diff --git a/doc/Section_howto.html b/doc/Section_howto.html index 3bac6be12..4ee014bda 100644 --- a/doc/Section_howto.html +++ b/doc/Section_howto.html @@ -1,2035 +1,2035 @@ <HTML> <CENTER><A HREF = "Section_accelerate.html">Previous Section</A> - <A HREF = "http://lammps.sandia.gov">LAMMPS WWW Site</A> - <A HREF = "Manual.html">LAMMPS Documentation</A> - <A HREF = "Section_commands.html#comm">LAMMPS Commands</A> - <A HREF = "Section_example.html">Next Section</A> </CENTER> <HR> <H3>6. How-to discussions </H3> <P>This section describes how to perform common tasks using LAMMPS. </P> 6.1 <A HREF = "#howto_1">Restarting a simulation</A><BR> 6.2 <A HREF = "#howto_2">2d simulations</A><BR> 6.3 <A HREF = "#howto_3">CHARMM, AMBER, and DREIDING force fields</A><BR> 6.4 <A HREF = "#howto_4">Running multiple simulations from one input script</A><BR> 6.5 <A HREF = "#howto_5">Multi-replica simulations</A><BR> 6.6 <A HREF = "#howto_6">Granular models</A><BR> 6.7 <A HREF = "#howto_7">TIP3P water model</A><BR> 6.8 <A HREF = "#howto_8">TIP4P water model</A><BR> 6.9 <A HREF = "#howto_9">SPC water model</A><BR> 6.10 <A HREF = "#howto_10">Coupling LAMMPS to other codes</A><BR> 6.11 <A HREF = "#howto_11">Visualizing LAMMPS snapshots</A><BR> 6.12 <A HREF = "#howto_12">Triclinic (non-orthogonal) simulation boxes</A><BR> 6.13 <A HREF = "#howto_13">NEMD simulations</A><BR> 6.14 <A HREF = "#howto_14">Extended spherical and aspherical particles</A><BR> 6.15 <A HREF = "#howto_15">Output from LAMMPS (thermo, dumps, computes, fixes, variables)</A><BR> 6.16 <A HREF = "#howto_16">Thermostatting, barostatting and computing temperature</A><BR> 6.17 <A HREF = "#howto_17">Walls</A><BR> 6.18 <A HREF = "#howto_18">Elastic constants</A><BR> 6.19 <A HREF = "#howto_19">Library interface to LAMMPS</A><BR> 6.20 <A HREF = "#howto_20">Calculating thermal conductivity</A><BR> 6.21 <A HREF = "#howto_21">Calculating viscosity</A> <BR> <P>The example input scripts included in the LAMMPS distribution and highlighted in <A HREF = "Section_example.html">Section_example</A> also show how to setup and run various kinds of simulations. </P> <HR> <HR> <A NAME = "howto_1"></A><H4>6.1 Restarting a simulation </H4> <P>There are 3 ways to continue a long LAMMPS simulation. Multiple <A HREF = "run.html">run</A> commands can be used in the same input script. Each run will continue from where the previous run left off. Or binary restart files can be saved to disk using the <A HREF = "restart.html">restart</A> command. At a later time, these binary files can be read via a <A HREF = "read_restart.html">read_restart</A> command in a new script. Or they can be converted to text data files and read by a <A HREF = "read_data.html">read_data</A> command in a new script. <A HREF = "Section_tools.html">This section</A> discusses the <I>restart2data</I> tool that is used to perform the conversion. </P> <P>Here we give examples of 2 scripts that read either a binary restart file or a converted data file and then issue a new run command to continue where the previous run left off. They illustrate what settings must be made in the new script. Details are discussed in the documentation for the <A HREF = "read_restart.html">read_restart</A> and <A HREF = "read_data.html">read_data</A> commands. </P> <P>Look at the <I>in.chain</I> input script provided in the <I>bench</I> directory of the LAMMPS distribution to see the original script that these 2 scripts are based on. If that script had the line </P> <PRE>restart 50 tmp.restart </PRE> <P>added to it, it would produce 2 binary restart files (tmp.restart.50 and tmp.restart.100) as it ran. </P> <P>This script could be used to read the 1st restart file and re-run the last 50 timesteps: </P> <PRE>read_restart tmp.restart.50 </PRE> <PRE>neighbor 0.4 bin neigh_modify every 1 delay 1 </PRE> <PRE>fix 1 all nve fix 2 all langevin 1.0 1.0 10.0 904297 </PRE> <PRE>timestep 0.012 </PRE> <PRE>run 50 </PRE> <P>Note that the following commands do not need to be repeated because their settings are included in the restart file: <I>units, atom_style, special_bonds, pair_style, bond_style</I>. However these commands do need to be used, since their settings are not in the restart file: <I>neighbor, fix, timestep</I>. </P> <P>If you actually use this script to perform a restarted run, you will notice that the thermodynamic data match at step 50 (if you also put a "thermo 50" command in the original script), but do not match at step 100. This is because the <A HREF = "fix_langevin.html">fix langevin</A> command uses random numbers in a way that does not allow for perfect restarts. </P> <P>As an alternate approach, the restart file could be converted to a data file using this tool: </P> <PRE>restart2data tmp.restart.50 tmp.restart.data </PRE> <P>Then, this script could be used to re-run the last 50 steps: </P> <PRE>units lj atom_style bond pair_style lj/cut 1.12 pair_modify shift yes bond_style fene special_bonds 0.0 1.0 1.0 </PRE> <PRE>read_data tmp.restart.data </PRE> <PRE>neighbor 0.4 bin neigh_modify every 1 delay 1 </PRE> <PRE>fix 1 all nve fix 2 all langevin 1.0 1.0 10.0 904297 </PRE> <PRE>timestep 0.012 </PRE> <PRE>reset_timestep 50 run 50 </PRE> <P>Note that nearly all the settings specified in the original <I>in.chain</I> script must be repeated, except the <I>pair_coeff</I> and <I>bond_coeff</I> commands since the new data file lists the force field coefficients. Also, the <A HREF = "reset_timestep.html">reset_timestep</A> command is used to tell LAMMPS the current timestep. This value is stored in restart files, but not in data files. </P> <HR> <A NAME = "howto_2"></A><H4>6.2 2d simulations </H4> <P>Use the <A HREF = "dimension.html">dimension</A> command to specify a 2d simulation. </P> <P>Make the simulation box periodic in z via the <A HREF = "boundary.html">boundary</A> command. This is the default. </P> <P>If using the <A HREF = "create_box.html">create box</A> command to define a simulation box, set the z dimensions narrow, but finite, so that the create_atoms command will tile the 3d simulation box with a single z plane of atoms - e.g. </P> <PRE><A HREF = "create_box.html">create box</A> 1 -10 10 -10 10 -0.25 0.25 </PRE> <P>If using the <A HREF = "read_data.html">read data</A> command to read in a file of atom coordinates, set the "zlo zhi" values to be finite but narrow, similar to the create_box command settings just described. For each atom in the file, assign a z coordinate so it falls inside the z-boundaries of the box - e.g. 0.0. </P> <P>Use the <A HREF = "fix_enforce2d.html">fix enforce2d</A> command as the last defined fix to insure that the z-components of velocities and forces are zeroed out every timestep. The reason to make it the last fix is so that any forces induced by other fixes will be zeroed out. </P> <P>Many of the example input scripts included in the LAMMPS distribution are for 2d models. </P> <P>IMPORTANT NOTE: Some models in LAMMPS treat particles as extended spheres, as opposed to point particles. In 2d, the particles will still be spheres, not disks, meaning their moment of inertia will be the same as in 3d. </P> <HR> <A NAME = "howto_3"></A><H4>6.3 CHARMM, AMBER, and DREIDING force fields </H4> <P>A force field has 2 parts: the formulas that define it and the coefficients used for a particular system. Here we only discuss formulas implemented in LAMMPS that correspond to formulas commonly used in the CHARMM, AMBER, and DREIDING force fields. Setting coefficients is done in the input data file via the <A HREF = "read_data.html">read_data</A> command or in the input script with commands like <A HREF = "pair_coeff.html">pair_coeff</A> or <A HREF = "bond_coeff.html">bond_coeff</A>. See <A HREF = "Section_tools.html">Section_tools</A> for additional tools that can use CHARMM or AMBER to assign force field coefficients and convert their output into LAMMPS input. </P> <P>See <A HREF = "#MacKerell">(MacKerell)</A> for a description of the CHARMM force field. See <A HREF = "#Cornell">(Cornell)</A> for a description of the AMBER force field. </P> <P>These style choices compute force field formulas that are consistent with common options in CHARMM or AMBER. See each command's documentation for the formula it computes. </P> <UL><LI><A HREF = "bond_harmonic.html">bond_style</A> harmonic <LI><A HREF = "angle_charmm.html">angle_style</A> charmm <LI><A HREF = "dihedral_charmm.html">dihedral_style</A> charmm <LI><A HREF = "pair_charmm.html">pair_style</A> lj/charmm/coul/charmm <LI><A HREF = "pair_charmm.html">pair_style</A> lj/charmm/coul/charmm/implicit <LI><A HREF = "pair_charmm.html">pair_style</A> lj/charmm/coul/long </UL> <UL><LI><A HREF = "special_bonds.html">special_bonds</A> charmm <LI><A HREF = "special_bonds.html">special_bonds</A> amber </UL> <P>DREIDING is a generic force field developed by the <A HREF = "http://www.wag.caltech.edu">Goddard group</A> at Caltech and is useful for predicting structures and dynamics of organic, biological and main-group inorganic molecules. The philosophy in DREIDING is to use general force constants and geometry parameters based on simple hybridization considerations, rather than individual force constants and geometric parameters that depend on the particular combinations of atoms involved in the bond, angle, or torsion terms. DREIDING has an <A HREF = "pair_hbond_dreiding.html">explicit hydrogen bond term</A> to describe interactions involving a hydrogen atom on very electronegative atoms (N, O, F). </P> <P>See <A HREF = "#Mayo">(Mayo)</A> for a description of the DREIDING force field </P> <P>These style choices compute force field formulas that are consistent with the DREIDING force field. See each command's documentation for the formula it computes. </P> <UL><LI><A HREF = "bond_harmonic.html">bond_style</A> harmonic <LI><A HREF = "bond_morse.html">bond_style</A> morse </UL> <UL><LI><A HREF = "angle_harmonic.html">angle_style</A> harmonic <LI><A HREF = "angle_cosine.html">angle_style</A> cosine <LI><A HREF = "angle_cosine_periodic.html">angle_style</A> cosine/periodic </UL> <UL><LI><A HREF = "dihedral_charmm.html">dihedral_style</A> charmm <LI><A HREF = "improper_umbrella.html">improper_style</A> umbrella </UL> <UL><LI><A HREF = "pair_buck.html">pair_style</A> buck <LI><A HREF = "pair_buck.html">pair_style</A> buck/coul/cut <LI><A HREF = "pair_buck.html">pair_style</A> buck/coul/long <LI><A HREF = "pair_lj.html">pair_style</A> lj/cut <LI><A HREF = "pair_lj.html">pair_style</A> lj/cut/coul/cut <LI><A HREF = "pair_lj.html">pair_style</A> lj/cut/coul/long </UL> <UL><LI><A HREF = "pair_hbond_dreiding.html">pair_style</A> hbond/dreiding/lj <LI><A HREF = "pair_hbond_dreiding.html">pair_style</A> hbond/dreiding/morse </UL> <UL><LI><A HREF = "special_bonds.html">special_bonds</A> dreiding </UL> <HR> <A NAME = "howto_4"></A><H4>6.4 Running multiple simulations from one input script </H4> <P>This can be done in several ways. See the documentation for individual commands for more details on how these examples work. </P> <P>If "multiple simulations" means continue a previous simulation for more timesteps, then you simply use the <A HREF = "run.html">run</A> command multiple times. For example, this script </P> <PRE>units lj atom_style atomic read_data data.lj run 10000 run 10000 run 10000 run 10000 run 10000 </PRE> <P>would run 5 successive simulations of the same system for a total of 50,000 timesteps. </P> <P>If you wish to run totally different simulations, one after the other, the <A HREF = "clear.html">clear</A> command can be used in between them to re-initialize LAMMPS. For example, this script </P> <PRE>units lj atom_style atomic read_data data.lj run 10000 clear units lj atom_style atomic read_data data.lj.new run 10000 </PRE> <P>would run 2 independent simulations, one after the other. </P> <P>For large numbers of independent simulations, you can use <A HREF = "variable.html">variables</A> and the <A HREF = "next.html">next</A> and <A HREF = "jump.html">jump</A> commands to loop over the same input script multiple times with different settings. For example, this script, named in.polymer </P> <PRE>variable d index run1 run2 run3 run4 run5 run6 run7 run8 shell cd $d read_data data.polymer run 10000 shell cd .. clear next d jump in.polymer </PRE> <P>would run 8 simulations in different directories, using a data.polymer file in each directory. The same concept could be used to run the same system at 8 different temperatures, using a temperature variable and storing the output in different log and dump files, for example </P> <PRE>variable a loop 8 variable t index 0.8 0.85 0.9 0.95 1.0 1.05 1.1 1.15 log log.$a read data.polymer velocity all create $t 352839 fix 1 all nvt $t $t 100.0 dump 1 all atom 1000 dump.$a run 100000 next t next a jump in.polymer </PRE> <P>All of the above examples work whether you are running on 1 or multiple processors, but assumed you are running LAMMPS on a single partition of processors. LAMMPS can be run on multiple partitions via the "-partition" command-line switch as described in <A HREF = "Section_start.html#start_7">this section</A> of the manual. </P> <P>In the last 2 examples, if LAMMPS were run on 3 partitions, the same scripts could be used if the "index" and "loop" variables were replaced with <I>universe</I>-style variables, as described in the <A HREF = "variable.html">variable</A> command. Also, the "next t" and "next a" commands would need to be replaced with a single "next a t" command. With these modifications, the 8 simulations of each script would run on the 3 partitions one after the other until all were finished. Initially, 3 simulations would be started simultaneously, one on each partition. When one finished, that partition would then start the 4th simulation, and so forth, until all 8 were completed. </P> <HR> <A NAME = "howto_5"></A><H4>6.5 Multi-replica simulations </H4> <P>Several commands in LAMMPS run mutli-replica simulations, meaning that multiple instances (replicas) of your simulation are run simultaneously, with small amounts of data exchanged between replicas periodically. </P> <P>These are the relevant commands: </P> <UL><LI><A HREF = "neb.html">neb</A> for nudged elastic band calculations <LI><A HREF = "prd.html">prd</A> for parallel replica dynamics <LI><A HREF = "tad.html">tad</A> for temperature accelerated dynamics <LI><A HREF = "temper.html">temper</A> for parallel tempering </UL> <P>NEB is a method for finding transition states and barrier energies. PRD and TAD are methods for performing accelerated dynamics to find and perform infrequent events. Parallel tempering or replica exchange runs different replicas at a series of temperature to facilitate rare-event sampling. </P> <P>These command can only be used if LAMMPS was built with the "replica" package. See the <A HREF = "Section_start.html#start_3">Making LAMMPS</A> section for more info on packages. </P> <P>In all these cases, you must run with one or more processors per replica. The processors assigned to each replica are determined at run-time by using the <A HREF = "Section_start.html#start_7">-partition command-line switch</A> to launch LAMMPS on multiple partitions, which in this context are the same as replicas. E.g. these commands: </P> <PRE>mpirun -np 16 lmp_linux -partition 8x2 -in in.temper mpirun -np 8 lmp_linux -partition 8x1 -in in.neb </PRE> <P>would each run 8 replicas, on either 16 or 8 processors. Note the use of the <A HREF = "Section_start.html#start_7">-in command-line switch</A> to specify the input script which is required when running in multi-replica mode. </P> <P>Also note that with MPI installed on a machine (e.g. your desktop), you can run on more (virtual) processors than you have physical processors. Thus the above commands could be run on a single-processor (or few-processor) desktop so that you can run a multi-replica simulation on more replicas than you have physical processors. </P> <HR> <A NAME = "howto_6"></A><H4>6.6 Granular models </H4> <P>Granular system are composed of spherical particles with a diameter, as opposed to point particles. This means they have an angular velocity and torque can be imparted to them to cause them to rotate. </P> <P>To run a simulation of a granular model, you will want to use the following commands: </P> <UL><LI><A HREF = "atom_style.html">atom_style sphere</A> <LI><A HREF = "fix_nve_sphere.html">fix nve/sphere</A> <LI><A HREF = "fix_gravity.html">fix gravity</A> </UL> <P>This compute </P> <UL><LI><A HREF = "compute_erotate_sphere.html">compute erotate/sphere</A> </UL> <P>calculates rotational kinetic energy which can be <A HREF = "Section_howto.html#howto_15">output with thermodynamic info</A>. </P> <P>Use one of these 3 pair potentials, which compute forces and torques between interacting pairs of particles: </P> <UL><LI><A HREF = "pair_style.html">pair_style</A> gran/history <LI><A HREF = "pair_style.html">pair_style</A> gran/no_history <LI><A HREF = "pair_style.html">pair_style</A> gran/hertzian </UL> <P>These commands implement fix options specific to granular systems: </P> <UL><LI><A HREF = "fix_freeze.html">fix freeze</A> <LI><A HREF = "fix_pour.html">fix pour</A> <LI><A HREF = "fix_viscous.html">fix viscous</A> <LI><A HREF = "fix_wall_gran.html">fix wall/gran</A> </UL> <P>The fix style <I>freeze</I> zeroes both the force and torque of frozen atoms, and should be used for granular system instead of the fix style <I>setforce</I>. </P> <P>For computational efficiency, you can eliminate needless pairwise computations between frozen atoms by using this command: </P> <UL><LI><A HREF = "neigh_modify.html">neigh_modify</A> exclude </UL> <HR> <A NAME = "howto_7"></A><H4>6.7 TIP3P water model </H4> <P>The TIP3P water model as implemented in CHARMM <A HREF = "#MacKerell">(MacKerell)</A> specifies a 3-site rigid water molecule with charges and Lennard-Jones parameters assigned to each of the 3 atoms. In LAMMPS the <A HREF = "fix_shake.html">fix shake</A> command can be used to hold the two O-H bonds and the H-O-H angle rigid. A bond style of <I>harmonic</I> and an angle style of <I>harmonic</I> or <I>charmm</I> should also be used. </P> <P>These are the additional parameters (in real units) to set for O and H atoms and the water molecule to run a rigid TIP3P-CHARMM model with a cutoff. The K values can be used if a flexible TIP3P model (without fix shake) is desired. If the LJ epsilon and sigma for HH and OH are set to 0.0, it corresponds to the original 1983 TIP3P model <A HREF = "#Jorgensen">(Jorgensen)</A>. </P> <P>O mass = 15.9994<BR> H mass = 1.008 <BR> </P> <P>O charge = -0.834<BR> H charge = 0.417 <BR> </P> <P>LJ epsilon of OO = 0.1521<BR> LJ sigma of OO = 3.1507<BR> LJ epsilon of HH = 0.0460<BR> LJ sigma of HH = 0.4000<BR> LJ epsilon of OH = 0.0836<BR> LJ sigma of OH = 1.7753 <BR> </P> <P>K of OH bond = 450<BR> r0 of OH bond = 0.9572 <BR> </P> <P>K of HOH angle = 55<BR> theta of HOH angle = 104.52 <BR> </P> <P>These are the parameters to use for TIP3P with a long-range Coulombic solver (Ewald or PPPM in LAMMPS), see <A HREF = "#Price">(Price)</A> for details: </P> <P>O mass = 15.9994<BR> H mass = 1.008 <BR> </P> <P>O charge = -0.830<BR> H charge = 0.415 <BR> </P> <P>LJ epsilon of OO = 0.102<BR> LJ sigma of OO = 3.188<BR> LJ epsilon, sigma of OH, HH = 0.0 <BR> </P> <P>K of OH bond = 450<BR> r0 of OH bond = 0.9572 <BR> </P> <P>K of HOH angle = 55<BR> theta of HOH angle = 104.52 <BR> </P> <P>Wikipedia also has a nice article on <A HREF = "http://en.wikipedia.org/wiki/Water_model">water models</A>. </P> <HR> <A NAME = "howto_8"></A><H4>6.8 TIP4P water model </H4> <P>The four-point TIP4P rigid water model extends the traditional three-point TIP3P model by adding an additional site, usually massless, where the charge associated with the oxygen atom is placed. This site M is located at a fixed distance away from the oxygen along the bisector of the HOH bond angle. A bond style of <I>harmonic</I> and an angle style of <I>harmonic</I> or <I>charmm</I> should also be used. </P> <P>A TIP4P model is run with LAMMPS using two commands: </P> <UL><LI><A HREF = "pair_lj.html">pair_style lj/cut/coul/long/tip4p</A> <LI><A HREF = "kspace_style.html">kspace_style pppm/tip4p</A> </UL> <P>Note that only a TIP4P model with long-range Coulombics is currently implemented. A cutoff version may be added in the future. for both models, the bond lengths and bond angles should be held fixed using the <A HREF = "fix_shake.html">fix shake</A> command. </P> <P>These are the additional parameters (in real units) to set for O and H atoms and the water molecule to run a rigid TIP4P model with a cutoff <A HREF = "#Jorgensen">(Jorgensen)</A>. Note that the OM distance is specified in the <A HREF = "pair_style.html">pair_style</A> command, not as part of the pair coefficients. </P> <P>O mass = 15.9994<BR> H mass = 1.008 <BR> </P> <P>O charge = -1.040<BR> H charge = 0.520 <BR> </P> <P>r0 of OH bond = 0.9572<BR> theta of HOH angle = 104.52 <BR> </P> <P>OM distance = 0.15 <BR> </P> <P>LJ epsilon of O-O = 0.1550<BR> LJ sigma of O-O = 3.1536<BR> LJ epsilon, sigma of OH, HH = 0.0 <BR> </P> <P>These are the parameters to use for TIP4P with a long-range Coulombic solver (Ewald or PPPM in LAMMPS): </P> <P>O mass = 15.9994<BR> H mass = 1.008 <BR> </P> <P>O charge = -1.0484<BR> H charge = 0.5242 <BR> </P> <P>r0 of OH bond = 0.9572<BR> theta of HOH angle = 104.52 <BR> </P> <P>OM distance = 0.1250 <BR> </P> <P>LJ epsilon of O-O = 0.16275<BR> LJ sigma of O-O = 3.16435<BR> LJ epsilon, sigma of OH, HH = 0.0 <BR> </P> <P>Note that the when using the TIP4P pair style, the neighobr list cutoff for Coulomb interactions is effectively extended by a distance 2 * (OM distance), to account for the offset distance of the fictitious charges on O atoms in water molecules. Thus it is typically best in an efficiency sense to use a LJ cutoff >= Coulomb cutoff + 2*(OM distance), to shrink the size of the neighbor list. This leads to slightly larger cost for the long-range calculation, so you can test the trade-off for your model. The OM distance and the LJ and Coulombic cutoffs are set in the <A HREF = "pair_lj.html">pair_style lj/cut/coul/long/tip4p</A> command. </P> <P>Wikipedia also has a nice article on <A HREF = "http://en.wikipedia.org/wiki/Water_model">water models</A>. </P> <HR> <A NAME = "howto_9"></A><H4>6.9 SPC water model </H4> <P>The SPC water model specifies a 3-site rigid water molecule with charges and Lennard-Jones parameters assigned to each of the 3 atoms. In LAMMPS the <A HREF = "fix_shake.html">fix shake</A> command can be used to hold the two O-H bonds and the H-O-H angle rigid. A bond style of <I>harmonic</I> and an angle style of <I>harmonic</I> or <I>charmm</I> should also be used. </P> <P>These are the additional parameters (in real units) to set for O and H atoms and the water molecule to run a rigid SPC model. </P> <P>O mass = 15.9994<BR> H mass = 1.008 <BR> </P> <P>O charge = -0.820<BR> H charge = 0.410 <BR> </P> <P>LJ epsilon of OO = 0.1553<BR> LJ sigma of OO = 3.166<BR> LJ epsilon, sigma of OH, HH = 0.0 <BR> </P> <P>r0 of OH bond = 1.0<BR> theta of HOH angle = 109.47 <BR> </P> <P>Note that as originally proposed, the SPC model was run with a 9 Angstrom cutoff for both LJ and Coulommbic terms. It can also be used with long-range Coulombics (Ewald or PPPM in LAMMPS), without changing any of the parameters above, though it becomes a different model in that mode of usage. </P> <P>The SPC/E (extended) water model is the same, except the partial charge assignemnts change: </P> <P>O charge = -0.8476<BR> H charge = 0.4238 <BR> </P> <P>See the <A HREF = "#Berendsen">(Berendsen)</A> reference for more details on both the SPC and SPC/E models. </P> <P>Wikipedia also has a nice article on <A HREF = "http://en.wikipedia.org/wiki/Water_model">water models</A>. </P> <HR> <A NAME = "howto_10"></A><H4>6.10 Coupling LAMMPS to other codes </H4> <P>LAMMPS is designed to allow it to be coupled to other codes. For example, a quantum mechanics code might compute forces on a subset of atoms and pass those forces to LAMMPS. Or a continuum finite element (FE) simulation might use atom positions as boundary conditions on FE nodal points, compute a FE solution, and return interpolated forces on MD atoms. </P> <P>LAMMPS can be coupled to other codes in at least 3 ways. Each has advantages and disadvantages, which you'll have to think about in the context of your application. </P> <P>(1) Define a new <A HREF = "fix.html">fix</A> command that calls the other code. In this scenario, LAMMPS is the driver code. During its timestepping, the fix is invoked, and can make library calls to the other code, which has been linked to LAMMPS as a library. This is the way the <A HREF = "http://www.rpi.edu/~anderk5/lab">POEMS</A> package that performs constrained rigid-body motion on groups of atoms is hooked to LAMMPS. See the <A HREF = "fix_poems.html">fix_poems</A> command for more details. See <A HREF = "Section_modify.html">this section</A> of the documentation for info on how to add a new fix to LAMMPS. </P> <P>(2) Define a new LAMMPS command that calls the other code. This is conceptually similar to method (1), but in this case LAMMPS and the other code are on a more equal footing. Note that now the other code is not called during the timestepping of a LAMMPS run, but between runs. The LAMMPS input script can be used to alternate LAMMPS runs with calls to the other code, invoked via the new command. The <A HREF = "run.html">run</A> command facilitates this with its <I>every</I> option, which makes it easy to run a few steps, invoke the command, run a few steps, invoke the command, etc. </P> <P>In this scenario, the other code can be called as a library, as in (1), or it could be a stand-alone code, invoked by a system() call made by the command (assuming your parallel machine allows one or more processors to start up another program). In the latter case the stand-alone code could communicate with LAMMPS thru files that the command writes and reads. </P> <P>See <A HREF = "Section_modify.html">Section_modify</A> of the documentation for how to add a new command to LAMMPS. </P> <P>(3) Use LAMMPS as a library called by another code. In this case the other code is the driver and calls LAMMPS as needed. Or a wrapper code could link and call both LAMMPS and another code as libraries. Again, the <A HREF = "run.html">run</A> command has options that allow it to be invoked with minimal overhead (no setup or clean-up) if you wish to do multiple short runs, driven by another program. </P> <P>Examples of driver codes that call LAMMPS as a library are included in the "couple" directory of the LAMMPS distribution; see couple/README for more details: </P> <UL><LI>simple: simple driver programs in C++ and C which invoke LAMMPS as a library <LI>lammps_quest: coupling of LAMMPS and <A HREF = "http://dft.sandia.gov/Quest">Quest</A>, to run classical MD with quantum forces calculated by a density functional code <LI>lammps_spparks: coupling of LAMMPS and <A HREF = "http://www.sandia.gov/~sjplimp/spparks.html">SPPARKS</A>, to couple a kinetic Monte Carlo model for grain growth using MD to calculate strain induced across grain boundaries </UL> <P><A HREF = "Section_start.html#start_5">This section</A> of the documentation describes how to build LAMMPS as a library. Once this is done, you can interface with LAMMPS either via C++, C, Fortran, or Python (or any other language that supports a vanilla C-like interface). For example, from C++ you could create one (or more) "instances" of LAMMPS, pass it an input script to process, or execute individual commands, all by invoking the correct class methods in LAMMPS. From C or Fortran you can make function calls to do the same things. See <A HREF = "Section_python.html">Section_python</A> of the manual for a description of the Python wrapper provided with LAMMPS that operates through the LAMMPS library interface. </P> <P>The files src/library.cpp and library.h contain the C-style interface to LAMMPS. See <A HREF = "Section_howto.html#howto_19">Section_howto 19</A> of the manual for a description of the interface and how to extend it for your needs. </P> <P>Note that the lammps_open() function that creates an instance of LAMMPS takes an MPI communicator as an argument. This means that instance of LAMMPS will run on the set of processors in the communicator. Thus the calling code can run LAMMPS on all or a subset of processors. For example, a wrapper script might decide to alternate between LAMMPS and another code, allowing them both to run on all the processors. Or it might allocate half the processors to LAMMPS and half to the other code and run both codes simultaneously before syncing them up periodically. Or it might instantiate multiple instances of LAMMPS to perform different calculations. </P> <HR> <A NAME = "howto_11"></A><H4>6.11 Visualizing LAMMPS snapshots </H4> <P>LAMMPS itself does not do visualization, but snapshots from LAMMPS simulations can be visualized (and analyzed) in a variety of ways. </P> <P>LAMMPS snapshots are created by the <A HREF = "dump.html">dump</A> command which can create files in several formats. The native LAMMPS dump format is a text file (see "dump atom" or "dump custom") which can be visualized by the <A HREF = "Section_tools.html#xmovie">xmovie</A> program, included with the LAMMPS package. This produces simple, fast 2d projections of 3d systems, and can be useful for rapid debugging of simulation geometry and atom trajectories. </P> <P>Several programs included with LAMMPS as auxiliary tools can convert native LAMMPS dump files to other formats. See the <A HREF = "Section_tools.html">Section_tools</A> doc page for details. The first is the <A HREF = "Section_tools.html#charmm">ch2lmp tool</A>, which contains a lammps2pdb Perl script which converts LAMMPS dump files into PDB files. The second is the <A HREF = "Section_tools.html#arc">lmp2arc tool</A> which converts LAMMPS dump files into Accelrys' Insight MD program files. The third is the <A HREF = "Section_tools.html#cfg">lmp2cfg tool</A> which converts LAMMPS dump files into CFG files which can be read into the <A HREF = "http://mt.seas.upenn.edu/Archive/Graphics/A">AtomEye</A> visualizer. </P> <P>A Python-based toolkit distributed by our group can read native LAMMPS dump files, including custom dump files with additional columns of user-specified atom information, and convert them to various formats or pipe them into visualization software directly. See the <A HREF = "http://www.sandia.gov/~sjplimp/pizza.html">Pizza.py WWW site</A> for details. Specifically, Pizza.py can convert LAMMPS dump files into PDB, XYZ, <A HREF = "http://www.ensight.com">Ensight</A>, and VTK formats. Pizza.py can pipe LAMMPS dump files directly into the Raster3d and RasMol visualization programs. Pizza.py has tools that do interactive 3d OpenGL visualization and one that creates SVG images of dump file snapshots. </P> <P>LAMMPS can create XYZ files directly (via "dump xyz") which is a simple text-based file format used by many visualization programs including <A HREF = "http://www.ks.uiuc.edu/Research/vmd">VMD</A>. </P> <P>LAMMPS can create DCD files directly (via "dump dcd") which can be read by <A HREF = "http://www.ks.uiuc.edu/Research/vmd">VMD</A> in conjunction with a CHARMM PSF file. Using this form of output avoids the need to convert LAMMPS snapshots to PDB files. See the <A HREF = "dump.html">dump</A> command for more information on DCD files. </P> <P>LAMMPS can create XTC files directly (via "dump xtc") which is GROMACS file format which can also be read by <A HREF = "http://www.ks.uiuc.edu/Research/vmd">VMD</A> for visualization. See the <A HREF = "dump.html">dump</A> command for more information on XTC files. </P> <HR> <A NAME = "howto_12"></A><H4>6.12 Triclinic (non-orthogonal) simulation boxes </H4> <P>By default, LAMMPS uses an orthogonal simulation box to encompass the particles. The <A HREF = "boundary.html">boundary</A> command sets the boundary conditions of the box (periodic, non-periodic, etc). The orthogonal box has its "origin" at (xlo,ylo,zlo) and is defined by 3 edge vectors starting from the origin given by <B>a</B> = (xhi-xlo,0,0); <B>b</B> = (0,yhi-ylo,0); <B>c</B> = (0,0,zhi-zlo). The 6 parameters (xlo,xhi,ylo,yhi,zlo,zhi) are defined at the time the simulation box is created, e.g. by the <A HREF = "create_box.html">create_box</A> or <A HREF = "read_data.html">read_data</A> or <A HREF = "read_restart.html">read_restart</A> commands. Additionally, LAMMPS defines box size parameters lx,ly,lz where lx = xhi-xlo, and similarly in the y and z dimensions. The 6 parameters, as well as lx,ly,lz, can be output via the <A HREF = "thermo_style.html">thermo_style custom</A> command. </P> <P>LAMMPS also allows simulations to be performed in triclinic (non-orthogonal) simulation boxes shaped as a parallelepiped with triclinic symmetry. The parallelepiped has its "origin" at (xlo,ylo,zlo) and is defined by 3 edge vectors starting from the origin given by <B>a</B> = (xhi-xlo,0,0); <B>b</B> = (xy,yhi-ylo,0); <B>c</B> = (xz,yz,zhi-zlo). <I>xy,xz,yz</I> can be 0.0 or positive or negative values and are called "tilt factors" because they are the amount of displacement applied to faces of an originally orthogonal box to transform it into the parallelepiped. In LAMMPS the triclinic simulation box edge vectors <B>a</B>, <B>b</B>, and <B>c</B> cannot be arbitrary vectors. As indicated, <B>a</B> must lie on the positive x axis. <B>b</B> must lie in the xy plane, with strictly positive y component. <B>c</B> may have any orientation with strictly positive z component. The requirement that <B>a</B>, <B>b</B>, and <B>c</B> have strictly positive x, y, and z components, respectively, ensures that <B>a</B>, <B>b</B>, and <B>c</B> form a complete right-handed basis. These restrictions impose no loss of generality, since it is possible to rotate/invert any set of 3 crystal basis vectors so that they conform to the restrictions. </P> <P>For example, assume that the 3 vectors <B>A</B>,<B>B</B>,<B>C</B> are the edge vectors of a general parallelepiped, where there is no restriction on <B>A</B>,<B>B</B>,<B>C</B> other than they form a complete right-handed basis i.e. <B>A</B> x <B>B</B> . <B>C</B> > 0. The equivalent LAMMPS <B>a</B>,<B>b</B>,<B>c</B> are a linear rotation of <B>A</B>, <B>B</B>, and <B>C</B> and can be computed as follows: </P> <CENTER><IMG SRC = "Eqs/transform.jpg"> </CENTER> <P>where A = |<B>A</B>| indicates the scalar length of <B>A</B>. The ^ hat symbol indicates the corresponding unit vector. <I>beta</I> and <I>gamma</I> are angles between the vectors described below. Note that by construction, <B>a</B>, <B>b</B>, and <B>c</B> have strictly positive x, y, and z components, respectively. If it should happen that <B>A</B>, <B>B</B>, and <B>C</B> form a left-handed basis, then the above equations are not valid for <B>c</B>. In this case, it is necessary to first apply an inversion. This can be achieved by interchanging two basis vectors or by changing the sign of one of them. </P> -<P>For consistency, the same rotation/inversion appied to the basis vectors +<P>For consistency, the same rotation/inversion applied to the basis vectors must also be applied to atom positions, velocities, and any other vector quantities. This can be conveniently achieved by first converting to fractional coordinates in the old basis and then converting to distance coordinates in the new basis. The transformation is given by the following equation: </P> <CENTER><IMG SRC = "Eqs/rotate.jpg"> </CENTER> <P>where <I>V</I> is the volume of the box, <B>X</B> is the original vector quantity and <B>x</B> is the vector in the LAMMPS basis. </P> <P>There is no requirement that a triclinic box be periodic in any dimension, though it typically should be in at least the 2nd dimension of the tilt (y in xy) if you want to enforce a shift in periodic boundary conditions across that boundary. Some commands that work with triclinic boxes, e.g. the <A HREF = "fix_deform.html">fix deform</A> and <A HREF = "fix_nh.html">fix npt</A> commands, require periodicity or non-shrink-wrap boundary conditions in specific dimensions. See the command doc pages for details. </P> <P>The 9 parameters (xlo,xhi,ylo,yhi,zlo,zhi,xy,xz,yz) are defined at the time the simluation box is created. This happens in one of 3 ways. If the <A HREF = "create_box.html">create_box</A> command is used with a region of style <I>prism</I>, then a triclinic box is setup. See the <A HREF = "region.html">region</A> command for details. If the <A HREF = "read_data.html">read_data</A> command is used to define the simulation box, and the header of the data file contains a line with the "xy xz yz" keyword, then a triclinic box is setup. See the <A HREF = "read_data.html">read_data</A> command for details. Finally, if the <A HREF = "read_restart.html">read_restart</A> command reads a restart file which was written from a simulation using a triclinic box, then a triclinic box will be setup for the restarted simulation. </P> <P>Note that you can define a triclinic box with all 3 tilt factors = 0.0, so that it is initially orthogonal. This is necessary if the box will become non-orthogonal, e.g. due to the <A HREF = "fix_nh.html">fix npt</A> or <A HREF = "fix_deform.html">fix deform</A> commands. Alternatively, you can use the <A HREF = "change_box.html">change_box</A> command to convert a simulation box from orthogonal to triclinic and vice versa. </P> <P>As with orthogonal boxes, LAMMPS defines triclinic box size parameters lx,ly,lz where lx = xhi-xlo, and similarly in the y and z dimensions. The 9 parameters, as well as lx,ly,lz, can be output via the <A HREF = "thermo_style.html">thermo_style custom</A> command. </P> <P>To avoid extremely tilted boxes (which would be computationally inefficient), no tilt factor can skew the box more than half the distance of the parallel box length, which is the 1st dimension in the tilt factor (x for xz). For example, if xlo = 2 and xhi = 12, then the x box length is 10 and the xy tilt factor must be between -5 and 5. Similarly, both xz and yz must be between -(xhi-xlo)/2 and +(yhi-ylo)/2. Note that this is not a limitation, since if the maximum tilt factor is 5 (as in this example), then configurations with tilt = ..., -15, -5, 5, 15, 25, ... are geometrically all equivalent. If the box tilt exceeds this limit during a dynamics run (e.g. via the <A HREF = "fix_deform.html">fix deform</A> command), then the box is "flipped" to an equivalent shape with a tilt factor within the bounds, so the run can continue. See the <A HREF = "fix_deform.html">fix deform</A> doc page for further details. </P> <P>The one exception to this rule is if the 1st dimension in the tilt factor (x for xy) is non-periodic. In that case, the limits on the tilt factor are not enforced, since flipping the box in that dimension does not change the atom positions due to non-periodicity. In this mode, if you tilt the system to extreme angles, the simulation will simply become inefficient, due to the highly skewed simulation box. </P> <P>Triclinic crystal structures are often defined using three lattice constants <I>a</I>, <I>b</I>, and <I>c</I>, and three angles <I>alpha</I>, <I>beta</I> and <I>gamma</I>. Note that in this nomenclature, the