/* * (C) Copyright 2013 ECMWF. * * This software is licensed under the terms of the Apache Licence Version 2.0 * which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. * In applying this licence, ECMWF does not waive the privileges and immunities * granted to it by virtue of its status as an intergovernmental organisation * nor does it submit to any jurisdiction. */ #include "atlas/grid/detail/partitioner/EqualRegionsPartitioner.h" #include #include #include #include #include #include #include "atlas/grid/Iterator.h" #include "atlas/grid/StructuredGrid.h" #include "atlas/parallel/mpi/Buffer.h" #include "atlas/parallel/mpi/mpi.h" #include "atlas/parallel/omp/sort.h" #include "atlas/runtime/Exception.h" #include "atlas/runtime/Log.h" #include "atlas/runtime/Trace.h" #include "atlas/util/MicroDeg.h" #include "atlas/util/vector.h" using atlas::util::microdeg; namespace atlas { namespace grid { namespace detail { namespace partitioner { namespace { static bool valid_mpi_size(size_t size) { return size < size_t(std::numeric_limits::max()); } template struct range_t { Iterator b; Sentinel e; Iterator begin() const { return b; } Sentinel end() const { return e; } // unused: // bool empty() const { return begin() == end(); } }; template range_t range(Iterator b, Sentinel e) { return {b, e}; } template range_t subrange( const Container& c, std::initializer_list&& _range) { const Int rbegin = _range.begin()[0]; const Int rend = _range.begin()[1]; return range(c.begin() + rbegin, c.begin() + rend); } } // namespace double gamma(const double& x) { double p[14]; double w, y; int k, n; p[0] = 0.999999999999999990e+00; p[1] = -0.422784335098466784e+00; p[2] = -0.233093736421782878e+00; p[3] = 0.191091101387638410e+00; p[4] = -0.024552490005641278e+00; p[5] = -0.017645244547851414e+00; p[6] = 0.008023273027855346e+00; p[7] = -0.000804329819255744e+00; p[8] = -0.000360837876648255e+00; p[9] = 0.000145596568617526e+00; p[10] = -0.000017545539395205e+00; p[11] = -0.000002591225267689e+00; p[12] = 0.000001337767384067e+00; p[13] = -0.000000199542863674e+00; n = round(x - 2); w = x - (n + 2); y = ((((((((((((p[13] * w + p[12]) * w + p[11]) * w + p[10]) * w + p[9]) * w + p[8]) * w + p[7]) * w + p[6]) * w + p[5]) * w + p[4]) * w + p[3]) * w + p[2]) * w + p[1]) * w + p[0]; if (n > 0) { w = x - 1; for (k = 2; k <= n; ++k) { w = w * (x - k); } } else { w = 1; for (k = 0; k <= -n - 1; ++k) { y = y * (x + k); } } return w / y; } double area_of_cap(const double& s_cap) { // // AREA_OF_CAP Area of spherical cap // // AREA_OF_CAP(S_CAP) sets AREA to be the area of an S^2 spherical // cap of spherical radius S_CAP. // return 4.0 * M_PI * std::pow(std::sin(0.5 * s_cap), 2); } double area_of_collar(const double& a_top, const double& a_bot) { // AREA_OF_COLLAR Area of spherical collar // // AREA_OF_COLLAR(A_TOP, A_BOT) sets AREA to be the area of an S^2 spherical // collar specified by A_TOP, A_BOT, where A_TOP is top (smaller) spherical // radius, // A_BOT is bottom (larger) spherical radius. // return area_of_cap(a_bot) - area_of_cap(a_top); } double sradius_of_cap(const double& area) { // SRADIUS_OF_CAP(AREA) returns the spherical radius of // an S^2 spherical cap of area AREA. // return 2. * std::asin(0.5 * std::sqrt(area / M_PI)); } double area_of_ideal_region(int N) { // // AREA_OF_IDEAL_REGION(N) sets AREA to be the area of one of N equal // area regions on S^2, that is 1/N times AREA_OF_SPHERE. // double area_of_sphere = 2. * std::pow(M_PI, 1.5) / gamma(1.5); return area_of_sphere / static_cast(N); } double polar_colat(int N) { // // Given N, determine the colatitude of