ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Updating documentation/examples, patching comm_modify cutoff/multi command

Browse Source
This commit is contained in:
Joel Clemmer
2021-02-02 14:50:30 -07:00
parent 0676c953c0
commit 852e4efc6f
9 changed files with 113 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
doc/src/comm_modify.rst
View File

@ -15,12 +15,15 @@ Syntax
.. parsed-literal::
*mode* value = *single* or *multi* = communicate atoms within a single or multiple distances
*mode* value = *single*, *multi*, or *multi/old* = communicate atoms within a single or multiple distances
*cutoff* value = Rcut (distance units) = communicate atoms from this far away
*cutoff/multi* type value
*cutoff/multi* collection value
collection = atom collection or collection range (supports asterisk notation)
value = Rcut (distance units) = communicate atoms for selected types from this far away
*reduce/multi* arg = none = reduce number of communicated ghost atoms for multi style
*cutoff/multi/old* type value
type = atom type or type range (supports asterisk notation)
value = Rcut (distance units) = communicate atoms for selected types from this far away
*multi/reduce* arg = none = reduce number of communicated ghost atoms for multi style
*group* value = group-ID = only communicate atoms in the group
*vel* value = *yes* or *no* = do or do not communicate velocity info with ghost atoms
@ -29,7 +32,7 @@ Examples
.. code-block:: LAMMPS
comm_modify mode multi
comm_modify mode multi reduce/multi
comm_modify mode multi group solvent
comm_modift mode multi cutoff/multi 1 10.0 cutoff/multi 2*4 15.0
comm_modify vel yes
@ -63,13 +66,18 @@ sub-domain. The distance is by default the maximum of the neighbor
cutoff across all atom type pairs.
For many systems this is an efficient algorithm, but for systems with
widely varying cutoffs for different type pairs, the *multi* mode can
be faster. In this case, each atom type is assigned its own distance
widely varying cutoffs for different type pairs, the *multi* or *multi/old* mode can
be faster. In *multi*, each atom is assigned to a collection which should
correspond to a set of atoms with similar interaction cutoffs.
In this case, each atom collection is assigned its own distance
cutoff for communication purposes, and fewer atoms will be
communicated. See the :doc:`neighbor multi <neighbor>` command for a
neighbor list construction option that may also be beneficial for
simulations of this kind. The *multi* mode is compatable with both the
*multi* and *multi/old* neighbor styles.
communicated. in *multi/old*, a similar technique is used but atoms
are grouped by atom type. See the :doc:`neighbor multi <neighbor>` and
:doc:`neighbor multi/old <neighbor>` commands for
neighbor list construction options that may also be beneficial for
simulations of this kind. The *multi* communiction mode is only compatable
with the *multi* neighbor style. The *multi/old* communication mode is compatble
with both the *multi* and *multi/old* neighbor styles.
The *cutoff* keyword allows you to extend the ghost cutoff distance
for communication mode *single*\ , which is the distance from the borders
@ -89,18 +97,24 @@ warning is printed, if this bond based estimate is larger than the
communication cutoff used.
The *cutoff/multi* option is equivalent to *cutoff*\ , but applies to
communication mode *multi* instead. Since in this case the communication
cutoffs are determined per atom type, a type specifier is needed and
cutoff for one or multiple types can be extended. Also ranges of types
using the usual asterisk notation can be given. For granular pair styles,
the default cutoff is set to the sum of the current maximum atomic radii
for each type.
communication mode *multi* instead. Since the communication
cutoffs are determined per atom collections, a collection specifier is needed and
cutoff for one or multiple collections can be extended. Also ranges of collections
using the usual asterisk notation can be given.
Note that the arguments for *cutoff/multi* are parsed right before each
simulation to account for potential changes in the number of collections.
Custom cutoffs are preserved between runs but if collections are redefined,
one may want to respecify communication cutoffs.
For granular pair styles,the default cutoff is set to the sum of the
current maximum atomic radii for each collection.
The *cutoff/multi/old* option is similar to *cutoff/multi* except it
operates on atom types as opposed to collections.
The *multi/reduce* option applies to *multi* and sets communication
cutoffs for different sized particles based on the largest interaction distance
between two particles in the same multi grouping. This reduces the number of
The *reduce/multi* option applies to *multi* and sets the communication
cutoff for a particle equal to the maximum interaction distance between particles
in the same collection. This reduces the number of
ghost atoms that need to be communicated. This method is only compatible with the
*multi* neighbor style and requires only half neighbor lists and Newton on.
*multi* neighbor style and requires a half neighbor list and Newton on.
See the :doc:`neighbor multi <neighbor>` command for more information.
These are simulation scenarios in which it may be useful or even