a, b, and c lattice constants are the scalar lengths of the edge vectors <B>a</B>, <B>b</B>, and <B>c</B> defined above. The relationship between these 6 quantities (a,b,c,alpha,beta,gamma) and the LAMMPS box sizes (lx,ly,lz) = (xhi-xlo,yhi-ylo,zhi-zlo) and tilt factors (xy,xz,yz) is as follows: </P> <CENTER><IMG SRC = "Eqs/box.jpg"> </CENTER> <P>The inverse relationship can be written as follows: </P> <CENTER><IMG SRC = "Eqs/box_inverse.jpg"> </CENTER> <P>The values of <I>a</I>, <I>b</I>, <I>c</I> , <I>alpha</I>, <I>beta</I> , and <I>gamma</I> can be printed out or accessed by computes using the <A HREF = "thermo_style.html">thermo_style custom</A> keywords <I>cella</I>, <I>cellb</I>, <I>cellc</I>, <I>cellalpha</I>, <I>cellbeta</I>, <I>cellgamma</I>, respectively. </P> <P>As discussed on the <A HREF = "dump.html">dump</A> command doc page, when the BOX BOUNDS for a snapshot is written to a dump file for a triclinic box, an orthogonal bounding box which encloses the triclinic simulation box is output, along with the 3 tilt factors (xy, xz, yz) of the triclinic box, formatted as follows: </P> <PRE>ITEM: BOX BOUNDS xy xz yz xlo_bound xhi_bound xy ylo_bound yhi_bound xz zlo_bound zhi_bound yz </PRE> <P>This bounding box is convenient for many visualization programs and is calculated from the 9 triclinic box parameters (xlo,xhi,ylo,yhi,zlo,zhi,xy,xz,yz) as follows: </P> <PRE>xlo_bound = xlo + MIN(0.0,xy,xz,xy+xz) xhi_bound = xhi + MAX(0.0,xy,xz,xy+xz) ylo_bound = ylo + MIN(0.0,yz) yhi_bound = yhi + MAX(0.0,yz) zlo_bound = zlo zhi_bound = zhi </PRE> <P>These formulas can be inverted if you need to convert the bounding box back into the triclinic box parameters, e.g. xlo = xlo_bound - MIN(0.0,xy,xz,xy+xz). </P> <P>One use of triclinic simulation boxes is to model solid-state crystals with triclinic symmetry. The <A HREF = "lattice.html">lattice</A> command can be used with non-orthogonal basis vectors to define a lattice that will tile a triclinic simulation box via the <A HREF = "create_atoms.html">create_atoms</A> command. </P> <P>A second use is to run Parinello-Rahman dyanamics via the <A HREF = "fix_nh.html">fix npt</A> command, which will adjust the xy, xz, yz tilt factors to compensate for off-diagonal components of the pressure tensor. The analalog for an <A HREF = "minimize.html">energy minimization</A> is the <A HREF = "fix_box_relax.html">fix box/relax</A> command. </P> <P>A third use is to shear a bulk solid to study the response of the material. The <A HREF = "fix_deform.html">fix deform</A> command can be used for this purpose. It allows dynamic control of the xy, xz, yz tilt factors as a simulation runs. This is discussed in the next section on non-equilibrium MD (NEMD) simulations. </P> <HR> <A NAME = "howto_13"></A><H4>6.13 NEMD simulations </H4> <P>Non-equilibrium molecular dynamics or NEMD simulations are typically used to measure a fluid's rheological properties such as viscosity. In LAMMPS, such simulations can be performed by first setting up a non-orthogonal simulation box (see the preceding Howto section). </P> <P>A shear strain can be applied to the simulation box at a desired strain rate by using the <A HREF = "fix_deform.html">fix deform</A> command. The <A HREF = "fix_nvt_sllod.html">fix nvt/sllod</A> command can be used to thermostat the sheared fluid and integrate the SLLOD equations of motion for the system. Fix nvt/sllod uses <A HREF = "compute_temp_deform.html">compute temp/deform</A> to compute a thermal temperature by subtracting out the streaming velocity of the shearing atoms. The velocity profile or other properties of the fluid can be monitored via the <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> command. </P> <P>As discussed in the previous section on non-orthogonal simulation boxes, the amount of tilt or skew that can be applied is limited by LAMMPS for computational efficiency to be 1/2 of the parallel box length. However, <A HREF = "fix_deform.html">fix deform</A> can continuously strain a box by an arbitrary amount. As discussed in the <A HREF = "fix_deform.html">fix deform</A> command, when the tilt value reaches a limit, the box is flipped to the opposite limit which is an equivalent tiling of periodic space. The strain rate can then continue to change as before. In a long NEMD simulation these box re-shaping events may occur many times. </P> <P>In a NEMD simulation, the "remap" option of <A HREF = "fix_deform.html">fix deform</A> should be set to "remap v", since that is what <A HREF = "fix_nvt_sllod.html">fix nvt/sllod</A> assumes to generate a velocity profile consistent with the applied shear strain rate. </P> <P>An alternative method for calculating viscosities is provided via the <A HREF = "fix_viscosity.html">fix viscosity</A> command. </P> <HR> <A NAME = "howto_14"></A><H4>6.14 Extended spherical and aspherical particles </H4> <P>Typical MD models treat atoms or particles as point masses. Sometimes, however, it is desirable to have a model with finite-size particles such as spheres or aspherical ellipsoids. The difference is that such particles have a moment of inertia, rotational energy, and angular momentum. Rotation is induced by torque from interactions with other particles. </P> <P>LAMMPS has several options for running simulations with these kinds of particles. The following aspects are discussed in turn: </P> <UL><LI>atom styles <LI>pair potentials <LI>time integration <LI>computes, thermodynamics, and dump output <LI>rigid bodies composed of extended particles </UL> <H5>Atom styles </H5> <P>There are 2 <A HREF = "atom_style.html">atom styles</A> that allow for definition of finite-size particles: sphere and ellipsoid. The peri atom style also treats particles as having a volume, but that is internal to the <A HREF = "pair_peri.html">pair_style peri</A> potentials. The dipole atom style is most often used in conjunction with finite-size particles. </P> <P>The sphere style defines particles that are spheriods and each particle can have a unique diameter and mass (or density). These particles store an angular velocity (omega) and can be acted upon by torque. The "set" command can be used to modify the diameter and mass of individual particles, after then are created. </P> <P>The ellipsoid style defines particles that are ellipsoids and thus can be aspherical. Each particle has a shape, specified by 3 diameters, and mass (or density). These particles store an angular momentum and their orientation (quaternion), and can be acted upon by torque. They do not store an angular velocity (omega), which can be in a different direction than angular momentum, rather they compute it as needed. The "set" command can be used to modify the diameter, orientation, and mass of individual particles, after then are created. It also has a brief explanation of what quaternions are. </P> <P>The dipole style does not define extended particles, but is often used in conjunction with spherical particles, via a command like </P> <PRE>atom_style hybrid sphere dipole </PRE> <P>This is because when dipoles interact with each other, they induce torques, and a particle must be extended (i.e. have a moment of inertia) in order to respond and rotate. See the <A HREF = "atom_style.html">atom_style dipole</A> command for details. The "set" command can be used to modify the orientation and length of the dipole moment of individual particles, after then are created. </P> <P>Note that if one of these atom styles is used (or multiple styles via the <A HREF = "atom_style.html">atom_style hybrid</A> command), not all particles in the system are required to be finite-size or aspherical. For example, if the 3 shape parameters are set to the same value, the particle will be a sphere rather than an ellipsoid. If the 3 shape parameters are all set to 0.0 or if the diameter is set to 0.0, it will be a point particle. If the length of the dipole moment is set to zero, the particle will not have a point dipole associated with it. The pair styles used to compute pairwise interactions will typically compute the correct interaction in these simplified (cheaper) cases. <A HREF = "pair_hybrid.html">Pair_style hybrid</A> can be used to insure the correct interactions are computed for the appropriate style of interactions. Likewise, using groups to partition particles (ellipsoids versus spheres versus point particles) will allow you to use the appropriate time integrators and temperature computations for each class of particles. See the doc pages for various commands for details. </P> <P>Also note that for <A HREF = "dimension.html">2d simulations</A>, finite-size spheres and ellipsoids are still treated as 3d particles, rather than as circular disks or ellipses. This means they have the same moment of inertia for a 3d extended object. When their temperature is coomputed, the correct degrees of freedom are used for rotation in a 2d versus 3d system. </P> <H5>Pair potentials </H5> <P>When a system with extended particles is defined, the particles will only rotate and experience torque if the force field computes such interactions. These are the various <A HREF = "pair_style.html">pair styles</A> that generate torque: </P> <UL><LI><A HREF = "pair_gran.html">pair_style gran/history</A> <LI><A HREF = "pair_gran.html">pair_style gran/hertzian</A> <LI><A HREF = "pair_gran.html">pair_style gran/no_history</A> <LI><A HREF = "pair_dipole.html">pair_style dipole/cut</A> <LI><A HREF = "pair_gayberne.html">pair_style gayberne</A> <LI><A HREF = "pair_resquared.html">pair_style resquared</A> <LI><A HREF = "pair_lubricate.html">pair_style lubricate</A> </UL> <P>The <A HREF = "pair_gran.html">granular pair styles</A> are used with spherical particles. The <A HREF = "pair_dipole.html">dipole pair style</A> is used with <A HREF = "atom_style.html">atom_style dipole</A>, which could be applied to spherical or ellipsoidal particles. The <A HREF = "pair_gayberne.html">GayBerne</A> and <A HREF = "pair_resquared.html">REsquared</A> potentials require ellipsoidal particles, though they will also work if the 3 shape parameters are the same (a sphere). The <A HREF = "pair_lubricate.html">lubrication potential</A> works with spherical particles. </P> <H5>Time integration </H5> <P>There are 3 fixes that perform time integration on extended spherical particles, meaning the integrators update the rotational orientation and angular velocity or angular momentum of the particles: </P> <UL><LI><A HREF = "fix_nve_sphere.html">fix nve/sphere</A> <LI><A HREF = "fix_nvt_sphere.html">fix nvt/sphere</A> <LI><A HREF = "fix_npt_sphere.html">fix npt/sphere</A> </UL> <P>Likewise, there are 3 fixes that perform time integration on ellipsoids as extended aspherical particles: </P> <UL><LI><A HREF = "fix_nve_asphere.html">fix nve/asphere</A> <LI><A HREF = "fix_nvt_asphere.html">fix nvt/asphere</A> <LI><A HREF = "fix_npt_asphere.html">fix npt/asphere</A> </UL> <P>The advantage of these fixes is that those which thermostat the particles include the rotational degrees of freedom in the temperature calculation and thermostatting. Other thermostats can be used with fix nve/sphere or fix nve/asphere, such as fix langevin or fix temp/berendsen, but those thermostats only operate on the translational kinetic energy of the extended particles. </P> <P>Note that for mixtures of point and extended particles, you should only use these integration fixes on <A HREF = "group.html">groups</A> which contain extended particles. </P> <H5>Computes, thermodynamics, and dump output </H5> <P>There are 4 computes that calculate the temperature or rotational energy of extended spherical or aspherical particles (ellipsoids): </P> <UL><LI><A HREF = "compute_temp_sphere.html">compute temp/sphere</A> <LI><A HREF = "compute_temp_asphere.html">compute temp/asphere</A> <LI><A HREF = "compute_erotate_sphere.html">compute erotate/sphere</A> <LI><A HREF = "compute_erotate_asphere.html">compute erotate/asphere</A> </UL> <P>These include rotational degrees of freedom in their computation. If you wish the thermodynamic output of temperature or pressure to use one of these computes (e.g. for a system entirely composed of extended particles), then the compute can be defined and the <A HREF = "thermo_modify.html">thermo_modify</A> command used. Note that by default thermodynamic quantities will be calculated with a temperature that only includes translational degrees of freedom. See the <A HREF = "thermo_style.html">thermo_style</A> command for details. </P> <P>The <A HREF = "dump.html">dump custom</A> command can output various attributes of extended particles, including the dipole moment (mu), the angular velocity (omega), the angular momentum (angmom), the quaternion (quat), and the torque (tq) on the particle. </P> <H5>Rigid bodies composed of extended particles </H5> <P>The <A HREF = "fix_rigid.html">fix rigid</A> command treats a collection of particles as a rigid body, computes its inertia tensor, sums the total force and torque on the rigid body each timestep due to forces on its constituent particles, and integrates the motion of the rigid body. </P> <P>If any of the constituent particles of a rigid body are extended particles (spheres or ellipsoids), then their contribution to the inertia tensor of the body is different than if they were point particles. This means the rotational dynamics of the rigid body will be different. Thus a model of a dimer is different if the dimer consists of two point masses versus two extended sphereoids, even if the two particles have the same mass. Extended particles that experience torque due to their interaction with other particles will also impart that torque to a rigid body they are part of. </P> <P>See the "fix rigid" command for example of complex rigid-body models it is possible to define in LAMMPS. </P> <P>Note that the <A HREF = "fix_shake.html">fix shake</A> command can also be used to treat 2, 3, or 4 particles as a rigid body, but it always assumes the particles are point masses. </P> <HR> <A NAME = "howto_15"></A><H4>6.15 Output from LAMMPS (thermo, dumps, computes, fixes, variables) </H4> <P>There are four basic kinds of LAMMPS output: </P> <UL><LI><A HREF = "thermo_style.html">Thermodynamic output</A>, which is a list of quantities printed every few timesteps to the screen and logfile. <LI><A HREF = "dump.html">Dump files</A>, which contain snapshots of atoms and various per-atom values and are written at a specified frequency. <LI>Certain fixes can output user-specified quantities to files: <A HREF = "fix_ave_time.html">fix ave/time</A> for time averaging, <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> for spatial averaging, and <A HREF = "fix_print.html">fix print</A> for single-line output of <A HREF = "variable.html">variables</A>. Fix print can also output to the screen. <LI><A HREF = "restart.html">Restart files</A>. </UL> <P>A simulation prints one set of thermodynamic output and (optionally) restart files. It can generate any number of dump files and fix output files, depending on what <A HREF = "dump.html">dump</A> and <A HREF = "fix.html">fix</A> commands you specify. </P> <P>As discussed below, LAMMPS gives you a variety of ways to determine what quantities are computed and printed when the thermodynamics, dump, or fix commands listed above perform output. Throughout this discussion, note that users can also <A HREF = "Section_modify.html">add their own computes and fixes to LAMMPS</A> which can then generate values that can then be output with these commands. </P> <P>The following sub-sections discuss different LAMMPS command related to output and the kind of data they operate on and produce: </P> <UL><LI><A HREF = "#global">Global/per-atom/local data</A> <LI><A HREF = "#scalar">Scalar/vector/array data</A> <LI><A HREF = "#thermo">Thermodynamic output</A> <LI><A HREF = "#dump">Dump file output</A> <LI><A HREF = "#fixoutput">Fixes that write output files</A> <LI><A HREF = "#computeoutput">Computes that process output quantities</A> <LI><A HREF = "#fixoutput">Fixes that process output quantities</A> <LI><A HREF = "#compute">Computes that generate values to output</A> <LI><A HREF = "#fix">Fixes that generate values to output</A> <LI><A HREF = "#variable">Variables that generate values to output</A> <LI><A HREF = "#table">Summary table of output options and data flow between commands</A> </UL> <H5><A NAME = "global"></A>Global/per-atom/local data </H5> <P>Various output-related commands work with three different styles of data: global, per-atom, or local. A global datum is one or more system-wide values, e.g. the temperature of the system. A per-atom datum is one or more values per atom, e.g. the kinetic energy of each atom. Local datums are calculated by each processor based on the atoms it owns, but there may be zero or more per atom, e.g. a list of bond distances. </P> <H5><A NAME = "scalar"></A>Scalar/vector/array data </H5> <P>Global, per-atom, and local datums can each come in three kinds: a single scalar value, a vector of values, or a 2d array of values. The doc page for a "compute" or "fix" or "variable" that generates data will specify both the style and kind of data it produces, e.g. a per-atom vector. </P> <P>When a quantity is accessed, as in many of the output commands discussed below, it can be referenced via the following bracket notation, where ID in this case is the ID of a compute. The leading "c_" would be replaced by "f_" for a fix, or "v_" for a variable: </P> -<DIV ALIGN=center><TABLE WIDTH="0%" BORDER=1 > +<DIV ALIGN=center><TABLE BORDER=1 > <TR><TD >c_ID </TD><TD > entire scalar, vector, or array</TD></TR> <TR><TD >c_ID[I] </TD><TD > one element of vector, one column of array</TD></TR> <TR><TD >c_ID[I][J] </TD><TD > one element of array </TD></TR></TABLE></DIV> <P>In other words, using one bracket reduces the dimension of the data once (vector -> scalar, array -> vector). Using two brackets reduces the dimension twice (array -> scalar). Thus a command that uses scalar values as input can typically also process elements of a vector or array. </P> <H5><A NAME = "thermo"></A>Thermodynamic output </H5> <P>The frequency and format of thermodynamic output is set by the <A HREF = "thermo.html">thermo</A>, <A HREF = "thermo_style.html">thermo_style</A>, and <A HREF = "thermo_modify.html">thermo_modify</A> commands. The <A HREF = "thermo_style.html">thermo_style</A> command also specifies what values are calculated and written out. Pre-defined keywords can be specified (e.g. press, etotal, etc). Three additional kinds of keywords can also be specified (c_ID, f_ID, v_name), where a <A HREF = "compute.html">compute</A> or <A HREF = "fix.html">fix</A> or <A HREF = "variable.html">variable</A> provides the value to be output. In each case, the compute, fix, or variable must generate global values for input to the <A HREF = "dump.html">thermo_style custom</A> command. </P> <H5><A NAME = "dump"></A>Dump file output </H5> <P>Dump file output is specified by the <A HREF = "dump.html">dump</A> and <A HREF = "dump_modify.html">dump_modify</A> commands. There are several pre-defined formats (dump atom, dump xtc, etc). </P> <P>There is also a <A HREF = "dump.html">dump custom</A> format where the user specifies what values are output with each atom. Pre-defined atom attributes can be specified (id, x, fx, etc). Three additional kinds of keywords can also be specified (c_ID, f_ID, v_name), where a <A HREF = "compute.html">compute</A> or <A HREF = "fix.html">fix</A> or <A HREF = "variable.html">variable</A> provides the values to be output. In each case, the compute, fix, or variable must generate per-atom values for input to the <A HREF = "dump.html">dump custom</A> command. </P> <P>There is also a <A HREF = "dump.html">dump local</A> format where the user specifies what local values to output. A pre-defined index keyword can be specified to enumuerate the local values. Two additional kinds of keywords can also be specified (c_ID, f_ID), where a <A HREF = "compute.html">compute</A> or <A HREF = "fix.html">fix</A> or <A HREF = "variable.html">variable</A> provides the values to be output. In each case, the compute or fix must generate local values for input to the <A HREF = "dump.html">dump local</A> command. </P> <H5><A NAME = "fixoutput"></A>Fixes that write output files </H5> <P>Sevarl fixes take various quantities as input and can write output files: <A HREF = "fix_ave_time.html">fix ave/time</A>, <A HREF = "fix_ave_spatial.html">fix ave/spatial</A>, <A HREF = "fix_ave_histo.html">fix ave/histo</A>, <A HREF = "fix_ave_correlate.html">fix ave/correlate</A>, and <A HREF = "fix_print.html">fix print</A>. </P> <P>The <A HREF = "fix_ave_time.html">fix ave/time</A> command enables direct output to a file and/or time-averaging of global scalars or vectors. The user specifies one or more quantities as input. These can be global <A HREF = "compute.html">compute</A> values, global <A HREF = "fix.html">fix</A> values, or <A HREF = "variable.html">variables</A> of any style except the atom style which produces per-atom values. Since a variable can refer to keywords used by the <A HREF = "thermo_style.html">thermo_style custom</A> command (like temp or press) and individual per-atom values, a wide variety of quantities can be time averaged and/or output in this way. If the inputs are one or more scalar values, then the fix generate a global scalar or vector of output. If the inputs are one or more vector values, then the fix generates a global vector or array of output. The time-averaged output of this fix can also be used as input to other output commands. </P> <P>The <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> command enables direct output to a file of spatial-averaged per-atom quantities like those output in dump files, within 1d layers of the simulation box. The per-atom quantities can be atom density (mass or number) or atom attributes such as position, velocity, force. They can also be per-atom quantities calculated by a <A HREF = "compute.html">compute</A>, by a <A HREF = "fix.html">fix</A>, or by an atom-style <A HREF = "variable.html">variable</A>. The spatial-averaged output of this fix can also be used as input to other output commands. </P> <P>The <A HREF = "fix_ave_histo.html">fix ave/histo</A> command enables direct output to a file of histogrammed quantities, which can be global or per-atom or local quantities. The histogram output of this fix can also be used as input to other output commands. </P> <P>The <A HREF = "fix_ave_histo.html">fix ave/correlate</A> command enables direct output to a file of time-correlated quantities, which can be global scalars. The correlation matrix output of this fix can also be used as input to other output commands. </P> <P>The <A HREF = "fix_print.html">fix print</A> command can generate a line of output written to the screen and log file or to a separate file, periodically during a running simulation. The line can contain one or more <A HREF = "variable.html">variable</A> values for any style variable except the atom style). As explained above, variables themselves can contain references to global values generated by <A HREF = "thermo_style.html">thermodynamic keywords</A>, <A HREF = "compute.html">computes</A>, <A HREF = "fix.html">fixes</A>, or other <A HREF = "variable.html">variables</A>, or to per-atom values for a specific atom. Thus the <A HREF = "fix_print.html">fix print</A> command is a means to output a wide variety of quantities separate from normal thermodynamic or dump file output. </P> <H5><A NAME = "computeoutput"></A>Computes that process output quantities </H5> <P>The <A HREF = "compute_reduce.html">compute reduce</A> and <A HREF = "compute_reduce.html">compute reduce/region</A> commands take one or more per-atom or local vector quantities as inputs and "reduce" them (sum, min, max, ave) to scalar quantities. These are produced as output values which can be used as input to other output commands. </P> <P>The <A HREF = "compute_slice.html">compute slice</A> command take one or more global vector or array quantities as inputs and extracts a subset of their values to create a new vector or array. These are produced as output values which can be used as input to other output commands. </P> <P>The <A HREF = "compute_property_atom.html">compute property/atom</A> command takes a list of one or more pre-defined atom attributes (id, x, fx, etc) and stores the values in a per-atom vector or array. These are produced as output values which can be used as input to other output commands. The list of atom attributes is the same as for the <A HREF = "dump.html">dump custom</A> command. </P> <P>The <A HREF = "compute_property_local.html">compute property/local</A> command takes a list of one or more pre-defined local attributes (bond info, angle info, etc) and stores the values in a local vector or array. These are produced as output values which can be used as input to other output commands. </P> <P>The <A HREF = "compute_atom_molecule.html">compute atom/molecule</A> command takes a list of one or more per-atom quantities (from a compute, fix, per-atom variable) and sums the quantities on a per-molecule basis. It produces a global vector or array as output values which can be used as input to other output commands. </P> <H5><A NAME = "fixoutput"></A>Fixes that process output quantities </H5> <P>The <A HREF = "fix_ave_atom.html">fix ave/atom</A> command performs time-averaging of per-atom vectors. The per-atom quantities can be atom attributes such as position, velocity, force. They can also be per-atom quantities calculated by a <A HREF = "compute.html">compute</A>, by a <A HREF = "fix.html">fix</A>, or by an atom-style <A HREF = "variable.html">variable</A>. The time-averaged per-atom output of this fix can be used as input to other output commands. </P> <P>The <A HREF = "fix_store_state.html">fix store/state</A> command can archive one or more per-atom attributes at a particular time, so that the old values can be used in a future calculation or output. The list of atom attributes is the same as for the <A HREF = "dump.html">dump custom</A> command, including per-atom quantities calculated by a <A HREF = "compute.html">compute</A>, by a <A HREF = "fix.html">fix</A>, or by an atom-style <A HREF = "variable.html">variable</A>. The output of this fix can be used as input to other output commands. </P> <H5><A NAME = "compute"></A>Computes that generate values to output </H5> <P>Every <A HREF = "compute.html">compute</A> in LAMMPS produces either global or per-atom or local values. The values can be scalars or vectors or arrays of data. These values can be output using the other commands described in this section. The doc page for each compute command describes what it produces. Computes that produce per-atom or local values have the word "atom" or "local" in their style name. Computes without the word "atom" or "local" produce global values. </P> <H5><A NAME = "fix"></A>Fixes that generate values to output </H5> <P>Some <A HREF = "fix.html">fixes</A> in LAMMPS produces either global or per-atom or local values which can be accessed by other commands. The values can be scalars or vectors or arrays of data. These values can be output using the other commands described in this section. The doc page for each fix command tells whether it produces any output quantities and describes them. </P> <H5><A NAME = "variable"></A>Variables that generate values to output </H5> <P>Every <A HREF = "variable.html">variables</A> defined in an input script generates either a global scalar value or a per-atom vector (only atom-style variables) when it is accessed. The formulas used to define equal- and atom-style variables can contain references to the thermodynamic keywords and to global and per-atom data generated by computes, fixes, and other variables. The values generated by variables can be output using the other commands described in this section. </P> <H5><A NAME = "table"></A>Summary table of output options and data flow between commands </H5> <P>This table summarizes the various commands that can be used for generating output from LAMMPS. Each command produces output data of some kind and/or writes data to a file. Most of the commands can take data from other commands as input. Thus you can link many of these commands together in pipeline form, where data produced by one command is used as input to another command and eventually written to the screen or to a file. Note that to hook two commands together the output and input data types must match, e.g. global/per-atom/local data and scalar/vector/array data. </P> <P>Also note that, as described above, when a command takes a scalar as input, that could be an element of a vector or array. Likewise a vector input could be a column of an array. </P> -<DIV ALIGN=center><TABLE WIDTH="0%" BORDER=1 > +<DIV ALIGN=center><TABLE BORDER=1 > <TR><TD >Command</TD><TD > Input</TD><TD > Output</TD><TD ></TD></TR> <TR><TD ><A HREF = "thermo_style.html">thermo_style custom</A></TD><TD > global scalars</TD><TD > screen, log file</TD><TD ></TD></TR> <TR><TD ><A HREF = "dump.html">dump custom</A></TD><TD > per-atom vectors</TD><TD > dump file</TD><TD ></TD></TR> <TR><TD ><A HREF = "dump.html">dump local</A></TD><TD > local vectors</TD><TD > dump file</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_print.html">fix print</A></TD><TD > global scalar from variable</TD><TD > screen, file</TD><TD ></TD></TR> <TR><TD ><A HREF = "print.html">print</A></TD><TD > global scalar from variable</TD><TD > screen</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute.html">computes</A></TD><TD > N/A</TD><TD > global/per-atom/local scalar/vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix.html">fixes</A></TD><TD > N/A</TD><TD > global/per-atom/local scalar/vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "variable.html">variables</A></TD><TD > global scalars, per-atom vectors</TD><TD > global scalar, per-atom vector</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute_reduce.html">compute reduce</A></TD><TD > per-atom/local vectors</TD><TD > global scalar/vector</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute_slice.html">compute slice</A></TD><TD > global vectors/arrays</TD><TD > global vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute_property_atom.html">compute property/atom</A></TD><TD > per-atom vectors</TD><TD > per-atom vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute_property_local.html">compute property/local</A></TD><TD > local vectors</TD><TD > local vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "compute_atom_molecule.html">compute atom/molecule</A></TD><TD > per-atom vectors</TD><TD > global vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_ave_atom.html">fix ave/atom</A></TD><TD > per-atom vectors</TD><TD > per-atom vector/array</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_ave_time.html">fix ave/time</A></TD><TD > global scalars/vectors</TD><TD > global scalar/vector/array, file</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_ave_spatial.html">fix ave/spatial</A></TD><TD > per-atom vectors</TD><TD > global array, file</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_ave_histo.html">fix ave/histo</A></TD><TD > global/per-atom/local scalars and vectors</TD><TD > global array, file</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_ave_correlate.html">fix ave/correlate</A></TD><TD > global scalars</TD><TD > global array, file</TD><TD ></TD></TR> <TR><TD ><A HREF = "fix_store_state.html">fix store/state</A></TD><TD > per-atom vectors</TD><TD > per-atom vector/array</TD><TD ></TD></TR> <TR><TD > </TD></TR></TABLE></DIV> <HR> <A NAME = "howto_16"></A><H4>6.16 Thermostatting, barostatting, and computing temperature </H4> <P>Thermostatting means controlling the temperature of particles in an MD simulation. Barostatting means controlling the pressure. Since the pressure includes a kinetic component due to particle velocities, both these operations require calculation of the temperature. Typically a target temperature (T) and/or pressure (P) is specified by the user, and the thermostat or barostat attempts to equilibrate the system to the requested T and/or P. </P> <P>Temperature is computed as kinetic energy divided by some number of degrees of freedom (and the Boltzmann constant). Since kinetic energy is a function of particle velocity, there is often a need to distinguish between a particle's advection velocity (due to some aggregate motiion of particles) and its thermal velocity. The sum of the two is the particle's total velocity, but the latter is often what is wanted to compute a temperature. </P> <P>LAMMPS has several options for computing temperatures, any of which can be used in thermostatting and barostatting. These <A HREF = "compute.html">compute commands</A> calculate temperature, and the <A HREF = "compute_pressure.html">compute pressure</A> command calculates pressure. </P> <UL><LI><A HREF = "compute_temp.html">compute temp</A> <LI><A HREF = "compute_temp_sphere.html">compute temp/sphere</A> <LI><A HREF = "compute_temp_asphere.html">compute temp/asphere</A> <LI><A HREF = "compute_temp_com.html">compute temp/com</A> <LI><A HREF = "compute_temp_deform.html">compute temp/deform</A> <LI><A HREF = "compute_temp_partial.html">compute temp/partial</A> <LI><A HREF = "compute_temp_profile.html">compute temp/profile</A> <LI><A HREF = "compute_temp_ramp.html">compute temp/ramp</A> <LI><A HREF = "compute_temp_region.html">compute temp/region</A> </UL> <P>All but the first 3 calculate velocity biases (i.e. advection velocities) that are removed when computing the thermal temperature. <A HREF = "compute_temp_sphere.html">Compute temp/sphere</A> and <A HREF = "compute_temp_asphere.html">compute temp/asphere</A> compute kinetic energy for extended particles that includes rotational degrees of freedom. They both allow, as an extra argument, which is another temperature compute that subtracts a velocity bias. This allows the translational velocity of extended spherical or aspherical particles to be adjusted in prescribed ways. </P> <P>Thermostatting in LAMMPS is performed by <A HREF = "fix.html">fixes</A>, or in one case by a pair style. Four thermostatting fixes are currently available: Nose-Hoover (nvt), Berendsen, Langevin, and direct rescaling (temp/rescale). Dissipative particle dynamics (DPD) thermostatting can be invoked via the <I>dpd/tstat</I> pair style: </P> <UL><LI><A HREF = "fix_nh.html">fix nvt</A> <LI><A HREF = "fix_nvt_sphere.html">fix nvt/sphere</A> <LI><A HREF = "fix_nvt_asphere.html">fix nvt/asphere</A> <LI><A HREF = "fix_nvt_sllod.html">fix nvt/sllod</A> <LI><A HREF = "fix_temp_berendsen.html">fix temp/berendsen</A> <LI><A HREF = "fix_langevin.html">fix langevin</A> <LI><A HREF = "fix_temp_rescale.html">fix temp/rescale</A> <LI><A HREF = "pair_dpd.html">pair_style dpd/tstat</A> </UL> <P><A HREF = "fix_nh.html">Fix nvt</A> only thermostats the translational velocity of particles. <A HREF = "fix_nvt_sllod.html">Fix nvt/sllod</A> also does this, except that it subtracts out a velocity bias due to a deforming box and integrates the SLLOD equations of motion. See the <A HREF = "#howto_13">NEMD simulations</A> section of this page for further details. <A HREF = "fix_nvt_sphere.html">Fix nvt/sphere</A> and <A HREF = "fix_nvt_asphere.html">fix nvt/asphere</A> thermostat not only translation velocities but also rotational velocities for spherical and aspherical particles. </P> <P>DPD thermostatting alters pairwise interactions in a manner analagous to the per-particle thermostatting of <A HREF = "fix_langevin.html">fix langevin</A>. </P> <P>Any of the thermostatting fixes can use temperature computes that remove bias for two purposes: (a) computing the current temperature to compare to the requested target temperature, and (b) adjusting only the thermal temperature component of the particle's velocities. See the doc pages for the individual fixes and for the <A HREF = "fix_modify.html">fix_modify</A> command for instructions on how to assign a temperature compute to a thermostatting fix. For example, you can apply a thermostat to only the x and z components of velocity by using it in conjunction with <A HREF = "compute_temp_partial.html">compute temp/partial</A>. </P> <P>IMPORTANT NOTE: Only the nvt fixes perform time integration, meaning they update the velocities and positions of particles due to forces and velocities respectively. The other thermostat fixes only adjust velocities; they do NOT perform time integration updates. Thus they should be used in conjunction with a constant NVE integration fix such as these: </P> <UL><LI><A HREF = "fix_nve.html">fix nve</A> <LI><A HREF = "fix_nve_sphere.html">fix nve/sphere</A> <LI><A HREF = "fix_nve_asphere.html">fix nve/asphere</A> </UL> <P>Barostatting in LAMMPS is also performed by <A HREF = "fix.html">fixes</A>. Two barosttating methods are currently available: Nose-Hoover (npt and nph) and Berendsen: </P> <UL><LI><A HREF = "fix_nh.html">fix npt</A> <LI><A HREF = "fix_npt_sphere.html">fix npt/sphere</A> <LI><A HREF = "fix_npt_asphere.html">fix npt/asphere</A> <LI><A HREF = "fix_nh.html">fix nph</A> <LI><A HREF = "fix_press_berendsen.html">fix press/berendsen</A> </UL> <P>The <A HREF = "fix_nh.html">fix npt</A> commands include a Nose-Hoover thermostat and barostat. <A HREF = "fix_nh.html">Fix nph</A> is just a Nose/Hoover barostat; it does no thermostatting. Both <A HREF = "fix_nh.html">fix nph</A> and <A HREF = "fix_press_berendsen.html">fix press/bernendsen</A> can be used in conjunction with any of the thermostatting fixes. </P> <P>As with the thermostats, <A HREF = "fix_nh.html">fix npt</A> and <A HREF = "fix_nh.html">fix nph</A> only use translational motion of the particles in computing T and P and performing thermo/barostatting. <A HREF = "fix_npt_sphere.html">Fix npt/sphere</A> and <A HREF = "fix_npt_asphere.html">fix npt/asphere</A> thermo/barostat using not only translation velocities but also rotational velocities for spherical and aspherical particles. </P> <P>All of the barostatting fixes use the <A HREF = "compute_pressure.html">compute pressure</A> compute to calculate a current pressure. By default, this compute is created with a simple <A HREF = "compute_temp.html">compute temp</A> (see the last argument of the <A HREF = "compute_pressure.html">compute pressure</A> command), which is used to calculated the kinetic componenet of the pressure. The barostatting fixes can also use temperature computes that remove bias for the purpose of computing the kinetic componenet which contributes to the current pressure. See the doc pages for the individual fixes and for the <A HREF = "fix_modify.html">fix_modify</A> command for instructions on how to assign a temperature or pressure compute to a barostatting fix. </P> <P>IMPORTANT NOTE: As with the thermostats, the Nose/Hoover methods (<A HREF = "fix_nh.html">fix npt</A> and <A HREF = "fix_nh.html">fix nph</A>) perform time integration. <A HREF = "fix_press_berendsen.html">Fix press/berendsen</A> does NOT, so it should be used with one of the constant NVE fixes or with one of the NVT fixes. </P> <P>Finally, thermodynamic output, which can be setup via the <A HREF = "thermo_style.html">thermo_style</A> command, often includes temperature and