From 104f989e7ff687e905684b893f7099d0ce825d8a Mon Sep 17 00:00:00 2001
From: Pietro Incardona <incardon@mpi-cbg.de>
Date: Mon, 2 Nov 2015 15:16:05 +0100
Subject: [PATCH] New classes
---
src/Decomposition/BasicDecomposition.hpp | 1125 +++++++++++++++++
.../BasicDecomposition_unit_test.hpp | 83 ++
src/parmetis_util.hpp | 486 +++++++
3 files changed, 1694 insertions(+)
create mode 100644 src/Decomposition/BasicDecomposition.hpp
create mode 100644 src/Decomposition/BasicDecomposition_unit_test.hpp
create mode 100644 src/parmetis_util.hpp
diff --git a/src/Decomposition/BasicDecomposition.hpp b/src/Decomposition/BasicDecomposition.hpp
new file mode 100644
index 000000000..594b85df3
--- /dev/null
+++ b/src/Decomposition/BasicDecomposition.hpp
@@ -0,0 +1,1125 @@
+/*
+ * BasicDecomposition.hpp
+ *
+ * Created on: Oct 07, 2015
+ * Author: Antonio Leo
+ */
+
+#ifndef BASICDECOMPOSITION_HPP
+#define BASICDECOMPOSITION_HPP
+
+#include "config.h"
+#include <cmath>
+#include "VCluster.hpp"
+#include "Graph/CartesianGraphFactory.hpp"
+#include "Decomposition.hpp"
+#include "Vector/map_vector.hpp"
+#include <vector>
+#include <initializer_list>
+#include "SubdomainGraphNodes.hpp"
+#include "parmetis_util.hpp"
+#include "dec_optimizer.hpp"
+#include "Space/Shape/Box.hpp"
+#include "Space/Shape/Point.hpp"
+#include "NN/CellList/CellDecomposer.hpp"
+#include <unordered_map>
+#include "NN/CellList/CellList.hpp"
+#include "Space/Ghost.hpp"
+#include "common.hpp"
+#include "ie_loc_ghost.hpp"
+#include "ie_ghost.hpp"
+#include "nn_processor.hpp"
+#include "GraphMLWriter.hpp"
+
+#define BASICDEC_ERROR 2000lu
+
+// Macro that decide what to do in case of error
+#ifdef STOP_ON_ERROR
+#define ACTION_ON_ERROR() exit(1);
+#elif defined(THROW_ON_ERROR)
+#define ACTION_ON_ERROR() throw BASICDEC_ERROR;
+#else
+#define ACTION_ON_ERROR()
+#endif
+
+/**
+ * \brief This class decompose a space into subspaces
+ *
+ * \tparam dim is the dimensionality of the physical domain we are going to decompose.
+ * \tparam T type of the space we decompose, Real, Integer, Complex ...
+ * \tparam Memory Memory factory used to allocate memory
+ * \tparam Domain Structure that contain the information of your physical domain
+ *
+ * Given an N-dimensional space, this class decompose the space into a Cartesian grid of small
+ * sub-sub-domain. At each sub-sub-domain is assigned an id that identify which processor is
+ * going to take care of that part of space (in general the space assigned to a processor is
+ * simply connected), a second step merge several sub-sub-domain with same id into bigger region
+ * sub-domain with the id. Each sub-domain has an extended space called ghost part
+ *
+ * Assuming that VCluster.getProcessUnitID(), equivalent to the MPI processor rank, return the processor local
+ * processor id, we define
+ *
+ * * local processor: processor rank
+ * * local sub-domain: sub-domain given to the local processor
+ * * external ghost box: (or ghost box) are the boxes that compose the ghost space of the processor, or the
+ * boxes produced expanding every local sub-domain by the ghost extension and intersecting with the sub-domain
+ * of the other processors
+ * * Near processors are the processors adjacent to the local processor, where with adjacent we mean all the processor
+ * that has a non-zero intersection with the ghost part of the local processor, or all the processors that
+ * produce non-zero external boxes with the local processor, or all the processor that should communicate
+ * in case of ghost data synchronization
+ * * internal ghost box: is the part of ghost of the near processor that intersect the space of the
+ * processor, or the boxes produced expanding the sub-domain of the near processors with the local sub-domain
+ * * Near processor sub-domain: is a sub-domain that live in the a near (or contiguous) processor
+ * * Near processor list: the list of all the near processor of the local processor (each processor has a list
+ * of the near processor)
+ * * Local ghosts interal or external are all the ghosts that does not involve inter-processor communications
+ *
+ * \see calculateGhostBoxes() for a visualization of internal and external ghost boxes
+ *
+ * ### Create a Cartesian decomposition object on a Box space, distribute, calculate internal and external ghost boxes
+ * \snippet BasicDecomposition_unit_test.hpp Create BasicDecomposition
+ *
+ */
+
+template<unsigned int dim, typename T, typename Memory = HeapMemory, template<unsigned int, typename > class Domain = Box>
+class BasicDecomposition: public ie_loc_ghost<dim, T>, public nn_prcs<dim, T>, public ie_ghost<dim, T> {
+
+public:
+
+ //! Type of the domain we are going to decompose
+ typedef T domain_type;
+
+ //! It simplify to access the SpaceBox element
+ typedef SpaceBox<dim, T> Box;
+
+private:
+
+ //! This is the key type to access data_s, for example in the case of vector
+ //! acc_key is size_t
+ typedef typename openfpm::vector<SpaceBox<dim, T>, Memory, openfpm::vector_grow_policy_default,
+ openfpm::vect_isel<SpaceBox<dim, T>>::value>::access_key acc_key;
+
+ //! the set of all local sub-domain as vector
+ openfpm::vector<SpaceBox<dim, T>> sub_domains;
+
+ //! for each sub-domain, contain the list of the neighborhood processors
+ openfpm::vector<openfpm::vector<long unsigned int> > box_nn_processor;
+
+ //! Structure that contain for each sub-sub-domain box the processor id
+ //! exist for efficient global communication
+ openfpm::vector<size_t> fine_s;
+
+ //! Structure that store the cartesian grid information
+ grid_sm<dim, void> gr;
+
+ //! Structure that decompose your structure into cell without creating them
+ //! useful to convert positions to CellId or sub-domain id in this case
+ CellDecomposer_sm<dim, T> cd;
+
+ //! rectangular domain to decompose
+ Domain<dim, T> domain;
+
+ //! Box Spacing
+ T spacing[dim];
+
+ //! Runtime virtual cluster machine
+ Vcluster & v_cl;
+
+ //! Cell-list that store the geometrical information of the local internal ghost boxes
+ CellList<dim, T, FAST> lgeo_cell;
+
+ /*! \brief Constructor, it decompose and distribute the sub-domains across the processors
+ *
+ * \param v_cl Virtual cluster, used internally for communications
+ *
+ */
