lijiext
/
lammps
mirror of https://github.com/lammps/lammps.git
1
0
Fork 1
Code Issues Packages Projects Releases Wiki Activity

changes to imbalance weight factors

This commit is contained in:
Steve Plimpton 2016-10-05 10:33:39 -06:00
parent 11c2892e54
commit c46be7db62
36 changed files with 241 additions and 302 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
doc/src/Manual.txt
View File

@ -1,7 +1,7 @@
<!-- HTML_ONLY -->
<HEAD>
<TITLE>LAMMPS Users Manual</TITLE>
<META NAME="docnumber" CONTENT="30 Sep 2016 version">
<META NAME="docnumber" CONTENT="5 Oct 2016 version">
<META NAME="author" CONTENT="http://lammps.sandia.gov - Sandia National Laboratories">
<META NAME="copyright" CONTENT="Copyright (2003) Sandia Corporation. This software and manual is distributed under the GNU General Public License.">
</HEAD>
@ -21,7 +21,7 @@
<H1></H1>
LAMMPS Documentation :c,h3
30 Sep 2016 version :c,h4
5 Oct 2016 version :c,h4
Version info: :h4

108
doc/src/balance.txt
View File

@ -319,24 +319,25 @@ accurately would be impractical and slow down the computation.
Instead the {weight} keyword implements several ways to influence the
per-particle weights empirically by properties readily available or
using the user's knowledge of the system. Note that the absolute
value of the weights are not important; their ratio is what is used to
assign particles to processors. A particle with a weight of 2.5 is
assumed to require 5x more computational than a particle with a weight
of 0.5.
value of the weights are not important; only their relative ratios
affect which particle is assigned to which processor. A particle with
a weight of 2.5 is assumed to require 5x more computational than a
particle with a weight of 0.5. For all the options below the weight
assigned to a particle must be a positive value; an error will be be
generated if a weight is <= 0.0.
Below is a list of possible weight options with a short description of
their usage and some example scenarios where they might be applicable.
It is possible to apply multiple weight flags and the weightins they
It is possible to apply multiple weight flags and the weightings they
induce will be combined through multiplication. Most of the time,
however, it is sufficient to use just one method.
The {group} weight style assigns weight factors to specified
"groups"_group.html of particles. The {group} style keyword is
followed by the number of groups, then pairs of group IDs and the
corresponding weight factor. If a particle belongs to none of the
corresponding weight factor. If a particle belongs to none of the
specified groups, its weight is not changed. If it belongs to
multiple groups, its weight is the product of the weight factors.
The weight factors have to be positive.
This weight style is useful in combination with pair style
"hybrid"_pair_hybrid.html, e.g. when combining a more costly manybody
@ -347,14 +348,24 @@ the computational cost for each group remains constant over time.
This is a purely empirical weighting, so a series test runs to tune
the assigned weight factors for optimal performance is recommended.
The {neigh} weight style assigns a weight to each particle equal to
its number of neighbors divided by the avergage number of neighbors
for all particles. The {factor} setting is then appied as an overall
scale factor to all the {neigh} weights which allows tuning of the
impact of this style. A {factor} smaller than 1.0 (e.g. 0.8) often
results in the best performance, since the number of neighbors is
likely to overestimate the ideal weight. The factor has to be between
0.0 and 2.0.
The {neigh} weight style assigns the same weight to each particle
owned by a processor based on the total count of neighbors in the
neighbor list owned by that processor. The motivation is that more
neighbors means a higher computational cost. The style does not use
neighbors per atom to assign a unique weight to each atom, because
that value can vary depending on how the neighbor list is built.
The {factor} setting is applied as an overall scale factor to the
{neigh} weights which allows adjustment of their impact on the
balancing operation. The specified {factor} value must be positive.
A value > 1.0 will increase the weights so that the ratio of max
weight to min weight increases by {factor}. A value < 1.0 will
decrease the weights so that the ratio of max weight to min weight
decreases by {factor}. In both cases the intermediate weight values
increase/decrease proportionally as well. A value = 1.0 has no effect
on the {neigh} weights. As a rule of thumb, we have found a {factor}
of about 0.8 often results in the best performance, since the number
of neighbors is likely to overestimate the ideal weight.
