engine.c

/*******************************************************************************
 * This file is part of SWIFT.
 * Coypright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
 * 
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 * 
 * This program 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 Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 * 
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* Some standard headers. */
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <pthread.h>
#include <math.h>
#include <float.h>
#include <limits.h>
#include <omp.h>
#include <sched.h>

/* MPI headers. */
#ifdef WITH_MPI
    #include <mpi.h>
#endif

/* METIS headers. */
#ifdef HAVE_METIS
    #include <metis.h>
#endif

/* Local headers. */
#include "const.h"
#include "cycle.h"
#include "atomic.h"
#include "timers.h"
#include "const.h"
#include "vector.h"
#include "lock.h"
#include "task.h"
#include "part.h"
#include "debug.h"
#include "space.h"
#include "multipole.h"
#include "cell.h"
#include "queue.h"
#include "scheduler.h"
#include "engine.h"
#include "runner.h"
#include "proxy.h"
#include "error.h"

#ifdef LEGACY_GADGET2_SPH
#include "runner_iact_legacy.h"
#else
#include "runner_iact.h"
#endif


/* Convert cell location to ID. */
#define cell_getid( cdim , i , j , k ) ( (int)(k) + (cdim)[2]*( (int)(j) + (cdim)[1]*(int)(i) ) )


/** The rank of the engine as a global variable (for messages). */
int engine_rank;


/**
 * @brief Link a density/force task to a cell.
 *
 * @param e The #engine.
 * @param l The #link.
 * @param t The #task.
 *
 * @return The new #link pointer.
 */
 
struct link *engine_addlink( struct engine *e , struct link *l , struct task *t ) {

    struct link *res = &e->links[ atomic_inc( &e->nr_links ) ];
    res->next = l;
    res->t = t;
    return res;

    }


/**
 * @brief Generate the ghost and kick tasks for a hierarchy of cells.
 *
 * @param e The #engine.
 * @param c The #cell.
 * @param super The super #cell.
 */
 
void engine_mkghosts ( struct engine *e , struct cell *c , struct cell *super ) {

    int k;
    struct scheduler *s = &e->sched;

    /* Am I the super-cell? */
    if ( super == NULL && c->nr_tasks > 0 ) {
    
        /* Remember me. */
        super = c;
        
        /* Local tasks only... */
        if ( c->nodeID == e->nodeID ) {
        
            /* Generate the ghost task. */
            c->ghost = scheduler_addtask( s , task_type_ghost , task_subtype_none , 0 , 0 , c , NULL , 0 );

            /* Add the kick2 task. */
            c->kick2 = scheduler_addtask( s , task_type_kick2 , task_subtype_none , 0 , 0 , c , NULL , 0 );

            /* Add the kick1 task if needed. */
            if ( !(e->policy & engine_policy_fixdt) )
                c->kick1 = scheduler_addtask( s , task_type_kick1 , task_subtype_none , 0 , 0 , c , NULL , 0 );
                
            }
            
        }
        
    /* Set the super-cell. */
    c->super = super;
        
    /* Recurse. */
    if ( c->split )
        for ( k = 0 ; k < 8 ; k++ )
            if ( c->progeny[k] != NULL )
                engine_mkghosts( e , c->progeny[k] , super );
    
    }


/**
 * @brief Redistribute the particles amongst the nodes accorind
 *      to their cell's node IDs.
 *
 * @param e The #engine.
 */
 
void engine_redistribute ( struct engine *e ) {

#ifdef WITH_MPI

    int i, j, k, cid;
    int nr_nodes = e->nr_nodes, nodeID = e->nodeID;
    struct space *s = e->s;
    int my_cells = 0;
    int *cdim = s->cdim;
    struct cell *cells = s->cells;
    int nr_cells = s->nr_cells;

    /* Start by sorting the particles according to their nodes and
       getting the counts. */
    int *counts, *dest;
    struct part *parts = s->parts;
    double ih[3];
    ih[0] = s->ih[0]; ih[1] = s->ih[1]; ih[2] = s->ih[2];
    if ( ( counts = (int *)malloc( sizeof(int) * nr_nodes * nr_nodes ) ) == NULL ||
         ( dest = (int *)malloc( sizeof(int) * s->nr_parts ) ) == NULL )
        error( "Failed to allocate count and dest buffers." );
    bzero( counts , sizeof(int) * nr_nodes * nr_nodes );
    for ( k = 0 ; k < s->nr_parts ; k++ ) {
        cid = cell_getid( cdim , parts[k].x[0]*ih[0] , parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] );
        dest[k] = cells[ cid ].nodeID;
        counts[ nodeID*nr_nodes + dest[k] ] += 1;
        }
    parts_sort( s->parts , s->xparts , dest , s->nr_parts , 0 , nr_nodes-1 );
    
    /* Get all the counts from all the nodes. */
    if ( MPI_Allreduce( MPI_IN_PLACE , counts , nr_nodes * nr_nodes , MPI_INT , MPI_SUM , MPI_COMM_WORLD ) != MPI_SUCCESS )
        error( "Failed to allreduce particle transfer counts." );

