How can I do faster a loop using the GPU? (GPU: RTX 3070)

n=700;

A=rand(n,n);

%%

Ms=gpuArray(M);

As=gpuArray(A);

% First questions: this loop takes to much time. How can I do this loop using the cores available in the GPU

tic

for i=1:100;

AA(:,:,i)=As; %this is just an example, in practice, As is different for each i of the loop

MM(:,:,i)=Ms; %Ms is the same for all the indexes i.

end

toc

%I m using this array to make 2 multiplications and a pointwise produt

% This are the multiplications

AGpage=gpuArray(AA);

Fs=gpuArray(fftshift(dftmtx(m)));%

Gs=gpuArray(fftshift(dftmtx(n)));

CG=pagefun(@mtimes,AGpage,Fs);

Hs=pagefun(@mtimes,Gs,CG);

% what I need next is this result: sum(Ms.*Hs,[2 1]);

%But, I use a reshape to use all the cores in the pointwise multiplication

Ms2=gpuArray(MM);

MMT=reshape(Ms2,l,l,49*100);

ST=reshape(Hs,l,l,49*100);

for iter=1:30

aG2=sum(MMT.*ST,[2 1]);

%%

% I need to go back to recover the sum of the pointwise multiplication sum(Ms.*Hs,[2 1]);

%First questions: But now, again this loop is too slow

for index=1:100

tramo=aG2((index-1)*49+1:49*index);

aG3(index)=sum(tramo);

end

%how can I do faster a loop using the GPU?

Answer 1

[Ishu]

Hi Sandra,

I understand that you want to make the comutation for "for" loop faster. To make the computation faster you can use Parallel computation in MATLAB. For this you have to replace "for" loop with "parfor", this will execute for-loop iterations in parallel on workers and making your computations faster.

In addition to this, you can improve code execution time by preallocating the maximum amount of space required for the arrays.

%Example 
x = 0;
for k = 2:10000
   x(k) = x(k-1) + 5;
end

% preallocating x for faster computations
x = zeros(1,10000);
for k = 2:10000
   x(k) = x(k-1) + 5;
end

For more information you can refer below documentation:

parfor: https://www.mathworks.com/help/parallel-computing/parfor.html

preallocation: https://www.mathworks.com/help/matlab/matlab_prog/preallocating-arrays.html

Hope it helps!