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 |
+ |
namespace OpenMD { |
46 |
|
|
47 |
< |
|
51 |
< |
/* -*- c++ -*- */ |
52 |
< |
#include "config.h" |
53 |
< |
#include <stdlib.h> |
47 |
> |
void ForceDecomposition::distributeInitialData() { |
48 |
|
#ifdef IS_MPI |
55 |
– |
#include <mpi.h> |
56 |
– |
#endif |
49 |
|
|
50 |
< |
#include <iostream> |
51 |
< |
#include <vector> |
60 |
< |
#include <algorithm> |
61 |
< |
#include <cmath> |
62 |
< |
#include "parallel/ForceDecomposition.hpp" |
50 |
> |
int nAtoms; |
51 |
> |
int nGroups; |
52 |
|
|
53 |
+ |
AtomCommRealI = new Communicator<Row,RealType>(nAtoms); |
54 |
+ |
AtomCommVectorI = new Communicator<Row,Vector3d>(nAtoms); |
55 |
+ |
AtomCommMatrixI = new Communicator<Row,Mat3x3d>(nAtoms); |
56 |
|
|
57 |
< |
using namespace std; |
58 |
< |
using namespace OpenMD; |
57 |
> |
AtomCommRealJ = new Communicator<Column,RealType>(nAtoms); |
58 |
> |
AtomCommVectorJ = new Communicator<Column,Vector3d>(nAtoms); |
59 |
> |
AtomCommMatrixJ = new Communicator<Column,Mat3x3d>(nAtoms); |
60 |
|
|
61 |
< |
//__static |
62 |
< |
#ifdef IS_MPI |
63 |
< |
static vector<MPI:Comm> communictors; |
61 |
> |
cgCommVectorI = new Communicator<Row,Vector3d>(nGroups); |
62 |
> |
cgCommVectorJ = new Communicator<Column,Vector3d>(nGroups); |
63 |
> |
// more to come |
64 |
|
#endif |
65 |
+ |
} |
66 |
+ |
|
67 |
|
|
73 |
– |
//____ MPITypeTraits |
74 |
– |
template<typename T> |
75 |
– |
struct MPITypeTraits; |
68 |
|
|
69 |
+ |
void ForceDecomposition::distributeData() { |
70 |
|
#ifdef IS_MPI |
71 |
< |
template<> |
72 |
< |
struct MPITypeTraits<RealType> { |
73 |
< |
static const MPI::Datatype datatype; |
74 |
< |
}; |
75 |
< |
const MPI_Datatype MPITypeTraits<RealType>::datatype = MY_MPI_REAL; |
76 |
< |
|
77 |
< |
template<> |
78 |
< |
struct MPITypeTraits<int> { |
79 |
< |
static const MPI::Datatype datatype; |
80 |
< |
}; |
81 |
< |
const MPI::Datatype MPITypeTraits<int>::datatype = MPI_INT; |
82 |
< |
#endif |
83 |
< |
|
84 |
< |
/** |
85 |
< |
* Constructor for ForceDecomposition Parallel Decomposition Method |
86 |
< |
* Will try to construct a symmetric grid of processors. Ideally, the |
87 |
< |
* number of processors will be a square ex: 4, 9, 16, 25. |
88 |
< |
* |
89 |
< |
*/ |
90 |
< |
|
91 |
< |
ForceDecomposition::ForceDecomposition() { |
92 |
< |
|
71 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
72 |
> |
|
73 |
> |
// gather up the atomic positions |
74 |
> |
AtomCommVectorI->gather(snap->atomData.position, |
75 |
> |
snap->atomIData.position); |
76 |
> |
AtomCommVectorJ->gather(snap->atomData.position, |
77 |
> |
snap->atomJData.position); |
78 |
> |
|
79 |
> |
// gather up the cutoff group positions |
80 |
> |
cgCommVectorI->gather(snap->cgData.position, |
81 |
> |
snap->cgIData.position); |
82 |
> |
cgCommVectorJ->gather(snap->cgData.position, |
83 |
> |
snap->cgJData.position); |
84 |
> |
|
85 |
> |
// if needed, gather the atomic rotation matrices |
86 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslAmat) { |
87 |
> |
AtomCommMatrixI->gather(snap->atomData.aMat, |
88 |
> |
snap->atomIData.aMat); |
89 |
> |
AtomCommMatrixJ->gather(snap->atomData.aMat, |
90 |
> |
snap->atomJData.aMat); |
91 |
> |
} |
92 |
> |
|
93 |
> |
// if needed, gather the atomic eletrostatic frames |
94 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslElectroFrame) { |
95 |
> |
AtomCommMatrixI->gather(snap->atomData.electroFrame, |
96 |
> |
snap->atomIData.electroFrame); |
97 |
> |
AtomCommMatrixJ->gather(snap->atomData.electroFrame, |
98 |
> |
snap->atomJData.electroFrame); |
99 |
> |
} |
100 |
> |
#endif |
101 |
> |
} |
102 |
> |
|
103 |
> |
void ForceDecomposition::collectIntermediateData() { |
104 |
|
#ifdef IS_MPI |
105 |
< |