    /* Get the new number of parts for this node, be generous in allocating. */
    int nr_parts = 0;
    for ( k = 0 ; k < nr_nodes ; k++ )
        nr_parts += counts[ k*nr_nodes + nodeID ];
    struct part *parts_new;
    struct xpart *xparts_new, *xparts = s->xparts;
    if ( posix_memalign( (void **)&parts_new , part_align , sizeof(struct part) * nr_parts * 1.2 ) != 0 ||
         posix_memalign( (void **)&xparts_new , part_align , sizeof(struct xpart) * nr_parts * 1.2 ) != 0 )
        error( "Failed to allocate new part data." );
        
    /* Emit the sends and recvs for the particle data. */
    MPI_Request *reqs;
    if ( ( reqs = (MPI_Request *)malloc( sizeof(MPI_Request) * 4 * nr_nodes ) ) == NULL )
        error( "Failed to allocate MPI request list." );
    for ( k = 0 ; k < 4*nr_nodes ; k++ )
        reqs[k] = MPI_REQUEST_NULL;
    for ( i = 0 , j = 0 , k = 0 ; k < nr_nodes ; k++ ) {
        if ( k == nodeID && counts[ nodeID*nr_nodes + k ] > 0 ) {
            memcpy( &parts_new[j] , &parts[i] , sizeof(struct part) * counts[ k*nr_nodes + nodeID ] );
            memcpy( &xparts_new[j] , &xparts[i] , sizeof(struct xpart) * counts[ k*nr_nodes + nodeID ] );
            i += counts[ nodeID*nr_nodes + k ];
            j += counts[ k*nr_nodes + nodeID ];
            }
        if ( k != nodeID && counts[ nodeID*nr_nodes + k ] > 0 ) {
            if ( MPI_Isend( &parts[i] , sizeof(struct part) * counts[ nodeID*nr_nodes + k ] , MPI_BYTE , k , 2*(nodeID*nr_nodes + k) + 0 , MPI_COMM_WORLD , &reqs[4*k] ) != MPI_SUCCESS )
                error( "Failed to isend parts to node %i." , k );
            if ( MPI_Isend( &xparts[i] , sizeof(struct xpart) * counts[ nodeID*nr_nodes + k ] , MPI_BYTE , k , 2*(nodeID*nr_nodes + k) + 1 , MPI_COMM_WORLD , &reqs[4*k+1] ) != MPI_SUCCESS )
                error( "Failed to isend xparts to node %i." , k );
            i += counts[ nodeID*nr_nodes + k ];
            }
        if ( k != nodeID && counts[ k*nr_nodes + nodeID ] > 0 ) {
            if ( MPI_Irecv( &parts_new[j] , sizeof(struct part) * counts[ k*nr_nodes + nodeID ] , MPI_BYTE , k , 2*(k*nr_nodes + nodeID) + 0 , MPI_COMM_WORLD , &reqs[4*k+2] ) != MPI_SUCCESS )
                error( "Failed to emit irecv of parts from node %i." , k );
            if ( MPI_Irecv( &xparts_new[j] , sizeof(struct xpart) * counts[ k*nr_nodes + nodeID ] , MPI_BYTE , k , 2*(k*nr_nodes + nodeID) + 1 , MPI_COMM_WORLD , &reqs[4*k+3] ) != MPI_SUCCESS )
                error( "Failed to emit irecv of parts from node %i." , k );
            j += counts[ k*nr_nodes + nodeID ];
            }
        }
        
    /* Wait for all the sends and recvs to tumble in. */
    MPI_Status stats[4*nr_nodes];
    int res;
    if ( ( res = MPI_Waitall( 4*nr_nodes , reqs , stats ) ) != MPI_SUCCESS ) {
        for ( k = 0 ; k < 4*nr_nodes ; k++ ) {
        char buff[ MPI_MAX_ERROR_STRING ];
        int res;
        MPI_Error_string( stats[k].MPI_ERROR , buff , &res );
        message( "request %i has error '%s'." , k , buff );
        }
    message( "counts is [ %i %i %i %i ]." , counts[0] , counts[1] , counts[2] , counts[3] );
        error( "Failed during waitall for part data." );
        }

    /* Verify that all parts are in the right place. */
    /* for ( k = 0 ; k < nr_parts ; k++ ) {
        cid = cell_getid( cdim , parts_new[k].x[0]*ih[0] , parts_new[k].x[1]*ih[1] , parts_new[k].x[2]*ih[2] );
        if ( cells[ cid ].nodeID != nodeID )
            error( "Received particle (%i) that does not belong here (nodeID=%i)." , k , cells[ cid ].nodeID );
        } */
        
    /* Set the new part data, free the old. */
    free( parts );
    free( xparts );
    s->parts = parts_new;
    s->xparts = xparts_new;
    s->nr_parts = nr_parts;
    s->size_parts = 1.2*nr_parts;
    
    /* Be verbose about what just happened. */
    for ( k = 0 ; k < nr_cells ; k++ )
        if ( cells[k].nodeID == nodeID )
            my_cells += 1;
    message( "node %i now has %i parts in %i cells." , nodeID , nr_parts , my_cells );
    
    /* Clean up other stuff. */
    free( reqs );
    free( counts );
    free( dest );
        
