Comments on: A Brief History of polyval

By: Brian McIntosh

Brian McIntosh — Tue, 14 Jul 2009 14:53:49 +0000

Nice topic, Loren. A few years ago while working with “ln” near unity, I ginned up the following functions.

function b = exp1(a)
%EXP1 (exp(a)-1) function
% does not use HP-48 naming convention
% since expm is taken
% for use when exp(a) is nearly one
% should work for any matrix [a]

% Author(s): B. McIntosh 2004/03/01: 2004/03/30
% Private user

rsf = size(a);
a = reshape(a,1,prod(rsf));
b = exp(a)-1;
ii = find(abs(a) < 0.25);
b(ii) = (1./cumprod(1:12))*cumprod(ones(12,1)*a(ii));
b = reshape(b,rsf);
%% end of function

function b = lnp1(a)
%LNP1 log(a+1) function
% uses HP-48 naming convention
% for use when (a) is nearly zero
% should work for any matrix [a]

% Author(s): B. McIntosh 2004/03/01: 2004/03/10: 2004/03/30
% Private user

rsf = size(a);
a = reshape(a,1,prod(rsf));
b = log(a+1);
ii = find(abs(a) < 0.065);
b(ii) = -(1./(1:12))*cumprod(-1*ones(12,1)*a(ii));
b = reshape(b,rsf);
%% end of function

By: James Ross

James Ross — Mon, 13 Jul 2009 18:48:52 +0000

Loren,
Good question! I tend to write one algorithm and try to make it solid – the right answer is far more important than execution time, especially with the speed of today’s PC’s.

If I feel Matlab is taking too long to execute my routines, out comes the Profiler and I attack the big hitters until the execution time is back under control. Recently, a colleague and I were able to reduce algorithm run-time from 12.5 hours down to 2.5 hours with a handful of changes based on the Profiler! With a few additional changes based on how we used Simulink, the final execution time was 30 minutes.

While I’ve always been a big proponent of optimization to minimize end-user pain, that type of work has to be weighed against robustness and the effort required to eke out that last bit of performance.

By: Loren

Loren — Thu, 09 Jul 2009 10:52:49 +0000

Daniel-

Thanks for your thoughts. No spammers list, but the comments are moderated so I can weed out ones that are totally off-topic. Please keep on commenting!

–Loren

By: Daniel

Daniel — Thu, 09 Jul 2009 07:18:13 +0000

As a sometimes follower of Dougs mantra, that one should not optimize if it sacrifices readibility, I tend to very carefully avoid multiple algorithms that do the same thing. Not only does the code become complex, but one must also verify both implementations, and the algorithm which selects which one to use.

Now for end-user programs which are properly verified, that may be alright, but in my lab, one faulty calculation at the wrong time costs way more than a little sluggishness here and there.

–DA
PS. I seem to have ended up on the spamers-list for this blog. Is that because I comment too much? :P DS