+ void CreateDecomposition(Vcluster & v_cl) {
+#ifdef SE_CLASS1
+ if (&v_cl == NULL)
+ {
+ std::cerr << __FILE__ << ":" << __LINE__ << " error VCluster instance is null, check that you ever initialized it \n";
+ ACTION_ON_ERROR()
+ }
+#endif
+ // Calculate the total number of box and and the spacing
+ // on each direction
+ // Get the box containing the domain
+ SpaceBox<dim, T> bs = domain.getBox();
+
+ for (unsigned int i = 0; i < dim; i++) {
+ // Calculate the spacing
+ spacing[i] = (bs.getHigh(i) - bs.getLow(i)) / gr.size(i);
+ }
+
+ //! MPI initialization
+ int MPI_PROC_ID = 0;
+ MPI_Comm_rank(MPI_COMM_WORLD, &MPI_PROC_ID);
+
+ // Here we use PARMETIS
+ // Create a cartesian grid graph
+ CartesianGraphFactory<dim, Graph_CSR<nm_v, nm_e>> g_factory_part;
+
+ // Processor graph
+ Graph_CSR<nm_v, nm_e> gp = g_factory_part.template construct<NO_EDGE, nm_v::id, T, dim - 1, 0, 1, 2>(gr.getSize(), domain);
+
+ //! Add computation information to each vertex (shape can change)
+ addWeights(gp, HYPERBOLOID);
+
+ //! Init ids vector
+ gp.init_map_ids();
+
+ // Get the number of processing units
+ size_t Np = v_cl.getProcessingUnits();
+
+ // Division of vertices in Np graphs
+ // Put (div+1) vertices in mod graphs
+ // Put div vertices in the rest of the graphs
+ size_t mod_v = gp.getNVertex() % Np;
+ size_t div_v = gp.getNVertex() / Np;
+
+ //! Init vtxdist needed for Parmetis
+ idx_t *vtxdist = new idx_t[Np + 1];
+
+ for (int i = 0; i <= Np; i++) {
+ if (i < mod_v)
+ vtxdist[i] = (div_v + 1) * (i);
+ else
+ vtxdist[i] = (div_v) * (i) + mod_v;
+ }
+
+ // Just for output purpose
+ if (MPI_PROC_ID == 0) {
+ VTKWriter<Graph_CSR<nm_v, nm_e>, GRAPH> gv2(gp);
+ gv2.write("test_graph_0.vtk");
+ }
+
+ //TODO transform in factory
+ //! Create subgraph
+ Graph_CSR<nm_v, nm_e> sub_g;
+
+ //! Put vertices into processor graph (different for each processor)
+ fillSubGraph(sub_g, gp, vtxdist, MPI_PROC_ID, Np);
+
+ // Convert the graph to parmetis format
+ Parmetis<Graph_CSR<nm_v, nm_e>> parmetis_graph(sub_g, Np);
+
+ // Decompose
+ parmetis_graph.decompose<nm_v::proc_id>(vtxdist);
+
+ // Get result partition for this processors
+ idx_t *partition = parmetis_graph.getPartition();
+
+ // Prepare MPI objects for communication
+ MPI_Request *requests_recv = new MPI_Request[Np];
+ MPI_Request *requests_send = new MPI_Request[Np];
+ MPI_Status *statuses = new MPI_Status[Np];
+
+ //-----------------------------------------------
+
+ /*
+
+ /* create a type for struct nm_v */
+ const int nitems = 9;
+ int blocklengths[9] = {1,1};
+ MPI_Datatype types[9] = {MPI_INT, MPI_INT, MPI_INT, MPI_INT, MPI_INT, MPI_INT, MPI_INT, MPI_INT, MPI_INT};
+ MPI_Datatype MPI_VERTEX_TYPE;
+ MPI_Aint offsets[9];
+
+ offsets[0] = offsetof(nm_v, x);
+ offsets[1] = offsetof(nm_v, y);
+ offsets[2] = offsetof(nm_v, z);
+ offsets[3] = offsetof(nm_v, communication);
+ offsets[4] = offsetof(nm_v, computation);
+ offsets[5] = offsetof(nm_v, memory);
+ offsets[6] = offsetof(nm_v, id);
+ offsets[7] = offsetof(nm_v, sub_id);
+ offsets[8] = offsetof(nm_v, proc_id);
+
+ MPI_Type_create_struct(nitems, blocklengths, offsets, types, &MPI_VERTEX_TYPE);
+ MPI_Type_commit(&MPI_VERTEX_TYPE);
+
+ //-----------------------------------------------
+
+ // Prepare vector of arrays to contain all partitions
+ int **partitions = new int*[Np];
+ partitions[MPI_PROC_ID] = partition;
+ for (int i = 0; i < Np; i++) {
+ if (i != MPI_PROC_ID)
+ partitions[i] = new int[gp.getNVertex()];
+ }
+
+ // Exchange partitions with other processors
+ exchangePartitions(partitions, gp.getNVertex(), sub_g.getNVertex(), requests_recv, requests_send, statuses, MPI_PROC_ID, Np);
+
+ // Init data structure to keep trace of new vertices distribution in processors (needed to update main graph)
+ std::vector<size_t> *v_per_proc = new std::vector<size_t>[Np];
+
+ // Update graphs with the new distributions
+ updateGraphs(partitions, gp, sub_g, v_per_proc, vtxdist, statuses, MPI_PROC_ID, Np);
+
+ if (MPI_PROC_ID == 0) {
+ VTKWriter<Graph_CSR<nm_v, nm_e>, GRAPH> gv2(gp);
+ gv2.write("test_graph_2.vtk");
+ }
+
+ // Renumbering subgraph
+ sub_g.reset_map_ids();
+ for (size_t j = vtxdist[MPI_PROC_ID], i = 0; j < vtxdist[MPI_PROC_ID + 1]; j++, i++) {
+ sub_g.set_map_ids(j, sub_g.vertex(i).template get<nm_v::id>());
+ sub_g.vertex(i).template get<nm_v::id>() = j;
+ }
+
+ gp.reset_map_ids();
+ // Renumbering main graph
+ for (size_t p = 0; p < Np; p++) {
+ for (size_t j = vtxdist[p], i = 0; j < vtxdist[p + 1]; j++, i++) {
+ gp.set_map_ids(j, gp.vertex(v_per_proc[p][i]).template get<nm_v::id>());
+ gp.vertex(v_per_proc[p][i]).template get<nm_v::id>() = j;
+ }
+ }
+
+ if (MPI_PROC_ID == 0) {
+ VTKWriter<Graph_CSR<nm_v, nm_e>, GRAPH> gv2(gp);
+ gv2.write("test_graph_3.vtk");
+ }
+
+ // Reset parmetis graph and reconstruct it
+ parmetis_graph.reset(gp);
+
+ // Refine
+ parmetis_graph.refine<nm_v::proc_id>(vtxdist);
+
+ // Get result partition for this processor
+ partition = parmetis_graph.getPartition();
+
+ // Reset partitions array
+ partitions[MPI_PROC_ID] = partition;
+ for (int i = 0; i < Np; i++) {
+ if (i != MPI_PROC_ID) {
+ delete partitions[i];
+ partitions[i] = new int[gp.getNVertex()];
+ }
+ }
+
+ // Reset data structure to keep trace of new vertices distribution in processors (needed to update main graph)
+ for (int i = 0; i < Np; ++i) {
+ v_per_proc[i].clear();
+ }
+
+ // Exchange partitions with other processors
+ exchangePartitions(partitions, gp.getNVertex(), sub_g.getNVertex(), requests_recv, requests_send, statuses, MPI_PROC_ID, Np);
+
+ // Update graphs with the new distributions
+ updateGraphs(partitions, gp, sub_g, v_per_proc, vtxdist, statuses, MPI_PROC_ID, Np);
+
+
+ if (MPI_PROC_ID == 0) {
+ VTKWriter<Graph_CSR<nm_v, nm_e>, GRAPH> gv2(gp);
+ gv2.write("test_graph_4.vtk");
+ }
+
+ /*
+ // fill the structure that store the processor id for each sub-domain
+ fine_s.resize(gr.size());
+
+ // Optimize the decomposition creating bigger spaces
+ // And reducing Ghost over-stress
+ dec_optimizer<dim,Graph_CSR<nm_v,nm_e>> d_o(gp,gr.getSize());
+
+ // set of Boxes produced by the decomposition optimizer
+ openfpm::vector<::Box<dim,size_t>> loc_box;
+
+ // optimize the decomposition
+ d_o.template optimize<nm_v::sub_id,nm_v::id>(gp,p_id,loc_box,box_nn_processor);