#else
    error( "SWIFT was not compiled with MPI and METIS support." );
#endif

    }


/**
 * @breif Repartition the cells amongst the nodes.
 *
 * @param e The #engine.
 */
 
void engine_repartition ( struct engine *e ) {

#if defined(WITH_MPI) && defined(HAVE_METIS)

    int i, j, k, l, cid, cjd, ii, jj, kk, res, w;
    idx_t *inds, *nodeIDs;
    idx_t *weights_v, *weights_e;
    struct space *s = e->s;
    int nr_cells = s->nr_cells, my_cells = 0;
    struct cell *cells = s->cells;
    int ind[3], *cdim = s->cdim;
    struct task *t, *tasks = e->sched.tasks;
    struct cell *ci, *cj;
    int nr_nodes = e->nr_nodes, nodeID = e->nodeID;
    float wscale = 0.0001, vscale = 0.001;
    
    /* Clear the repartition flag. */
    e->forcerepart = 0;
    
    /* Allocate the inds and weights. */
    if ( ( inds = (idx_t *)malloc( sizeof(idx_t) * 26*nr_cells ) ) == NULL ||
         ( weights_v = (idx_t *)malloc( sizeof(idx_t) * nr_cells ) ) == NULL ||
         ( weights_e = (idx_t *)malloc( sizeof(idx_t) * 26*nr_cells ) ) == NULL ||
         ( nodeIDs = (idx_t *)malloc( sizeof(idx_t) * nr_cells ) ) == NULL )
        error( "Failed to allocate inds and weights arrays." );
        
    /* Fill the inds array. */
    for ( cid = 0 ; cid < nr_cells ; cid++ ) {
        ind[0] = cells[cid].loc[0] / s->cells[cid].h[0] + 0.5;
        ind[1] = cells[cid].loc[1] / s->cells[cid].h[1] + 0.5;
        ind[2] = cells[cid].loc[2] / s->cells[cid].h[2] + 0.5;
        l = 0;
        for ( i = -1 ; i <= 1 ; i++ ) {
            ii = ind[0] + i;
            if ( ii < 0 ) ii += cdim[0];
            else if ( ii >= cdim[0] ) ii -= cdim[0];
            for ( j = -1 ; j <= 1 ; j++ ) {
                jj = ind[1] + j;
                if ( jj < 0 ) jj += cdim[1];
                else if ( jj >= cdim[1] ) jj -= cdim[1];
                for ( k = -1 ; k <= 1 ; k++ ) {
                    kk = ind[2] + k;
                    if ( kk < 0 ) kk += cdim[2];
                    else if ( kk >= cdim[2] ) kk -= cdim[2];
                    if ( i || j || k ) {
                        inds[ cid*26 + l ] = cell_getid( cdim , ii , jj , kk );
                        l += 1;
                        }
                    }
                }
            }
        }
        
    /* Init the weights arrays. */
    /* bzero( weights_e , sizeof(idx_t) * 26*nr_cells );
    bzero( weights_v , sizeof(idx_t) * nr_cells ); */
    for ( k = 0 ; k < 26*nr_cells ; k++ )
        weights_e[k] = 1;
    for ( k = 0 ; k < nr_cells ; k++ )
        weights_v[k] = 1;
    
    /* Loop over the tasks... */
    for ( j = 0 ; j < e->sched.nr_tasks ; j++ ) {
    
        /* Get a pointer to the kth task. */
        t = &tasks[j];
        
        /* Skip un-interesting tasks. */
        if ( t->type != task_type_self &&
             t->type != task_type_pair &&
             t->type != task_type_sub &&
             t->type != task_type_ghost &&
             t->type != task_type_kick1 &&
             t->type != task_type_kick2 )
            continue;
            
        /* Get the task weight. */
        w = ( t->toc - t->tic ) * wscale;
        if ( w < 0 )
            error( "Bad task weight (%i)." , w );
        
        /* Get the top-level cells involved. */
        for ( ci = t->ci ; ci->parent != NULL ; ci = ci->parent );
        if ( t->cj != NULL )
            for ( cj = t->cj ; cj->parent != NULL ; cj = cj->parent );
        else
            cj = NULL;
            
        /* Get the cell IDs. */
        cid = ci - cells;
            
        /* Different weights for different tasks. */
        if ( t->type == task_type_ghost ||
             t->type == task_type_kick1 || 
             t->type == task_type_kick2 ) {
             
            /* Particle updates add only to vertex weight. */
            weights_v[cid] += w;
            
            }
        
        /* Self interaction? */     
        else if ( ( t->type == task_type_self && ci->nodeID == nodeID ) ||
             ( t->type == task_type_sub && cj == NULL && ci->nodeID == nodeID ) ) {
        
            /* Self interactions add only to vertex weight. */
            weights_v[cid] += w;
            
            }
            
        /* Pair? */
        else if ( t->type == task_type_pair ||
                  ( t->type == task_type_sub && cj != NULL ) ) {
                  
            /* In-cell pair? */
            if ( ci == cj ) {
            
                /* Add weight to vertex for ci. */
                weights_v[cid] += w;
            
                }
                
            /* Distinct cells with local ci? */
            else if ( ci->nodeID == nodeID ) {
            
                /* Index of the jth cell. */
                cjd = cj - cells;
                
