How to run my code in GPU instead of CPU?

Hello everyone,

I'm trying to run my for loop using GPU instead of CPU, but i searched a lot and it turned out that I should rewrite my code again

but some people said it's not necessary to rewrite it, you could use some tool box or something like that

so, if anyone could help me how to use these tool boxes or how can i rewrite my code in GPU form, please answer me

This is my code:

x = [1 5 9];% input features
x = [1/9 5/9 1]; % normlize 
y = [0.49 0.51]; % outputs
min_error = 100;
for i = 1:10000
w1 = rand(3,3); 
b1 = rand(1,3); % first bias , 1 value going to 3 neurons
w2 = rand(3,3);
b2 = rand(1,3); % second bias
w3 = rand(3,2);
b3 = rand(1,2); % going to 2 neurons
% L1 input layer, L2 and L3 hidden layers , L4 output layer
% L1:
a1 = x;
% L1 - L2:
zh1 = (x * w1) + b1;
a2 = logsig(zh1);
% L2 - L3:
zh2 = (a2 * w2) + b2;
a3 = logsig(zh2);
% L3 - L4:
zh3 = (a3 * w3) + b3;
a4 = logsig(zh3);
err = abs(y-a4);
percentage(i) = (err / y) * 100;
if percentage(i) < min_error 
    min_error = percentage(i);
    cw1 = w1;
    cw2 = w2;
    cw3 = w3;
    bw1 = b1;
    bw2 = b2;
    bw3 = b3;
end
end
display(min_error)

Answer 1

[Walter Roberson]

It is correct that you would have to rewrite your code to use GPU. At the very least you would have to make some of your computations into gpuarray() .

However, your operations are too small to be worth doing on the GPU. GPU is best at doing the same thing to many data points, and at its worst when doing short calculations, because of the communications and synchronization overheads.

You would, I think, do much better to vectorize your calculations.

In particular, R2020b and later have pagemtimes() https://www.mathworks.com/help/matlab/ref/pagemtimes.html

so you could

N = 10000;
w1 = rand(3,3,N); 
b1 = rand(1,3,N); % first bias , 1 value going to 3 neurons

zh1 = pagemtimes(x, w1) + b1;
a2 = logsig(zh1);

and similar operations. At the end you would get a vector of err, and would min() that extracting the index and using the index to pull out the appropriate page from w1, b1, and so on.

If that does not in itself speed things up enough, then those vectorized calculations would be suitable for processing on the GPU.

[Abdulaziz Alotaibi]

Thanks a lot , I'm just getting started so I thought it's a good idea to use GPU instead of CPU because in further application i'm gonna biuld a CNN model and train it and so on.

GPU will be much faster in that kind of applications right?