void box_partition_rcb( const int np ,
const int my_p ,
const int root_box[][2] ,
const int ghost ,
int (**pbox)[3][2] ,
int ** map_local_id ,
int ** map_recv_pc ,
int ** map_send_pc ,
int ** map_send_id )
{
*pbox = (int (*)[3][2]) malloc( sizeof(int) * np * 3 * 2 );
box_partition( 0 , np , 2 , root_box , *pbox );
box_partition_maps( np , my_p , (const int (*)[3][2]) *pbox , ghost ,
map_local_id , map_recv_pc ,
map_send_pc , map_send_id );
}
static FEMeshType create( const size_t proc_count ,
const size_t proc_local ,
const size_t gang_count ,
const size_t elems_x ,
const size_t elems_y ,
const size_t elems_z ,
const double x_coord_curve = 1 ,
const double y_coord_curve = 1 ,
const double z_coord_curve = 1 )
{
const size_t vertices_x = elems_x + 1 ;
const size_t vertices_y = elems_y + 1 ;
const size_t vertices_z = elems_z + 1 ;
const BoxBoundsLinear vertex_box_bounds ;
const ElementSpec element ;
// Partition based upon vertices:
BoxType vertex_box_global ;
std::vector< BoxType > vertex_box_parts( proc_count );
vertex_box_global[0][0] = 0 ; vertex_box_global[0][1] = vertices_x ;
vertex_box_global[1][0] = 0 ; vertex_box_global[1][1] = vertices_y ;
vertex_box_global[2][0] = 0 ; vertex_box_global[2][1] = vertices_z ;
box_partition_rcb( vertex_box_global , vertex_box_parts );
const BoxType vertex_box_local_owned = vertex_box_parts[ proc_local ];
// Determine interior and used vertices:
BoxType vertex_box_local_interior ;
BoxType vertex_box_local_used ;
vertex_box_bounds.apply( vertex_box_global ,
vertex_box_local_owned ,
vertex_box_local_interior ,
vertex_box_local_used );
// Element counts:
const long local_elems_x =
( vertex_box_local_used[0][1] - vertex_box_local_used[0][0] ) - 1 ;
const long local_elems_y =
( vertex_box_local_used[1][1] - vertex_box_local_used[1][0] ) - 1 ;
const long local_elems_z =
( vertex_box_local_used[2][1] - vertex_box_local_used[2][0] ) - 1 ;
const size_t elem_count_total = std::max( long(0) , local_elems_x ) *
std::max( long(0) , local_elems_y ) *
std::max( long(0) , local_elems_z );
const long interior_elems_x =
( vertex_box_local_owned[0][1] - vertex_box_local_owned[0][0] ) - 1 ;
const long interior_elems_y =
( vertex_box_local_owned[1][1] - vertex_box_local_owned[1][0] ) - 1 ;
const long interior_elems_z =
( vertex_box_local_owned[2][1] - vertex_box_local_owned[2][0] ) - 1 ;
const size_t elem_count_interior = std::max( long(0) , interior_elems_x ) *
std::max( long(0) , interior_elems_y ) *
std::max( long(0) , interior_elems_z );
// Expand vertex boxes to node boxes:
BoxType node_box_global ;
BoxType node_box_local_used ;
std::vector< BoxType > node_box_parts ;
element.create_node_boxes_from_vertex_boxes(
vertex_box_global , vertex_box_parts ,
node_box_global , node_box_parts );
// Node communication maps:
size_t node_count_interior = 0 ;
size_t node_count_owned = 0 ;
size_t node_count_total = 0 ;
std::vector<size_t> node_used_id_map ;
std::vector<size_t> node_part_counts ;
std::vector< std::vector<size_t> > node_send_map ;
box_partition_maps( node_box_global ,
node_box_parts ,
element.box_bounds ,
proc_local ,
node_box_local_used ,
node_used_id_map ,
node_count_interior ,
node_count_owned ,
node_count_total ,
node_part_counts ,
node_send_map );
size_t node_count_send = 0 ;
for ( size_t i = 0 ; i < node_send_map.size() ; ++i ) {
node_count_send += node_send_map[i].size();
}
size_t recv_msg_count = 0 ;
size_t send_msg_count = 0 ;
size_t send_count = 0 ;
for ( size_t i = 1 ; i < proc_count ; ++i ) {
if ( node_part_counts[i] ) ++recv_msg_count ;
if ( node_send_map[i].size() ) {
++send_msg_count ;
send_count += node_send_map[i].size();
}
}
// Finite element mesh:
FEMeshType mesh ;
if ( node_count_total ) {
mesh.node_coords = node_coords_type( "node_coords", node_count_total );
}
if ( elem_count_total ) {
mesh.elem_node_ids =
elem_node_ids_type( "elem_node_ids", elem_count_total );
}
mesh.parallel_data_map.assign( node_count_interior ,
node_count_owned ,
node_count_total ,