                /* Add half of weight to each cell. */
                if ( ci->nodeID == nodeID )
                    weights_v[cid] += 0.5 * w;
                if ( cj->nodeID == nodeID )
                    weights_v[cjd] += 0.5 * w;
                    
                /* Add Weight to edge. */
                for ( k = 26*cid ; inds[k] != cjd ; k++ );
                weights_e[ k ] += w;
                for ( k = 26*cjd ; inds[k] != cid ; k++ );
                weights_e[ k ] += w;
            
                }
                  
            }
    
        }
        
    /* Merge the weights arrays accross all nodes. */
#if IDXTYPEWIDTH==32
    if ( ( res = MPI_Reduce( ( nodeID == 0 ) ? MPI_IN_PLACE : weights_v , weights_v , nr_cells , MPI_INT , MPI_SUM , 0 , MPI_COMM_WORLD ) ) != MPI_SUCCESS ) {
#else
    if ( ( res = MPI_Reduce( ( nodeID == 0 ) ? MPI_IN_PLACE : weights_v , weights_v , nr_cells , MPI_LONG_LONG_INT , MPI_SUM , 0 , MPI_COMM_WORLD ) ) != MPI_SUCCESS ) {
#endif
        char buff[ MPI_MAX_ERROR_STRING ];
        MPI_Error_string( res , buff , &i );
        error( "Failed to allreduce vertex weights (%s)." , buff );
        }
#if IDXTYPEWIDTH==32
    if ( MPI_Reduce( ( nodeID == 0 ) ? MPI_IN_PLACE : weights_e , weights_e , 26*nr_cells , MPI_INT , MPI_SUM , 0 , MPI_COMM_WORLD ) != MPI_SUCCESS )
#else
    if ( MPI_Reduce( ( nodeID == 0 ) ? MPI_IN_PLACE : weights_e , weights_e , 26*nr_cells , MPI_LONG_LONG_INT , MPI_SUM , 0 , MPI_COMM_WORLD ) != MPI_SUCCESS )
#endif
       error( "Failed to allreduce edge weights." );
        
    /* As of here, only one node needs to compute the partition. */
    if ( nodeID == 0 ) {
    
        /* Check that the edge weights are fully symmetric. */
        /* for ( cid = 0 ; cid < nr_cells ; cid++ )
            for ( k = 0 ; k < 26 ; k++ ) {
                cjd = inds[ cid*26 + k ];
                for ( j = 26*cjd ; inds[j] != cid ; j++ );
                if ( weights_e[ cid*26+k ] != weights_e[ j ] )
                    error( "Unsymmetric edge weights detected (%i vs %i)." , weights_e[ cid*26+k ] , weights_e[ j ] );
                } */
        /* int w_min = weights_e[0], w_max = weights_e[0], w_tot = weights_e[0];
        for ( k = 1 ; k < 26*nr_cells ; k++ ) {
            w_tot += weights_e[k];
            if ( weights_e[k] < w_min )
                w_min = weights_e[k];
            else if ( weights_e[k] > w_max )
                w_max = weights_e[k];
            }
        message( "edge weights in [ %i , %i ], tot=%i." , w_min , w_max , w_tot );
        w_min = weights_e[0], w_max = weights_e[0]; w_tot = weights_v[0];
        for ( k = 1 ; k < nr_cells ; k++ ) {
            w_tot += weights_v[k];
            if ( weights_v[k] < w_min )
                w_min = weights_v[k];
            else if ( weights_v[k] > w_max )
                w_max = weights_v[k];
            }
        message( "vertex weights in [ %i , %i ], tot=%i." , w_min , w_max , w_tot ); */
                
        /* Make sure there are no zero weights. */
        for ( k = 0 ; k < 26*nr_cells ; k++ )
            if ( weights_e[k] == 0 )
                weights_e[k] = 1;
        for ( k = 0 ; k < nr_cells ; k++ )
            if ( ( weights_v[k] *= vscale ) == 0 )
                weights_v[k] = 1;
    
        /* Allocate and fill the connection array. */
        idx_t *offsets;
        if ( ( offsets = (idx_t *)malloc( sizeof(idx_t) * (nr_cells + 1) ) ) == NULL )
            error( "Failed to allocate offsets buffer." );
        offsets[0] = 0;
        for ( k = 0 ; k < nr_cells ; k++ )
            offsets[k+1] = offsets[k] + 26;
            
        /* Set the METIS options. */
        idx_t options[METIS_NOPTIONS];
        METIS_SetDefaultOptions( options );
        options[ METIS_OPTION_OBJTYPE ] = METIS_OBJTYPE_CUT;
        options[ METIS_OPTION_NUMBERING ] = 0;
        options[ METIS_OPTION_CONTIG ] = 1;
        options[ METIS_OPTION_NCUTS ] = 10;
        options[ METIS_OPTION_NITER ] = 20;
        // options[ METIS_OPTION_UFACTOR ] = 1;
        
        /* Set the initial partition, although this is probably ignored. */
        for ( k = 0 ; k < nr_cells ; k++ )
            nodeIDs[k] = cells[k].nodeID;
            
