problem with making ocaml look nice and pretty for numerical 
> code is that there is no overloading (of functions or operators), and 
> camlp4 doesn't have access to types, so you can't have both:
> 
> s*a (scalar times matrix)
> a*b (matrix times matrix)
> 
> What code would the * operator generate in camlp4?  This needs to be fixed 
> at a deeper level than syntax (which is where camlp4 operates...and 
> operates well! :).  Althought hopefully it's fixed in a way that doesn't 
> require a runtime type-check if the information is available at compile time.
> 
> Or, maybe I'm missing something...it would be awesome if I was and this was 
> possible!
> 

I'd like to ask a stupid question here: how important is operator 
overloading?  Remember, before you answer, that FORTRAN managed to be 
crowned king of numerical languages for decades (and may still hold the 
crown depending upon who you talk to), with no operator or function 
overloading.  You had to call functions with names like cgbmv() (pop quiz- 
what does that function do?) and dgesvd().

I'm still an ocaml newbie, so I don't know how ocaml handles operator 
overloading.  I do know how C++ handles operator overloading.  Consider 
the 'innocent' statement:
    a = b + c + d;
What you want is for the compiler to produce code like:
    a = b;
    a += c;
    a += d;
Instead, what the compiler instead does is:
    t1 = b + c;
    t2 = t1 + d;
    a = t2;
generating two unnessecary temporaries.  If your objects are complex 
variables (2 floats) or even short 3D vectors (3 floats) this isn't too 
bad.  If your objects are 10,000 element vectors or matricies, creating 
two unnecessary temporaries (or even 1 unnecessary temporary) is bad.

So why not just code it as:
    a = b;
    a += c;
    a += d;
?  Well, I ask- is that code all that much more understandable then:

    matrix.assign a b ;
    matrix.addto a c;
    matrix.addto a d;
?  The advantage of operator overloading is the ability to express complex 
equations "obviously"-
    a = b + c + d;
is way more understandable than the two examples above or:
    matrix.assign a (matrix.add (matrix.add b c) d)
which is the equivelent.  But if you can't optimize it, you're just asking 
to produce bad code.

The other cent I'd like to throw into this discussion is a pointer at the 
OoLaLa project- which is attempting to build a whole new linear algebra 
library in Java with an OO design (instead of the thin wrappers around 
FORTRAN-era BLAS libraries and various operator overloading proposals):
http://citeseer.nj.nec.com/luj00oolala.html
Yes, I know this isn't a functional design.  But I'm throwing it out there 
as a springboard for ideas.

Brian


-------------------
To unsubscribe, mail caml-list-request@inria.fr Archives: http://caml.inria.fr
Bug reports: http://caml.inria.fr/bin/caml-bugs FAQ: http://caml.inria.fr/FAQ/
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners