for / parfor within if

Is there a way to write something like:

if parallelloop
  parfor i=1:n
else
  for i=1:n
end
...iterations
end

where parallellop is a boolean.

This code is not accepted by Matlab because the same end for both parfor and for.

A solution I do not accept for many reasons would be:

if parallelloop
  parfor i=1:n
...iterations
  end
else
  for i=1:n
...iterations
  end
end

Another solution I would like to have is an #ifdef like approach, but is there any?

Thanks by advance.

[Philippe Lebel]

what are the reasons why you dont want to put the end statments after the parfor and the for?

[Eric]

Because the code between for and end statements is too long and also if this writing can be done it can be written in a function with a genric name that will perform a parfor loop or a simple for loop, depending on the capabilities of the Matlab.

Just as a comparizon, I have the same approach with openMP and C++, the same code with preprocessor directives and if openMP is installed it is used, otherwise it is not. And I think that many people using distcomp would like to have this flexibility: a single code for many platforms/architectures.

Answer 1

[Jason Ross]

You can control if a loop is run with parallel workers or not by changing the number of workers. See "Test parfor-Loops by Switching Between Parallel and Serial Execution" in the parfor documentation. Note that even with 0 workers, the parfor is still nondeterministic.

To make your example work, you would set the number of workers to 0 if parallelloop is false.

[Eric]

Yes I agree, but you need distcomp installed.

If I want to have this writing it is because I have servers running Matlab with distcomp and laptops running it alone, but I want my code to run on both.

[Jason Ross]

You don't need to do that -- parfor will be essentially run as a for without Parallel Computing Toolbox installed. See here.

[Eric]

You are partially right because:

1. computational time is not the same (as seen in your link) and
2. parfor is not known on old versions of Matlab (mine is R2007 and parfor is not known).

[Jason Ross]

There have been a lot of changes since then -- the computational time issue might have been a bug, and fixed. The overall point I was getting at was that parfor is interpreted as a for in the absence of Parallel Computing Toolbox being present.

This also does make the implicit assumption that parfor introduces new functionality that would be incompatible with releases that don't know what parfor is, but that's going to be the case with pretty much any arbitrary software package.