+
+ // Initialize ss_box and bbox
+ if (loc_box.size() >= 0)
+ {
+ SpaceBox<dim,size_t> sub_dc = loc_box.get(0);
+ SpaceBox<dim,T> sub_d(sub_dc);
+ sub_d.mul(spacing);
+ sub_d.expand(spacing);
+
+ // Fixing sub-domains to cover all the domain
+
+ // Fixing sub_d
+ // if (loc_box) is a the boundary we have to ensure that the box span the full
+ // domain (avoiding rounding off error)
+ for (size_t i = 0 ; i < dim ; i++)
+ {
+ if (sub_dc.getHigh(i) == cd.getGrid().size(i) - 1)
+ {
+ sub_d.setHigh(i,domain.getHigh(i));
+ }
+ }
+
+ // add the sub-domain
+ sub_domains.add(sub_d);
+
+ ss_box = sub_d;
+ ss_box -= ss_box.getP1();
+ bbox = sub_d;
+ }
+
+ // convert into sub-domain
+ for (size_t s = 1 ; s < loc_box.size() ; s++)
+ {
+ SpaceBox<dim,size_t> sub_dc = loc_box.get(s);
+ SpaceBox<dim,T> sub_d(sub_dc);
+
+ // re-scale and add spacing (the end is the starting point of the next domain + spacing)
+ sub_d.mul(spacing);
+ sub_d.expand(spacing);
+
+ // Fixing sub-domains to cover all the domain
+
+ // Fixing sub_d
+ // if (loc_box) is a the boundary we have to ensure that the box span the full
+ // domain (avoiding rounding off error)
+ for (size_t i = 0 ; i < dim ; i++)
+ {
+ if (sub_dc.getHigh(i) == cd.getGrid().size(i) - 1)
+ {
+ sub_d.setHigh(i,domain.getHigh(i));
+ }
+ }
+
+ // add the sub-domain
+ sub_domains.add(sub_d);
+
+ // Calculate the bound box
+ bbox.enclose(sub_d);
+
+ // Create the smallest box contained in all sub-domain
+ ss_box.contained(sub_d);
+ }
+
+ nn_prcs<dim,T>::create(box_nn_processor, sub_domains);
+
+ // fill fine_s structure
+ // fine_s structure contain the processor id for each sub-sub-domain
+ // with sub-sub-domain we mean the sub-domain decomposition before
+ // running dec_optimizer (before merging sub-domains)
+ it = gp.getVertexIterator();
+
+ while (it.isNext())
+ {
+ size_t key = it.get();
+
+ // fill with the fine decomposition
+ fine_s.get(key) = gp.template vertex_p<nm_v::id>(key);
+
+ ++it;
+ }
+
+ // Get the smallest sub-division on each direction
+ ::Box<dim,T> unit = getSmallestSubdivision();
+ // Get the processor bounding Box
+ ::Box<dim,T> bound = getProcessorBounds();
+
+ // calculate the sub-divisions
+ size_t div[dim];
+ for (size_t i = 0 ; i < dim ; i++)
+ div[i] = (size_t)((bound.getHigh(i) - bound.getLow(i)) / unit.getHigh(i));
+
+ // Create shift
+ Point<dim,T> orig;
+
+ // p1 point of the Processor bound box is the shift
+ for (size_t i = 0 ; i < dim ; i++)
+ orig.get(i) = bound.getLow(i);
+
+ // Initialize the geo_cell structure
+ ie_ghost<dim,T>::Initialize_geo_cell(domain,div,orig);
+ lgeo_cell.Initialize(domain,div,orig);
+ */
+ }
+
+ // Save the ghost boundaries
+ Ghost<dim, T> ghost;
+
+ /*! \brief Create the subspaces that decompose your domain
+ *
+ */
+ void CreateSubspaces() {
+ // Create a grid where each point is a space
+ grid_sm<dim, void> g(div);
+
+ // create a grid_key_dx iterator
+ grid_key_dx_iterator < dim > gk_it(g);
+
+ // Divide the space into subspaces
+ while (gk_it.isNext()) {
+ //! iterate through all subspaces
+ grid_key_dx < dim > key = gk_it.get();
+
+ //! Create a new subspace
+ SpaceBox<dim, T> tmp;
+
+ //! fill with the Margin of the box
+ for (int i = 0; i < dim; i++) {
+ tmp.setHigh(i, (key.get(i) + 1) * spacing[i]);
+ tmp.setLow(i, key.get(i) * spacing[i]);
+ }
+
+ //! add the space box
+ sub_domains.add(tmp);
+
+ // add the iterator
+ ++gk_it;
+ }
+ }
+
+ /* /brief types of weights distributions
+ *
+ */
+ enum weightShape {
+ UNIFORM, SPHERE, HYPERBOLOID
+ };
+
+ /* \brief add vertex weights to the main domain, follow a shape
+ *
+ ** 0 - weights are all 1 on all vertices
+ ** 1 - weights are distributed as a sphere
+ *
+ * \param i id of the shape
+ *
+ */
+ void addWeights(Graph_CSR<nm_v, nm_e> & gp, int i)
+ {
+ float c_x = 0, c_y = 0, c_z = 0 , radius2, eq;
+ float x = 0, y = 0, z = 0;
+
+ switch (i) {
+ case UNIFORM:
+
+ // Add computation information to each vertex
+ for (int i = 0; i < gp.getNVertex(); i++) {
+ gp.vertex(i).template get<nm_v::computation>() = 1;
+ }
+ break;
+ case SPHERE:
+
+ // Fill vertices weights with a sphere (if dim=2 is a circle)
+ radius2 = pow(4, 2);
+ c_x = 2;
+ c_y = 2;
+
+ if(dim == 3)
+ c_z = 2;
+
+ for (int i = 0; i < gp.getNVertex(); i++) {
+ x = gp.vertex(i).template get<nm_v::x>() * 10;
+ y = gp.vertex(i).template get<nm_v::y>() * 10;
+
+ if(dim == 3)
+ z = gp.vertex(i).template get<nm_v::z>() * 10;
+
+ eq = pow((x - c_x), 2) + pow((y - c_y), 2) + pow((z - c_z), 2);
+
+ if (eq <= radius2) {
+ gp.vertex(i).template get<nm_v::computation>() = 5;
+ } else {
+ gp.vertex(i).template get<nm_v::computation>() = 1;
+ }
+ }
+ break;
+ case HYPERBOLOID:
+
+ // Fill vertices weights with a elliptic hyperboloid (if dim=2 is an hyperbole)
+ c_x = 5;
+ c_y = 5;
+
+ if(dim == 3)
+ c_z = 5;
+ for (int i = 0; i < gp.getNVertex(); i++) {
+ x = gp.vertex(i).template get<nm_v::x>() * 10;
+ y = gp.vertex(i).template get<nm_v::y>() * 10;
+
+ if(dim == 3)
+ z = gp.vertex(i).template get<nm_v::z>() * 10;
+
+ eq = - pow((x - c_x), 2)/3 - pow((y - c_y), 2)/3 + pow((z - c_z), 2)/2;
+
+ if (eq >= 1) {
+ gp.vertex(i).template get<nm_v::computation>() = 5;
+ } else {
+ gp.vertex(i).template get<nm_v::computation>() = 1;
+ }
+ }
+ break;
+ }
+ }
+
+ /* \brief fill the graph of the processor with the first decomposition (linear)
+ *
+ * Put vertices into processor graph (different for each processor)
+ *
+ * \param sub_g sub graph to fill
+ * \param gp mai graph, source for the vertices
+ * \param vtxdist array with the distribution of vertices through processors
+ * \param proc_id rank of the processor
+ * \param Np total number of processors
+ */
+ void fillSubGraph(Graph_CSR<nm_v, nm_e> &sub_g, Graph_CSR<nm_v, nm_e> &gp, idx_t* &vtxdist, int proc_id, int Np)
+ {
+
+ for (size_t j = vtxdist[proc_id], local_j = 0; j < vtxdist[proc_id + 1]; j++, local_j++) {
+
+ // Add vertex
+ nm_v pv = gp.vertexById(j);
+ sub_g.addVertex(pv);
+
+ // Add edges of vertex
+ for (size_t s = 0; s < gp.getNChilds(j); s++) {
+ sub_g.template addEdge<NoCheck>(local_j, gp.getChildByVertexId(j, s));
+ }
+ }
+
+ // Just for output purpose
+ if (proc_id == 0) {
+ for (int i = 0; i < Np; i++) {
+ for (size_t j = vtxdist[i]; j < vtxdist[i + 1]; j++) {
