Julia equivalent of Python's "fsum" for floating point summation

[mbauman]

This is so great.

@ffevotte

It looks like, on some systems, the linear summation (that occurs in the leafs of the pairwise summation tree, when blocks are sufficiently small) is vectorized. But on some other systems, it is not. And I could not determine any pattern to say what caused the vectorization to happen or to fail.

That’s entirely possible. Newer chips have much more capable SIMD vectorization units. The newest instruction sets are AVX2 (most new laptop chips) and AVX512 (only on beefy Xeon server chips). You can check which instruction sets your chips implement in the “flags” fields of /proc/cpuinfo (on Linux) or sysctl -a | grep machdep.cpu.features (on macOS).