Look into why Brian is slower than Auryn

[thesamovar]

From a quick look, Auryn doesn't seem to be doing anything particularly different to Brian except that they are explicitly managing SIMD instructions, e.g. in this file. I thought that Brian should also be using them the way we had set things up, but perhaps not as efficiently as I think?

@fzenke any comment on this?

[fzenke]

How big are the differences we are currently talking about? Has there been a recent benchmark comparison? Since I have little time to improve Auryn these days, most of its performance improvements are due to updates to the GNU C++ compiler.

[thesamovar]

I think the differences are not huge (not orders of magnitude) but if it would be easy for us to just throw in a few SIMD instructions here and there or make sure a few arrays are allocated in a properly aligned way, it might be that we could get a speed boost at relatively low effort.

[fzenke]

In my experience, the speed-up is non-significant for modern compiler versions which automatically use SIMD.

[thesamovar]

That was my impression, but someone sent me some (unpublished) numebrs where they carefully controlled compiler version etc. and Auryn was still quite a bit faster than Brian for at least some models. Having had a look at the code, there doesn't seem to be anything super clever you're doing different to Brian, other than the SIMD stuff, so I'm a bit baffled as to why the speed difference!