Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!natinst!sequoia!memqa!r91400
From: r91400@memqa.uucp (Michael C. Grant)
Newsgroups: comp.lang.c++
Subject: Re: parameter[][]
Message-ID: <5044@memqa.uucp>
Date: 8 Aug 90 10:27:17 GMT
References: <1545@m1.cs.man.ac.uk> <26B9BF5A.7291@orion.oac.uci.edu> <37961@ucbvax.BERKELEY.EDU> <65983@lll-winken.LLNL.GOV>
Organization: Memory R&QA, Motorola SPD
Lines: 35

In article <65983@lll-winken.LLNL.GOV>, jac@gandalf..llnl.gov (James Crotinger) writes:
> In article <37961@ucbvax.BERKELEY.EDU> jbuck@galileo.berkeley.edu (Joe Buck) writes:
>>
>>
>>double **mat_alloc (int nrows, int ncols) {
>>	double **t = new double* [nrows];
>>	for (int i = 0; i < ncols; i++)
>>		t[i] = new double [ncols];
>>}
>>
>   There are at least a couple of problems with this approach. The
> first is that, as written, the inter-row spacing is not constant. This
> may hurt your performance on some architectures when you try to do a
> loop over the first index. (This is particularly true of vectorizing
> architectures but can also give a performance hit on systems with
> interleaved memory access).  In fact optimization in general is
> greatly hindered by the use of a double ** to represent a simple
> square or rectanglar array since the optimizer no longer has any way
> of knowing that it is, indeed, dealing with such a simple layout of
> the data.

It really depends on how you access the data.  If you are performing
sequential access, then yes I agree that a simple rectangular array
would speed things up quite a bit--but for relatively random access,
you really can't beat this double ** approach.  Some processors 
can implement this as a single indirect addressing with offset instruction.
Most could do it in two instructions.  But, the simple rectangular 
array would require a multiply and addition IN addition to the access.

Of course, if access time is only a small portion of your program's
execution time, then optimizing access won't to diddley to improve
the performance.  For most programs, it's worthless to quibble
over it (see Amdahl's law...)

Michael C. Grant