How to accelerate exponent (exp) calculation in MATLAB?

Hello everyone. I developed a program and I find it spends lots of time on calculating exp(-ax). So is there any way to accelerate such calculations exp(-ax)? Many thanks~

[KSSV]

It depends on how you have coded....we can help you only on knowing how your code is.

[Walter Roberson]

Are you doing a symbolic calculation?

[Tian]

Not a symbolic calculation, just need to calculate tens of thousands of simple exp(ax). I have tried to use arrayfun, or matrix calculation, but too much memory is required.

[Walter Roberson]

ZZZ = rand(1,1E6);
timeit(@() exp(ZZZ),0)

This is about 0.0045 seconds, so even 10^8 exp() would take less than 1/2 second.

Answer 1

[Jan]

No. exp() is an expensive operation. The only way to accelerate it is to reduce repeated calls with the same argument.

I've experimented with the implementation of netlib.org, but did not found a significant acceleration and the results are lesss accurate. I assume this is the reason why it is not used in Matlab (in opposite to e.g. acos, which is taken from netlib).

A parfor loop might help to distribute the calculation to multiple cores.

Note: Think twice. If there would be a way to accelerate the computation of a specific function, it would have been included by Mathworks certainly. It is very pitty, but there are no magic tricks to make Matlab faster.

[Tian]

Thank you for your answer. I will try other ways. I've tried parfor, but it is not so efficient in my code. Thanks again~

[Adam]

Make sure you vectorise calls to exp as much as yo can too.

[Tian]

OK, thanks a lot.

Answer 2

[Rangeet Mitra]

Try replacing exp() by a custom function using a Hermite approximation for exp() (which is learnt a-priori for the dynamic range of your code). Without loss of performance, you will get a considerable speedup!!!