+ gp.vertexById(j).template get<nm_v::proc_id>() = i;
+ }
+ }
+ VTKWriter<Graph_CSR<nm_v, nm_e>, GRAPH> gv2(gp);
+ gv2.write("test_graph_1.vtk");
+ }
+ }
+
+ /* \brief exchange partitions with other processors
+ *
+ * \param partitions array to store all the partitions
+ * \param gp_nv number of vertices on main graph
+ * \param sub_g_nv number of vertices on sub graph
+ * \param requests_recv array of requests
+ * \param requests_send array of requests
+ * \param statuses array of statsu objects
+ * \param proc_id current processors rank
+ * \param Np total umber of processors
+ */
+ void exchangePartitions(int** &partitions, int gp_nv, int sub_g_nv, MPI_Request* &requests_recv, MPI_Request* &requests_send,
+ MPI_Status* &statuses, int proc_id, int Np)
+ {
+
+ // Receive other partitions, each partition can contain max NVertex of main graph
+ for (int i = 0; i < Np; i++) {
+ if (i != proc_id)
+ MPI_Irecv(partitions[i], gp_nv, MPI_INT, i, 0, MPI_COMM_WORLD, &requests_recv[i]);
+ }
+
+ // Send processor partition to other processors
+ for (int i = 0; i < Np; i++) {
+ if (i != proc_id)
+ MPI_Isend(partitions[proc_id], sub_g_nv, MPI_INT, i, 0, MPI_COMM_WORLD, &requests_send[i]);
+ }
+
+ // Wait for all partitions from other processors
+ for (int i = 0; i < Np; i++) {
+ if (i != proc_id)
+ MPI_Wait(&requests_recv[i], &statuses[i]);
+ }
+ }
+
+ /* \brief update main graph ad subgraph with the partition in partitions param
+ *
+ * \param partitions array storing all the partitions
+ * \param gp main graph
+ * \param sub_g sub graph
+ * \param v_per_proc array needed to recontruct the main graph
+ * \param vtxdist array with the distribution of vertices through processors
+ * \param statuses array of statsu objects
+ * \param proc_id current processors rank
+ * \param Np total umber of processors
+ */
+ void updateGraphs(int** &partitions,Graph_CSR<nm_v, nm_e> &gp, Graph_CSR<nm_v, nm_e> &sub_g, std::vector<size_t>* &v_per_proc, idx_t* &vtxdist, MPI_Status* &statuses, int proc_id, int Np) {
+
+ int size_p = sub_g.getNVertex();
+ int local_j = 0;
+ sub_g.clear();
+
+ // Init n_vtxdist to gather informations about the new decomposition
+ idx_t *n_vtxdist = new idx_t[Np + 1];
+ for (int i = 0; i <= Np; i++)
+ n_vtxdist[i] = 0;
+
+ // Update main graph with other partitions made by Parmetis in other processors and the local partition
+ for (int i = 0; i < Np; i++) {
+
+ int ndata = 0;
+ MPI_Get_count(&statuses[i], MPI_INT, &ndata);
+
+ if (i == proc_id)
+ ndata = size_p;
+
+ // Update the main graph with received informations
+ for (int k = 0, l = vtxdist[i]; k < ndata && l < vtxdist[i + 1]; k++, l++) {
+
+ // Create new n_vtxdist (1) (just count processors vertices)
+ n_vtxdist[partitions[i][k] + 1]++;
+
+ // Update proc id in the vertex
+ gp.vertexById(l).template get<nm_v::proc_id>() = partitions[i][k];
+
+ // Add vertex to temporary structure of distribution (needed to update main graph)
+ v_per_proc[partitions[i][k]].push_back(gp.vertexById(l).template get<nm_v::id>());
+
+ // Add vertices belonging to this processor in sub graph
+ if (partitions[i][k] == proc_id) {
+
+ nm_v pv = gp.vertexById(l);
+ sub_g.addVertex(pv);
+
+ // Add edges of vertex
+ for (size_t s = 0; s < gp.getNChildsByVertexId(l); s++) {
+ sub_g.template addEdge<NoCheck>(local_j, gp.getChildByVertexId(l, s));
+ }
+
+ local_j++;
+ }
+ }
+ }
+
+ // Create new n_vtxdist (2) (write boundaries)
+ for (int i = 2; i <= Np; i++) {
+ n_vtxdist[i] += n_vtxdist[i - 1];
+ }
+ // Copy the new decomposition in the main vtxdist
+ for (int i = 0; i <= Np; i++) {
+ vtxdist[i] = n_vtxdist[i];
+ }
+ }
+
+ // Heap memory receiver
+ HeapMemory hp_recv;
+
+ // vector v_proc
+ openfpm::vector<size_t> v_proc;
+
+ // Receive counter
+ size_t recv_cnt;
+
+public:
+
+ /*! \brief Basic decomposition constructor
+ *
+ * \param v_cl Virtual cluster, used internally to handle or pipeline communication
+ *
+ */
+ BasicDecomposition(Vcluster & v_cl) :
+ nn_prcs<dim, T>(v_cl), v_cl(v_cl) {
+ // Reset the box to zero
+ bbox.zero();
+ }
+
+ //! Basic decomposition destructor
+ ~BasicDecomposition() {
+ }
+
+ // openfpm::vector<size_t> ids;
+
+ /*! \brief class to select the returned id by ghost_processorID
+ *
+ */
+ class box_id {
+ public:
+ /*! \brief Return the box id
+ *
+ * \param p structure containing the id informations
+ * \param b_id box_id
+ *
+ * \return box id
+ *
+ */
+ inline static size_t id(p_box<dim, T> & p, size_t b_id) {
+ return b_id;
+ }
+ };
+
+ /*! \brief class to select the returned id by ghost_processorID
+ *
+ */
+ class processor_id {
+ public:
+ /*! \brief Return the processor id
+ *
+ * \param p structure containing the id informations
+ * \param b_id box_id
+ *
+ * \return processor id
+ *
+ */
+ inline static size_t id(p_box<dim, T> & p, size_t b_id) {
+ return p.proc;
+ }
+ };
+
+ /*! \brief class to select the returned id by ghost_processorID
+ *
+ */
+ class lc_processor_id {
+ public:
+ /*! \brief Return the near processor id
+ *
+ * \param p structure containing the id informations
+ * \param b_id box_id
+ *
+ * \return local processor id
+ *
+ */
+ inline static size_t id(p_box<dim, T> & p, size_t b_id) {
+ return p.lc_proc;
+ }
+ };
+
+ /*! It calculate the internal ghost boxes
+ *
+ * Example: Processor 10 calculate
+ * B8_0 B9_0 B9_1 and B5_0
+ *
+ *
+ *
+ \verbatim
+
+ +----------------------------------------------------+
+ | |
+ | Processor 8 |
+ | Sub-domain 0 +-----------------------------------+
+ | | |
+ | | |
+ ++--------------+---+---------------------------+----+ Processor 9 |
+ | | | B8_0 | | Subdomain 0 |
+ | +------------------------------------+ |
+ | | | | | |
+ | | | XXXXXXXXXXXXX XX |B9_0| |
+ | | B | X Processor 10 X | | |
+ | Processor 5 | 5 | X Sub-domain 0 X | | |
+ | Subdomain 0 | _ | X X +----------------------------------------+
+ | | 0 | XXXXXXXXXXXXXXXX | | |
+ | | | | | |
+ | | | | | Processor 9 |
+ | | | |B9_1| Subdomain 1 |
+ | | | | | |
+ | | | | | |
+ | | | | | |
+ +--------------+---+---------------------------+----+ |
+ | |
+ +-----------------------------------+
+
+ \endverbatim
+
+ and also
+ G8_0 G9_0 G9_1 G5_0 (External ghost boxes)
+
+ \verbatim
+
+ +----------------------------------------------------+
+ | |
+ | Processor 8 |
+ | Sub-domain 0 +-----------------------------------+
