A beginners guide to using Python for performance computing

> Finally, for comparison we implemented this in simple C++ (nothing fancy) without any Python. One would expect that the C++ code would be faster but surprisingly, not by much! Given the fact that it's so easy to develop with Python, this speed reduction is not very significant.

This is the problem I have with Python+Scipy v. MATLAB/C++/YourFavoritePlatform comparisons. I agree that Python code can be clean and clear, but once you start to worry about vectorizing, inlining, manual loop optimizations, or a C or Fortran FFI, then you end up with the same solution that say MATLAB users have to deal with by writing mex functions in C.

Granted there are differences between Language A v. Language B, one may be more concise or their functions are _even higher-order_ than the other's. Point is, once you go down the road of "let's optimize this dynamic language script," your choices are not that different. You lose the "ease" of developing with Python, and if you really performance then it's ultimately inferior to just writing it in C++ or even Fortran.

Many times performance computing is about 1% optimizable code and 99% setup. Python is much lovelier for that 99% than matlab is.

I often wonder why there is so little interest in a high level language that would compile to pure C++, with a syntax that makes code more succinct and friendlier, just like CoffeeScript compiles to JavaScript. We have a clear problem: higher level languages make programmers more productive, but they have performance issues. Because of those, developers that need performance fall back to C++ (look at videogame development, C++ is there to stay).

So why not make a language that is high level, but compiles to pure, cross platform, C++? Think of the productivity win: a higher level language that rarely needs optimizations. Obviously it would have to come with some lower level semantics, like memory management (although it could maybe be solved somehow like Apple did with Automatic Reference Counting, which is basically a preprocessor) or static typing (although the compiler can sometimes be instructed to guess the type, see the := operator in ooc [1]), but even if some lower level semantics are unavoidable, it still seems like a huge win. Yet every time I saw the idea mentioned online it's been either mostly ignored or treated as stupid idea, and the very few projects that attempted to go in that direction never took off. Maybe I'm missing something? If so I'd love to know what it is.

No matter how fancy and awesome higher level languages are, we keep going back to C++ for performance, and that's probably never going to change (at least until processors are stupidly fast). So why not make it easier? The CoffeeScript approach has largely been proven to work.

[1] http://docs.ooc-lang.org/language/syntax.html#declarations (note that ooc is a good step in this direction, it compiles to C99, but then takes step back: it depends on a garbage collector)

The weave bits, and Pyrex bits, found in this article are actually snippets of C/C++ written inline in the Python code (or very simplified Python), and able to access variables and such defined in the Python earlier (with caveats and when used with care), and later Python code is able to use the resulting compiled functions.

So, it's high level Python code, except when you need to write low level, super fast code.

The reality of making a very high level language compile down to high performance C/C++ is a much harder problem than "compiling" CoffeeScript to JavaScript. Both are quite high level, and CoffeeScript is merely very concise syntactic sugar for a number of common patterns in JavaScript, and maybe a few higher level constructs tacked on for good measure. But, JavaScript is effectively Scheme with C-like syntax; it was designed to be built up in this way, just like Lisp and Scheme (and it's only the lack of control over the parser and syntax that forces it to be a compiler at all; all the capabilities are there in JavaScript: closures, first class functions, code as data, etc.). Building DSLs is what this kind of language is for. C and C++ is not for building DSLs, and you don't magically get C performance by converting high level code to low level code.

One example is B::C, a Perl to C compiler: http://www-rohan.sdsu.edu/doc/perldoc-html/B/C.html

In short, the CoffeeScript approach is not at all comparable to a high-level to low-level compiler. It is comparable to a, say, Lua to Perl compiler, perhaps. From one very high level language to another very high level language.

It's not lack of interest. Weave and Pyrex, which are discussed in this article, are proof positive that people want the speed of low level and the convenience of high level languages. It's just not possible to get it via the means you've described. JITs seem to be the currently fashionable way to get closer to that goal, though it's still miles away.

Oh, and there are Lisp compilers that compile to lower level languages like C or compile very fast Lisp. Lisp is quite high level. So, when it goes fast, it is impressive. I'm not knowledgeable enough to know whether they could be used in the same contexts of SciPy, Weave, Pyrex, etc. I imagine there are big math libraries for Lisp, though...and they're probably pretty fast.

As for ooc and garbage collection: How would you write a compiler for a high level garbage collected language without including a garbage collector? And, do you believe that garbage collection is the primary reason dynamic languages are slower than compiled non-GC languages? (Hint: It is not. GC is very far down the list of resource users in every garbage collected language I'm aware of.)

I have done that in the past. But writing a mex function is a lot more tedious than writing a weave.inline string. Think of all the type conversions you have to put up with in mex. Think about how you have to set up a compile infrastructure and how you have to recompile manually for every target platform.

In most cases, I will use weave.inline only for the innermost loop, and all the compilation issues and type conversions will be taken care of for me.