How to vectorize three for loop or fast

Question

Eb Bed 2018-6-11

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/405110-how-to-vectorize-three-for-loop-or-fast

评论： Jan 2018-6-12

% this is to do the grid-based central finite difference for the two vectors dimension as u(lon, lat, time) and v(lon,lat,time). the for loops works fine but it is too slow. I need a helpful loop that is faster than me from Matlab community.

    Ur = zeros(291,243,1459);
    [NR, NC]=size(lat);
    for k = 1: length(t)
    for i = 2:NR-1
    for j = 2:NC-1
         Ur(k,i,j)=sqrt((u(i+1,j,k)-2*u(i,j,k)+u(i-1,j,k))./(lon(i)-lon(i-1)^2) ...
                 +(v(i,j+1,k)-2*v(i,j,k)+v(i,j-1,k))./(lat(j)-lat(j-1)^2));
    end
    end
    end

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Geoff Hayes 2018-6-11

Eb - please clarify/quantify what you mean by too slow. Does this take seconds, minutes or hours to complete? Have you considered deferring the square root until after Ur is populated for all i,j,k? (Just to remove one computation that isn't strictly necessary until later...)

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Jan 2018-6-11

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/405110-how-to-vectorize-three-for-loop-or-fast#answer_324197

编辑：Jan 2018-6-11

在 MATLAB Online 中打开

Avoid repeated calculations.
Process the data in column order.

Ur = zeros(291,243,1459);
[NR, NC]=size(lat);
for j = 2:NC-1
   latj = lat(j) - lat(j-1)^2;
   for i = 2:NR-1
      loni = lon(i) - lon(i-1)^2;
      for k = 1: length(t)
         Ur(k,i,j) = sqrt((u(i+1,j,k) - 2 * u(i,j,k) + u(i-1,j,k)) ./ loni ...
                        + (v(i,j+1,k) - 2 * v(i,j,k) + v(i,j-1,k)) ./ latj);
      end
   end
end

If you provide all input arguments, e.g. created by some rand() command, we could check the speed. I assume the inner loop is easy to vectorize:

Ur = zeros(291,243,1459);
[NR, NC]=size(lat);
for j = 2:NC-1
   latj = lat(j) - lat(j-1)^2;
   for i = 2:NR-1
      loni = lon(i) - lon(i-1)^2;
      Ur(:,i,j) = sqrt((u(i+1,j,:) - 2 * u(i,j,:) + u(i-1,j,:)) ./ loni + ...
                       (v(i,j+1,:) - 2 * v(i,j,:) + v(i,j-1,:)) ./ latj);
   end
end

A further vectorization is possible, but I hesitate to implement it: It looks strange, that you treat lat as 2D in [NR, NC]=size(lat), but access the first column only by using 1 index later. Does the code really do, what you want?

2 个评论
显示无隐藏无

Eb Bed 2018-6-11

Daer Jan, you figure out it for me. My data formats are both lat 2D and 1D. for 1D is it more efficient code?? If there is I am happy to hear you. Thanks

Jan 2018-6-12

在 MATLAB Online 中打开

@Eb Bde: As said already, if you provide the input arguments such that this piece of code runs, it can be improved further. All you need to do is to post some rand commands to create lat, u and v:

lat = rand(...)
u   = rand(...)
v   = rand(...)

Please replace the ... with the real dimension. Of course I could guess some values, but I'm unsure due to the suspicious indexing and I do not understand: "My data formats are both lat 2D and 1D".

请先登录，再进行评论。