LBPM/common/Communication.hpp
2021-03-24 21:32:08 -04:00

409 lines
15 KiB
C++

/*
Copyright 2013--2018 James E. McClure, Virginia Polytechnic & State University
Copyright Equnior ASA
This file is part of the Open Porous Media project (OPM).
OPM is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
OPM is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with OPM. If not, see <http://www.gnu.org/licenses/>.
*/
/*
Copyright 2013--2018 James E. McClure, Virginia Polytechnic & State University
Copyright Equnior ASA
This file is part of the Open Porous Media project (OPM).
OPM is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
OPM is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with OPM. If not, see <http://www.gnu.org/licenses/>.
*/
#ifndef COMMUNICATION_HPP_INC
#define COMMUNICATION_HPP_INC
#include "common/Communication.h"
#include "common/MPI.h"
#include "common/Utilities.h"
/********************************************************
* Redistribute data between two grids *
********************************************************/
template<class TYPE>
Array<TYPE> redistribute( const RankInfoStruct& src_rank, const Array<TYPE>& src_data,
const RankInfoStruct& dst_rank, std::array<int,3> dst_size, const Utilities::MPI& comm )
{
if ( comm.getSize() == 1 ) {
return src_data.subset( { 0, (size_t) dst_size[0]-1, 0, (size_t) dst_size[1]-1, 0, (size_t) dst_size[2]-1 } );
}
// Get the src size
std::array<int,3> src_size;
int size0[3] = { (int) src_data.size(0), (int) src_data.size(1), (int) src_data.size(2) };
comm.maxReduce( size0, src_size.data(), 3 );
if ( !src_data.empty() )
ASSERT( src_size[0] == size0[0] && src_size[1] == size0[1] && src_size[2] == size0[2] );
// Check that dst_size matches on all ranks
comm.maxReduce( dst_size.data(), size0, 3 );
ASSERT( dst_size[0] == size0[0] && dst_size[1] == size0[1] && dst_size[2] == size0[2] );
// Function to get overlap range
auto calcOverlap = []( int i1[3], int i2[3], int j1[3], int j2[3] ) {
std::vector<size_t> index;
if ( i1[0] > j2[0] || i2[0] < j1[0] || i1[1] > j2[1] || i2[1] < j1[1] || i1[2] > j2[2] || i2[2] < j1[2] )
return index;
index.resize( 6 );
index[0] = std::max( j1[0] - i1[0], 0 );
index[1] = std::min( j2[0] - i1[0], i2[0] - i1[0] );
index[2] = std::max( j1[1] - i1[1], 0 );
index[3] = std::min( j2[1] - i1[1], i2[1] - i1[1] );
index[4] = std::max( j1[2] - i1[2], 0 );
index[5] = std::min( j2[2] - i1[2], i2[2] - i1[2] );
return index;
};
// Pack and send my data to the appropriate ranks (including myself)
std::vector<int> send_rank;
std::vector<Array<TYPE>> send_data;
if ( !src_data.empty() ) {
int i1[3] = { src_size[0] * src_rank.ix, src_size[1] * src_rank.jy, src_size[2] * src_rank.kz };
int i2[3] = { i1[0] + src_size[0] - 1, i1[1] + src_size[1] - 1, i1[2] + src_size[2] - 1 };
for ( int i=0; i<dst_rank.nx; i++ ) {
for ( int j=0; j<dst_rank.ny; j++ ) {
for ( int k=0; k<dst_rank.nz; k++ ) {