This weight style is useful for systems where there are different
cutoffs used for different pairs of interations, or the density
@ -370,35 +381,48 @@ weights are computed. Inserting a "run 0 post no"_run.html command
before issuing the {balance} command, may be a workaround for this
case, as it will induce the neighbor list to be built.
The {time} weight style uses "timer data"_timer.html to estimate a
weight for each particle. It uses the same information as is used for
the "MPI task timing breakdown"_Section_start.html#start_8, namely,
the timings for sections {Pair}, {Bond}, {Kspace}, and {Neigh}. The
time spent in these sections of the timestep are measured for each MPI
rank, summed up, then converted into a cost for each MPI rank relative
to the average cost over all MPI ranks for the same sections. That
cost then evenly distributed over all the particles owned by that
rank. Finally, the {factor} setting is then appied as an overall
scale factor to all the {time} weights as a way to fine tune the
impact of this weight style. Good {factor} values to use are
typically between 0.5 and 1.2. Allowed are values between 0.0 and 2.0.
The {time} weight style uses "timer data"_timer.html to estimate
weights. It assigns the same weight to each particle owned by a
processor based on the total computational time spent by that
processor. See details below on what time window is used. It uses
the same timing information as is used for the "MPI task timing
breakdown"_Section_start.html#start_8, namely, for sections {Pair},
{Bond}, {Kspace}, and {Neigh}. The time spent in those portions of
the timestep are measured for each MPI rank, summed, then divided by
the number of particles owned by that processor. I.e. the weight is
an effective CPU time/particle averaged over the particles on that
processor.
For the {balance} command the timing data is taken from the preceding
run command, i.e. the timings are for the entire previous run. For
the {fix balance} command the timing data is for only the timesteps
since the last balancing operation was performed. If timing
information for the required sections is not available, e.g. at the
beginning of a run, or when the "timer"_timer.html command is set to
either {loop} or {off}, a warning is issued. In this case no weights
are computed.
The {factor} setting is applied as an overall scale factor to the
{time} weights which allows adjustment of their impact on the
balancing operation. The specified {factor} value must be positive.
A value > 1.0 will increase the weights so that the ratio of max
weight to min weight increases by {factor}. A value < 1.0 will
decrease the weights so that the ratio of max weight to min weight
decreases by {factor}. In both cases the intermediate weight values
increase/decrease proportionally as well. A value = 1.0 has no effect
on the {time} weights. As a rule of thumb, effective values to use
are typicall between 0.5 and 1.2. Note that the timer quantities
mentioned above can be affected by communication which occurs in the
middle of the operations, e.g. pair styles with intermediate exchange
of data witin the force computation, and likewise for KSpace solves.
This weight style is the most generic one, and should be tried first,
if neither the {group} or {neigh} styles are easily applicable.
However, since the computed cost function is averaged over all local
particles this weight style may not be highly accurate. This style
can also be effective as a secondary weight in combination with either
{group} or {neigh} to offset some of inaccuracies in either of those
heuristics.
When using the {time} weight style with the {balance} command, the
timing data is taken from the preceding run command, i.e. the timings
are for the entire previous run. For the {fix balance} command the
timing data is for only the timesteps since the last balancing
operation was performed. If timing information for the required
sections is not available, e.g. at the beginning of a run, or when the
"timer"_timer.html command is set to either {loop} or {off}, a warning
is issued. In this case no weights are computed.
NOTE: The {time} weight style is the most generic option, and should
be tried first, unless the {group} style is easily applicable.
However, since the computed cost function is averaged over all
particles on a processor, the weights may not be highly accurate.
This style can also be effective as a secondary weight in combination
with either {group} or {neigh} to offset some of inaccuracies in
either of those heuristics.
The {var} weight style assigns per-particle weights by evaluating an
"atom-style variable"_variable.html specified by {name}. This is