pressure values. As explained on the doc page for the <A HREF = "thermo_style.html">thermo_style</A> command, the default T and P are setup by the thermo command itself. They are NOT the ones associated with any thermostatting or barostatting fix you have defined or with any compute that calculates a temperature or pressure. Thus if you want to view these values of T and P, you need to specify them explicitly via a <A HREF = "thermo_style.html">thermo_style custom</A> command. Or you can use the <A HREF = "thermo_modify.html">thermo_modify</A> command to re-define what temperature or pressure compute is used for default thermodynamic output. </P> <HR> <A NAME = "howto_17"></A><H4>6.17 Walls </H4> <P>Walls in an MD simulation are typically used to bound particle motion, i.e. to serve as a boundary condition. </P> <P>Walls in LAMMPS can be of rough (made of particles) or idealized surfaces. Ideal walls can be smooth, generating forces only in the normal direction, or frictional, generating forces also in the tangential direction. </P> <P>Rough walls, built of particles, can be created in various ways. The particles themselves can be generated like any other particle, via the <A HREF = "lattice.html">lattice</A> and <A HREF = "create_atoms.html">create_atoms</A> commands, or read in via the <A HREF = "read_data.html">read_data</A> command. </P> <P>Their motion can be constrained by many different commands, so that they do not move at all, move together as a group at constant velocity or in response to a net force acting on them, move in a prescribed fashion (e.g. rotate around a point), etc. Note that if a time integration fix like <A HREF = "fix_nve.html">fix nve</A> or <A HREF = "fix_nh.html">fix nvt</A> is not used with the group that contains wall particles, their positions and velocities will not be updated. </P> <UL><LI><A HREF = "fix_aveforce.html">fix aveforce</A> - set force on particles to average value, so they move together <LI><A HREF = "fix_setforce.html">fix setforce</A> - set force on particles to a value, e.g. 0.0 <LI><A HREF = "fix_freeze.html">fix freeze</A> - freeze particles for use as granular walls <LI><A HREF = "fix_nve_noforce.html">fix nve/noforce</A> - advect particles by their velocity, but without force <LI><A HREF = "fix_move.html">fix move</A> - prescribe motion of particles by a linear velocity, oscillation, rotation, variable </UL> <P>The <A HREF = "fix_move.html">fix move</A> command offers the most generality, since the motion of individual particles can be specified with <A HREF = "variable.html">variable</A> formula which depends on time and/or the particle position. </P> <P>For rough walls, it may be useful to turn off pairwise interactions between wall particles via the <A HREF = "neigh_modify.html">neigh_modify exclude</A> command. </P> <P>Rough walls can also be created by specifying frozen particles that do not move and do not interact with mobile particles, and then tethering other particles to the fixed particles, via a <A HREF = "bond_style.html">bond</A>. The bonded particles do interact with other mobile particles. </P> <P>Idealized walls can be specified via several fix commands. <A HREF = "fix_wall_gran.html">Fix wall/gran</A> creates frictional walls for use with granular particles; all the other commands create smooth walls. </P> <UL><LI><A HREF = "fix_wall_reflect.html">fix wall/reflect</A> - reflective flat walls <LI><A HREF = "fix_wall.html">fix wall/lj93</A> - flat walls, with Lennard-Jones 9/3 potential <LI><A HREF = "fix_wall.html">fix wall/lj126</A> - flat walls, with Lennard-Jones 12/6 potential <LI><A HREF = "fix_wall.html">fix wall/colloid</A> - flat walls, with <A HREF = "pair_colloid.html">pair_style colloid</A> potential <LI><A HREF = "fix_wall.html">fix wall/harmonic</A> - flat walls, with repulsive harmonic spring potential <LI><A HREF = "fix_wall_region.html">fix wall/region</A> - use region surface as wall <LI><A HREF = "fix_wall_gran.html">fix wall/gran</A> - flat or curved walls with <A HREF = "pair_gran.html">pair_style granular</A> potential </UL> <P>The <I>lj93</I>, <I>lj126</I>, <I>colloid</I>, and <I>harmonic</I> styles all allow the flat walls to move with a constant velocity, or oscillate in time. The <A HREF = "fix_wall_region.html">fix wall/region</A> command offers the most generality, since the region surface is treated as a wall, and the geometry of the region can be a simple primitive volume (e.g. a sphere, or cube, or plane), or a complex volume made from the union and intersection of primitive volumes. <A HREF = "region.html">Regions</A> can also specify a volume "interior" or "exterior" to the specified primitive shape or <I>union</I> or <I>intersection</I>. <A HREF = "region.html">Regions</A> can also be "dynamic" meaning they move with constant velocity, oscillate, or rotate. </P> <P>The only frictional idealized walls currently in LAMMPS are flat or curved surfaces specified by the <A HREF = "fix_wall_gran.html">fix wall/gran</A> command. At some point we plan to allow regoin surfaces to be used as frictional walls, as well as triangulated surfaces. </P> <HR> <A NAME = "howto_18"></A><H4>6.18 Elastic constants </H4> <P>Elastic constants characterize the stiffness of a material. The formal definition is provided by the linear relation that holds between the stress and strain tensors in the limit of infinitesimal deformation. In tensor notation, this is expressed as s_ij = C_ijkl * e_kl, where the repeated indices imply summation. s_ij are the elements of the symmetric stress tensor. e_kl are the elements of the symmetric strain tensor. C_ijkl are the elements of the fourth rank tensor of elastic constants. In three dimensions, this tensor has 3^4=81 elements. Using Voigt notation, the tensor can be written as a 6x6 matrix, where C_ij is now the derivative of s_i w.r.t. e_j. Because s_i is itself a derivative w.r.t. e_i, it follows that C_ij is also symmetric, with at most 7*6/2 = 21 distinct elements. </P> <P>At zero temperature, it is easy to estimate these derivatives by deforming the simulation box in one of the six directions using the <A HREF = "change_box.html">change_box</A> command and measuring the change in the stress tensor. A general-purpose script that does this is given in the examples/elastic directory described in <A HREF = "Section_example.html">this section</A>. </P> <P>Calculating elastic constants at finite temperature is more challenging, because it is necessary to run a simulation that perfoms time averages of differential properties. One way to do this is to measure the change in average stress tensor in an NVT simulations when the cell volume undergoes a finite deformation. In order to balance the systematic and statistical errors in this method, the magnitude of the deformation must be chosen judiciously, and care must be taken to fully equilibrate the deformed cell before sampling the stress tensor. Another approach is to sample the triclinic cell fluctuations that occur in an NPT simulation. This method can also be slow to converge and requires careful post-processing <A HREF = "#Shinoda">(Shinoda)</A> </P> <HR> <A NAME = "howto_19"></A><H4>6.19 Library interface to LAMMPS </H4> <P>As described in <A HREF = "Section_start.html#start_5">Section_start 4</A>, LAMMPS can be built as a library, so that it can be called by another code, used in a <A HREF = "Section_howto.html#howto_10">coupled manner</A> with other codes, or driven through a <A HREF = "Section_python.html">Python interface</A>. </P> <P>All of these methodologies use a C-style interface to LAMMPS that is provided in the files src/library.cpp and src/library.h. The functions therein have a C-style argument list, but contain C++ code you could write yourself in a C++ application that was invoking LAMMPS directly. The C++ code in the functions illustrates how to invoke internal LAMMPS operations. Note that LAMMPS classes are defined within a LAMMPS namespace (LAMMPS_NS) if you use them from another C++ application. </P> <P>Library.cpp contains these 4 functions: </P> <PRE>void lammps_open(int, char **, MPI_Comm, void **); void lammps_close(void *); void lammps_file(void *, char *); char *lammps_command(void *, char *); </PRE> <P>The lammps_open() function is used to initialize LAMMPS, passing in a list of strings as if they were <A HREF = "Section_start.html#start_7">command-line arguments</A> when LAMMPS is run in stand-alone mode from the command line, and a MPI communicator for LAMMPS to run under. It returns a ptr to the LAMMPS object that is created, and which is used in subsequent library calls. The lammps_open() function can be called multiple times, to create multiple instances of LAMMPS. </P> <P>LAMMPS will run on the set of processors in the communicator. This means the calling code can run LAMMPS on all or a subset of processors. For example, a wrapper script might decide to alternate between LAMMPS and another code, allowing them both to run on all the processors. Or it might allocate half the processors to LAMMPS and half to the other code and run both codes simultaneously before syncing them up periodically. Or it might instantiate multiple instances of LAMMPS to perform different calculations. </P> <P>The lammps_close() function is used to shut down an instance of LAMMPS and free all its memory. </P> <P>The lammps_file() and lammps_command() functions are used to pass a file or string to LAMMPS as if it were an input script or single command in an input script. Thus the calling code can read or generate a series of LAMMPS commands one line at a time and pass it thru the library interface to setup a problem and then run it, interleaving the lammps_command() calls with other calls to extract information from LAMMPS, perform its own operations, or call another code's library. </P> <P>Other useful functions are also included in library.cpp. For example: </P> <PRE>void *lammps_extract_global(void *, char *) void *lammps_extract_atom(void *, char *) void *lammps_extract_compute(void *, char *, int, int) void *lammps_extract_fix(void *, char *, int, int, int, int) void *lammps_extract_variable(void *, char *, char *) int lammps_get_natoms(void *) void lammps_get_coords(void *, double *) void lammps_put_coords(void *, double *) </PRE> <P>These can extract various global or per-atom quantities from LAMMPS as well as values calculated by a compute, fix, or variable. The "get" and "put" operations can retrieve and reset atom coordinates. See the library.cpp file and its associated header file library.h for details. </P> <P>The key idea of the library interface is that you can write any functions you wish to define how your code talks to LAMMPS and add them to src/library.cpp and src/library.h, as well as to the <A HREF = "Section_python.html">Python interface</A>. The routines you add can access or change any LAMMPS data you wish. The couple and python directories have example C++ and C and Python codes which show how a driver code can link to LAMMPS as a library, run LAMMPS on a subset of processors, grab data from LAMMPS, change it, and put it back into LAMMPS. </P> <HR> <A NAME = "howto_20"></A><H4>6.20 Calculating thermal conductivity </H4> <P>The thermal conductivity kappa of a material can be measured in at least 3 ways using various options in LAMMPS. (See <A HREF = "Section_howto.html#howto_21">this section</A> of the manual for an analogous discussion for viscosity). The thermal conducitivity tensor kappa is a measure of the propensity of a material to transmit heat energy in a diffusive manner as given by Fourier's law </P> <P>J = -kappa grad(T) </P> <P>where J is the heat flux in units of energy per area per time and grad(T) is the spatial gradient of temperature. The thermal conductivity thus has units of energy per distance per time per degree K and is often approximated as an isotropic quantity, i.e. as a scalar. </P> <P>The first method is to setup two thermostatted regions at opposite ends of a simulation box, or one in the middle and one at the end of a periodic box. By holding the two regions at different temperatures with a <A HREF = "Section_howto.html#howto_13">thermostatting fix</A>, the energy added to the hot region should equal the energy subtracted from the cold region and be proportional to the heat flux moving between the regions. See the paper by <A HREF = "#Ikeshoji">Ikeshoji and Hafskjold</A> for details of this idea. Note that thermostatting fixes such as <A HREF = "fix_nh.html">fix nvt</A>, <A HREF = "fix_langevin.html">fix langevin</A>, and <A HREF = "fix_temp_rescale.html">fix temp/rescale</A> store the cumulative energy they add/subtract. Alternatively, the <A HREF = "fix_heat.html">fix heat</A> command can used in place of thermostats on each of two regions, and the resulting temperatures of the two regions monitored with the "compute temp/region" command or the temperature profile of the intermediate region monitored with the <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> and <A HREF = "compute_ke_atom.html">compute ke/atom</A> commands. </P> <P>The second method is to perform a reverse non-equilibrium MD simulation using the <A HREF = "fix_thermal_conductivity.html">fix thermal/conductivity</A> command which implements the rNEMD algorithm of Muller-Plathe. Kinetic energy is swapped between atoms in two different layers of the simulation box. This induces a temperature gradient between the two layers which can be monitored with the <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> and <A HREF = "compute_ke_atom.html">compute ke/atom</A> commands. The fix tallies the cumulative energy transfer that it performs. See the <A HREF = "fix_thermal_conductivity.html">fix thermal/conductivity</A> command for details. </P> <P>The third method is based on the Green-Kubo (GK) formula which relates the ensemble average of the auto-correlation of the heat flux to kappa. The heat flux can be calculated from the fluctuations of per-atom potential and kinetic energies and per-atom stress tensor in a steady-state equilibrated simulation. This is in contrast to the two preceding non-equilibrium methods, where energy flows continuously between hot and cold regions of the simulation box. </P> <P>The <A HREF = "compute_heat_flux.html">compute heat/flux</A> command can calculate the needed heat flux and describes how to implement the Green_Kubo formalism using additional LAMMPS commands, such as the <A HREF = "fix_ave_correlate.html">fix ave/correlate</A> command to calculate the needed auto-correlation. See the doc page for the <A HREF = "compute_heat_flux.html">compute heat/flux</A> command for an example input script that calculates the thermal conductivity of solid Ar via the GK formalism. </P> <HR> <A NAME = "howto_21"></A><H4>6.21 Calculating viscosity </H4> <P>The shear viscosity eta of a fluid can be measured in at least 3 ways using various options in LAMMPS. (See <A HREF = "Section_howto.html#howto_20">this section</A> of the manual for an analogous discussion for thermal conductivity). Eta is a measure of the propensity of a fluid to transmit momentum in a direction perpendicular to the direction of velocity or momentum flow. Alternatively it is the resistance the fluid has to being sheared. It is given by </P> <P>J = -eta grad(Vstream) </P> <P>where J is the momentum flux in units of momentum per area per time. and grad(Vstream) is the spatial gradient of the velocity of the fluid moving in another direction, normal to the area through which the momentum flows. Viscosity thus has units of pressure-time. </P> <P>The first method is to perform a non-equlibrium MD (NEMD) simulation by shearing the simulation box via the <A HREF = "fix_deform.html">fix deform</A> command, and using the <A HREF = "fix_nvt_sllod.html">fix nvt/sllod</A> command to thermostat the fluid via the SLLOD equations of motion. The velocity profile setup in the fluid by this procedure can be monitored by the <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> command, which determines grad(Vstream) in the equation above. E.g. the derivative in the y-direction of the Vx component of fluid motion or grad(Vstream) = dVx/dy. In this case, the Pxy off-diagonal component of the pressure or stress tensor, as calculated by the <A HREF = "compute_pressure.html">compute pressure</A> command, can also be monitored, which is the J term in the equation above. See <A HREF = "Section_howto.html#howto_13">this section</A> of the manual for details on NEMD simulations. </P> <P>The second method is to perform a reverse non-equilibrium MD simulation using the <A HREF = "fix_viscosity.html">fix viscosity</A> command which implements the rNEMD algorithm of Muller-Plathe. Momentum in one dimension is swapped between atoms in two different layers of the simulation box in a different dimension. This induces a velocity gradient which can be monitored with the <A HREF = "fix_ave_spatial.html">fix ave/spatial</A> command. The fix tallies the cummulative momentum transfer that it performs. See the <A HREF = "fix_viscosity.html">fix viscosity</A> command for details. </P> <P>The third method is based on the Green-Kubo (GK) formula which relates the ensemble average of the auto-correlation of the stress/pressure tensor to eta. This can be done in a steady-state equilibrated simulation which is in contrast to the two preceding non-equilibrium methods, where momentum flows continuously through the simulation box. </P> <P>Here is an example input script that calculates the viscosity of liquid Ar via the GK formalism: </P> <PRE># Sample LAMMPS input script for viscosity of liquid Ar </PRE> <PRE>units real variable T equal 86.4956 variable V equal vol variable dt equal 4.0 variable p equal 400 # correlation length variable s equal 5 # sample interval variable d equal $p*$s # dump interval </PRE> <PRE># convert from LAMMPS real units to SI </PRE> <PRE>variable kB equal 1.3806504e-23 # [J/K/</B> Boltzmann variable atm2Pa equal 101325.0 variable A2m equal 1.0e-10 variable fs2s equal 1.0e-15 variable convert equal ${atm2Pa}*${atm2Pa}*${fs2s}*${A2m}*${A2m}*${A2m} </PRE> <PRE># setup problem </PRE> <PRE>dimension 3 boundary p p p lattice fcc 5.376 orient x 1 0 0 orient y 0 1 0 orient z 0 0 1 region box block 0 4 0 4 0 4 create_box 1 box create_atoms 1 box mass 1 39.948 pair_style lj/cut 13.0 pair_coeff * * 0.2381 3.405 timestep ${dt} thermo $d </PRE> <PRE># equilibration and thermalization </PRE> <PRE>velocity all create $T 102486 mom yes rot yes dist gaussian fix NVT all nvt temp $T $T 10 drag 0.2 run 8000 </PRE> <PRE># viscosity calculation, switch to NVE if desired </PRE> <PRE>#unfix NVT #fix NVE all nve </PRE> <PRE>reset_timestep 0 variable pxy equal pxy variable pxz equal pxz variable pyz equal pyz fix SS all ave/correlate $s $p $d & v_pxy v_pxz v_pyz type auto file S0St.dat ave running variable scale equal ${convert}/(${kB}*$T)*$V*$s*${dt} variable v11 equal trap(f_SS[3/</B>)*${scale} variable v22 equal trap(f_SS[4/</B>)*${scale} variable v33 equal trap(f_SS[5/</B>)*${scale} thermo_style custom step temp press v_pxy v_pxz v_pyz v_v11 v_v22 v_v33 run 100000 variable v equal (v_v11+v_v22+v_v33)/3.0 variable ndens equal count(all)/vol print "average viscosity: $v [Pa.s/</B> @ $T K, ${ndens} /A^3" </PRE> <HR> <HR> <A NAME = "Berendsen"></A> <P><B>(Berendsen)</B> Berendsen, Grigera, Straatsma, J Phys Chem, 91, 6269-6271 (1987). </P> <A NAME = "Cornell"></A> <P><B>(Cornell)</B> Cornell, Cieplak, Bayly, Gould, Merz, Ferguson, Spellmeyer, Fox, Caldwell, Kollman, JACS 117, 5179-5197 (1995). </P> <A NAME = "Horn"></A> <P><B>(Horn)</B> Horn, Swope, Pitera, Madura, Dick, Hura, and Head-Gordon, J Chem Phys, 120, 9665 (2004). </P> <A NAME = "Ikeshoji"></A> <P><B>(Ikeshoji)</B> Ikeshoji and Hafskjold, Molecular Physics, 81, 251-261 (1994). </P> <A NAME = "MacKerell"></A> <P><B>(MacKerell)</B> MacKerell, Bashford, Bellott, Dunbrack, Evanseck, Field, Fischer, Gao, Guo, Ha, et al, J Phys Chem, 102, 3586 (1998). </P> <A NAME = "Mayo"></A> <P><B>(Mayo)</B> Mayo, Olfason, Goddard III, J Phys Chem, 94, 8897-8909 (1990). </P> <A NAME = "Jorgensen"></A> <P><B>(Jorgensen)</B> Jorgensen, Chandrasekhar, Madura, Impey, Klein, J Chem Phys, 79, 926 (1983). </P> <A NAME = "Price"></A> <P><B>(Price)</B> Price and Brooks, J Chem Phys, 121, 10096 (2004). </P> <A NAME = "Shinoda"></A> <P><B>(Shinoda)</B> Shinoda, Shiga, and Mikami, Phys Rev B, 69, 134103 (2004). </P> </HTML> diff --git a/doc/kspace_modify.html b/doc/kspace_modify.html index 44ff5d864..266a54b48 100644 --- a/doc/kspace_modify.html +++ b/doc/kspace_modify.html @@ -1,121 +1,131 @@ <HTML> <CENTER><A HREF = "http://lammps.sandia.gov">LAMMPS WWW Site</A> - <A HREF = "Manual.html">LAMMPS Documentation</A> - <A HREF = "Section_commands.html#comm">LAMMPS Commands</A> </CENTER> <HR> <H3>kspace_modify command </H3> <P><B>Syntax:</B> </P> <PRE>kspace_modify keyword value ... </PRE> <UL><LI>one or more keyword/value pairs may be listed <LI>keyword = <I>mesh</I> or <I>order</I> or <I>gewald</I> or <I>slab</I> <PRE> <I>mesh</I> value = x y z x,y,z = PPPM FFT grid size in each dimension <I>order</I> value = N N = grid extent of Gaussian for PPPM mapping of each charge <I>force</I> value = accuracy (force units) <I>gewald</I> value = rinv (1/distance units) rinv = PPPM G-ewald parameter <I>slab</I> value = volfactor volfactor = ratio of the total extended volume used in the - 2d approximation compared with the volume of the simulation domain + 2d approximation compared with the volume of the simulation domain + <I>compute</I> value = <I>yes</I> or <I>no</I> </PRE> </UL> <P><B>Examples:</B> </P> <PRE>kspace_modify mesh 24 24 30 order 6 kspace_modify slab 3.0 </PRE> <P><B>Description:</B> </P> <P>Set parameters used by the kspace solvers defined by the <A HREF = "kspace_style.html">kspace_style</A> command. Not all parameters are relevant to all kspace styles. </P> <P>The <I>mesh</I> keyword sets the 3d FFT grid size for kspace style pppm. Each dimension must be factorizable into powers of 2, 3, and 5. When this option is not set, the PPPM solver chooses its own grid size, consistent with the user-specified accuracy and pairwise cutoff. Values for x,y,z of 0,0,0 unset the option. </P> <P>The <I>order</I> keyword determines how many grid spacings an atom's charge extends when it is mapped to the FFT grid in kspace style pppm. The default for this parameter is 5, which means each charge spans 5 grid cells in each dimension. The larger the value of this parameter, the smaller the FFT grid will need to be to achieve the requested precision. Conversely, the smaller the order value, the larger the grid will be. Note that there is an inherent trade-off involved: a small grid will lower the cost of FFTs, but a large order parameter will increase the cost of intepolating charge/fields to/from the grid. And vice versa. </P> <P>The order parameter may be reset by LAMMPS when it sets up the PPPM FFT grid if the implied grid stencil extends beyond the grid cells owned by neighboring processors. Typically this will only occur when small problems are run on large numbers of processors. A warning will be generated indicating the order parameter is being reduced to allow LAMMPS to run the problem. </P> <P>The <I>force</I> keyword overrides the relative accuracy parameter set by the <A HREF = "kspace_style.html">kspace_style</A> command with an absolute accuracy. The accuracy determines the RMS error in per-atom forces calculated by the long-range solver and is thus specified in force units. A negative value for the accuracy setting means to use the relative accuracy parameter. The accuracy setting is used in conjunction with the pairwise cutoff to determine the number of K-space vectors for style <I>ewald</I> or the FFT grid size for style <I>pppm</I>. </P> <P>The <I>gewald</I> keyword sets the value of the Ewald or PPPM G-ewald parameter as <I>rinv</I> in reciprocal distance units. Without this setting, LAMMPS chooses the parameter automatically as a function of cutoff, precision, grid spacing, etc. This means it can vary from one simulation to the next which may not be desirable for matching a KSpace solver to a pre-tabulated pairwise potential. This setting can also be useful if Ewald or PPPM fails to choose a good grid spacing and G-ewald parameter automatically. If the value is set to 0.0, LAMMPS will choose the G-ewald parameter automatically. </P> <P>The <I>slab</I> keyword allows an Ewald or PPPM solver to be used for a systems that are periodic in x,y but non-periodic in z - a <A HREF = "boundary.html">boundary</A> setting of "boundary p p f". This is done by treating the system as if it were periodic in z, but inserting empty volume between atom slabs and removing dipole inter-slab interactions so that slab-slab interactions are effectively turned off. The volfactor value sets the ratio of the extended dimension in z divided by the actual dimension in z. The recommended value is 3.0. A larger value is inefficient; a smaller value introduces unwanted slab-slab interactions. The use of fixed boundaries in z means that the user must prevent particle migration beyond the initial z-bounds, typically by providing a wall-style fix. The methodology behind the <I>slab</I> option is explained in the paper by <A HREF = "#Yeh">(Yeh)</A>. </P> +<P>The <I>compute</I> keyword allows Kspace computations to be turned off, +even though a <A HREF = "kspace_style.html">kspace_style</A> is defined. This is +not useful for running a real simulation, but can be useful for +debugging purposes or for computing only partial forces that do not +include the Kspace contribution. You can also do this by simply not +defining a <A HREF = "kspace_style.html">kspace_style</A>, but a Kspace-compatible +<A HREF = "pair_style.html">pair_style</A> requires a kspace_style to be defined. +This keyword gives you that option. +</P> <P><B>Restrictions:</B> none </P> <P><B>Related commands:</B> </P> <P><A HREF = "kspace_style.html">kspace_style</A>, <A HREF = "boundary.html">boundary</A> </P> <P><B>Default:</B> </P> <P>The option defaults are mesh = 0 0 0, order = 5, force = -1.0, gewald -= 0.0, and slab = 1.0. += 0.0, slab = 1.0, and compute = yes. </P> <HR> <A NAME = "Yeh"></A> <P><B>(Yeh)</B> Yeh and Berkowitz, J Chem Phys, 111, 3155 (1999). </P> </HTML> diff --git a/doc/kspace_modify.txt b/doc/kspace_modify.txt index 6a4b54d93..b679c67e5 100644 --- a/doc/kspace_modify.txt +++ b/doc/kspace_modify.txt @@ -1,112 +1,122 @@ "LAMMPS WWW Site"_lws - "LAMMPS Documentation"_ld - "LAMMPS Commands"_lc :c :link(lws,http://lammps.sandia.gov) :link(ld,Manual.html) :link(lc,Section_commands.html#comm) :line kspace_modify command :h3 [Syntax:] kspace_modify keyword value ... :pre one or more keyword/value pairs may be listed :ulb,l keyword = {mesh} or {order} or {gewald} or {slab} :l {mesh} value = x y z x,y,z = PPPM FFT grid size in each dimension {order} value = N N = grid extent of Gaussian for PPPM mapping of each charge {force} value = accuracy (force units) {gewald} value = rinv (1/distance units) rinv = PPPM G-ewald parameter {slab} value = volfactor volfactor = ratio of the total extended volume used in the - 2d approximation compared with the volume of the simulation domain :pre + 2d approximation compared with the volume of the simulation domain + {compute} value = {yes} or {no} :pre :ule [Examples:] kspace_modify mesh 24 24 30 order 6 kspace_modify slab 3.0 :pre [Description:] Set parameters used by the kspace solvers defined by the "kspace_style"_kspace_style.html command. Not all parameters are relevant to all kspace styles. The {mesh} keyword sets the 3d FFT grid size for kspace style pppm. Each dimension must be factorizable into powers of 2, 3, and 5. When this option is not set, the PPPM solver chooses its own grid size, consistent with the user-specified accuracy and pairwise cutoff. Values for x,y,z of 0,0,0 unset the option. The {order} keyword determines how many grid spacings an atom's charge extends when it is mapped to the FFT grid in kspace style pppm. The default for this parameter is 5, which means each charge spans 5 grid cells in each dimension. The larger the value of this parameter, the smaller the FFT grid will need to be to achieve the requested precision. Conversely, the smaller the order value, the larger the grid will be. Note that there is an inherent trade-off involved: a small grid will lower the cost of FFTs, but a large order parameter will increase the cost of intepolating charge/fields to/from the grid. And vice versa. The order parameter may be reset by LAMMPS when it sets up the PPPM FFT grid if the implied grid stencil extends beyond the grid cells owned by neighboring processors. Typically this will only occur when small problems are run on large numbers of processors. A warning will be generated indicating the order parameter is being reduced to allow LAMMPS to run the problem. The {force} keyword overrides the relative accuracy parameter set by the "kspace_style"_kspace_style.html command with an absolute accuracy. The accuracy determines the RMS error in per-atom forces calculated by the long-range solver and is thus specified in force units. A negative value for the accuracy setting means to use the relative accuracy parameter. The accuracy setting is used in conjunction with the pairwise cutoff to determine the number of K-space vectors for style {ewald} or the FFT grid size for style {pppm}. The {gewald} keyword sets the value of the Ewald or PPPM G-ewald parameter as {rinv} in reciprocal distance units. Without this setting, LAMMPS chooses the parameter automatically as a function of cutoff, precision, grid spacing, etc. This means it can vary from one simulation to the next which may not be desirable for matching a KSpace solver to a pre-tabulated pairwise potential. This setting can also be useful if Ewald or PPPM fails to choose a good grid spacing and G-ewald parameter automatically. If the value is set to 0.0, LAMMPS will choose the G-ewald parameter automatically. The {slab} keyword allows an Ewald or PPPM solver to be used for a systems that are periodic in x,y but non-periodic in z - a "boundary"_boundary.html setting of "boundary p p f". This is done by treating the system as if it were periodic in z, but inserting empty volume between atom slabs and removing dipole inter-slab interactions so that slab-slab interactions are effectively turned off. The volfactor value sets the ratio of the extended dimension in z divided by the actual dimension in z. The recommended value is 3.0. A larger value is inefficient; a smaller value introduces unwanted slab-slab interactions. The use of fixed boundaries in z means that the user must prevent particle migration beyond the initial z-bounds, typically by providing a wall-style fix. The methodology behind the {slab} option is explained in the paper by "(Yeh)"_#Yeh. +The {compute} keyword allows Kspace computations to be turned off, +even though a "kspace_style"_kspace_style.html is defined. This is +not useful for running a real simulation, but can be useful for +debugging purposes or for computing only partial forces that do not +include the Kspace contribution. You can also do this by simply not +defining a "kspace_style"_kspace_style.html, but a Kspace-compatible +"pair_style"_pair_style.html requires a kspace_style to be defined. +This keyword gives you that option. + [Restrictions:] none [Related commands:] "kspace_style"_kspace_style.html, "boundary"_boundary.html [Default:] The option defaults are mesh = 0 0 0, order = 5, force = -1.0, gewald -= 0.0, and slab = 1.0. += 0.0, slab = 1.0, and compute = yes. :line :link(Yeh) [(Yeh)] Yeh and Berkowitz, J Chem Phys, 111, 3155 (1999). diff --git a/doc/pair_modify.html b/doc/pair_modify.html index 4aa2b2648..8e163f6aa 100644 --- a/doc/pair_modify.html +++ b/doc/pair_modify.html @@ -1,175 +1,185 @@ <HTML> <CENTER><A HREF = "http://lammps.sandia.gov">LAMMPS WWW Site</A> - <A HREF = "Manual.html">LAMMPS Documentation</A> - <A HREF = "Section_commands.html#comm">LAMMPS Commands</A> </CENTER> <HR> <H3>pair_modify command </H3> <P><B>Syntax:</B> </P> <PRE>pair_modify keyword value ... </PRE> <UL><LI>one or more keyword/value pairs may be listed -<LI>keyword = <I>shift</I> or <I>mix</I> or <I>table</I> or <I>tabinner</I> or <I>tail</I> +<LI>keyword = <I>shift</I> or <I>mix</I> or <I>table</I> or <I>tabinner</I> or <I>tail</I> or <I>compute</I> <PRE> <I>mix</I> value = <I>geometric</I> or <I>arithmetic</I> or <I>sixthpower</I> <I>shift</I> value = <I>yes</I> or <I>no</I> <I>table</I> value = N 2^N = # of values in table <I>tabinner</I> value = cutoff cutoff = inner cutoff at which to begin table (distance units) - <I>tail</I> value = <I>yes</I> or <I>no</I> + <I>tail</I> value = <I>yes</I> or <I>no</I> + <I>compute</I> value = <I>yes</I> or <I>no</I> </PRE> </UL> <P><B>Examples:</B> </P> <PRE>pair_modify shift yes mix geometric pair_modify tail yes pair_modify table 12 </PRE> <P><B>Description:</B> </P> <P>Modify the parameters of the currently defined pair style. Not all parameters are relevant to all pair styles. </P> <P>The <I>mix</I> keyword affects pair coefficients for interactions between atoms of type I and J, when I != J and the coefficients are not explicitly set in the input script. Note that coefficients for I = J must be set explicitly, either in the input script via the "pair_coeff" command or in the "Pair Coeffs" section of the <A HREF = "read_data.html">data file</A>. For some pair styles it is not necessary to specify coefficients when I != J, since a "mixing" rule will create them from the I,I and J,J settings. The pair_modify <I>mix</I> value determines what formulas are used to compute the mixed coefficients. In each case, the cutoff distance is mixed the same way as sigma. </P> <P>Note that not all pair styles support mixing. Also, some mix options are not available for certain pair styles. See the doc page for individual pair styles for those restrictions. Note also that the <A HREF = "pair_coeff.html">pair_coeff</A> command also can be to directly set coefficients for a specific I != J pairing, in which case no mixing is performed. </P> <P>mix <I>geometric</I> </P> <PRE>epsilon_ij = sqrt(epsilon_i * epsilon_j) sigma_ij = sqrt(sigma_i * sigma_j) </PRE> <P>mix <I>arithmetic</I> </P> <PRE>epsilon_ij = sqrt(epsilon_i * epsilon_j) sigma_ij = (sigma_i + sigma_j) / 2 </PRE> <P>mix <I>sixthpower</I> </P> <PRE>epsilon_ij = (2 * sqrt(epsilon_i*epsilon_j) * sigma_i^3 * sigma_j^3) / (sigma_i^6 + sigma_j^6) sigma_ij = ((sigma_i**6 + sigma_j**6) / 2) ^ (1/6) </PRE> <P>The <I>shift</I> keyword determines whether a Lennard-Jones potential is shifted at its cutoff to 0.0. If so, this adds an energy term to each pairwise interaction which will be included in the thermodynamic output, but does not affect pair forces or atom trajectories. See the doc page for individual pair styles to see which ones support this option. </P> <P>The <I>table</I> keyword applies to pair styles with a long-range Coulombic term; see the doc page for individual styles to see which potentials support this option. If N is non-zero, a table of length 2^N is pre-computed for forces and energies, which can shrink their computational cost by up to a factor of 2. The table is indexed via a bit-mapping technique <A HREF = "#Wolff">(Wolff)</A> and a linear interpolation is performed between adjacent table values. In our experiments with different table styles (lookup, linear, spline), this method typically gave the best performance in terms of speed and accuracy. </P> <P>The choice of table length is a tradeoff in accuracy versus speed. A larger N yields more accurate force computations, but requires more memory which can slow down the computation due to cache misses. A reasonable value of N is between 8 and 16. The default value of 12 (table of length 4096) gives approximately the same accuracy as the no-table (N = 0) option. For N = 0, forces and energies are computed directly, using a polynomial fit for the needed erfc() function evaluation, which is what earlier versions of LAMMPS did. Values greater than 16 typically slow down the simulation and will not improve accuracy; values from 1 to 8 give unreliable results. </P> <P>The <I>tabinner</I> keyword sets an inner cutoff above which the pairwise computation is done by table lookup (if tables are invoked). The smaller this value is set, the less accurate the table becomes (for a given number of table values), which can require use of larger tables. The default cutoff value is sqrt(2.0) distance units which means nearly all pairwise interactions are computed via table lookup for simulations with "real" units, but some close pairs may be computed directly (non-table) for simulations with "lj" units. </P> <P>When the <I>tail</I> keyword is set to <I>yes</I>, certain pair styles will add a long-range VanderWaals tail "correction" to the energy and pressure. See the doc page for individual styles to see which support this option. These corrections are included in the calculation and printing of thermodynamic quantities (see the <A HREF = "thermo_style.html">thermo_style</A> command). Their effect will also be included in constant NPT or NPH simulations where the pressure influences the simulation box dimensions (e.g. the <A HREF = "fix_nh.html">fix npt</A> and <A HREF = "fix_nh.html">fix nph</A> commands). The formulas used for the long-range corrections come from equation 5 of <A HREF = "#Sun">(Sun)</A>. </P> <P>Several assumptions are inherent in using tail corrections, including the following: </P> <UL><LI>The simulated system is a 3d bulk homogeneous liquid. This option should not be used for systems that are non-liquid, 2d, have a slab geometry (only 2d periodic), or inhomogeneous. <LI>G(r), the radial distribution function (rdf), is unity beyond the cutoff, so a fairly large cutoff should be used (i.e. 2.5 sigma for an LJ fluid), and it is probably a good idea to verify this assumption by checking the rdf. The rdf is not exactly unity beyond the cutoff for each pair of interaction types, so the tail correction is necessarily an approximation. <LI>Thermophysical properties obtained from calculations with this option enabled will not be thermodynamically consistent with the truncated force-field that was used. In other words, atoms do not feel any LJ pair interactions beyond the cutoff, but the energy and pressure reported by the simulation include an estimated contribution from those interactions. </UL> +<P>The <I>compute</I> keyword allows pairwise computations to be turned off, +even though a <A HREF = "pair_style.html">pair_style</A> is defined. This is not +useful for running a real simulation, but can be useful for debugging +purposes or for computing only partial forces that do not include the +pairwise contribution. You can also do this by simply not defining a +<A HREF = "pair_style.html">pair_style</A>, but a Kspace-compatible pair_style is +required if you also want to define a +<A HREF = "kspace_style.html">kspace_style</A>. This keyword gives you that option. +</P> <P><B>Restrictions:</B> none </P> <P>You cannot use <I>shift</I> yes with <I>tail</I> yes, since those are conflicting options. You cannot use <I>tail</I> yes with 2d simulations. </P> <P><B>Related commands:</B> </P> <P><A HREF = "pair_style.html">pair_style</A>, <A