Using the GPU through the parallel comp. toolbox to optimize matrix inversion

Hello,

In my code I impliment:

A=B\c;

B is an invertable matrix (14400x14400), c is a vector.

Since B is large this takes very long. I am using a windows 7, 64bit, 4 GB RAM, 2 cores.

Would I be able to shorten the time by using the GPU through the parallel comp. toolbox?

Would I need to write my own version of parallel matrix inversion, or has this problem been faced before and there is a posted solution?

Thank you.

Answer 1

[Jill Reese]

As of MATLAB R2010b this functionality has been available on the GPU in the Parallel Computing Toolbox. In order to use it, at least one of your variables B and c must be transferred to the GPU before calling mldivide (\).

In the latest release of MATLAB (R2012b) you can use this code to solve your problem:

B=gpuArray(B); % transfer B to the GPU
c=gpuArray(c); % transfer c to the GPU

A = B \ c; % Solve the linear system on the GPU and store A on the GPU
           % This line of your existing code doesn't change at all

% you can continue to perform work on the GPU using A, B, and c or

A=gather(A); % transfer A back to the MATLAB workspace

Answer 2

[lior]

Great,

So mldivide knows how to exploit the GPU and run in parallel?

I guess that is true for every function in MATLAB, its being "parallelized" when I use it on "gpuArrey" variables without having to optimize it myself?

How about functions I write myself? if I write a simple "for" loop with basic math operators, then the GPU would not be used even if the variables are "gpuArrey"?

[Matt J]

If the operations done in each pass through the loop are element-wise, then you can use arrayfun to parallelize the loop on the GPU.