57
doc/src/dump_modify.txt
View File

@ -49,8 +49,8 @@ keyword = {append} or {buffer} or {element} or {every} or {fileper} or {first} o
-N = sort per-atom lines in descending order by the Nth column
{thresh} args = attribute operation value
attribute = same attributes (x,fy,etotal,sxx,etc) used by dump custom style
operation = "<" or "<=" or ">" or ">=" or "==" or "!="
value = numeric value to compare to
operation = "<" or "<=" or ">" or ">=" or "==" or "!=" or "|^"
value = numeric value to compare to, or LAST
these 3 args can be replaced by the word "none" to turn off thresholding
{unwrap} arg = {yes} or {no} :pre
these keywords apply only to the {image} and {movie} "styles"_dump_image.html :l
@ -458,16 +458,59 @@ as well as memory, versus unsorted output.
The {thresh} keyword only applies to the dump {custom}, {cfg},
{image}, and {movie} styles. Multiple thresholds can be specified.
Specifying "none" turns off all threshold criteria. If thresholds are
Specifying {none} turns off all threshold criteria. If thresholds are
specified, only atoms whose attributes meet all the threshold criteria
are written to the dump file or included in the image. The possible
attributes that can be tested for are the same as those that can be
specified in the "dump custom"_dump.html command, with the exception
of the {element} attribute, since it is not a numeric value. Note
that different attributes can be output by the dump custom command
than are used as threshold criteria by the dump_modify command.
E.g. you can output the coordinates and stress of atoms whose energy
is above some threshold.
that a different attributes can be used than those output by the "dump
custom"_dump.html command. E.g. you can output the coordinates and
stress of atoms whose energy is above some threshold.
If an atom-style variable is used as the attribute, then it can
produce continuous numeric values or effective Boolean 0/1 values
which may be useful for the comparision operation. Boolean values can
be generated by variable formulas that use comparison or Boolean math
operators or special functions like gmask() and rmask() and grmask().
See the "variable"_variable.html command doc page for details.
NOTE: The LAST option, discussed below, is not yet implemented. It
will be soon.
The specified value must be a simple numeric value or the word LAST.
If LAST is used, it refers to the value of the attribute the last time
the dump command was invoked to produce a snapshot. This is a way to
only dump atoms whose attribute has changed (or not changed).
Three examples follow.
dump_modify ... thresh ix != LAST :pre
This will dump atoms which have crossed the periodic x boundary of the
simulation box since the last dump. (Note that atoms that crossed
once and then crossed back between the two dump timesteps would not be
included.)
region foo sphere 10 20 10 15
variable inregion atom rmask(foo)
dump_modify ... thresh v_inregion |^ LAST
This will dump atoms which crossed the boundary of the spherical
region since the last dump.
variable charge atom "(q > 0.5) || (q < -0.5)"
dump_modify ... thresh v_charge |^ LAST
This will dump atoms whose charge has changed from an absolute value
less than 1/2 to greater than 1/2 (or vice versa) since the last dump.
E.g. due to reactions and subsequent charge equilibration in a
reactive force field.
The choice of operations are the usual comparison operators. The XOR
operation (exclusive or) is also included as "|^". In this context,
XOR means that if either the attribute or value is 0.0 and the other
is non-zero, then the result is "true" and the threshold criterion is
met. Otherwise it is not met.
:line

10
examples/body/in.body
View File

@ -11,13 +11,19 @@ velocity all create 1.44 87287 loop geom
pair_style body 5.0
pair_coeff * * 1.0 1.0
neighbor 0.3 bin
neighbor 0.5 bin
neigh_modify every 1 delay 0 check yes
fix 1 all nve/body
#fix 1 all nvt/body temp 1.44 1.44 1.0
fix 2 all enforce2d
#compute 1 all body/local type 1 2 3
#dump 1 all local 100 dump.body index c_1[1] c_1[2] c_1[3] c_1[4]
thermo 500
#dump 2 all image 1000 image.*.jpg type type &
# zoom 1.6 adiam 1.5 body type 1.0 0
#dump_modify 2 pad 5
thermo 100
run 10000

3
src/angle.cpp
View File

@ -40,9 +40,6 @@ Angle::Angle(LAMMPS *lmp) : Pointers(lmp)
vatom = NULL;
setflag = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/angle.h
View File

@ -29,10 +29,9 @@ class Angle : protected Pointers {
double energy; // accumulated energies
double virial[6]; // accumlated virial
double *eatom,**vatom; // accumulated per-atom energy/virial
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
int copymode;
@ -51,9 +50,6 @@ class Angle : protected Pointers {
virtual double single(int, int, int, int) = 0;
virtual double memory_usage();
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
protected:
int suffix_flag; // suffix compatibility flag

3
src/atom.cpp
View File

@ -208,9 +208,6 @@ Atom::Atom(LAMMPS *lmp) : Pointers(lmp)
atom_style = NULL;
avec = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
avec_map = new AtomVecCreatorMap();
#define ATOM_CLASS

5
src/atom.h
View File

@ -124,11 +124,6 @@ class Atom : protected Pointers {
char **iname,**dname;
int nivector,ndvector;
// used by USER-CUDA to flag used per-atom arrays
unsigned int datamask;
unsigned int datamask_ext;
// atom style and per-atom array existence flags
// customize by adding new flag

4
src/atom_vec.h
View File

@ -156,10 +156,6 @@ E: Invalid atom_style command
Self-explanatory.
E: USER-CUDA package requires a cuda enabled atom_style
Self-explanatory.
E: KOKKOS package requires a kokkos enabled atom_style
Self-explanatory.

