Accelerating MATLAB Performance

479Memory-Related Techniques

for yIdx = 1:size(data,2)

total = total + data(xIdx,yIdx,zIdx);

end

end

end

⇨ Elapsed time is 0.665435 seconds.

% Much better - permute dimensions: loop over z,y,x (5x faster)

data = permute(data, [3,2,1]); % in-place to prevent memory spikes

for zIdx = 1:size(data,3)

for yIdx = 1:size(data,2)

for xIdx = 1:size(data,1)

total = total + data(xIdx,yIdx,zIdx);

end

end

end

data = permute(data, [3,2,1]); % reshape back to original

⇨ Elapsed time is 0.137673 seconds.

Loop order effects can also be seen with internal MATLAB code, even in cases where we

might expect implicit parallelization to negate such effects. For example, ( B'*A')' might

be faster than the equivalent A*B., if A has many m