How to reduce the computation time for adding 3D-array?

2 次查看（过去 30 天）
jae lee
jae lee 2020-12-9
评论： Bruno Luong 2021-1-7
Hi, I am trying to add multiple 3D-arrays to a bigger 3D-array at a specific index (x,y,z)
Below is the code, and it does work and compute the answer but it seems very ineffecient.
In this example, i only have 4 sets of coordiantes (x,y,z) but in real code, i have more than 1e6 sets of points.
It takes very long to compute the result with that many points.
Is there any way to reduce the computation time?
Thank you in advance
Regards
J
Big=zeros(500,500,500); %%%% Bigger Array
Small=rand(250,250,250);
x=[245; 220; 256; 270];
y=[245; 220; 256; 270];
z=[245; 220; 256; 270];
for n = 1 : length(x)
x_cord=x(n)-length(Small)/2;
y_cord=y(n)-length(Small)/2;
z_cord=z(n)-length(Small)/2;
x_end= x_cord + length(Small) -1 ;
y_end= y_cord + length(Small) -1;
z_end=z_cord + length(Small) -1;
if x_end <= length(Big)
Big(x_cord:x_end, y_cord:y_end,z_cord:z_end)=Big(x_cord:x_end, y_cord:y_end,z_cord:z_end)+Small();
end
end
  2 个评论
Walter Roberson
Walter Roberson 2020-12-9
could also be done with accumarray, but I am not sure that would be faster considering the time to generate the coordinate matrices... though I did just think of a shortcut for that.
jae lee
jae lee 2020-12-9
Hi Walter. Thank you for your comment.
What would be the coordinate matrices??
Does it mean x,y,z?
Thank you

请先登录，再进行评论。

请先登录，再回答此问题。

采纳的回答

Bruno Luong
Bruno Luong 2020-12-24
编辑：Bruno Luong 2020-12-24
You can reformulate the loop as convolution of
A = accumarray([x(:) y(:) z(:)]-length(Small)/2,1,[500 500 500])
and
B = flip(flip(flip(Small,1),2),3)
(I left out the detail of overflowed for simplicity)
You might look at convn function. The problem I see is that A is sparse (1/125 in density) and might not be efficient as for loop.
You also might try this FEX to see if you can exploit the sparsity
https://www.mathworks.com/matlabcentral/fileexchange/41310-sparse-matrix-convolution
Or this one using different method of compute convolution
https://www.mathworks.com/matlabcentral/fileexchange/24504-fft-based-convolution
  5 个评论
显示 3更早的评论
jae lee
jae lee 2021-1-7
I tried the above method but it went into the infinite loop.
A = accumarray([x(:) y(:) z(:)]-length(Small)/2,1,[500 500 500])
B = flip(flip(flip(Small,1),2),3)
C=convn(A,B);
Bruno Luong
Bruno Luong 2021-1-7
You might try the alternative convolution implementations in the links I post above.
As I said above, the alternative ways I have proposed might not be fater than your for-loop.

请先登录，再进行评论。

更多回答（1 个）

Amrtanshu Raj
Amrtanshu Raj 2020-12-24
Hi,
You can use the parfor loop to use parallel processing and get higher computation speeds. However you will have to modify your for loop to be used for parfor loop.
Hope this helps !!

类别

在 Help Center 和 File Exchange 中查找有关 Loops and Conditional Statements 的更多信息

标签

产品


版本

R2020b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by