{"id":2080,"date":"2016-08-01T09:26:20","date_gmt":"2016-08-01T13:26:20","guid":{"rendered":"https:\/\/blogs.mathworks.com\/steve\/?p=2080"},"modified":"2019-11-01T16:44:44","modified_gmt":"2019-11-01T20:44:44","slug":"ineffective-preallocation","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/steve\/2016\/08\/01\/ineffective-preallocation\/","title":{"rendered":"Ineffective preallocation"},"content":{"rendered":"

Today I want to explain a MATLAB coding mistake that I have seen even experienced MATLAB users make.<\/p>

I was looking at some code on the File Exchange recently, and these lines caught my eye:<\/p>

cm_data = zeros(m,3);\r\nhsv=rgb2hsv(cm);\r\ncm_data=interp1(linspace(0,1,size(cm,1)),hsv,linspace(0,1,m));\r\ncm_data=hsv2rgb(cm_data);<\/pre>
Specifically, I was struck by the first and third lines. The first line creates an all-zeros matrix. Then the third line throws away the matrix created by the first line, replacing it with the output of interp1<\/tt>.<\/p>
I have seen code like this before. It results from a misunderstanding of the concept of preallocation<\/i> in MATLAB.<\/p>
In MATLAB, preallocation refers to creating a matrix (usually with zeros<\/tt>) prior to entering a loop that would otherwise repeatedly resize the matrix.<\/p>
Here is an example. In the for-loop below, the size of the matrix x<\/tt> is increased by one element each time through the body of the loop.<\/p>
x = 0;\r\nfor<\/span> k = 2:1000000\r\n   x(k) = x(k-1) + 5;\r\nend<\/span>\r\n<\/pre>
In versions of MATLAB before R2011a, this kind of loop could be quite slow. The reason is that MATLAB would reallocate the memory space for the matrix x<\/tt> each time through the loop, resulting in $O(N^2)$ execution time overall. To avoid this performance penalty, the usual recommendation was to create the full-size matrix x<\/tt> before entering the loop, like this:<\/p>
x = zeros(1,1000000);\r\nfor<\/span> k = 2:1000000\r\n   x(k) = x(k-1) + 5;\r\nend<\/span>\r\n<\/pre>
Although automatic array growth got much more efficient in MATLAB R2011a<\/a> (it is no longer $O(N^2)$), there is still some performance benefit to preallocation, which is why it is still described in the documentation<\/a>.<\/p>
In the code from the File Exchange, though, the matrix cm_data<\/tt> is not being grown in a loop. The result of the initial assignment, cm_data = zeros(m,3)<\/tt>, is just discarded without ever being used.<\/p>
The MATLAB Editor tries to tell you about this. Do you see the orange squiggle on the first line below?<\/p>
\"\" <\/p>
If you hover over the squiggle with the mouse, you see this:<\/p>
\"\" <\/p>
And if you then click on the Details button, you get a full explanation:<\/p>
\"\" <\/p>
In summary, don't worry about putting in that call to zeros<\/tt> unless you are growing a matrix in a loop. If you later replace the variable with a completely new matrix returned by another function call, then the call to zeros<\/tt> is not helping you, and it might even slow down your code.<\/p>