How To Write Fast Numerical Code: Fill & Download for Free

Personally after some years in C# indeed I’m considering going back to C++. Not only has it become as easy to write as C# (vectors, the new for loops, lambdas, shared_ptr, etc.), but it’s actually much easier in some regards (for example, try implementing a formula that works with all numeric types in C++ and in C# and you’ll see - in C++ you can just use a template, but you can’t use a generic in C# as there is no common interface to all numeric types and you can’t get access to the arithmetic operators).And, while performance isn’t always a top concern… who would say “no” to a 2x-10x improvement, given that the effort to write the program is the same (or smaller; at least to me writing fast C++ seems easier than writing fast C#, you don’t need to know how the entire .NET works just to be able to bang out some fast code).

Julia used to be my go to solution for fast, custom, numerical code that I could call from Python if I didn't want to write it in C or Cython. The LLVM JIT compiler is built in to Julia, and when it came out, I was pretty happy. Suddenly it was fun and fast to write something custom. Having to switch to a different syntax kind of sucked, including 1- indexing.And then came Numba for Python.With a single autojit decorator, I can transform my proof of concept code into something faster by compilation to LLVM. If I want to make it even faster, I can specify specific constraints that basically make things as fast or faster than calling Julia from Python.And that's all using Python syntax, with some special decorator knowledge handy.I don't use Julia anymore. Since most of my work for the past while has needed to be in Python for various reasons, except for custom code, there's simply been no need.

It’s still kind of niche, but it is really good as a platform for writing and using scientific numerical packages. Most languages used for that task are only good for using them (Python, R, Matlab, etc) but terrible for writing them, or they are okay for writing them (C++) but very non-ergonomic for using them.Julia has the advantage of being amazingly expressive for writing fast numerical packages, and at least as good as any other scripting language for using them. As long as it maintains a decently large community for a long enough time, it will eventually have a bigger ecosystem of quality packages simply because writing them in Julia and exposing a good interface to users is so much easier. It already has best-in-class packages for some specific domains, and the situation will probably just keep getting better with the recent Julia 1.0 release.You also get more composability between packages and it is easier to write packages that depend on other packages (for example, a differential equations package can use custom high-precision float types from another package without needing any change to its code as long as they implement all required methods).