the North polar spherical cap. // double polar_c(0); if (N == 1) { polar_c = M_PI; } if (N == 2) { polar_c = 0.5 * M_PI; } if (N > 2) { polar_c = sradius_of_cap(area_of_ideal_region(N)); } return polar_c; } double ideal_collar_angle(int N) { // // IDEAL_COLLAR_ANGLE The ideal angle for spherical collars of an EQ partition // // IDEAL_COLLAR_ANGLE(N) sets ANGLE to the ideal angle for the // spherical collars of an EQ partition of the unit sphere S^2 into N regions. // return std::sqrt(area_of_ideal_region(N)); } void ideal_region_list(int N, const double& c_polar, int n_collars, double r_regions[]) { // // IDEAL_REGION_LIST The ideal real number of regions in each zone // // List the ideal real number of regions in each collar, plus the polar caps. // // Given N, c_polar and n_collars, determine r_regions, a list of the ideal // real // number of regions in each collar, plus the polar caps. // The number of elements is n_collars+2. // r_regions[1] is 1. // r_regions[n_collars+2] is 1. // The sum of r_regions is N. // // real(wp),intent(out) :: r_regions(n_collars+2) double ideal_region_area, ideal_collar_area; r_regions[0] = 1.; if (n_collars > 0) { // // Based on n_collars and c_polar, determine a_fitting, // the collar angle such that n_collars collars fit between the polar caps. // double a_fitting = (M_PI - 2. * c_polar) / static_cast(n_collars); ideal_region_area = area_of_ideal_region(N); for (int collar_n = 0; collar_n < n_collars; ++collar_n) { ideal_collar_area = area_of_collar(c_polar + collar_n * a_fitting, c_polar + (collar_n + 1) * a_fitting); r_regions[1 + collar_n] = ideal_collar_area / ideal_region_area; } } r_regions[2 + n_collars - 1] = 1.; } int num_collars(int N, const double& c_polar, const double& a_ideal) { // // NUM_COLLARS The number of collars between the polar caps // // Given N, an ideal angle, and c_polar, // determine n_collars, the number of collars between the polar caps. // bool enough = (N > 2) && (a_ideal > 0); if (enough) { return std::max(1, static_cast(round((M_PI - 2. * c_polar) / a_ideal))); } else { return 0; } } void round_to_naturals(int N, int ncollars, double r_regions[], int n_regions[]) { // ROUND_TO_NATURALS Round off a given list of numbers of regions // // Given N and r_regions, determine n_regions, // a list of the natural number of regions in each collar and the polar caps. // This list is as close as possible to r_regions, using rounding. // The number of elements is n_collars+2. // n_regions[1] is 1. // n_regions[n_collars+2] is 1. // The sum of n_regions is N. // double discrepancy = 0.; for (int zone_n = 0; zone_n < ncollars + 2; ++zone_n) { n_regions[zone_n] = round(r_regions[zone_n] + discrepancy); discrepancy += r_regions[zone_n] - n_regions[zone_n]; } } void cap_colats(int N, int n_collars, const double& c_polar, int n_regions[], double c_caps[]) { // CAP_COLATS Colatitudes of spherical caps enclosing cumulative sum of // regions // // Given dim, N, c_polar and n_regions, determine c_caps, // an increasing list of colatitudes of spherical caps which enclose the same // area // as that given by the cumulative sum of regions. // The number of elements is n_collars+2. // c_caps[1] is c_polar. // c_caps[n_collars+1] is Pi-c_polar. // c_caps[n_collars+2] is Pi. // // c_caps = cap_colats(dim,N,c_polar,n_regions); c_caps[0] = c_polar; double ideal_region_area = area_of_ideal_region(N); int subtotal_n_regions = 1; for (int collar_n = 0; collar_n < n_collars; ++collar_n) { subtotal_n_regions = subtotal_n_regions + n_regions[1 + collar_n]; c_caps[collar_n + 1] = sradius_of_cap(subtotal_n_regions * ideal_region_area); } c_caps[n_collars + 1] = M_PI; } // Disable optimisation for Cray (See ATLAS-327) #ifdef _CRAYC #pragma _CRI noopt #endif void eq_caps(int N, std::vector& n_regions, std::vector& s_cap) { // // eq_regions uses the zonal equal area sphere partitioning algorithm to // partition // the surface of a sphere into N regions of equal area and small diameter. // if (N == 1) { // // We have only one region, which must be the whole sphere. // n_regions.resize(1); s_cap.resize(1); n_regions[0] = 1; s_cap[0] = M_PI; // int n_regions_ns=1; } else { // // Given N, determine c_polar // the colatitude of the North polar spherical cap. // const double c_polar = polar_colat(N); // // Given N, determine the ideal angle for spherical collars. // Based on N, this ideal angle, and c_polar, // determine n_collars, the number of collars between the polar caps. // const int n_collars = num_collars(N, c_polar, ideal_collar_angle(N)); // int n_regions_ns=n_collars+2; // // Given N, c_polar and n_collars, determine r_regions, // a list of the ideal real number of regions in each collar, // plus the polar caps. // The number of elements is n_collars+2. // r_regions[0] is 1. // r_regions[2+n_collars-1] is 1. // The sum of r_regions is N. std::vector r_regions(n_collars + 2); ideal_region_list(N, c_polar, n_collars, r_regions.data()); // // Given N and r_regions, determine n_regions, a list of the natural number // of regions in each collar and the polar caps. // This list is as close as possible to r_regions. // The number of elements is n_collars+2. // n_regions[0] is 1. // n_regions[2+n_collars-1] is 1. // The sum of n_regions is N. // n_regions.resize(n_collars + 2); round_to_naturals(N, n_collars, r_regions.data(), n_regions.data()); // // Given dim, N, c_polar and n_regions, determine s_cap, // an increasing list of colatitudes of spherical caps which enclose the // same area // as that given by the cumulative sum of regions. // The number of elements is n_collars+2. // s_cap[0] is c_polar. // s_cap[n_collars] is Pi-c_polar. // s_cap[n_collars+1] is Pi. // s_cap.resize(n_collars + 2); cap_colats(N, n_collars, c_polar, n_regions.data(), s_cap.data()); } // int n_regions_ew=maxval(n_regions(:)); } // Reenable optimisation for Cray (See ATLAS-327) #ifdef _CRAYC #pragma _CRI opt #endif void eq_regions(int N, double xmin[], double xmax[], double ymin[], double ymax[]) { // EQ_REGIONS Recursive zonal equal area (EQ) partition of sphere // // Syntax // [regions,dim_1_rot] = eq_regions(dim,N,options); // // Description // REGIONS = EQ_REGIONS(dim,N) uses the recursive zonal equal area sphere // partitioning algorithm to partition S^dim (the unit sphere in dim+1 // dimensional space) into N regions of equal area and small diameter. // // The arguments dim and N must be positive integers. // // The result REGIONS is a (dim by 2 by N) array, representing the regions // of S^dim. Each element represents a pair of vertex points in spherical // polar // coordinates. // // Each region is defined as a product of intervals in spherical polar // coordinates. The pair of vertex points regions(:,1,n) and regions(:,2,n) // give // the lower and upper limits of each interval. // // REGIONS = EQ_REGIONS(dim,N,'offset','extra') uses experimental extra // offsets for S^2 and S^3 to try to minimize energy. If dim > 3, extra // offsets // are not used. // // REGIONS = EQ_REGIONS(dim,N,extra_offset) uses experimental extra offsets // if extra_offset is true or non-zero. // // [REGIONS,DIM_1_ROT] = EQ_REGIONS(dim,N) also returns DIM_1_ROT, a cell // array containing N rotation matrices, one