MATLAB Central Newsreader - very slow vectorized code

very slow vectorized code

Julian Reichl — Thu, 24 Apr 2008 06:50:04 -0400

Hi all,
I have a problem that I can't figure out. I have two
versions of some code, one vectorised, the other not. The
vectorised version runs WAY slower for single repeats - ~17
seconds cf. 0.3 seconds.

The profiler tells me that pretty much every line of the
code is a lot slower in the vectorised version. Even simple
things like

e=repmat(1,1);
for i=1:28305
e(:,:)=0;
end

(the 1:28305 is a time-stepping loop and can't be removed)
takes 0.39 s inside the vectorised code (not including the
repmat), whereas in the command line it takes less than a
10th of that time.
if, in the same code (the vectorised version), I replace it
with

for i=1:28305
e=0;
end

(which is fine for this example because it's a 1x1, but not
for my application), the time for that function drops,
again, to ~ a 10th.

I really hope someone has an idea!!!

thanks

Julian

Re: very slow vectorized code

Julian Reichl — Thu, 24 Apr 2008 07:07:02 -0400

I just tried commenting out all the other code within the loop
i=1:28305
...
end
and it bought that e(:,:)=0;
statement back to 0.04s. So could this mean that there is
some memory issue?

"Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
<fupais$8av$1@fred.mathworks.com>...
> Hi all,
> I have a problem that I can't figure out. I have two
> versions of some code, one vectorised, the other not. The
> vectorised version runs WAY slower for single repeats - ~17
> seconds cf. 0.3 seconds.
>
> The profiler tells me that pretty much every line of the
> code is a lot slower in the vectorised version. Even simple
> things like
>
> e=repmat(1,1);
> for i=1:28305
> e(:,:)=0;
> end
>
> (the 1:28305 is a time-stepping loop and can't be removed)
> takes 0.39 s inside the vectorised code (not including the
> repmat), whereas in the command line it takes less than a
> 10th of that time.
> if, in the same code (the vectorised version), I replace it
> with
>
> for i=1:28305
> e=0;
> end
>
> (which is fine for this example because it's a 1x1, but not
> for my application), the time for that function drops,
> again, to ~ a 10th.
>
> I really hope someone has an idea!!!
>
> thanks
>
> Julian
>

Re: very slow vectorized code

Bill Nell — Fri, 25 Apr 2008 15:26:17 -0400

Julian Reichl wrote:
> I just tried commenting out all the other code within the loop
> i=1:28305
> ...
> end
> and it bought that e(:,:)=0;
> statement back to 0.04s. So could this mean that there is
> some memory issue?
>
> "Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
> <fupais$8av$1@fred.mathworks.com>...
>> Hi all,
>> I have a problem that I can't figure out. I have two
>> versions of some code, one vectorised, the other not. The
>> vectorised version runs WAY slower for single repeats - ~17
>> seconds cf. 0.3 seconds.
>>
>> The profiler tells me that pretty much every line of the
>> code is a lot slower in the vectorised version. Even simple
>> things like
>>
>> e=repmat(1,1);
>> for i=1:28305
>> e(:,:)=0;
>> end
>>
>> (the 1:28305 is a time-stepping loop and can't be removed)
>> takes 0.39 s inside the vectorised code (not including the
>> repmat), whereas in the command line it takes less than a
>> 10th of that time.
>> if, in the same code (the vectorised version), I replace it
>> with
>>
>> for i=1:28305
>> e=0;
>> end
>>
>> (which is fine for this example because it's a 1x1, but not
>> for my application), the time for that function drops,
>> again, to ~ a 10th.
>>
>> I really hope someone has an idea!!!
>>
>> thanks
>>
>> Julian
>>
>

Julian,
If you are just trying to set the contents of e to 0 then just use
e(:) = 0 instead of e(:,:) = 0.

