Optimised pair_lj_charmm_coul* code

[ibethune]

Some small optimisations to the pair_lj_charmm_coul compute() functions, principally storing reciprocal values to avoid excessive floating point divisions. In my tests (for a system dominated by these pair interactions), it gave a 4% speedup.

Before:

Loop time of 120.903 on 1 procs for 1000 steps with 7018 atoms

Performance: 5.717 ns/day, 4.198 hours/ns, 8.271 timesteps/s
99.3% CPU use with 1 MPI tasks x no OpenMP threads

MPI task timing breakdown:
Section |  min time  |  avg time  |  max time  |%varavg| %total
---------------------------------------------------------------
Pair    | 103.35     | 103.35     | 103.35     |   0.0 | 85.48
Bond    | 1.095      | 1.095      | 1.095      |   0.0 |  0.91
Kspace  | 4.3415     | 4.3415     | 4.3415     |   0.0 |  3.59
Neigh   | 7.2393     | 7.2393     | 7.2393     |   0.0 |  5.99
Comm    | 1.1778     | 1.1778     | 1.1778     |   0.0 |  0.97
Output  | 0.0020359  | 0.0020359  | 0.0020359  |   0.0 |  0.00
Modify  | 2.9979     | 2.9979     | 2.9979     |   0.0 |  2.48
Other   |            | 0.6963     |            |       |  0.58

After:

Loop time of 116.697 on 1 procs for 1000 steps with 7018 atoms

Performance: 5.923 ns/day, 4.052 hours/ns, 8.569 timesteps/s
99.4% CPU use with 1 MPI tasks x no OpenMP threads

MPI task timing breakdown:
Section |  min time  |  avg time  |  max time  |%varavg| %total
---------------------------------------------------------------
Pair    | 99.448     | 99.448     | 99.448     |   nan | 85.22
Bond    | 1.0763     | 1.0763     | 1.0763     |   0.0 |  0.92
Kspace  | 4.0912     | 4.0912     | 4.0912     |   0.0 |  3.51
Neigh   | 7.3013     | 7.3013     | 7.3013     |   0.0 |  6.26
Comm    | 1.1464     | 1.1464     | 1.1464     |   0.0 |  0.98
Output  | 0.00031209 | 0.00031209 | 0.00031209 |   0.0 |  0.00
Modify  | 2.9559     | 2.9559     | 2.9559     |   0.0 |  2.53
Other   |            | 0.6776     |            |       |  0.58

[akohlmey]

FYI, there is quite a bit more speedup possible when optimizing data access and making invariant if statements constant at compile time via templates. see the modifications done in the OPT and USER-OMP package (when compiling USER-OMP without OpenMP enabled, you should pretty much get what is in OPT or better).