per region, each of size dim by // dim. // These describe the R^dim rotation needed to place the region in its final // position. // // [REGIONS,DIM_1_ROT] = EQ_REGIONS(dim,N,'offset','extra') partitions S^dim // into N regions, using extra offsets, and also returning DIM_1_ROT, as // above. // if (N == 1) { // // We have only one region, which must be the whole sphere. // xmin[0] = 0.; ymin[0] = -0.5 * M_PI; xmax[0] = 2. * M_PI; ymax[0] = 0.5 * M_PI; return; } // // Start the partition of the sphere into N regions by partitioning // to caps defined in the current dimension. // std::vector n_regions; std::vector s_cap; eq_caps(N, n_regions, s_cap); // // s_cap is an increasing list of colatitudes of the caps. // // // We have a number of zones: two polar caps and a number of collars. // n_regions is the list of the number of regions in each zone. // int n_collars = n_regions.size() - 2; // // Start with the top cap. // xmin[0] = 0.; ymin[0] = 0.5 * M_PI - s_cap[0]; xmax[0] = 2. * M_PI; ymax[0] = 0.5 * M_PI; int region_n = 1; for (int collar_n = 0; collar_n < n_collars; ++collar_n) { for (int region_ew = 0; region_ew < n_regions[collar_n + 1]; ++region_ew) { xmin[region_n] = 2. * M_PI / (static_cast(n_regions[collar_n + 1])) * region_ew; ymin[region_n] = 0.5 * M_PI - s_cap[collar_n + 1]; xmax[region_n] = 2. * M_PI / (static_cast(n_regions[collar_n + 1])) * (region_ew + 1.); ymax[region_n] = 0.5 * M_PI - s_cap[collar_n]; ++region_n; } } // // End with the bottom cap. // xmin[N - 1] = 0.; ymin[N - 1] = -0.5 * M_PI; xmax[N - 1] = 2. * M_PI; ymax[N - 1] = 0.5 * M_PI - s_cap[s_cap.size() - 2]; } void EqualRegionsPartitioner::init() { N_ = nb_partitions(); std::vector s_cap; eq_caps(N_, sectors_, s_cap); bands_.resize(s_cap.size()); for (size_t n = 0; n < s_cap.size(); ++n) { bands_[n] = 0.5 * M_PI - s_cap[n]; } } EqualRegionsPartitioner::EqualRegionsPartitioner(): Partitioner() { init(); } EqualRegionsPartitioner::EqualRegionsPartitioner(int N): EqualRegionsPartitioner(N, util::NoConfig()) { } EqualRegionsPartitioner::EqualRegionsPartitioner(int N, const eckit::Parametrisation& config): Partitioner(N,config) { std::string crds; if (config.get("coordinates", crds)) { if (crds == "lonlat") { coordinates_ = Coordinates::LONLAT; } } init(); } EqualRegionsPartitioner::EqualRegionsPartitioner(const eckit::Parametrisation& config): Partitioner(config) { std::string crds; if (config.get("coordinates", crds)) { if (crds == "lonlat") { coordinates_ = Coordinates::LONLAT; } } init(); } int EqualRegionsPartitioner::partition(const double& x, const double& y) const { int b = band(y); int p = 0; for (int n = 0; n < b; ++n) { p += sectors_[n]; } return p + sector(b, x); } int EqualRegionsPartitioner::band(const double& y) const { return std::distance(bands_.begin(), std::lower_bound(bands_.begin(), bands_.end(), y, std::greater())); } int EqualRegionsPartitioner::sector(int band, const double& x) const { constexpr double two_pi = 2. * M_PI; constexpr double div_two_pi = 1. / (2. * M_PI + 1.e-8); double xreg = x; if (x < 0.) { xreg += two_pi; } else if (x > two_pi) { xreg -= two_pi; } return std::floor(xreg * sectors_[band] * div_two_pi); } void EqualRegionsPartitioner::where(int partition, int& band, int& sector) const { int p = 0; for (size_t b = 0; b < bands_.size(); ++b) { for (int s = 0; s < sectors_[b]; ++s) { if (partition == p) { band = b; sector = s; return; } ++p; } } } bool compare_NS_WE(const EqualRegionsPartitioner::NodeInt& node1, const EqualRegionsPartitioner::NodeInt& node2) { if (node1.y > node2.y) { return true; } if (node1.y == node2.y) { return (node1.x < node2.x); } return false; } bool compare_WE_NS(const EqualRegionsPartitioner::NodeInt& node1, const EqualRegionsPartitioner::NodeInt& node2) { if (node1.x < node2.x) { return true; } if (node1.x == node2.x) { return (node1.y > node2.y); } return false; } void EqualRegionsPartitioner::partition(int nb_nodes, NodeInt nodes[], int part[]) const { ATLAS_TRACE("EqualRegionsPartitioner::partition"); // std::clock_t init, final; // init=std::clock(); // std::cout << "partition start (" << nb_nodes << " points)" << std::endl; int nb_parts = N_; int p; int i; int begin; int end; int band; /* Sort nodes from north to south, and west to east. Now we can easily split the points in bands. Note, for RGG, this should not be necessary, as it is already by construction in this order, but then sorting is really fast */ // std::sort( nodes, nodes+nb_nodes, compare_NS_WE); /* For every band, now sort from west to east, and north to south. Inside every band we can now easily split nodes in sectors. */ int chunk_size = nb_nodes / nb_parts; int chunk_remainder = nb_nodes - chunk_size * nb_parts; int remainder = chunk_remainder; std::vector count; count.reserve(nb_parts); end = 0; for (band = 0; band < nb_bands(); ++band) { begin = end; for (p = 0; p < nb_regions(band); ++p) { count.push_back(chunk_size + (remainder-- > 0 ? 1 : 0)); end += count.back(); } std::sort(nodes + begin, nodes + end, compare_WE_NS); } /* Create list that tells in original node numbering which part the node belongs to */ end = 0; for (p = 0; p < nb_parts; ++p) { begin = end; end = begin + count[p]; for (i = begin; i < end; ++i) { part[nodes[i].n] = p; } } /* This piece of code is to reorder back along latitudes */ // begin = 0; // for( p=0; p(comm.rank()); int mpi_size = static_cast(comm.size()); atlas::vector nodes(grid.size()); int* nodes_buffer = reinterpret_cast(nodes.data()); long nb_workers = comm.size(); /* Sort nodes from north to south, and west to east. Now we can easily split the points in bands. Note, for StructuredGrid, this should not be necessary, as it is already by construction in this order, but then sorting is really fast */ if (StructuredGrid(grid) && (coordinates_ == Coordinates::XY)) { // The grid comes sorted from north to south and west to east by // construction // ATLAS_ASSERT to make sure. StructuredGrid structured_grid(grid); //ATLAS_ASSERT( structured_grid.y( 1 ) < structured_grid.y( 0 ) ); ATLAS_ASSERT(structured_grid.x(1, 0) > structured_grid.x(0, 0)); ATLAS_TRACE("Take shortcut"); if (atlas_omp_get_max_threads() > 1) { atlas_omp_parallel { const idx_t num_threads = atlas_omp_get_num_threads(); const idx_t thread_num = atlas_omp_get_thread_num(); const idx_t begin = static_cast(thread_num * size_t(structured_grid.size()) / num_threads); const idx_t end = static_cast((thread_num + 1) * size_t(structured_grid.size()) / num_threads); idx_t thread_j_begin = 0; std::vector thread_i_begin(structured_grid.ny()); std::vector thread_i_end(structured_grid.ny()); idx_t n = 0; for (idx_t j = 0; j < structured_grid.ny(); ++j) { if (n + structured_grid.nx(j) > begin) { thread_j_begin = j; thread_i_begin[j] = begin - n; break; } n += structured_grid.nx(j); } idx_t thread_j_end{thread_j_begin}; for (idx_t j = thread_j_begin; j < structured_grid.ny(); ++j) { idx_t i_end = end - n; if (j > thread_j_begin) { thread_i_begin[j] = 0; } if (i_end > structured_grid.nx(j)) { thread_i_end[j] = structured_grid.nx(j); n += structured_grid.nx(j); } else { thread_i_end[j] = i_end; thread_j_end = j + 1; break; } } int nn = begin; for (idx_t j = thread_j_begin; j < thread_j_end; ++j) { int