int j1[3] = { i * dst_size[0], j * dst_size[1], k * dst_size[2] };
int j2[3] = { j1[0] + dst_size[0] - 1, j1[1] + dst_size[1] - 1, j1[2] + dst_size[2] - 1 };
auto index = calcOverlap( i1, i2, j1, j2 );
if ( index.empty() )
continue;
send_rank.push_back( dst_rank.getRankForBlock(i,j,k) );
send_data.push_back( src_data.subset( index ) );
}
}
}
}
std::vector<MPI_Request> send_request( send_rank.size() );
for (size_t i=0; i<send_rank.size(); i++)
send_request[i] = comm.Isend( send_data[i].data(), send_data[i].length(), send_rank[i], 5462 );
// Unpack data from the appropriate ranks (including myself)
Array<TYPE> dst_data( dst_size[0], dst_size[1], dst_size[2] );
int i1[3] = { dst_size[0] * dst_rank.ix, dst_size[1] * dst_rank.jy, dst_size[2] * dst_rank.kz };
int i2[3] = { i1[0] + dst_size[0] - 1, i1[1] + dst_size[1] - 1, i1[2] + dst_size[2] - 1 };
for ( int i=0; i<src_rank.nx; i++ ) {
for ( int j=0; j<src_rank.ny; j++ ) {
for ( int k=0; k<src_rank.nz; k++ ) {
int j1[3] = { i * src_size[0], j * src_size[1], k * src_size[2] };
int j2[3] = { j1[0] + src_size[0] - 1, j1[1] + src_size[1] - 1, j1[2] + src_size[2] - 1 };
auto index = calcOverlap( i1, i2, j1, j2 );
if ( index.empty() )
continue;
int rank = src_rank.getRankForBlock(i,j,k);
Array<TYPE> data( index[1] - index[0] + 1, index[3] - index[2] + 1, index[5] - index[4] + 1 );
comm.recv( data.data(), data.length(), rank, 5462 );
dst_data.copySubset( index, data );
}
}
}
// Free data
comm.waitAll( send_request.size(), send_request.data() );
return dst_data;
}
/********************************************************
* Structure to fill halo cells *
********************************************************/
template<class TYPE>
fillHalo<TYPE>::fillHalo( const Utilities::MPI& comm_, const RankInfoStruct& info_,
std::array<int,3> n_, std::array<int,3> ng_, int tag0, int depth_,
std::array<bool,3> fill, std::array<bool,3> periodic ):
comm(comm_), info(info_), n(n_), ng(ng_), depth(depth_)
{
// Set the fill pattern
memset(fill_pattern,0,sizeof(fill_pattern));
if ( fill[0] ) {
fill_pattern[0][1][1] = true;
fill_pattern[2][1][1] = true;
fill_pattern[1][0][1] = true;
fill_pattern[1][2][1] = true;
fill_pattern[1][1][0] = true;
fill_pattern[1][1][2] = true;
}
if ( fill[1] ) {
fill_pattern[0][0][1] = true;
fill_pattern[0][2][1] = true;
fill_pattern[2][0][1] = true;
fill_pattern[2][2][1] = true;
fill_pattern[0][1][0] = true;
fill_pattern[0][1][2] = true;
fill_pattern[2][1][0] = true;
fill_pattern[2][1][2] = true;
fill_pattern[1][0][0] = true;
fill_pattern[1][0][2] = true;
fill_pattern[1][2][0] = true;
fill_pattern[1][2][2] = true;
}
if ( fill[2] ) {
fill_pattern[0][0][0] = true;
fill_pattern[0][0][2] = true;
fill_pattern[0][2][0] = true;
fill_pattern[0][2][2] = true;
fill_pattern[2][0][0] = true;
fill_pattern[2][0][2] = true;
fill_pattern[2][2][0] = true;
fill_pattern[2][2][2] = true;
}
