| 1 |
/* |
| 2 |
* (C) Copyright 2013 ECMWF. |
| 3 |
* |
| 4 |
* This software is licensed under the terms of the Apache Licence Version 2.0 |
| 5 |
* which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. |
| 6 |
* In applying this licence, ECMWF does not waive the privileges and immunities |
| 7 |
* granted to it by virtue of its status as an intergovernmental organisation |
| 8 |
* nor does it submit to any jurisdiction. |
| 9 |
*/ |
| 10 |
|
| 11 |
#include "DistributionArray.h" |
| 12 |
|
| 13 |
#include <algorithm> |
| 14 |
#include <numeric> |
| 15 |
#include <ostream> |
| 16 |
#include <vector> |
| 17 |
|
| 18 |
#include "eckit/types/Types.h" |
| 19 |
#include "eckit/utils/Hash.h" |
| 20 |
|
| 21 |
#include "atlas/grid/Grid.h" |
| 22 |
#include "atlas/grid/Partitioner.h" |
| 23 |
#include "atlas/parallel/omp/omp.h" |
| 24 |
#include "atlas/util/Config.h" |
| 25 |
|
| 26 |
namespace atlas { |
| 27 |
namespace grid { |
| 28 |
namespace detail { |
| 29 |
namespace distribution { |
| 30 |
|
| 31 |
namespace { |
| 32 |
std::string distribution_type(int N, const Partitioner& p = Partitioner()) { |
| 33 |
if (N == 1) { |
| 34 |
return "serial"; |
| 35 |
} |
| 36 |
if (not p) { |
| 37 |
return "custom"; |
| 38 |
} |
| 39 |
return p.type(); |
| 40 |
} |
| 41 |
} // namespace |
| 42 |
|
| 43 |
DistributionArray::DistributionArray(const Grid& grid, const eckit::Parametrisation& config): |
| 44 |
DistributionArray(grid, Partitioner{config}) {} |
| 45 |
|
| 46 |
|
| 47 |
DistributionArray::DistributionArray(const Grid& grid, const Partitioner& partitioner) { |
| 48 |
part_.resize(grid.size()); |
| 49 |
partitioner.partition(grid, part_.data()); |
| 50 |
nb_partitions_ = partitioner.nb_partitions(); |
| 51 |
|
| 52 |
size_t size = part_.size(); |
| 53 |
nb_pts_.resize(nb_partitions_, 0); |
| 54 |
atlas_omp_parallel { |
| 55 |
std::vector<int> nb_pts(nb_partitions_, 0); // thread-local |
| 56 |
atlas_omp_for(size_t j = 0; j < size; ++j) { |
| 57 |
++nb_pts[part_[j]]; |
| 58 |
} |
| 59 |
atlas_omp_critical { |
| 60 |
for (int p = 0; p < nb_partitions_; ++p) { |
| 61 |
nb_pts_[p] += nb_pts[p]; |
| 62 |
} |
| 63 |
} |
| 64 |
} |
| 65 |
|
| 66 |
ATLAS_ASSERT( std::accumulate(nb_pts_.begin(),nb_pts_.end(),idx_t{0}) == part_.size() ); |
| 67 |
|
| 68 |
max_pts_ = *std::max_element(nb_pts_.begin(), nb_pts_.end()); |
| 69 |
min_pts_ = *std::min_element(nb_pts_.begin(), nb_pts_.end()); |
| 70 |
type_ = distribution_type(nb_partitions_, partitioner); |
| 71 |
} |
| 72 |
|
| 73 |
DistributionArray::DistributionArray(int nb_partitions, idx_t npts, int part[], int part0) { |
| 74 |
part_.assign(part, part + npts); |
| 75 |
if (nb_partitions == 0) { |
| 76 |
std::set<int> partset(part_.begin(), part_.end()); |
| 77 |
nb_partitions_ = static_cast<idx_t>(partset.size()); |
| 78 |
} |
| 79 |
else { |
| 80 |
nb_partitions_ = nb_partitions; |
| 81 |
} |
| 82 |
nb_pts_.resize(nb_partitions_, 0); |
| 83 |
for (idx_t j = 0, size = static_cast<idx_t>(part_.size()); j < size; ++j) { |
| 84 |
part_[j] -= part0; |
| 85 |
++nb_pts_[part_[j]]; |
| 86 |
} |
| 87 |
max_pts_ = *std::max_element(nb_pts_.begin(), nb_pts_.end()); |
| 88 |
min_pts_ = *std::min_element(nb_pts_.begin(), nb_pts_.end()); |
| 89 |
type_ = distribution_type(nb_partitions_); |
| 90 |
} |
| 91 |
|
| 92 |
DistributionArray::DistributionArray(int nb_partitions, partition_t&& part): |
| 93 |
nb_partitions_(nb_partitions), part_(std::move(part)), nb_pts_(nb_partitions_, 0) { |
| 94 |
size_t size = part_.size(); |
| 95 |
int num_threads = atlas_omp_get_max_threads(); |
| 96 |
std::vector<std::vector<int> > nb_pts_per_thread(num_threads, std::vector<int>(nb_partitions_)); |
| 97 |
atlas_omp_parallel { |
| 98 |
int thread = atlas_omp_get_thread_num(); |
| 99 |
auto& nb_pts = nb_pts_per_thread[thread]; |
| 100 |
atlas_omp_for(size_t j = 0; j < size; ++j) { |
| 101 |
int p = part_[j]; |
| 102 |
++nb_pts[p]; |
| 103 |
} |
| 104 |
} |
| 105 |
for (int thread = 0; thread < num_threads; ++thread) { |
| 106 |
for (int p = 0; p < nb_partitions_; ++p) { |
| 107 |
nb_pts_[p] += nb_pts_per_thread[thread][p]; |
| 108 |
} |
| 109 |
} |
| 110 |
|
| 111 |
max_pts_ = *std::max_element(nb_pts_.begin(), nb_pts_.end()); |
| 112 |
min_pts_ = *std::min_element(nb_pts_.begin(), nb_pts_.end()); |
| 113 |
type_ = distribution_type(nb_partitions_); |
| 114 |
} |
| 115 |
|
| 116 |
DistributionArray::~DistributionArray() = default; |
| 117 |
|
| 118 |
void DistributionArray::print(std::ostream& s) const { |
| 119 |
auto print_partition = [&](std::ostream& s) { |
| 120 |
eckit::output_list<int> list_printer(s); |
| 121 |
for (size_t i = 0, size = part_.size(); i < size; i++) { |
| 122 |
list_printer.push_back(part_[i]); |
| 123 |
} |
| 124 |
}; |
| 125 |
s << "Distribution( " |
| 126 |
<< "type: " << type_ << ", nb_points: " << size() << ", nb_partitions: " << nb_pts_.size() << ", parts : "; |
| 127 |
print_partition(s); |
| 128 |
} |
| 129 |
|
| 130 |
void DistributionArray::hash(eckit::Hash& hash) const { |
| 131 |
for (size_t i = 0; i < part_.size(); i++) { |
| 132 |
hash.add(part_[i]); |
| 133 |
} |
| 134 |
} |
| 135 |
|
| 136 |
|
| 137 |
} // namespace distribution |
| 138 |
} // namespace detail |
| 139 |
} // namespace grid |
| 140 |
} // namespace atlas |