| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "leonova_a_most_diff_neigh_vec_elems/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <array> | ||
| 7 | #include <cmath> | ||
| 8 | #include <cstdint> | ||
| 9 | #include <cstdlib> | ||
| 10 | #include <limits> | ||
| 11 | #include <tuple> | ||
| 12 | #include <vector> | ||
| 13 | |||
| 14 | #include "leonova_a_most_diff_neigh_vec_elems/common/include/common.hpp" | ||
| 15 | |||
| 16 | namespace leonova_a_most_diff_neigh_vec_elems { | ||
| 17 | |||
| 18 |
1/2✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
|
14 | LeonovaAMostDiffNeighVecElemsMPI::LeonovaAMostDiffNeighVecElemsMPI(const InType &in) { |
| 19 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 20 |
1/2✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
|
14 | GetInput() = in; |
| 21 | GetOutput() = std::tuple<int, int>(0, 0); | ||
| 22 | 14 | } | |
| 23 | |||
| 24 | 28 | bool LeonovaAMostDiffNeighVecElemsMPI::ValidationImpl() { | |
| 25 | 28 | int rank = 0; | |
| 26 | 28 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 27 | |||
| 28 | 28 | bool is_valid = false; | |
| 29 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 14 times.
|
28 | if (rank == 0) { |
| 30 | 14 | is_valid = !GetInput().empty(); | |
| 31 | } | ||
| 32 | |||
| 33 | 28 | MPI_Bcast(&is_valid, 1, MPI_C_BOOL, 0, MPI_COMM_WORLD); | |
| 34 | 28 | return is_valid; | |
| 35 | } | ||
| 36 | |||
| 37 | 14 | bool LeonovaAMostDiffNeighVecElemsMPI::PreProcessingImpl() { | |
| 38 | 14 | return true; | |
| 39 | } | ||
| 40 | |||
| 41 | 14 | bool LeonovaAMostDiffNeighVecElemsMPI::RunImpl() { | |
| 42 | const auto &input_vec = GetInput(); | ||
| 43 | |||
| 44 |
1/2✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
|
14 | if (!ValidationImpl()) { |
| 45 | return false; | ||
| 46 | } | ||
| 47 | |||
| 48 | 14 | int rank = 0; | |
| 49 | 14 | int size = 0; | |
| 50 | 14 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 51 | 14 | MPI_Comm_size(MPI_COMM_WORLD, &size); | |
| 52 | |||
| 53 | 14 | int total_size = static_cast<int>(input_vec.size()); | |
| 54 | 14 | ProcessWithMultipleProcesses(rank, size, total_size, input_vec); | |
| 55 | return true; | ||
| 56 | } | ||
| 57 | |||
| 58 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 1 times.
|
14 | void LeonovaAMostDiffNeighVecElemsMPI::ProcessWithMultipleProcesses(int rank, int size, int total_size, |
| 59 | const std::vector<int> &input_vec) { | ||
| 60 | int actual_processes = std::min(size, total_size); | ||
| 61 | 14 | int local_max_diff = -1; | |
| 62 | 14 | int local_first = 0; | |
| 63 | 14 | int local_second = 0; | |
| 64 | |||
| 65 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 1 times.
|
14 | if (rank < actual_processes) { |
| 66 | 13 | ProcessLocalData(rank, actual_processes, total_size, input_vec, local_max_diff, local_first, local_second, size); | |
| 67 | } | ||
| 68 | |||
| 69 | 14 | GatherAndProcessResults(rank, actual_processes, local_max_diff, local_first, local_second, size); | |
| 70 | 14 | } | |
| 71 | |||
| 72 | 13 | void LeonovaAMostDiffNeighVecElemsMPI::ProcessLocalData(int rank, int actual_processes, int total_size, | |
| 73 | const std::vector<int> &input_vec, int &local_max_diff, | ||
| 74 | int &local_first, int &local_second, int size) { | ||
| 75 | 13 | int chunk_size = total_size / actual_processes; | |
| 76 | 13 | int remainder = total_size % actual_processes; | |
| 77 | |||
| 78 | 13 | std::vector<int> sizes(actual_processes); | |
| 79 |
1/4✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
13 | std::vector<int> offsets(actual_processes); |
| 80 | |||
| 81 | int offset = 0; | ||
| 82 |
2/2✓ Branch 0 taken 25 times.
