1 |
< |
/** |
2 |
< |
* @file ForceDecomposition.cpp |
3 |
< |
* @author Charles Vardeman <cvardema.at.nd.edu> |
4 |
< |
* @date 08/18/2010 |
5 |
< |
* @time 11:56am |
6 |
< |
* @version 1.0 |
1 |
> |
/* |
2 |
> |
* Copyright (c) 2005 The University of Notre Dame. All Rights Reserved. |
3 |
|
* |
8 |
– |
* @section LICENSE |
9 |
– |
* Copyright (c) 2010 The University of Notre Dame. All Rights Reserved. |
10 |
– |
* |
4 |
|
* The University of Notre Dame grants you ("Licensee") a |
5 |
|
* non-exclusive, royalty free, license to use, modify and |
6 |
|
* redistribute this software in source and binary code form, provided |
38 |
|
* [3] Sun, Lin & Gezelter, J. Chem. Phys. 128, 24107 (2008). |
39 |
|
* [4] Vardeman & Gezelter, in progress (2009). |
40 |
|
*/ |
41 |
+ |
#include "parallel/ForceDecomposition.hpp" |
42 |
+ |
#include "parallel/Communicator.hpp" |
43 |
+ |
#include "math/SquareMatrix3.hpp" |
44 |
|
|
45 |
+ |
using namespace std; |
46 |
+ |
namespace OpenMD { |
47 |
|
|
48 |
< |
|
51 |
< |
/* -*- c++ -*- */ |
52 |
< |
#include "config.h" |
53 |
< |
#include <stdlib.h> |
48 |
> |
void ForceDecomposition::distributeInitialData() { |
49 |
|
#ifdef IS_MPI |
50 |
< |
#include <mpi.h> |
51 |
< |
#endif |
50 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
51 |
> |
int nAtoms = snap->getNumberOfAtoms(); |
52 |
> |
int nGroups = snap->getNumberOfCutoffGroups(); |
53 |
|
|
54 |
< |
#include <iostream> |
55 |
< |
#include <vector> |
56 |
< |
#include <algorithm> |
61 |
< |
#include <cmath> |
62 |
< |
#include "parallel/ForceDecomposition.hpp" |
54 |
> |
AtomCommRealI = new Communicator<Row,RealType>(nAtoms); |
55 |
> |
AtomCommVectorI = new Communicator<Row,Vector3d>(nAtoms); |
56 |
> |
AtomCommMatrixI = new Communicator<Row,Mat3x3d>(nAtoms); |
57 |
|
|
58 |
+ |
AtomCommRealJ = new Communicator<Column,RealType>(nAtoms); |
59 |
+ |
AtomCommVectorJ = new Communicator<Column,Vector3d>(nAtoms); |
60 |
+ |
AtomCommMatrixJ = new Communicator<Column,Mat3x3d>(nAtoms); |
61 |
|
|
62 |
< |
using namespace std; |
63 |
< |
using namespace OpenMD; |
62 |
> |
cgCommVectorI = new Communicator<Row,Vector3d>(nGroups); |
63 |
> |
cgCommVectorJ = new Communicator<Column,Vector3d>(nGroups); |
64 |
|
|
65 |
< |
//__static |
66 |
< |
#ifdef IS_MPI |
70 |
< |
static vector<MPI:Comm> communictors; |
71 |
< |
#endif |
65 |
> |
int nInRow = AtomCommRealI.getSize(); |
66 |
> |
int nInCol = AtomCommRealJ.getSize(); |
67 |
|
|
68 |
< |
//____ MPITypeTraits |
69 |
< |
template<typename T> |
70 |
< |
struct MPITypeTraits; |
68 |
> |
vector<vector<RealType> > pot_row(LR_POT_TYPES, |
69 |
> |
vector<RealType> (nInRow, 0.0)); |
70 |
> |
vector<vector<RealType> > pot_col(LR_POT_TYPES, |
71 |
> |
vector<RealType> (nInCol, 0.0)); |
72 |
|
|
73 |
< |
#ifdef IS_MPI |
74 |
< |
template<> |
79 |
< |
struct MPITypeTraits<RealType> { |
80 |
< |
static const MPI::Datatype datatype; |
81 |
< |
}; |
82 |
< |
const MPI_Datatype MPITypeTraits<RealType>::datatype = MY_MPI_REAL; |
73 |
> |
vector<vector<RealType> > pot_local(LR_POT_TYPES, |
74 |
> |
vector<RealType> (nAtoms, 0.0)); |
75 |
|
|
84 |
– |
template<> |
