| 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 Comm<I,RealType>(nAtoms); |
| 54 |
+ |
AtomCommVectorI = new Comm<I,Vector3d>(nAtoms); |
| 55 |
+ |
AtomCommMatrixI = new Comm<I,Mat3x3d>(nAtoms); |
| 56 |
|
|
| 57 |
< |
using namespace std; |
| 58 |
< |
using namespace OpenMD; |
| 57 |
> |
AtomCommRealJ = new Comm<J,RealType>(nAtoms); |
| 58 |
> |
AtomCommVectorJ = new Comm<J,Vector3d>(nAtoms); |
| 59 |
> |
AtomCommMatrixJ = new Comm<J,Mat3x3d>(nAtoms); |
| 60 |
|
|
| 61 |
< |
//__static |
| 62 |
< |
#ifdef IS_MPI |
| 63 |
< |
static vector<MPI:Comm> communictors; |
| 61 |
> |
cgCommVectorI = new Comm<I,Vector3d>(nGroups); |
| 62 |
> |
cgCommVectorJ = new Comm<J,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<> |
| 79 |
< |
struct MPITypeTraits<RealType> { |
| 80 |
< |
static const MPI::Datatype datatype; |
| 81 |
< |
}; |
| 82 |
< |
const MPI_Datatype MPITypeTraits<RealType>::datatype = MY_MPI_REAL; |
| 71 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
| 72 |
|
|
| 73 |
< |
template<> |
| 74 |
< |
struct MPITypeTraits<int> { |
| 75 |
< |
static const MPI::Datatype datatype; |
| 76 |
< |
}; |
| 77 |
< |
const MPI::Datatype MPITypeTraits<int>::datatype = MPI_INT; |
| 78 |
< |
#endif |
| 79 |
< |
|
| 80 |
< |
/** |
| 81 |
< |
* Constructor for ForceDecomposition Parallel Decomposition Method |
| 82 |
< |
* Will try to construct a symmetric grid of processors. Ideally, the |
| 83 |
< |
* number of processors will be a square ex: 4, 9, 16, 25. |
| 84 |
< |
* |
| 85 |
< |
*/ |
| 86 |
< |
|
| 87 |
< |
ForceDecomposition::ForceDecomposition() { |
| 88 |
< |
|
| 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 |
> |
// gather up the atomic positions |
| 107 |
> |
|
| 108 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslDensity) { |
| 109 |
> |
AtomCommRealI->scatter(snap->atomIData.density, |
| 110 |
> |
snap->atomData.density); |
| 111 |
> |
std::vector<RealType> rho_tmp; |
| 112 |
> |
int n = snap->getNumberOfAtoms(); |
| 113 |
> |
rho_tmp.reserve( n ); |
| 114 |
> |
AtomCommRealJ->scatter(snap->atomJData.density, rho_tmp); |
| 115 |
> |
for (int i = 0; i < n; i++) |
| 116 |
> |
snap->atomData.density[i] += rho_tmp[i]; |
| 117 |
> |
} |
| 118 |
|
#endif |
| 119 |
< |
|
| 120 |
< |
// First time through, construct column stride. |
| 121 |
< |
if (communicators.size() == 0) |
| 122 |
< |
{ |
| 123 |
< |
int nColumnsMax = (int) round(sqrt((float) nProcs)); |
| 124 |
< |
for (int i = 0; i < nProcs; ++i) |
| 125 |
< |
{ |
| 126 |
< |
if (nProcs%i==0) nColumns=i; |
| 119 |
> |
} |
| 120 |
> |
|
| 121 |
> |
void ForceDecomposition::distributeIntermediateData() { |
| 122 |
> |
#ifdef IS_MPI |
| 123 |
> |
Snapshot* snap = sman_->getCurrentSnapshot(); |
| 124 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctional) { |
| 125 |
> |
AtomCommRealI->gather(snap->atomData.functional, |
| 126 |
> |
snap->atomIData.functional); |
| 127 |
> |
AtomCommRealJ->gather(snap->atomData.functional, |
| 128 |
> |
snap->atomJData.functional); |
| 129 |
|
} |
| 130 |
< |
|
| 131 |
< |
int nRows = nProcs/nColumns; |
| 132 |
< |
myRank_ = (int) worldRank%nColumns; |
| 130 |
> |
|
| 131 |
> |
if (snap->atomData.getStorageLayout() & DataStorage::dslFunctionalDerivative) { |
| 132 |
> |
AtomCommRealI->gather(snap->atomData.functionalDerivative, |
| 133 |
> |
snap->atomIData.functionalDerivative); |
| 134 |
> |
AtomCommRealJ->gather(snap->atomData.functionalDerivative, |
| 135 |
> |
snap->atomJData.functionalDerivative); |
| 136 |
> |
} |
| 137 |
> |
#endif |
| 138 |
|
} |
| 117 |
– |
else |
| 118 |
– |
{ |
| 119 |
– |
myRank_ = myRank/nColumns; |
| 120 |
– |
} |
| 121 |
– |
MPI::Comm newComm = MPI:COMM_WORLD.Split(myRank_,0); |
| 139 |
|
|
| 123 |
– |
isColumn_ = false; |
| 140 |
|
|
| 141 |
< |
} |
| 142 |
< |
|
| 143 |
< |
ForceDecomposition::gather(sendbuf, receivebuf){ |
| 144 |
< |
communicators(myIndex_).Allgatherv(); |
| 145 |
< |
} |
| 146 |
< |
|
| 131 |
< |
|
| 132 |
< |
|
| 133 |
< |
ForceDecomposition::scatter(sbuffer, rbuffer){ |
| 134 |
< |
communicators(myIndex_).Reduce_scatter(sbuffer, recevbuf. recvcounts, MPI::DOUBLE, MPI::SUM); |
| 135 |
< |
} |
| 136 |
< |
|
| 137 |
< |
|
| 141 |
> |
void ForceDecomposition::collectData() { |
| 142 |
> |
#ifdef IS_MPI |
| 143 |
> |
#endif |
| 144 |
> |
} |
| 145 |
> |
|
| 146 |
> |
} //end namespace OpenMD |