        /* Call METIS. */
        idx_t one = 1, idx_nr_cells = nr_cells, idx_nr_nodes = nr_nodes;
        idx_t objval;
        if ( METIS_PartGraphRecursive( &idx_nr_cells , &one , offsets , inds , weights_v , NULL , weights_e , &idx_nr_nodes , NULL , NULL , options , &objval , nodeIDs ) != METIS_OK )
            error( "Call to METIS_PartGraphKway failed." );

        /* Dump the 3d array of cell IDs. */
        printf( "engine_repartition: nodeIDs = reshape( [" );
        for ( i = 0 ; i < cdim[0]*cdim[1]*cdim[2] ; i++ )
            printf( "%i " , (int)nodeIDs[ i ] );
        printf("] ,%i,%i,%i);\n",cdim[0],cdim[1],cdim[2]);
    
        }
        
    /* Broadcast the result of the partition. */
    if ( MPI_Bcast( nodeIDs , nr_cells , MPI_INT , 0 , MPI_COMM_WORLD ) != MPI_SUCCESS )
        error( "Failed to bcast the node IDs." );
        
    /* Set the cell nodeIDs and clear any non-local parts. */
    for ( k = 0 ; k < nr_cells ; k++ ) {
        cells[k].nodeID = nodeIDs[k];
        if ( nodeIDs[k] == nodeID )
            my_cells += 1;
        }
        
    /* Clean up. */
    free( inds );
    free( weights_v );
    free( weights_e );
    free( nodeIDs );
        
    /* Now comes the tricky part: Exchange particles between all nodes.
       This is done in two steps, first allreducing a matrix of 
       how many particles go from where to where, then re-allocating
       the parts array, and emiting the sends and receives.
       Finally, the space, tasks, and proxies need to be rebuilt. */
       
    /* Redistribute the particles between the nodes. */
    engine_redistribute( e );
        
    /* Make the proxies. */
    engine_makeproxies( e );
        
    /* Tell the engine it should re-build whenever possible */
    e->forcerebuild = 1;
    
#else
    error( "SWIFT was not compiled with MPI and METIS support." );
#endif

    }
    
    
/**
 * @brief Add up/down gravity tasks to a cell hierarchy.
 *
 * @param e The #engine.
 * @param c The #cell
 * @param up The upward gravity #task.
 * @param down The downward gravity #task.
 */
 
void engine_addtasks_grav ( struct engine *e , struct cell *c , struct task *up , struct task *down ) {

    /* Link the tasks to this cell. */
    c->grav_up = up;
    c->grav_down = down;
    
    /* Recurse? */
    if ( c->split )
        for ( int k = 0 ; k < 8 ; k++ )
            if ( c->progeny[k] != NULL )
                engine_addtasks_grav( e , c->progeny[k] , up , down );

    }


/**
 * @brief Add send tasks to a hierarchy of cells.
 *
 * @param e The #engine.
 * @param c The #cell.
 * @param t_xv The send_xv #task, if it has already been created.
 * @param t_rho The send_rho #task, if it has already been created.
 */

void engine_addtasks_send ( struct engine *e , struct cell *ci , struct cell *cj ) {

    int k;
    struct link *l = NULL;
    struct scheduler *s = &e->sched;

    /* Check if any of the density tasks are for the target node. */
    for ( l = ci->density ; l != NULL ; l = l->next )
        if ( l->t->ci->nodeID == cj->nodeID ||
             ( l->t->cj != NULL && l->t->cj->nodeID == cj->nodeID ) )
            break;

    /* If so, attach send tasks. */
    if ( l != NULL ) {

        /* Create the tasks. */
        struct task *t_xv = scheduler_addtask( &e->sched , task_type_send , task_subtype_none , 2*ci->tag , 0 , ci , cj , 0 );
        struct task *t_rho = scheduler_addtask( &e->sched , task_type_send , task_subtype_none , 2*ci->tag + 1 , 0 , ci , cj , 0 );

        /* The send_rho task depends on the cell's ghost task. */
        scheduler_addunlock( s , ci->super->ghost , t_rho );

        /* The send_rho task should unlock the super-cell's kick2 task. */
        scheduler_addunlock( s , t_rho , ci->super->kick2 );

        /* The send_xv task should unlock the super-cell's ghost task. */
        scheduler_addunlock( s , t_xv , ci->super->ghost );

        }
        
    /* Recurse? */
    else if ( ci->split )
        for ( k = 0 ; k < 8 ; k++ )
            if ( ci->progeny[k] != NULL )
                engine_addtasks_send( e , ci->progeny[k] , cj );

    }


/**
 * @brief Add recv tasks to a hierarchy of cells.
 *
 * @param e The #engine.
 * @param c The #cell.
 * @param t_xv The recv_xv #task, if it has already been created.
 * @param t_rho The recv_rho #task, if it has already been created.
 */

void engine_addtasks_recv ( struct engine *e , struct cell *c , struct task *t_xv , struct task *t_rho ) {

    int k;
    struct scheduler *s = &e->sched;

    /* Do we need to construct a recv task? */
    if ( t_xv == NULL && c->nr_density > 0 ) {
    
        /* Create the tasks. */
        t_xv = c->recv_xv = scheduler_addtask( &e->sched , task_type_recv , task_subtype_none , 2*c->tag , 0 , c , NULL , 0 );
        t_rho = c->recv_rho = scheduler_addtask( &e->sched , task_type_recv , task_subtype_none , 2*c->tag + 1 , 0 , c , NULL , 0 );
        