85 |
– |
struct MPITypeTraits<int> { |
86 |
– |
static const MPI::Datatype datatype; |
87 |
– |
}; |
88 |
– |
const MPI::Datatype MPITypeTraits<int>::datatype = MPI_INT; |
76 |
|
#endif |
77 |
+ |
} |
78 |
+ |
|
79 |
|
|
91 |
– |
/** |
92 |
– |
* Constructor for ForceDecomposition Parallel Decomposition Method |
93 |
– |
* Will try to construct a symmetric grid of processors. Ideally, the |
94 |
– |
* number of processors will be a square ex: 4, 9, 16, 25. |
95 |
– |
* |
96 |
– |
*/ |
80 |
|
|
81 |
< |
ForceDecomposition::ForceDecomposition() { |
99 |
< |
|
81 |
> |
void ForceDecomposition::distributeData() { |
82 |
|
#ifdef IS_MPI |
83 |
< |
int nProcs = MPI::COMM_WORLD.Get_size(); |
84 |
< |
int worldRank = MPI::COMM_WORLD.Get_rank(); |
85 |
< |
#endif |
86 |
< |
|
87 |
< |
// First time through, construct column stride. |
88 |
< |
if (communicators.size() == 0) |
89 |
< |
{ |
90 |
< |
int nColumnsMax = (int) round(sqrt((float) nProcs)); |
91 |
< |
for (int i = 0; i < nProcs; ++i) |
92 |
< |
{ |
93 |
< |
if (nProcs%i==0) nColumns=i; |
83 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
84 |
> |
|
85 |
> |
// gather up the atomic positions |
86 |
> |
AtomCommVectorI->gather(snap->atomData.position, |
87 |
> |
snap->atomIData.position); |
88 |
> |
AtomCommVectorJ->gather(snap->atomData.position, |
89 |
> |
snap->atomJData.position); |
90 |
> |
|
91 |
> |
// gather up the cutoff group positions |
92 |
> |
cgCommVectorI->gather(snap->cgData.position, |
93 |
> |
snap->cgIData.position); |
94 |
> |
cgCommVectorJ->gather(snap->cgData.position, |
95 |
> |
snap->cgJData.position); |
96 |
> |
|
97 |
> |
// if needed, gather the atomic rotation matrices |
98 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslAmat) { |
99 |
> |
AtomCommMatrixI->gather(snap->atomData.aMat, |
100 |
> |
snap->atomIData.aMat); |
101 |
> |
AtomCommMatrixJ->gather(snap->atomData.aMat, |
102 |
> |
snap->atomJData.aMat); |
103 |
|
} |
104 |
+ |
|
105 |
+ |
// if needed, gather the atomic eletrostatic frames |
106 |
+ |
if (snap->atomData.getStorageLayout() & DataStorage::dslElectroFrame) { |
107 |
+ |
AtomCommMatrixI->gather(snap->atomData.electroFrame, |
108 |
+ |
snap->atomIData.electroFrame); |
109 |
+ |
AtomCommMatrixJ->gather(snap->atomData.electroFrame, |
110 |
+ |
snap->atomJData.electroFrame); |
111 |
+ |
} |
112 |
+ |
#endif |
113 |
+ |
} |
114 |
+ |
|
115 |
+ |
void ForceDecomposition::collectIntermediateData() { |
116 |
+ |
#ifdef IS_MPI |
117 |
+ |
Snapshot* snap = sman_->getCurrentSnapshot(); |
118 |
+ |
|
119 |
+ |
if (snap->atomData.getStorageLayout() & DataStorage::dslDensity) { |
120 |
|
|
121 |
< |
int nRows = nProcs/nColumns; |
122 |
< |
myRank_ = (int) worldRank%nColumns; |
121 |
> |
AtomCommRealI->scatter(snap->atomIData.density, |
122 |
> |
snap->atomData.density); |
123 |
> |
|
124 |
> |
int n = snap->atomData.density.size(); |
125 |
> |
std::vector<RealType> rho_tmp(n, 0.0); |
126 |
> |
AtomCommRealJ->scatter(snap->atomJData.density, rho_tmp); |
127 |
> |
for (int i = 0; i < n; i++) |
128 |
> |
snap->atomData.density[i] += rho_tmp[i]; |
129 |
> |
} |
130 |
> |
#endif |
131 |
