| 1 |
/* $Id$ */ |
| 2 |
|
| 3 |
/******************************************************* |
| 4 |
* |
| 5 |
* Copyright 2003-2007 by ACceSS MNRF |
| 6 |
* Copyright 2007 by University of Queensland |
| 7 |
* |
| 8 |
* http://esscc.uq.edu.au |
| 9 |
* Primary Business: Queensland, Australia |
| 10 |
* Licensed under the Open Software License version 3.0 |
| 11 |
* http://www.opensource.org/licenses/osl-3.0.php |
| 12 |
* |
| 13 |
*******************************************************/ |
| 14 |
|
| 15 |
/************************************************************************/ |
| 16 |
|
| 17 |
/* Finley: Mesh: optimizes the labeling of the DOFs on each processor */ |
| 18 |
|
| 19 |
/************************************************************************/ |
| 20 |
|
| 21 |
#include "Mesh.h" |
| 22 |
#include "IndexList.h" |
| 23 |
|
| 24 |
/**************************************************************/ |
| 25 |
|
| 26 |
void Finley_Mesh_optimizeDOFLabeling(Finley_Mesh* in,dim_t *distribution) { |
| 27 |
|
| 28 |
index_t myFirstVertex,myLastVertex, *newGlobalDOFID=NULL, firstVertex, lastVertex; |
| 29 |
register index_t k; |
| 30 |
dim_t mpiSize, myNumVertices,len, p, i; |
| 31 |
Paso_Pattern *pattern=NULL; |
| 32 |
Paso_MPI_rank myRank,dest,source,current_rank; |
| 33 |
Finley_IndexList* index_list=NULL; |
| 34 |
#ifdef PASO_MPI |
| 35 |
MPI_Status status; |
| 36 |
#endif |
| 37 |
|
| 38 |
if (in==NULL) return; |
| 39 |
if (in->Nodes == NULL) return; |
| 40 |
|
| 41 |
myRank=in->MPIInfo->rank; |
| 42 |
mpiSize=in->MPIInfo->size; |
| 43 |
myFirstVertex=distribution[myRank]; |
| 44 |
myLastVertex=distribution[myRank+1]; |
| 45 |
myNumVertices=myLastVertex-myFirstVertex; |
| 46 |
len=0; |
| 47 |
for (p=0;p<mpiSize;++p) len=MAX(len,distribution[p+1]-distribution[p]); |
| 48 |
|
| 49 |
index_list=TMPMEMALLOC(myNumVertices,Finley_IndexList); |
| 50 |
newGlobalDOFID=TMPMEMALLOC(len,index_t); |
| 51 |
/* create the adjacency structure xadj and adjncy */ |
| 52 |
if (! ( Finley_checkPtr(index_list) || Finley_checkPtr(newGlobalDOFID) ) ) { |
| 53 |
#pragma omp parallel private(i) |
| 54 |
{ |
| 55 |
#pragma omp for schedule(static) |
| 56 |
for(i=0;i<myNumVertices;++i) { |
| 57 |
index_list[i].extension=NULL; |
| 58 |
index_list[i].n=0; |
| 59 |
} |
| 60 |
/* insert contributions from element matrices into colums index index_list: */ |
| 61 |
Finley_IndexList_insertElementsWithRowRangeNoMainDiagonal(index_list, myFirstVertex, myLastVertex, |
| 62 |
in->Elements,in->Nodes->globalDegreesOfFreedom, |
| 63 |
in->Nodes->globalDegreesOfFreedom); |
| 64 |
Finley_IndexList_insertElementsWithRowRangeNoMainDiagonal(index_list, myFirstVertex, myLastVertex, |
| 65 |
