Thread Subject:

Indexing, modifying, then returning values

Good morning,

Thank you again for your help. I get a bit confused when I try to vectorize the following "for" loops. The purpose of the loops is to index into P_4d_BB, grab out the correct values, modify the values, then put them back into the correct index where they came from. Index_beam_range is a 3d table that is indexed i,j,k where the values in index_beam_range(i,j,:) are the index values of interest in P_4d_BB. The "for" loops work but as my index_beam_range table gets larger it will obviously take much more time. P_3d_BB is a 3d matrix here P_3d_BB(m,n,:) is indexed by index_beam_range. The line for ibrt strips any NaN values in that particular (i,j,:) "vector" of index_beam_range.

for rb = 1:size(index_beam_range,2)%step through range bin table index
%second loop beam pointing
for bc = 1:size(index_beam_range,1)%step through the beam pointing angles

ibrt = index_beam_range(bc,rb,~isnan(index_beam_range(bc,rb,:)));
af = size(ibrt,3);
npw = sum(P_3d(ibrt))/af;
P_3d_BB(ibrt) = npw;
clear ibrt

Thank you,

Tony

end%beam pointing angle
end%range bin table index

I feel like I must have written the problem in a confusing way. I thought this would be a really ez vectorizing problem to figure out. I have been staring at this since yesterday without any improvement...

Thanks

T

>
> for rb = 1:size(index_beam_range,2)%step through range bin table index
> %second loop beam pointing
> for bc = 1:size(index_beam_range,1)%step through the beam pointing angles
>
> ibrt = index_beam_range(bc,rb,~isnan(index_beam_range(bc,rb,:)));
> af = size(ibrt,3);
> npw = sum(P_3d(ibrt))/af;
> P_3d_BB(ibrt) = npw;
> clear ibrt
>
> Thank you,
>
> Tony
>

Ok, so i've been thinking about this and it looks like to vectorize the for look I would need many temporary matrices to hold each value of index_beam_range.

Now I'm looking at parallelizing the problem. I have a quadcore I am working with so I could hope to split is locally 4 times. But I get an error:
??? Error: File: /home/otw/hopf/matlab/code/TEMP10/beambincorrect.m Line: 20 Column: 1
The variable P_4d_BB in a parfor cannot be classified.
 See <a href="matlab: helpview([docroot
 '/toolbox/distcomp/distcomp_ug.map'],'MIR_ERROR_PARFOR_BAD_TEMPORARY_VARIABLE')">Parallel for
 Loops in MATLAB, "Overview"</a>.

It looks like I can't slice up P_4d_BB because multiple points are indexed at a time... but because each internal for loop acts independently shouldn't I be able to parallelize it? I know I can't clear ibrt in the parfor, so i also took that out.

Any help?

Thank you,

Tony
>
> for rb = 1:size(index_beam_range,2)%step through range bin table index
> %second loop beam pointing
> for bc = 1:size(index_beam_range,1)%step through the beam pointing angles
>
> ibrt = index_beam_range(bc,rb,~isnan(index_beam_range(bc,rb,:)));
> af = size(ibrt,3);
> npw = sum(P_3d(ibrt))/af;
> P_3d_BB(ibrt) = npw;
>
> end%beam pointing angle
> end%range bin table index

"Anthony Hopf" <anthony.hopf@gmail.com> writes:

> Ok, so i've been thinking about this and it looks like to vectorize
> the for look I would need many temporary matrices to hold each value
> of index_beam_range.
>
> Now I'm looking at parallelizing the problem. I have a quadcore I am working with so I could hope to split is locally 4 times. But I get an error:
> ??? Error: File: /home/otw/hopf/matlab/code/TEMP10/beambincorrect.m Line: 20 Column: 1
> The variable P_4d_BB in a parfor cannot be classified.
> See <a href="matlab: helpview([docroot
> '/toolbox/distcomp/distcomp_ug.map'],'MIR_ERROR_PARFOR_BAD_TEMPORARY_VARIABLE')">Parallel for
> Loops in MATLAB, "Overview"</a>.
>
> It looks like I can't slice up P_4d_BB because multiple points are
> indexed at a time... but because each internal for loop acts
> independently shouldn't I be able to parallelize it?

Could you post an example showing the actual PARFOR loop you're trying
to run which is failing?

Cheers,

Edric.

Edric,

Sorry about that. It is the same code as above but the inner loop is the parfor. I have a feeling I can't run the parfor loop within another for loop. I need to reshape index_beam_range so I can write one for loop to step through the entire array... it will be i*j in length. But assuming we can reshape index_beam_range in this way... basically just one parfor loop the code would look like this:

P_4d_BB = P_4d;
aif = ones(size(P_4d)); %antenna integration factor (to be applied to informativeness)
%first loop range bin
parfor bc = 1:length(index_beam_range)%step through the beam pointing angles
ibrt = index_beam_range(bc,~isnan(index_beam_range(bc,:)));
aif(ibrt) = size(ibrt,3);
npw = nansum(P_4d(ibrt));
P_4d_BB(ibrt) = npw;
end%beam pointing angle



Edric M Ellis <eellis@mathworks.com> wrote in message <ytwwrv0nldv.fsf@uk-eellis-deb5-64.mathworks.co.uk>...
> "Anthony Hopf" <anthony.hopf@gmail.com> writes:
>
> > Ok, so i've been thinking about this and it looks like to vectorize
> > the for look I would need many temporary matrices to hold each value
> > of index_beam_range.
> >
> > Now I'm looking at parallelizing the problem. I have a quadcore I am working with so I could hope to split is locally 4 times. But I get an error:
> > ??? Error: File: /home/otw/hopf/matlab/code/TEMP10/beambincorrect.m Line: 20 Column: 1
> > The variable P_4d_BB in a parfor cannot be classified.
> > See <a href="matlab: helpview([docroot
> > '/toolbox/distcomp/distcomp_ug.map'],'MIR_ERROR_PARFOR_BAD_TEMPORARY_VARIABLE')">Parallel for
> > Loops in MATLAB, "Overview"</a>.
> >
> > It looks like I can't slice up P_4d_BB because multiple points are
> > indexed at a time... but because each internal for loop acts
> > independently shouldn't I be able to parallelize it?
>
> Could you post an example showing the actual PARFOR loop you're trying
> to run which is failing?
>
> Cheers,
>
> Edric.

