OpenMP option default setting

[Tectu]

While preparing the nextpnr-0.9 release for FreeBSD, I came across the CMake option USE_OPENMP which appears to be OFF by default.

What is the reason for this to be off by default?
Is this considered an "experimental" or "unstable" feature? Or could I just set it to ON in the FreeBSD port without any worries?

[Ravenslofty]

My understanding is that enabling OpenMP is meant to help accelerate the Eigen matrix solver used in the heap placer. Honestly, I don't know if it actually does anything; if it does, I think there might be another win possible by enabling wider SIMD through compiler flags.

(But writing CMake to do this terrifies me.)

[whitequark]

I can write the CMake for it.

[Ravenslofty]

If you wish to do so, I would perhaps start with detecting AVX512 support, and enabling it through a mechanism like -march=x86-64-v4 or so. If there's a measurable gain for enabling SIMD support, we should find it there.

(and if not...perhaps we have a performance improvement available)

[Tectu]

I have always been using nextpnr with -march=broadwell (equivalent to -march=x86-64-v3) and I have never experienced a single issue with it.

[Ravenslofty]

The question isn't whether it breaks anything, the question is whether it improves anything. It has the potential to, and if so we should of course enable it. It's just that nextpnr has no infrastructure for that at the moment.

[gatecat]

The only place where OpenMP is used at the moment is not within nextpnr itself, but inside Eigen3 which speeds up sparse matrix solving during HeAP a little bit. In general it's a fairly small benefit (we already get about 2x parallelism there by solving X and Y in separate threads regardless) hence it not being enabled by default.