// Remove communication for non-perioidic directions
if ( !periodic[0] && info.ix==0 ) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++)
fill_pattern[0][j][k] = false;
}
}
if ( !periodic[0] && info.ix==info.nx-1 ) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++)
fill_pattern[2][j][k] = false;
}
}
if ( !periodic[1] && info.jy==0 ) {
for (int i=0; i<3; i++) {
for (int k=0; k<3; k++)
fill_pattern[i][0][k] = false;
}
}
if ( !periodic[1] && info.jy==info.ny-1 ) {
for (int i=0; i<3; i++) {
for (int k=0; k<3; k++)
fill_pattern[i][2][k] = false;
}
}
if ( !periodic[2] && info.kz==0 ) {
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++)
fill_pattern[i][j][0] = false;
}
}
if ( !periodic[2] && info.kz==info.nz-1 ) {
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++)
fill_pattern[i][j][2] = false;
}
}
// Determine the number of elements for each send/recv
for (int i=0; i<3; i++) {
int ni = (i-1)==0 ? n[0]:ng[0];
for (int j=0; j<3; j++) {
int nj = (j-1)==0 ? n[1]:ng[1];
for (int k=0; k<3; k++) {
int nk = (k-1)==0 ? n[2]:ng[2];
if ( fill_pattern[i][j][k] )
N_send_recv[i][j][k] = ni*nj*nk;
else
N_send_recv[i][j][k] = 0;
}
}
}
// Create send/recv buffers
size_t N_mem=0;
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++)
N_mem += N_send_recv[i][j][k];
}
}
mem = new TYPE[2*depth*N_mem];
size_t index = 0;
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++) {
send[i][j][k] = &mem[index];
index += depth*N_send_recv[i][j][k];
recv[i][j][k] = &mem[index];
index += depth*N_send_recv[i][j][k];
}
}
}
// Create the tags
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++) {
tag[i][j][k] = tag0 + i + j*3 + k*9;
}
}
}
}
template<class TYPE>
fillHalo<TYPE>::~fillHalo( )
{
delete [] mem;
}
template<class TYPE>
void fillHalo<TYPE>::fill( Array<TYPE>& data )
{
//PROFILE_START("fillHalo::fill",1);
int depth2 = data.size(3);
ASSERT((int)data.size(0)==n[0]+2*ng[0]);
ASSERT((int)data.size(1)==n[1]+2*ng[1]);
ASSERT((int)data.size(2)==n[2]+2*ng[2]);
ASSERT(depth2<=depth);
ASSERT(data.ndim()==3||data.ndim()==4);
// Start the recieves
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++) {
if ( !fill_pattern[i][j][k] )
continue;
recv_req[i][j][k] = comm.Irecv( recv[i][j][k], depth2*N_send_recv[i][j][k],
info.rank[i][j][k], tag[2-i][2-j][2-k] );
}
}
}
// Pack the src data and start the sends
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++) {
if ( !fill_pattern[i][j][k] )
continue;
pack( data, i-1, j-1, k-1, send[i][j][k] );
send_req[i][j][k] = comm.Isend( send[i][j][k], depth2*N_send_recv[i][j][k],
info.rank[i][j][k], tag[i][j][k] );
}
}
}
// Recv the dst data and unpack (we recive in reverse order to match the sends)
for (int i=2; i>=0; i--) {
for (int j=2; j>=0; j--) {
for (int k=2; k>=0; k--) {
if ( !fill_pattern[i][j][k] )
continue;
comm.wait( recv_req[i][j][k] );
unpack( data, i-1, j-1, k-1, recv[i][j][k] );
}
}
}
// Wait until all sends have completed