✓ Branch 1 taken 13 times.
|
38 | for (int i = 0; i < actual_processes; ++i) { |
| 83 |
2/2✓ Branch 0 taken 21 times.
✓ Branch 1 taken 4 times.
|
46 | sizes[i] = chunk_size + (i < remainder ? 1 : 0) + 1; |
| 84 | 25 | offsets[i] = offset; | |
| 85 | 25 | offset += sizes[i] - 1; | |
| 86 | } | ||
| 87 | |||
| 88 |
1/2✓ Branch 0 taken 13 times.
✗ Branch 1 not taken.
|
13 | if (actual_processes > 0) { |
| 89 | 13 | sizes[actual_processes - 1] = total_size - offsets[actual_processes - 1]; | |
| 90 | } | ||
| 91 | |||
| 92 |
2/4✓ Branch 0 taken 13 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 13 times.
✗ Branch 3 not taken.
|
13 | int my_size = (rank < actual_processes) ? sizes[rank] : 0; |
| 93 | |||
| 94 |
1/2✓ Branch 0 taken 13 times.
✗ Branch 1 not taken.
|
13 | if (my_size > 0) { |
| 95 |
1/4✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
13 | std::vector<int> local_data(my_size); |
| 96 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | ReceiveLocalData(rank, actual_processes, input_vec, sizes, offsets, local_data, size); |
| 97 | 13 | FindLocalMaxDiff(local_data, local_max_diff, local_first, local_second); | |
| 98 | } | ||
| 99 | 13 | } | |
| 100 | |||
| 101 | 13 | void LeonovaAMostDiffNeighVecElemsMPI::ReceiveLocalData(int rank, int actual_processes, | |
| 102 | const std::vector<int> &input_vec, | ||
| 103 | const std::vector<int> &sizes, const std::vector<int> &offsets, | ||
| 104 | std::vector<int> &local_data, int size) { | ||
| 105 | 13 | std::vector<int> send_counts(size, 0); | |
| 106 |
1/4✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
13 | std::vector<int> displacements(size, 0); |
| 107 |
1/4✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
13 | std::vector<int> recv_counts(size, 0); |
| 108 | |||
| 109 |
2/2✓ Branch 0 taken 25 times.
✓ Branch 1 taken 13 times.
|
38 | for (int i = 0; i < actual_processes; ++i) { |
| 110 | 25 | send_counts[i] = sizes[i]; | |
| 111 | 25 | displacements[i] = offsets[i]; | |
| 112 | 25 | recv_counts[i] = sizes[i]; | |
| 113 | } | ||
| 114 | |||
| 115 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 6 times.
|
13 | if (rank == 0) { |
| 116 |
1/2✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
|
7 | if (!local_data.empty()) { |
| 117 | 7 | std::copy(input_vec.begin() + offsets[0], input_vec.begin() + offsets[0] + sizes[0], local_data.begin()); | |
| 118 | } | ||
| 119 | } | ||
| 120 | |||
| 121 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 6 times.
|
13 | int recv_count = static_cast<int>(local_data.size()); |
| 122 |
3/4✓ Branch 0 taken 7 times.
✓ Branch 1 taken 6 times.
✓ Branch 3 taken 13 times.
✗ Branch 4 not taken.
|
20 | MPI_Scatterv((rank == 0) ? input_vec.data() : nullptr, send_counts.data(), displacements.data(), MPI_INT, |
| 123 | local_data.data(), recv_count, MPI_INT, 0, MPI_COMM_WORLD); | ||
| 124 | 13 | } | |
| 125 | |||
| 126 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 11 times.
|
13 | void LeonovaAMostDiffNeighVecElemsMPI::FindLocalMaxDiff(const std::vector<int> &local_data, int &local_max_diff, |
| 127 | int &local_first, int &local_second) { | ||
| 128 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 11 times.
|
13 | if (local_data.size() == 1) { |
| 129 | 2 | local_max_diff = 0; | |
| 130 | 2 | local_first = local_data[0]; | |
| 131 | 2 | local_second = local_data[0]; | |
| 132 | 2 | return; | |
| 133 | } | ||
| 134 | |||
| 135 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 11 times.
|
31 | for (int index = 0; index < static_cast<int>(local_data.size()) - 1; ++index) { |
| 136 |
2/2✓ Branch 0 taken 12 times.