y = microdeg(structured_grid.y(j)); for (idx_t i = thread_i_begin[j]; i < thread_i_end[j]; ++i, ++nn) { nodes[nn].x = microdeg(structured_grid.x(i, j)); nodes[nn].y = y; nodes[nn].n = nn; } } ATLAS_ASSERT(nn == end); } } else { int n(0); for (idx_t j = 0; j < structured_grid.ny(); ++j) { int y = microdeg(structured_grid.y(j)); for (idx_t i = 0; i < structured_grid.nx(j); ++i, ++n) { nodes[n].x = microdeg(structured_grid.x(i, j)); nodes[n].y = y; nodes[n].n = n; } } } } else { ATLAS_TRACE("sort all"); std::vector requests; for (int w = 0; w < nb_workers; ++w) { idx_t w_begin = w * grid.size() / N_; idx_t w_end = (w + 1) * grid.size() / N_; if (w == nb_workers - 1) { w_end = grid.size(); } idx_t w_size = w_end - w_begin; int work_rank = std::min(w, mpi_size - 1); if (mpi_rank == 0) { ATLAS_ASSERT(valid_mpi_size(w_size * 3)); requests.push_back(comm.iReceive(nodes_buffer + w_begin * 3, w_size * 3, /* source= */ work_rank, /* tag= */ 0)); } if (mpi_rank == work_rank) { atlas::vector w_nodes(w_size); int* w_nodes_buffer = reinterpret_cast(w_nodes.data()); if (coordinates_ == Coordinates::XY) { ATLAS_TRACE_SCOPE("create one bit - Coordinates::XY") { if (true) // optimized experimental when true (still need to // benchmark) { int i = w_begin; int j(0); for (const auto& point : subrange(grid.xy(), {w_begin, w_end})) { w_nodes[j].x = microdeg(point[0]); w_nodes[j].y = microdeg(point[1]); w_nodes[j].n = i++; ++j; } } else { idx_t i(0); idx_t j(0); for (const auto& point : grid.xy()) { if (i >= w_begin && i < w_end) { w_nodes[j].x = microdeg(point[0]); w_nodes[j].y = microdeg(point[1]); w_nodes[j].n = i; ++j; } ++i; } } } } else if (coordinates_ == Coordinates::LONLAT) { ATLAS_TRACE_SCOPE("create one bit - Coordinates::LONLAT") { if (true) // optimized experimental when true (still need to // benchmark) { int i = w_begin; int j(0); for (const auto& point : subrange(grid.lonlat(), {w_begin, w_end})) { w_nodes[j].x = microdeg(point[0]); w_nodes[j].y = microdeg(point[1]); w_nodes[j].n = i++; ++j; } } else { idx_t i(0); idx_t j(0); for (const auto& point : grid.xy()) { if (i >= w_begin && i < w_end) { w_nodes[j].x = microdeg(point[0]); w_nodes[j].y = microdeg(point[1]); w_nodes[j].n = i; ++j; } ++i; } } } } else { ATLAS_THROW_EXCEPTION("Should not be here"); } ATLAS_TRACE_SCOPE("sort one bit") { omp::sort(w_nodes.begin(), w_nodes.end(), compare_NS_WE); } ATLAS_TRACE_SCOPE("send to rank0") { ATLAS_ASSERT(valid_mpi_size(w_size * 3)); comm.send(w_nodes_buffer, w_size * 3, /* dest= */ 0, /* tag= */ 0); } } } ATLAS_TRACE_MPI(WAIT) { for (auto request : requests) { comm.wait(request); } } ATLAS_TRACE_SCOPE("merge sorted") { for (int w = 0; w < nb_workers; ++w) { int w_begin = w * grid.size() / N_; int w_end = (w + 1) * grid.size() / N_; if (w == nb_workers - 1) { w_end = grid.size(); } if (w != 0) { std::inplace_merge(nodes.begin(), nodes.begin() + w_begin, nodes.begin() + w_end, compare_NS_WE); } } } ATLAS_TRACE_MPI(BROADCAST) { ATLAS_ASSERT(valid_mpi_size(grid.size() * 3)); comm.broadcast(nodes_buffer, grid.size() * 3, /* root= */ 0); } } // sort all /* For every band, now sort from west to east, and north to south. Inside every band we can now easily split nodes in sectors. */ ATLAS_TRACE_SCOPE("sort bands") { ATLAS_TRACE_BARRIERS(false); // avoid deadlock std::vector requests; int nb_parts = N_; size_t nb_nodes = grid.size(); size_t chunk_size = nb_nodes / nb_parts; size_t chunk_remainder = nb_nodes - chunk_size * nb_parts; int remainder = chunk_remainder; std::vector count; count.reserve(nb_parts); std::vector displs; displs.reserve(nb_parts + 1); std::vector