"Anthony Hopf" <anthony.hopf@gmail.com> writes:

> Sorry about that. It is the same code as above but the inner loop is
> the parfor. I have a feeling I can't run the parfor loop within
> another for loop. I need to reshape index_beam_range so I can write
> one for loop to step through the entire array... it will be i*j in
> length. But assuming we can reshape index_beam_range in this
> way... basically just one parfor loop the code would look like this:
>
> P_4d_BB = P_4d;
> aif = ones(size(P_4d)); %antenna integration factor (to be applied to informativeness)
> %first loop range bin
> parfor bc = 1:length(index_beam_range)%step through the beam pointing angles
> ibrt = index_beam_range(bc,~isnan(index_beam_range(bc,:)));
> aif(ibrt) = size(ibrt,3);
> npw = nansum(P_4d(ibrt));
> P_4d_BB(ibrt) = npw;
> end%beam pointing angle


It looks like you're trying to calculate "aif" and
"P_4d_BB". Unfortunately, the way you're indexing those is not
compatible with PARFOR - the output variables from a PARFOR loop must be
constructed either as "sliced" outputs, or "reduction" outputs. I think
"sliced" is most appropriate for what you're trying to do - a sliced
output variable basically means that you assign into it using the loop
variable (and possibly some other constant indices). For example, here's
an approximation to what I think you're trying to do:

data = rand(10);
parfor ii = 1:10
  transformedIdx = 11 - ii; % this is your "ibrt" calculation
  output( transformedIdx ) = sum( data( transformedIdx, : ) );
end

This fails because "output" is not sliced correctly. One way to fix
things is like this:

data = rand(10);
parfor ii = 1:10
  transformedIdx = 11 - ii; % this is your "ibrt" calculation
  % Store the mapping so it can be reversed later
  idxMap( ii ) = transformedIdx;
  % index into output using the loop variable "ii"
  output( ii ) = sum( data( transformedIdx, : ) );
end

% At this point, "output" is scrambled, but you can descramble it like
% so:
[y, i] = sort( idxMap );
output = output( i );

For more details, see the doc on sliced variables etc.:

<http://www.mathworks.com/access/helpdesk/help/toolbox/distcomp/brdqtjj-1.html>

Cheers,

Edric.

Edric,

Thank you for the response!! I will work on this...

Tony

Edric M Ellis <eellis@mathworks.com> wrote in message <ytwocgbnj92.fsf@uk-eellis-deb5-64.mathworks.co.uk>...
> "Anthony Hopf" <anthony.hopf@gmail.com> writes:
>
> > Sorry about that. It is the same code as above but the inner loop is
> > the parfor. I have a feeling I can't run the parfor loop within
> > another for loop. I need to reshape index_beam_range so I can write
> > one for loop to step through the entire array... it will be i*j in
> > length. But assuming we can reshape index_beam_range in this
> > way... basically just one parfor loop the code would look like this:
> >
> > P_4d_BB = P_4d;
> > aif = ones(size(P_4d)); %antenna integration factor (to be applied to informativeness)
> > %first loop range bin
> > parfor bc = 1:length(index_beam_range)%step through the beam pointing angles
> > ibrt = index_beam_range(bc,~isnan(index_beam_range(bc,:)));
> > aif(ibrt) = size(ibrt,3);
> > npw = nansum(P_4d(ibrt));
> > P_4d_BB(ibrt) = npw;
> > end%beam pointing angle
>
>
> It looks like you're trying to calculate "aif" and
> "P_4d_BB". Unfortunately, the way you're indexing those is not
> compatible with PARFOR - the output variables from a PARFOR loop must be
> constructed either as "sliced" outputs, or "reduction" outputs. I think
> "sliced" is most appropriate for what you're trying to do - a sliced
> output variable basically means that you assign into it using the loop
> variable (and possibly some other constant indices). For example, here's
> an approximation to what I think you're trying to do:
>
> data = rand(10);
> parfor ii = 1:10
> transformedIdx = 11 - ii; % this is your "ibrt" calculation
> output( transformedIdx ) = sum( data( transformedIdx, : ) );
> end
>
> This fails because "output" is not sliced correctly. One way to fix
> things is like this:
>
> data = rand(10);
> parfor ii = 1:10
> transformedIdx = 11 - ii; % this is your "ibrt" calculation
> % Store the mapping so it can be reversed later
> idxMap( ii ) = transformedIdx;
> % index into output using the loop variable "ii"
> output( ii ) = sum( data( transformedIdx, : ) );
> end
>
> % At this point, "output" is scrambled, but you can descramble it like
> % so:
> [y, i] = sort( idxMap );
> output = output( i );
>
> For more details, see the doc on sliced variables etc.:
>
> <http://www.mathworks.com/access/helpdesk/help/toolbox/distcomp/brdqtjj-1.html>
>
> Cheers,
>
> Edric.