3
src/bond.cpp
View File

@ -44,9 +44,6 @@ Bond::Bond(LAMMPS *lmp) : Pointers(lmp)
vatom = NULL;
setflag = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/bond.h
View File

@ -29,10 +29,9 @@ class Bond : protected Pointers {
double energy; // accumulated energies
double virial[6]; // accumlated virial
double *eatom,**vatom; // accumulated per-atom energy/virial
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
int copymode;
@ -51,9 +50,6 @@ class Bond : protected Pointers {
virtual double single(int, double, int, int, double &) = 0;
virtual double memory_usage();
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
void write_file(int, char**);
protected:

4
src/comm_tiled.h
View File

@ -155,10 +155,6 @@ class CommTiled : public Comm {
/* ERROR/WARNING messages:
E: USER-CUDA package does not yet support comm_style tiled
Self-explanatory.
E: KOKKOS package does not yet support comm_style tiled
Self-explanatory.

3
src/compute.cpp
View File

@ -99,9 +99,6 @@ Compute::Compute(LAMMPS *lmp, int narg, char **arg) : Pointers(lmp),
// data masks
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/compute.h
View File

@ -84,9 +84,6 @@ class Compute : protected Pointers {
int comm_reverse; // size of reverse communication (0 if none)
int dynamic_group_allow; // 1 if can be used with dynamic group, else 0
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
@ -140,9 +137,6 @@ class Compute : protected Pointers {
double, double, double,
double, double, double) {}
virtual int unsigned data_mask() {return datamask;}
virtual int unsigned data_mask_ext() {return datamask_ext;}
protected:
int instance_me; // which Compute class instantiation I am

3
src/dihedral.cpp
View File

@ -41,9 +41,6 @@ Dihedral::Dihedral(LAMMPS *lmp) : Pointers(lmp)
vatom = NULL;
setflag = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/dihedral.h
View File

@ -29,10 +29,9 @@ class Dihedral : protected Pointers {
double energy; // accumulated energy
double virial[6]; // accumlated virial
double *eatom,**vatom; // accumulated per-atom energy/virial
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
int copymode;
@ -49,9 +48,6 @@ class Dihedral : protected Pointers {
virtual void write_data(FILE *) {}
virtual double memory_usage();
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
protected:
int suffix_flag; // suffix compatibility flag

8
src/dump_custom.cpp
View File

@ -43,7 +43,7 @@ enum{ID,MOL,PROC,PROCP1,TYPE,ELEMENT,MASS,
OMEGAX,OMEGAY,OMEGAZ,ANGMOMX,ANGMOMY,ANGMOMZ,
TQX,TQY,TQZ,
COMPUTE,FIX,VARIABLE,INAME,DNAME};
enum{LT,LE,GT,GE,EQ,NEQ};
enum{LT,LE,GT,GE,EQ,NEQ,XOR};
enum{INT,DOUBLE,STRING,BIGINT}; // same as in DumpCFG
#define INVOKED_PERATOM 8
@ -947,6 +947,11 @@ int DumpCustom::count()
} else if (thresh_op[ithresh] == NEQ) {
for (i = 0; i < nlocal; i++, ptr += nstride)
if (choose[i] && *ptr == value) choose[i] = 0;
} else if (thresh_op[ithresh] == XOR) {
for (i = 0; i < nlocal; i++, ptr += nstride)
if (choose[i] && (*ptr == 0.0 && value == 0.0) ||
(*ptr != 0.0 && value != 0.0))
choose[i] = 0;
}
}
}
@ -1835,6 +1840,7 @@ int DumpCustom::modify_param(int narg, char **arg)
else if (strcmp(arg[2],">=") == 0) thresh_op[nthresh] = GE;
else if (strcmp(arg[2],"==") == 0) thresh_op[nthresh] = EQ;
else if (strcmp(arg[2],"!=") == 0) thresh_op[nthresh] = NEQ;
else if (strcmp(arg[2],"|^") == 0) thresh_op[nthresh] = XOR;
else error->all(FLERR,"Invalid dump_modify threshold operator");
// set threshold value

5
src/fix.cpp
View File

@ -95,10 +95,7 @@ id(NULL), style(NULL), eatom(NULL), vatom(NULL)
maxeatom = maxvatom = 0;
vflag_atom = 0;
// CUDA and KOKKOS per-fix data masks
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
// KOKKOS per-fix data masks
execution_space = Host;
datamask_read = ALL_MASK;