+ | +---------------------------------------------+ |
+ | | G8_0 | | |
+ ++--------------+------------------------------------+ | Processor 9 |
+ | | | | | Subdomain 0 |
+ | | | |G9_0| |
+ | | | | | |
+ | | | XXXXXXXXXXXXX XX | | |
+ | | | X Processor 10 X | | |
+ | Processor|5 | X Sub-domain 0 X | | |
+ | Subdomain|0 | X X +-----------------------------------+
+ | | | XXXXXXXXXXXXXXXX | | |
+ | | G | | | |
+ | | 5 | | | Processor 9 |
+ | | | | | | Subdomain 1 |
+ | | 0 | |G9_1| |
+ | | | | | |
+ | | | | | |
+ +--------------+------------------------------------+ | |
+ | | | |
+ +----------------------------------------+----+------------------------------+
+
+
+ \endverbatim
+
+ *
+ *
+ *
+ * \param ghost margins for each dimensions (p1 negative part) (p2 positive part)
+ *
+ *
+ \verbatim
+ ^ p2[1]
+ |
+ |
+ +----+----+
+ | |
+ | |
+ p1[0]<-----+ +----> p2[0]
+ | |
+ | |
+ +----+----+
+ |
+ v p1[1]
+
+ \endverbatim
+
+ *
+ *
+ */
+ void calculateGhostBoxes() {
+#ifdef DEBUG
+ // the ghost margins are assumed to be smaller
+ // than one sub-domain
+
+ for (size_t i = 0; i < dim; i++) {
+ if (ghost.template getLow(i) >= domain.template getHigh(i) / gr.size(i)
+ || ghost.template getHigh(i) >= domain.template getHigh(i) / gr.size(i)) {
+ std::cerr << "Error " << __FILE__ << ":" << __LINE__ << " : Ghost are bigger than one domain" << "\n";
+ }
+ }
+#endif
+
+ // Intersect all the local sub-domains with the sub-domains of the contiguous processors
+
+ // create the internal structures that store ghost information
+ ie_ghost<dim, T>::create_box_nn_processor_ext(v_cl, ghost, sub_domains, box_nn_processor, *this);
+ ie_ghost<dim, T>::create_box_nn_processor_int(v_cl, ghost, sub_domains, box_nn_processor, *this);
+
+ // ebox must come after ibox (in this case)
+
+ ie_loc_ghost<dim, T>::create_loc_ghost_ibox(ghost, sub_domains);
+ ie_loc_ghost<dim, T>::create_loc_ghost_ebox(ghost, sub_domains);
+
+ // get the smallest sub-domain dimension on each direction
+ for (size_t i = 0; i < dim; i++) {
+ if (ghost.template getLow(i) >= ss_box.getHigh(i) || ghost.template getHigh(i) >= domain.template getHigh(i) / gr.size(i)) {
+ std::cerr << "Error " << __FILE__ << ":" << __LINE__ << " : Ghost are bigger than one domain" << "\n";
+ }
+ }
+ }
+
+ /*! \brief The default grid size
+ *
+ * The default grid is always an isotropic grid that adapt with the number of processors,
+ * it define in how many cell it will be divided the space for a particular required minimum
+ * number of sub-domain
+ *
+ */
+ static size_t getDefaultGrid(size_t n_sub) {
+ // Calculate the number of sub-sub-domain on
+ // each dimension
+ return openfpm::math::round_big_2(pow(n_sub, 1.0 / dim));
+ }
+
+ /*! \brief Given a point return in which processor the particle should go
+ *
+ * \return processorID
+ *
+ */
+ template<typename Mem> size_t inline processorID(encapc<1, Point<dim, T>, Mem> p) {
+ return fine_s.get(cd.getCell(p));
+ }
+
+ // Smallest subdivision on each direction
+ ::Box<dim, T> ss_box;
+
+ /*! \brief Get the smallest subdivision of the domain on each direction
+ *
+ * \return a box p1 is set to zero
+ *
+ */
+ const ::Box<dim, T> & getSmallestSubdivision() {
+ return ss_box;
+ }
+
+ /*! \brief Given a point return in which processor the particle should go
+ *
+ * \return processorID
+ *
+ */
+
+ size_t inline processorID(const T (&p)[dim]) const {
+ return fine_s.get(cd.getCell(p));
+ }
+
+ /*! \brief Set the parameter of the decomposition
+ *
+ * \param div_ storing into how many domain to decompose on each dimension
+ * \param domain_ domain to decompose
+ *
+ */
+ void setParameters(const size_t (&div_)[dim], Domain<dim, T> domain_, Ghost<dim, T> ghost = Ghost<dim, T>()) {
+ // set the ghost
+ this->ghost = ghost;
+ // Set the decomposition parameters
+
+ gr.setDimensions(div_);
+ domain = domain_;
+ cd.setDimensions(domain, div_, 0);
+
+ //! Create the decomposition
+
+ CreateDecomposition(v_cl);
+ }
+
+ /*! \brief Get the number of local sub-domains
+ *
+ * \return the number of sub-domains
+ *
+ */
+ size_t getNLocalHyperCube() {
+ return sub_domains.size();
+ }
+
+ /*! \brief Get the local sub-domain
+ *
+ * \param i (each local processor can have more than one sub-domain)
+ * \return the sub-domain
+ *
+ */
+ SpaceBox<dim, T> getLocalHyperCube(size_t lc) {
+ // Create a space box
+ SpaceBox<dim, T> sp;
+
+ // fill the space box
+
+ for (size_t k = 0; k < dim; k++) {
+ // create the SpaceBox Low and High
+ sp.setLow(k, sub_domains.template get < Box::p1 > (lc)[k]);
+ sp.setHigh(k, sub_domains.template get < Box::p2 > (lc)[k]);
+ }
+
+ return sp;
+ }
+
+ /*! \brief Get the local sub-domain with ghost extension
+ *
+ * \param i (each local processor can have more than one sub-domain)
+ * \return the sub-domain
+ *
+ */
+ SpaceBox<dim, T> getSubDomainWithGhost(size_t lc) {
+ // Create a space box
+ SpaceBox<dim, T> sp = sub_domains.get(lc);
+
+ // enlarge with ghost
+ sp.enlarge(ghost);
+
+ return sp;
+ }
+
+ /*! \brief Return the structure that store the physical domain
+ *
+ * \return The physical domain
+ *
+ */
+ Domain<dim, T> & getDomain() {
+ return domain;
+ }
+
+ /*! \brief Check if the particle is local
+ *
+ * \param p object position
+ *
+ * \return true if it is local
+ *
+ */
+ template<typename Mem> bool isLocal(const encapc<1, Point<dim, T>, Mem> p) const {
+ return processorID<Mem>(p) == v_cl.getProcessUnitID();
+ }
+
+ /*! \brief Check if the particle is local
+ *
+ * \param p object position
+ *
+ * \return true if it is local
+ *
+ */
+ bool isLocal(const T (&pos)[dim]) const {
+ return processorID(pos) == v_cl.getProcessUnitID();
+ }
+
+ ::Box<dim, T> bbox;
+
+ /*! \brief Return the bounding box containing union of all the sub-domains for the local processor
+ *
+ * \return The bounding box
+ *
+ */
+ ::Box<dim, T> & getProcessorBounds() {
+ return bbox;
+ }
+
+ ////////////// Functions to get decomposition information ///////////////
+
+ /*! \brief Write the decomposition as VTK file
+ *
+ * The function generate several files
+ *
+ * * subdomains_X.vtk domain for the local processor (X) as union of sub-domain
+ * * subdomains_adjacent_X.vtk sub-domains adjacent to the local processor (X)
+ * * internal_ghost_X.vtk Internal ghost boxes for the local processor (X)
