function ann = my_nnmex(A, B)
% function ann = my_nnmex(A, B,~, ~, ~, ~, ~, ~, ~, ~, ~,~,nnfield_prev)
%function ann = nnmex(A, B, [algo='cpu'], [patch_w=7], [nn_iters=5], [rs_max=100000], [rs_min=1], [rs_ratio=0.5], [rs_iters=1.0], [cores=2], [bmask=NULL]...
% [win_size=[INT_MAX INT_MAX]], [nnfield_prev=[]], [nnfield_prior=[]], [prior_winsize=[]], [knn=1], [scalerange=4])
%      (Pass [] to leave a parameter at its default).

%ann = nnmex(res, img,  algo, [], [], [], [], [], [], cores, [], [], ann);

% ann size = B size

% BEGIN DEBUGGING CODE
% clear;
% clc;
% close all;
% tic
% image = imread('SimakovFarmer.png');
% imageLab = rgb2lab(image); % Convert the source and target Images
% imageLab = double(imageLab)/255;
% targetHeight = 35;
% res = imresize(imageLab,[targetHeight,NaN]);
% stage1 = res;
% H = 35;
% % disp('Retargeting at the coarsest scale:')
%
% W = 30;
% fprintf('\nH=%d, W=%d\n',H,W);
% stage2 = imresize(stage1,[H,W]);
% A = stage2; % small image
% B = res; % large image

% END DEBUGGING CODE


sizeA = size(A);
sizeB = size(B);
w = 7;
w1 = w-1;
xALimit = sizeA(2) - w1;
yALimit = sizeA(1) - w1;
xBLimit = sizeB(2) - w1;
yBLimit = sizeB(1) - w1;
cost = zeros(yBLimit,xBLimit);
ann = zeros(size(A));
ann(:,:,3) = ones(sizeA(1),sizeA(2))*2147483647;
% class(A)
pause(0.01)
% if (~ isa(A,'double'))
%    A = double(A)/255;
% end
% maxA = max(A(:))
% maxB = max(B(:))
% class(A)
% class(B)
A255 = A*255;
B255 = B*255;

% patchB_cache = cell(yALimit,xALimit);
% for xB = 1:xBLimit
%     for yB = 1:yBLimit
%         patchB_cache{yB,xB}= B255(yB:yB+w1,xB:xB+w1);
%     end
% end
% patchB_cache(yALimit,xALimit) = struct('patch',zeros(w,w));
% patchB_cache(1:yALimit,1:xALimit) = struct('patch',zeros(w,w));
patchB.p=zeros(w,w);
patchB_array = repmat(patchB,yBLimit,xBLimit);

% patchB_cache(yALimit,xALimit)=struct();
for xB = 1:xBLimit
    for yB = 1:yBLimit
          patchB_array(yB,xB).p = B255(yB:yB+w1,xB:xB+w1);
    end
end
% tic
for xA = 1:xALimit
%     fprintf('xA=%d\n',xA)
    disp(xA)
    for yA = 1:yALimit
       patchA = A255(yA:yA+w1,xA:xA+w1); % small image
        for xB = 1:xBLimit
            for yB = 1:yBLimit
%                 patchB = B255(yB:yB+w1,xB:xB+w1); % large image
%                 costM = (patchB - patchA).^2;
%                 patchB_array(yB,xB).p
                costM = (patchB_array(yB,xB).p - patchA).^2;

                cost(yB,xB) = sum(costM(:));
            end
        end
        [minVal,minLoc] = min(cost(:));
        [locY,locX]=ind2sub(size(cost),minLoc);
        ann(yA,xA,1)=locX - 1;
        ann(yA,xA,2)=locY - 1;
        ann(yA,xA,3)=minVal;
%         ann(yA,xA,3)=(locY-yA)^2 + (locX - xA)^2;
%         ann(yA,xA,3)= sum(costM(:))
%         d = (patchB - patchA)
%         ann(yA,xA,3)= d*d'
    end
end
% toc

ann = int32(ann);
% ann2 = nnmex(A, B, [], [], [], [], [], [], [], []);
% whos ann
% pause(0.01)
% toc
end