8
src/fix.h
View File

@ -99,11 +99,6 @@ class Fix : protected Pointers {
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
// USER-CUDA per-fix data masks
unsigned int datamask;
unsigned int datamask_ext;
Fix(class LAMMPS *, int, char **);
virtual ~Fix();
void modify_params(int, char **);
@ -211,9 +206,6 @@ class Fix : protected Pointers {
virtual double memory_usage() {return 0.0;}
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
protected:
int instance_me; // which Fix class instantiation I am

12
src/imbalance_group.cpp
View File

@ -18,12 +18,11 @@
#include "error.h"
using namespace LAMMPS_NS;
#define SMALL 0.001
/* -------------------------------------------------------------------- */
ImbalanceGroup::ImbalanceGroup(LAMMPS *lmp) : Imbalance(lmp),
id(0), factor(0), num(0) {}
ImbalanceGroup::ImbalanceGroup(LAMMPS *lmp) : Imbalance(lmp), id(0), factor(0)
{}
/* -------------------------------------------------------------------- */
@ -50,7 +49,7 @@ int ImbalanceGroup::options(int narg, char **arg)
if (id[i] < 0)
error->all(FLERR,"Unknown group in balance weight command");
factor[i] = force->numeric(FLERR,arg[2*i+2]);
if (factor[i] < 0.0) error->all(FLERR,"Illegal balance weight command");
if (factor[i] <= 0.0) error->all(FLERR,"Illegal balance weight command");
}
return 2*num+1;
}
@ -67,13 +66,10 @@ void ImbalanceGroup::compute(double *weight)
for (int i = 0; i < nlocal; ++i) {
const int imask = mask[i];
double iweight = weight[i];
for (int j = 0; j < num; ++j) {
if (imask & bitmask[id[j]])
iweight *= factor[j];
weight[i] *= factor[j];
}
if (iweight < SMALL) weight[i] = SMALL;
else weight[i] = iweight;
}
}

54
src/imbalance_neigh.cpp
View File

@ -22,14 +22,14 @@
#include "error.h"
using namespace LAMMPS_NS;
#define SMALL 0.001
#define BIG 1.0e20
/* -------------------------------------------------------------------- */
ImbalanceNeigh::ImbalanceNeigh(LAMMPS *lmp) : Imbalance(lmp)
{
did_warn = 0;
factor = 1.0;
}
/* -------------------------------------------------------------------- */
@ -38,8 +38,7 @@ int ImbalanceNeigh::options(int narg, char **arg)
{
if (narg < 1) error->all(FLERR,"Illegal balance weight command");
factor = force->numeric(FLERR,arg[0]);
if ((factor < 0.0) || (factor > 2.0))
error->all(FLERR,"Illegal balance weight command");
if (factor <= 0.0) error->all(FLERR,"Illegal balance weight command");
return 1;
}
@ -52,7 +51,7 @@ void ImbalanceNeigh::compute(double *weight)
if (factor == 0.0) return;
// find suitable neighbor list
// we can only make use of certain (conventional) neighbor lists
// can only use certain conventional neighbor lists
for (req = 0; req < neighbor->old_nrequest; ++req) {
if ((neighbor->old_requests[req]->half ||
@ -65,37 +64,46 @@ void ImbalanceNeigh::compute(double *weight)
if (req >= neighbor->old_nrequest || neighbor->ago < 0) {
if (comm->me == 0 && !did_warn)
error->warning(FLERR,"No suitable neighbor list found. "
"Neighbor weighted balancing skipped");
error->warning(FLERR,"Balance weight neigh skipped b/c no list found");
did_warn = 1;
return;
}
// neighsum = total neigh count for atoms on this proc
// localwt = weight assigned to each owned atom
NeighList *list = neighbor->lists[req];
bigint neighsum = 0;
const int inum = list->inum;
const int * const ilist = list->ilist;
const int * const numneigh = list->numneigh;
int nlocal = atom->nlocal;
// first pass: get local number of neighbors
bigint neighsum = 0;
for (int i = 0; i < inum; ++i) neighsum += numneigh[ilist[i]];
double localwt = 0.0;
if (nlocal) localwt = 1.0*neighsum/nlocal;
double allatoms = static_cast <double>(atom->natoms);
if (allatoms == 0.0) allatoms = 1.0;
double allavg;
double myavg = static_cast<double>(neighsum)/allatoms;
MPI_Allreduce(&myavg,&allavg,1,MPI_DOUBLE,MPI_SUM,world);
// second pass: compute and apply weights
if (nlocal && localwt <= 0.0) error->one(FLERR,"Balance weight <= 0.0");
double scale = 1.0/allavg;
for (int ii = 0; ii < inum; ++ii) {
const int i = ilist[ii];
weight[i] *= (1.0-factor) + factor*scale*numneigh[i];
if (weight[i] < SMALL) weight[i] = SMALL;
// apply factor if specified != 1.0
// wtlo,wthi = lo/hi values excluding 0.0 due to no atoms on this proc
// lo value does not change
// newhi = new hi value to give hi/lo ratio factor times larger/smaller
// expand/contract all localwt values from lo->hi to lo->newhi
if (factor != 1.0) {
double wtlo,wthi;
if (localwt == 0.0) localwt = BIG;
MPI_Allreduce(&localwt,&wtlo,1,MPI_DOUBLE,MPI_MIN,world);
if (localwt == BIG) localwt = 0.0;
MPI_Allreduce(&localwt,&wthi,1,MPI_DOUBLE,MPI_MAX,world);
if (wtlo == wthi) return;
double newhi = wthi*factor;
localwt = wtlo + ((localwt-wtlo)/(wthi-wtlo)) * (newhi-wtlo);
}
for (int i = 0; i < nlocal; i++) weight[i] *= localwt;
}
/* -------------------------------------------------------------------- */

