Thread Subject:

Time consuming power operation

Hi there matlab users.

I have a program that im running some optimization routines on using the profiler. I appears that the following line takes more time than I initially expected:

change = sqrt(sum((vectorChange).^2,2));

The variable "vectorChange" is just a three component (x,y,z) vector. The line is processed many million times during execution. My question is: What is taking so much time in this line, and why? And what can i do? I find it a little puzzling that this line takes so much time.

Regards, Morten.

"Morten Gjerding" <s082177@student.dtu.dk> wrote in message <i4m2so$4fb$1@fred.mathworks.com>...
> Hi there matlab users.
>
> I have a program that im running some optimization routines on using the profiler. I appears that the following line takes more time than I initially expected:
>
> change = sqrt(sum((vectorChange).^2,2));
>
> The variable "vectorChange" is just a three component (x,y,z) vector. The line is processed many million times during execution. My question is: What is taking so much time in this line, and why? And what can i do? I find it a little puzzling that this line takes so much time.
>
> Regards, Morten.

Well it is being called millions of times as you said. That would take some execution time.

What else are you doing, can you vectorize other things to minimize how many times this is called?

On 20/08/10 9:19 AM, Morten Gjerding wrote:

> I have a program that im running some optimization routines on using the
> profiler. I appears that the following line takes more time than I
> initially expected:
>
> change = sqrt(sum((vectorChange).^2,2));
> The variable "vectorChange" is just a three component (x,y,z) vector.
> The line is processed many million times during execution. My question
> is: What is taking so much time in this line, and why? And what can i
> do? I find it a little puzzling that this line takes so much time.

You might get better performance if you were to switch the vectors from
being row vectors to being column vectors.

Also, you might get better performance if you were to store the sum in a
temporary variable and multiply that variable by itself rather than
using .^2 .

"Sean " <sean.dewolski@nospamplease.umit.maine.edu> wrote in message <i4m3jp$l2p$1@fred.mathworks.com>...
> "Morten Gjerding" <s082177@student.dtu.dk> wrote in message <i4m2so$4fb$1@fred.mathworks.com>...
> > Hi there matlab users.
> >
> > I have a program that im running some optimization routines on using the profiler. I appears that the following line takes more time than I initially expected:
> >
> > change = sqrt(sum((vectorChange).^2,2));
> >
> > The variable "vectorChange" is just a three component (x,y,z) vector. The line is processed many million times during execution. My question is: What is taking so much time in this line, and why? And what can i do? I find it a little puzzling that this line takes so much time.
> >
> > Regards, Morten.
>
> Well it is being called millions of times as you said. That would take some execution time.
>
> What else are you doing, can you vectorize other things to minimize how many times this is called?


True :), of course it takes some time when it is processed so many times, but relative to other lines this takes much more time to process than expected.

Unfortunately, I have already done all the vectorization optimizations that i could think of.

Walter Roberson <roberson@hushmail.com> wrote in message <bswbo.58634$f_3.21887@newsfe17.iad>...
> On 20/08/10 9:19 AM, Morten Gjerding wrote:
>
> > I have a program that im running some optimization routines on using the
> > profiler. I appears that the following line takes more time than I
> > initially expected:
> >
> > change = sqrt(sum((vectorChange).^2,2));
> > The variable "vectorChange" is just a three component (x,y,z) vector.
> > The line is processed many million times during execution. My question
> > is: What is taking so much time in this line, and why? And what can i
> > do? I find it a little puzzling that this line takes so much time.
>
> You might get better performance if you were to switch the vectors from
> being row vectors to being column vectors.
>
> Also, you might get better performance if you were to store the sum in a
> temporary variable and multiply that variable by itself rather than
> using .^2 .

Ok i Will try to see of it will be less time consuming implementing the changes you mention. One question: Why would it make a difference to change to column vectors?

On 21/08/10 7:48 AM, Morten Gjerding wrote:
> Walter Roberson <roberson@hushmail.com> wrote in message

>> > change = sqrt(sum((vectorChange).^2,2));

>> You might get better performance if you were to switch the vectors
>> from being row vectors to being column vectors.

> Ok i Will try to see of it will be less time consuming implementing the
> changes you mention. One question: Why would it make a difference to
> change to column vectors?

MATLAB is optimized for memory access down columns rather than across
rows. The CPU code to access memory locations known to be adjacent can
be a bit more efficient. I don't promise that you will be able to
measure the difference (as I do not know the implementation details of
MATLAB) but there is the theoretical possibility that might add up when
you do the computation repeatedly.