Is there a way to run this loop in gpu and will it consume lesser time compared to parallel cpu computation time?

N=50;
for n=1:8
    b=rand(300,103);
    w=rand(300,1);
    [Q,R]=qr(b,0);
    A(:,:,n);=R
end

[Joss Knight]

I don't want to put this as an answer to your question because it might seem like an obvious point: have you tried using gpuArray data to see if things go faster? Something as simple as b=rand(300,103,'gpuArray') would get you started.

[KSSV]

First place...is the code working ?

[Naveen kumar Elumalai]

Yeah, at final step we converted A to 3d array back to 2d array. Gpu Array direct usage is little slow since the loop above mentioned, contains for-loop to compute multiplication of every column of matrix A.

[Joss Knight]

But you're not doing a multiplication, you're doing a QR decomposition. Decomposing a 300-by-103 matrix should be sufficient computation that you get an advantage using the GPU, even when you're doing it 8 times.

[Naveen kumar Elumalai]

yes, you are right but i have code like this now

D=rand(300,52)+i*rand(300,52);

B =rand(300,36);
F=rand(300,36)+i*rand(300,36);

A=zeros(300,103);

for n=1:36

      for m=1:52
        A(1:300,m)=B(1:300,n).*D(1:300,m);
      end

      for m=1:51
        A(1:300,52+m)=-B(1:300,n).*D(1:300,m).*F(1:300,n).';    
      end

      A=[real(A);imag(A)];

      [Q,R]=qr(A,0);
    AA(:,:,n)=R
  end

I am not sure what should be done in this case, eventhough i do a gpuarray its taking a same time to that of the normal cpu . Is there a way to compute the whole main for loop in one shoot in gpu ?

[Joss Knight]

Okay, well this

      for m=1:52
        A(1:300,m)=B(1:300,n).*D(1:300,m);
      end

is equivalent to this

A = B(1:300,n).*D;

If you have enough memory you could even vectorize the outer loop since this

for n=1:36
    for m=1:52
      A(1:300,m)=B(1:300,n).*D(1:300,m);
    end
end

is equivalent to this

A = reshape(B,size(B,1),1,size(B,2)) .* D;

and then you'd access A page-by-page, like

[Q,R] = qr(A(:,:,n),0);

[Naveen kumar Elumalai]

Could I be able to access it at once, instead of using page by page

%code

for n=1:36

[Q(:,:,n),R(:,:,n)]=qr(A2(:,:,n),0);

end

I wanted this to happen without for loop. Here A2, is 600*103*36,

R needs to be of the size 103*103*36 and

Q of the size 600*103*36

Any answer would be appreciated. Kindly do reply

[Joss Knight]

No, you cannot. See answers to the equivalent questions from you and others e.g.

https://uk.mathworks.com/matlabcentral/answers/391752-i-need-to-do-this-without-for-loop-for-gpu-computing

https://uk.mathworks.com/matlabcentral/answers/391345-how-to-convert-this-without-using-for-loop-for-gpu-computing

Answer 1

[Naveen kumar Elumalai]

This is taking so much time than the normal cpu execution