Speed up a program

Question

Vlad 2014-12-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/165428-speed-up-a-program

评论： Vlad 2014-12-5

It is required to speed up a very simple program:

for i=1:1000
   for j=1:1000
      A=fun(i, j) % matrix 3x3
      D=eig(A);
   end
end

Execution time for fun(i, j) is much less then time for eig(A). A mean execution time of whole program is about 40 minutes. I wrote a mex-file for calculating eigenvalues and improved the speed by 4 times. So now the mean time is about 10 minutes. I tried parfor-loop, it did not work. CUDA did not help as well =(

I will be very appreciated for any idea.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Zoltán Csáti 2014-12-4

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/165428-speed-up-a-program#answer_161275

Why didn't parfor and CUDA work? Do you meet the requirements?

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Vlad 2014-12-5

I suppose, that matrix 3x3 is very small for gpu-calculations, that's why there is no advantage of using it.

请先登录，再进行评论。

Answer 2

Edric Ellis 2014-12-5

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/165428-speed-up-a-program#answer_161365

Here's what I tried with PARFOR, and this worked:

function D = pfeg
N = 100;
D = zeros(N, N, 3);
parfor i = 1:N
  for j = 1:N
      A = fun(i, j);
      D(i, j, :) = eig(A);
  end
end
end
function A = fun(i, j)
A = rand(3) .* i + j;
end

I'm not surprised that the GPU didn't help much - 3x3 matrices are simply too small to take good advantage of the GPU capabilities (unless you can use pagefun - which unfortunately you cannot in this case).