b_count; b_count.reserve(nb_bands()); std::vector b_displs; b_displs.reserve(nb_bands() + 1); { size_t end = 0; for (int band = 0; band < nb_bands(); ++band) { b_displs.emplace_back(end); size_t b_size(0); for (int p = 0; p < nb_regions(band); ++p) { size_t w_size = chunk_size + (remainder-- > 0 ? size_t(1) : size_t(0)); displs.emplace_back(end); count.emplace_back(w_size); end += w_size; b_size += w_size; } b_count.emplace_back(b_size); } displs.emplace_back(end); b_displs.emplace_back(end); } int w(0); for (int band = 0; band < nb_bands(); ++band) { int w0 = w; int w0_work_rank = std::min(w0, mpi_size - 1); for (int p = 0; p < nb_regions(band); ++p) { size_t w_begin = displs[w]; size_t w_size = count[w]; size_t w_end = w_begin + w_size; int work_rank = std::min(w, mpi_size - 1); if (mpi_rank == w0_work_rank) { if (mpi_rank != work_rank) { ATLAS_ASSERT(valid_mpi_size(w_size * 3)); requests.push_back(comm.iReceive(nodes_buffer + w_begin * 3, w_size * 3, /* source= */ work_rank, /* tag= */ 0)); } } if (work_rank == mpi_rank) { ATLAS_TRACE_SCOPE("sort bit of band on each MPI rank") { omp::sort(nodes.data() + w_begin, nodes.data() + w_end, compare_WE_NS); } if (mpi_rank != w0_work_rank) { ATLAS_TRACE_SCOPE("send bit of band to band leader " + std::to_string(w0_work_rank)) { ATLAS_ASSERT(valid_mpi_size(w_size * 3)); comm.send(nodes_buffer + w_begin * 3, w_size * 3, /* dest= */ w0_work_rank, /* tag= */ 0); } } } ++w; } } ATLAS_TRACE_MPI(WAIT) { for (auto request : requests) { comm.wait(request); } } requests.clear(); std::vector w0_band(nb_bands()); for (int band = 0, w = 0; band < nb_bands(); ++band) { w0_band[band] = w; w += nb_regions(band); } for (int band = 0; band < nb_bands(); ++band) { int w0 = w0_band[band]; int w0_work_rank = std::min(w0, mpi_size - 1); if (mpi_rank == w0_work_rank) { auto nodes_band_begin = nodes.begin() + b_displs[band]; auto nodes_band_end = nodes_band_begin + b_count[band]; auto blocks_size_begin = count.begin() + w0; auto blocks_size_end = blocks_size_begin + nb_regions(band); ATLAS_TRACE_SCOPE("band leader merging bits for band " + std::to_string(band)) { omp::merge_blocks(nodes_band_begin, nodes_band_end, blocks_size_begin, blocks_size_end, compare_WE_NS); } } } for (int band = 0; band < nb_bands(); ++band) { int w0 = w0_band[band]; size_t w0_begin = b_displs[band]; size_t w0_size = b_count[band]; int w0_work_rank = std::min(w0, mpi_size - 1); if (mpi_rank == 0) { if (mpi_rank != w0_work_rank) { ATLAS_ASSERT(valid_mpi_size(w0_size * 3)); requests.push_back(comm.iReceive(nodes_buffer + w0_begin * 3, w0_size * 3, /* source= */ w0_work_rank, /* tag= */ 0)); } } if (mpi_rank == w0_work_rank) { ATLAS_TRACE_SCOPE("send band to rank 0") { if (mpi_rank != 0) { ATLAS_ASSERT(valid_mpi_size(w0_size * 3)); comm.send(nodes_buffer + w0_begin * 3, w0_size * 3, /* dest= */ 0, /* tag= */ 0); } } } } if (mpi_rank == 0) { ATLAS_TRACE_SCOPE("rank 0 waiting for all bands to come in") { ATLAS_TRACE_MPI(WAIT) { for (auto request : requests) { comm.wait(request); } } } } /* Create list that tells in original node numbering which part the node belongs to */ for (int p = 0; p < nb_parts; ++p) { size_t begin = displs[p]; size_t end = begin + count[p]; atlas_omp_parallel_for(size_t i = begin; i < end; ++i) { part[nodes[i].n] = p; } } ATLAS_TRACE_MPI(BROADCAST) { ATLAS_ASSERT(valid_mpi_size(nb_nodes)); comm.broadcast(part, nb_nodes, 0); } } // sort bands } // else } } // namespace partitioner } // namespace detail } // namespace grid } // namespace atlas namespace { atlas::grid::detail::partitioner::PartitionerBuilder __EqualRegions("equal_regions"); }