[whahnsr]

Summary

Optimize the two-mask radial distribution calculation for orthogonal
cells.

Profiling a two-mask radial calculation showed essentially all runtime
concentrated in Action_Radial::calcRDF_twoMask() and the distance
routines it calls.

For orthogonal cells, this change caches the box lengths once per frame
and performs the orthogonal minimum-image distance calculation directly
in the inner pair loop. Other imaging modes continue to use the existing
DIST2() path.

Using std::fabs() for the coordinate differences also allows the
compiler to generate branchless absolute-value operations in this hot
loop.

Performance

Test case:

• test/Test_AutoImage/walp_2.strip.dcd
• 11,670 atoms
• 71 frames
• two masks of 5,520 atoms each
• radial DOPC2 :39-78 :79-118 out ... 0.25 15.0
• Serial GNU build with -O3

CPU user time:

• Before: 17.10 s
• After: 10.30 s

This is approximately a 39.8% reduction in CPU time, or a 1.66x speedup.

Validation

• All 17 comparisons in test/Test_Radial pass.
• RDF output from the benchmark is byte-for-byte identical to the output
  produced by the original code.
• The modified Action_Radial.cpp was compiled successfully with
  -fopenmp to verify the OpenMP code path.
• git diff --check reports no whitespace errors.

Non-orthogonal and no-image calculations retain the existing DIST2()
path.

[drroe]

Thanks for these! I'm surprised the overhead from the function call/branching is so high. Probably should look into refactoring the routines in DistRoutines.h so that more Actions etc can benefit from this kind of speedup.

[drroe]

FYI I've merged #1195 into this branch.