Numba was started in 2012 by Travis Oliphant, the original author of NumPy, as a library for compiling individual Python functions at runtime using the Low-Level Virtual Machine (LLVM) toolchain.
LLVM is a set of tools designed to write compilers. LLVM is language agnostic and is used to write compilers for a wide range of languages (an important example is the clang compiler). One of the core aspects of LLVM is the intermediate representation (the LLVM IR), a very low-level platform-agnostic language similar to assembly, that can be compiled to machine code for the specific target platform.
Numba works by inspecting Python functions and by compiling them, using LLVM, to the IR. As we have already seen in the last chapter, the speed gains can be obtained when we introduce types for variables and functions. Numba implements clever algorithms to guess the types (this is called type inference) and compiles type-aware versions of the functions for fast execution.
Note that Numba was developed to improve the performance of numerical code. The development efforts often prioritize the optimization of applications that intensively use NumPy arrays.
Numba is evolving really fast and can have substantial improvements between releases and, sometimes, backward incompatible changes. To keep up, ensure that you refer to the release notes for each version. In the rest of this chapter, we will use Numba version 0.30.1; ensure that you install the correct version to avoid any error.
The complete code examples in this chapter can be found in the Numba.ipynb notebook.