HREF = "pair_coeff.html">pair_coeff</A>, <A HREF = "thermo_style.html">thermo_style</A> </P> <P><B>Default:</B> </P> <P>The option defaults are mix = geometric, shift = no, table = 12, -tabinner = sqrt(2.0), tail = no. +tabinner = sqrt(2.0), tail = no, and compute = yes. </P> <P>Note that some pair styles perform mixing, but only a certain style of mixing. See the doc pages for individual pair styles for details. </P> <HR> <A NAME = "Wolff"></A> <P><B>(Wolff)</B> Wolff and Rudd, Comp Phys Comm, 120, 200-32 (1999). </P> <A NAME = "Sun"></A> <P><B>(Sun)</B> Sun, J Phys Chem B, 102, 7338-7364 (1998). </P> </HTML> diff --git a/doc/pair_modify.txt b/doc/pair_modify.txt index b09044548..2614e93db 100644 --- a/doc/pair_modify.txt +++ b/doc/pair_modify.txt @@ -1,165 +1,175 @@ "LAMMPS WWW Site"_lws - "LAMMPS Documentation"_ld - "LAMMPS Commands"_lc :c :link(lws,http://lammps.sandia.gov) :link(ld,Manual.html) :link(lc,Section_commands.html#comm) :line pair_modify command :h3 [Syntax:] pair_modify keyword value ... :pre one or more keyword/value pairs may be listed :ulb,l -keyword = {shift} or {mix} or {table} or {tabinner} or {tail} :l +keyword = {shift} or {mix} or {table} or {tabinner} or {tail} or {compute} :l {mix} value = {geometric} or {arithmetic} or {sixthpower} {shift} value = {yes} or {no} {table} value = N 2^N = # of values in table {tabinner} value = cutoff cutoff = inner cutoff at which to begin table (distance units) - {tail} value = {yes} or {no} :pre + {tail} value = {yes} or {no} + {compute} value = {yes} or {no} :pre :ule [Examples:] pair_modify shift yes mix geometric pair_modify tail yes pair_modify table 12 :pre [Description:] Modify the parameters of the currently defined pair style. Not all parameters are relevant to all pair styles. The {mix} keyword affects pair coefficients for interactions between atoms of type I and J, when I != J and the coefficients are not explicitly set in the input script. Note that coefficients for I = J must be set explicitly, either in the input script via the "pair_coeff" command or in the "Pair Coeffs" section of the "data file"_read_data.html. For some pair styles it is not necessary to specify coefficients when I != J, since a "mixing" rule will create them from the I,I and J,J settings. The pair_modify {mix} value determines what formulas are used to compute the mixed coefficients. In each case, the cutoff distance is mixed the same way as sigma. Note that not all pair styles support mixing. Also, some mix options are not available for certain pair styles. See the doc page for individual pair styles for those restrictions. Note also that the "pair_coeff"_pair_coeff.html command also can be to directly set coefficients for a specific I != J pairing, in which case no mixing is performed. mix {geometric} epsilon_ij = sqrt(epsilon_i * epsilon_j) sigma_ij = sqrt(sigma_i * sigma_j) :pre mix {arithmetic} epsilon_ij = sqrt(epsilon_i * epsilon_j) sigma_ij = (sigma_i + sigma_j) / 2 :pre mix {sixthpower} epsilon_ij = (2 * sqrt(epsilon_i*epsilon_j) * sigma_i^3 * sigma_j^3) / (sigma_i^6 + sigma_j^6) sigma_ij = ((sigma_i**6 + sigma_j**6) / 2) ^ (1/6) :pre The {shift} keyword determines whether a Lennard-Jones potential is shifted at its cutoff to 0.0. If so, this adds an energy term to each pairwise interaction which will be included in the thermodynamic output, but does not affect pair forces or atom trajectories. See the doc page for individual pair styles to see which ones support this option. The {table} keyword applies to pair styles with a long-range Coulombic term; see the doc page for individual styles to see which potentials support this option. If N is non-zero, a table of length 2^N is pre-computed for forces and energies, which can shrink their computational cost by up to a factor of 2. The table is indexed via a bit-mapping technique "(Wolff)"_#Wolff and a linear interpolation is performed between adjacent table values. In our experiments with different table styles (lookup, linear, spline), this method typically gave the best performance in terms of speed and accuracy. The choice of table length is a tradeoff in accuracy versus speed. A larger N yields more accurate force computations, but requires more memory which can slow down the computation due to cache misses. A reasonable value of N is between 8 and 16. The default value of 12 (table of length 4096) gives approximately the same accuracy as the no-table (N = 0) option. For N = 0, forces and energies are computed directly, using a polynomial fit for the needed erfc() function evaluation, which is what earlier versions of LAMMPS did. Values greater than 16 typically slow down the simulation and will not improve accuracy; values from 1 to 8 give unreliable results. The {tabinner} keyword sets an inner cutoff above which the pairwise computation is done by table lookup (if tables are invoked). The smaller this value is set, the less accurate the table becomes (for a given number of table values), which can require use of larger tables. The default cutoff value is sqrt(2.0) distance units which means nearly all pairwise interactions are computed via table lookup for simulations with "real" units, but some close pairs may be computed directly (non-table) for simulations with "lj" units. When the {tail} keyword is set to {yes}, certain pair styles will add a long-range VanderWaals tail "correction" to the energy and pressure. See the doc page for individual styles to see which support this option. These corrections are included in the calculation and printing of thermodynamic quantities (see the "thermo_style"_thermo_style.html command). Their effect will also be included in constant NPT or NPH simulations where the pressure influences the simulation box dimensions (e.g. the "fix npt"_fix_nh.html and "fix nph"_fix_nh.html commands). The formulas used for the long-range corrections come from equation 5 of "(Sun)"_#Sun. Several assumptions are inherent in using tail corrections, including the following: The simulated system is a 3d bulk homogeneous liquid. This option should not be used for systems that are non-liquid, 2d, have a slab geometry (only 2d periodic), or inhomogeneous. :ulb,l G(r), the radial distribution function (rdf), is unity beyond the cutoff, so a fairly large cutoff should be used (i.e. 2.5 sigma for an LJ fluid), and it is probably a good idea to verify this assumption by checking the rdf. The rdf is not exactly unity beyond the cutoff for each pair of interaction types, so the tail correction is necessarily an approximation. :l Thermophysical properties obtained from calculations with this option enabled will not be thermodynamically consistent with the truncated force-field that was used. In other words, atoms do not feel any LJ pair interactions beyond the cutoff, but the energy and pressure reported by the simulation include an estimated contribution from those interactions. :l,ule +The {compute} keyword allows pairwise computations to be turned off, +even though a "pair_style"_pair_style.html is defined. This is not +useful for running a real simulation, but can be useful for debugging +purposes or for computing only partial forces that do not include the +pairwise contribution. You can also do this by simply not defining a +"pair_style"_pair_style.html, but a Kspace-compatible pair_style is +required if you also want to define a +"kspace_style"_kspace_style.html. This keyword gives you that option. + [Restrictions:] none You cannot use {shift} yes with {tail} yes, since those are conflicting options. You cannot use {tail} yes with 2d simulations. [Related commands:] "pair_style"_pair_style.html, "pair_coeff"_pair_coeff.html, "thermo_style"_thermo_style.html [Default:] The option defaults are mix = geometric, shift = no, table = 12, -tabinner = sqrt(2.0), tail = no. +tabinner = sqrt(2.0), tail = no, and compute = yes. Note that some pair styles perform mixing, but only a certain style of mixing. See the doc pages for individual pair styles for details. :line :link(Wolff) [(Wolff)] Wolff and Rudd, Comp Phys Comm, 120, 200-32 (1999). :link(Sun) [(Sun)] Sun, J Phys Chem B, 102, 7338-7364 (1998). diff --git a/src/integrate.cpp b/src/integrate.cpp index 334333718..beef03c30 100644 --- a/src/integrate.cpp +++ b/src/integrate.cpp @@ -1,132 +1,147 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #include "stdlib.h" #include "integrate.h" #include "update.h" +#include "force.h" +#include "pair.h" +#include "kspace.h" #include "modify.h" #include "compute.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- */ Integrate::Integrate(LAMMPS *lmp, int narg, char **arg) : Pointers(lmp) { elist_global = elist_atom = NULL; vlist_global = vlist_atom = NULL; external_force_clear = 0; } /* ---------------------------------------------------------------------- */ Integrate::~Integrate() { delete [] elist_global; delete [] elist_atom; delete [] vlist_global; delete [] vlist_atom; } +/* ---------------------------------------------------------------------- */ + +void Integrate::init() +{ + // allow pair and Kspace compute() to be turned off via modify flags + + if (force->pair && force->pair->compute_flag) pair_compute_flag = 1; + else pair_compute_flag = 0; + if (force->kspace && force->kspace->compute_flag) kspace_compute_flag = 1; + else kspace_compute_flag = 0; +} + /* ---------------------------------------------------------------------- setup lists of computes for global and per-atom PE and pressure ------------------------------------------------------------------------- */ void Integrate::ev_setup() { delete [] elist_global; delete [] elist_atom; delete [] vlist_global; delete [] vlist_atom; elist_global = elist_atom = NULL; vlist_global = vlist_atom = NULL; nelist_global = nelist_atom = 0; nvlist_global = nvlist_atom = 0; for (int i = 0; i < modify->ncompute; i++) { if (modify->compute[i]->peflag) nelist_global++; if (modify->compute[i]->peatomflag) nelist_atom++; if (modify->compute[i]->pressflag) nvlist_global++; if (modify->compute[i]->pressatomflag) nvlist_atom++; } if (nelist_global) elist_global = new Compute*[nelist_global]; if (nelist_atom) elist_atom = new Compute*[nelist_atom]; if (nvlist_global) vlist_global = new Compute*[nvlist_global]; if (nvlist_atom) vlist_atom = new Compute*[nvlist_atom]; nelist_global = nelist_atom = 0; nvlist_global = nvlist_atom = 0; for (int i = 0; i < modify->ncompute; i++) { if (modify->compute[i]->peflag) elist_global[nelist_global++] = modify->compute[i]; if (modify->compute[i]->peatomflag) elist_atom[nelist_atom++] = modify->compute[i]; if (modify->compute[i]->pressflag) vlist_global[nvlist_global++] = modify->compute[i]; if (modify->compute[i]->pressatomflag) vlist_atom[nvlist_atom++] = modify->compute[i]; } } /* ---------------------------------------------------------------------- set eflag,vflag for current iteration invoke matchstep() on all timestep-dependent computes to clear their arrays eflag/vflag based on computes that need info on this ntimestep eflag = 0 = no energy computation eflag = 1 = global energy only eflag = 2 = per-atom energy only eflag = 3 = both global and per-atom energy vflag = 0 = no virial computation (pressure) vflag = 1 = global virial with pair portion via sum of pairwise interactions vflag = 2 = global virial with pair portion via F dot r including ghosts vflag = 4 = per-atom virial only vflag = 5 or 6 = both global and per-atom virial ------------------------------------------------------------------------- */ void Integrate::ev_set(bigint ntimestep) { int i,flag; flag = 0; int eflag_global = 0; for (i = 0; i < nelist_global; i++) if (elist_global[i]->matchstep(ntimestep)) flag = 1; if (flag) eflag_global = 1; flag = 0; int eflag_atom = 0; for (i = 0; i < nelist_atom; i++) if (elist_atom[i]->matchstep(ntimestep)) flag = 1; if (flag) eflag_atom = 2; if (eflag_global) update->eflag_global = ntimestep; if (eflag_atom) update->eflag_atom = ntimestep; eflag = eflag_global + eflag_atom; flag = 0; int vflag_global = 0; for (i = 0; i < nvlist_global; i++) if (vlist_global[i]->matchstep(ntimestep)) flag = 1; if (flag) vflag_global = virial_style; flag = 0; int vflag_atom = 0; for (i = 0; i < nvlist_atom; i++) if (vlist_atom[i]->matchstep(ntimestep)) flag = 1; if (flag) vflag_atom = 4; if (vflag_global) update->vflag_global = ntimestep; if (vflag_atom) update->vflag_atom = ntimestep; vflag = vflag_global + vflag_atom; } diff --git a/src/integrate.h b/src/integrate.h index 1a8e0124c..7195a677c 100644 --- a/src/integrate.h +++ b/src/integrate.h @@ -1,54 +1,57 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifndef LMP_INTEGRATE_H #define LMP_INTEGRATE_H #include "pointers.h" namespace LAMMPS_NS { class Integrate : protected Pointers { public: Integrate(class LAMMPS *, int, char **); virtual ~Integrate(); - virtual void init() = 0; + virtual void init(); virtual void setup() = 0; virtual void setup_minimal(int) = 0; virtual void run(int) = 0; virtual void cleanup() {} virtual void reset_dt() {} virtual bigint memory_usage() {return 0;} protected: int eflag,vflag; // flags for energy/virial computation int virial_style; // compute virial explicitly or implicitly int external_force_clear; // clear forces locally or externally int nelist_global,nelist_atom; // # of PE,virial computes to check int nvlist_global,nvlist_atom; class Compute **elist_global; // lists of PE,virial Computes class Compute **elist_atom; class Compute **vlist_global; class Compute **vlist_atom; + int pair_compute_flag; // 0 if pair->compute is skipped + int kspace_compute_flag; // 0 if kspace->compute is skipped + void ev_setup(); void ev_set(bigint); }; } #endif /* ERROR/WARNING messages: */ diff --git a/src/kspace.cpp b/src/kspace.cpp index 9913b4480..4c527f887 100644 --- a/src/kspace.cpp +++ b/src/kspace.cpp @@ -1,162 +1,180 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #include "stdlib.h" #include "string.h" #include "kspace.h" #include "atom.h" #include "comm.h" #include "force.h" #include "memory.h" #include "error.h" #include "suffix.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- */ KSpace::KSpace(LAMMPS *lmp, int narg, char **arg) : Pointers(lmp) { energy = 0.0; + virial[0] = virial[1] = virial[2] = virial[3] = virial[4] = virial[5] = 0.0; + + compute_flag = 1; order = 5; gridflag = 0; gewaldflag = 0; slabflag = 0; slab_volfactor = 1; suffix_flag = Suffix::NONE; accuracy_absolute = -1.0; two_charge_force = force->qqr2e * (force->qelectron * force->qelectron) / (force->angstrom * force->angstrom); maxeatom = maxvatom = 0; eatom = NULL; vatom = NULL; } /* ---------------------------------------------------------------------- */ KSpace::~KSpace() { memory->destroy(eatom); memory->destroy(vatom); } +/* ---------------------------------------------------------------------- */ + +void KSpace::compute_dummy(int eflag, int vflag) +{ + if (eflag || vflag) ev_setup(eflag,vflag); + else evflag = evflag_atom = eflag_global = vflag_global = + eflag_atom = vflag_atom = 0; +} + /* ---------------------------------------------------------------------- setup for energy, virial computation see integrate::ev_set() for values of eflag (0-3) and vflag (0-6) ------------------------------------------------------------------------- */ void KSpace::ev_setup(int eflag, int vflag) { int i,n; evflag = 1; eflag_either = eflag; eflag_global = eflag % 2; eflag_atom = eflag / 2; vflag_either = vflag; vflag_global = vflag % 4; vflag_atom = vflag / 4; if (eflag_atom || vflag_atom) evflag_atom = 1; else evflag_atom = 0; // reallocate per-atom arrays if necessary if (eflag_atom && atom->nlocal > maxeatom) { maxeatom = atom->nmax; memory->destroy(eatom); memory->create(eatom,maxeatom,"kspace:eatom"); } if (vflag_atom && atom->nlocal > maxvatom) { maxvatom = atom->nmax; memory->destroy(vatom); memory->create(vatom,maxvatom,6,"kspace:vatom"); } // zero accumulators if (eflag_global) energy = 0.0; if (vflag_global) for (i = 0; i < 6; i++) virial[i] = 0.0; if (eflag_atom) { n = atom->nlocal; for (i = 0; i < n; i++) eatom[i] = 0.0; } if (vflag_atom) { n = atom->nlocal; for (i = 0; i < n; i++) { vatom[i][0] = 0.0; vatom[i][1] = 0.0; vatom[i][2] = 0.0; vatom[i][3] = 0.0; vatom[i][4] = 0.0; vatom[i][5] = 0.0; } } } /* ---------------------------------------------------------------------- modify parameters of the KSpace style ------------------------------------------------------------------------- */ void KSpace::modify_params(int narg, char **arg) { int iarg = 0; while (iarg < narg) { if (strcmp(arg[iarg],"mesh") == 0) { if (iarg+4 > narg) error->all(FLERR,"Illegal kspace_modify command"); nx_pppm = atoi(arg[iarg+1]); ny_pppm = atoi(arg[iarg+2]); nz_pppm = atoi(arg[iarg+3]); if (nx_pppm == 0 && ny_pppm == 0 && nz_pppm == 0) gridflag = 0; else gridflag = 1; iarg += 4; } else if (strcmp(arg[iarg],"order") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal kspace_modify command"); order = atoi(arg[iarg+1]); iarg += 2; } else if (strcmp(arg[iarg],"force") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal kspace_modify command"); accuracy_absolute = atof(arg[iarg+1]); iarg += 2; } else if (strcmp(arg[iarg],"gewald") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal kspace_modify command"); g_ewald = atof(arg[iarg+1]); if (g_ewald == 0.0) gewaldflag = 0; else gewaldflag = 1; iarg += 2; } else if (strcmp(arg[iarg],"slab") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal kspace_modify command"); slab_volfactor = atof(arg[iarg+1]); iarg += 2; if (slab_volfactor <= 1.0) error->all(FLERR,"Bad kspace_modify slab parameter"); if (slab_volfactor < 2.0 && comm->me == 0) error->warning(FLERR,"Kspace_modify slab param < 2.0 may " "cause unphysical behavior"); slabflag = 1; + } else if (strcmp(arg[iarg],"compute") == 0) { + if (iarg+2 > narg) error->all(FLERR,"Illegal kspace_modify command"); + if (strcmp(arg[iarg+1],"yes") == 0) compute_flag = 1; + else if (strcmp(arg[iarg+1],"no") == 0) compute_flag = 0; + else error->all(FLERR,"Illegal kspace_modify command"); + iarg += 2; } else error->all(FLERR,"Illegal kspace_modify command"); } } /* ---------------------------------------------------------------------- */ void *KSpace::extract(const char *str) { if (strcmp(str,"scale") == 0) return (void *) &scale; return NULL; } diff --git a/src/kspace.h b/src/kspace.h index 30605b419..78c8b477e 100644 --- a/src/kspace.h +++ b/src/kspace.h @@ -1,87 +1,92 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifndef LMP_KSPACE_H #define LMP_KSPACE_H #include "pointers.h" namespace LAMMPS_NS { class KSpace : protected Pointers { friend class ThrOMP; friend class FixOMP; public: double energy; // accumulated energy double virial[6]; // accumlated virial double *eatom,**vatom; // accumulated per-atom energy/virial double g_ewald; int nx_pppm,ny_pppm,nz_pppm; + int compute_flag; // 0 if skip compute() + KSpace(class LAMMPS *, int, char **); virtual ~KSpace(); void modify_params(int, char **); void *extract(const char *); + void compute_dummy(int, int); + + // general child-class methods virtual void init() = 0; virtual void setup() = 0; virtual void compute(int, int) = 0; virtual void timing(int, double &, double &) {} virtual double memory_usage() {return 0.0;} protected: int gridflag,gewaldflag; int order; int slabflag; int suffix_flag; // suffix compatibility flag double scale; double slab_volfactor; double accuracy; // accuracy of KSpace solver (force units) double accuracy_absolute; // user-specifed accuracy in force units double accuracy_relative; // user-specified dimensionless accuracy // accurary = acc_rel * two_charge_force double two_charge_force; // force in user units of two point // charges separated by 1 Angstrom int evflag,evflag_atom; int eflag_either,eflag_global,eflag_atom; int vflag_either,vflag_global,vflag_atom; int maxeatom,maxvatom; void ev_setup(int, int); }; } #endif /* ERROR/WARNING messages: E: Illegal ... command Self-explanatory. Check the input script syntax and compare to the documentation for the command. You can use -echo screen as a command-line option when running LAMMPS to see the offending line. E: Bad kspace_modify slab parameter Kspace_modify value for the slab/volume keyword must be >= 2.0. W: Kspace_modify slab param < 2.0 may cause unphysical behavior The kspace_modify slab parameter should be larger to insure periodic grids padded with empty space do not overlap. */ diff --git a/src/min.cpp b/src/min.cpp index 35c33b3c2..7c5d5a2cc 100644 --- a/src/min.cpp +++ b/src/min.cpp @@ -1,776 +1,783 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- Contributing author: Aidan Thompson (SNL) improved CG and backtrack ls, added quadratic ls Sources: Numerical Recipes frprmn routine "Conjugate Gradient Method Without the Agonizing Pain" by JR Shewchuk, http://www-2.cs.cmu.edu/~jrs/jrspapers.html#cg ------------------------------------------------------------------------- */ #include "math.h" #include "stdlib.h" #include "string.h" #include "min.h" #include "atom.h" #include "domain.h" #include "comm.h" #include "update.h" #include "modify.h" #include "fix_minimize.h" #include "compute.h" #include "neighbor.h" #include "force.h" #include "pair.h" #include "bond.h" #include "angle.h" #include "dihedral.h" #include "improper.h" #include "kspace.h" #include "output.h" #include "thermo.h" #include "timer.h" #include "memory.h" #include "error.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- */ Min::Min(LAMMPS *lmp) : Pointers(lmp) { dmax = 0.1; searchflag = 0; linestyle = 0; elist_global = elist_atom = NULL; vlist_global = vlist_atom = NULL; nextra_global = 0; fextra = NULL; nextra_atom = 0; xextra_atom = fextra_atom = NULL; extra_peratom = extra_nlen = NULL; extra_max = NULL; requestor = NULL; external_force_clear = 0; } /* ---------------------------------------------------------------------- */ Min::~Min() { delete [] elist_global; delete [] elist_atom; delete [] vlist_global; delete [] vlist_atom; delete [] fextra; memory->sfree(xextra_atom); memory->sfree(fextra_atom); memory->destroy(extra_peratom); memory->destroy(extra_nlen); memory->destroy(extra_max); memory->sfree(requestor); } /* ---------------------------------------------------------------------- */ void Min::init() { // create fix needed for storing atom-based quantities // will delete it at end of run char **fixarg = new char*[3]; fixarg[0] = (char *) "MINIMIZE"; fixarg[1] = (char *) "all"; fixarg[2] = (char *) "MINIMIZE"; modify->add_fix(3,fixarg); delete [] fixarg; fix_minimize = (FixMinimize *) modify->fix[modify->nfix-1]; // clear out extra global and per-atom dof // will receive requests for new per-atom dof during pair init() // can then add vectors to fix_minimize in setup() nextra_global = 0; delete [] fextra; fextra = NULL; nextra_atom = 0; memory->sfree(xextra_atom); memory->sfree(fextra_atom); memory->destroy(extra_peratom); memory->destroy(extra_nlen); memory->destroy(extra_max); memory->sfree(requestor); xextra_atom = fextra_atom = NULL; extra_peratom = extra_nlen = NULL; extra_max = NULL; requestor = NULL; // virial_style: // 1 if computed explicitly by pair->compute via sum over pair interactions // 2 if computed implicitly by pair->virial_compute via sum over ghost atoms if (force->newton_pair) virial_style = 2; else virial_style = 1; // setup lists of computes for global and per-atom PE and pressure ev_setup(); // detect if fix omp is present for clearing force arrays int ifix = modify->find_fix("package_omp"); if (ifix >= 0) external_force_clear = 1; // set flags for what arrays to clear in force_clear() // need to clear additionals arrays if they exist torqueflag = 0; if (atom->torque_flag) torqueflag = 1; erforceflag = 0; if (atom->erforce_flag) erforceflag = 1; e_flag = 0; if (atom->e_flag) e_flag = 1; rho_flag = 0; if (atom->rho_flag) rho_flag = 1; + // allow pair and Kspace compute() to be turned off via modify flags + + if (force->pair && force->pair->compute_flag) pair_compute_flag = 1; + else pair_compute_flag = 0; + if (force->kspace && force->kspace->compute_flag) kspace_compute_flag = 1; + else kspace_compute_flag = 0; + // orthogonal vs triclinic simulation box triclinic = domain->triclinic; // reset reneighboring criteria if necessary neigh_every = neighbor->every; neigh_delay = neighbor->delay; neigh_dist_check = neighbor->dist_check; if (neigh_every != 1 || neigh_delay != 0 || neigh_dist_check != 1) { if (comm->me == 0) error->warning(FLERR, "Resetting reneighboring criteria during minimization"); } neighbor->every = 1; neighbor->delay = 0; neighbor->dist_check = 1; niter = neval = 0; - - // style-specific initialization - - init_style(); } /* ---------------------------------------------------------------------- setup before run ------------------------------------------------------------------------- */ void Min::setup() { if (comm->me == 0 && screen) fprintf(screen,"Setting up minimization ...\n"); // setup extra global dof due to fixes // cannot be done in init() b/c update init() is before modify init() nextra_global = modify->min_dof(); if (nextra_global) fextra = new double[nextra_global]; // compute for potential energy int id = modify->find_compute("thermo_pe"); if (id < 0) error->all(FLERR,"Minimization could not find thermo_pe compute"); pe_compute = modify->compute[id]; // style-specific setup does two tasks // setup extra global dof vectors // setup extra per-atom dof vectors due to requests from Pair classes // cannot be done in init() b/c update init() is before modify/pair init() setup_style(); // ndoftotal = total dof for entire minimization problem // dof for atoms, extra per-atom, extra global bigint ndofme = 3*atom->nlocal; for (int m = 0; m < nextra_atom; m++) ndofme += extra_peratom[m]*atom->nlocal; MPI_Allreduce(&ndofme,&ndoftotal,1,MPI_LMP_BIGINT,MPI_SUM,world); ndoftotal += nextra_global; // setup domain, communication and neighboring // acquire ghosts // build neighbor lists atom->setup(); if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); if (atom->sortfreq > 0) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; // remove these restriction eventually if (nextra_global && searchflag == 0) error->all(FLERR, "Cannot use a damped dynamics min style with fix box/relax"); if (nextra_atom && searchflag == 0) error->all(FLERR, "Cannot use a damped dynamics min style with per-atom DOF"); // atoms may have migrated in comm->exchange() reset_vectors(); // compute all forces ev_set(update->ntimestep); force_clear(); modify->setup_pre_force(vflag); - if (force->pair) force->pair->compute(eflag,vflag); + if (pair_compute_flag) force->pair->compute(eflag,vflag); + else if (force->pair) force->pair->compute_dummy(eflag,vflag); if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); } if (force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); + else force->kspace->compute_dummy(eflag,vflag); } if (force->newton) comm->reverse_comm(); // update per-atom minimization variables stored by pair styles if (nextra_atom) for (int m = 0; m < nextra_atom; m++) requestor[m]->min_xf_get(m); modify->setup(vflag); output->setup(1); // stats for Finish to print ecurrent = pe_compute->compute_scalar(); if (nextra_global) ecurrent += modify->min_energy(fextra); if (output->thermo->normflag) ecurrent /= atom->natoms; einitial = ecurrent; fnorm2_init = sqrt(fnorm_sqr()); fnorminf_init = fnorm_inf(); } /* ---------------------------------------------------------------------- setup without output or one-time post-init setup flag = 0 = just force calculation flag = 1 = reneighbor and force calculation ------------------------------------------------------------------------- */ void Min::setup_minimal(int flag) { // setup domain, communication and neighboring // acquire ghosts // build neighbor lists if (flag) { if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; } // atoms may have migrated in comm->exchange() reset_vectors(); // compute all forces ev_set(update->ntimestep); force_clear(); modify->setup_pre_force(vflag); - if (force->pair) force->pair->compute(eflag,vflag); + if (pair_compute_flag) force->pair->compute(eflag,vflag); + else if (force->pair) force->pair->compute_dummy(eflag,vflag); if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); } if (force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); + else force->kspace->compute_dummy(eflag,vflag); } if (force->newton) comm->reverse_comm(); // update per-atom minimization variables stored by pair styles if (nextra_atom) for (int m = 0; m < nextra_atom; m++) requestor[m]->min_xf_get(m); modify->setup(vflag); // stats for Finish to print ecurrent = pe_compute->compute_scalar(); if (nextra_global) ecurrent += modify->min_energy(fextra); if (output->thermo->normflag) ecurrent /= atom->natoms; einitial = ecurrent; fnorm2_init = sqrt(fnorm_sqr()); fnorminf_init = fnorm_inf(); } /* ---------------------------------------------------------------------- perform minimization, calling iterate() for N steps ------------------------------------------------------------------------- */ void Min::run(int n) { // minimizer iterations stop_condition = iterate(n); stopstr = stopstrings(stop_condition); // if early exit from iterate loop: // set update->nsteps to niter for Finish stats to print // set output->next values to this timestep // call energy_force() to insure vflag is set when forces computed // output->write does final output for thermo, dump, restart files // add ntimestep to all computes that store invocation times // since are hardwiring call to thermo/dumps and computes may not be ready if (stop_condition) { update->nsteps = niter; if (update->restrict_output == 0) { for (int idump = 0; idump < output->ndump; idump++) output->next_dump[idump] = update->ntimestep; output->next_dump_any = update->ntimestep; if (output->restart_every) output->next_restart = update->ntimestep; } output->next_thermo = update->ntimestep; modify->addstep_compute_all(update->ntimestep); ecurrent = energy_force(0); output->write(update->ntimestep); } } /* ---------------------------------------------------------------------- */ void Min::cleanup() { // stats for Finish to print efinal = ecurrent; fnorm2_final = sqrt(fnorm_sqr()); fnorminf_final = fnorm_inf(); // reset reneighboring criteria neighbor->every = neigh_every; neighbor->delay = neigh_delay; neighbor->dist_check = neigh_dist_check; // delete fix at end of run, so its atom arrays won't persist modify->delete_fix("MINIMIZE"); } /* ---------------------------------------------------------------------- evaluate potential energy and forces may migrate atoms due to reneighboring return new energy, which should include nextra_global dof return negative gradient stored in atom->f return negative gradient for nextra_global dof in fextra ------------------------------------------------------------------------- */ double Min::energy_force(int resetflag) { // check for reneighboring // always communicate since minimizer moved atoms int nflag = neighbor->decide(); if (nflag == 0) { timer->stamp(); comm->forward_comm(); timer->stamp(Timer::COMM); } else { if (modify->n_min_pre_exchange) { timer->stamp(); modify->min_pre_exchange(); timer->stamp(Timer::MODIFY); } if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); if (domain->box_change) { domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); } timer->stamp(); comm->exchange(); if (atom->sortfreq > 0 && update->ntimestep >= atom->nextsort) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); timer->stamp(Timer::COMM); neighbor->build(); timer->stamp(Timer::NEIGHBOR); } ev_set(update->ntimestep); force_clear(); timer->stamp(); if (modify->n_min_pre_force) { modify->min_pre_force(vflag); timer->stamp(Timer::MODIFY); } - if (force->pair) { + if (pair_compute_flag) { force->pair->compute(eflag,vflag); timer->stamp(Timer::PAIR); } if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); timer->stamp(Timer::BOND); } - if (force->kspace) { + if (kspace_compute_flag) { force->kspace->compute(eflag,vflag); timer->stamp(Timer::KSPACE); } if (force->newton) { comm->reverse_comm(); timer->stamp(Timer::COMM); } // update per-atom minimization variables stored by pair styles if (nextra_atom) for (int m = 0; m < nextra_atom; m++) requestor[m]->min_xf_get(m); // fixes that affect minimization if (modify->n_min_post_force) { timer->stamp(); modify->min_post_force(vflag); timer->stamp(Timer::MODIFY); } // compute potential energy of system // normalize if thermo PE does double energy = pe_compute->compute_scalar(); if (nextra_global) energy += modify->min_energy(fextra); if (output->thermo->normflag) energy /= atom->natoms; // if reneighbored, atoms migrated // if resetflag = 1, update x0 of atoms crossing PBC // reset vectors used by lo-level minimizer if (nflag) { if (resetflag) fix_minimize->reset_coords(); reset_vectors(); } return energy; } /* ---------------------------------------------------------------------- clear force on own & ghost atoms setup and clear other arrays as needed ------------------------------------------------------------------------- */ void Min::force_clear() { if (external_force_clear) return; int i; if (external_force_clear) return; // clear global force array // nall includes ghosts only if either newton flag is set int nall; if (force->newton) nall = atom->nlocal + atom->nghost; else nall = atom->nlocal; size_t nbytes = sizeof(double) * nall; if (nbytes) { memset(&(atom->f[0][0]),0,3*nbytes); if (torqueflag) memset(&(atom->torque[0][0]),0,3*nbytes); if (erforceflag) memset(&(atom->erforce[0]), 0, nbytes); if (e_flag) memset(&(atom->de[0]), 0, nbytes); if (rho_flag) memset(&(atom->drho[0]), 0, nbytes); } } /* ---------------------------------------------------------------------- pair style makes request to add a per-atom variables to minimization requestor stores callback to pair class to invoke during min to get current variable and forces on it and to update the variable return flag that pair can use if it registers multiple variables ------------------------------------------------------------------------- */ int Min::request(Pair *pair, int peratom, double maxvalue) { int n = nextra_atom + 1; xextra_atom = (double **) memory->srealloc(xextra_atom,n*sizeof(double *), "min:xextra_atom"); fextra_atom = (double **) memory->srealloc(fextra_atom,n*sizeof(double *), "min:fextra_atom"); memory->grow(extra_peratom,n,"min:extra_peratom"); memory->grow(extra_nlen,n,"min:extra_nlen"); memory->grow(extra_max,n,"min:extra_max"); requestor = (Pair **) memory->srealloc(requestor,n*sizeof(Pair *), "min:requestor"); requestor[nextra_atom] = pair; extra_peratom[nextra_atom] = peratom; extra_max[nextra_atom] = maxvalue; nextra_atom++; return nextra_atom-1; } /* ---------------------------------------------------------------------- */ void Min::modify_params(int narg, char **arg) { if (narg == 0) error->all(FLERR,"Illegal min_modify command"); int iarg = 0; while (iarg < narg) { if (strcmp(arg[iarg],"dmax") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal min_modify command"); dmax = atof(arg[iarg+1]); iarg += 2; } else if (strcmp(arg[iarg],"line") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal min_modify command"); if (strcmp(arg[iarg+1],"backtrack") == 0) linestyle = 0; else if (strcmp(arg[iarg+1],"quadratic") == 0) linestyle = 1; else if (strcmp(arg[iarg+1],"forcezero") == 0) linestyle = 2; else error->all(FLERR,"Illegal min_modify command"); iarg += 2; } else error->all(FLERR,"Illegal min_modify command"); } } /* ---------------------------------------------------------------------- setup lists of computes for global and per-atom PE and pressure ------------------------------------------------------------------------- */ void Min::ev_setup() { delete [] elist_global; delete [] elist_atom; delete [] vlist_global; delete [] vlist_atom; elist_global = elist_atom = NULL; vlist_global = vlist_atom = NULL; nelist_global = nelist_atom = 0; nvlist_global = nvlist_atom = 0; for (int i = 0; i < modify->ncompute; i++) { if (modify->compute[i]->peflag) nelist_global++; if (modify->compute[i]->peatomflag) nelist_atom++; if (modify->compute[i]->pressflag) nvlist_global++; if (modify->compute[i]->pressatomflag) nvlist_atom++; } if (nelist_global) elist_global = new Compute*[nelist_global]; if (nelist_atom) elist_atom = new Compute*[nelist_atom]; if (nvlist_global) vlist_global = new Compute*[nvlist_global]; if (nvlist_atom) vlist_atom = new Compute*[nvlist_atom]; nelist_global = nelist_atom = 0; nvlist_global = nvlist_atom = 0; for (int i = 0; i < modify->ncompute; i++) { if (modify->compute[i]->peflag) elist_global[nelist_global++] = modify->compute[i]; if (modify->compute[i]->peatomflag) elist_atom[nelist_atom++] = modify->compute[i]; if (modify->compute[i]->pressflag) vlist_global[nvlist_global++] = modify->compute[i]; if (modify->compute[i]->pressatomflag) vlist_atom[nvlist_atom++] = modify->compute[i]; } } /* ---------------------------------------------------------------------- set eflag,vflag for current iteration invoke matchstep() on all timestep-dependent computes to clear their arrays eflag/vflag based on computes that need info on this ntimestep always set eflag_global = 1, since need energy every iteration eflag = 0 = no energy computation eflag = 1 = global energy only eflag = 2 = per-atom energy only eflag = 3 = both global and per-atom energy vflag = 0 = no virial computation (pressure) vflag = 1 = global virial with pair portion via sum of pairwise interactions vflag = 2 = global virial with pair portion via F dot r including ghosts vflag = 4 = per-atom virial only vflag = 5 or 6 = both global and per-atom virial ------------------------------------------------------------------------- */ void Min::ev_set(bigint ntimestep) { int i,flag; int eflag_global = 1; for (i = 0; i < nelist_global; i++) elist_global[i]->matchstep(ntimestep); flag = 0; int eflag_atom = 0; for (i = 0; i < nelist_atom; i++) if (elist_atom[i]->matchstep(ntimestep)) flag = 1; if (flag) eflag_atom = 2; if (eflag_global) update->eflag_global = update->ntimestep; if (eflag_atom) update->eflag_atom = update->ntimestep; eflag = eflag_global + eflag_atom; flag = 0; int vflag_global = 0; for (i = 0; i < nvlist_global; i++) if (vlist_global[i]->matchstep(ntimestep)) flag = 1; if (flag) vflag_global = virial_style; flag = 0; int vflag_atom = 0; for (i = 0; i < nvlist_atom; i++) if (vlist_atom[i]->matchstep(ntimestep)) flag = 1; if (flag) vflag_atom = 4; if (vflag_global) update->vflag_global = update->ntimestep; if (vflag_atom) update->vflag_atom = update->ntimestep; vflag = vflag_global + vflag_atom; } /* ---------------------------------------------------------------------- compute and return ||force||_2^2 ------------------------------------------------------------------------- */ double Min::fnorm_sqr() { int i,n; double *fatom; double local_norm2_sqr = 0.0; for (i = 0; i < nvec; i++) local_norm2_sqr += fvec[i]*fvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) local_norm2_sqr += fatom[i]*fatom[i]; } } double norm2_sqr = 0.0; MPI_Allreduce(&local_norm2_sqr,&norm2_sqr,1,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) for (i = 0; i < nextra_global; i++) norm2_sqr += fextra[i]*fextra[i]; return norm2_sqr; } /* ---------------------------------------------------------------------- compute and return ||force||_inf ------------------------------------------------------------------------- */ double Min::fnorm_inf() { int i,n; double *fatom; double local_norm_inf = 0.0; for (i = 0; i < nvec; i++) local_norm_inf = MAX(fabs(fvec[i]),local_norm_inf); if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) local_norm_inf = MAX(fabs(fatom[i]),local_norm_inf); } } double norm_inf = 0.0; MPI_Allreduce(&local_norm_inf,&norm_inf,1,MPI_DOUBLE,MPI_MAX,world); if (nextra_global) for (i = 0; i < nextra_global; i++) norm_inf = MAX(fabs(fextra[i]),norm_inf); return norm_inf; } /* ---------------------------------------------------------------------- possible stop conditions ------------------------------------------------------------------------- */ char *Min::stopstrings(int n) { const char *strings[] = {"max iterations", "max force evaluations", "energy tolerance", "force tolerance", "search direction is not downhill", "linesearch alpha is zero", "forces are zero", "quadratic factors are zero", "trust region too small", "HFTN minimizer error"}; return (char *) strings[n]; } diff --git a/src/min.h b/src/min.h index c297de643..17a082e39 100644 --- a/src/min.h +++ b/src/min.h @@ -1,141 +1,144 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifndef LMP_MIN_H #define LMP_MIN_H #include "pointers.h" namespace LAMMPS_NS { class Min : protected Pointers { public: double einitial,efinal,eprevious; double fnorm2_init,fnorminf_init,fnorm2_final,fnorminf_final; double alpha_final; int niter,neval; int stop_condition; char *stopstr; int searchflag; // 0 if damped dynamics, 1 if sub-cycles on local search Min(class LAMMPS *); virtual ~Min(); - void init(); + virtual void init(); void setup(); void setup_minimal(int); void run(int); void cleanup(); int request(class Pair *, int, double); virtual bigint memory_usage() {return 0;} void modify_params(int, char **); double fnorm_sqr(); double fnorm_inf(); virtual void init_style() {} virtual void setup_style() = 0; virtual void reset_vectors() = 0; virtual int iterate(int) = 0; protected: int eflag,vflag; // flags for energy/virial computation int virial_style; // compute virial explicitly or implicitly int external_force_clear; // clear forces locally or externally double dmax; // max dist to move any atom in one step int linestyle; // 0 = backtrack, 1 = quadratic, 2 = forcezero int nelist_global,nelist_atom; // # of PE,virial computes to check int nvlist_global,nvlist_atom; class Compute **elist_global; // lists of PE,virial Computes class Compute **elist_atom; class Compute **vlist_global; class Compute **vlist_atom; int triclinic; // 0 if domain is orthog, 1 if triclinic int pairflag; int torqueflag,erforceflag; int e_flag,rho_flag; + int pair_compute_flag; // 0 if pair->compute is skipped + int kspace_compute_flag; // 0 if kspace->compute is skipped + int narray; // # of arrays stored by fix_minimize class FixMinimize *fix_minimize; // fix that stores auxiliary data class Compute *pe_compute; // compute for potential energy double ecurrent; // current potential energy bigint ndoftotal; // total dof for entire problem int nvec; // local atomic dof = length of xvec double *xvec; // variables for atomic dof, as 1d vector double *fvec; // force vector for atomic dof, as 1d vector int nextra_global; // # of extra global dof due to fixes double *fextra; // force vector for extra global dof // xextra is stored by fix int nextra_atom; // # of extra per-atom variables double **xextra_atom; // ptr to the variable double **fextra_atom; // ptr to the force on the variable int *extra_peratom; // # of values in variable, e.g. 3 in x int *extra_nlen; // total local length of variable, e.g 3*nlocal double *extra_max; // max allowed change per iter for atom's var class Pair **requestor; // Pair that stores/manipulates the variable int neigh_every,neigh_delay,neigh_dist_check; // neighboring params double energy_force(int); void force_clear(); double compute_force_norm_sqr(); double compute_force_norm_inf(); void ev_setup(); void ev_set(bigint); char *stopstrings(int); }; } #endif /* ERROR/WARNING messages: W: Resetting reneighboring criteria during minimization Minimization requires that neigh_modify settings be delay = 0, every = 1, check = yes. Since these settings were not in place, LAMMPS changed them and will restore them to their original values after the minimization. E: Minimization could not find thermo_pe compute This compute is created by the thermo command. It must have been explicitly deleted by a uncompute command. E: Cannot use a damped dynamics min style with fix box/relax This is a current restriction in LAMMPS. Use another minimizer style. E: Cannot use a damped dynamics min style with per-atom DOF This is a current restriction in LAMMPS. Use another minimizer style. E: Illegal ... command Self-explanatory. Check the input script syntax and compare to the documentation for the command. You can use -echo screen as a command-line option when running LAMMPS to see the offending line. */ diff --git a/src/min_fire.cpp b/src/min_fire.cpp index 0a67f4809..61946529d 100644 --- a/src/min_fire.cpp +++ b/src/min_fire.cpp @@ -1,270 +1,272 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #include "math.h" #include "min_fire.h" #include "universe.h" #include "atom.h" #include "force.h" #include "update.h" #include "output.h" #include "timer.h" #include "error.h" using namespace LAMMPS_NS; // EPS_ENERGY = minimum normalization for energy tolerance #define EPS_ENERGY 1.0e-8 // same as in other min classes enum{MAXITER,MAXEVAL,ETOL,FTOL,DOWNHILL,ZEROALPHA,ZEROFORCE,ZEROQUAD}; #define DELAYSTEP 5 #define DT_GROW 1.1 #define DT_SHRINK 0.5 #define ALPHA0 0.1 #define ALPHA_SHRINK 0.99 #define TMAX 10.0 /* ---------------------------------------------------------------------- */ MinFire::MinFire(LAMMPS *lmp) : Min(lmp) {} /* ---------------------------------------------------------------------- */ -void MinFire::init_style() +void MinFire::init() { + Min::init(); + dt = update->dt; } /* ---------------------------------------------------------------------- */ void MinFire::setup_style() { double **v = atom->v; int nlocal = atom->nlocal; for (int i = 0; i < nlocal; i++) v[i][0] = v[i][1] = v[i][2] = 0.0; } /* ---------------------------------------------------------------------- set current vector lengths and pointers called after atoms have migrated ------------------------------------------------------------------------- */ void MinFire::reset_vectors() { // atomic dof nvec = 3 * atom->nlocal; if (nvec) xvec = atom->x[0]; if (nvec) fvec = atom->f[0]; } /* ---------------------------------------------------------------------- */ int MinFire::iterate(int maxiter) { bigint ntimestep; double vmax,vdotf,vdotfall,vdotv,vdotvall,fdotf,fdotfall; double scale1,scale2; double dtvone,dtv,dtfm; int flag,flagall; alpha_final = 0.0; double alpha = ALPHA0; double dtmax = TMAX * dt; bigint last_negative = update->ntimestep; for (int iter = 0; iter < maxiter; iter++) { ntimestep = ++update->ntimestep; niter++; // vdotfall = v dot f double **v = atom->v; double **f = atom->f; int nlocal = atom->nlocal; vdotf = 0.0; for (int i = 0; i < nlocal; i++) vdotf += v[i][0]*f[i][0] + v[i][1]*f[i][1] + v[i][2]*f[i][2]; MPI_Allreduce(&vdotf,&vdotfall,1,MPI_DOUBLE,MPI_SUM,world); // sum vdotf over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { vdotf = vdotfall; MPI_Allreduce(&vdotf,&vdotfall,1,MPI_DOUBLE,MPI_SUM,universe->uworld); } // if (v dot f) > 0: // v = (1-alpha) v + alpha |v| Fhat // |v| = length of v, Fhat = unit f // if more than DELAYSTEP since v dot f was negative: // increase timestep and decrease alpha if (vdotfall > 0.0) { vdotv = 0.0; for (int i = 0; i < nlocal; i++) vdotv += v[i][0]*v[i][0] + v[i][1]*v[i][1] + v[i][2]*v[i][2]; MPI_Allreduce(&vdotv,&vdotvall,1,MPI_DOUBLE,MPI_SUM,world); // sum vdotv over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { vdotv = vdotvall; MPI_Allreduce(&vdotv,&vdotvall,1,MPI_DOUBLE,MPI_SUM,universe->uworld); } fdotf = 0.0; for (int i = 0; i < nlocal; i++) fdotf += f[i][0]*f[i][0] + f[i][1]*f[i][1] + f[i][2]*f[i][2]; MPI_Allreduce(&fdotf,&fdotfall,1,MPI_DOUBLE,MPI_SUM,world); // sum fdotf over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { fdotf = fdotfall; MPI_Allreduce(&fdotf,&fdotfall,1,MPI_DOUBLE,MPI_SUM,universe->uworld); } scale1 = 1.0 - alpha; if (fdotfall == 0.0) scale2 = 0.0; else scale2 = alpha * sqrt(vdotvall/fdotfall); for (int i = 0; i < nlocal; i++) { v[i][0] = scale1*v[i][0] + scale2*f[i][0]; v[i][1] = scale1*v[i][1] + scale2*f[i][1]; v[i][2] = scale1*v[i][2] + scale2*f[i][2]; } if (ntimestep - last_negative > DELAYSTEP) { dt = MIN(dt*DT_GROW,dtmax); alpha *= ALPHA_SHRINK; } // else (v dot f) <= 0: // decrease timestep, reset alpha, set v = 0 } else { last_negative = ntimestep; dt *= DT_SHRINK; alpha = ALPHA0; for (int i = 0; i < nlocal; i++) v[i][0] = v[i][1] = v[i][2] = 0.0; } // limit timestep so no particle moves further than dmax double *rmass = atom->rmass; double *mass = atom->mass; int *type = atom->type; dtvone = dt; for (int i = 0; i < nlocal; i++) { vmax = MAX(fabs(v[i][0]),fabs(v[i][1])); vmax = MAX(vmax,fabs(v[i][2])); if (dtvone*vmax > dmax) dtvone = dmax/vmax; } MPI_Allreduce(&dtvone,&dtv,1,MPI_DOUBLE,MPI_MIN,world); // min dtv over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { dtvone = dtv; MPI_Allreduce(&dtvone,&dtv,1,MPI_DOUBLE,MPI_MIN,universe->uworld); } // Euler integration step double **x = atom->x; if (rmass) { for (int i = 0; i < nlocal; i++) { dtfm = dtv / rmass[i]; x[i][0] += dtv * v[i][0]; x[i][1] += dtv * v[i][1]; x[i][2] += dtv * v[i][2]; v[i][0] += dtfm * f[i][0]; v[i][1] += dtfm * f[i][1]; v[i][2] += dtfm * f[i][2]; } } else { for (int i = 0; i < nlocal; i++) { dtfm = dtv / mass[type[i]]; x[i][0] += dtv * v[i][0]; x[i][1] += dtv * v[i][1]; x[i][2] += dtv * v[i][2]; v[i][0] += dtfm * f[i][0]; v[i][1] += dtfm * f[i][1]; v[i][2] += dtfm * f[i][2]; } } eprevious = ecurrent; ecurrent = energy_force(0); neval++; // energy tolerance criterion // only check after DELAYSTEP elapsed since velocties reset to 0 // sync across replicas if running multi-replica minimization if (update->etol > 0.0 && ntimestep-last_negative > DELAYSTEP) { if (update->multireplica == 0) { if (fabs(ecurrent-eprevious) < update->etol * 0.5*(fabs(ecurrent) + fabs(eprevious) + EPS_ENERGY)) return ETOL; } else { if (fabs(ecurrent-eprevious) < update->etol * 0.5*(fabs(ecurrent) + fabs(eprevious) + EPS_ENERGY)) flag = 0; else flag = 1; MPI_Allreduce(&flag,&flagall,1,MPI_INT,MPI_SUM,universe->uworld); if (flagall == 0) return ETOL; } } // force tolerance criterion // sync across replicas if running multi-replica minimization if (update->ftol > 0.0) { fdotf = fnorm_sqr(); if (update->multireplica == 0) { if (fdotf < update->ftol*update->ftol) return FTOL; } else { if (fdotf < update->ftol*update->ftol) flag = 0; else flag = 1; MPI_Allreduce(&flag,&flagall,1,MPI_INT,MPI_SUM,universe->uworld); if (flagall == 0) return FTOL; } } // output for thermo, dump, restart files if (output->next == ntimestep) { timer->stamp(); output->write(ntimestep); timer->stamp(Timer::OUTPUT); } } return MAXITER; } diff --git a/src/min_fire.h b/src/min_fire.h index 358d7ba22..7d54357ed 100644 --- a/src/min_fire.h +++ b/src/min_fire.h @@ -1,43 +1,43 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifdef MINIMIZE_CLASS MinimizeStyle(fire,MinFire) #else #ifndef LMP_MIN_FIRE_H #define LMP_MIN_FIRE_H #include "min.h" namespace LAMMPS_NS { class MinFire : public Min { public: MinFire(class LAMMPS *); ~MinFire() {} - void init_style(); + void init(); void setup_style(); void reset_vectors(); int iterate(int); private: double dt; }; } #endif #endif diff --git a/src/min_hftn.cpp b/src/min_hftn.cpp index f8f36d9b6..71ac7f909 100644 --- a/src/min_hftn.cpp +++ b/src/min_hftn.cpp @@ -1,1691 +1,1693 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- Author: Todd Plantenga (SNL) Sources: "Numerical Optimization", Nocedal and Wright, 2nd Ed, p170 "Parallel Unconstrained Min", Plantenga, SAND98-8201 ------------------------------------------------------------------------- */ #include "math.h" #include "string.h" #include "atom.h" #include "fix_minimize.h" #include "min_hftn.h" #include "modify.h" #include "output.h" #include "pair.h" #include "update.h" #include "timer.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- * This class performs Hessian-free truncated Newton minimization on an * unconstrained molecular potential. The algorithm avoids computing the * Hessian matrix, but obtains a near-quadratic rate of convergence. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- File local data ------------------------------------------------------------------------- */ //---- CONSTANTS MAP TO stopstrings DECLARED IN Min.run (min.cpp). static const int STOP_MAX_ITER = 0; //-- MAX ITERATIONS EXCEEDED static const int STOP_MAX_FORCE_EVALS = 1; //-- MAX FORCE EVALUATIONS EXCEEDED static const int STOP_ENERGY_TOL = 2; //-- STEP DID NOT CHANGE ENERGY static const int STOP_FORCE_TOL = 3; //-- CONVERGED TO DESIRED FORCE TOL static const int STOP_TR_TOO_SMALL = 8; //-- TRUST REGION TOO SMALL static const int STOP_ERROR = 9; //-- INTERNAL ERROR static const int NO_CGSTEP_BECAUSE_F_TOL_SATISFIED = 0; static const int CGSTEP_NEWTON = 1; static const int CGSTEP_TO_TR = 2; static const int CGSTEP_TO_DMAX = 3; static const int CGSTEP_NEGATIVE_CURVATURE = 4; static const int CGSTEP_MAX_INNER_ITERS = 5; static const int CGSTEP_UNDETERMINED = 6; //---- WHEN TESTING ENERGY_TOL, THE ENERGY MAGNITUDE MUST BE AT LEAST THIS BIG. static const double MIN_ETOL_MAG = 1.0e-8; //---- MACHINE PRECISION IS SOMETIMES DEFINED BY THE C RUNTIME. #ifdef DBL_EPSILON #define MACHINE_EPS DBL_EPSILON #else #define MACHINE_EPS 2.220446049250313e-16 #endif /* ---------------------------------------------------------------------- Constructor ------------------------------------------------------------------------- */ MinHFTN::MinHFTN(LAMMPS *lmp) : Min(lmp) { searchflag = 1; for (int i = 1; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daExtraGlobal[i] = NULL; for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daExtraAtom[i] = NULL; _fpPrint = NULL; return; } /* ---------------------------------------------------------------------- Destructor ------------------------------------------------------------------------- */ MinHFTN::~MinHFTN (void) { for (int i = 1; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) if (_daExtraGlobal[i] != NULL) delete [] _daExtraGlobal[i]; for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) if (_daExtraAtom[i] != NULL) delete [] _daExtraAtom[i]; return; } /* ---------------------------------------------------------------------- - Public method init_style + Public method init ------------------------------------------------------------------------- */ -void MinHFTN::init_style() +void MinHFTN::init() { + Min::init(); + for (int i = 1; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) { if (_daExtraGlobal[i] != NULL) delete [] _daExtraGlobal[i]; _daExtraGlobal[i] = NULL; } for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) { if (_daExtraAtom[i] != NULL) delete [] _daExtraAtom[i]; _daExtraAtom[i] = NULL; } return; } /* ---------------------------------------------------------------------- Public method setup_style ------------------------------------------------------------------------- */ void MinHFTN::setup_style() { //---- ALLOCATE MEMORY FOR ATOMIC DEGREES OF FREEDOM. for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) fix_minimize->add_vector(3); //---- ALLOCATE MEMORY FOR EXTRA GLOBAL DEGREES OF FREEDOM. //---- THE FIX MODULE TAKES CARE OF THE FIRST VECTOR, X0 (XK). if (nextra_global) { for (int i = 1; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daExtraGlobal[i] = new double[nextra_global]; } //---- ALLOCATE MEMORY FOR EXTRA PER-ATOM DEGREES OF FREEDOM. if (nextra_atom) { for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daExtraAtom[i] = new double*[nextra_atom]; for (int m = 0; m < nextra_atom; m++) { for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) fix_minimize->add_vector (extra_peratom[m]); } } return; } /* ---------------------------------------------------------------------- Public method reset_vectors After an energy/force calculation, atoms may migrate from one processor to another. Any local vector correlated with atom positions or forces must also be migrated. This is accomplished by a subclass of Fix. This method updates local pointers to the latest Fix copies. ------------------------------------------------------------------------- */ void MinHFTN::reset_vectors() { nvec = 3 * atom->nlocal; //---- ATOMIC DEGREES OF FREEDOM. if (nvec > 0) { xvec = atom->x[0]; fvec = atom->f[0]; } for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daAVectors[i] = fix_minimize->request_vector (i); //---- EXTRA PER-ATOM DEGREES OF FREEDOM. if (nextra_atom) { int n = NUM_HFTN_ATOM_BASED_VECTORS; for (int m = 0; m < nextra_atom; m++) { extra_nlen[m] = extra_peratom[m] * atom->nlocal; requestor[m]->min_xf_pointers(m,&xextra_atom[m],&fextra_atom[m]); for (int i = 0; i < NUM_HFTN_ATOM_BASED_VECTORS; i++) _daExtraAtom[i][m] = fix_minimize->request_vector (n++); } } return; } /* ---------------------------------------------------------------------- Public method iterate Upon entry, Min::setup() and Min::run have executed, and energy has already been evaluated at the initial point. Return an integer code that maps to a stop condition in min.cpp. ------------------------------------------------------------------------- */ int MinHFTN::iterate(int) { //---- TURN THIS ON TO GENERATE AN OPTIMIZATION PROGRESS FILE. bool bPrintProgress = false; if (bPrintProgress) open_hftn_print_file_(); double dFinalEnergy = 0.0; double dFinalFnorm2 = 0.0; modify->min_clearstore(); int nStopCode = execute_hftn_ (bPrintProgress, einitial, fnorm2_init, dFinalEnergy, dFinalFnorm2); modify->min_clearstore(); if (bPrintProgress) close_hftn_print_file_(); return( nStopCode ); } /* ---------------------------------------------------------------------- Private method execute_hftn_ @param[in] bPrintProgress - if true then print progress to a file @param[in] dInitialEnergy - energy at input x @param[in] dInitialForce2 - |F|_2 at input x @param[out] dFinalEnergy - energy at output x @param[out] dFinalForce2 - |F|_2 at output x Return stop code described in the enumeration at the top of this file, and the following: atom->x - positions at output x atom->f - forces evaluated at output x ------------------------------------------------------------------------- */ int MinHFTN::execute_hftn_(const bool bPrintProgress, const double dInitialEnergy, const double dInitialForce2, double & dFinalEnergy, double & dFinalForce2) { //---- DEFINE OUTPUTS PRINTED BY "Finish". eprevious = dInitialEnergy; alpha_final = 0.0; dFinalEnergy = dInitialEnergy; dFinalForce2 = dInitialForce2; if (dInitialForce2 < update->ftol) return( STOP_FORCE_TOL ); //---- SAVE ATOM POSITIONS BEFORE AN ITERATION. fix_minimize->store_box(); for (int i = 0; i < nvec; i++) _daAVectors[VEC_XK][i] = xvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xkAtom[i] = xatom[i]; } } if (nextra_global) modify->min_store(); double dXInf = calc_xinf_using_mpi_(); //---- FIND THE NUMBER OF UNKNOWNS. int nLocalNumUnknowns = nvec + nextra_atom; MPI_Allreduce (&nLocalNumUnknowns, &_nNumUnknowns, 1, MPI_INT, MPI_SUM, world); //---- INITIALIZE THE TRUST RADIUS BASED ON THE GRADIENT. double dTrustRadius = 1.5 * dInitialForce2; //---- TRUST RADIUS MUST KEEP STEPS FROM LETTING ATOMS MOVE SO FAR THEY //---- VIOLATE PHYSICS OR JUMP BEYOND A PARALLEL PROCESSING DOMAIN. //---- LINE SEARCH METHODS DO THIS BY RESTRICTING THE LARGEST CHANGE //---- OF ANY ATOM'S COMPONENT TO dmax. AN EXACT CHECK IS MADE LATER, //---- BUT THIS GUIDES DETERMINATION OF A MAX TRUST RADIUS. double dMaxTrustRadius = dmax * sqrt((double) _nNumUnknowns); dTrustRadius = MIN (dTrustRadius, dMaxTrustRadius); double dLastNewtonStep2 = dMaxTrustRadius; if (bPrintProgress) hftn_print_line_ (false, -1, neval, dInitialEnergy, dInitialForce2, -1, dTrustRadius, 0.0, 0.0, 0.0); bool bHaveEvaluatedAtX = true; double dCurrentEnergy = dInitialEnergy; double dCurrentForce2 = dInitialForce2; for (niter = 0; niter < update->nsteps; niter++) { (update->ntimestep)++; //---- CALL THE INNER LOOP TO GET THE NEXT TRUST REGION STEP. double dCgForce2StopTol = MIN ((dCurrentForce2 / 2.0), 0.1 / (niter+1)); dCgForce2StopTol = MAX (dCgForce2StopTol, update->ftol); double dNewEnergy; double dNewForce2; int nStepType; double dStepLength2; double dStepLengthInf; if (compute_inner_cg_step_ (dTrustRadius, dCgForce2StopTol, update->max_eval, bHaveEvaluatedAtX, dCurrentEnergy, dCurrentForce2, dNewEnergy, dNewForce2, nStepType, dStepLength2, dStepLengthInf) == false) { //---- THERE WAS AN ERROR. RESTORE TO LAST ACCEPTED STEP. if (nextra_global) modify->min_step (0.0, _daExtraGlobal[VEC_CG_P]); for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_XK][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = xkAtom[i]; requestor[m]->min_x_set(m); } } dFinalEnergy = energy_force (0); neval++; dFinalForce2 = sqrt (fnorm_sqr()); return( STOP_ERROR ); } //---- STOP IF THE CURRENT POSITION WAS FOUND TO BE ALREADY GOOD ENOUGH. //---- IN THIS CASE THE ENERGY AND FORCES ARE ALREADY COMPUTED. if (nStepType == NO_CGSTEP_BECAUSE_F_TOL_SATISFIED) { if (bPrintProgress) hftn_print_line_ (true, niter+1, neval, dNewEnergy, dNewForce2, nStepType, dTrustRadius, dStepLength2, 0.0, 0.0); dFinalEnergy = dNewEnergy; dFinalForce2 = dNewForce2; return( STOP_FORCE_TOL ); } //---- COMPUTE THE DIRECTIONAL DERIVATIVE H(x_k) p. bool bUseForwardDiffs = (dCurrentForce2 > 1000.0 * sqrt (MACHINE_EPS)); evaluate_dir_der_ (bUseForwardDiffs, VEC_CG_P, VEC_CG_HD, true, dCurrentEnergy); //---- COMPUTE p^T grad(x_k) AND SAVE IT FOR PRED. double dGradDotP = calc_grad_dot_v_using_mpi_ (VEC_CG_P); //---- MOVE TO THE NEW POINT AND EVALUATE ENERGY AND FORCES. //---- THIS IS THE PLACE WHERE energy_force IS ALLOWED TO RESET. for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_XK][i] + _daAVectors[VEC_CG_P][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; double * pAtom = _daExtraAtom[VEC_CG_P][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = xkAtom[i] + pAtom[i]; requestor[m]->min_x_set(m); } } if (nextra_global) modify->min_step (1.0, _daExtraGlobal[VEC_CG_P]); dNewEnergy = energy_force (1); neval++; dNewForce2 = sqrt (fnorm_sqr()); double dAred = dCurrentEnergy - dNewEnergy; //---- STOP IF THE FORCE TOLERANCE IS MET. if (dNewForce2 < update->ftol) { if (bPrintProgress) hftn_print_line_ (true, niter+1, neval, dNewEnergy, dNewForce2, nStepType, dTrustRadius, dStepLength2, dAred, -1.0); //---- (IMPLICITLY ACCEPT THE LAST STEP TO THE NEW POINT.) dFinalEnergy = dNewEnergy; dFinalForce2 = dNewForce2; return( STOP_FORCE_TOL ); } //---- STOP IF THE ACTUAL ENERGY REDUCTION IS TINY. if (nStepType != CGSTEP_TO_DMAX) { double dMag = 0.5 * (fabs (dCurrentEnergy) + fabs (dNewEnergy)); dMag = MAX (dMag, MIN_ETOL_MAG); if ( (fabs (dAred) < (update->etol * dMag)) || (dStepLengthInf == 0.0) ) { if (bPrintProgress) hftn_print_line_ (true, niter+1, neval, dNewEnergy, dNewForce2, nStepType, dTrustRadius, dStepLength2, dAred, -1.0); //---- (IMPLICITLY ACCEPT THE LAST STEP TO THE NEW POINT.) dFinalEnergy = dNewEnergy; dFinalForce2 = dNewForce2; return( STOP_ENERGY_TOL ); } } //---- COMPUTE THE PREDICTED REDUCTION - p^T grad - 0.5 p^T Hp double dPHP = calc_dot_prod_using_mpi_ (VEC_CG_P, VEC_CG_HD); double dPred = - dGradDotP - (0.5 * dPHP); //---- ACCEPT OR REJECT THE STEP PROPOSED BY THE INNER CG LOOP. //---- WHEN NEAR A SOLUTION, THE FORCE NORM IS PROBABLY MORE ACCURATE, //---- SO DON'T ACCEPT A STEP THAT REDUCES ENERGY SOME TINY AMOUNT //---- WHILE INCREASING THE FORCE NORM. bool bStepAccepted = (dAred > 0.0) && ( (dNewForce2 < dCurrentForce2) || (dCurrentForce2 > 1.0e-6)); if (bStepAccepted) { //---- THE STEP IS ACCEPTED. if (bPrintProgress) hftn_print_line_ (true, niter+1, neval, dNewEnergy, dNewForce2, nStepType, dTrustRadius, dStepLength2, dAred, dPred); fix_minimize->store_box(); modify->min_clearstore(); for (int i = 0; i < nvec; i++) _daAVectors[VEC_XK][i] = xvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xkAtom[i] = xatom[i]; } } if (nextra_global) modify->min_store(); if (niter > 0) eprevious = dCurrentEnergy; dCurrentEnergy = dNewEnergy; dCurrentForce2 = dNewForce2; bHaveEvaluatedAtX = true; if (nStepType == CGSTEP_NEWTON) dLastNewtonStep2 = dStepLength2; //---- UPDATE THE TRUST REGION BASED ON AGREEMENT BETWEEN //---- THE ACTUAL REDUCTION AND THE PREDICTED MODEL REDUCTION. if ((dAred > 0.75 * dPred) && (dStepLength2 >= 0.99 * dTrustRadius)) dTrustRadius = 2.0 * dTrustRadius; dTrustRadius = MIN (dTrustRadius, dMaxTrustRadius); //---- DMAX VIOLATIONS TRUNCATE THE CG STEP WITHOUT COMPARISONS; //---- BETTER TO ADJUST THE TRUST REGION SO DMAX STOPS HAPPENING. if (nStepType == CGSTEP_TO_DMAX) { if (dStepLength2 <= MACHINE_EPS) dTrustRadius = 0.1 * dTrustRadius; else dTrustRadius = MIN (dTrustRadius, 2.0 * dStepLength2); } } else { //---- THE STEP IS REJECTED. if (bPrintProgress) hftn_print_line_ (false, niter+1, neval, dCurrentEnergy, dCurrentForce2, nStepType, dTrustRadius, dStepLength2, dAred, dPred); //---- RESTORE THE LAST X_K POSITION. if (nextra_global) modify->min_step (0.0, _daExtraGlobal[VEC_CG_P]); for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_XK][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = xkAtom[i]; requestor[m]->min_x_set(m); } } modify->min_clearstore(); bHaveEvaluatedAtX = false; //---- UPDATE THE TRUST REGION. //---- EXPERIMENTS INDICATE NEGATIVE CURVATURE CAN TAKE A BAD //---- STEP A LONG WAY, SO BE MORE AGGRESSIVE IN THIS CASE. //---- ALSO, IF NEAR A SOLUTION AND DONE WITH NEWTON STEPS, //---- THEN REDUCE TO SOMETHING NEAR THE LAST GOOD NEWTON STEP. if ((nStepType == CGSTEP_NEGATIVE_CURVATURE) && (-dAred > dPred)) dTrustRadius = 0.10 * MIN (dTrustRadius, dStepLength2); else if ( (nStepType == CGSTEP_TO_DMAX) && (dStepLength2 <= MACHINE_EPS)) dTrustRadius = 0.10 * dTrustRadius; else if (-dAred > dPred) dTrustRadius = 0.20 * MIN (dTrustRadius, dStepLength2); else dTrustRadius = 0.25 * MIN (dTrustRadius, dStepLength2); if ( (nStepType != CGSTEP_NEWTON) && (dCurrentForce2 < sqrt (MACHINE_EPS))) dTrustRadius = MIN (dTrustRadius, 2.0 * dLastNewtonStep2); dLastNewtonStep2 = dMaxTrustRadius; //---- STOP IF THE TRUST RADIUS IS TOO SMALL TO CONTINUE. if ( (dTrustRadius <= 0.0) || (dTrustRadius <= MACHINE_EPS * MAX (1.0, dXInf))) { dFinalEnergy = dCurrentEnergy; dFinalForce2 = dCurrentForce2; return( STOP_TR_TOO_SMALL ); } } //---- OUTPUT FOR thermo, dump, restart FILES. if (output->next == update->ntimestep) { //---- IF THE LAST STEP WAS REJECTED, THEN REEVALUATE ENERGY AND //---- FORCES AT THE OLD POINT SO THE OUTPUT DOES NOT DISPLAY //---- THE INCREASED ENERGY OF THE REJECTED STEP. if (bStepAccepted == false) { dCurrentEnergy = energy_force (1); neval++; } timer->stamp(); output->write (update->ntimestep); timer->stamp (Timer::OUTPUT); } //---- RETURN IF NUMBER OF EVALUATIONS EXCEEDED. if (neval >= update->max_eval) { dFinalEnergy = dCurrentEnergy; dFinalForce2 = dCurrentForce2; return( STOP_MAX_FORCE_EVALS ); } } //-- END for LOOP OVER niter dFinalEnergy = dCurrentEnergy; dFinalForce2 = dCurrentForce2; return( STOP_MAX_ITER ); } /* ---------------------------------------------------------------------- Private method compute_inner_cg_step_ Execute CG using Hessian-vector products approximated by finite difference directional derivatives. On input these must be defined: atom->x - positions at x atom->f - ignored VEC_XK - positions at x On output these are defined: atom->x - unchanged atom->f - forces evaluated at x, but only if nStepType == NO_CGSTEP VEC_XK - unchanged VEC_CG_P - step from VEC_XK to new positions During processing these are modified: VEC_CG_D - conjugate gradient inner loop step VEC_CG_HD - Hessian-vector product VEC_CG_R - residual of inner loop step VEC_DIF1 - temp storage VEC_DIF2 - temp storage @param[in] dTrustRadius - trust region radius for this subiteration @param[in] dForceTol - stop tolerance on |F|_2 for this subiteration @param[in] nMaxEvals - total energy/force evaluations allowed @param[in] bHaveEvalAtXin - true if forces are valid at input x @param[in] dEnergyAtXin - energy at input x, if bHaveEvalAtXin is true @param[in] dForce2AtXin - |F|_2 at input x, if bHaveEvalAtXin is true @param[out] dEnergyAtXout - energy at output x, if NO_CGSTEP (see below) @param[out] dForce2AtXout - |F|_2 at output x, if NO_CGSTEP (see below) @param[out] nStepType - step type for hftn_print_line_() @param[out] dStepLength2 - |step|_2 @param[out] dStepLengthInf - |step|_inf Return false if there was a fatal error. If nStepType equals NO_CGSTEP_BECAUSE_F_TOL_SATISFIED, then the energy and forces are evaluated and returned in dEnergyAtXout, dForce2AtXout; else energy and forces are not evaluated. ------------------------------------------------------------------------- */ bool MinHFTN::compute_inner_cg_step_(const double dTrustRadius, const double dForceTol, const int nMaxEvals, const bool bHaveEvalAtXin, const double dEnergyAtXin, const double dForce2AtXin, double & dEnergyAtXout, double & dForce2AtXout, int & nStepType, double & dStepLength2, double & dStepLengthInf) { //---- SET p_0 = 0. if (nextra_global) { for (int i = 0; i < nextra_global; i++) _daExtraGlobal[VEC_CG_P][i] = 0.0; } for (int i = 0; i < nvec; i++) _daAVectors[VEC_CG_P][i] = 0.0; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) pAtom[i] = 0.0; } } double dPP = 0.0; //---- OBTAIN THE ENERGY AND FORCES AT THE INPUT POSITION. double dEnergyAtX = dEnergyAtXin; double dForce2AtX = dForce2AtXin; if (bHaveEvalAtXin == false) { dEnergyAtX = energy_force (0); neval++; dForce2AtX = sqrt (fnorm_sqr()); } //---- RETURN IMMEDIATELY IF THE FORCE TOLERANCE IS ALREADY MET. //---- THE STEP TYPE INFORMS THE CALLER THAT ENERGY AND FORCES HAVE //---- BEEN EVALUATED. if (dForce2AtX <= dForceTol) { dEnergyAtXout = dEnergyAtX; dForce2AtXout = dForce2AtX; nStepType = NO_CGSTEP_BECAUSE_F_TOL_SATISFIED; dStepLength2 = 0.0; dStepLengthInf = 0.0; return( true ); } //---- r_0 = -grad (FIRST SEARCH DIRECTION IS STEEPEST DESCENT) //---- d_0 = r_0 //---- REMEMBER THAT FORCES = -GRADIENT. if (nextra_global) { for (int i = 0; i < nextra_global; i++) { _daExtraGlobal[VEC_CG_R][i] = fextra[i]; _daExtraGlobal[VEC_CG_D][i] = fextra[i]; } } for (int i = 0; i < nvec; i++) { _daAVectors[VEC_CG_R][i] = fvec[i]; _daAVectors[VEC_CG_D][i] = fvec[i]; } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * fatom = fextra_atom[m]; double * rAtom = _daExtraAtom[VEC_CG_R][m]; double * dAtom = _daExtraAtom[VEC_CG_D][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) { rAtom[i] = fatom[i]; dAtom[i] = fatom[i]; } } } double dRR = dForce2AtX * dForce2AtX; double dR0norm2 = sqrt (dRR); //---- LIMIT THE NUMBER OF INNER CG ITERATIONS. //---- BASE IT ON THE NUMBER OF UNKNOWNS, OR MAXIMUM EVALUATIONS ASSUMING //---- FORWARD DIFFERENCES ARE USED. //---- NOTE THAT SETTING MAX=1 GIVES STEEPEST DESCENT. int nLimit1 = _nNumUnknowns / 5; if (nLimit1 < 100) nLimit1 = MIN (_nNumUnknowns, 100); int nLimit2 = (nMaxEvals - neval) / 2; int nMaxInnerIters = MIN (nLimit1, nLimit2); //---- FURTHER LIMIT ITERATIONS IF NEAR MACHINE ROUNDOFF. //---- THE METHOD CAN WASTE A LOT EVALUATIONS WITH LITTLE PAYOFF PROSPECT. if (dForce2AtX < (sqrt (MACHINE_EPS) * MAX (1.0, fabs (dEnergyAtX))) ) nMaxInnerIters = MIN (nMaxInnerIters, _nNumUnknowns / 20); bool bUseForwardDiffs = (dForce2AtX > 1000.0 * sqrt (MACHINE_EPS)); //---- MAIN CG LOOP. for (int nInnerIter = 0; nInnerIter < nMaxInnerIters; nInnerIter++) { //---- COMPUTE HESSIAN-VECTOR PRODUCT: H(x_k) d_i. double dDummyEnergy; evaluate_dir_der_ (bUseForwardDiffs, VEC_CG_D, VEC_CG_HD, false, dDummyEnergy); //---- CALCULATE d_i^T H d_i AND d_i^T d_i. double dDHD; double dDD; calc_dhd_dd_using_mpi_ (dDHD, dDD); //---- HANDLE NEGATIVE CURVATURE. if (dDHD <= (MACHINE_EPS * dDD)) { //---- PROJECT BOTH DIRECTIONS TO THE TRUST RADIUS AND DECIDE //---- WHICH MAKES A BETTER PREDICTED REDUCTION. //---- p_i^T H(x_k) d_i AND grad_i^T d_i. double dPdotD = calc_dot_prod_using_mpi_ (VEC_CG_P, VEC_CG_D); double dPdotHD = calc_dot_prod_using_mpi_ (VEC_CG_P, VEC_CG_HD); //---- MOVE TO X_K AND COMPUTE ENERGY AND FORCES. if (nextra_global) modify->min_step (0.0, _daExtraGlobal[VEC_CG_P]); for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_XK][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * xkAtom = _daExtraAtom[VEC_XK][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = xkAtom[i]; requestor[m]->min_x_set(m); } } dEnergyAtX = energy_force (0); neval++; double dGradDotD = calc_grad_dot_v_using_mpi_ (VEC_CG_D); double tau = compute_to_tr_ (dPP, dPdotD, dDD, dTrustRadius, true, dDHD, dPdotHD, dGradDotD); //---- MOVE THE POINT. if (nextra_global) { double * pGlobal = _daExtraGlobal[VEC_CG_P]; double * dGlobal = _daExtraGlobal[VEC_CG_D]; for (int i = 0; i < nextra_global; i++) { pGlobal[i] += tau * dGlobal[i]; } } for (int i = 0; i < nvec; i++) _daAVectors[VEC_CG_P][i] += tau * _daAVectors[VEC_CG_D][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; double * dAtom = _daExtraAtom[VEC_CG_D][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) pAtom[i] += tau * dAtom[i]; } } nStepType = CGSTEP_NEGATIVE_CURVATURE; calc_plengths_using_mpi_ (dStepLength2, dStepLengthInf); return( true ); } //---- COMPUTE THE OPTIMAL STEP LENGTH BASED ON THE QUADRATIC CG MODEL. double dAlpha = dRR / dDHD; //---- MIGHT WANT TO ENABLE THIS TO DEBUG INTERNAL CG STEPS. //fprintf (_fpPrint, " alpha = %11.8f neval=%4d\n", dAlpha, neval); //---- p_i+1 = p_i + alpha_i d_i //---- (SAVE THE CURRENT p_i IN CASE THE STEP HAS TO BE SHORTENED.) if (nextra_global) { double * pGlobal = _daExtraGlobal[VEC_CG_P]; double * dGlobal = _daExtraGlobal[VEC_CG_D]; double * d1Global = _daExtraGlobal[VEC_DIF1]; for (int i = 0; i < nextra_global; i++) { d1Global[i] = pGlobal[i]; pGlobal[i] += dAlpha * dGlobal[i]; } } for (int i = 0; i < nvec; i++) { _daAVectors[VEC_DIF1][i] = _daAVectors[VEC_CG_P][i]; _daAVectors[VEC_CG_P][i] += dAlpha * _daAVectors[VEC_CG_D][i]; } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; double * dAtom = _daExtraAtom[VEC_CG_D][m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) { d1Atom[i] = pAtom[i]; pAtom[i] += dAlpha * dAtom[i]; } } } //---- COMPUTE VECTOR PRODUCTS p_i+1^T p_i+1 AND p_i^T d_i. double dPnewDotPnew; double dPoldDotD; calc_ppnew_pdold_using_mpi_ (dPnewDotPnew, dPoldDotD); nStepType = CGSTEP_UNDETERMINED; //---- IF STEP LENGTH IS TOO LARGE, THEN REDUCE IT AND RETURN. double tau; if (step_exceeds_TR_ (dTrustRadius, dPP, dPoldDotD, dDD, tau)) { adjust_step_to_tau_ (tau); nStepType = CGSTEP_TO_TR; } if (step_exceeds_DMAX_()) { adjust_step_to_tau_ (0.0); nStepType = CGSTEP_TO_DMAX; } if ((nStepType == CGSTEP_TO_TR) || (nStepType == CGSTEP_TO_DMAX)) { calc_plengths_using_mpi_ (dStepLength2, dStepLengthInf); return( true ); } dStepLength2 = sqrt (dPnewDotPnew); //---- r_i+1 = r_i - alpha * H d_i if (nextra_global) { double * rGlobal = _daExtraGlobal[VEC_CG_R]; double * hdGlobal = _daExtraGlobal[VEC_CG_HD]; for (int i = 0; i < nextra_global; i++) rGlobal[i] -= dAlpha * hdGlobal[i]; } for (int i = 0; i < nvec; i++) _daAVectors[VEC_CG_R][i] -= dAlpha * _daAVectors[VEC_CG_HD][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * rAtom = _daExtraAtom[VEC_CG_R][m]; double * hdAtom = _daExtraAtom[VEC_CG_HD][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) rAtom[i] -= dAlpha * hdAtom[i]; } } double dRnewDotRnew = calc_dot_prod_using_mpi_ (VEC_CG_R, VEC_CG_R); //---- IF RESIDUAL IS SMALL ENOUGH, THEN RETURN THE CURRENT STEP. if (sqrt (dRnewDotRnew) < dForceTol * dR0norm2) { nStepType = CGSTEP_NEWTON; calc_plengths_using_mpi_ (dStepLength2, dStepLengthInf); return( true ); } //---- beta = r_i+1^T r_i+1 / r_i^T r_i //---- d_i+1 = r_i+1 + beta d_i double dBeta = dRnewDotRnew / dRR; if (nextra_global) { double * rGlobal = _daExtraGlobal[VEC_CG_R]; double * dGlobal = _daExtraGlobal[VEC_CG_D]; for (int i = 0; i < nextra_global; i++) dGlobal[i] = rGlobal[i] + dBeta * dGlobal[i]; } for (int i = 0; i < nvec; i++) _daAVectors[VEC_CG_D][i] = _daAVectors[VEC_CG_R][i] + dBeta * _daAVectors[VEC_CG_D][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * rAtom = _daExtraAtom[VEC_CG_R][m]; double * dAtom = _daExtraAtom[VEC_CG_D][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) dAtom[i] = rAtom[i] + dBeta * dAtom[i]; } } //---- CONTINUE THE LOOP. dRR = dRnewDotRnew; dPP = dPnewDotPnew; } nStepType = CGSTEP_MAX_INNER_ITERS; calc_plengths_using_mpi_ (dStepLength2, dStepLengthInf); return( true ); } /* ---------------------------------------------------------------------- Private method calc_xinf_using_mpi_ ------------------------------------------------------------------------- */ double MinHFTN::calc_xinf_using_mpi_(void) const { double dXInfLocal = 0.0; for (int i = 0; i < nvec; i++) dXInfLocal = MAX(dXInfLocal,fabs(xvec[i])); double dXInf; MPI_Allreduce (&dXInfLocal, &dXInf, 1, MPI_DOUBLE, MPI_MAX, world); if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; int n = extra_nlen[m]; double dXInfLocalExtra = 0.0; for (int i = 0; i < n; i++) dXInfLocalExtra = MAX (dXInfLocalExtra, fabs (xatom[i])); double dXInfExtra; MPI_Allreduce (&dXInfLocalExtra, &dXInfExtra, 1, MPI_DOUBLE, MPI_MAX, world); dXInf = MAX (dXInf, dXInfExtra); } } return( dXInf ); } /* ---------------------------------------------------------------------- Private method calc_dot_prod_using_mpi_ ------------------------------------------------------------------------- */ double MinHFTN::calc_dot_prod_using_mpi_(const int nIx1, const int nIx2) const { double dDotLocal = 0.0; for (int i = 0; i < nvec; i++) dDotLocal += _daAVectors[nIx1][i] * _daAVectors[nIx2][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * i1Atom = _daExtraAtom[nIx1][m]; double * i2Atom = _daExtraAtom[nIx2][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) dDotLocal += i1Atom[i] * i2Atom[i]; } } double dDot; MPI_Allreduce (&dDotLocal, &dDot, 1, MPI_DOUBLE, MPI_SUM, world); if (nextra_global) { for (int i = 0; i < nextra_global; i++) { double * i1Global = _daExtraGlobal[nIx1]; double * i2Global = _daExtraGlobal[nIx2]; dDot += i1Global[i] * i2Global[i]; } } return( dDot ); } /* ---------------------------------------------------------------------- Private method calc_grad_dot_v_using_mpi_ ------------------------------------------------------------------------- */ double MinHFTN::calc_grad_dot_v_using_mpi_(const int nIx) const { //---- ASSUME THAT FORCES HAVE BEEN EVALUATED AT DESIRED ATOM POSITIONS. //---- REMEMBER THAT FORCES = -GRADIENT. double dGradDotVLocal = 0.0; for (int i = 0; i < nvec; i++) dGradDotVLocal += - _daAVectors[nIx][i] * fvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * fatom = fextra_atom[m]; double * iAtom = _daExtraAtom[nIx][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) dGradDotVLocal += - iAtom[i] * fatom[i]; } } double dGradDotV; MPI_Allreduce (&dGradDotVLocal, &dGradDotV, 1, MPI_DOUBLE, MPI_SUM, world); if (nextra_global) { for (int i = 0; i < nextra_global; i++) { double * iGlobal = _daExtraGlobal[nIx]; dGradDotV += - iGlobal[i] * fextra[i]; } } return( dGradDotV ); } /* ---------------------------------------------------------------------- Private