recv_msg_count ,
send_msg_count ,
send_count );
typename node_coords_type::HostMirror node_coords =
Kokkos::create_mirror( mesh.node_coords );
typename elem_node_ids_type::HostMirror elem_node_ids =
Kokkos::create_mirror( mesh.elem_node_ids );
//------------------------------------
// set node coordinates to grid location for subsequent verification
for ( size_t iz = node_box_local_used[2][0] ;
iz < node_box_local_used[2][1] ; ++iz ) {
for ( size_t iy = node_box_local_used[1][0] ;
iy < node_box_local_used[1][1] ; ++iy ) {
for ( size_t ix = node_box_local_used[0][0] ;
ix < node_box_local_used[0][1] ; ++ix ) {
const size_t node_local_id =
box_map_id( node_box_local_used , node_used_id_map , ix , iy , iz );
node_coords( node_local_id , 0 ) = ix ;
node_coords( node_local_id , 1 ) = iy ;
node_coords( node_local_id , 2 ) = iz ;
}}}
//------------------------------------
// Initialize element-node connectivity:
if ( 1 < gang_count ) {
layout_elements_partitioned( vertex_box_local_used ,
vertex_box_local_owned ,
node_box_local_used ,
node_used_id_map ,
element ,
gang_count ,
elem_node_ids );
}
else {
layout_elements_interior_exterior( vertex_box_local_used ,
vertex_box_local_owned ,
node_box_local_used ,
node_used_id_map ,
element ,
elem_count_interior ,
elem_node_ids );
}
//------------------------------------
// Populate node->element connectivity:
std::vector<size_t> node_elem_work( node_count_total , (size_t) 0 );
for ( size_t i = 0 ; i < elem_count_total ; ++i ) {
for ( size_t n = 0 ; n < element_node_count ; ++n ) {
++node_elem_work[ elem_node_ids(i,n) ];
}
}
mesh.node_elem_ids =
Kokkos::create_staticcrsgraph< node_elem_ids_type >( "node_elem_ids" , node_elem_work );
typename node_elem_ids_type::HostMirror
node_elem_ids = Kokkos::create_mirror( mesh.node_elem_ids );
for ( size_t i = 0 ; i < node_count_total ; ++i ) {
node_elem_work[i] = node_elem_ids.row_map[i];
}
// Looping in element order insures the list of elements
// is sorted by element index.
for ( size_t i = 0 ; i < elem_count_total ; ++i ) {
for ( size_t n = 0 ; n < element_node_count ; ++n ) {
const unsigned nid = elem_node_ids(i, n);
const unsigned j = node_elem_work[nid] ; ++node_elem_work[nid] ;
node_elem_ids.entries( j , 0 ) = i ;
node_elem_ids.entries( j , 1 ) = n ;
}
}
//------------------------------------
// Verify setup with node coordinates matching grid indices.
verify( node_coords , elem_node_ids , node_elem_ids );
//------------------------------------
// Scale node coordinates to problem extent with
// nonlinear mapping.
{
const double problem_extent[3] =
{ static_cast<double>( vertex_box_global[0][1] - 1 ) ,
static_cast<double>( vertex_box_global[1][1] - 1 ) ,
static_cast<double>( vertex_box_global[2][1] - 1 ) };
const double grid_extent[3] =
{ static_cast<double>( node_box_global[0][1] - 1 ) ,
static_cast<double>( node_box_global[1][1] - 1 ) ,
static_cast<double>( node_box_global[2][1] - 1 ) };
for ( size_t i = 0 ; i < node_count_total ; ++i ) {
const double x_unit = node_coords(i,0) / grid_extent[0] ;
const double y_unit = node_coords(i,1) / grid_extent[1] ;
const double z_unit = node_coords(i,2) / grid_extent[2] ;
node_coords(i,0) = coordinate_scalar_type( problem_extent[0] * std::pow( x_unit , x_coord_curve ) );
node_coords(i,1) = coordinate_scalar_type( problem_extent[1] * std::pow( y_unit , y_coord_curve ) );
node_coords(i,2) = coordinate_scalar_type( problem_extent[2] * std::pow( z_unit , z_coord_curve ) );
}
}
Kokkos::deep_copy( mesh.node_coords , node_coords );
Kokkos::deep_copy( mesh.elem_node_ids , elem_node_ids );
Kokkos::deep_copy( mesh.node_elem_ids.entries , node_elem_ids.entries );
//------------------------------------
// Communication lists:
{
recv_msg_count = 0 ;
send_msg_count = 0 ;
send_count = 0 ;
for ( size_t i = 1 ; i < proc_count ; ++i ) {
// Order sending starting with the local processor rank
// to try to smooth out the amount of messages simultaneously
// send to a particular processor.
