Improve benchmarking

[andyferris]

Current benchmarking is rather ad-hoc. I would like to use BenchmarkTools.jl but the results from perf/benchmark3.jl is rather different from my own (to the point where those results appear incorrect).

Can also steal code from FixedSizeArrays.

[KristofferC]

I had to use JuliaCI/BenchmarkTools.jl#28 to get good results. Else things were being constant elided.

[andyferris]

Thanks @KristofferC. Actually, I was aware of that PR and it didn't make a difference for me... i'd like to understand this better.

[ChrisRackauckas]

What about setting this up with https://github.com/JuliaCI/PkgBenchmark.jl ?

[andyferris]

Would be a nice idea.

[KristofferC]

I just want to point out that PkgBenchmark.jl master is way ahead of the tagged version, it is not tagged because it hasn't gotten enough testing by others so I'm not sure how well the API works.

Also, in the new PkgBenchmark.jl, the package is only used when doing measurements. It is not needed when writing the actual benchmarks. That is done by just defining a BenchmarkGroup.