loopiloop outsourceQ

[behinger]

I added all my improvements here.

On one test I got a speedup from 0.17s to 0.12s (fortran 0.3/0.4) - per iteration

I did remove all the intel optimizations, I don't know if they would help even further, but I wanted to focus on other things first.

The code is not fully type stable, but I improved quite a bit. E.g. all the learningrate stuff is not type stable which afaik makes "B" also not type stable.

I mainly looked at the first loop, nt the update-parameter loop

• Update typings to be more type-stable
• Reduce memory allocation and implement other performance improvements
• add array dimensions
• re-enable multi threading
• make Calculate_Q type stable
• use AppleAccelerate in certain places
• add intelvectormath
• minor
• more optimized methods
• minor
• log
• overflow fixed, debug statements added, comparson to fortran/matlab semi-automated
• fixes, make fp local
• update project.toml
• precalculate y^rho
• things are faster now
• optimize more sums
• fix sphering LL
• speed improvement from 272µs to 640ns
• insane additional 100x speed improvement in calculate_y
• this should work, but havent tested -oops
• added missing broadcast to first minus, 340 to 230 µs
• actual 10% slower, but less allocs. try later again to use?
• 15% or so improvement, much better on allocations
• tried to pull out generation of Q
• improved y_rho, removed experimental exp/pow

[behinger]

for a large problem I get ~12s/iteration in julia vs. 3s in fortran. So speedup of fortran is still factor 4 or so