Parallelized PINV
in [Matlab]
|
From: sara on 15 May 2010 10:48 Hello, Please let me know if the PINV function is supported on a parallel platform. i.e Is this operation legit: x = pinv(a) b where a is a codistributed array. Thank you.
From: Edric M Ellis on 17 May 2010 03:18 "sara " <ssalha(a)ucla.edu> writes: > Please let me know if the PINV function is supported on a parallel platform. i.e > Is this operation legit: x = pinv(a) b where a is a codistributed array. No, the PINV function is not currently supported for codistributed arrays (you can see a full list by executing "methods codistributed"). What's your application for using the pseudo-inverse? Could you use MATLAB's backslash operator instead? (This is supported for codistributed arrays). Cheers, Edric.
From: sara on 25 May 2010 01:38 Thank you, linsolve + qr would be really handy operating in parallel taking advantage of the spmd.
From: sara on 25 May 2010 15:13 I mean, Yes, backslash operator could do the job, however, in its attempt to find which decomposition to use, backslash takes long time, and also sometimes it ends up reducing the matrix incorrectly. QR is a lower level operator though than \ , pinv and SVD, so it might be easier to parallelize it? I wonder why it is not available. "sara " <ssalha(a)ucla.edu> wrote in message <htfnnr$kjn$1(a)fred.mathworks.com>... > Thank you, linsolve + qr would be really handy operating in parallel taking advantage of the spmd.
From: Bruno Luong on 25 May 2010 15:24
"sara " <ssalha(a)ucla.edu> wrote in message <hth7gk$gt3$1(a)fred.mathworks.com>... > > > QR is a lower level operator though than \ , pinv and SVD, so it might be easier to parallelize it? I wonder why it is not available. What you mean by "it is not available"? Of course it is. And I disagree, it is not lower level than SVD. Both are basic matrix factorizations. Bruno |