in->FaceElements,in->Nodes->globalDegreesOfFreedom, |
| 66 |
in->Nodes->globalDegreesOfFreedom); |
| 67 |
Finley_IndexList_insertElementsWithRowRangeNoMainDiagonal(index_list, myFirstVertex, myLastVertex, |
| 68 |
in->ContactElements,in->Nodes->globalDegreesOfFreedom, |
| 69 |
in->Nodes->globalDegreesOfFreedom); |
| 70 |
Finley_IndexList_insertElementsWithRowRangeNoMainDiagonal(index_list, myFirstVertex, myLastVertex, |
| 71 |
in->Points,in->Nodes->globalDegreesOfFreedom, |
| 72 |
in->Nodes->globalDegreesOfFreedom); |
| 73 |
} |
| 74 |
/* create the local matrix pattern */ |
| 75 |
pattern=Finley_IndexList_createPattern(0,myNumVertices,index_list,myFirstVertex, myLastVertex,-myFirstVertex); |
| 76 |
|
| 77 |
/* clean up index list */ |
| 78 |
if (index_list!=NULL) { |
| 79 |
#pragma omp parallel for private(i) |
| 80 |
for(i=0;i<myNumVertices;++i) Finley_IndexList_free(index_list[i].extension); |
| 81 |
} |
| 82 |
|
| 83 |
if (Finley_noError()) Paso_Pattern_reduceBandwidth(pattern,newGlobalDOFID); |
| 84 |
|
| 85 |
Paso_Pattern_free(pattern); |
| 86 |
} |
| 87 |
Paso_MPIInfo_noError(in->MPIInfo); |
| 88 |
if (Finley_noError()) { |
| 89 |
/* shift new labeling to create a global id */ |
| 90 |
#pragma omp parallel for private(i) |
| 91 |
for (i=0;i<myNumVertices;++i) newGlobalDOFID[i]+=myFirstVertex; |
| 92 |
|
| 93 |
|
| 94 |
/* distribute new labeling to other processors */ |
| 95 |
dest=Paso_MPIInfo_mod(mpiSize, myRank + 1); |
| 96 |
source=Paso_MPIInfo_mod(mpiSize, myRank - 1); |
| 97 |
current_rank=myRank; |
| 98 |
for (p=0; p< mpiSize; ++p) { |
| 99 |
firstVertex=distribution[current_rank]; |
| 100 |
lastVertex=distribution[current_rank+1]; |
| 101 |
#pragma omp parallel for private(i,k) |
| 102 |
for (i=0;i<in->Nodes->numNodes;++i) { |
| 103 |
k=in->Nodes->globalDegreesOfFreedom[i]; |
| 104 |
if ( (firstVertex<=k) && (k<lastVertex)) { |
| 105 |
in->Nodes->globalDegreesOfFreedom[i]=newGlobalDOFID[k-firstVertex]; |
| 106 |
} |
| 107 |
} |
| 108 |
|
| 109 |
if (p<mpiSize-1) { /* the final send can be skipped */ |
| 110 |
#ifdef PASO_MPI |
| 111 |
MPI_Sendrecv_replace(newGlobalDOFID,len, MPI_INT, |
| 112 |
dest, in->MPIInfo->msg_tag_counter, |
| 113 |
source, in->MPIInfo->msg_tag_counter, |
| 114 |
in->MPIInfo->comm,&status); |
| 115 |
#endif |
| 116 |
in->MPIInfo->msg_tag_counter++; |
| 117 |
current_rank=Paso_MPIInfo_mod(mpiSize, current_rank-1); |
| 118 |
} |
| 119 |
} |
| 120 |
} |
| 121 |
TMPMEMFREE(index_list); |
| 122 |
TMPMEMFREE(newGlobalDOFID); |
| 123 |
#if 0 |
| 124 |
for (i=0;i<in->Nodes->numNodes;++i) printf("%d ",in->Nodes->globalDegreesOfFreedom[i]); |
| 125 |
printf("\n"); |
| 126 |
#endif |
| 127 |
return; |
| 128 |
} |
Parent Directory
| 
Revision Log