int nProcs = MPI::COMM_WORLD.Get_size(); |
106 |
< |
int worldRank = MPI::COMM_WORLD.Get_rank(); |
105 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
106 |
> |
|
107 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslDensity) { |
108 |
> |
AtomCommRealI->scatter(snap->atomIData.density, |
109 |
> |
snap->atomData.density); |
110 |
> |
std::vector<RealType> rho_tmp; |
111 |
> |
int n = snap->getNumberOfAtoms(); |
112 |
> |
rho_tmp.reserve( n ); |
113 |
> |
AtomCommRealJ->scatter(snap->atomJData.density, rho_tmp); |
114 |
> |
for (int i = 0; i < n; i++) |
115 |
> |
snap->atomData.density[i] += rho_tmp[i]; |
116 |
> |
} |
117 |
|
#endif |
118 |
< |
|
119 |
< |
// First time through, construct column stride. |
120 |
< |
if (communicators.size() == 0) |
121 |
< |
{ |
122 |
< |
int nColumnsMax = (int) round(sqrt((float) nProcs)); |
123 |
< |
for (int i = 0; i < nProcs; ++i) |
124 |
< |
{ |
125 |
< |
if (nProcs%i==0) nColumns=i; |
118 |
> |
} |
119 |
> |
|
120 |
> |
void ForceDecomposition::distributeIntermediateData() { |
121 |
> |
#ifdef IS_MPI |
122 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
123 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctional) { |
124 |
> |
AtomCommRealI->gather(snap->atomData.functional, |
125 |
> |
snap->atomIData.functional); |
126 |
> |
AtomCommRealJ->gather(snap->atomData.functional, |
127 |
> |
snap->atomJData.functional); |
128 |
|
} |
129 |
< |
|
130 |
< |
int nRows = nProcs/nColumns; |
131 |
< |
myRank_ = (int) worldRank%nColumns; |
129 |
> |
|
130 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctionalDerivative) { |
131 |
> |
AtomCommRealI->gather(snap->atomData.functionalDerivative, |
132 |
> |
snap->atomIData.functionalDerivative); |
133 |
> |
AtomCommRealJ->gather(snap->atomData.functionalDerivative, |
134 |
> |
snap->atomJData.functionalDerivative); |
135 |
> |
} |
136 |
> |
#endif |
137 |
|
} |
117 |
– |
else |
118 |
– |
{ |
119 |
– |
myRank_ = myRank/nColumns; |
120 |
– |
} |
121 |
– |
MPI::Comm newComm = MPI:COMM_WORLD.Split(myRank_,0); |
138 |
|
|
123 |
– |
isColumn_ = false; |
139 |
|
|
140 |
< |
} |
140 |
> |
void ForceDecomposition::collectData() { |
141 |
> |
#ifdef IS_MPI |
142 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
143 |
> |
int n = snap->getNumberOfAtoms(); |
144 |
|
|
145 |
< |
ForceDecomposition::gather(sendbuf, receivebuf){ |
146 |
< |
communicators(myIndex_).Allgatherv(); |
147 |
< |
} |
145 |
> |
std::vector<Vector3d> frc_tmp; |
146 |
> |
frc_tmp.reserve( n ); |
147 |
> |
|
148 |
> |
AtomCommVectorI->scatter(snap->atomIData.force, frc_tmp); |
149 |
> |
for (int i = 0; i < n; i++) |
150 |
> |
snap->atomData.force[i] += frc_tmp[i]; |
151 |
> |
|
152 |
> |
AtomCommVectorJ->scatter(snap->atomJData.force, frc_tmp); |
153 |
> |
for (int i = 0; i < n; i++) |
154 |
> |
snap->atomData.force[i] += frc_tmp[i]; |
155 |
> |
|
156 |
> |
|
157 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslTorque) { |
158 |
> |
std::vector<Vector3d> trq_tmp; |
159 |
> |
trq_tmp.reserve( n ); |
160 |
> |
|
161 |
> |
AtomCommVectorI->scatter(snap->atomIData.torque, trq_tmp); |
162 |
> |
for (int i = 0; i < n; i++) |
163 |
> |
snap->atomData.torque[i] += trq_tmp[i]; |
164 |
> |
|
165 |
> |
AtomCommVectorJ->scatter(snap->atomJData.torque, trq_tmp); |
166 |
> |
for (int i = 0; i < n; i++) |
167 |
> |
snap->atomData.torque[i] += trq_tmp[i]; |
168 |
> |
} |
169 |
> |
|
170 |
> |
// Still need pot! |
171 |
> |
|
172 |
|
|
173 |
|
|
174 |
< |
|
175 |
< |
ForceDecomposition::scatter(sbuffer, rbuffer){ |
176 |
< |
communicators(myIndex_).Reduce_scatter(sbuffer, recevbuf. recvcounts, MPI::DOUBLE, MPI::SUM); |
177 |
< |
} |
136 |
< |
|
137 |
< |
|
174 |
> |
#endif |
175 |
> |
} |
176 |
> |
|
177 |
> |
} //end namespace OpenMD |