        }
        
    /* Add dependencies. */
    for ( struct link *l = c->density ; l != NULL ; l = l->next ) {
        scheduler_addunlock( s , t_xv , l->t );
        scheduler_addunlock( s , l->t , t_rho );
        }
    for ( struct link *l = c->force ; l != NULL ; l = l->next )
        scheduler_addunlock( s , t_rho , l->t );
    if ( c->sorts != NULL )
        scheduler_addunlock( s , t_xv , c->sorts );
    
    /* Recurse? */
    if ( c->split )
        for ( k = 0 ; k < 8 ; k++ )
            if ( c->progeny[k] != NULL )
                engine_addtasks_recv( e , c->progeny[k] , t_xv , t_rho );

    }


/**
 * @brief Exchange cell structures with other nodes.
 *
 * @param e The #engine.
 */
 
void engine_exchange_cells ( struct engine *e ) {

#ifdef WITH_MPI

    int j, k, pid, count = 0;
    struct pcell *pcells;
    struct space *s = e->s;
    struct cell *cells = s->cells;
    int nr_cells = s->nr_cells;
    int nr_proxies = e->nr_proxies;
    int offset[ nr_cells ];
    MPI_Request reqs_in[ engine_maxproxies ];
    MPI_Request reqs_out[ engine_maxproxies ];
    MPI_Status status;
    struct part *parts = &s->parts[ s->nr_parts ];
    
    /* Run through the cells and get the size of the ones that will be sent off. */
    for ( k = 0 ; k < nr_cells ; k++ ) {
        offset[k] = count;
        if ( cells[k].sendto )
            count += ( cells[k].pcell_size = cell_getsize( &cells[k] ) );
        }
        
    /* Allocate the pcells. */
    if ( ( pcells = (struct pcell *)malloc( sizeof(struct pcell) * count ) ) == NULL )
        error( "Failed to allocate pcell buffer." );
        
    /* Pack the cells. */
    for ( k = 0 ; k < nr_cells ; k++ )
        if ( cells[k].sendto ) {
            cell_pack( &cells[k] , &pcells[ offset[k] ] );
            cells[k].pcell = &pcells[ offset[k] ];
            }

    /* Launch the proxies. */
    for ( k = 0 ; k < nr_proxies ; k++ ) {
        proxy_cells_exch1( &e->proxies[k] );
        reqs_in[k] = e->proxies[k].req_cells_count_in;
        reqs_out[k] = e->proxies[k].req_cells_count_out;
        }
        
    /* Wait for each count to come in and start the recv. */
    for ( k = 0 ; k < nr_proxies ; k++ ) {
        if ( MPI_Waitany( nr_proxies , reqs_in , &pid , &status ) != MPI_SUCCESS ||
             pid == MPI_UNDEFINED )
            error( "MPI_Waitany failed." );
        // message( "request from proxy %i has arrived." , pid );
        proxy_cells_exch2( &e->proxies[pid] );
        }
        
    /* Wait for all the sends to have finnished too. */
    if ( MPI_Waitall( nr_proxies , reqs_out , MPI_STATUSES_IGNORE ) != MPI_SUCCESS )
        error( "MPI_Waitall on sends failed." );
        
    /* Set the requests for the cells. */
    for ( k = 0 ; k < nr_proxies ; k++ ) {
        reqs_in[k] = e->proxies[k].req_cells_in;
        reqs_out[k] = e->proxies[k].req_cells_out;
        }
    
    /* Wait for each pcell array to come in from the proxies. */
    for ( k = 0 ; k < nr_proxies ; k++ ) {
        if ( MPI_Waitany( nr_proxies , reqs_in , &pid , &status ) != MPI_SUCCESS ||
             pid == MPI_UNDEFINED )
            error( "MPI_Waitany failed." );
        // message( "cell data from proxy %i has arrived." , pid );
        for ( count = 0 , j = 0 ; j < e->proxies[pid].nr_cells_in ; j++ )
            count += cell_unpack( &e->proxies[pid].pcells_in[count] , e->proxies[pid].cells_in[j] , e->s );
        }
        
    /* Wait for all the sends to have finnished too. */
    if ( MPI_Waitall( nr_proxies , reqs_out , MPI_STATUSES_IGNORE ) != MPI_SUCCESS )
        error( "MPI_Waitall on sends failed." );
        
    /* Count the number of particles we need to import and re-allocate
       the buffer if needed. */
    for ( count = 0 , k = 0 ; k < nr_proxies ; k++ )
        for ( j = 0 ; j < e->proxies[k].nr_cells_in ; j++ )
            count += e->proxies[k].cells_in[j]->count;
    if ( count > s->size_parts_foreign ) {
        if ( s->parts_foreign != NULL )
            free( s->parts_foreign );
        s->size_parts_foreign = 1.1 * count;
        if ( posix_memalign( (void **)&s->parts_foreign , part_align , sizeof(struct part) * s->size_parts_foreign ) != 0 )
            error( "Failed to allocate foreign part data." );
        }
        