+ * * external_ghost_X.vtk External ghost boxes for the local processor (X)
+ * * local_internal_ghost_X.vtk internal local ghost boxes for the local processor (X)
+ * * local_external_ghost_X.vtk external local ghost boxes for the local processor (X)
+ *
+ * where X is the local processor rank
+ *
+ * \param output directory where to write the files
+ *
+ */
+ bool write(std::string output) const {
+ //! subdomains_X.vtk domain for the local processor (X) as union of sub-domain
+ VTKWriter<openfpm::vector<::SpaceBox<dim, T>>, VECTOR_BOX> vtk_box1;
+ vtk_box1.add(sub_domains);
+ vtk_box1.write(output + std::string("subdomains_") + std::to_string(v_cl.getProcessUnitID()) + std::string(".vtk"));
+
+ nn_prcs<dim, T>::write(output);
+ ie_ghost<dim, T>::write(output, v_cl.getProcessUnitID());
+ ie_loc_ghost<dim, T>::write(output, v_cl.getProcessUnitID());
+
+ return true;
+ }
+
+ /*! \brief function to check the consistency of the information of the decomposition
+ *
+ * \return false if is inconsistent
+ *
+ */
+ bool check_consistency() {
+ if (ie_loc_ghost<dim, T>::check_consistency(getNLocalHyperCube()) == false)
+ return false;
+
+ return true;
+ }
+
+ void debugPrint() {
+ std::cout << "Subdomains\n";
+ for (size_t p = 0; p < sub_domains.size(); p++) {
+ std::cout << ::SpaceBox<dim, T>(sub_domains.get(p)).toString() << "\n";
+ }
+
+ std::cout << "External ghost box\n";
+
+ for (size_t p = 0; p<nn_prcs < dim, T>::getNNProcessors(); p++) {
+ for (size_t i = 0; i<ie_ghost < dim, T>::getProcessorNEGhost(p); i++) {
+ std::cout << ie_ghost<dim, T>::getProcessorEGhostBox(p, i).toString() << " prc=" << nn_prcs<dim, T>::IDtoProc(p)
+ << " id=" << ie_ghost<dim, T>::getProcessorEGhostId(p, i) << "\n";
+ }
+ }
+
+ std::cout << "Internal ghost box\n";
+
+ for (size_t p = 0; p<nn_prcs < dim, T>::getNNProcessors(); p++) {
+ for (size_t i = 0; i<ie_ghost < dim, T>::getProcessorNIGhost(p); i++) {
+ std::cout << ie_ghost<dim, T>::getProcessorIGhostBox(p, i).toString() << " prc=" << nn_prcs<dim, T>::IDtoProc(p)
+ << " id=" << ie_ghost<dim, T>::getProcessorIGhostId(p, i) << "\n";
+ }
+ }
+ }
+
+};
+
+#endif
diff --git a/src/Decomposition/BasicDecomposition_unit_test.hpp b/src/Decomposition/BasicDecomposition_unit_test.hpp
new file mode 100644
index 000000000..9ae1a7d94
--- /dev/null
+++ b/src/Decomposition/BasicDecomposition_unit_test.hpp
@@ -0,0 +1,83 @@
+#ifndef BASICDECOMPOSITION_UNIT_TEST_HPP
+#define BASICDECOMPOSITION_UNIT_TEST_HPP
+
+#include "BasicDecomposition.hpp"
+#include "util/mathutil.hpp"
+
+BOOST_AUTO_TEST_SUITE( BasicDecomposition_test )
+
+#define SUB_UNIT_FACTOR 1024
+
+BOOST_AUTO_TEST_CASE( BasicDecomposition_test_use)
+{
+ // Vcluster
+ Vcluster & vcl = *global_v_cluster;
+
+ // Initialize the global VCluster
+ init_global_v_cluster(&boost::unit_test::framework::master_test_suite().argc,&boost::unit_test::framework::master_test_suite().argv);
+
+ //! [Create CartDecomposition]
+ BasicDecomposition<3,float> dec(vcl);
+
+ // Physical domain
+ Box<3,float> box({0.0,0.0,0.0},{1.0,1.0,1.0});
+ size_t div[3];
+
+ // Get the number of processor and calculate the number of sub-domain
+ // for each processor (SUB_UNIT_FACTOR=64)
+ size_t n_proc = vcl.getProcessingUnits();
+ size_t n_sub = n_proc * SUB_UNIT_FACTOR;
+
+ // Set the number of sub-domains on each dimension (in a scalable way)
+ for (int i = 0 ; i < 3 ; i++)
+ {div[i] = openfpm::math::round_big_2(pow(n_sub,1.0/3));}
+
+ // Define ghost
+ Ghost<3,float> g(0.01);
+
+ // Decompose
+ dec.setParameters(div,box,g);
+
+ // create a ghost border
+ dec.calculateGhostBoxes();
+
+ //! [Create BasicDecomposition]
+
+ // For each calculated ghost box
+ for (size_t i = 0 ; i < dec.getNIGhostBox() ; i++)
+ {
+ SpaceBox<3,float> b = dec.getIGhostBox(i);
+ size_t proc = dec.getIGhostBoxProcessor(i);
+
+ // sample one point inside the box
+ Point<3,float> p = b.rnd();
+
+ // Check that ghost_processorsID return that processor number
+ const openfpm::vector<size_t> & pr = dec.template ghost_processorID<BasicDecomposition<3,float>::processor_id>(p);
+
+ bool found = false;
+
+ for (size_t j = 0; j < pr.size() ; j++)
+ {
+ if (pr.get(j) == proc)
+ {found = true; break;}
+ }
+
+ if (found == false)
+ {
+ const openfpm::vector<size_t> pr3 = dec.template ghost_processorID<BasicDecomposition<3,float>::processor_id>(p);
+ }
+
+ BOOST_REQUIRE_EQUAL(found,true);
+ }
+
+ // Check the consistency
+
+ bool val = dec.check_consistency();
+ BOOST_REQUIRE_EQUAL(val,true);
+}
+
+BOOST_AUTO_TEST_SUITE_END()
+
+
+#endif
diff --git a/src/parmetis_util.hpp b/src/parmetis_util.hpp
new file mode 100644
index 000000000..83bf2bca1
--- /dev/null
+++ b/src/parmetis_util.hpp
@@ -0,0 +1,486 @@
+/*
+ * parmetis_util.hpp
+ *
+ * Created on: Oct 07, 2015
+ * Author: Antonio Leo
+ */
+
+#ifndef PARMETIS_UTIL_HPP
+#define PARMETIS_UTIL_HPP
+
+#include <iostream>
+#include "parmetis.h"
+#include "VTKWriter.hpp"
+
+/*! \brief Metis graph structure
+ *
+ * Metis graph structure
+ *
+ */
+struct Parmetis_graph {
+ //! The number of vertices in the graph
+ idx_t * nvtxs;
+
+ //! number of balancing constrains
+ //! more practical, are the number of weights for each vertex
+ //! PS even we you specify vwgt == NULL ncon must be set at leat to
+ //! one
+ idx_t * ncon;
+
+ //! For each vertex it store the adjacency lost start for the vertex i
+ idx_t * xadj;
+
+ //! For each vertex it store a list of all neighborhood vertex
+ idx_t * adjncy;
+
+ //! Array that store the weight for each vertex
+ idx_t * vwgt;
+
+ //! Array of the vertex size, basically is the total communication amount
+ idx_t * vsize;
+
+ //! The weight of the edge
+ idx_t * adjwgt;
+
+ //! number of part to partition the graph
+ idx_t * nparts;
+
+ //! Desired weight for each partition (one for each constrain)
+ real_t * tpwgts;
+
+ //! For each partition load imbalance tollerated
+ real_t * ubvec;
+
+ //! Additional option for the graph partitioning
+ idx_t * options;
+
+ //! return the total comunication cost for each partition
+ idx_t * objval;
+
+ //! Is a output vector containing the partition for each vertex
+ idx_t * part;
+
+ //! Upon successful completion, the number of edges that are cut by the partitioning is written to this parameter.