Bill

Re: very slow vectorized code

Yi Cao — Fri, 25 Apr 2008 19:18:01 -0400

"Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
<fupbim$hs6$1@fred.mathworks.com>...
> I just tried commenting out all the other code within the
loop
> i=1:28305
> ...
> end
> and it bought that e(:,:)=0;
> statement back to 0.04s. So could this mean that there is
> some memory issue?
>
> "Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
> <fupais$8av$1@fred.mathworks.com>...
> > Hi all,
> > I have a problem that I can't figure out. I have two
> > versions of some code, one vectorised, the other not.
The
> > vectorised version runs WAY slower for single repeats -
~17
> > seconds cf. 0.3 seconds.
> >
> > The profiler tells me that pretty much every line of the
> > code is a lot slower in the vectorised version. Even
simple
> > things like
> >
> > e=repmat(1,1);
> > for i=1:28305
> > e(:,:)=0;
> > end
> >
> > (the 1:28305 is a time-stepping loop and can't be
removed)
> > takes 0.39 s inside the vectorised code (not including
the
> > repmat), whereas in the command line it takes less than
a
> > 10th of that time.
> > if, in the same code (the vectorised version), I
replace it
> > with
> >
> > for i=1:28305
> > e=0;
> > end
> >
> > (which is fine for this example because it's a 1x1, but
not
> > for my application), the time for that function drops,
> > again, to ~ a 10th.
> >
> > I really hope someone has an idea!!!
> >
> > thanks
> >
> > Julian
> >
>

I guess this is because JIT accelerator, i.e. your
vectorized code is quite complicated, hence the JIT
accelerator is not able to accelerate it, while your simple
loop version is in favour for the JIT to accelerate. When
you remove all other lines in the loop, the vectorized
version becomes simple as well, hence the JIT can
accelerate it now. That is why you see speed improvement.
To demonstrate this, you can try run your code with JIT on
and off by using the command:

feature accel on

or

feature accel off

to see if this is the case.

hth
Yi Cao

Re: very slow vectorized code

Julian Reichl — Tue, 29 Apr 2008 03:24:01 -0400

Thanks Bill and Yi Cao for your replies...
I'll check out the JIT thing and let you know how I go,

cheers
Julian

"Yi Cao" <y.cao@cranfield.ac.uk> wrote in message
<futap9$f0d$1@fred.mathworks.com>...
> "Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
> <fupbim$hs6$1@fred.mathworks.com>...
> > I just tried commenting out all the other code within the
> loop
> > i=1:28305
> > ...
> > end
> > and it bought that e(:,:)=0;
> > statement back to 0.04s. So could this mean that there is
> > some memory issue?
> >
> > "Julian Reichl" <julesreichl@yahoo.co.uk> wrote in message
> > <fupais$8av$1@fred.mathworks.com>...
> > > Hi all,
> > > I have a problem that I can't figure out. I have two
> > > versions of some code, one vectorised, the other not.
> The
> > > vectorised version runs WAY slower for single repeats -
> ~17
> > > seconds cf. 0.3 seconds.
> > >
> > > The profiler tells me that pretty much every line of the
> > > code is a lot slower in the vectorised version. Even
> simple
> > > things like
> > >
> > > e=repmat(1,1);
> > > for i=1:28305
> > > e(:,:)=0;
> > > end
> > >
> > > (the 1:28305 is a time-stepping loop and can't be
> removed)
> > > takes 0.39 s inside the vectorised code (not including
> the
> > > repmat), whereas in the command line it takes less than
> a
> > > 10th of that time.
> > > if, in the same code (the vectorised version), I
> replace it
> > > with
> > >
> > > for i=1:28305
> > > e=0;
> > > end
> > >
> > > (which is fine for this example because it's a 1x1, but
> not
> > > for my application), the time for that function drops,
> > > again, to ~ a 10th.
> > >
> > > I really hope someone has an idea!!!
> > >
> > > thanks
> > >
> > > Julian
> > >
> >
>
> I guess this is because JIT accelerator, i.e. your
> vectorized code is quite complicated, hence the JIT
> accelerator is not able to accelerate it, while your simple
> loop version is in favour for the JIT to accelerate. When
> you remove all other lines in the loop, the vectorized
> version becomes simple as well, hence the JIT can
> accelerate it now. That is why you see speed improvement.
> To demonstrate this, you can try run your code with JIT on
> and off by using the command:
>
> feature accel on
>
> or
>
> feature accel off
>
> to see if this is the case.
>
> hth
> Yi Cao
>