for (int i=0; i<3; i++) {
for (int j=0; j<3; j++) {
for (int k=0; k<3; k++) {
if ( !fill_pattern[i][j][k] )
continue;
comm.wait( send_req[i][j][k] );
}
}
}
//PROFILE_STOP("fillHalo::fill",1);
}
template<class TYPE>
void fillHalo<TYPE>::pack( const Array<TYPE>& data, int i0, int j0, int k0, TYPE *buffer )
{
int depth2 = data.size(3);
int ni = i0==0 ? n[0]:ng[0];
int nj = j0==0 ? n[1]:ng[1];
int nk = k0==0 ? n[2]:ng[2];
int is = i0==0 ? ng[0]:((i0==-1)?ng[0]:n[0]);
int js = j0==0 ? ng[1]:((j0==-1)?ng[1]:n[1]);
int ks = k0==0 ? ng[2]:((k0==-1)?ng[2]:n[2]);
for (int d=0; d<depth2; d++) {
for (int k=0; k<nk; k++) {
for (int j=0; j<nj; j++) {
for (int i=0; i<ni; i++) {
buffer[i+j*ni+k*ni*nj+d*ni*nj*nk] = data(i+is,j+js,k+ks,d);
}
}
}
}
}
template<class TYPE>
void fillHalo<TYPE>::unpack( Array<TYPE>& data, int i0, int j0, int k0, const TYPE *buffer )
{
int depth2 = data.size(3);
int ni = i0==0 ? n[0]:ng[0];
int nj = j0==0 ? n[1]:ng[1];
int nk = k0==0 ? n[2]:ng[2];
int is = i0==0 ? ng[0]:((i0==-1)?0:n[0]+ng[0]);
int js = j0==0 ? ng[1]:((j0==-1)?0:n[1]+ng[1]);
int ks = k0==0 ? ng[2]:((k0==-1)?0:n[2]+ng[2]);
for (int d=0; d<depth2; d++) {
for (int k=0; k<nk; k++) {
for (int j=0; j<nj; j++) {
for (int i=0; i<ni; i++) {
data(i+is,j+js,k+ks,d) = buffer[i+j*ni+k*ni*nj+d*ni*nj*nk];
}
}
}
}
}
/********************************************************
* Function to remove the ghost halo *
********************************************************/
template<class TYPE>
template<class TYPE1, class TYPE2>
void fillHalo<TYPE>::copy( const Array<TYPE1>& src, Array<TYPE2>& dst )
{
//PROFILE_START("fillHalo::copy",1);
ASSERT( (int)src.size(0)==n[0] || (int)src.size(0)==n[0]+2*ng[0] );
ASSERT( (int)dst.size(0)==n[0] || (int)dst.size(0)==n[0]+2*ng[0] );
bool src_halo = (int)src.size(0)==n[0]+2*ng[0];
bool dst_halo = (int)dst.size(0)==n[0]+2*ng[0];
if ( src_halo ) {
ASSERT((int)src.size(0)==n[0]+2*ng[0]);
ASSERT((int)src.size(1)==n[1]+2*ng[1]);
ASSERT((int)src.size(2)==n[2]+2*ng[2]);
} else {
ASSERT((int)src.size(0)==n[0]);
ASSERT((int)src.size(1)==n[1]);
ASSERT((int)src.size(2)==n[2]);
}
if ( dst_halo ) {
ASSERT((int)dst.size(0)==n[0]+2*ng[0]);
ASSERT((int)dst.size(1)==n[1]+2*ng[1]);
ASSERT((int)dst.size(2)==n[2]+2*ng[2]);
} else {
ASSERT((int)dst.size(0)==n[0]);
ASSERT((int)dst.size(1)==n[1]);
ASSERT((int)dst.size(2)==n[2]);
}
if ( src_halo == dst_halo ) {
// Src and dst halos match
for (size_t i=0; i<src.length(); i++)
dst(i) = src(i);
} else if ( src_halo && !dst_halo ) {
// Src has halos
for (int k=0; k<n[2]; k++) {
for (int j=0; j<n[1]; j++) {
for (int i=0; i<n[0]; i++) {
dst(i,j,k) = src(i+ng[0],j+ng[1],k+ng[2]);
}
}
}
} else if ( !src_halo && dst_halo ) {
// Dst has halos
for (int k=0; k<n[2]; k++) {
for (int j=0; j<n[1]; j++) {
for (int i=0; i<n[0]; i++) {
dst(i+ng[0],j+ng[1],k+ng[2]) = src(i,j,k);
}
}
}
fill(dst);
}
//PROFILE_STOP("fillHalo::copy",1);
}
#endif