Using GPU, multiply a 3D matrix by a 2D matrix (slicewise)

Hi everyone,

I am trying to vectorize the code for my neural network so that it can be quickly executed on the GPU.

I need to multiply each slice of 3D matrix X by a 2D matrix T.

X = 3D matrix.

T = 2D matrix.

Is there a good, fast way to do this using the GPU? I have heard it suggested to use 'repmat' to create a 3D matrix from the 2D matrix (duplicating it many times). But it feels wasteful and inefficient.

Thanks!

Answer 1

[Edric Ellis]

In this case, you can use pagefun. For example:

X = rand(10, 10, 4, 'gpuArray');
T = rand(10, 'gpuArray');
pagefun(@mtimes, X, T)

[Brad Hesse]

Oh my god! I am speechless! My entire forward/backward propagation algorithm worked, on the very first try, after completely re-writing it to be vectorized for GPU execution. This is at least a 40-50 fold speed improvement (my original code obviously wasn't even very well optimized for CPU execution).

I cannot believe how fast this is.

Thank you so much for your help Edric. I had actually already tried using the pagefun function, but it failed and I assumed it didn't work with 3D matrices x 2D matrices.

[Edric Ellis]

Glad it worked for you!