How to output a vector that is the sum of each slice or page from a 3D array on GPU

i7 quadcore, GTX Geforce 960 running CUDA driver, Matlab 2014b

My process I am trying to accomplish:

1. Send a 3D array to the GPU
2. Distribute each 'page' (:,:,i) for summing on the GPU
3. Return the output vector to the CPU

% My stab at it:
 array = ones(3,3,5,'gpuArray');   % create a 3x3x5 array of ones on GPU
 array = pagefun(@sum,array);      % for each page, sum all elements
 array = gather(array);            % return resulting 1x1x5 array to CPU 
% Desired output: array = 1x1x5 vector of 9's
% This throws an error that pagefun does not like the summation function.

On the CPU, a similar process works just fine. It works in a PARFOR or FOR loop just fine on the GPU, but this does not vectorize the process for speed. Would a CUDA kernel using a PTX file be more able to do something like this? Is there a better way to do this? Is this more suitable for a cluster rather than a GPU?

Help appreciated, Will

Answer 1

[Edric Ellis]

There's a much simpler solution that involves only a single summation, like so:

    array = ones(3, 3, 5, 'gpuArray');
    result = sum(reshape(array, [], size(array, 3)));
    result = gather(result);

In this case, result will not be the same shape as in your solution, but you can fix this if you wish like so

    result = reshape(result, 1, 1, []);

That way should be considerably faster on the GPU (I'm assuming your real data is much larger) because reshape can be done without manipulating the data (unlike rot90), and there's only a single call to sum.

[Will Kinsman]

Excellent, I have been looking for improvements like this. Thank you. Will

Answer 2

[Will Kinsman]

Solution:

array = ones(3,3,5,'gpuArray');   % create a 3x3x5 array of ones on GPU
array = sum(array);               % sum all column vectors to create 1x3x5
array = rot90(array);             % rot all pages to create 3x1x5
array = sum(array);               % sum all column vectors to create 1x1x5
array = gather(array);            % return resulting 1x1x5 array to CPU

unfortunately I asked the question using the sum function as a simpler proxy to determine how to do perform corrcoef on each page, however the question in the title above has been answered so I will mark this as closed.

[Will Kinsman]

I was also successful in my corrcoef function, it decomposing the non-gpu functions to using gpu-only functions