    /* Unpack the cells and link to the particle data. */
    parts = s->parts_foreign;
    for ( k = 0 ; k < nr_proxies ; k++ ) {
        for ( count = 0 , j = 0 ; j < e->proxies[k].nr_cells_in ; j++ ) {
            count += cell_link( e->proxies[k].cells_in[j] , parts );
            parts = &parts[ e->proxies[k].cells_in[j]->count ];
            }
        }
    s->nr_parts_foreign = parts - s->parts_foreign;
        
    /* Is the parts buffer large enough? */
    if ( s->nr_parts_foreign > s->size_parts_foreign )
        error( "Foreign parts buffer too small." );
        
    /* Free the pcell buffer. */
    free( pcells );
    
#else
    error( "SWIFT was not compiled with MPI support." );
#endif

    }


/**
 * @brief Exchange straying parts with other nodes.
 *
 * @param e The #engine.
 * @param parts An array of straying parts.
 * @param xparts The corresponding xparts.
 * @param ind The ID of the foreign #cell.
 * @param N The number of stray parts.
 *
 * @return The number of arrived parts copied to parts and xparts.
 */
 
int engine_exchange_strays ( struct engine *e , struct part *parts , struct xpart *xparts , int *ind , int N ) {

#ifdef WITH_MPI

    int k, pid, count = 0, nr_in = 0, nr_out = 0;
    MPI_Request reqs_in[ 2*engine_maxproxies ];
    MPI_Request reqs_out[ 2*engine_maxproxies ];
    MPI_Status status;
    struct proxy *p;

    /* Re-set the proxies. */
    for ( k = 0 ; k < e->nr_proxies ; k++ )
        e->proxies[k].nr_parts_out = 0;
    
    /* Put the parts into the corresponding proxies. */
    for ( k = 0 ; k < N ; k++ ) {
        pid = e->proxy_ind[ e->s->cells[ ind[k] ].nodeID ];
        if ( pid < 0 )
            error( "Do not have a proxy for the requested nodeID." );
        proxy_parts_load( &e->proxies[pid] , &parts[k] , &xparts[k] , 1 );
        }
    
    /* Launch the proxies. */
    for ( k = 0 ; k < e->nr_proxies ; k++ ) {
        proxy_parts_exch1( &e->proxies[k] );
        reqs_in[k] = e->proxies[k].req_parts_count_in;
        reqs_out[k] = e->proxies[k].req_parts_count_out;
        }
        
    /* Wait for each count to come in and start the recv. */
    for ( k = 0 ; k < e->nr_proxies ; k++ ) {
        if ( MPI_Waitany( e->nr_proxies , reqs_in , &pid , &status ) != MPI_SUCCESS ||
             pid == MPI_UNDEFINED )
            error( "MPI_Waitany failed." );
        // message( "request from proxy %i has arrived." , pid );
        proxy_parts_exch2( &e->proxies[pid] );
        }
        
    /* Wait for all the sends to have finnished too. */
    if ( MPI_Waitall( e->nr_proxies , reqs_out , MPI_STATUSES_IGNORE ) != MPI_SUCCESS )
        error( "MPI_Waitall on sends failed." );
        
    /* Return the number of harvested parts. */
    /* Set the requests for the particle data. */
    for ( k = 0 ; k < e->nr_proxies ; k++ ) {
        if ( e->proxies[k].nr_parts_in > 0 ) {
            reqs_in[2*k] = e->proxies[k].req_parts_in;
            reqs_in[2*k+1] = e->proxies[k].req_xparts_in;
            nr_in += 1;
            }
        else
            reqs_in[2*k] = reqs_in[2*k+1] = MPI_REQUEST_NULL;
        if ( e->proxies[k].nr_parts_out > 0 ) {
            reqs_out[2*k] = e->proxies[k].req_parts_out;
            reqs_out[2*k+1] = e->proxies[k].req_xparts_out;
            nr_out += 1;
            }
        else
            reqs_out[2*k] = reqs_out[2*k+1] = MPI_REQUEST_NULL;
        }
    
    /* Wait for each part array to come in and collect the new
       parts from the proxies. */
    for ( k = 0 ; k < 2*(nr_in + nr_out) ; k++ ) {
        int err;
        if ( ( err = MPI_Waitany( 2*e->nr_proxies , reqs_in , &pid , &status ) ) != MPI_SUCCESS ) {
        char buff[ MPI_MAX_ERROR_STRING ];
        int res;
        MPI_Error_string( err , buff , &res );
            error( "MPI_Waitany failed (%s)." , buff );
        }
    if ( pid == MPI_UNDEFINED )
        break;
        if ( pid == MPI_UNDEFINED )
            break;
        // message( "request from proxy %i has arrived." , pid );
        if ( reqs_in[pid & ~1] == MPI_REQUEST_NULL &&
             reqs_in[pid | 1 ] == MPI_REQUEST_NULL ) {
            p = &e->proxies[pid/2];
            memcpy( &parts[count] , p->parts_in , sizeof(struct part) * p->nr_parts_in );
            memcpy( &xparts[count] , p->xparts_in , sizeof(struct xpart) * p->nr_parts_in );
            count += p->nr_parts_in;
            /* for ( int k = 0 ; k < p->nr_parts_in ; k++ )
                message( "received particle %lli, x=[%.3e %.3e %.3e], h=%.3e, from node %i." ,
                    p->parts_in[k].id , p->parts_in[k].x[0] , p->parts_in[k].x[1] , p->parts_in[k].x[2] ,
                    p->parts_in[k].h , p->nodeID ); */
            }
        }
    