38
doc/src/neigh_modify.rst
View File

@ -14,7 +14,7 @@ Syntax
.. parsed-literal::
keyword = *delay* or *every* or *check* or *once* or *cluster* or *include* or *exclude* or *page* or *one* or *binsize* or *multi/custom*
keyword = *delay* or *every* or *check* or *once* or *cluster* or *include* or *exclude* or *page* or *one* or *binsize* or *collection/type* or *collection/interval*
*delay* value = N
N = delay building until this many steps since last build
*every* value = M
@ -47,9 +47,12 @@ Syntax
N = max number of neighbors of one atom
*binsize* value = size
size = bin size for neighbor list construction (distance units)
*multi/custom* values = N arg1 ... argN
N = number of custom groups
arg = N separate types or ranges of types (see below)
*collection/type* values = N arg1 ... argN
N = number of custom collections
arg = N separate lists of types (see below)
*collection/interval* values = N arg1 ... argN
N = number of custom collections
arg = N separate cutoffs for intervals (see below)
Examples
""""""""
@ -61,7 +64,8 @@ Examples
neigh_modify exclude group frozen frozen check no
neigh_modify exclude group residue1 chain3
neigh_modify exclude molecule/intra rigid
neigh_modify multi/custom 2 1*2 3*4
neigh_modify collection/type 2 1*2,5 3*4
neigh_modify collection/interval 2 1.0 10.0
Description
"""""""""""
@ -192,8 +196,9 @@ atom can have.
The *binsize* option allows you to specify what size of bins will be
used in neighbor list construction to sort and find neighboring atoms.
By default, for :doc:`neighbor style bin <neighbor>`, LAMMPS uses bins
that are 1/2 the size of the maximum pair cutoff. For :doc:`neighbor style multi <neighbor>`, the bins are 1/2 the size of the minimum pair
cutoff. Typically these are good values for minimizing the time for
that are 1/2 the size of the maximum pair cutoff. For :doc:`neighbor style multi <neighbor>`,
the bins are 1/2 the size of the collection interaction cutoff.
Typically these are good values for minimizing the time for
neighbor list construction. This setting overrides the default.
If you make it too big, there is little overhead due to
looping over bins, but more atoms are checked. If you make it too
@ -201,19 +206,28 @@ small, the optimal number of atoms is checked, but bin overhead goes
up. If you set the binsize to 0.0, LAMMPS will use the default
binsize of 1/2 the cutoff.
The *multi/custom* option allows you to define custom groups of atom
The *collection/type* option allows you to define custom collections of atom
types for the *multi* neighbor mode. By grouping atom types with
similar cutoffs, one may be able to improve performance by reducing
overhead. You must first specify the number of custom groups N to be
defined followed by N ranges of types. The range can be specified as a
overhead. You must first specify the number of custom collections N to be
defined followed by N lists of types. Each list consists of a series of type
ranges separated by commas. The range can be specified as a
single numeric value, or a wildcard asterisk can be used to specify a range
of values. This takes the form "\*" or "\*n" or "n\*" or "m\*n". For
example, if M = the number of atom types, then an asterisk with no numeric
values means all types from 1 to M. A leading asterisk means all types
from 1 to n (inclusive). A trailing asterisk means all types from n to M
(inclusive). A middle asterisk means all types from m to n (inclusive).
Note that any atom types not included in a custom group will be automatically
placed within a separate group.
Note that all atom types must be included in a custom collection.
The *collection/interval* option provides a similar capability.
This command allows a user to define custom collections by specifying a
series of cutoff intervals. LAMMPS will automatically sort atoms into these intervals
based on their type-dependent cutoffs or their finite size.
You must first specify the number of custom collections N to be
defined followed by N values representing the upper cutoff of each interval.
This command is particularly useful for granular pairstyles where the interaction distance
of particles depends on their radius and may not depend on their atom type.
Restrictions
""""""""""""

