Does Matlab automatically vectorize for loops?
7 次查看(过去 30 天)
显示 更早的评论
I currently have Python code that takes 1.5 hours to calculate 450+ million Levenshtein distances between 30,011 text labels of no more than 20 characters each. Currently migrating it to Matlab to see the JiT compilation allows optimizations of flow control. The Python implementation uses Anaconda's CPython, so no JiT, and there is a distinct possibility that Matlab could be much faster.
Years ago, I recall reading that Matlab automatically vectorizes for loops. I cannot find any such mentions now. Is this just assumed to be part of the JiT compilation?
There will be 2 levels of for-loops, one nested within the other, to iterate over matrix rows and columns. Each iteration will call editDistance, so I don't know if that will prevent vectorization, even if automated vectorization was available.
0 个评论
采纳的回答
Walter Roberson
2024-6-13
Years ago, I recall reading that Matlab automatically vectorizes for loops. I cannot find any such mentions now. Is this just assumed to be part of the JiT compilation?
That is not done. (Or if it is done, it is in fairly narrow circumstances.)
0 个评论
更多回答(1 个)
Steven Lord
2024-6-12
In general we don't discuss what the JIT does internally as we don't want users to try to code against that always-moving target. [What if you write your code in a somewhat "unnatural" way to avoid some JIT limitations and the next release we remove those limitations and make the more "natural" implementation much faster? Would you ever notice the change?]
Searching the online documentation for the term "Levenshtein" found four functions, two each in Text Analytics Toolbox and Computer Vision Toolbox, whose reference pages include that term. One or more of those may be useful to you. There are also a few hits in MATLAB Answers and in the File Exchange, if you look at the search results without the scoping to just functions in MathWorks products.
4 个评论
Paul
2024-6-14
editDistance is vectorized from a user interface perspective, but it's not really. As best I can tell, it basically uses arrayfun under the hood to call itself in a loop for all of the pairs of scalars in the input. And on each call it does argument validation and various other checks that really only need to be done once. I haven't done any testing, but I suspect editDistance could be modified to be (much?) more efficient.
另请参阅
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!