Run custom function on GPU with Parallel Computing Toolbox?

I have a function which needs to find min( b -A*x ). The dimensions of b = 5 X 1; A = 5 X 2; x = 2 X 10; The output of A*x would be 5 X 10. and since b is a column vector, the final output will be 5 X 10, [b - A*x] and if we apply min operator we would get 5 X 1. When I try to use the arrayfun function it gives error :- Error using parallel.gpu.GPUArray/arrayfun Matrix dimensions must agree for 'mtimes'

Answer 1

[Jill Reese]

Arrayfun on the GPU only supports elementwise operations, so it is not the command that you want to use in this case.

You need to use min and * (mtimes) directly, which are overloaded to run on the GPU. If at least one of A, b, or x are stored on the GPU (i.e. of type parallel.gpu.GPUArray), then your code snippet will run on the GPU.

[Subhrajit Debnath]

Thanks a lot. I am trying to do bsxfun(@minus,b,A*x);

It is giving error as: Undefined function 'bsxfun' for input arguments of type 'parallel.gpu.GPUArray'.

It seems this function is not supported. I am using Parallel Computing Toolbox Version 5.2 (R2011b).

But this function is listed in the built in functions that support gpuArray.

http://www.mathworks.in/help/toolbox/distcomp/bsic4fr-1.html#bsloua3-1

[Jill Reese]

The link you provide is for R2012a. If you upgrade your version of MATLAB to R2012a then you will have access to bsxfun on the gpu.

Answer 2

[Walter Roberson]

column vector minus a matrix is not defined.

You will need to show your arrayfun code; it sounds as if you are using the wrong arguments or wrong function.

[Subhrajit Debnath]

yes it can be done.

bsxfun(@minus,b,A*x); will do matrix column-wise subtraction.