Perform mldivide between 3x3 matrix M and every RGB pixel in a image in GPU
3 views (last 30 days)
I have a rgb image (2048x2048x3), and I wan to perform mldivide on every rgb (3,1) pixels with a M matrix (3x3).
So each pixel will do a M\pixel operation. The easy way to do this is using two for loops to go over each pixels. However, I am wondering if I can use arrayfun or pagefun for this so I can use GPU.
Right now in order to use arrayfun, I have to hard code mldivide so M becomes 9x4194304(2048*2048) array which takes lots of vram. I am wondering if there is a way I can run this without having a big M gpu array as input.
Stephen23 on 8 Jul 2022
S = 5;
M = rand(3,3);
I = rand(S,S,3);
A = nan(S,S,3);
for ii = 1:S
for jj = 1:S
pixel = reshape(I(ii,jj,:),3,1);
A(ii,jj,:) = M\pixel;
Remember that RESHAPE is a very fast operation (it does not change your data in memory, only the shape meta-data in its header). In contrast you want to avoid operations that rearrange your data in memory (e.g. TRANSPOSE, PEMUTE). Note that using PAGEFUN would require permutimh your data.
% RESHAPE() and TRANSPOSE() and MLDIVIDE():
B = reshape((M\reshape(I,S*S,3).').',S,S,3)
% we can simplify the above into RESHAPE() and MRDIVIDE():
C = reshape((reshape(I,S*S,3)/M.'),S,S,3) % recommended
More Answers (2)
Joss Knight on 9 Jul 2022
Edited: Joss Knight on 10 Jul 2022
I feel like I'm missing something - this is just a single backslash with multiple right-hand sides, or to avoid permutation a single mrdivide (edit: you still have to transpose M, but that's very quick):
[h,w] = size(im,[1 2]);
imout = reshape(reshape(im,,3)/(M.'), h, w, 3);