method calc_dhd_dd_using_mpi_ ------------------------------------------------------------------------- */ void MinHFTN::calc_dhd_dd_using_mpi_(double & dDHD, double & dDD) const { double dDHDLocal = 0.0; double dDDLocal = 0.0; for (int i = 0; i < nvec; i++) { dDHDLocal += _daAVectors[VEC_CG_D][i] * _daAVectors[VEC_CG_HD][i]; dDDLocal += _daAVectors[VEC_CG_D][i] * _daAVectors[VEC_CG_D][i]; } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * dAtom = _daExtraAtom[VEC_CG_D][m]; double * hdAtom = _daExtraAtom[VEC_CG_HD][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) { dDHDLocal += dAtom[i] * hdAtom[i]; dDDLocal += dAtom[i] * dAtom[i]; } } } double daDotsLocal[2]; daDotsLocal[0] = dDHDLocal; daDotsLocal[1] = dDDLocal; double daDots[2]; MPI_Allreduce (daDotsLocal, daDots, 2, MPI_DOUBLE, MPI_SUM, world); if (nextra_global) { double * dGlobal = _daExtraGlobal[VEC_CG_D]; double * hdGlobal = _daExtraGlobal[VEC_CG_HD]; for (int i = 0; i < nextra_global; i++) { daDots[0] += dGlobal[i] * hdGlobal[i]; daDots[1] += dGlobal[i] * dGlobal[i]; } } dDHD = daDots[0]; dDD = daDots[1]; return; } /* ---------------------------------------------------------------------- Private method calc_ppnew_pdold_using_mpi_ ------------------------------------------------------------------------- */ void MinHFTN::calc_ppnew_pdold_using_mpi_(double & dPnewDotPnew, double & dPoldDotD) const { double dPnewDotPnewLocal = 0.0; double dPoldDotDLocal = 0.0; for (int i = 0; i < nvec; i++) { dPnewDotPnewLocal += _daAVectors[VEC_CG_P][i] * _daAVectors[VEC_CG_P][i]; dPoldDotDLocal += _daAVectors[VEC_DIF1][i] * _daAVectors[VEC_CG_D][i]; } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * dAtom = _daExtraAtom[VEC_CG_D][m]; double * pAtom = _daExtraAtom[VEC_CG_P][m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) { dPnewDotPnewLocal += pAtom[i] * pAtom[i]; dPoldDotDLocal += d1Atom[i] * dAtom[i]; } } } double daDotsLocal[2]; daDotsLocal[0] = dPnewDotPnewLocal; daDotsLocal[1] = dPoldDotDLocal; double daDots[2]; MPI_Allreduce (daDotsLocal, daDots, 2, MPI_DOUBLE, MPI_SUM, world); if (nextra_global) { for (int i = 0; i < nextra_global; i++) { double * dGlobal = _daExtraGlobal[VEC_CG_D]; double * pGlobal = _daExtraGlobal[VEC_CG_P]; double * d1Global = _daExtraGlobal[VEC_DIF1]; daDots[0] += pGlobal[i] * pGlobal[i]; daDots[1] += d1Global[i] * dGlobal[i]; } } dPnewDotPnew = daDots[0]; dPoldDotD = daDots[1]; return; } /* ---------------------------------------------------------------------- Private method calc_plengths_using_mpi_ ------------------------------------------------------------------------- */ void MinHFTN::calc_plengths_using_mpi_(double & dStepLength2, double & dStepLengthInf) const { double dPPLocal = 0.0; double dPInfLocal = 0.0; for (int i = 0; i < nvec; i++) { dPPLocal += _daAVectors[VEC_CG_P][i] * _daAVectors[VEC_CG_P][i]; dPInfLocal = MAX (dPInfLocal, fabs (_daAVectors[VEC_CG_P][i])); } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) { dPPLocal += pAtom[i] * pAtom[i]; dPInfLocal = MAX (dPInfLocal, fabs (pAtom[i])); } } } double dPP; MPI_Allreduce (&dPPLocal, &dPP, 1, MPI_DOUBLE, MPI_SUM, world); double dPInf; MPI_Allreduce (&dPInfLocal, &dPInf, 1, MPI_DOUBLE, MPI_MAX, world); if (nextra_global) { for (int i = 0; i < nextra_global; i++) { double * pGlobal = _daExtraGlobal[VEC_CG_P]; dPP += pGlobal[i] * pGlobal[i]; dPInf = MAX (dPInf, fabs (pGlobal[i])); } } dStepLength2 = sqrt (dPP); dStepLengthInf = dPInf; return; } /* ---------------------------------------------------------------------- Private method step_exceeds_TR_ ------------------------------------------------------------------------- */ bool MinHFTN::step_exceeds_TR_(const double dTrustRadius, const double dPP, const double dPD, const double dDD, double & dTau) const { double dPnewNorm2; double dPnewNormInf; calc_plengths_using_mpi_ (dPnewNorm2, dPnewNormInf); if (dPnewNorm2 > dTrustRadius) { dTau = compute_to_tr_ (dPP, dPD, dDD, dTrustRadius, false, 0.0, 0.0, 0.0); return( true ); } //---- STEP LENGTH IS NOT TOO LONG. dTau = 0.0; return( false ); } /* ---------------------------------------------------------------------- Private method step_exceeds_DMAX_ Check that atoms do not move too far: for atom coordinates: limit is dmax for extra per-atom DOF: limit is extra_max[] for extra global DOF: limit is set by modify->max_alpha, which calls fix_box_relax->max_alpha ------------------------------------------------------------------------- */ bool MinHFTN::step_exceeds_DMAX_(void) const { double dAlpha = dmax * sqrt((double) _nNumUnknowns); double dPInfLocal = 0.0; for (int i = 0; i < nvec; i++) dPInfLocal = MAX (dPInfLocal, fabs (_daAVectors[VEC_CG_P][i])); double dPInf; MPI_Allreduce (&dPInfLocal, &dPInf, 1, MPI_DOUBLE, MPI_MAX, world); if (dPInf > dmax) return( true ); if (dPInf > MACHINE_EPS) dAlpha = MIN (dAlpha, dmax / dPInf); if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; dPInfLocal = 0.0; int n = extra_nlen[m]; for (int i = 0; i < n; i++) dPInfLocal = MAX (dPInfLocal, fabs (pAtom[i])); MPI_Allreduce (&dPInfLocal, &dPInf, 1, MPI_DOUBLE, MPI_MAX, world); if (dPInf > extra_max[m]) return( true ); if (dPInf > MACHINE_EPS) dAlpha = MIN (dAlpha, extra_max[m] / dPInf); } } if (nextra_global) { //---- IF THE MAXIMUM DISTANCE THAT THE GLOBAL BOX CONSTRAINT WILL //---- ALLOW IS SMALLER THAN THE PROPOSED DISTANCE, THEN THE STEP //---- IS TOO LONG. PROPOSED DISTANCE IS ESTIMATED BY |P|_INF. double dAlphaExtra = modify->max_alpha (_daExtraGlobal[VEC_CG_P]); if (dAlphaExtra < dAlpha) return( true ); } //---- STEP LENGTH IS NOT TOO LONG. return( false ); } /* ---------------------------------------------------------------------- Private method adjust_step_to_tau_ Adjust the step so that VEC_CG_P = VEC_DIF1 + tau * VEC_CG_D. ------------------------------------------------------------------------- */ void MinHFTN::adjust_step_to_tau_(const double tau) { if (nextra_global) { double * pGlobal = _daExtraGlobal[VEC_CG_P]; double * dGlobal = _daExtraGlobal[VEC_CG_D]; double * d1Global = _daExtraGlobal[VEC_DIF1]; for (int i = 0; i < nextra_global; i++) pGlobal[i] = d1Global[i] + (tau * dGlobal[i]); } for (int i = 0; i < nvec; i++) { _daAVectors[VEC_CG_P][i] = _daAVectors[VEC_DIF1][i] + (tau * _daAVectors[VEC_CG_D][i]); } if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * pAtom = _daExtraAtom[VEC_CG_P][m]; double * dAtom = _daExtraAtom[VEC_CG_D][m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) pAtom[i] = d1Atom[i] + (tau * dAtom[i]); } } return; } /* ---------------------------------------------------------------------- Private method compute_to_tr_ Return the value tau that solves || p + tau*d ||_2 = dTrustRadius If both roots are considered, the TR method chooses the one that minimizes grad^T (p + tau*d) + 0.5 (p + tau*d)^T H (p + tau*d) @param[in] dPP - p^T p @param[in] dPD - p^T d @param[in] dDD - d^T d @param[in] dTrustRadius - distance to match @param[in] bConsiderBothRoots - evaluate both roots, or return the positive @param[in] dDHD - d^T H d @param[in] dPdotHD - p^T H d @param[in] dGradDotD - grad(x_k)^T d ------------------------------------------------------------------------- */ double MinHFTN::compute_to_tr_(const double dPP, const double dPD, const double dDD, const double dTrustRadius, const bool bConsiderBothRoots, const double dDHD, const double dPdotHD, const double dGradDotD) const { //---- SOLVE A QUADRATIC EQUATION FOR TAU. //---- THE COEFFICIENTS ARE SUCH THAT THERE ARE ALWAYS TWO REAL ROOTS, //---- ONE POSITIVE AND ONE NEGATIVE. //---- CHECK FOR ERRONEOUS DATA. if ( (dDD <= 0.0) || (dPP < 0.0) || (dTrustRadius < 0.0) || (dTrustRadius * dTrustRadius < dPP) ) { printf ("HFTN internal error - bad data given to compute_to_tr_()\n"); return( 0.0 ); } double dTRsqrd = dTrustRadius * dTrustRadius; double dDiscr = (dPD * dPD) - (dDD * (dPP - dTRsqrd)); dDiscr = MAX (0.0, dDiscr); //-- SHOULD NEVER BE NEGATIVE dDiscr = sqrt (dDiscr); double dRootPos = (-dPD + dDiscr) / dDD; double dRootNeg = (-dPD - dDiscr) / dDD; if (bConsiderBothRoots == false) return( dRootPos ); //---- EVALUATE THE CG OBJECTIVE FUNCTION FOR EACH ROOT. double dTmpTerm = dGradDotD + dPdotHD; double dCgRedPos = (dRootPos * dTmpTerm) + (0.5 * dRootPos*dRootPos * dDHD); double dCgRedNeg = (dRootNeg * dTmpTerm) + (0.5 * dRootNeg*dRootNeg * dDHD); if ((-dCgRedPos) > (-dCgRedNeg)) return( dRootPos ); else return( dRootNeg ); } /* ---------------------------------------------------------------------- Private method evaluate_dir_der_ Compute the directional derivative using a finite difference approximation. This is equivalent to the Hessian times direction vector p. As a side effect, the method computes the energy and forces at x. On input these must be defined: atom->x - positions at x atom->f - ignored nIxDir - VEC_ index of the direction p nIxResult - ignored On output these are defined: atom->x - unchanged atom->f - forces evaluated at x, only if bEvaluateAtX is true nIxDir - unchanged nIxResult - directional derivative Hp During processing these are modified: VEC_DIF1 VEC_DIF2 @param[in] bUseForwardDiffs - if true use forward difference approximation, else use central difference @param[in] nIxDir - VEC_ index of the direction @param[in] nIxResult - VEC_ index to place the result (it is acceptable for nIxDir = nIxResult) @param[in] bEvaluateAtX - if true, then evaluate at x before returning @param[out] dNewEnergy - energy at x, if bEvaluateAtX is true @param[out] dNewForce2 - |F|_2 at x, if bEvaluateAtX is true ------------------------------------------------------------------------- */ void MinHFTN::evaluate_dir_der_(const bool bUseForwardDiffs, const int nIxDir, const int nIxResult, const bool bEvaluateAtX, double & dNewEnergy) { //---- COMPUTE THE MAGNITUDE OF THE DIRECTION VECTOR: |p|_2. double dDirNorm2SqrdLocal = 0.0; for (int i = 0; i < nvec; i++) dDirNorm2SqrdLocal += _daAVectors[nIxDir][i] * _daAVectors[nIxDir][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * iAtom = _daExtraAtom[nIxDir][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) dDirNorm2SqrdLocal += iAtom[i] * iAtom[i]; } } double dDirNorm2Sqrd = 0.0; MPI_Allreduce (&dDirNorm2SqrdLocal, &dDirNorm2Sqrd, 1, MPI_DOUBLE, MPI_SUM, world); if (nextra_global) { for (int i = 0; i < nextra_global; i++) { double * iGlobal = _daExtraGlobal[nIxDir]; dDirNorm2Sqrd += iGlobal[i] * iGlobal[i]; } } double dDirNorm2 = sqrt (dDirNorm2Sqrd); //---- IF THE STEP IS TOO SMALL, RETURN ZERO FOR THE DERIVATIVE. if (dDirNorm2 == 0.0) { for (int i = 0; i < nvec; i++) _daAVectors[nIxResult][i] = 0.0; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * iAtom = _daExtraAtom[nIxDir][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) iAtom[i] = 0; } } if (nextra_global) { for (int i = 0; i < nextra_global; i++) _daExtraGlobal[nIxDir][i] = 0.0; } if (bEvaluateAtX) { dNewEnergy = energy_force (0); neval++; } return; } //---- FORWARD DIFFERENCES ARE LESS ACCURATE THAN CENTRAL DIFFERENCES, //---- BUT REQUIRE ONLY 2 ENERGY+FORCE EVALUATIONS VERSUS 3 EVALUATIONS. //---- STORAGE REQUIREMENTS ARE THE SAME. if (bUseForwardDiffs) { //---- EQUATION IS FROM THE OLD LAMMPS VERSION, SAND98-8201. double dEps = 2.0 * sqrt (1000.0 * MACHINE_EPS) / dDirNorm2; //---- SAVE A COPY OF x. fix_minimize->store_box(); for (int i = 0; i < nvec; i++) _daAVectors[VEC_DIF1][i] = xvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) d1Atom[i] = xatom[i]; } } if (nextra_global) { modify->min_pushstore(); modify->min_store(); } //---- EVALUATE FORCES AT x + eps*p. if (nextra_global) modify->min_step (dEps, _daExtraGlobal[nIxDir]); for (int i = 0; i < nvec; i++) xvec[i] += dEps * _daAVectors[nIxDir][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * iAtom = _daExtraAtom[nIxDir][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] += dEps * iAtom[i]; requestor[m]->min_x_set(m); } } energy_force (0); neval++; //---- STORE THE FORCE IN DIF2. if (nextra_global) { for (int i = 0; i < nextra_global; i++) _daExtraGlobal[VEC_DIF2][i] = fextra[i]; } for (int i = 0; i < nvec; i++) _daAVectors[VEC_DIF2][i] = fvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * fatom = fextra_atom[m]; double * d2Atom = _daExtraAtom[VEC_DIF2][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) d2Atom[i] = fatom[i]; } } //---- MOVE BACK TO x AND EVALUATE FORCES. if (nextra_global) { modify->min_step (0.0, _daExtraGlobal[VEC_DIF1]); modify->min_popstore(); } for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_DIF1][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] += d1Atom[i]; requestor[m]->min_x_set(m); } } dNewEnergy = energy_force (0); neval++; //---- COMPUTE THE DIFFERENCE VECTOR: [grad(x + eps*p) - grad(x)] / eps. //---- REMEMBER THAT FORCES = -GRADIENT. for (int i = 0; i < nvec; i++) _daAVectors[nIxResult][i] = (fvec[i] - _daAVectors[VEC_DIF2][i]) / dEps; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * iAtom = _daExtraAtom[nIxResult][m]; double * d2Atom = _daExtraAtom[VEC_DIF2][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) iAtom[i] = (fextra_atom[m][i] - d2Atom[i]) / dEps; } } if (nextra_global) { for (int i = 0; i < nextra_global; i++) _daExtraGlobal[nIxResult][i] = (fextra[i] - _daExtraGlobal[VEC_DIF2][i]) / dEps; } } else { //-- bUseForwardDiffs == false //---- EQUATION IS FROM THE OLD LAMMPS VERSION, SAND98-8201. double dEps = pow (3000.0 * MACHINE_EPS, 0.33333333) / dDirNorm2; //---- SAVE A COPY OF x. fix_minimize->store_box(); for (int i = 0; i < nvec; i++) _daAVectors[VEC_DIF1][i] = xvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) d1Atom[i] = xatom[i]; } } if (nextra_global) { modify->min_pushstore(); modify->min_store(); } //---- EVALUATE FORCES AT x + eps*p. if (nextra_global) modify->min_step (dEps, _daExtraGlobal[nIxDir]); for (int i = 0; i < nvec; i++) xvec[i] += dEps * _daAVectors[nIxDir][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * iAtom = _daExtraAtom[nIxDir][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] += dEps * iAtom[i]; requestor[m]->min_x_set(m); } } energy_force (0); neval++; //---- STORE THE FORCE IN DIF2. if (nextra_global) { for (int i = 0; i < nextra_global; i++) _daExtraGlobal[VEC_DIF2][i] = fextra[i]; } for (int i = 0; i < nvec; i++) _daAVectors[VEC_DIF2][i] = fvec[i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * fatom = fextra_atom[m]; double * d2Atom = _daExtraAtom[VEC_DIF2][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) d2Atom[i] = fatom[i]; } } //---- EVALUATE FORCES AT x - eps*p. if (nextra_global) modify->min_step (-dEps, _daExtraGlobal[nIxDir]); for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_DIF1][i] - dEps * _daAVectors[nIxDir][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * iAtom = _daExtraAtom[nIxDir][m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = d1Atom[i] - dEps * iAtom[i]; requestor[m]->min_x_set(m); } } energy_force (0); neval++; //---- COMPUTE THE DIFFERENCE VECTOR: //---- [grad(x + eps*p) - grad(x - eps*p)] / 2*eps. //---- REMEMBER THAT FORCES = -GRADIENT. if (nextra_global) { double * iGlobal = _daExtraGlobal[nIxResult]; double * d2Global = _daExtraGlobal[VEC_DIF2]; for (int i = 0; i < nextra_global; i++) iGlobal[i] = (fextra[i] - d2Global[i]) / (2.0 + dEps); } for (int i = 0; i < nvec; i++) _daAVectors[nIxResult][i] = (fvec[i] - _daAVectors[VEC_DIF2][i]) / (2.0 * dEps); if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * fatom = fextra_atom[m]; double * iAtom = _daExtraAtom[nIxResult][m]; double * d2Atom = _daExtraAtom[VEC_DIF2][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) iAtom[i] = (fatom[i] - d2Atom[i]) / (2.0 + dEps); } } if (bEvaluateAtX) { //---- EVALUATE FORCES AT x. if (nextra_global) { modify->min_step (0.0, _daExtraGlobal[VEC_DIF1]); modify->min_popstore(); } for (int i = 0; i < nvec; i++) xvec[i] = _daAVectors[VEC_DIF1][i]; if (nextra_atom) { for (int m = 0; m < nextra_atom; m++) { double * xatom = xextra_atom[m]; double * d1Atom = _daExtraAtom[VEC_DIF1][m]; int n = extra_nlen[m]; for (int i = 0; i < n; i++) xatom[i] = d1Atom[i]; requestor[m]->min_x_set(m); } } dNewEnergy = energy_force (0); neval++; } } return; } /* ---------------------------------------------------------------------- Private method open_hftn_print_file_ ------------------------------------------------------------------------- */ void MinHFTN::open_hftn_print_file_(void) { int nMyRank; MPI_Comm_rank (world, &nMyRank); char szTmp[50]; sprintf (szTmp, "progress_MinHFTN_%d.txt", nMyRank); _fpPrint = fopen (szTmp, "w"); if (_fpPrint == NULL) { printf ("*** MinHFTN cannot open file '%s'\n", szTmp); printf ("*** continuing...\n"); return; } fprintf (_fpPrint, " Iter Evals Energy |F|_2" " Step TR used |step|_2 ared pred\n"); return; } /* ---------------------------------------------------------------------- Private method hftn_print_line_ Step types: 1 - Nw (inner iteration converged like a Newton step) 2 - TR (inner iteration reached the trust region boundary) 3 - Neg (inner iteration ended with negative curvature) ------------------------------------------------------------------------- */ void MinHFTN::hftn_print_line_(const bool bIsStepAccepted, const int nIteration, const int nTotalEvals, const double dEnergy, const double dForce2, const int nStepType, const double dTrustRadius, const double dStepLength2, const double dActualRed, const double dPredictedRed) const { const char sFormat1[] = " %4d %5d %14.8f %11.5e\n"; const char sFormatA[] = " %4d %5d %14.8f %11.5e %3s %9.3e %8.2e %10.3e %10.3e\n"; const char sFormatR[] = "r %4d %5d %14.8f %11.5e %3s %9.3e %8.2e %10.3e %10.3e\n"; if (_fpPrint == NULL) return; char sStepType[4]; if (nStepType == NO_CGSTEP_BECAUSE_F_TOL_SATISFIED) strcpy (sStepType, " - "); else if (nStepType == CGSTEP_NEWTON) strcpy (sStepType, "Nw "); else if (nStepType == CGSTEP_TO_TR) strcpy (sStepType, "TR "); else if (nStepType == CGSTEP_TO_DMAX) strcpy (sStepType, "dmx"); else if (nStepType == CGSTEP_NEGATIVE_CURVATURE) strcpy (sStepType, "Neg"); else if (nStepType == CGSTEP_MAX_INNER_ITERS) strcpy (sStepType, "its"); else strcpy (sStepType, "???"); if (nIteration == -1) { fprintf (_fpPrint, sFormat1, 0, nTotalEvals, dEnergy, dForce2); } else { if (bIsStepAccepted) fprintf (_fpPrint, sFormatA, nIteration, nTotalEvals, dEnergy, dForce2, sStepType, dTrustRadius, dStepLength2, dActualRed, dPredictedRed); else fprintf (_fpPrint, sFormatR, nIteration, nTotalEvals, dEnergy, dForce2, sStepType, dTrustRadius, dStepLength2, dActualRed, dPredictedRed); } fflush (_fpPrint); return; } /* ---------------------------------------------------------------------- Private method close_hftn_print_file_ ------------------------------------------------------------------------- */ void MinHFTN::close_hftn_print_file_(void) { if (_fpPrint != NULL) fclose (_fpPrint); return; } diff --git a/src/min_hftn.h b/src/min_hftn.h index 6630b0d89..1387581cc 100644 --- a/src/min_hftn.h +++ b/src/min_hftn.h @@ -1,126 +1,125 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifdef MINIMIZE_CLASS MinimizeStyle(hftn,MinHFTN) #else #ifndef LMP_MIN_HFTN_H #define LMP_MIN_HFTN_H #include "min.h" namespace LAMMPS_NS { //---- THE ALGORITHM NEEDS TO STORE THIS MANY ATOM-BASED VECTORS, //---- IN ADDITION TO ATOM POSITIONS AND THE FORCE VECTOR. static const int NUM_HFTN_ATOM_BASED_VECTORS = 7; static const int VEC_XK = 0; //-- ATOM POSITIONS AT SUBITER START static const int VEC_CG_P = 1; //-- STEP p IN CG SUBITER static const int VEC_CG_D = 2; //-- DIRECTION d IN CG SUBITER static const int VEC_CG_HD = 3; //-- HESSIAN-VECTOR PRODUCT Hd static const int VEC_CG_R = 4; //-- RESIDUAL r IN CG SUBITER static const int VEC_DIF1 = 5; //-- FOR FINITE DIFFERENCING static const int VEC_DIF2 = 6; //-- FOR FINITE DIFFERENCING class MinHFTN : public Min { public: MinHFTN (LAMMPS *); ~MinHFTN (void); - - void init_style(); + void init(); void setup_style(); void reset_vectors(); int iterate (int); private: //---- ATOM-BASED STORAGE VECTORS. double * _daAVectors[NUM_HFTN_ATOM_BASED_VECTORS]; double ** _daExtraAtom[NUM_HFTN_ATOM_BASED_VECTORS]; //---- GLOBAL DOF STORAGE. ELEMENT [0] IS X0 (XK), NOT USED IN THIS ARRAY. double * _daExtraGlobal[NUM_HFTN_ATOM_BASED_VECTORS]; int _nNumUnknowns; FILE * _fpPrint; int execute_hftn_ (const bool bPrintProgress, const double dInitialEnergy, const double dInitialForce2, double & dFinalEnergy, double & dFinalForce2); bool compute_inner_cg_step_ (const double dTrustRadius, const double dForceTol, const int nMaxEvals, const bool bHaveEvalAtXin, const double dEnergyAtXin, const double dForce2AtXin, double & dEnergyAtXout, double & dForce2AtXout, int & nStepType, double & dStepLength2, double & dStepLengthInf); double calc_xinf_using_mpi_ (void) const; double calc_dot_prod_using_mpi_ (const int nIx1, const int nIx2) const; double calc_grad_dot_v_using_mpi_ (const int nIx) const; void calc_dhd_dd_using_mpi_ (double & dDHD, double & dDD) const; void calc_ppnew_pdold_using_mpi_ (double & dPnewDotPnew, double & dPoldDotD) const; void calc_plengths_using_mpi_ (double & dStepLength2, double & dStepLengthInf) const; bool step_exceeds_TR_ (const double dTrustRadius, const double dPP, const double dPD, const double dDD, double & dTau) const; bool step_exceeds_DMAX_ (void) const; void adjust_step_to_tau_ (const double tau); double compute_to_tr_ (const double dPP, const double dPD, const double dDD, const double dTrustRadius, const bool bConsiderBothRoots, const double dDHD, const double dPdotHD, const double dGradDotD) const; void evaluate_dir_der_ (const bool bUseForwardDiffs, const int nIxDir, const int nIxResult, const bool bEvaluateAtX, double & dNewEnergy); void open_hftn_print_file_ (void); void hftn_print_line_ (const bool bIsStepAccepted, const int nIteration, const int nTotalEvals, const double dEnergy, const double dForce2, const int nStepType, const double dTrustRadius, const double dStepLength2, const double dActualRed, const double dPredictedRed) const; void close_hftn_print_file_ (void); }; } #endif #endif diff --git a/src/min_linesearch.cpp b/src/min_linesearch.cpp index 73d9ef83f..b50971a35 100644 --- a/src/min_linesearch.cpp +++ b/src/min_linesearch.cpp @@ -1,924 +1,926 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- Contributing author: Aidan Thompson (SNL) improved CG and backtrack ls, added quadratic ls Contributing author: Asad Hasan (CMU) added forcezero ls Sources: Numerical Recipes frprmn routine "Conjugate Gradient Method Without the Agonizing Pain" by JR Shewchuk, http://www-2.cs.cmu.edu/~jrs/jrspapers.html#cg ------------------------------------------------------------------------- */ #include "math.h" #include "min_linesearch.h" #include "atom.h" #include "update.h" #include "neighbor.h" #include "domain.h" #include "modify.h" #include "fix_minimize.h" #include "pair.h" #include "output.h" #include "thermo.h" #include "timer.h" #include "error.h" using namespace LAMMPS_NS; // ALPHA_MAX = max alpha allowed to avoid long backtracks // ALPHA_REDUCE = reduction ratio, should be in range [0.5,1) // BACKTRACK_SLOPE, should be in range (0,0.5] // QUADRATIC_TOL = tolerance on alpha0, should be in range [0.1,1) // EMACH = machine accuracy limit of energy changes (1.0e-8) // EPS_QUAD = tolerance for quadratic projection #define ALPHA_MAX 1.0 #define ALPHA_REDUCE 0.5 #define BACKTRACK_SLOPE 0.4 #define QUADRATIC_TOL 0.1 #define EMACH 1.0e-8 #define EPS_QUAD 1.0e-28 // same as in other min classes enum{MAXITER,MAXEVAL,ETOL,FTOL,DOWNHILL,ZEROALPHA,ZEROFORCE,ZEROQUAD}; /* ---------------------------------------------------------------------- */ MinLineSearch::MinLineSearch(LAMMPS *lmp) : Min(lmp) { searchflag = 1; gextra = hextra = NULL; x0extra_atom = gextra_atom = hextra_atom = NULL; } /* ---------------------------------------------------------------------- */ MinLineSearch::~MinLineSearch() { delete [] gextra; delete [] hextra; delete [] x0extra_atom; delete [] gextra_atom; delete [] hextra_atom; } /* ---------------------------------------------------------------------- */ -void MinLineSearch::init_style() +void MinLineSearch::init() { + Min::init(); + if (linestyle == 0) linemin = &MinLineSearch::linemin_backtrack; else if (linestyle == 1) linemin = &MinLineSearch::linemin_quadratic; else if (linestyle == 2) linemin = &MinLineSearch::linemin_forcezero; delete [] gextra; delete [] hextra; gextra = hextra = NULL; delete [] x0extra_atom; delete [] gextra_atom; delete [] hextra_atom; x0extra_atom = gextra_atom = hextra_atom = NULL; } /* ---------------------------------------------------------------------- */ void MinLineSearch::setup_style() { // memory for x0,g,h for atomic dof fix_minimize->add_vector(3); fix_minimize->add_vector(3); fix_minimize->add_vector(3); // memory for g,h for extra global dof, fix stores x0 if (nextra_global) { gextra = new double[nextra_global]; hextra = new double[nextra_global]; } // memory for x0,g,h for extra per-atom dof if (nextra_atom) { x0extra_atom = new double*[nextra_atom]; gextra_atom = new double*[nextra_atom]; hextra_atom = new double*[nextra_atom]; for (int m = 0; m < nextra_atom; m++) { fix_minimize->add_vector(extra_peratom[m]); fix_minimize->add_vector(extra_peratom[m]); fix_minimize->add_vector(extra_peratom[m]); } } } /* ---------------------------------------------------------------------- set current vector lengths and pointers called after atoms have migrated ------------------------------------------------------------------------- */ void MinLineSearch::reset_vectors() { // atomic dof nvec = 3 * atom->nlocal; if (nvec) xvec = atom->x[0]; if (nvec) fvec = atom->f[0]; x0 = fix_minimize->request_vector(0); g = fix_minimize->request_vector(1); h = fix_minimize->request_vector(2); // extra per-atom dof if (nextra_atom) { int n = 3; for (int m = 0; m < nextra_atom; m++) { extra_nlen[m] = extra_peratom[m] * atom->nlocal; requestor[m]->min_xf_pointers(m,&xextra_atom[m],&fextra_atom[m]); x0extra_atom[m] = fix_minimize->request_vector(n++); gextra_atom[m] = fix_minimize->request_vector(n++); hextra_atom[m] = fix_minimize->request_vector(n++); } } } /* ---------------------------------------------------------------------- line minimization methods find minimum-energy starting at x along h direction input args: eoriginal = energy at initial x input extra: n,x,x0,f,h for atomic, extra global, extra per-atom dof output args: return 0 if successful move, non-zero alpha return non-zero if failed alpha = distance moved along h for x at min eng config update neval counter of eng/force function evaluations output extra: if fail, energy_force() of original x if succeed, energy_force() at x + alpha*h atom->x = coords at new configuration atom->f = force at new configuration ecurrent = energy of new configuration ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- linemin: backtracking line search (Proc 3.1, p 41 in Nocedal and Wright) uses no gradient info, but should be very robust start at maxdist, backtrack until energy decrease is sufficient ------------------------------------------------------------------------- */ int MinLineSearch::linemin_backtrack(double eoriginal, double &alpha) { int i,m,n; double fdothall,fdothme,hme,hmax,hmaxall; double de_ideal,de; double *xatom,*x0atom,*fatom,*hatom; // fdothall = projection of search dir along downhill gradient // if search direction is not downhill, exit with error fdothme = 0.0; for (i = 0; i < nvec; i++) fdothme += fvec[i]*h[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) fdothme += fatom[i]*hatom[i]; } MPI_Allreduce(&fdothme,&fdothall,1,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) for (i = 0; i < nextra_global; i++) fdothall += fextra[i]*hextra[i]; if (output->thermo->normflag) fdothall /= atom->natoms; if (fdothall <= 0.0) return DOWNHILL; // set alpha so no dof is changed by more than max allowed amount // for atom coords, max amount = dmax // for extra per-atom dof, max amount = extra_max[] // for extra global dof, max amount is set by fix // also insure alpha <= ALPHA_MAX // else will have to backtrack from huge value when forces are tiny // if all search dir components are already 0.0, exit with error hme = 0.0; for (i = 0; i < nvec; i++) hme = MAX(hme,fabs(h[i])); MPI_Allreduce(&hme,&hmaxall,1,MPI_DOUBLE,MPI_MAX,world); alpha = MIN(ALPHA_MAX,dmax/hmaxall); if (nextra_atom) for (m = 0; m < nextra_atom; m++) { hatom = hextra_atom[m]; n = extra_nlen[m]; hme = 0.0; for (i = 0; i < n; i++) hme = MAX(hme,fabs(hatom[i])); MPI_Allreduce(&hme,&hmax,1,MPI_DOUBLE,MPI_MAX,world); alpha = MIN(alpha,extra_max[m]/hmax); hmaxall = MAX(hmaxall,hmax); } if (nextra_global) { double alpha_extra = modify->max_alpha(hextra); alpha = MIN(alpha,alpha_extra); for (i = 0; i < nextra_global; i++) hmaxall = MAX(hmaxall,fabs(hextra[i])); } if (hmaxall == 0.0) return ZEROFORCE; // store box and values of all dof at start of linesearch fix_minimize->store_box(); for (i = 0; i < nvec; i++) x0[i] = xvec[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { xatom = xextra_atom[m]; x0atom = x0extra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) x0atom[i] = xatom[i]; } if (nextra_global) modify->min_store(); // important diagnostic: test the gradient against energy // double etmp; // double alphatmp = alphamax*1.0e-4; // etmp = alpha_step(alphatmp,1); // printf("alpha = %g dele = %g dele_force = %g err = %g\n", // alphatmp,etmp-eoriginal,-alphatmp*fdothall, // etmp-eoriginal+alphatmp*fdothall); // alpha_step(0.0,1); // backtrack with alpha until energy decrease is sufficient while (1) { ecurrent = alpha_step(alpha,1); // if energy change is better than ideal, exit with success de_ideal = -BACKTRACK_SLOPE*alpha*fdothall; de = ecurrent - eoriginal; if (de <= de_ideal) { if (nextra_global) { int itmp = modify->min_reset_ref(); if (itmp) ecurrent = energy_force(1); } return 0; } // reduce alpha alpha *= ALPHA_REDUCE; // backtracked all the way to 0.0 // reset to starting point, exit with error if (alpha <= 0.0 || de_ideal >= -EMACH) { ecurrent = alpha_step(0.0,0); return ZEROALPHA; } } } /* ---------------------------------------------------------------------- // linemin: quadratic line search (adapted from Dennis and Schnabel) // The objective function is approximated by a quadratic // function in alpha, for sufficiently small alpha. // This idea is the same as that used in the well-known secant // method. However, since the change in the objective function // (difference of two finite numbers) is not known as accurately // as the gradient (which is close to zero), all the expressions // are written in terms of gradients. In this way, we can converge // the LAMMPS forces much closer to zero. // // We know E,Eprev,fh,fhprev. The Taylor series about alpha_prev // truncated at the quadratic term is: // // E = Eprev - del_alpha*fhprev + (1/2)del_alpha^2*Hprev // // and // // fh = fhprev - del_alpha*Hprev // // where del_alpha = alpha-alpha_prev // // We solve these two equations for Hprev and E=Esolve, giving: // // Esolve = Eprev - del_alpha*(f+fprev)/2 // // We define relerr to be: // // relerr = |(Esolve-E)/Eprev| // = |1.0 - (0.5*del_alpha*(f+fprev)+E)/Eprev| // // If this is accurate to within a reasonable tolerance, then // we go ahead and use a secant step to fh = 0: // // alpha0 = alpha - (alpha-alphaprev)*fh/delfh; // ------------------------------------------------------------------------- */ int MinLineSearch::linemin_quadratic(double eoriginal, double &alpha) { int i,m,n; double fdothall,fdothme,hme,hmax,hmaxall; double de_ideal,de; double delfh,engprev,relerr,alphaprev,fhprev,ff,fh,alpha0; double dot[2],dotall[2]; double *xatom,*x0atom,*fatom,*hatom; double alphamax; // fdothall = projection of search dir along downhill gradient // if search direction is not downhill, exit with error fdothme = 0.0; for (i = 0; i < nvec; i++) fdothme += fvec[i]*h[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) fdothme += fatom[i]*hatom[i]; } MPI_Allreduce(&fdothme,&fdothall,1,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) for (i = 0; i < nextra_global; i++) fdothall += fextra[i]*hextra[i]; if (output->thermo->normflag) fdothall /= atom->natoms; if (fdothall <= 0.0) return DOWNHILL; // set alphamax so no dof is changed by more than max allowed amount // for atom coords, max amount = dmax // for extra per-atom dof, max amount = extra_max[] // for extra global dof, max amount is set by fix // also insure alphamax <= ALPHA_MAX // else will have to backtrack from huge value when forces are tiny // if all search dir components are already 0.0, exit with error hme = 0.0; for (i = 0; i < nvec; i++) hme = MAX(hme,fabs(h[i])); MPI_Allreduce(&hme,&hmaxall,1,MPI_DOUBLE,MPI_MAX,world); alphamax = MIN(ALPHA_MAX,dmax/hmaxall); if (nextra_atom) for (m = 0; m < nextra_atom; m++) { hatom = hextra_atom[m]; n = extra_nlen[m]; hme = 0.0; for (i = 0; i < n; i++) hme = MAX(hme,fabs(hatom[i])); MPI_Allreduce(&hme,&hmax,1,MPI_DOUBLE,MPI_MAX,world); alphamax = MIN(alphamax,extra_max[m]/hmax); hmaxall = MAX(hmaxall,hmax); } if (nextra_global) { double alpha_extra = modify->max_alpha(hextra); alphamax = MIN(alphamax,alpha_extra); for (i = 0; i < nextra_global; i++) hmaxall = MAX(hmaxall,fabs(hextra[i])); } if (hmaxall == 0.0) return ZEROFORCE; // store box and values of all dof at start of linesearch fix_minimize->store_box(); for (i = 0; i < nvec; i++) x0[i] = xvec[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { xatom = xextra_atom[m]; x0atom = x0extra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) x0atom[i] = xatom[i]; } if (nextra_global) modify->min_store(); // backtrack with alpha until energy decrease is sufficient // or until get to small energy change, then perform quadratic projection alpha = alphamax; fhprev = fdothall; engprev = eoriginal; alphaprev = 0.0; // important diagnostic: test the gradient against energy // double etmp; // double alphatmp = alphamax*1.0e-4; // etmp = alpha_step(alphatmp,1); // printf("alpha = %g dele = %g dele_force = %g err = %g\n", // alphatmp,etmp-eoriginal,-alphatmp*fdothall, // etmp-eoriginal+alphatmp*fdothall); // alpha_step(0.0,1); while (1) { ecurrent = alpha_step(alpha,1); // compute new fh, alpha, delfh dot[0] = dot[1] = 0.0; for (i = 0; i < nvec; i++) { dot[0] += fvec[i]*fvec[i]; dot[1] += fvec[i]*h[i]; } if (nextra_atom) for (m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) { dot[0] += fatom[i]*fatom[i]; dot[1] += fatom[i]*hatom[i]; } } MPI_Allreduce(dot,dotall,2,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) { for (i = 0; i < nextra_global; i++) { dotall[0] += fextra[i]*fextra[i]; dotall[1] += fextra[i]*hextra[i]; } } ff = dotall[0]; fh = dotall[1]; if (output->thermo->normflag) { ff /= atom->natoms; fh /= atom->natoms; } delfh = fh - fhprev; // if fh or delfh is epsilon, reset to starting point, exit with error if (fabs(fh) < EPS_QUAD || fabs(delfh) < EPS_QUAD) { ecurrent = alpha_step(0.0,0); return ZEROQUAD; } // Check if ready for quadratic projection, equivalent to secant method // alpha0 = projected alpha relerr = fabs(1.0-(0.5*(alpha-alphaprev)*(fh+fhprev)+ecurrent)/engprev); alpha0 = alpha - (alpha-alphaprev)*fh/delfh; if (relerr <= QUADRATIC_TOL && alpha0 > 0.0 && alpha0 < alphamax) { ecurrent = alpha_step(alpha0,1); if (ecurrent - eoriginal < EMACH) { if (nextra_global) { int itmp = modify->min_reset_ref(); if (itmp) ecurrent = energy_force(1); } return 0; } } // if backtracking energy change is better than ideal, exit with success de_ideal = -BACKTRACK_SLOPE*alpha*fdothall; de = ecurrent - eoriginal; if (de <= de_ideal) { if (nextra_global) { int itmp = modify->min_reset_ref(); if (itmp) ecurrent = energy_force(1); } return 0; } // save previous state fhprev = fh; engprev = ecurrent; alphaprev = alpha; // reduce alpha alpha *= ALPHA_REDUCE; // backtracked all the way to 0.0 // reset to starting point, exit with error if (alpha <= 0.0 || de_ideal >= -EMACH) { ecurrent = alpha_step(0.0,0); return ZEROALPHA; } } } /* ---------------------------------------------------------------------- forcezero linesearch method - seeks a zero of force in a robust manner. (motivated by a line minimization routine of f77 DYNAMO code) central idea: In each linesearch we attempt to converge to a zero of force (usual case) or reduces forces (worst case). Energy does not play any role in the search procedure, except we ensure that it doesn't increase. pseudo code: i) Fix an alpha max: // also account for nextra atom & global alpha_max <= dmax/hmaxall ii) Initialize: fhCurr = current_force.dot.search_direction fhoriginal = fhCurr // try decreasing the energy to 1/10 of initial alpha_init = 0.1*fabs(eoriginal)/fhCurr; // initial alpha is smaller than alpha_max alpha_del = MIN(alpha_init, 0.5*alpha_max); alpha = 0.0 iii) Loop: backtrack = false alpha += alpha_del if (alpha > alpha_max): // we have done enough in the search space EXIT with success Step with the new alpha Compute: current energy and 'fhCurr' de = ecurrent - eprev // ZERO_ENERGY = 1e-12, is max allowed energy increase if (de > ZERO_ENERGY): bactrack = true // GRAD_TOL = 0.1 if ( (not backtrack) && (fabs(fhCurr/fh0) <= GRAD_TOL) ): // forces sufficiently reduced without energy increase EXIT with success // projected force changed sign but didn't become small enough if ( fhCurr < 0): backtrack = true if (bactrack): // forces along search direction changed sign if (fhCurr < 0): Get alpha_del by solving for zero of force (1D Newton's Method) else: // force didn't change sign but only energy increased, // we overshot a minimum which is very close to a // maximum (or there is an inflection point) // New alpha_del should be much smaller // ALPHA_FACT = 0.1 alpha_del *= ALPHA_FACT // Check to see if new 'alpha_del' isn't too small if (alpha_del < MIN_ALPHA): EXIT with failure("linesearch alpha is zero") Undo the step of alpha. // continue the loop with a new alpha_del else: Get new alpha_del by linearizing force and solving for its zero ---------------------------------------------------------------------- */ int MinLineSearch::linemin_forcezero(double eoriginal, double &alpha) { int i,m,n; double fdothall,fdothme,hme,hmax,hmaxall; double de_ideal,de; double *xatom,*x0atom,*fatom,*hatom; double alpha_max, alpha_init, alpha_del; // projection of: force on itself, current force on search direction, double ffCurr, fhCurr; // previous force on search direction, initial force on search direction double fhPrev, fhoriginal; // current energy, previous energy double engCurr, engPrev; bool backtrack; // hardcoded constants // factor by which alpha is reduced when backtracking double ALPHA_FACT = 0.1; // maximum amount by which we'll permit energy increase double ZERO_ENERGY = 1e-12; // fraction to which we want to reduce the directional derivative double GRAD_TOL = 0.1; // largest alpha increment which will trigger a failed_linesearch double MIN_ALPHA_FAC = 1e-20; double LIMIT_BOOST = 4.0; // fdothall = projection of search dir along downhill gradient // if search direction is not downhill, exit with error fdothme = 0.0; for (i = 0; i < nvec; i++) fdothme += fvec[i]*h[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) fdothme += fatom[i]*hatom[i]; } MPI_Allreduce(&fdothme,&fdothall,1,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) for (i = 0; i < nextra_global; i++) fdothall += fextra[i]*hextra[i]; if (output->thermo->normflag) fdothall /= atom->natoms; if (fdothall <= 0.0) return DOWNHILL; // set alpha so no dof is changed by more than