vectorize nested loop using imcrop for gpu usage

Jason Dortch

2018 3 7

0 个回答

5 次查看（30 天）

0 个投票

Hello

I am working with a large dataset (2e+12) and need to clip ~200,000 small matrices (~30*30 pixels) out. I have a functioning nested loop but would like to take advantage of vectorising the code for dual 6 core cpus OR and even taking advantage of the quadro 4000 (8gb ram on card) in my desktop. The large dataset is a DEM cut into several parts and stored in a multidimensional matrix [MD_DEMs]. I thought this would allow gpuArray to be utilized by keeping file size down.

I tried to vectorize a couple of different ways, but can’t seem to get imcrop to play nicely. Any help is appreciated.

    S=3; % changeable depending on matrix size and available ram
%fake datasets
MD_DEMs=rand(1000,1000,S); % random matrix
M{1}=[1 546 218 29 29 % made up locations for imcrop
        2   128 51  29 29 % [label Xmin Ymin width height]
        3   146 58  29 29
        4   5   253 29 29
        5   45  18  29 29];
M{2}=[6 5   850 29 29
        7   546 218 29 29
        8   5   12  29 29];
M{3}=[9 5   2   29 29
        10  546 218 29 29];
MD_Clips=zeros(30,30,10);  % preallocate DEM clips
MD_Lists=zeros(10,1);  % preallocate the matrix for row identifier (label)
%code snipet
k=0;
for i=1:S;
    [m,~]=size(M{i});
    for j=1:m;
        k=k+1; % counter so data is not overwritten
        MD_Clips(:,:,k)= imcrop(MD_DEMs(:,:,i),M{[i]}(j,2:5));
        MD_Lists(k,1)= M{[i]}(j,1);
    end
end