    /* Wait for all the sends to have finnished too. */
    if ( nr_out > 0 )
        if ( MPI_Waitall( 2*e->nr_proxies , reqs_out , MPI_STATUSES_IGNORE ) != MPI_SUCCESS )
            error( "MPI_Waitall on sends failed." );
        
    /* Return the number of harvested parts. */
    return count;
    
#else
    error( "SWIFT was not compiled with MPI support." );
    return 0;
#endif

    }


/**
 * @brief Fill the #space's task list.
 *
 * @param s The #space we are working in.
 */
 
void engine_maketasks ( struct engine *e ) {

    struct space *s = e->s;
    struct scheduler *sched = &e->sched;
    struct cell *cells = s->cells;
    int nr_cells = s->nr_cells;
    int nodeID = e->nodeID;
    int i, j, k, ii, jj, kk, iii, jjj, kkk, cid, cjd, sid;
    int *cdim = s->cdim;
    struct task *t, *t2;
    struct cell *ci, *cj;

    /* Re-set the scheduler. */
    scheduler_reset( sched , s->tot_cells * engine_maxtaskspercell );
    
    /* Run through the highest level of cells and add pairs. */
    for ( i = 0 ; i < cdim[0] ; i++ )
        for ( j = 0 ; j < cdim[1] ; j++ )
            for ( k = 0 ; k < cdim[2] ; k++ ) {
                cid = cell_getid( cdim , i , j , k );
                if ( cells[cid].count == 0 )
                    continue;
                ci = &cells[cid];
                if ( ci->count == 0 )
                    continue;
                if ( ci->nodeID == nodeID )
                    scheduler_addtask( sched , task_type_self , task_subtype_density , 0 , 0 , ci , NULL , 0 );
                for ( ii = -1 ; ii < 2 ; ii++ ) {
                    iii = i + ii;
                    if ( !s->periodic && ( iii < 0 || iii >= cdim[0] ) )
                        continue;
                    iii = ( iii + cdim[0] ) % cdim[0];
                    for ( jj = -1 ; jj < 2 ; jj++ ) {
                        jjj = j + jj;
                        if ( !s->periodic && ( jjj < 0 || jjj >= cdim[1] ) )
                            continue;
                        jjj = ( jjj + cdim[1] ) % cdim[1];
                        for ( kk = -1 ; kk < 2 ; kk++ ) {
                            kkk = k + kk;
                            if ( !s->periodic && ( kkk < 0 || kkk >= cdim[2] ) )
                                continue;
                            kkk = ( kkk + cdim[2] ) % cdim[2];
                            cjd = cell_getid( cdim , iii , jjj , kkk );
                            cj = &cells[cjd];
                            if ( cid >= cjd || cj->count == 0 || 
                                 ( ci->nodeID != nodeID && cj->nodeID != nodeID ) )
                                continue;
                            sid = sortlistID[ (kk+1) + 3*( (jj+1) + 3*(ii+1) ) ];
                            scheduler_addtask( sched , task_type_pair , task_subtype_density , sid , 0 , ci , cj , 1 );
                            }
                        }
                    }
                }
                
    /* Add the gravity mm tasks. */
    for ( i = 0 ; i < nr_cells ; i++ )
        if ( cells[i].gcount > 0 ) {
            scheduler_addtask( sched , task_type_grav_mm , task_subtype_none , -1 , 0 , &cells[i] , NULL , 0 );
            for ( j = i+1 ; j < nr_cells ; j++ )
                if ( cells[j].gcount > 0 )
                    scheduler_addtask( sched , task_type_grav_mm , task_subtype_none , -1 , 0 , &cells[i] , &cells[j] , 0 );
            }
        
    /* Split the tasks. */
    scheduler_splittasks( sched );
    
    /* Allocate the list of cell-task links. */
    if ( e->links != NULL )
        free( e->links );
    if ( ( e->links = malloc( sizeof(struct link) * s->tot_cells * 27 ) ) == NULL )
        error( "Failed to allocate cell-task links." );
    e->nr_links = 0;
    
    /* Add the gravity up/down tasks at the top-level cells and push them down. */
    for ( k = 0 ; k < nr_cells ; k++ )
        if ( cells[k].nodeID == nodeID && cells[k].gcount > 0 ) {
        
            /* Create tasks at top level. */
            struct task *up = scheduler_addtask( sched , task_type_grav_up , task_subtype_none , 0 , 0 , &cells[k] , NULL , 0 );
            struct task *down = scheduler_addtask( sched , task_type_grav_down , task_subtype_none , 0 , 0 , &cells[k] , NULL , 0 );
            
            /* Push tasks down the cell hierarchy. */
            engine_addtasks_grav( e , &cells[k] , up , down );
            
            }
    
    /* Count the number of tasks associated with each cell and
       store the density tasks in each cell, and make each sort
       depend on the sorts of its progeny. */
    // #pragma omp parallel for private(t,j)