85
src/imbalance_time.cpp
View File

@ -19,15 +19,16 @@
#include "timer.h"
#include "error.h"
// DEBUG
#include "update.h"
using namespace LAMMPS_NS;
#define SMALL 0.001
#define BIG 1.0e20
/* -------------------------------------------------------------------- */
ImbalanceTime::ImbalanceTime(LAMMPS *lmp) : Imbalance(lmp)
{
factor = 1.0;
}
ImbalanceTime::ImbalanceTime(LAMMPS *lmp) : Imbalance(lmp) {}
/* -------------------------------------------------------------------- */
@ -35,8 +36,7 @@ int ImbalanceTime::options(int narg, char **arg)
{
if (narg < 1) error->all(FLERR,"Illegal balance weight command");
factor = force->numeric(FLERR,arg[0]);
if ((factor < 0.0) || (factor > 2.0))
error->all(FLERR,"Illegal balance weight command");
if (factor <= 0.0) error->all(FLERR,"Illegal balance weight command");
return 1;
}
@ -53,37 +53,60 @@ void ImbalanceTime::init()
void ImbalanceTime::compute(double *weight)
{
const int nlocal = atom->nlocal;
const bigint natoms = atom->natoms;
if (!timer->has_normal()) return;
if (factor == 0.0) return;
// cost = CPU time for relevant timers since last invocation
// localwt = weight assigned to each owned atom
// just return if no time yet tallied
// compute the cost function of based on relevant timers
if (timer->has_normal()) {
double cost = -last;
cost += timer->get_wall(Timer::PAIR);
cost += timer->get_wall(Timer::NEIGH);
cost += timer->get_wall(Timer::BOND);
cost += timer->get_wall(Timer::KSPACE);
double cost = -last;
cost += timer->get_wall(Timer::PAIR);
cost += timer->get_wall(Timer::NEIGH);
cost += timer->get_wall(Timer::BOND);
cost += timer->get_wall(Timer::KSPACE);
double allcost;
MPI_Allreduce(&cost,&allcost,1,MPI_DOUBLE,MPI_SUM,world);
/*
printf("TIME %ld %d %g %g: %g %g %g %g\n",
update->ntimestep,atom->nlocal,last,cost,
timer->get_wall(Timer::PAIR),
timer->get_wall(Timer::NEIGH),
timer->get_wall(Timer::BOND),
timer->get_wall(Timer::KSPACE));
*/
if ((allcost > 0.0) && (nlocal > 0)) {
const double avgcost = allcost/natoms;
const double localcost = cost/nlocal;
const double scale = (1.0-factor) + factor*localcost/avgcost;
for (int i = 0; i < nlocal; ++i) {
weight[i] *= scale;
if (weight[i] < SMALL) weight[i] = SMALL;
}
}
double maxcost;
MPI_Allreduce(&cost,&maxcost,1,MPI_DOUBLE,MPI_MAX,world);
if (maxcost <= 0.0) return;
// record time up to this point
int nlocal = atom->nlocal;
double localwt = 0.0;
if (nlocal) localwt = cost/nlocal;
last += cost;
if (nlocal && localwt <= 0.0) error->one(FLERR,"Balance weight <= 0.0");
// apply factor if specified != 1.0
// wtlo,wthi = lo/hi values excluding 0.0 due to no atoms on this proc
// lo value does not change
// newhi = new hi value to give hi/lo ratio factor times larger/smaller
// expand/contract all localwt values from lo->hi to lo->newhi
if (factor != 1.0) {
double wtlo,wthi;
if (localwt == 0.0) localwt = BIG;
MPI_Allreduce(&localwt,&wtlo,1,MPI_DOUBLE,MPI_MIN,world);
if (localwt == BIG) localwt = 0.0;
MPI_Allreduce(&localwt,&wthi,1,MPI_DOUBLE,MPI_MAX,world);
if (wtlo == wthi) return;
double newhi = wthi*factor;
localwt = wtlo + ((localwt-wtlo)/(wthi-wtlo)) * (newhi-wtlo);
}
for (int i = 0; i < nlocal; i++) weight[i] *= localwt;
// record time up to this point
last += cost;
}
/* -------------------------------------------------------------------- */

