Thread Subject: vectorize loop exponential

Hi All

Any ways to vectorize this loop better

N = 100000

for i = 1 to N
Ymat(:,:,i) = exp(Alphamat + Betamat.*(Xmat(:,:,i))
end

All the matrices are 4*4 matrices.

I tried..

Alphamat1 = repmat(Alphamat,[1 1 N])
Betamat1 = repmat(Betamat, [1 1 N])
Ymat = exp(bsxfun(@plus,Alphamat1,bsxfun(@times,Betamat1,Xmat)))

It didn't really help. Actually it resulted in an increase in time.I tried splitting it up using a temporary variable, but that didn't help either. On profiling, the exp operation is the biggest bottleneck, but I'm hoping vectorizing could reduce the number of calls and thus save time.

Thanks for any suggestions..

"Rajgopal " <runraj@gmail.com> wrote in message <h6b5lm$99t$1@fred.mathworks.com>...
> Hi All
>
> Any ways to vectorize this loop better
>
> N = 100000
>
> for i = 1 to N
> Ymat(:,:,i) = exp(Alphamat + Betamat.*(Xmat(:,:,i))
> end
>
> All the matrices are 4*4 matrices.
>
> I tried..
>
> Alphamat1 = repmat(Alphamat,[1 1 N])
> Betamat1 = repmat(Betamat, [1 1 N])
> Ymat = exp(bsxfun(@plus,Alphamat1,bsxfun(@times,Betamat1,Xmat)))
>
> It didn't really help. Actually it resulted in an increase in time.I tried splitting it up using a temporary variable, but that didn't help either. On profiling, the exp operation is the biggest bottleneck, but I'm hoping vectorizing could reduce the number of calls and thus save time.
>
> Thanks for any suggestions..

one of the solutions
- the for loop most likely is the fastest candidate
- if(f) you take the costly EXP out of it and run the function at the end...

% 1) loop
for i=1:n
     y(:,:,i)=bm.*x(:,:,i);
end
% 2) or
     y=bsxfun(@times,x,bm)
% 3) final pathway...
     y=exp(am+y);

us

Thanks a lot for the tip. The bsxfun version is faster after adjusting the dimensions of bm.. The exp outside saves a lot of time too!

"us " <us@neurol.unizh.ch> wrote in message <h6baha$j4f$1@fred.mathworks.com>...
> "Rajgopal " <runraj@gmail.com> wrote in message <h6b5lm$99t$1@fred.mathworks.com>...
> > Hi All
> >
> > Any ways to vectorize this loop better
> >
> > N = 100000
> >
> > for i = 1 to N
> > Ymat(:,:,i) = exp(Alphamat + Betamat.*(Xmat(:,:,i))
> > end
> >
> > All the matrices are 4*4 matrices.
> >
> > I tried..
> >
> > Alphamat1 = repmat(Alphamat,[1 1 N])
> > Betamat1 = repmat(Betamat, [1 1 N])
> > Ymat = exp(bsxfun(@plus,Alphamat1,bsxfun(@times,Betamat1,Xmat)))
> >
> > It didn't really help. Actually it resulted in an increase in time.I tried splitting it up using a temporary variable, but that didn't help either. On profiling, the exp operation is the biggest bottleneck, but I'm hoping vectorizing could reduce the number of calls and thus save time.
> >
> > Thanks for any suggestions..
>
> one of the solutions
> - the for loop most likely is the fastest candidate
> - if(f) you take the costly EXP out of it and run the function at the end...
>
> % 1) loop
> for i=1:n
> y(:,:,i)=bm.*x(:,:,i);
> end
> % 2) or
> y=bsxfun(@times,x,bm)
> % 3) final pathway...
> y=exp(am+y);
>
> us