+ idx_t * edgecut;
+
+ //! This parameter describes the ratio of inter-processor communication time compared to data redistri- bution time. It should be set between 0.000001 and 1000000.0. If ITR is set high, a repartitioning with a low edge-cut will be computed. If it is set low, a repartitioning that requires little data redistri- bution will be computed. Good values for this parameter can be obtained by dividing inter-processor communication time by data redistribution time. Otherwise, a value of 1000.0 is recommended.
+ real_t * itr;
+
+ //! This is used to indicate the numbering scheme that is used for the vtxdist, xadj, adjncy, and part arrays. (0 for C-style, start from 0 index)
+ idx_t * numflag;
+
+ //! This is used to indicate if the graph is weighted. wgtflag can take one of four values:
+ // 0 No weights (vwgt and adjwgt are both NULL).
+ // 1 Weights on the edges only (vwgt is NULL).
+ // 2 Weights on the vertices only (adjwgt is NULL).
+ // 3 Weights on both the vertices and edges.
+ idx_t * wgtflag;
+};
+
+//! Balance communication and computation
+#define BALANCE_CC 1
+//! Balance communication computation and memory
+#define BALANCE_CCM 2
+//! Balance computation and comunication and others
+#define BALANCE_CC_O(c) c+1
+
+/*! \brief Helper class to define Metis graph
+ *
+ * \tparam graph structure that store the graph
+ *
+ */
+template<typename Graph>
+class Parmetis {
+ // Graph in metis reppresentation
+ Parmetis_graph Mg;
+
+ // Original graph
+ Graph & g;
+
+ // Communticator for OpenMPI
+ MPI_Comm comm = NULL;
+
+ // Process rank information
+ int MPI_PROC_ID = 0;
+
+ /*! \brief Construct Adjacency list
+ *
+ * \param g Reference graph to get informations
+ *
+ */
+ void constructAdjList(Graph &refGraph) {
+ // create xadj and adjlist
+ Mg.vwgt = new idx_t[g.getNVertex()];
+ Mg.xadj = new idx_t[g.getNVertex() + 1];
+ Mg.adjncy = new idx_t[g.getNEdge()];
+
+ //! starting point in the adjacency list
+ size_t prev = 0;
+
+ // actual position
+ size_t id = 0;
+
+ // property id
+ size_t real_id;
+
+ // boolan to check if ref is the main graph
+ bool main = refGraph.getNVertex() != g.getNVertex();
+
+ // for each vertex calculate the position of the starting point in the adjacency list
+ for (size_t i = 0; i < g.getNVertex(); i++) {
+ // Add weight to vertex
+ Mg.vwgt[i] = g.vertex(i).template get<nm_v::computation>();
+
+ // Calculate the starting point in the adjacency list
+ Mg.xadj[id] = prev;
+
+ if (main)
+ real_id = g.get_real_id(g.vertex(i).template get<nm_v::id>());
+ else
+ real_id = i;
+
+ // Create the adjacency list and the weights for edges
+ for (size_t s = 0; s < refGraph.getNChilds(real_id); s++) {
+ size_t child = refGraph.getChild(real_id, s);
+
+ if (main)
+ Mg.adjncy[prev + s] = refGraph.vertex(child).template get<nm_v::id>();
+ else
+ Mg.adjncy[prev + s] = child;
+ }
+
+ // update the position for the next vertex
+ prev += refGraph.getNChilds(real_id);
+
+ id++;
+ }
+
+ // Fill the last point
+ Mg.xadj[id] = prev;
+
+ /*
+ std::cout << MPI_PROC_ID << "\n";
+ for(int i=0; i<= g.getNVertex();i++){
+ std::cout << Mg.xadj[i] << " ";
+ }
+ std::cout << "\n\n";
+ for(int i=0; i< g.getNEdge();i++){
+ std::cout << Mg.adjncy[i] << " ";
+ }
+ std::cout << "\n\n";
+ */
+
+ }
+
+ /*! \brief Construct Adjacency list
+ *
+ * \param g Reference graph to get informations
+ *
+ */
+ /*
+ void constructAdjList(Graph &refGraph, idx_t* &old_vtxdist ) {
+ // create xadj and adjlist
+ Mg.vwgt = new idx_t[g.getNVertex()];
+ Mg.xadj = new idx_t[g.getNVertex() + 1];
+ Mg.adjncy = new idx_t[g.getNEdge()];
+
+ //! starting point in the adjacency list
+ size_t prev = 0;
+
+ // actual position
+ size_t id = 0;
+
+ // for each vertex calculate the position of the starting point in the adjacency list
+ for (size_t i = 0; i < g.getNVertex(); i++) {
+
+ // Add weight to vertex
+ Mg.vwgt[i] = g.vertex(i).template get<nm_v::computation>();
+
+ // Calculate the starting point in the adjacency list
+ Mg.xadj[id] = prev;
+
+ // Create the adjacency list and the weights for edges
+ for (size_t s = 0; s < refGraph.getNChilds(i); s++) {
+
+ size_t child = refGraph.getChild(i, s);
+
+ // Check if child is not in this processor
+ if(child > old_vtxdist[MPI_PROC_ID+1] || child < old_vtxdist[MPI_PROC_ID])
+
+ Mg.adjncy[prev + s] = child;
+ }
+
+ // update the position for the next vertex
+ prev += refGraph.getNChilds(i);
+
+ id++;
+ }
+
+ // Fill the last point
+ Mg.xadj[id] = prev;
+
+
+ std::cout << MPI_PROC_ID << "\n";
+ for(int i=0; i<= g.getNVertex();i++){
+ std::cout << Mg.xadj[i] << " ";
+ }
+ std::cout << "\n\n";
+ for(int i=0; i< g.getNEdge();i++){
+ std::cout << Mg.adjncy[i] << " ";
+ }
+ std::cout << "\n\n";
+
+
+ }
+ */
+public:
+
+ /*! \brief Constructor
+ *
+ * Construct a metis graph from Graph_CSR
+ *
+ * \param g Graph we want to convert to decompose