16
src/imbalance_var.cpp
View File

@ -24,11 +24,10 @@
#include "update.h"
using namespace LAMMPS_NS;
#define SMALL 0.001
/* -------------------------------------------------------------------- */
ImbalanceVar::ImbalanceVar(LAMMPS *lmp) : Imbalance(lmp), name(0), id(0) {}
ImbalanceVar::ImbalanceVar(LAMMPS *lmp) : Imbalance(lmp), name(0) {}
/* -------------------------------------------------------------------- */
@ -76,10 +75,15 @@ void ImbalanceVar::compute(double *weight)
memory->create(values,nlocal,"imbalance:values");
input->variable->compute_atom(id,all,values,1,0);
for (int i = 0; i < nlocal; ++i) {
weight[i] *= values[i];
if (weight[i] < SMALL) weight[i] = SMALL;
}
int flag = 0;
for (int i = 0; i < nlocal; i++)
if (values[i] <= 0.0) flag = 1;
int flagall;
MPI_Allreduce(&flag,&flagall,1,MPI_INT,MPI_SUM,world);
if (flagall) error->one(FLERR,"Balance weight <= 0.0");
for (int i = 0; i < nlocal; i++) weight[i] *= values[i];
memory->destroy(values);
}

3
src/improper.cpp
View File

@ -38,9 +38,6 @@ Improper::Improper(LAMMPS *lmp) : Pointers(lmp)
vatom = NULL;
setflag = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/improper.h
View File

@ -29,10 +29,9 @@ class Improper : protected Pointers {
double energy; // accumulated energies
double virial[6]; // accumlated virial
double *eatom,**vatom; // accumulated per-atom energy/virial
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
int copymode;
@ -49,9 +48,6 @@ class Improper : protected Pointers {
virtual void write_data(FILE *) {}
virtual double memory_usage();
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
protected:
int suffix_flag; // suffix compatibility flag

6
src/input.h
View File

@ -328,12 +328,6 @@ E: Package command after simulation box is defined
The package command cannot be used afer a read_data, read_restart, or
create_box command.
E: Package cuda command without USER-CUDA package enabled
The USER-CUDA package must be installed via "make yes-user-cuda"
before LAMMPS is built, and the "-c on" must be used to enable the
package.
E: Package gpu command without GPU package installed
The GPU package must be installed via "make yes-gpu" before LAMMPS is

3
src/kspace.cpp
View File

@ -88,9 +88,6 @@ KSpace::KSpace(LAMMPS *lmp, int narg, char **arg) : Pointers(lmp)
eatom = NULL;
vatom = NULL;
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

4
src/kspace.h
View File

@ -80,10 +80,8 @@ class KSpace : protected Pointers {
int group_group_enable; // 1 if style supports group/group calculation
unsigned int datamask;
unsigned int datamask_ext;
// KOKKOS host/device flag and data masks
ExecutionSpace execution_space;
unsigned int datamask_read,datamask_modify;
int copymode;

9
src/lammps.h
View File

@ -168,19 +168,10 @@ E: Cannot use -cuda on and -kokkos on together
This is not allowed since both packages can use GPUs.
E: Cannot use -cuda on without USER-CUDA installed
The USER-CUDA package must be installed via "make yes-user-cuda"
before LAMMPS is built.
E: Cannot use -kokkos on without KOKKOS installed
Self-explanatory.
E: Using suffix cuda without USER-CUDA package enabled
Self-explanatory.
E: Using suffix gpu without GPU package installed
Self-explanatory.