6
doc/src/neighbor.rst
View File

@ -56,13 +56,13 @@ the largest cutoff distance between any pair of atom types and a
single set of bins is defined to search over for all atom types. This
can be inefficient if one pair of types has a very long cutoff, but
other type pairs have a much shorter cutoff. The *multi* style uses
different sized bins for groups of different sized particles. Different
different sized bins for collections of different sized particles. Different
sets of bins are then used to construct the neighbor lists as as further
described by Shire, Hanley, and Stratford :ref:`(Shire) <bytype-Shire>`.
This imposes some extra setup overhead, but the searches themselves
may be much faster. By default, separate groups of particles are defined
may be much faster. By default, separate collections of particles are defined
for each atom type. For systems with two or more types with similar
cutoffs, one can reduce the extra overhead by defining custom groupings
cutoffs, one can reduce the extra overhead by defining custom collections
using the :doc:`neigh_modify <neigh_modify>` command. See the
:doc:`comm_modify mode bytype <comm_modify>` command for compatible
communication options that may be beneficial for simulations of this kind.

4
examples/multi/in.colloid
View File

@ -30,8 +30,8 @@ velocity all create 1.44 87287 loop geom
# multi neighbor and comm for efficiency
neighbor 1 multi #multi/old
neigh_modify delay 0 multi/custom 2 1*4 5
comm_modify mode multi multi/reduce
neigh_modify delay 0 collection/type 2 1*4 5
comm_modify mode multi reduce/multi
# colloid potential

2
examples/multi/in.granular
View File

@ -29,7 +29,7 @@ velocity all create 1.44 87287 loop geom
neighbor 1 multi #multi/old
neigh_modify delay 0
comm_modify mode multi vel yes multi/reduce
comm_modify mode multi vel yes reduce/multi
# colloid potential

18
src/comm.cpp
View File

@ -35,7 +35,7 @@
#include "update.h"
#include <cstring>
#include <vector>
#ifdef _OPENMP
#include <omp.h>
#endif
@ -59,6 +59,7 @@ Comm::Comm(LAMMPS *lmp) : Pointers(lmp)
cutghostuser = 0.0;
cutusermulti = nullptr;
cutusermultiold = nullptr;
ncollections_prior = 0;
ghost_velocity = 0;
user_procgrid[0] = user_procgrid[1] = user_procgrid[2] = 0;
@ -342,21 +343,14 @@ void Comm::modify_params(int narg, char **arg)
if (domain->box_exist == 0)
error->all(FLERR,
"Cannot set cutoff/multi before simulation box is defined");
ncollections = neighbor->ncollections;
if (iarg+3 > narg)
error->all(FLERR,"Illegal comm_modify command");
if (cutusermulti == nullptr) {
memory->create(cutusermulti,ncollections,"comm:cutusermulti");
for (i=0; i < ncollections; ++i)
cutusermulti[i] = -1.0;
}
utils::bounds(FLERR,arg[iarg+1],1,ncollections,nlo,nhi,error);
// save arguments so they can be parsed in comm->setup()
// ncollections can be changed by neigh_modify commands
cut = utils::numeric(FLERR,arg[iarg+2],false,lmp);
cutghostuser = MAX(cutghostuser,cut);
if (cut < 0.0)
error->all(FLERR,"Invalid cutoff in comm_modify command");
for (i=nlo; i<=nhi; ++i)
cutusermulti[i] = cut;
usermultiargs.emplace_back(arg[iarg+1], cut);
iarg += 3;
} else if (strcmp(arg[iarg],"cutoff/multi/old") == 0) {
int i,nlo,nhi;

