Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!thunder.mcrcim.mcgill.edu!snorkelwacker.mit.edu!usc!samsung!uunet!microsoft!jimad
From: jimad@microsoft.UUCP (Jim ADCOCK)
Newsgroups: comp.lang.c++
Subject: Re: why is this program slow?
Message-ID: <60395@microsoft.UUCP>
Date: 10 Jan 91 18:32:51 GMT
References: <1991Jan9.002244.23398@news.cs.indiana.edu> <1991Jan9.101212@Pkg.Mcc.COM>
Reply-To: jimad@microsoft.UUCP (Jim ADCOCK)
Organization: Microsoft Corp., Redmond WA
Lines: 40

In article <1991Jan9.101212@Pkg.Mcc.COM> steve@Pkg.Mcc.COM (Steve Madere) writes
|This is a problem with C++ that I have known about for some time.
|The lure of simplified notation leads one to try to implement
|vector and matrix operations with special classes that will hide
|the complexity.  Unfortunately, all of those temporary variables
|that must be created to realize any mathematical operations really
|waste CPU time.  It takes a finite amount of time to allocate space,
|copy data, and free up the space again.  In many cases it is a
|significant percentage of the total run time of the program.
|
|A friend of mine at UCSD has worked on this problem for some time
|and concluded that what one needs is some kind of matrix awareness
|in the compiler to recognize when operations can be somehow 
|combined.  Clearly this is not the kind of thing that C++ was designed
|to do.  It is appropriate to add matrix awareness to a computational
|language like FORTRAN but C++ is not a scientific computation
|language.  (no flames please, I just mean that its PRIMARY purpose
|is not scientific computation).
|
|Any suggestions from the peanut gallery?

My claim would be that if one is serious about working with large objects --
say large matrices and/or large vectors, then one can afford to 
"do the job right" in C++ -- which means probably using reference counting
on the internal storage of the vectors and/or matrices, while maintaining
an external polymorphic wrapper over those guts which represents the 
exact form of the vector or matrix: is this matrix a "plain old matrix",
tridiagonal, identity, .... 

[In fact, probably some vendors already provide such libraries]

If you have a parallel computer, and a highly parallelizing fortran compiler,
you might want to write some of the guts of the internal representation of
these matrices in fortran -- but better yet, you'll probably find that 
your vendor already provides highly optimized routines for performing
those guts -- linpack-type stuff.

There's no reason why you can't have both the nice syntax of C++, and
the speed of fortran -- if you're willing to do a little serious work.
On the other hand, if you don't care, well, then you don't care.