3
src/pair.cpp
View File

@ -100,9 +100,6 @@ Pair::Pair(LAMMPS *lmp) : Pointers(lmp)
// KOKKOS per-fix data masks
datamask = ALL_MASK;
datamask_ext = ALL_MASK;
execution_space = Host;
datamask_read = ALL_MASK;
datamask_modify = ALL_MASK;

6
src/pair.h
View File

@ -97,9 +97,6 @@ class Pair : protected Pointers {
class NeighList *listmiddle;
class NeighList *listouter;
unsigned int datamask;
unsigned int datamask_ext;
int allocated; // 0/1 = whether arrays are allocated
// public so external driver can check
int compute_flag; // 0 if skip compute()
@ -191,9 +188,6 @@ class Pair : protected Pointers {
virtual void min_xf_get(int) {}
virtual void min_x_set(int) {}
virtual unsigned int data_mask() {return datamask;}
virtual unsigned int data_mask_ext() {return datamask_ext;}
// management of callbacks to be run from ev_tally()
protected:

5
src/suffix.h
View File

@ -20,9 +20,8 @@ namespace Suffix {
static const int NONE = 0;
static const int OPT = 1<<0;
static const int GPU = 1<<1;
static const int CUDA = 1<<2;
static const int OMP = 1<<3;
static const int INTEL = 1<<4;
static const int OMP = 1<<2;
static const int INTEL = 1<<3;
}
}

8
src/update.h
View File

@ -81,14 +81,6 @@ class Update : protected Pointers {
/* ERROR/WARNING messages:
E: USER-CUDA mode requires CUDA variant of run style
CUDA mode is enabled, so the run style must include a cuda suffix.
E: USER-CUDA mode requires CUDA variant of min style
CUDA mode is enabled, so the min style must include a cuda suffix.
E: Illegal ... command
Self-explanatory. Check the input script syntax and compare to the

66
src/variable.cpp
View File

@ -4813,72 +4813,6 @@ double Variable::evaluate_boolean(char *str)
return argstack[0].value;
}
/* ---------------------------------------------------------------------- */
unsigned int Variable::data_mask(int ivar)
{
if (eval_in_progress[ivar]) return EMPTY_MASK;
eval_in_progress[ivar] = 1;
unsigned int datamask = data_mask(data[ivar][0]);
eval_in_progress[ivar] = 0;
return datamask;
}
/* ---------------------------------------------------------------------- */
unsigned int Variable::data_mask(char *str)
{
unsigned int datamask = EMPTY_MASK;
for (unsigned int i = 0; i < strlen(str)-2; i++) {
int istart = i;
while (isalnum(str[i]) || str[i] == '_') i++;
int istop = i-1;
int n = istop - istart + 1;
char *word = new char[n+1];
strncpy(word,&str[istart],n);
word[n] = '\0';
// ----------------
// compute
// ----------------
if ((strncmp(word,"c_",2) == 0) && (i>0) && (!isalnum(str[i-1]))) {
if (domain->box_exist == 0)
error->all(FLERR,
"Variable evaluation before simulation box is defined");
int icompute = modify->find_compute(word+2);
if (icompute < 0)
error->all(FLERR,"Invalid compute ID in variable formula");
datamask &= modify->compute[icompute]->data_mask();
}
if ((strncmp(word,"f_",2) == 0) && (i>0) && (!isalnum(str[i-1]))) {
if (domain->box_exist == 0)
error->all(FLERR,
"Variable evaluation before simulation box is defined");
int ifix = modify->find_fix(word+2);
if (ifix < 0) error->all(FLERR,"Invalid fix ID in variable formula");
datamask &= modify->fix[ifix]->data_mask();
}
if ((strncmp(word,"v_",2) == 0) && (i>0) && (!isalnum(str[i-1]))) {
int ivar = find(word+2);
if (ivar < 0) error->all(FLERR,"Invalid variable name in variable formula");
datamask &= data_mask(ivar);
}
delete [] word;
}
return datamask;
}
/* ----------------------------------------------------------------------
class to read variable values from a file
for flag = SCALARFILE, reads one value per line

3
src/variable.h
View File

@ -49,9 +49,6 @@ class Variable : protected Pointers {
tagint int_between_brackets(char *&, int);
double evaluate_boolean(char *);
unsigned int data_mask(int ivar);
unsigned int data_mask(char *str);
private:
int me;
int nvar; // # of defined variables

2
src/version.h
View File

@ -1 +1 @@
#define LAMMPS_VERSION "30 Sep 2016"
#define LAMMPS_VERSION "5 Oct 2016"