2
src/comm.h
View File

@ -33,6 +33,8 @@ class Comm : protected Pointers {
double cutghost[3]; // cutoffs used for acquiring ghost atoms
double cutghostuser; // user-specified ghost cutoff (mode == 0)
double *cutusermulti; // per collection user ghost cutoff (mode == 1)
std::vector<std::pair<std::string, double>> usermultiargs;
// collection args for custom ghost cutoffs
double *cutusermultiold; // per type user ghost cutoff (mode == 2)
int recv_from_partition; // recv proc layout from this partition
int send_to_partition; // send my proc layout to this partition

22
src/comm_brick.cpp
View File

@ -30,6 +30,7 @@
#include <cmath>
#include <cstring>
#include <vector>
using namespace LAMMPS_NS;
@ -195,10 +196,7 @@ void CommBrick::setup()
if (mode == Comm::MULTI) {
// build initial collection array
neighbor->build_collection(0);
if(cutusermulti and ncollections != neighbor->ncollections)
error->all(FLERR, "Cannot change number of collections after defining comm_modify multi/cutoff");
else ncollections = neighbor->ncollections;
ncollections = neighbor->ncollections;
// reallocate memory for multi-style communication at setup if ncollections change
if(ncollections_prior != ncollections){
@ -207,9 +205,25 @@ void CommBrick::setup()
allocate_multi(maxswap);
memory->create(cutghostmulti,ncollections,3,"comm:cutghostmulti");
memory->grow(cutusermulti,ncollections,"comm:cutusermulti");
for(i = ncollections_prior; i < ncollections; i++)
cutusermulti[i] = -1.0;
ncollections_prior = ncollections;
}
// parse any cutoff/multi commands
int nhi, nlo;
for(auto it = usermultiargs.begin(); it != usermultiargs.end(); it ++) {
utils::bounds(FLERR,it->first,1,ncollections,nlo,nhi,error);
if(nhi >= ncollections)
error->all(FLERR, "Unused collection id in comm_modify cutoff/multi command");
for (j=nlo; j<=nhi; ++j)
cutusermulti[j] = it->second;
}
usermultiargs.clear();
double **cutcollectionsq = neighbor->cutcollectionsq;
// If using multi/reduce, communicate particles a distance equal
// to the max cutoff with equally sized or smaller collections

21
src/comm_tiled.cpp
View File

@ -31,6 +31,7 @@
#include <cmath>
#include <cstring>
#include <vector>
using namespace LAMMPS_NS;
@ -182,10 +183,6 @@ void CommTiled::setup()
if (mode == Comm::MULTI) {
// build collection from scratch as it is needed for atom exchange
neighbor->build_collection(0);
if(cutusermulti and ncollections != neighbor->ncollections)
error->all(FLERR, "Cannot change number of collections after defining comm_modify multi/cutoff");
ncollections = neighbor->ncollections;
// allocate memory for multi-style communication at setup as ncollections can change
@ -197,9 +194,25 @@ void CommTiled::setup()
for(i = 0; i < maxswap; i ++)
grow_swap_send_multi(i,DELTA_PROCS);
memory->grow(cutusermulti,ncollections,"comm:cutusermulti");
for(i = ncollections_prior; i < ncollections; i++)
cutusermulti[i] = -1.0;
ncollections_prior = ncollections;
}
// parse any cutoff/multi commands
int nhi, nlo;
for(auto it = usermultiargs.begin(); it != usermultiargs.end(); it ++) {
utils::bounds(FLERR,it->first,1,ncollections,nlo,nhi,error);
if(nhi >= ncollections)
error->all(FLERR, "Unused collection id in comm_modify cutoff/multi command");
for (j=nlo; j<=nhi; ++j)
cutusermulti[j] = it->second;
}
usermultiargs.clear();
double **cutcollectionsq = neighbor->cutcollectionsq;
// If using multi/reduce, communicate particles a distance equal
// to the max cutoff with equally sized or smaller collections