max allowed amount // for atom coords, max amount = dmax // for extra per-atom dof, max amount = extra_max[] // for extra global dof, max amount is set by fix // also insure alpha <= ALPHA_MAX else will have // to backtrack from huge value when forces are tiny // if all search dir components are already 0.0, exit with error hme = 0.0; for (i = 0; i < nvec; i++) hme = MAX(hme,fabs(h[i])); MPI_Allreduce(&hme,&hmaxall,1,MPI_DOUBLE,MPI_MAX,world); alpha_max = dmax/hmaxall; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { hatom = hextra_atom[m]; n = extra_nlen[m]; hme = 0.0; for (i = 0; i < n; i++) hme = MAX(hme,fabs(hatom[i])); MPI_Allreduce(&hme,&hmax,1,MPI_DOUBLE,MPI_MAX,world); alpha_max = MIN(alpha_max,extra_max[m]/hmax); hmaxall = MAX(hmaxall,hmax); } if (nextra_global) { double alpha_extra = modify->max_alpha(hextra); alpha_max = MIN(alpha_max, alpha_extra); for (i = 0; i < nextra_global; i++) hmaxall = MAX(hmaxall,fabs(hextra[i])); } if (hmaxall == 0.0) return ZEROFORCE; // store box and values of all dof at start of linesearch fix_minimize->store_box(); for (i = 0; i < nvec; i++) x0[i] = xvec[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { xatom = xextra_atom[m]; x0atom = x0extra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) x0atom[i] = xatom[i]; } if (nextra_global) modify->min_store(); // initialize important variables before main linesearch loop ffCurr = 0.0; fhCurr = fdothall; fhoriginal = fhCurr; engCurr = eoriginal; // stores energy difference due to the current move de = 0.0; // choosing the initial alpha that we'll use // rough estimate that'll decrease energy to 1/10 alpha_init = 0.1*fabs(eoriginal)/fdothall; // initialize aplha to 0.0 alpha = 0.0; // compute increment to alpha, ensure that we // don't take the largest allowed alpha // first alpha that will actually apply alpha_del = MIN(alpha_init,0.5*alpha_max); // main linesearch loop while (1) { backtrack = false; fhPrev = fhCurr; engPrev = engCurr; // apply the increment to alpha, but first // check whether we are still in allowed search space alpha += alpha_del; if (alpha > alpha_max) { // undo the increment alpha -= alpha_del; if (nextra_global) { int itmp = modify->min_reset_ref(); if (itmp) ecurrent = energy_force(1); } // exit linesearch with success: have done // enough in allowed search space return 0; } // move the system // '1' updates coordinates of atoms which cross PBC engCurr = alpha_step(alpha,1); ecurrent = engCurr; // compute the new directional derivative and also f_dot_f fhCurr = compute_dir_deriv(ffCurr); // energy change de = engCurr - engPrev; // if the function value increases measurably, // then we have to reduce alpha if (de >= ZERO_ENERGY) backtrack = true; // check if the directional derivative has sufficiently decreased // NOTE: the fabs is essential here if ((!backtrack) && (fabs(fhCurr/fhoriginal) <= GRAD_TOL)) { if (nextra_global) { int itmp = modify->min_reset_ref(); if (itmp) ecurrent = energy_force(1); } // we are done return 0; } // check if the directional derivative changed sign // but it's not small: we overshot the minima -- BACKTRACK if (fhCurr < 0.0) backtrack = true; // backtrack by undoing step and choosing a new alpha if (backtrack) { // move back alpha -= alpha_del; // choose new alpha // if the force changed sign, linearize force and // solve for new alpha_del if (fhCurr < 0.0) alpha_del *= fhPrev/(fhPrev - fhCurr); else // force didn't change sign but only energy increased, // we overshot a minimum which is very close to a maxima // (or there is an inflection point) // new alpha_del should be much smaller alpha_del *= ALPHA_FACT; // since we moved back ... engCurr = engPrev; ecurrent = engCurr; fhCurr = fhPrev; // if new move is too small then we have failed; // exit with 'failed_linesearch' if (hmaxall*alpha_del <= MIN_ALPHA_FAC) { // undo all line minization moves engCurr = alpha_step(0.0,1); ecurrent= engCurr; return ZEROALPHA; } } else { // get a new alpha by linearizing force and start over double boostFactor = LIMIT_BOOST; // avoids problems near an energy inflection point if (fhPrev > fhCurr) boostFactor = fhCurr/(fhPrev - fhCurr); // don't want to boost too much boostFactor = MIN(boostFactor, LIMIT_BOOST); alpha_del *= boostFactor; } } } /* ---------------------------------------------------------------------- */ double MinLineSearch::alpha_step(double alpha, int resetflag) { int i,n,m; double *xatom,*x0atom,*hatom; // reset to starting point if (nextra_global) modify->min_step(0.0,hextra); for (i = 0; i < nvec; i++) xvec[i] = x0[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { xatom = xextra_atom[m]; x0atom = x0extra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) xatom[i] = x0atom[i]; requestor[m]->min_x_set(m); } // step forward along h if (alpha > 0.0) { if (nextra_global) modify->min_step(alpha,hextra); for (i = 0; i < nvec; i++) xvec[i] += alpha*h[i]; if (nextra_atom) for (m = 0; m < nextra_atom; m++) { xatom = xextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) xatom[i] += alpha*hatom[i]; requestor[m]->min_x_set(m); } } // compute and return new energy neval++; return energy_force(resetflag); } /* ---------------------------------------------------------------------- */ // compute projection of force on: itself and the search direction double MinLineSearch::compute_dir_deriv(double &ff) { int i,m,n; double *xatom,*hatom, *fatom; double dot[2],dotall[2]; double fh; // compute new fh, alpha, delfh dot[0] = dot[1] = 0.0; for (i = 0; i < nvec; i++) { dot[0] += fvec[i]*fvec[i]; dot[1] += fvec[i]*h[i]; } if (nextra_atom) for (m = 0; m < nextra_atom; m++) { fatom = fextra_atom[m]; hatom = hextra_atom[m]; n = extra_nlen[m]; for (i = 0; i < n; i++) { dot[0] += fatom[i]*fatom[i]; dot[1] += fatom[i]*hatom[i]; } } MPI_Allreduce(dot,dotall,2,MPI_DOUBLE,MPI_SUM,world); if (nextra_global) { for (i = 0; i < nextra_global; i++) { dotall[0] += fextra[i]*fextra[i]; dotall[1] += fextra[i]*hextra[i]; } } ff = dotall[0]; fh = dotall[1]; if (output->thermo->normflag) { ff /= atom->natoms; fh /= atom->natoms; } return fh; } diff --git a/src/min_linesearch.h b/src/min_linesearch.h index cabf5f14d..bd539747d 100644 --- a/src/min_linesearch.h +++ b/src/min_linesearch.h @@ -1,57 +1,57 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifndef LMP_MIN_LSRCH_H #define LMP_MIN_LSRCH_H #include "min.h" namespace LAMMPS_NS { class MinLineSearch : public Min { public: MinLineSearch(class LAMMPS *); ~MinLineSearch(); - void init_style(); + void init(); void setup_style(); void reset_vectors(); protected: // vectors needed by linesearch minimizers // allocated and stored by fix_minimize // x,f are stored by parent or Atom class or Pair class double *x0; // coords at start of linesearch double *g; // old gradient vector double *h; // search direction vector double *gextra; // g,h for extra global dof, x0 is stored by fix double *hextra; double **x0extra_atom; // x0,g,h for extra per-atom dof double **gextra_atom; double **hextra_atom; typedef int (MinLineSearch::*FnPtr)(double, double &); FnPtr linemin; int linemin_backtrack(double, double &); int linemin_quadratic(double, double &); int linemin_forcezero(double, double &); double alpha_step(double, int); double compute_dir_deriv(double &); }; } #endif diff --git a/src/min_quickmin.cpp b/src/min_quickmin.cpp index 07a887073..e12379827 100644 --- a/src/min_quickmin.cpp +++ b/src/min_quickmin.cpp @@ -1,237 +1,239 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #include "lmptype.h" #include "mpi.h" #include "math.h" #include "min_quickmin.h" #include "universe.h" #include "atom.h" #include "force.h" #include "update.h" #include "output.h" #include "timer.h" #include "error.h" using namespace LAMMPS_NS; // EPS_ENERGY = minimum normalization for energy tolerance #define EPS_ENERGY 1.0e-8 // same as in other min classes enum{MAXITER,MAXEVAL,ETOL,FTOL,DOWNHILL,ZEROALPHA,ZEROFORCE,ZEROQUAD}; #define DELAYSTEP 5 /* ---------------------------------------------------------------------- */ MinQuickMin::MinQuickMin(LAMMPS *lmp) : Min(lmp) {} /* ---------------------------------------------------------------------- */ -void MinQuickMin::init_style() +void MinQuickMin::init() { + Min::init(); + dt = update->dt; } /* ---------------------------------------------------------------------- */ void MinQuickMin::setup_style() { double **v = atom->v; int nlocal = atom->nlocal; for (int i = 0; i < nlocal; i++) v[i][0] = v[i][1] = v[i][2] = 0.0; } /* ---------------------------------------------------------------------- set current vector lengths and pointers called after atoms have migrated ------------------------------------------------------------------------- */ void MinQuickMin::reset_vectors() { // atomic dof nvec = 3 * atom->nlocal; if (nvec) xvec = atom->x[0]; if (nvec) fvec = atom->f[0]; } /* ---------------------------------------------------------------------- minimization via QuickMin damped dynamics ------------------------------------------------------------------------- */ int MinQuickMin::iterate(int maxiter) { bigint ntimestep; double vmax,vdotf,vdotfall,fdotf,fdotfall,scale; double dtvone,dtv,dtfm; int flag,flagall; alpha_final = 0.0; bigint last_negative = update->ntimestep; for (int iter = 0; iter < maxiter; iter++) { ntimestep = ++update->ntimestep; niter++; // zero velocity if anti-parallel to force // else project velocity in direction of force double **v = atom->v; double **f = atom->f; int nlocal = atom->nlocal; vdotf = 0.0; for (int i = 0; i < nlocal; i++) vdotf += v[i][0]*f[i][0] + v[i][1]*f[i][1] + v[i][2]*f[i][2]; MPI_Allreduce(&vdotf,&vdotfall,1,MPI_DOUBLE,MPI_SUM,world); // sum vdotf over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { vdotf = vdotfall; MPI_Allreduce(&vdotf,&vdotfall,1,MPI_DOUBLE,MPI_SUM,universe->uworld); } if (vdotfall < 0.0) { last_negative = ntimestep; for (int i = 0; i < nlocal; i++) v[i][0] = v[i][1] = v[i][2] = 0.0; } else { fdotf = 0.0; for (int i = 0; i < nlocal; i++) fdotf += f[i][0]*f[i][0] + f[i][1]*f[i][1] + f[i][2]*f[i][2]; MPI_Allreduce(&fdotf,&fdotfall,1,MPI_DOUBLE,MPI_SUM,world); // sum fdotf over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { fdotf = fdotfall; MPI_Allreduce(&fdotf,&fdotfall,1,MPI_DOUBLE,MPI_SUM,universe->uworld); } if (fdotfall == 0.0) scale = 0.0; else scale = vdotfall/fdotfall; for (int i = 0; i < nlocal; i++) { v[i][0] = scale * f[i][0]; v[i][1] = scale * f[i][1]; v[i][2] = scale * f[i][2]; } } // limit timestep so no particle moves further than dmax double *rmass = atom->rmass; double *mass = atom->mass; int *type = atom->type; dtvone = dt; for (int i = 0; i < nlocal; i++) { vmax = MAX(fabs(v[i][0]),fabs(v[i][1])); vmax = MAX(vmax,fabs(v[i][2])); if (dtvone*vmax > dmax) dtvone = dmax/vmax; } MPI_Allreduce(&dtvone,&dtv,1,MPI_DOUBLE,MPI_MIN,world); // min dtv over replicas, if necessary // this communicator would be invalid for multiprocess replicas if (update->multireplica == 1) { dtvone = dtv; MPI_Allreduce(&dtvone,&dtv,1,MPI_DOUBLE,MPI_MIN,universe->uworld); } // Euler integration step double **x = atom->x; if (rmass) { for (int i = 0; i < nlocal; i++) { dtfm = dtv / rmass[i]; x[i][0] += dtv * v[i][0]; x[i][1] += dtv * v[i][1]; x[i][2] += dtv * v[i][2]; v[i][0] += dtfm * f[i][0]; v[i][1] += dtfm * f[i][1]; v[i][2] += dtfm * f[i][2]; } } else { for (int i = 0; i < nlocal; i++) { dtfm = dtv / mass[type[i]]; x[i][0] += dtv * v[i][0]; x[i][1] += dtv * v[i][1]; x[i][2] += dtv * v[i][2]; v[i][0] += dtfm * f[i][0]; v[i][1] += dtfm * f[i][1]; v[i][2] += dtfm * f[i][2]; } } eprevious = ecurrent; ecurrent = energy_force(0); neval++; // energy tolerance criterion // only check after DELAYSTEP elapsed since velocties reset to 0 // sync across replicas if running multi-replica minimization if (update->etol > 0.0 && ntimestep-last_negative > DELAYSTEP) { if (update->multireplica == 0) { if (fabs(ecurrent-eprevious) < update->etol * 0.5*(fabs(ecurrent) + fabs(eprevious) + EPS_ENERGY)) return ETOL; } else { if (fabs(ecurrent-eprevious) < update->etol * 0.5*(fabs(ecurrent) + fabs(eprevious) + EPS_ENERGY)) flag = 0; else flag = 1; MPI_Allreduce(&flag,&flagall,1,MPI_INT,MPI_SUM,universe->uworld); if (flagall == 0) return ETOL; } } // force tolerance criterion // sync across replicas if running multi-replica minimization if (update->ftol > 0.0) { fdotf = fnorm_sqr(); if (update->multireplica == 0) { if (fdotf < update->ftol*update->ftol) return FTOL; } else { if (fdotf < update->ftol*update->ftol) flag = 0; else flag = 1; MPI_Allreduce(&flag,&flagall,1,MPI_INT,MPI_SUM,universe->uworld); if (flagall == 0) return FTOL; } } // output for thermo, dump, restart files if (output->next == ntimestep) { timer->stamp(); output->write(ntimestep); timer->stamp(Timer::OUTPUT); } } return MAXITER; } diff --git a/src/min_quickmin.h b/src/min_quickmin.h index 2ced7dce5..e46bcb25e 100644 --- a/src/min_quickmin.h +++ b/src/min_quickmin.h @@ -1,43 +1,43 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifdef MINIMIZE_CLASS MinimizeStyle(quickmin,MinQuickMin) #else #ifndef LMP_MIN_QUICKMIN_H #define LMP_MIN_QUICKMIN_H #include "min.h" namespace LAMMPS_NS { class MinQuickMin : public Min { public: MinQuickMin(class LAMMPS *); ~MinQuickMin() {} - void init_style(); + void init(); void setup_style(); void reset_vectors(); int iterate(int); private: double dt; }; } #endif #endif diff --git a/src/pair.cpp b/src/pair.cpp index 234235211..722feec10 100644 --- a/src/pair.cpp +++ b/src/pair.cpp @@ -1,1141 +1,1156 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- Contributing author: Paul Crozier (SNL) ------------------------------------------------------------------------- */ #include "mpi.h" #include "float.h" #include "limits.h" #include "math.h" #include "stdio.h" #include "stdlib.h" #include "string.h" #include "pair.h" #include "atom.h" #include "neighbor.h" #include "neigh_list.h" #include "domain.h" #include "comm.h" #include "force.h" #include "update.h" #include "accelerator_cuda.h" #include "memory.h" #include "error.h" #include "suffix.h" using namespace LAMMPS_NS; enum{GEOMETRIC,ARITHMETIC,SIXTHPOWER}; enum{R,RSQ,BMP}; /* ---------------------------------------------------------------------- */ Pair::Pair(LAMMPS *lmp) : Pointers(lmp) { THIRD = 1.0/3.0; eng_vdwl = eng_coul = 0.0; comm_forward = comm_reverse = comm_reverse_off = 0; single_enable = 1; restartinfo = 1; respa_enable = 0; one_coeff = 0; no_virial_fdotr_compute = 0; ghostneigh = 0; nextra = 0; pvector = NULL; single_extra = 0; svector = NULL; // pair_modify settings + compute_flag = 1; offset_flag = 0; mix_flag = GEOMETRIC; tail_flag = 0; etail = ptail = etail_ij = ptail_ij = 0.0; ncoultablebits = 12; tabinner = sqrt(2.0); allocated = 0; suffix_flag = Suffix::NONE; maxeatom = maxvatom = 0; eatom = NULL; vatom = NULL; } /* ---------------------------------------------------------------------- */ Pair::~Pair() { memory->destroy(eatom); memory->destroy(vatom); } /* ---------------------------------------------------------------------- modify parameters of the pair style pair_hybrid has its own version of this routine for its sub-styles ------------------------------------------------------------------------- */ void Pair::modify_params(int narg, char **arg) { if (narg == 0) error->all(FLERR,"Illegal pair_modify command"); int iarg = 0; while (iarg < narg) { if (strcmp(arg[iarg],"mix") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); if (strcmp(arg[iarg+1],"geometric") == 0) mix_flag = GEOMETRIC; else if (strcmp(arg[iarg+1],"arithmetic") == 0) mix_flag = ARITHMETIC; else if (strcmp(arg[iarg+1],"sixthpower") == 0) mix_flag = SIXTHPOWER; else error->all(FLERR,"Illegal pair_modify command"); iarg += 2; } else if (strcmp(arg[iarg],"shift") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); if (strcmp(arg[iarg+1],"yes") == 0) offset_flag = 1; else if (strcmp(arg[iarg+1],"no") == 0) offset_flag = 0; else error->all(FLERR,"Illegal pair_modify command"); iarg += 2; } else if (strcmp(arg[iarg],"table") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); ncoultablebits = atoi(arg[iarg+1]); if (ncoultablebits > sizeof(float)*CHAR_BIT) error->all(FLERR,"Too many total bits for bitmapped lookup table"); iarg += 2; } else if (strcmp(arg[iarg],"tabinner") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); tabinner = atof(arg[iarg+1]); iarg += 2; } else if (strcmp(arg[iarg],"tail") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); if (strcmp(arg[iarg+1],"yes") == 0) tail_flag = 1; else if (strcmp(arg[iarg+1],"no") == 0) tail_flag = 0; else error->all(FLERR,"Illegal pair_modify command"); iarg += 2; + } else if (strcmp(arg[iarg],"compute") == 0) { + if (iarg+2 > narg) error->all(FLERR,"Illegal pair_modify command"); + if (strcmp(arg[iarg+1],"yes") == 0) compute_flag = 1; + else if (strcmp(arg[iarg+1],"no") == 0) compute_flag = 0; + else error->all(FLERR,"Illegal pair_modify command"); + iarg += 2; } else error->all(FLERR,"Illegal pair_modify command"); } } /* ---------------------------------------------------------------------- */ void Pair::init() { int i,j; if (offset_flag && tail_flag) error->all(FLERR,"Cannot have both pair_modify shift and tail set to yes"); if (tail_flag && domain->dimension == 2) error->all(FLERR,"Cannot use pair tail corrections with 2d simulations"); if (tail_flag && domain->nonperiodic && comm->me == 0) error->warning(FLERR,"Using pair tail corrections with nonperiodic system"); if (!allocated) error->all(FLERR,"All pair coeffs are not set"); // I,I coeffs must be set // init_one() will check if I,J is set explicitly or inferred by mixing for (i = 1; i <= atom->ntypes; i++) if (setflag[i][i] == 0) error->all(FLERR,"All pair coeffs are not set"); // style-specific initialization init_style(); // call init_one() for each I,J // set cutsq for each I,J, used to neighbor // cutforce = max of all I,J cutoffs cutforce = 0.0; etail = ptail = 0.0; double cut; for (i = 1; i <= atom->ntypes; i++) for (j = i; j <= atom->ntypes; j++) { cut = init_one(i,j); cutsq[i][j] = cutsq[j][i] = cut*cut; cutforce = MAX(cutforce,cut); if (tail_flag) { etail += etail_ij; ptail += ptail_ij; if (i != j) { etail += etail_ij; ptail += ptail_ij; } } } } /* ---------------------------------------------------------------------- reset all type-based params by invoking init_one() for each I,J called by fix adapt after it changes one or more params ------------------------------------------------------------------------- */ void Pair::reinit() { int i,j; double tmp; etail = ptail = 0.0; for (i = 1; i <= atom->ntypes; i++) for (j = i; j <= atom->ntypes; j++) { tmp = init_one(i,j); if (tail_flag) { etail += etail_ij; ptail += ptail_ij; if (i != j) { etail += etail_ij; ptail += ptail_ij; } } } } /* ---------------------------------------------------------------------- init specific to a pair style specific pair style can override this function if needs its own error checks if needs another kind of neighbor list request default neighbor list = half list ------------------------------------------------------------------------- */ void Pair::init_style() { neighbor->request(this); } /* ---------------------------------------------------------------------- neighbor callback to inform pair style of neighbor list to use specific pair style can override this function ------------------------------------------------------------------------- */ void Pair::init_list(int which, NeighList *ptr) { list = ptr; } /* ---------------------------------------------------------------------- mixing of pair potential prefactors (epsilon) ------------------------------------------------------------------------- */ double Pair::mix_energy(double eps1, double eps2, double sig1, double sig2) { double value; if (mix_flag == GEOMETRIC) value = sqrt(eps1*eps2); else if (mix_flag == ARITHMETIC) value = sqrt(eps1*eps2); else if (mix_flag == SIXTHPOWER) value = 2.0 * sqrt(eps1*eps2) * pow(sig1,3.0) * pow(sig2,3.0) / (pow(sig1,6.0) + pow(sig2,6.0)); return value; } /* ---------------------------------------------------------------------- mixing of pair potential distances (sigma, cutoff) ------------------------------------------------------------------------- */ double Pair::mix_distance(double sig1, double sig2) { double value; if (mix_flag == GEOMETRIC) value = sqrt(sig1*sig2); else if (mix_flag == ARITHMETIC) value = 0.5 * (sig1+sig2); else if (mix_flag == SIXTHPOWER) value = pow((0.5 * (pow(sig1,6.0) + pow(sig2,6.0))),1.0/6.0); return value; } +/* ---------------------------------------------------------------------- */ + +void Pair::compute_dummy(int eflag, int vflag) +{ + if (eflag || vflag) ev_setup(eflag,vflag); + else evflag = 0; +} + /* ---------------------------------------------------------------------- setup for energy, virial computation see integrate::ev_set() for values of eflag (0-3) and vflag (0-6) ------------------------------------------------------------------------- */ void Pair::ev_setup(int eflag, int vflag) { int i,n; evflag = 1; eflag_either = eflag; eflag_global = eflag % 2; eflag_atom = eflag / 2; vflag_either = vflag; vflag_global = vflag % 4; vflag_atom = vflag / 4; // reallocate per-atom arrays if necessary if (eflag_atom && atom->nmax > maxeatom) { maxeatom = atom->nmax; memory->destroy(eatom); memory->create(eatom,comm->nthreads*maxeatom,"pair:eatom"); } if (vflag_atom && atom->nmax > maxvatom) { maxvatom = atom->nmax; memory->destroy(vatom); memory->create(vatom,comm->nthreads*maxvatom,6,"pair:vatom"); } // zero accumulators // use force->newton instead of newton_pair // b/c some bonds/dihedrals call pair::ev_tally with pairwise info if (eflag_global) eng_vdwl = eng_coul = 0.0; if (vflag_global) for (i = 0; i < 6; i++) virial[i] = 0.0; if (eflag_atom) { n = atom->nlocal; if (force->newton) n += atom->nghost; for (i = 0; i < n; i++) eatom[i] = 0.0; } if (vflag_atom) { n = atom->nlocal; if (force->newton) n += atom->nghost; for (i = 0; i < n; i++) { vatom[i][0] = 0.0; vatom[i][1] = 0.0; vatom[i][2] = 0.0; vatom[i][3] = 0.0; vatom[i][4] = 0.0; vatom[i][5] = 0.0; } } // if vflag_global = 2 and pair::compute() calls virial_fdotr_compute() // compute global virial via (F dot r) instead of via pairwise summation // unset other flags as appropriate if (vflag_global == 2 && no_virial_fdotr_compute == 0) { vflag_fdotr = 1; vflag_global = 0; if (vflag_atom == 0) vflag_either = 0; if (vflag_either == 0 && eflag_either == 0) evflag = 0; } else vflag_fdotr = 0; if (lmp->cuda) lmp->cuda->evsetup_eatom_vatom(eflag_atom,vflag_atom); } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators need i < nlocal test since called by bond_quartic and dihedral_charmm ------------------------------------------------------------------------- */ void Pair::ev_tally(int i, int j, int nlocal, int newton_pair, double evdwl, double ecoul, double fpair, double delx, double dely, double delz) { double evdwlhalf,ecoulhalf,epairhalf,v[6]; if (eflag_either) { if (eflag_global) { if (newton_pair) { eng_vdwl += evdwl; eng_coul += ecoul; } else { evdwlhalf = 0.5*evdwl; ecoulhalf = 0.5*ecoul; if (i < nlocal) { eng_vdwl += evdwlhalf; eng_coul += ecoulhalf; } if (j < nlocal) { eng_vdwl += evdwlhalf; eng_coul += ecoulhalf; } } } if (eflag_atom) { epairhalf = 0.5 * (evdwl + ecoul); if (newton_pair || i < nlocal) eatom[i] += epairhalf; if (newton_pair || j < nlocal) eatom[j] += epairhalf; } } if (vflag_either) { v[0] = delx*delx*fpair; v[1] = dely*dely*fpair; v[2] = delz*delz*fpair; v[3] = delx*dely*fpair; v[4] = delx*delz*fpair; v[5] = dely*delz*fpair; if (vflag_global) { if (newton_pair) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } else { if (i < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } if (j < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } } } if (vflag_atom) { if (newton_pair || i < nlocal) { vatom[i][0] += 0.5*v[0]; vatom[i][1] += 0.5*v[1]; vatom[i][2] += 0.5*v[2]; vatom[i][3] += 0.5*v[3]; vatom[i][4] += 0.5*v[4]; vatom[i][5] += 0.5*v[5]; } if (newton_pair || j < nlocal) { vatom[j][0] += 0.5*v[0]; vatom[j][1] += 0.5*v[1]; vatom[j][2] += 0.5*v[2]; vatom[j][3] += 0.5*v[3]; vatom[j][4] += 0.5*v[4]; vatom[j][5] += 0.5*v[5]; } } } } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators can use this version with full neighbor lists ------------------------------------------------------------------------- */ void Pair::ev_tally_full(int i, double evdwl, double ecoul, double fpair, double delx, double dely, double delz) { double v[6]; if (eflag_either) { if (eflag_global) { eng_vdwl += 0.5*evdwl; eng_coul += 0.5*ecoul; } if (eflag_atom) eatom[i] += 0.5 * (evdwl + ecoul); } if (vflag_either) { v[0] = 0.5*delx*delx*fpair; v[1] = 0.5*dely*dely*fpair; v[2] = 0.5*delz*delz*fpair; v[3] = 0.5*delx*dely*fpair; v[4] = 0.5*delx*delz*fpair; v[5] = 0.5*dely*delz*fpair; if (vflag_global) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } if (vflag_atom) { vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; } } } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators for virial, have delx,dely,delz and fx,fy,fz ------------------------------------------------------------------------- */ void Pair::ev_tally_xyz(int i, int j, int nlocal, int newton_pair, double evdwl, double ecoul, double fx, double fy, double fz, double delx, double dely, double delz) { double evdwlhalf,ecoulhalf,epairhalf,v[6]; if (eflag_either) { if (eflag_global) { if (newton_pair) { eng_vdwl += evdwl; eng_coul += ecoul; } else { evdwlhalf = 0.5*evdwl; ecoulhalf = 0.5*ecoul; if (i < nlocal) { eng_vdwl += evdwlhalf; eng_coul += ecoulhalf; } if (j < nlocal) { eng_vdwl += evdwlhalf; eng_coul += ecoulhalf; } } } if (eflag_atom) { epairhalf = 0.5 * (evdwl + ecoul); if (newton_pair || i < nlocal) eatom[i] += epairhalf; if (newton_pair || j < nlocal) eatom[j] += epairhalf; } } if (vflag_either) { v[0] = delx*fx; v[1] = dely*fy; v[2] = delz*fz; v[3] = delx*fy; v[4] = delx*fz; v[5] = dely*fz; if (vflag_global) { if (newton_pair) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } else { if (i < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } if (j < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } } } if (vflag_atom) { if (newton_pair || i < nlocal) { vatom[i][0] += 0.5*v[0]; vatom[i][1] += 0.5*v[1]; vatom[i][2] += 0.5*v[2]; vatom[i][3] += 0.5*v[3]; vatom[i][4] += 0.5*v[4]; vatom[i][5] += 0.5*v[5]; } if (newton_pair || j < nlocal) { vatom[j][0] += 0.5*v[0]; vatom[j][1] += 0.5*v[1]; vatom[j][2] += 0.5*v[2]; vatom[j][3] += 0.5*v[3]; vatom[j][4] += 0.5*v[4]; vatom[j][5] += 0.5*v[5]; } } } } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators for virial, have delx,dely,delz and fx,fy,fz called when using full neighbor lists ------------------------------------------------------------------------- */ void Pair::ev_tally_xyz_full(int i, double evdwl, double ecoul, double fx, double fy, double fz, double delx, double dely, double delz) { double evdwlhalf,ecoulhalf,epairhalf,v[6]; if (eflag_either) { if (eflag_global) { evdwlhalf = 0.5*evdwl; ecoulhalf = 0.5*ecoul; eng_vdwl += evdwlhalf; eng_coul += ecoulhalf; } if (eflag_atom) { epairhalf = 0.5 * (evdwl + ecoul); eatom[i] += epairhalf; } } if (vflag_either) { v[0] = 0.5*delx*fx; v[1] = 0.5*dely*fy; v[2] = 0.5*delz*fz; v[3] = 0.5*delx*fy; v[4] = 0.5*delx*fz; v[5] = 0.5*dely*fz; if (vflag_global) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } if (vflag_atom) { vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; } } } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators called by SW and hbond potentials, newton_pair is always on virial = riFi + rjFj + rkFk = (rj-ri) Fj + (rk-ri) Fk = drji*fj + drki*fk ------------------------------------------------------------------------- */ void Pair::ev_tally3(int i, int j, int k, double evdwl, double ecoul, double *fj, double *fk, double *drji, double *drki) { double epairthird,v[6]; if (eflag_either) { if (eflag_global) { eng_vdwl += evdwl; eng_coul += ecoul; } if (eflag_atom) { epairthird = THIRD * (evdwl + ecoul); eatom[i] += epairthird; eatom[j] += epairthird; eatom[k] += epairthird; } } if (vflag_either) { v[0] = drji[0]*fj[0] + drki[0]*fk[0]; v[1] = drji[1]*fj[1] + drki[1]*fk[1]; v[2] = drji[2]*fj[2] + drki[2]*fk[2]; v[3] = drji[0]*fj[1] + drki[0]*fk[1]; v[4] = drji[0]*fj[2] + drki[0]*fk[2]; v[5] = drji[1]*fj[2] + drki[1]*fk[2]; if (vflag_global) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } if (vflag_atom) { vatom[i][0] += THIRD*v[0]; vatom[i][1] += THIRD*v[1]; vatom[i][2] += THIRD*v[2]; vatom[i][3] += THIRD*v[3]; vatom[i][4] += THIRD*v[4]; vatom[i][5] += THIRD*v[5]; vatom[j][0] += THIRD*v[0]; vatom[j][1] += THIRD*v[1]; vatom[j][2] += THIRD*v[2]; vatom[j][3] += THIRD*v[3]; vatom[j][4] += THIRD*v[4]; vatom[j][5] += THIRD*v[5]; vatom[k][0] += THIRD*v[0]; vatom[k][1] += THIRD*v[1]; vatom[k][2] += THIRD*v[2]; vatom[k][3] += THIRD*v[3]; vatom[k][4] += THIRD*v[4]; vatom[k][5] += THIRD*v[5]; } } } /* ---------------------------------------------------------------------- tally eng_vdwl and virial into global and per-atom accumulators called by AIREBO potential, newton_pair is always on ------------------------------------------------------------------------- */ void Pair::ev_tally4(int i, int j, int k, int m, double evdwl, double *fi, double *fj, double *fk, double *drim, double *drjm, double *drkm) { double epairfourth,v[6]; if (eflag_either) { if (eflag_global) eng_vdwl += evdwl; if (eflag_atom) { epairfourth = 0.25 * evdwl; eatom[i] += epairfourth; eatom[j] += epairfourth; eatom[k] += epairfourth; eatom[m] += epairfourth; } } if (vflag_atom) { v[0] = 0.25 * (drim[0]*fi[0] + drjm[0]*fj[0] + drkm[0]*fk[0]); v[1] = 0.25 * (drim[1]*fi[1] + drjm[1]*fj[1] + drkm[1]*fk[1]); v[2] = 0.25 * (drim[2]*fi[2] + drjm[2]*fj[2] + drkm[2]*fk[2]); v[3] = 0.25 * (drim[0]*fi[1] + drjm[0]*fj[1] + drkm[0]*fk[1]); v[4] = 0.25 * (drim[0]*fi[2] + drjm[0]*fj[2] + drkm[0]*fk[2]); v[5] = 0.25 * (drim[1]*fi[2] + drjm[1]*fj[2] + drkm[1]*fk[2]); vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; vatom[j][0] += v[0]; vatom[j][1] += v[1]; vatom[j][2] += v[2]; vatom[j][3] += v[3]; vatom[j][4] += v[4]; vatom[j][5] += v[5]; vatom[k][0] += v[0]; vatom[k][1] += v[1]; vatom[k][2] += v[2]; vatom[k][3] += v[3]; vatom[k][4] += v[4]; vatom[k][5] += v[5]; vatom[m][0] += v[0]; vatom[m][1] += v[1]; vatom[m][2] += v[2]; vatom[m][3] += v[3]; vatom[m][4] += v[4]; vatom[m][5] += v[5]; } } /* ---------------------------------------------------------------------- tally ecoul and virial into each of n atoms in list called by TIP4P potential, newton_pair is always on changes v values by dividing by n ------------------------------------------------------------------------- */ void Pair::ev_tally_list(int n, int *list, double ecoul, double *v) { int i,j; if (eflag_either) { if (eflag_global) eng_coul += ecoul; if (eflag_atom) { double epairatom = ecoul/n; for (i = 0; i < n; i++) eatom[list[i]] += epairatom; } } if (vflag_either) { if (vflag_global) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } if (vflag_atom) { v[0] /= n; v[1] /= n; v[2] /= n; v[3] /= n; v[4] /= n; v[5] /= n; for (i = 0; i < n; i++) { j = list[i]; vatom[j][0] += v[0]; vatom[j][1] += v[1]; vatom[j][2] += v[2]; vatom[j][3] += v[3]; vatom[j][4] += v[4]; vatom[j][5] += v[5]; } } } } /* ---------------------------------------------------------------------- tally virial into per-atom accumulators called by AIREBO potential, newton_pair is always on fpair is magnitude of force on atom I ------------------------------------------------------------------------- */ void Pair::v_tally2(int i, int j, double fpair, double *drij) { double v[6]; v[0] = 0.5 * drij[0]*drij[0]*fpair; v[1] = 0.5 * drij[1]*drij[1]*fpair; v[2] = 0.5 * drij[2]*drij[2]*fpair; v[3] = 0.5 * drij[0]*drij[1]*fpair; v[4] = 0.5 * drij[0]*drij[2]*fpair; v[5] = 0.5 * drij[1]*drij[2]*fpair; vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; vatom[j][0] += v[0]; vatom[j][1] += v[1]; vatom[j][2] += v[2]; vatom[j][3] += v[3]; vatom[j][4] += v[4]; vatom[j][5] += v[5]; } /* ---------------------------------------------------------------------- tally virial into per-atom accumulators called by AIREBO and Tersoff potential, newton_pair is always on ------------------------------------------------------------------------- */ void Pair::v_tally3(int i, int j, int k, double *fi, double *fj, double *drik, double *drjk) { double v[6]; v[0] = THIRD * (drik[0]*fi[0] + drjk[0]*fj[0]); v[1] = THIRD * (drik[1]*fi[1] + drjk[1]*fj[1]); v[2] = THIRD * (drik[2]*fi[2] + drjk[2]*fj[2]); v[3] = THIRD * (drik[0]*fi[1] + drjk[0]*fj[1]); v[4] = THIRD * (drik[0]*fi[2] + drjk[0]*fj[2]); v[5] = THIRD * (drik[1]*fi[2] + drjk[1]*fj[2]); vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; vatom[j][0] += v[0]; vatom[j][1] += v[1]; vatom[j][2] += v[2]; vatom[j][3] += v[3]; vatom[j][4] += v[4]; vatom[j][5] += v[5]; vatom[k][0] += v[0]; vatom[k][1] += v[1]; vatom[k][2] += v[2]; vatom[k][3] += v[3]; vatom[k][4] += v[4]; vatom[k][5] += v[5]; } /* ---------------------------------------------------------------------- tally virial into per-atom accumulators called by AIREBO potential, newton_pair is always on ------------------------------------------------------------------------- */ void Pair::v_tally4(int i, int j, int k, int m, double *fi, double *fj, double *fk, double *drim, double *drjm, double *drkm) { double v[6]; v[0] = 0.25 * (drim[0]*fi[0] + drjm[0]*fj[0] + drkm[0]*fk[0]); v[1] = 0.25 * (drim[1]*fi[1] + drjm[1]*fj[1] + drkm[1]*fk[1]); v[2] = 0.25 * (drim[2]*fi[2] + drjm[2]*fj[2] + drkm[2]*fk[2]); v[3] = 0.25 * (drim[0]*fi[1] + drjm[0]*fj[1] + drkm[0]*fk[1]); v[4] = 0.25 * (drim[0]*fi[2] + drjm[0]*fj[2] + drkm[0]*fk[2]); v[5] = 0.25 * (drim[1]*fi[2] + drjm[1]*fj[2] + drkm[1]*fk[2]); vatom[i][0] += v[0]; vatom[i][1] += v[1]; vatom[i][2] += v[2]; vatom[i][3] += v[3]; vatom[i][4] += v[4]; vatom[i][5] += v[5]; vatom[j][0] += v[0]; vatom[j][1] += v[1]; vatom[j][2] += v[2]; vatom[j][3] += v[3]; vatom[j][4] += v[4]; vatom[j][5] += v[5]; vatom[k][0] += v[0]; vatom[k][1] += v[1]; vatom[k][2] += v[2]; vatom[k][3] += v[3]; vatom[k][4] += v[4]; vatom[k][5] += v[5]; vatom[m][0] += v[0]; vatom[m][1] += v[1]; vatom[m][2] += v[2]; vatom[m][3] += v[3]; vatom[m][4] += v[4]; vatom[m][5] += v[5]; } /* ---------------------------------------------------------------------- tally virial into global and per-atom accumulators called by pair lubricate potential with 6 tensor components ------------------------------------------------------------------------- */ void Pair::v_tally_tensor(int i, int j, int nlocal, int newton_pair, double vxx, double vyy, double vzz, double vxy, double vxz, double vyz) { double v[6]; v[0] = vxx; v[1] = vyy; v[2] = vzz; v[3] = vxy; v[4] = vxz; v[5] = vyz; if (vflag_global) { if (newton_pair) { virial[0] += v[0]; virial[1] += v[1]; virial[2] += v[2]; virial[3] += v[3]; virial[4] += v[4]; virial[5] += v[5]; } else { if (i < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } if (j < nlocal) { virial[0] += 0.5*v[0]; virial[1] += 0.5*v[1]; virial[2] += 0.5*v[2]; virial[3] += 0.5*v[3]; virial[4] += 0.5*v[4]; virial[5] += 0.5*v[5]; } } } if (vflag_atom) { if (newton_pair || i < nlocal) { vatom[i][0] += 0.5*v[0]; vatom[i][1] += 0.5*v[1]; vatom[i][2] += 0.5*v[2]; vatom[i][3] += 0.5*v[3]; vatom[i][4] += 0.5*v[4]; vatom[i][5] += 0.5*v[5]; } if (newton_pair || j < nlocal) { vatom[j][0] += 0.5*v[0]; vatom[j][1] += 0.5*v[1]; vatom[j][2] += 0.5*v[2]; vatom[j][3] += 0.5*v[3]; vatom[j][4] += 0.5*v[4]; vatom[j][5] += 0.5*v[5]; } } } /* ---------------------------------------------------------------------- compute global pair virial via summing F dot r over own & ghost atoms at this point, only pairwise forces have been accumulated in atom->f ------------------------------------------------------------------------- */ void Pair::virial_fdotr_compute() { double **x = atom->x; double **f = atom->f; // sum over force on all particles including ghosts if (neighbor->includegroup == 0) { int nall = atom->nlocal + atom->nghost; for (int i = 0; i < nall; i++) { virial[0] += f[i][0]*x[i][0]; virial[1] += f[i][1]*x[i][1]; virial[2] += f[i][2]*x[i][2]; virial[3] += f[i][1]*x[i][0]; virial[4] += f[i][2]*x[i][0]; virial[5] += f[i][2]*x[i][1]; } // neighbor includegroup flag is set // sum over force on initial nfirst particles and ghosts } else { int nall = atom->nfirst; for (int i = 0; i < nall; i++) { virial[0] += f[i][0]*x[i][0]; virial[1] += f[i][1]*x[i][1]; virial[2] += f[i][2]*x[i][2]; virial[3] += f[i][1]*x[i][0]; virial[4] += f[i][2]*x[i][0]; virial[5] += f[i][2]*x[i][1]; } nall = atom->nlocal + atom->nghost; for (int i = atom->nlocal; i < nall; i++) { virial[0] += f[i][0]*x[i][0]; virial[1] += f[i][1]*x[i][1]; virial[2] += f[i][2]*x[i][2]; virial[3] += f[i][1]*x[i][0]; virial[4] += f[i][2]*x[i][0]; virial[5] += f[i][2]*x[i][1]; } } } /* ---------------------------------------------------------------------- write a table of pair potential energy/force vs distance to a file ------------------------------------------------------------------------- */ void Pair::write_file(int narg, char **arg) { if (narg < 8) error->all(FLERR,"Illegal pair_write command"); if (single_enable == 0) error->all(FLERR,"Pair style does not support pair_write"); // parse arguments int itype = atoi(arg[0]); int jtype = atoi(arg[1]); if (itype < 1 || itype > atom->ntypes || jtype < 1 || jtype > atom->ntypes) error->all(FLERR,"Invalid atom types in pair_write command"); int n = atoi(arg[2]); int style; if (strcmp(arg[3],"r") == 0) style = R; else if (strcmp(arg[3],"rsq") == 0) style = RSQ; else if (strcmp(arg[3],"bitmap") == 0) style = BMP; else error->all(FLERR,"Invalid style in pair_write command"); double inner = atof(arg[4]); double outer = atof(arg[5]); if (inner <= 0.0 || inner >= outer) error->all(FLERR,"Invalid cutoffs in pair_write command"); // open file in append mode // print header in format used by pair_style table int me; MPI_Comm_rank(world,&me); FILE *fp; if (me == 0) { fp = fopen(arg[6],"a"); if (fp == NULL) error->one(FLERR,"Cannot open pair_write file"); fprintf(fp,"# Pair potential %s for atom types %d %d: i,r,energy,force\n", force->pair_style,itype,jtype); if (style == R) fprintf(fp,"\n%s\nN %d R %g %g\n\n",arg[7],n,inner,outer); if (style == RSQ) fprintf(fp,"\n%s\nN %d RSQ %g %g\n\n",arg[7],n,inner,outer); } // initialize potentials before evaluating pair potential // insures all pair coeffs are set and force constants force->init(); // if pair style = any of EAM, swap in dummy fp vector double eamfp[2]; eamfp[0] = eamfp[1] = 0.0; double *eamfp_hold; Pair *epair = force->pair_match("eam",0); if (epair) epair->swap_eam(eamfp,&eamfp_hold); // if atom style defines charge, swap in dummy q vec double q[2]; q[0] = q[1] = 1.0; if (narg == 10) { q[0] = atof(arg[8]); q[1] = atof(arg[9]); } double *q_hold; if (atom->q) { q_hold = atom->q; atom->q = q; } // evaluate energy and force at each of N distances int masklo,maskhi,nmask,nshiftbits; if (style == BMP) { init_bitmap(inner,outer,n,masklo,maskhi,nmask,nshiftbits); int ntable = 1 << n; if (me == 0) fprintf(fp,"\n%s\nN %d BITMAP %g %g\n\n",arg[7],ntable,inner,outer); n = ntable; } double r,e,f,rsq; union_int_float_t rsq_lookup; for (int i = 0; i < n; i++) { if (style == R) { r = inner + (outer-inner) * i/(n-1); rsq = r*r; } else if (style == RSQ) { rsq = inner*inner + (outer*outer - inner*inner) * i/(n-1); r = sqrt(rsq); } else if (style == BMP) { rsq_lookup.i = i << nshiftbits; rsq_lookup.i |= masklo; if (rsq_lookup.f < inner*inner) { rsq_lookup.i = i << nshiftbits; rsq_lookup.i |= maskhi; } rsq = rsq_lookup.f; r = sqrt(rsq); } if (rsq < cutsq[itype][jtype]) { e = single(0,1,itype,jtype,rsq,1.0,1.0,f); f *= r; } else e = f = 0.0; if (me == 0) fprintf(fp,"%d %g %g %g\n",i+1,r,e,f); } // restore original vecs that were swapped in for double *tmp; if (epair) epair->swap_eam(eamfp_hold,&tmp); if (atom->q) atom->q = q_hold; if (me == 0) fclose(fp); } /* ---------------------------------------------------------------------- define bitmap parameters based on inner and outer cutoffs ------------------------------------------------------------------------- */ void Pair::init_bitmap(double inner, double outer, int ntablebits, int &masklo, int &maskhi, int &nmask, int &nshiftbits) { if (sizeof(int) != sizeof(float)) error->all(FLERR,"Bitmapped lookup tables require int/float be same size"); if (ntablebits > sizeof(float)*CHAR_BIT) error->all(FLERR,"Too many total bits for bitmapped lookup table"); if (inner >= outer) error->warning(FLERR,"Table inner cutoff >= outer cutoff"); int nlowermin = 1; while (!