Table speedup

[brucefan1983]

• Speed up NEP calculation (both the LAMMPS version and the non-LAMMPS version) by using tables for the radial functions. See the PR in the GPUMD repo for details: https://github.com/brucefan1983/GPUMD/pull/392
• Also used openMP for some parts of the code, see #17

To enable the table, one can modify line 20 in src/nep.h:

// #define USE_TABLE_FOR_RADIAL_FUNCTIONS

The speedup is even higher in the CPU versions of NEP here because memory accessing in CPU is relatively cheaper compared to floating point arithmetics. From my tests, one can get a 2X speedup for typical models.

Force is only accurate at the level of 1e-6 eV/A compared to the analytical version:

[chem-william]

On my end, this is an awesome change! I compiled LAMMPS with the OneAPI intel compilers and Intel make file with 42 cpus and I see an almost two-times speedup (simulation time 4min 56sec -> 2min 32secs)

I ran a simulation with a bunch of gold and looking at the trajectory with ase gui they look qualitatively similar. There are differences, but I think they could be because its chaotic systems. I don't have any hard properties to look at yet.

[brucefan1983]

Thanks for the tests. At least you confirmed the expected 2x speedup and that the MD results are reasonale. Whether there will be any risk in some MD application is something that one should keep in mind. So far I have not seen any risk. For heat transport which is sensitive to energy conservation we also got consistent results between table and non-table versions. Because the default is not to enable table, I think it is safe to merge the PR now.