|
} |
132 |
< |
else |
133 |
< |
{ |
134 |
< |
myRank_ = myRank/nColumns; |
132 |
> |
|
133 |
> |
void ForceDecomposition::distributeIntermediateData() { |
134 |
> |
#ifdef IS_MPI |
135 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
136 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctional) { |
137 |
> |
AtomCommRealI->gather(snap->atomData.functional, |
138 |
> |
snap->atomIData.functional); |
139 |
> |
AtomCommRealJ->gather(snap->atomData.functional, |
140 |
> |
snap->atomJData.functional); |
141 |
> |
} |
142 |
> |
|
143 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctionalDerivative) { |
144 |
> |
AtomCommRealI->gather(snap->atomData.functionalDerivative, |
145 |
> |
snap->atomIData.functionalDerivative); |
146 |
> |
AtomCommRealJ->gather(snap->atomData.functionalDerivative, |
147 |
> |
snap->atomJData.functionalDerivative); |
148 |
> |
} |
149 |
> |
#endif |
150 |
|
} |
121 |
– |
MPI::Comm newComm = MPI:COMM_WORLD.Split(myRank_,0); |
151 |
|
|
123 |
– |
isColumn_ = false; |
152 |
|
|
153 |
< |
} |
153 |
> |
void ForceDecomposition::collectData() { |
154 |
> |
#ifdef IS_MPI |
155 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
156 |
> |
|
157 |
> |
int n = snap->atomData.force.size(); |
158 |
> |
std::vector<Vector3d> frc_tmp(n, 0.0); |
159 |
> |
|
160 |
> |
AtomCommVectorI->scatter(snap->atomIData.force, frc_tmp); |
161 |
> |
for (int i = 0; i < n; i++) { |
162 |
> |
snap->atomData.force[i] += frc_tmp[i]; |
163 |
> |
frc_tmp[i] = 0.0; |
164 |
> |
} |
165 |
> |
|
166 |
> |
AtomCommVectorJ->scatter(snap->atomJData.force, frc_tmp); |
167 |
> |
for (int i = 0; i < n; i++) |
168 |
> |
snap->atomData.force[i] += frc_tmp[i]; |
169 |
> |
|
170 |
> |
|
171 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslTorque) { |
172 |
|
|
173 |
< |
ForceDecomposition::gather(sendbuf, receivebuf){ |
174 |
< |
communicators(myIndex_).Allgatherv(); |
129 |
< |
} |
173 |
> |
int nt = snap->atomData.force.size(); |
174 |
> |
std::vector<Vector3d> trq_tmp(nt, 0.0); |
175 |
|
|
176 |
+ |
AtomCommVectorI->scatter(snap->atomIData.torque, trq_tmp); |
177 |
+ |
for (int i = 0; i < n; i++) { |
178 |
+ |
snap->atomData.torque[i] += trq_tmp[i]; |
179 |
+ |
trq_tmp[i] = 0.0; |
180 |
+ |
} |
181 |
+ |
|
182 |
+ |
AtomCommVectorJ->scatter(snap->atomJData.torque, trq_tmp); |
183 |
+ |
for (int i = 0; i < n; i++) |
184 |
+ |
snap->atomData.torque[i] += trq_tmp[i]; |
185 |
+ |
} |
186 |
+ |
|
187 |
+ |
|
188 |
+ |
vector<vector<RealType> > pot_temp(LR_POT_TYPES, |
189 |
+ |
vector<RealType> (nAtoms, 0.0)); |
190 |
+ |
|
191 |
+ |
for (int i = 0; i < LR_POT_TYPES; i++) { |
192 |
+ |
AtomCommRealI->scatter(pot_row[i], pot_temp[i]); |
193 |
+ |
for (int ii = 0; ii < pot_temp[i].size(); ii++ ) { |
194 |
+ |
pot_local[i] += pot_temp[i][ii]; |
195 |
+ |
} |
196 |
+ |
} |
197 |
+ |
|
198 |
|
|
199 |
|
|
200 |
< |
ForceDecomposition::scatter(sbuffer, rbuffer){ |
201 |
< |
communicators(myIndex_).Reduce_scatter(sbuffer, recevbuf. recvcounts, MPI::DOUBLE, MPI::SUM); |
202 |
< |
} |
203 |
< |
|
137 |
< |
|
200 |
> |
#endif |
201 |
> |
} |
202 |
> |
|
203 |
> |
} //end namespace OpenMD |