((pow(double(2),nlowermin) <= inner*inner) && (pow(double(2),nlowermin+1) > inner*inner))) { if (pow(double(2),nlowermin) <= inner*inner) nlowermin++; else nlowermin--; } int nexpbits = 0; double required_range = outer*outer / pow(double(2),nlowermin); double available_range = 2.0; while (available_range < required_range) { nexpbits++; available_range = pow(double(2),pow(double(2),nexpbits)); } int nmantbits = ntablebits - nexpbits; if (nexpbits > sizeof(float)*CHAR_BIT - FLT_MANT_DIG) error->all(FLERR,"Too many exponent bits for lookup table"); if (nmantbits+1 > FLT_MANT_DIG) error->all(FLERR,"Too many mantissa bits for lookup table"); if (nmantbits < 3) error->all(FLERR,"Too few bits for lookup table"); nshiftbits = FLT_MANT_DIG - (nmantbits+1); nmask = 1; for (int j = 0; j < ntablebits+nshiftbits; j++) nmask *= 2; nmask -= 1; union_int_float_t rsq_lookup; rsq_lookup.f = outer*outer; maskhi = rsq_lookup.i & ~(nmask); rsq_lookup.f = inner*inner; masklo = rsq_lookup.i & ~(nmask); } /* ---------------------------------------------------------------------- */ double Pair::memory_usage() { double bytes = comm->nthreads*maxeatom * sizeof(double); bytes += comm->nthreads*maxvatom*6 * sizeof(double); return bytes; } diff --git a/src/pair.h b/src/pair.h index 53fa3e9c4..d86b7005d 100644 --- a/src/pair.h +++ b/src/pair.h @@ -1,255 +1,258 @@ /* -*- c++ -*- ---------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #ifndef LMP_PAIR_H #define LMP_PAIR_H #include "pointers.h" namespace LAMMPS_NS { class Pair : protected Pointers { friend class AngleSDK; friend class AngleSDKOMP; friend class BondQuartic; friend class BondQuarticOMP; friend class DihedralCharmm; friend class DihedralCharmmOMP; friend class FixGPU; friend class FixOMP; friend class ThrOMP; public: double eng_vdwl,eng_coul; // accumulated energies double virial[6]; // accumulated virial double *eatom,**vatom; // accumulated per-atom energy/virial double cutforce; // max cutoff for all atom pairs double **cutsq; // cutoff sq for each atom pair int **setflag; // 0/1 = whether each i,j has been set int comm_forward; // size of forward communication (0 if none) int comm_reverse; // size of reverse communication (0 if none) int comm_reverse_off; // size of reverse comm even if newton off int single_enable; // 1 if single() routine exists int restartinfo; // 1 if pair style writes restart info int respa_enable; // 1 if inner/middle/outer rRESPA routines int one_coeff; // 1 if allows only one coeff * * call int no_virial_fdotr_compute; // 1 if does not invoke virial_fdotr_compute() int ghostneigh; // 1 if pair style needs neighbors of ghosts double **cutghost; // cutoff for each ghost pair int tail_flag; // pair_modify flag for LJ tail correction double etail,ptail; // energy/pressure tail corrections double etail_ij,ptail_ij; int evflag; // energy,virial settings int eflag_either,eflag_global,eflag_atom; int vflag_either,vflag_global,vflag_atom; int nextra; // # of extra quantities pair style calculates double *pvector; // vector of extra pair quantities int single_extra; // number of extra single values calculated double *svector; // vector of extra single quantities class NeighList *list; // standard neighbor list used by most pairs class NeighList *listhalf; // half list used by some pairs class NeighList *listfull; // full list used by some pairs class NeighList *listgranhistory; // granular history list used by some pairs class NeighList *listinner; // rRESPA lists used by some pairs class NeighList *listmiddle; class NeighList *listouter; + int compute_flag; // 0 if skip compute() + Pair(class LAMMPS *); virtual ~Pair(); // top-level Pair methods void init(); void reinit(); double mix_energy(double, double, double, double); double mix_distance(double, double); void write_file(int, char **); void init_bitmap(double, double, int, int &, int &, int &, int &); virtual void modify_params(int, char **); + void compute_dummy(int, int); // need to be public, so can be called by pair_style reaxc void ev_tally(int, int, int, int, double, double, double, double, double, double); void ev_tally3(int, int, int, double, double, double *, double *, double *, double *); void v_tally3(int, int, int, double *, double *, double *, double *); void v_tally4(int, int, int, int, double *, double *, double *, double *, double *, double *); void ev_tally_xyz(int, int, int, int, double, double, double, double, double, double, double, double); // general child-class methods virtual void compute(int, int) = 0; virtual void compute_inner() {} virtual void compute_middle() {} virtual void compute_outer(int, int) {} virtual double single(int, int, int, int, double, double, double, double &) {return 0.0;} virtual void settings(int, char **) = 0; virtual void coeff(int, char **) = 0; virtual void init_style(); virtual void init_list(int, class NeighList *); virtual double init_one(int, int) {return 0.0;} virtual void write_restart(FILE *) {} virtual void read_restart(FILE *) {} virtual void write_restart_settings(FILE *) {} virtual void read_restart_settings(FILE *) {} virtual int pack_comm(int, int *, double *, int, int *) {return 0;} virtual void unpack_comm(int, int, double *) {} virtual int pack_reverse_comm(int, int, double *) {return 0;} virtual void unpack_reverse_comm(int, int *, double *) {} virtual double memory_usage(); // specific child-class methods for certain Pair styles virtual void *extract(const char *, int &) {return NULL;} virtual void swap_eam(double *, double **) {} virtual void reset_dt() {} virtual void min_xf_pointers(int, double **, double **) {} virtual void min_xf_get(int) {} virtual void min_x_set(int) {} protected: int allocated; // 0/1 = whether arrays are allocated int suffix_flag; // suffix compatibility flag // pair_modify settings int offset_flag,mix_flag; // flags for offset and mixing int ncoultablebits; // size of Coulomb table double tabinner; // inner cutoff for Coulomb table // custom data type for accessing Coulomb tables typedef union {int i; float f;} union_int_float_t; double THIRD; int vflag_fdotr; int maxeatom,maxvatom; virtual void ev_setup(int, int); void ev_tally_full(int, double, double, double, double, double, double); void ev_tally_xyz_full(int, double, double, double, double, double, double, double, double); void ev_tally4(int, int, int, int, double, double *, double *, double *, double *, double *, double *); void ev_tally_list(int, int *, double, double *); void v_tally2(int, int, double, double *); void v_tally_tensor(int, int, int, int, double, double, double, double, double, double); void virial_fdotr_compute(); inline int sbmask(int j) { return j >> SBBITS & 3; } }; } #endif /* ERROR/WARNING messages: E: Illegal ... command Self-explanatory. Check the input script syntax and compare to the documentation for the command. You can use -echo screen as a command-line option when running LAMMPS to see the offending line. E: Too many total bits for bitmapped lookup table Table size specified via pair_modify command is too large. Note that a value of N generates a 2^N size table. E: Cannot have both pair_modify shift and tail set to yes These 2 options are contradictory. E: Cannot use pair tail corrections with 2d simulations The correction factors are only currently defined for 3d systems. W: Using pair tail corrections with nonperiodic system This is probably a bogus thing to do, since tail corrections are computed by integrating the density of a periodic system out to infinity. E: All pair coeffs are not set All pair coefficients must be set in the data file or by the pair_coeff command before running a simulation. E: Pair style does not support pair_write The pair style does not have a single() function, so it can not be invoked by pair write. E: Invalid atom types in pair_write command Atom types must range from 1 to Ntypes inclusive. E: Invalid style in pair_write command Self-explanatory. Check the input script. E: Invalid cutoffs in pair_write command Inner cutoff must be larger than 0.0 and less than outer cutoff. E: Cannot open pair_write file The specified output file for pair energies and forces cannot be opened. Check that the path and name are correct. E: Bitmapped lookup tables require int/float be same size Cannot use pair tables on this machine, because of word sizes. Use the pair_modify command with table 0 instead. W: Table inner cutoff >= outer cutoff You specified an inner cutoff for a Coulombic table that is longer than the global cutoff. Probably not what you wanted. E: Too many exponent bits for lookup table Table size specified via pair_modify command does not work with your machine's floating point representation. E: Too many mantissa bits for lookup table Table size specified via pair_modify command does not work with your machine's floating point representation. E: Too few bits for lookup table Table size specified via pair_modify command does not work with your machine's floating point representation. */ diff --git a/src/respa.cpp b/src/respa.cpp index 97ad3291d..013b25db6 100644 --- a/src/respa.cpp +++ b/src/respa.cpp @@ -1,699 +1,701 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- Contributing authors: Mark Stevens (SNL), Paul Crozier (SNL) ------------------------------------------------------------------------- */ #include "stdlib.h" #include "string.h" #include "respa.h" #include "neighbor.h" #include "atom.h" #include "domain.h" #include "comm.h" #include "atom.h" #include "force.h" #include "pair.h" #include "bond.h" #include "angle.h" #include "dihedral.h" #include "improper.h" #include "kspace.h" #include "output.h" #include "update.h" #include "modify.h" #include "compute.h" #include "fix_respa.h" #include "timer.h" #include "memory.h" #include "error.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- */ Respa::Respa(LAMMPS *lmp, int narg, char **arg) : Integrate(lmp, narg, arg) { if (narg < 1) error->all(FLERR,"Illegal run_style respa command"); nlevels = atoi(arg[0]); if (nlevels < 1) error->all(FLERR,"Respa levels must be >= 1"); if (narg < nlevels) error->all(FLERR,"Illegal run_style respa command"); loop = new int[nlevels]; for (int iarg = 1; iarg < nlevels; iarg++) { loop[iarg-1] = atoi(arg[iarg]); if (loop[iarg-1] <= 0) error->all(FLERR,"Illegal run_style respa command"); } loop[nlevels-1] = 1; // set level at which each force is computed // argument settings override defaults level_bond = level_angle = level_dihedral = level_improper = -1; level_pair = level_kspace = -1; level_inner = level_middle = level_outer = -1; int iarg = nlevels; while (iarg < narg) { if (strcmp(arg[iarg],"bond") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_bond = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"angle") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_angle = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"dihedral") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_dihedral = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"improper") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_improper = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"pair") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_pair = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"inner") == 0) { if (iarg+4 > narg) error->all(FLERR,"Illegal run_style respa command"); level_inner = atoi(arg[iarg+1]) - 1; cutoff[0] = atof(arg[iarg+2]); cutoff[1] = atof(arg[iarg+3]); iarg += 4; } else if (strcmp(arg[iarg],"middle") == 0) { if (iarg+4 > narg) error->all(FLERR,"Illegal run_style respa command"); level_middle = atoi(arg[iarg+1]) - 1; cutoff[2] = atof(arg[iarg+2]); cutoff[3] = atof(arg[iarg+3]); iarg += 4; } else if (strcmp(arg[iarg],"outer") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_outer = atoi(arg[iarg+1]) - 1; iarg += 2; } else if (strcmp(arg[iarg],"kspace") == 0) { if (iarg+2 > narg) error->all(FLERR,"Illegal run_style respa command"); level_kspace = atoi(arg[iarg+1]) - 1; iarg += 2; } else error->all(FLERR,"Illegal run_style respa command"); } // cannot specify both pair and inner/middle/outer if (level_pair >= 0 && (level_inner >= 0 || level_middle >= 0 || level_outer >= 0)) error->all(FLERR,"Cannot set both respa pair and inner/middle/outer"); // if either inner and outer is specified, then both must be if ((level_inner >= 0 && level_outer == -1) || (level_outer >= 0 && level_inner == -1)) error->all(FLERR,"Must set both respa inner and outer"); // middle cannot be set without inner/outer if (level_middle >= 0 && level_inner == -1) error->all(FLERR,"Cannot set respa middle without inner/outer"); // set defaults if user did not specify level // bond to innermost level // angle same as bond, dihedral same as angle, improper same as dihedral // pair to outermost level if no inner/middle/outer // inner/middle/outer have no defaults // kspace same as pair or outer if (level_bond == -1) level_bond = 0; if (level_angle == -1) level_angle = level_bond; if (level_dihedral == -1) level_dihedral = level_angle; if (level_improper == -1) level_improper = level_dihedral; if (level_pair == -1 && level_inner == -1) level_pair = nlevels-1; if (level_kspace == -1 && level_pair >= 0) level_kspace = level_pair; if (level_kspace == -1 && level_pair == -1) level_kspace = level_outer; // print respa levels if (comm->me == 0) { if (screen) { fprintf(screen,"Respa levels:\n"); for (int i = 0; i < nlevels; i++) { fprintf(screen," %d =",i); if (level_bond == i) fprintf(screen," bond"); if (level_angle == i) fprintf(screen," angle"); if (level_dihedral == i) fprintf(screen," dihedral"); if (level_improper == i) fprintf(screen," improper"); if (level_pair == i) fprintf(screen," pair"); if (level_inner == i) fprintf(screen," pair-inner"); if (level_middle == i) fprintf(screen," pair-middle"); if (level_outer == i) fprintf(screen," pair-outer"); if (level_kspace == i) fprintf(screen," kspace"); fprintf(screen,"\n"); } } if (logfile) { fprintf(logfile,"Respa levels:\n"); for (int i = 0; i < nlevels; i++) { fprintf(logfile," %d =",i); if (level_bond == i) fprintf(logfile," bond"); if (level_angle == i) fprintf(logfile," angle"); if (level_dihedral == i) fprintf(logfile," dihedral"); if (level_improper == i) fprintf(logfile," improper"); if (level_pair == i) fprintf(logfile," pair"); if (level_inner == i) fprintf(logfile," pair-inner"); if (level_middle == i) fprintf(logfile," pair-middle"); if (level_outer == i) fprintf(logfile," pair-outer"); if (level_kspace == i) fprintf(logfile," kspace"); fprintf(logfile,"\n"); } } } // check that levels are in correct order if (level_angle < level_bond || level_dihedral < level_angle || level_improper < level_dihedral) error->all(FLERR,"Invalid order of forces within respa levels"); if (level_pair >= 0) { if (level_pair < level_improper || level_kspace < level_pair) error->all(FLERR,"Invalid order of forces within respa levels"); } if (level_pair == -1 && level_middle == -1) { if (level_inner < level_improper || level_outer < level_inner || level_kspace != level_outer) error->all(FLERR,"Invalid order of forces within respa levels"); } if (level_pair == -1 && level_middle >= 0) { if (level_inner < level_improper || level_middle < level_inner || level_outer < level_inner || level_kspace != level_outer) error->all(FLERR,"Invalid order of forces within respa levels"); } // warn if any levels are devoid of forces int flag = 0; for (int i = 0; i < nlevels; i++) if (level_bond != i && level_angle != i && level_dihedral != i && level_improper != i && level_pair != i && level_inner != i && level_middle != i && level_outer != i && level_kspace != i) flag = 1; if (flag && comm->me == 0) error->warning(FLERR,"One or more respa levels compute no forces"); // check cutoff consistency if inner/middle/outer are enabled if (level_inner >= 0 && cutoff[1] < cutoff[0]) error->all(FLERR,"Respa inner cutoffs are invalid"); if (level_middle >= 0 && (cutoff[3] < cutoff[2] || cutoff[2] < cutoff[1])) error->all(FLERR,"Respa middle cutoffs are invalid"); // set outer pair of cutoffs to inner pair if middle is not enabled if (level_inner >= 0 && level_middle < 0) { cutoff[2] = cutoff[0]; cutoff[3] = cutoff[1]; } // allocate other needed arrays newton = new int[nlevels]; step = new double[nlevels]; } /* ---------------------------------------------------------------------- */ Respa::~Respa() { delete [] loop; delete [] newton; delete [] step; } /* ---------------------------------------------------------------------- initialization before run ------------------------------------------------------------------------- */ void Respa::init() { + Integrate::init(); + // warn if no fixes if (modify->nfix == 0 && comm->me == 0) error->warning(FLERR,"No fixes defined, atoms won't move"); // create fix needed for storing atom-based respa level forces // will delete it at end of run char **fixarg = new char*[4]; fixarg[0] = (char *) "RESPA"; fixarg[1] = (char *) "all"; fixarg[2] = (char *) "RESPA"; fixarg[3] = new char[8]; sprintf(fixarg[3],"%d",nlevels); modify->add_fix(4,fixarg); delete [] fixarg[3]; delete [] fixarg; fix_respa = (FixRespa *) modify->fix[modify->nfix-1]; // insure respa inner/middle/outer is using Pair class that supports it if (level_inner >= 0) if (force->pair && force->pair->respa_enable == 0) error->all(FLERR,"Pair style does not support rRESPA inner/middle/outer"); // virial_style = 1 (explicit) since never computed implicitly like Verlet virial_style = 1; // setup lists of computes for global and per-atom PE and pressure ev_setup(); // detect if fix omp is present and will clear force arrays for us int ifix = modify->find_fix("package_omp"); if (ifix >= 0) external_force_clear = 1; // set flags for what arrays to clear in force_clear() // need to clear additionals arrays if they exist torqueflag = 0; if (atom->torque_flag) torqueflag = 1; erforceflag = 0; if (atom->erforce_flag) erforceflag = 1; e_flag = 0; if (atom->e_flag) e_flag = 1; rho_flag = 0; if (atom->rho_flag) rho_flag = 1; // step[] = timestep for each level step[nlevels-1] = update->dt; for (int ilevel = nlevels-2; ilevel >= 0; ilevel--) step[ilevel] = step[ilevel+1]/loop[ilevel]; // set newton flag for each level for (int ilevel = 0; ilevel < nlevels; ilevel++) { newton[ilevel] = 0; if (force->newton_bond) { if (level_bond == ilevel || level_angle == ilevel || level_dihedral == ilevel || level_improper == ilevel) newton[ilevel] = 1; } if (force->newton_pair) { if (level_pair == ilevel || level_inner == ilevel || level_middle == ilevel || level_outer == ilevel) newton[ilevel] = 1; } } // orthogonal vs triclinic simulation box triclinic = domain->triclinic; } /* ---------------------------------------------------------------------- setup before run ------------------------------------------------------------------------- */ void Respa::setup() { if (comm->me == 0 && screen) fprintf(screen,"Setting up run ...\n"); update->setupflag = 1; // setup domain, communication and neighboring // acquire ghosts // build neighbor lists atom->setup(); modify->setup_pre_exchange(); if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); if (atom->sortfreq > 0) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; // compute all forces ev_set(update->ntimestep); for (int ilevel = 0; ilevel < nlevels; ilevel++) { force_clear(newton[ilevel]); modify->setup_pre_force_respa(vflag,ilevel); if (level_bond == ilevel && force->bond) force->bond->compute(eflag,vflag); if (level_angle == ilevel && force->angle) force->angle->compute(eflag,vflag); if (level_dihedral == ilevel && force->dihedral) force->dihedral->compute(eflag,vflag); if (level_improper == ilevel && force->improper) force->improper->compute(eflag,vflag); - if (level_pair == ilevel && force->pair) + if (level_pair == ilevel && pair_compute_flag) force->pair->compute(eflag,vflag); - if (level_inner == ilevel && force->pair) + if (level_inner == ilevel && pair_compute_flag) force->pair->compute_inner(); - if (level_middle == ilevel && force->pair) + if (level_middle == ilevel && pair_compute_flag) force->pair->compute_middle(); - if (level_outer == ilevel && force->pair) + if (level_outer == ilevel && pair_compute_flag) force->pair->compute_outer(eflag,vflag); if (level_kspace == ilevel && force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); } if (newton[ilevel]) comm->reverse_comm(); copy_f_flevel(ilevel); } modify->setup(vflag); sum_flevel_f(); output->setup(1); update->setupflag = 0; } /* ---------------------------------------------------------------------- setup without output flag = 0 = just force calculation flag = 1 = reneighbor and force calculation ------------------------------------------------------------------------- */ void Respa::setup_minimal(int flag) { update->setupflag = 1; // setup domain, communication and neighboring // acquire ghosts // build neighbor lists if (flag) { if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; } // compute all forces ev_set(update->ntimestep); for (int ilevel = 0; ilevel < nlevels; ilevel++) { force_clear(newton[ilevel]); modify->setup_pre_force_respa(vflag,ilevel); if (level_bond == ilevel && force->bond) force->bond->compute(eflag,vflag); if (level_angle == ilevel && force->angle) force->angle->compute(eflag,vflag); if (level_dihedral == ilevel && force->dihedral) force->dihedral->compute(eflag,vflag); if (level_improper == ilevel && force->improper) force->improper->compute(eflag,vflag); - if (level_pair == ilevel && force->pair) + if (level_pair == ilevel && pair_compute_flag) force->pair->compute(eflag,vflag); - if (level_inner == ilevel && force->pair) + if (level_inner == ilevel && pair_compute_flag) force->pair->compute_inner(); - if (level_middle == ilevel && force->pair) + if (level_middle == ilevel && pair_compute_flag) force->pair->compute_middle(); - if (level_outer == ilevel && force->pair) + if (level_outer == ilevel && pair_compute_flag) force->pair->compute_outer(eflag,vflag); if (level_kspace == ilevel && force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); } if (newton[ilevel]) comm->reverse_comm(); copy_f_flevel(ilevel); } modify->setup(vflag); sum_flevel_f(); update->setupflag = 0; } /* ---------------------------------------------------------------------- run for N steps ------------------------------------------------------------------------- */ void Respa::run(int n) { bigint ntimestep; for (int i = 0; i < n; i++) { ntimestep = ++update->ntimestep; ev_set(ntimestep); recurse(nlevels-1); if (modify->n_end_of_step) { timer->stamp(); modify->end_of_step(); timer->stamp(Timer::MODIFY); } if (ntimestep == output->next) { timer->stamp(); sum_flevel_f(); output->write(update->ntimestep); timer->stamp(Timer::OUTPUT); } } } /* ---------------------------------------------------------------------- delete rRESPA fix at end of run, so its atom arrays won't persist ------------------------------------------------------------------------- */ void Respa::cleanup() { modify->post_run(); modify->delete_fix("RESPA"); } /* ---------------------------------------------------------------------- */ void Respa::reset_dt() { step[nlevels-1] = update->dt; for (int ilevel = nlevels-2; ilevel >= 0; ilevel--) step[ilevel] = step[ilevel+1]/loop[ilevel]; } /* ---------------------------------------------------------------------- */ void Respa::recurse(int ilevel) { copy_flevel_f(ilevel); for (int iloop = 0; iloop < loop[ilevel]; iloop++) { timer->stamp(); modify->initial_integrate_respa(vflag,ilevel,iloop); if (modify->n_post_integrate_respa) modify->post_integrate_respa(ilevel,iloop); timer->stamp(Timer::MODIFY); if (ilevel) recurse(ilevel-1); // at outermost level, check on rebuilding neighbor list // at innermost level, communicate // at middle levels, do nothing if (ilevel == nlevels-1) { int nflag = neighbor->decide(); if (nflag) { if (modify->n_pre_exchange) { timer->stamp(); modify->pre_exchange(); timer->stamp(Timer::MODIFY); } if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); if (domain->box_change) { domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); } timer->stamp(); comm->exchange(); if (atom->sortfreq > 0 && update->ntimestep >= atom->nextsort) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); timer->stamp(Timer::COMM); if (modify->n_pre_neighbor) { modify->pre_neighbor(); timer->stamp(Timer::MODIFY); } neighbor->build(); timer->stamp(Timer::NEIGHBOR); } } else if (ilevel == 0) { timer->stamp(); comm->forward_comm(); timer->stamp(Timer::COMM); } force_clear(newton[ilevel]); if (modify->n_pre_force_respa) { timer->stamp(); modify->pre_force_respa(vflag,ilevel,iloop); timer->stamp(Timer::MODIFY); } timer->stamp(); if (level_bond == ilevel && force->bond) { force->bond->compute(eflag,vflag); timer->stamp(Timer::BOND); } if (level_angle == ilevel && force->angle) { force->angle->compute(eflag,vflag); timer->stamp(Timer::BOND); } if (level_dihedral == ilevel && force->dihedral) { force->dihedral->compute(eflag,vflag); timer->stamp(Timer::BOND); } if (level_improper == ilevel && force->improper) { force->improper->compute(eflag,vflag); timer->stamp(Timer::BOND); } - if (level_pair == ilevel && force->pair) { + if (level_pair == ilevel && pair_compute_flag) { force->pair->compute(eflag,vflag); timer->stamp(Timer::PAIR); } - if (level_inner == ilevel && force->pair) { + if (level_inner == ilevel && pair_compute_flag) { force->pair->compute_inner(); timer->stamp(Timer::PAIR); } - if (level_middle == ilevel && force->pair) { + if (level_middle == ilevel && pair_compute_flag) { force->pair->compute_middle(); timer->stamp(Timer::PAIR); } - if (level_outer == ilevel && force->pair) { + if (level_outer == ilevel && pair_compute_flag) { force->pair->compute_outer(eflag,vflag); timer->stamp(Timer::PAIR); } - if (level_kspace == ilevel && force->kspace) { + if (level_kspace == ilevel && kspace_compute_flag) { force->kspace->compute(eflag,vflag); timer->stamp(Timer::KSPACE); } if (newton[ilevel]) { comm->reverse_comm(); timer->stamp(Timer::COMM); } timer->stamp(); if (modify->n_post_force_respa) modify->post_force_respa(vflag,ilevel,iloop); modify->final_integrate_respa(ilevel,iloop); timer->stamp(Timer::MODIFY); } copy_f_flevel(ilevel); } /* ---------------------------------------------------------------------- clear force on own & ghost atoms ------------------------------------------------------------------------- */ void Respa::force_clear(int newtonflag) { if (external_force_clear) return; int i; if (external_force_clear) return; // clear global force array // nall includes ghosts only if newton flag is set int nall; if (newtonflag) nall = atom->nlocal + atom->nghost; else nall = atom->nlocal; size_t nbytes = sizeof(double) * nall; if (nbytes > 0 ) { memset(&(atom->f[0][0]),0,3*nbytes); if (torqueflag) memset(&(atom->torque[0][0]),0,3*nbytes); if (erforceflag) memset(&(atom->erforce[0]), 0, nbytes); if (e_flag) memset(&(atom->de[0]), 0, nbytes); if (rho_flag) memset(&(atom->drho[0]), 0, nbytes); } } /* ---------------------------------------------------------------------- copy force components from atom->f to FixRespa->f_level ------------------------------------------------------------------------- */ void Respa::copy_f_flevel(int ilevel) { double ***f_level = fix_respa->f_level; double **f = atom->f; int n = atom->nlocal; for (int i = 0; i < n; i++) { f_level[i][ilevel][0] = f[i][0]; f_level[i][ilevel][1] = f[i][1]; f_level[i][ilevel][2] = f[i][2]; } } /* ---------------------------------------------------------------------- copy force components from FixRespa->f_level to atom->f ------------------------------------------------------------------------- */ void Respa::copy_flevel_f(int ilevel) { double ***f_level = fix_respa->f_level; double **f = atom->f; int n = atom->nlocal; for (int i = 0; i < n; i++) { f[i][0] = f_level[i][ilevel][0]; f[i][1] = f_level[i][ilevel][1]; f[i][2] = f_level[i][ilevel][2]; } } /* ---------------------------------------------------------------------- sum all force components from FixRespa->f_level to create full atom->f ------------------------------------------------------------------------- */ void Respa::sum_flevel_f() { copy_flevel_f(0); double ***f_level = fix_respa->f_level; double **f = atom->f; int n = atom->nlocal; for (int ilevel = 1; ilevel < nlevels; ilevel++) { for (int i = 0; i < n; i++) { f[i][0] += f_level[i][ilevel][0]; f[i][1] += f_level[i][ilevel][1]; f[i][2] += f_level[i][ilevel][2]; } } } diff --git a/src/verlet.cpp b/src/verlet.cpp index 82b661283..8632b0fc2 100644 --- a/src/verlet.cpp +++ b/src/verlet.cpp @@ -1,425 +1,431 @@ /* ---------------------------------------------------------------------- LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator http://lammps.sandia.gov, Sandia National Laboratories Steve Plimpton, sjplimp@sandia.gov Copyright (2003) Sandia Corporation. Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains certain rights in this software. This software is distributed under the GNU General Public License. See the README file in the top-level LAMMPS directory. ------------------------------------------------------------------------- */ #include "string.h" #include "verlet.h" #include "neighbor.h" #include "domain.h" #include "comm.h" #include "atom.h" #include "force.h" #include "pair.h" #include "bond.h" #include "angle.h" #include "dihedral.h" #include "improper.h" #include "kspace.h" #include "output.h" #include "update.h" #include "modify.h" #include "compute.h" #include "fix.h" #include "timer.h" #include "memory.h" #include "error.h" using namespace LAMMPS_NS; /* ---------------------------------------------------------------------- */ Verlet::Verlet(LAMMPS *lmp, int narg, char **arg) : Integrate(lmp, narg, arg) {} /* ---------------------------------------------------------------------- initialization before run ------------------------------------------------------------------------- */ void Verlet::init() { + Integrate::init(); + // warn if no fixes if (modify->nfix == 0 && comm->me == 0) error->warning(FLERR,"No fixes defined, atoms won't move"); // virial_style: // 1 if computed explicitly by pair->compute via sum over pair interactions // 2 if computed implicitly by pair->virial_fdotr_compute via sum over ghosts if (force->newton_pair) virial_style = 2; else virial_style = 1; // setup lists of computes for global and per-atom PE and pressure ev_setup(); // detect if fix omp is present for clearing force arrays int ifix = modify->find_fix("package_omp"); if (ifix >= 0) external_force_clear = 1; // set flags for what arrays to clear in force_clear() // need to clear additionals arrays if they exist torqueflag = 0; if (atom->torque_flag) torqueflag = 1; erforceflag = 0; if (atom->erforce_flag) erforceflag = 1; e_flag = 0; if (atom->e_flag) e_flag = 1; rho_flag = 0; if (atom->rho_flag) rho_flag = 1; // orthogonal vs triclinic simulation box triclinic = domain->triclinic; } /* ---------------------------------------------------------------------- setup before run ------------------------------------------------------------------------- */ void Verlet::setup() { if (comm->me == 0 && screen) fprintf(screen,"Setting up run ...\n"); update->setupflag = 1; // setup domain, communication and neighboring // acquire ghosts // build neighbor lists atom->setup(); modify->setup_pre_exchange(); if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); if (atom->sortfreq > 0) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; // compute all forces ev_set(update->ntimestep); force_clear(); modify->setup_pre_force(vflag); - if (force->pair) force->pair->compute(eflag,vflag); + if (pair_compute_flag) force->pair->compute(eflag,vflag); + else if (force->pair) force->pair->compute_dummy(eflag,vflag); if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); } if (force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); + else force->kspace->compute_dummy(eflag,vflag); } if (force->newton) comm->reverse_comm(); modify->setup(vflag); output->setup(1); update->setupflag = 0; } /* ---------------------------------------------------------------------- setup without output flag = 0 = just force calculation flag = 1 = reneighbor and force calculation ------------------------------------------------------------------------- */ void Verlet::setup_minimal(int flag) { update->setupflag = 1; // setup domain, communication and neighboring // acquire ghosts // build neighbor lists if (flag) { if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); comm->exchange(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); neighbor->build(); neighbor->ncalls = 0; } // compute all forces ev_set(update->ntimestep); force_clear(); modify->setup_pre_force(vflag); - if (force->pair) force->pair->compute(eflag,vflag); + if (pair_compute_flag) force->pair->compute(eflag,vflag); + else if (force->pair) force->pair->compute_dummy(eflag,vflag); if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); } if (force->kspace) { force->kspace->setup(); - force->kspace->compute(eflag,vflag); + if (kspace_compute_flag) force->kspace->compute(eflag,vflag); + else force->kspace->compute_dummy(eflag,vflag); } if (force->newton) comm->reverse_comm(); modify->setup(vflag); update->setupflag = 0; } /* ---------------------------------------------------------------------- run for N steps ------------------------------------------------------------------------- */ void Verlet::run(int n) { bigint ntimestep; int nflag,sortflag; int n_post_integrate = modify->n_post_integrate; int n_pre_exchange = modify->n_pre_exchange; int n_pre_neighbor = modify->n_pre_neighbor; int n_pre_force = modify->n_pre_force; int n_post_force = modify->n_post_force; int n_end_of_step = modify->n_end_of_step; if (atom->sortfreq > 0) sortflag = 1; else sortflag = 0; for (int i = 0; i < n; i++) { ntimestep = ++update->ntimestep; ev_set(ntimestep); // initial time integration timer->stamp(); modify->initial_integrate(vflag); if (n_post_integrate) modify->post_integrate(); timer->stamp(Timer::MODIFY); // regular communication vs neighbor list rebuild nflag = neighbor->decide(); if (nflag == 0) { timer->stamp(); comm->forward_comm(); timer->stamp(Timer::COMM); } else { if (n_pre_exchange) { timer->stamp(); modify->pre_exchange(); timer->stamp(Timer::MODIFY); } if (triclinic) domain->x2lamda(atom->nlocal); domain->pbc(); if (domain->box_change) { domain->reset_box(); comm->setup(); if (neighbor->style) neighbor->setup_bins(); } timer->stamp(); comm->exchange(); if (sortflag && ntimestep >= atom->nextsort) atom->sort(); comm->borders(); if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost); timer->stamp(Timer::COMM); if (n_pre_neighbor) { modify->pre_neighbor(); timer->stamp(Timer::MODIFY); } neighbor->build(); timer->stamp(Timer::NEIGHBOR); } // force computations force_clear(); timer->stamp(); if (n_pre_force) { modify->pre_force(vflag); timer->stamp(Timer::MODIFY); } - if (force->pair) { + if (pair_compute_flag) { force->pair->compute(eflag,vflag); timer->stamp(Timer::PAIR); } if (atom->molecular) { if (force->bond) force->bond->compute(eflag,vflag); if (force->angle) force->angle->compute(eflag,vflag); if (force->dihedral) force->dihedral->compute(eflag,vflag); if (force->improper) force->improper->compute(eflag,vflag); timer->stamp(Timer::BOND); } - if (force->kspace) { + if (kspace_compute_flag) { force->kspace->compute(eflag,vflag); timer->stamp(Timer::KSPACE); } // reverse communication of forces if (force->newton) { comm->reverse_comm(); timer->stamp(Timer::COMM); } // force modifications, final time integration, diagnostics if (n_post_force) modify->post_force(vflag); modify->final_integrate(); if (n_end_of_step) modify->end_of_step(); timer->stamp(Timer::MODIFY); // all output if (ntimestep == output->next) { timer->stamp(); output->write(ntimestep); timer->stamp(Timer::OUTPUT); } } } /* ---------------------------------------------------------------------- */ void Verlet::cleanup() { modify->post_run(); } /* ---------------------------------------------------------------------- clear force on own & ghost atoms setup and clear other arrays as needed ------------------------------------------------------------------------- */ void Verlet::force_clear() { if (external_force_clear) return; int i; if (external_force_clear) return; // clear force on all particles // if either newton flag is set, also include ghosts // when using threads always clear all forces. if (neighbor->includegroup == 0) { int nall; if (force->newton) nall = atom->nlocal + atom->nghost; else nall = atom->nlocal; size_t nbytes = sizeof(double) * nall; if (nbytes > 0) { memset(&(atom->f[0][0]),0,3*nbytes); if (torqueflag) memset(&(atom->torque[0][0]),0,3*nbytes); if (erforceflag) memset(&(atom->erforce[0]), 0, nbytes); if (e_flag) memset(&(atom->de[0]), 0, nbytes); if (rho_flag) memset(&(atom->drho[0]), 0, nbytes); } // neighbor includegroup flag is set // clear force only on initial nfirst particles // if either newton flag is set, also include ghosts } else { int nall = atom->nfirst; double **f = atom->f; for (i = 0; i < nall; i++) { f[i][0] = 0.0; f[i][1] = 0.0; f[i][2] = 0.0; } if (torqueflag) { double **torque = atom->torque; for (i = 0; i < nall; i++) { torque[i][0] = 0.0; torque[i][1] = 0.0; torque[i][2] = 0.0; } } if (erforceflag) { double *erforce = atom->erforce; for (i = 0; i < nall; i++) erforce[i] = 0.0; } if (e_flag) { double *de = atom->de; for (i = 0; i < nall; i++) de[i] = 0.0; } if (rho_flag) { double *drho = atom->drho; for (i = 0; i < nall; i++) drho[i] = 0.0; } if (force->newton) { nall = atom->nlocal + atom->nghost; for (i = atom->nlocal; i < nall; i++) { f[i][0] = 0.0; f[i][1] = 0.0; f[i][2] = 0.0; } if (torqueflag) { double **torque = atom->torque; for (i = atom->nlocal; i < nall; i++) { torque[i][0] = 0.0; torque[i][1] = 0.0; torque[i][2] = 0.0; } } if (erforceflag) { double *erforce = atom->erforce; for (i = atom->nlocal; i < nall; i++) erforce[i] = 0.0; } if (e_flag) { double *de = atom->de; for (i = 0; i < nall; i++) de[i] = 0.0; } if (rho_flag) { double *drho = atom->drho; for (i = 0; i < nall; i++) drho[i] = 0.0; } } } } diff --git a/src/version.h b/src/version.h index 4412c3a44..05cb579ab 100644 --- a/src/version.h +++ b/src/version.h @@ -1 +1 @@ -#define LAMMPS_VERSION "10 Apr 2012" +#define LAMMPS_VERSION "12 Apr 2012"