+ * \param nc number of partitions
+ *
+ */
+ Parmetis(Graph & g, size_t nc) :
+ g(g) {
+ // Init OpenMPI structures
+
+ MPI_Comm_dup(MPI_COMM_WORLD, &comm);
+ MPI_Comm_rank(MPI_COMM_WORLD, &MPI_PROC_ID);
+
+ // Get the number of vertex
+
+ Mg.nvtxs = new idx_t[1];
+ Mg.nvtxs[0] = g.getNVertex();
+
+ // Set the number of constrains
+
+ Mg.ncon = new idx_t[1];
+ Mg.ncon[0] = 1;
+
+ // Set to null the weight of the vertex
+
+ Mg.vwgt = NULL;
+
+ // construct the adjacency list
+
+ constructAdjList(g);
+
+ // Put the total communication size to NULL
+
+ Mg.vsize = NULL;
+
+ // Set to null the weight of the edge
+
+ Mg.adjwgt = NULL;
+
+ // Set the total number of partitions
+
+ Mg.nparts = new idx_t[1];
+ Mg.nparts[0] = nc;
+
+ //! Set option for the graph partitioning (set as default)
+
+ Mg.options = new idx_t[4];
+ Mg.options[0] = 1;
+ Mg.options[1] = 3;
+ Mg.options[2] = 0;
+ Mg.options[3] = 0;
+
+ //! is an output vector containing the partition for each vertex
+
+ Mg.part = new idx_t[g.getNVertex()];
+ for (int i = 0; i < g.getNVertex(); i++)
+ Mg.part[i] = MPI_PROC_ID;
+
+ //! adaptiveRepart itr value
+ Mg.itr = new real_t[1];
+ Mg.itr[0] = 1000.0;
+
+ //! init tpwgts to have balanced vertices and ubvec
+
+ Mg.tpwgts = new real_t[Mg.nparts[0]];
+ Mg.ubvec = new real_t[Mg.nparts[0]];
+
+ for (int s = 0; s < Mg.nparts[0]; s++) {
+ Mg.tpwgts[s] = 1.0 / Mg.nparts[0];
+ Mg.ubvec[s] = 1.05;
+ }
+
+ Mg.edgecut = new idx_t[1];
+ Mg.edgecut[0] = 0;
+
+ //! This is used to indicate the numbering scheme that is used for the vtxdist, xadj, adjncy, and part arrays. (0 for C-style, start from 0 index)
+ Mg.numflag = new idx_t[1];
+ Mg.numflag[0] = 0;
+
+ //! This is used to indicate if the graph is weighted. wgtflag can take one of four values:
+ Mg.wgtflag = new idx_t[1];
+ Mg.wgtflag[0] = 2;
+ }
+
+ //TODO deconstruct new variables
+ /*! \brief destructor
+ *
+ * Destructor, It destroy all the memory allocated
+ *
+ */
+ ~Parmetis() {
+ // Deallocate the Mg structure
+ if (Mg.nvtxs != NULL) {
+ delete[] Mg.nvtxs;
+ }
+
+ if (Mg.ncon != NULL) {
+ delete[] Mg.ncon;
+ }
+
+ if (Mg.xadj != NULL) {
+ delete[] Mg.xadj;
+ }
+
+ if (Mg.adjncy != NULL) {
+ delete[] Mg.adjncy;
+ }
+
+ if (Mg.vwgt != NULL) {
+ delete[] Mg.vwgt;
+ }
+
+ if (Mg.adjwgt != NULL) {
+ delete[] Mg.adjwgt;
+ }
+
+ if (Mg.nparts != NULL) {
+ delete[] Mg.nparts;
+ }
+
+ if (Mg.tpwgts != NULL) {
+ delete[] Mg.tpwgts;
+ }
+
+ if (Mg.ubvec != NULL) {
+ delete[] Mg.ubvec;
+ }
+
+ if (Mg.options != NULL) {
+ delete[] Mg.options;
+ }
+
+ if (Mg.part != NULL) {
+ delete[] Mg.part;
+ }
+
+ if (Mg.edgecut != NULL) {
+ delete[] Mg.edgecut;
+ }
+
+ if (Mg.numflag != NULL) {
+ delete[] Mg.numflag;
+ }
+
+ if (Mg.wgtflag != NULL) {
+ delete[] Mg.wgtflag;
+ }
+ }
+
+ /*! \brief Decompose the graph
+ *
+ * \tparam i which property store the decomposition
+ *
+ */
+
+ template<unsigned int i>
+ void decompose(idx_t *vtxdist) {
+
+ // Decompose
+ ParMETIS_V3_PartKway(vtxdist, Mg.xadj, Mg.adjncy, Mg.vwgt, Mg.adjwgt, Mg.wgtflag, Mg.numflag, Mg.ncon, Mg.nparts, Mg.tpwgts,
+ Mg.ubvec, Mg.options, Mg.edgecut, Mg.part, &comm);
+ /*
+ ParMETIS_V3_AdaptiveRepart( vtxdist, Mg.xadj,Mg.adjncy,Mg.vwgt,Mg.vsize,Mg.adjwgt, &(Mg.wgtflag), &(Mg.numflag),
+ Mg.ncon, Mg.nparts, Mg.tpwgts, Mg.ubvec, &(Mg.itr), Mg.options, &(Mg.edgecut),
+ Mg.part, &comm );
+ */
+
+ // For each vertex store the processor that contain the data
+ for (size_t j = 0, id = 0; j < g.getNVertex(); j++, id++) {
+
+ g.vertex(j).template get<i>() = Mg.part[id];
+
+ }
+
+ }
+
+ /*! \brief Refine the graph
+ *
+ * \tparam i which property store the refined decomposition
+ *
+ */
+
+ template<unsigned int i>
+ void refine(idx_t *vtxdist) {
+
+ // Refine
+
+ ParMETIS_V3_RefineKway(vtxdist, Mg.xadj, Mg.adjncy, Mg.vwgt, Mg.adjwgt, Mg.wgtflag, Mg.numflag, Mg.ncon, Mg.nparts, Mg.tpwgts,
+ Mg.ubvec, Mg.options, Mg.edgecut, Mg.part, &comm);
+
+ // For each vertex store the processor that contain the data
+
+ for (size_t j = 0, id = 0; j < g.getNVertex(); j++, id++) {
+
+ g.vertex(j).template get<i>() = Mg.part[id];
+
+ }
+ }
+
+ /*! \brief Get graph partition vector
+ *
+ */
+ idx_t* getPartition() {
+ return Mg.part;
+ }
+
+ /*! \brief Reset graph and reconstruct it
+ *
+ */
+ void reset(Graph & mainGraph) {
+ // Deallocate the graph structures
+
+ if (Mg.xadj != NULL) {
+ delete[] Mg.xadj;
+ }
+
+ if (Mg.adjncy != NULL) {
+ delete[] Mg.adjncy;
+ }
+
+ if (Mg.vwgt != NULL) {
+ delete[] Mg.vwgt;
+ }
+
+ if (Mg.adjwgt != NULL) {
+ delete[] Mg.adjwgt;
+ }
+
+ if (Mg.part != NULL) {
+ delete[] Mg.part;
+ }
+
+ Mg.nvtxs[0] = g.getNVertex();
+
+ Mg.part = new idx_t[g.getNVertex()];
+
+ for (int i = 0; i < g.getNVertex(); i++)
+ Mg.part[i] = MPI_PROC_ID;
+
+ // construct the adjacency list
+ constructAdjList(mainGraph);
+
+ }
+
+};
+
+#endif
--
GitLab