const int proc = ( proc_local + i ) % proc_count ;
if ( node_part_counts[i] ) {
mesh.parallel_data_map.host_recv(recv_msg_count,0) = proc ;
mesh.parallel_data_map.host_recv(recv_msg_count,1) = node_part_counts[i] ;
++recv_msg_count ;
}
if ( node_send_map[i].size() ) {
mesh.parallel_data_map.host_send(send_msg_count,0) = proc ;
mesh.parallel_data_map.host_send(send_msg_count,1) = node_send_map[i].size() ;
for ( size_t j = 0 ; j < node_send_map[i].size() ; ++j , ++send_count ) {
mesh.parallel_data_map.host_send_item(send_count) = node_send_map[i][j] - node_count_interior ;
}
++send_msg_count ;
}
}
}
return mesh ;
}
static void test_maps( const int root_box[][2] , const int np )
{
const int ghost = 1 ;
const int nx_global = root_box[0][1] - root_box[0][0] ;
const int ny_global = root_box[1][1] - root_box[1][0] ;
int ieq , i , j ;
int (*pbox)[3][2] ;
int **local_values ;
int **map_local_id ;
int **map_recv_pc ;
int **map_send_pc ;
int **map_send_id ;
pbox = (int (*)[3][2]) malloc( sizeof(int) * np * 3 * 2 );
box_partition( 0 , np , 2 , root_box , pbox );
local_values = (int **) malloc( sizeof(int*) * np );
map_local_id = (int **) malloc( sizeof(int*) * np );
map_recv_pc = (int **) malloc( sizeof(int*) * np );
map_send_pc = (int **) malloc( sizeof(int*) * np );
map_send_id = (int **) malloc( sizeof(int*) * np );
/* Set each local value to the global equation number */
for ( ieq = i = 0 ; i < np ; ++i ) {
const int (*mybox)[2] = (const int (*)[2]) pbox[i] ;
const int nx = mybox[0][1] - mybox[0][0] ;
const int ny = mybox[1][1] - mybox[1][0] ;
const int nz = mybox[2][1] - mybox[2][0] ;
int ix , iy , iz ;
/* Generate the partition maps for this rank */
box_partition_maps( np , i , (const int (*)[3][2]) pbox , ghost ,
& map_local_id[i] , & map_recv_pc[i] ,
& map_send_pc[i] , & map_send_id[i] );
local_values[i] = (int *) malloc( sizeof(int) * map_recv_pc[i][np] );
for ( iz = -ghost ; iz < nz + ghost ; ++iz ) {
for ( iy = -ghost ; iy < ny + ghost ; ++iy ) {
for ( ix = -ghost ; ix < nx + ghost ; ++ix ) {
const int ieq = box_map_local(mybox,ghost,map_local_id[i],ix,iy,iz);
if ( 0 <= ieq ) {
const int ix_global = ix + mybox[0][0] ;
const int iy_global = iy + mybox[1][0] ;
const int iz_global = iz + mybox[2][0] ;
if ( root_box[0][0] <= ix_global && ix_global < root_box[0][1] &&
root_box[1][0] <= iy_global && iy_global < root_box[1][1] &&
root_box[2][0] <= iz_global && iz_global < root_box[2][1] ) {
local_values[i][ ieq ] = ix_global +
iy_global * nx_global +
iz_global * nx_global * ny_global ;
}
else {
local_values[i][ ieq ] = -1 ;
}
}
}
}
}
}
/* Pair-wise compare the local values */
/* i == receiving processor rank */
/* ip == sending processor rank */
/* j == receiving processor data entry for message from 'ip' */
/* jp == sending processor data entry for message to 'i' */
for ( i = 0 ; i < np ; ++i ) {
for ( j = 1 ; j < np ; ++j ) {
const int ip = ( i + j ) % np ;
const int jp = ( i + np - ip ) % np ;
const int nrecv = map_recv_pc[i] [j+1] - map_recv_pc[i] [j] ;
const int nsend = map_send_pc[ip][jp+1] - map_send_pc[ip][jp] ;
int k ;
if ( nrecv != nsend ) {
fprintf(stderr,"P%d recv %d from P%d\n",i,nrecv,ip);
fprintf(stderr,"P%d send %d to P%d\n",ip,nsend,i);
abort();
}
for ( k = 0 ; k < nrecv ; ++k ) {
const int irecv = map_recv_pc[i][j] + k ;
const int isend = map_send_pc[ip][jp] + k ;
const int val_irecv = local_values[i][irecv] ;
const int val_isend = local_values[ip][ map_send_id[ip][isend] ] ;
if ( val_irecv != val_isend ) {
fprintf(stderr,"P%d recv[%d] = %d , from P%d\n",i,k,val_irecv,ip);
fprintf(stderr,"P%d send[%d] = %d , to P%d\n",ip,k,val_isend,i);
abort();
}
}
}
}
for ( i = 0 ; i < np ; ++i ) {
free( map_local_id[i] );
free( map_recv_pc[i] );
free( map_send_pc[i] );
free( map_send_id[i] );
free( local_values[i] );
}
free( map_send_id );
free( map_send_pc );
free( map_recv_pc );
free( map_local_id );
free( local_values );
free( pbox );
}