✓ Branch 1 taken 8 times.
|
20 | int64_t diff = static_cast<int64_t>(local_data[index]) - static_cast<int64_t>(local_data[index + 1]); |
| 137 | 20 | int64_t curr_diff_ll = std::llabs(diff); // safe abs | |
| 138 | |||
| 139 | int curr_diff = (curr_diff_ll <= static_cast<int64_t>(std::numeric_limits<int>::max())) | ||
| 140 | 20 | ? static_cast<int>(curr_diff_ll) | |
| 141 | : std::numeric_limits<int>::max(); | ||
| 142 | |||
| 143 |
2/2✓ Branch 0 taken 12 times.
✓ Branch 1 taken 8 times.
|
20 | if (curr_diff > local_max_diff) { |
| 144 | 12 | local_max_diff = curr_diff; | |
| 145 | 12 | local_first = local_data[index]; | |
| 146 | 12 | local_second = local_data[index + 1]; | |
| 147 | } | ||
| 148 | } | ||
| 149 | } | ||
| 150 | |||
| 151 | 14 | void LeonovaAMostDiffNeighVecElemsMPI::GatherAndProcessResults(int rank, int actual_processes, int local_max_diff, | |
| 152 | int local_first, int local_second, int size) { | ||
| 153 | struct ProcessResult { | ||
| 154 | int diff; | ||
| 155 | int first; | ||
| 156 | int second; | ||
| 157 | }; | ||
| 158 | |||
| 159 | 14 | ProcessResult local_result{.diff = local_max_diff, .first = local_first, .second = local_second}; | |
| 160 |
1/2✓ Branch 2 taken 14 times.
✗ Branch 3 not taken.
|
14 | std::vector<ProcessResult> all_results(size); |
| 161 | |||
| 162 |
1/2✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
|
14 | MPI_Gather(&local_result, 3, MPI_INT, all_results.data(), 3, MPI_INT, 0, MPI_COMM_WORLD); |
| 163 | |||
| 164 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
|
14 | if (rank == 0) { |
| 165 | int global_max_diff = -1; | ||
| 166 | int best_index = -1; | ||
| 167 | |||
| 168 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 7 times.
|
20 | for (int index = 0; index < actual_processes; ++index) { |
| 169 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 6 times.
|
13 | if (all_results[index].diff > global_max_diff) { |
| 170 | global_max_diff = all_results[index].diff; | ||
| 171 | best_index = index; | ||
| 172 | } | ||
| 173 | } | ||
| 174 | |||
| 175 |
1/2✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
|
7 | if (best_index != -1) { |
| 176 | 7 | GetOutput() = std::make_tuple(all_results[best_index].first, all_results[best_index].second); | |
| 177 | } else { | ||
| 178 | GetOutput() = std::make_tuple(0, 0); | ||
| 179 | } | ||
| 180 | } | ||
| 181 | |||
| 182 |
1/2✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
|
14 | BroadcastResult(rank); |
| 183 | 14 | } | |
| 184 | |||
| 185 | 14 | void LeonovaAMostDiffNeighVecElemsMPI::BroadcastResult(int rank) { | |
| 186 | 14 | std::array<int, 2> result_data{0, 0}; | |
| 187 | |||
| 188 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
|
14 | if (rank == 0) { |
| 189 | 7 | result_data[0] = std::get<0>(GetOutput()); | |
| 190 | 7 | result_data[1] = std::get<1>(GetOutput()); | |
| 191 | } | ||
| 192 | |||
| 193 | 14 | MPI_Bcast(result_data.data(), 2, MPI_INT, 0, MPI_COMM_WORLD); | |
| 194 | |||
| 195 |
2/2✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
|
14 | if (rank != 0) { |
| 196 | GetOutput() = std::make_tuple(result_data[0], result_data[1]); | ||
| 197 | } | ||
| 198 | 14 | } | |
| 199 | |||
| 200 | 14 | bool LeonovaAMostDiffNeighVecElemsMPI::PostProcessingImpl() { | |
| 201 | 14 | return true; | |
| 202 | } | ||
| 203 | |||
| 204 | } // namespace leonova_a_most_diff_neigh_vec_elems | ||
| 205 |