In case you are using a lot of CUSTOM functions or switching functions, notice that these commands depend on the lepton library that is included in PLUMED. This library replaces libmatheval since PLUMED 2.5, and by itself it is significantly faster than libmatheval. However, you can make it even faster using a just-in-time compiler. As of PLUMED 2.6, the correct version of ASMJIT is embedded in PLUMED. In order to enable it it is sufficient to use a specific flag in configure:
./configure --enable-asmjit make make install
You are done!
Once ASMJIT has been configured, you can disable it at runtime setting the environment variable PLUMED_USE_ASMJIT
:
export PLUMED_USE_ASMJIT=no
In some case using a custom expression is almost as fast as using a hard-coded function. For instance, with an input that contained the following lines:
c: COORDINATIONGROUPA=1-108First list of atoms.GROUPB=1-108Second list of atoms (if empty, N*(N-1)/2 pairs in GROUPA are counted).R_0=1 d_fast: COORDINATIONcould not find this keywordGROUPA=1-108First list of atoms.GROUPB=1-108Second list of atoms (if empty, N*(N-1)/2 pairs in GROUPA are counted).SWITCH={CUSTOM FUNC=1/(1+x2^3) R_0=1}This keyword is used if you want to employ an alternative to the continuous switching function defined above.
I (GB) obtained the following timings (on a Macbook laptop):
... PLUMED: 4A 1 c 108 0.126592 0.001172 0.000701 0.002532 PLUMED: 4A 2 d_fast 108 0.135210 0.001252 0.000755 0.002623 ...
Notice the usage of x2
as a variable for the switching function (see switchingfunction), which avoids an unnecessary square root calculation (this is done automatically by the hard-coded switching functions when you use only even powers). The asmjit calculation (d_fast
) takes less than 10% more than the hard-coded one (c
).