Parfor error with bsxfun

Dear all,

I'm having a weird issue with using parfor and bsxfun. I provide a minimal working example below (tried on both 2011b and 2013a). The version with for is working fine, whereas if I switch to parfor, I receive the 'non-singleton dimensions must match' error for bsxfun. I'm not sure if this was already answered although I looked for possible related questions&answers. Thanks in advance.

Cem

var1=repmat((1:1000)',2,3);
var2=ones(1000,1000);
var3=ones(6000,1);
temporary_var=bsxfun(@times,var2(var1,:),var3);
result=zeros(1000,1000);
for i=1:1000
    result(i,:)=sum(bsxfun(@times,var2(var1,i),temporary_var));
end

Note: the issue persists both matlabpools open and closed.

Answer 1

[Jill Reese]

I think that the parfor analysis is getting a bit confused by the multidimensional indexing into var2. If you re-write the code as follows it should work:

var1=repmat((1:1000)',2,3);
var2=ones(1000,1000);
var3=ones(6000,1);
temporary_var=bsxfun(@times,var2(var1,:),var3);
result=zeros(1000,1000);
tmp = var2(var1,:);
parfor i=1:1000
    result(i,:)=sum(bsxfun(@times,tmp(:,i),temporary_var));
end

[Matt J]

@Cem,

If this is what you want, the parfor loop looks like a doubtful way to go about it. A loop-free computation of the same result is

   result=tmp.'*temporary_var;

which would likely have better internal optimization/parallelization than parfor.