Benchit | Minimalist benchmarking library for Android

There are several issue with the Numba implementation. One of them is that parallel loops breaks the constant propagation optimization in LLVM-Lite (the JIT-compiler used by Numba). This cause critical information like array strides not to be propagated resulting in a slow scalar implementation instead of an SIMD one, and additional unneded instructions so to compute the offsets. Such issue can also be seen in C code. Numpy added specific macros so help compilers to automatically vectorize the code (ie. use SIMD instructions) when the stride of the working dimension is actually 1.

A solution to overcome the constant propagation issue is to call another Numba function. This function must not be inlined. The signature should be manually provided so the compiler can know the stride of the array is 1 at compilation time and generate a faster code. Finally, the function should work on fixed-size chunks because function calls are expensive and the compiler can vectorize the code. Unrolling the loop with shifts also produce a faster code (although it is uglier). Here is an example: