Optional compile flags

In response to #920, and following the discussion in #949, we need some way to make the compile flags optional during setup. The -march=native is the probable cause of issues with the conda-forge build #647, and also causes problems for some installing in a cluster environment.

However, for most people installing on a local machine, -march=native is likely to have significant performance advantages. The suggestion is then to have some cmd line option to choose compile flags where the default is ‘optimal’, which would include -march=native.

@Ericgig we can discuss this next week

[slavoutich]

I might oppose to the fact that it should be default. The reason is that majority of people don’t read the documentation and don’t know what “segfault: illegal instruction” means. It might better to note somewhere that CFLAGS="-march=native" pip install qutip will lead to 30% increased performance in certain scenarios.

[jakelishman]

For future posterity: we simply can never issue -march=native commands to build wheels for distribution via pip or conda; we have to have these appeal to the lowest common denominator instruction set. We have hard-coded SSE2 instructions in zspmv.cpp, but these are ancient, and that’s probably the baseline we need to support at least for now (it’s been working without issue for a few years now). Computing clusters don’t have the latest and greatest hardware, but usually are just a weird amalgam of other bits.

The CFLAGS changes were merged already, so I’m closing this now.