| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "liulin_y_matrix_max_column/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <limits> | ||
| 8 | #include <vector> | ||
| 9 | |||
| 10 | #include "liulin_y_matrix_max_column/common/include/common.hpp" | ||
| 11 | |||
| 12 | namespace liulin_y_matrix_max_column { | ||
| 13 | |||
| 14 |
1/2✓ Branch 0 taken 49 times.
✗ Branch 1 not taken.
|
49 | int LiulinYMatrixMaxColumnMPI::TournamentMax(const std::vector<int> &column) { |
| 15 |
1/2✓ Branch 0 taken 49 times.
✗ Branch 1 not taken.
|
49 | if (column.empty()) { |
| 16 | return std::numeric_limits<int>::min(); | ||
| 17 | } | ||
| 18 | |||
| 19 | 49 | int size = static_cast<int>(column.size()); | |
| 20 | 49 | std::vector<int> temp = column; | |
| 21 | |||
| 22 |
2/2✓ Branch 0 taken 153 times.
✓ Branch 1 taken 49 times.
|
202 | while (size > 1) { |
| 23 | int new_size = 0; | ||
| 24 |
2/2✓ Branch 0 taken 594 times.
✓ Branch 1 taken 153 times.
|
747 | for (int i = 0; i < size; i += 2) { |
| 25 |
4/4✓ Branch 0 taken 522 times.
✓ Branch 1 taken 72 times.
✓ Branch 2 taken 250 times.
✓ Branch 3 taken 272 times.
|
844 | temp[new_size] = (i + 1 < size) ? std::max(temp[i], temp[i + 1]) : temp[i]; |
| 26 | 594 | ++new_size; | |
| 27 | } | ||
| 28 | size = new_size; | ||
| 29 | } | ||
| 30 | 49 | return temp[0]; | |
| 31 | } | ||
| 32 | |||
| 33 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | LiulinYMatrixMaxColumnMPI::LiulinYMatrixMaxColumnMPI(const InType &in) { |
| 34 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 35 | |||
| 36 | GetInput().clear(); | ||
| 37 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | GetInput().reserve(in.size()); |
| 38 |
2/2✓ Branch 0 taken 142 times.
✓ Branch 1 taken 24 times.
|
166 | for (const auto &row : in) { |
| 39 |
1/2✓ Branch 1 taken 142 times.
✗ Branch 2 not taken.
|
142 | GetInput().push_back(row); |
| 40 | } | ||
| 41 | |||
| 42 | GetOutput().clear(); | ||
| 43 | 24 | } | |
| 44 | |||
| 45 | 24 | bool LiulinYMatrixMaxColumnMPI::ValidationImpl() { | |
| 46 | const auto &in = GetInput(); | ||
| 47 | |||
| 48 | if (in.empty() || in[0].empty()) { | ||
| 49 | return true; | ||
| 50 | } | ||
| 51 | |||
| 52 | const std::size_t cols = in[0].size(); | ||
| 53 | |||
| 54 | for (const auto &row : in) { | ||
| 55 | if (row.size() != cols) { | ||
| 56 | return true; | ||
| 57 | } | ||
| 58 | } | ||
| 59 | |||
| 60 | return true; | ||
| 61 | } | ||
| 62 | |||
| 63 | 24 | bool LiulinYMatrixMaxColumnMPI::PreProcessingImpl() { | |
| 64 | 24 | return true; | |
| 65 | } | ||
| 66 | |||
| 67 | namespace { | ||
| 68 | 20 | void PrepareCounts(int rows, int cols, int world_size, std::vector<int> &sendcounts, std::vector<int> &displs) { | |
| 69 | 20 | sendcounts.assign(world_size, 0); | |
| 70 | 20 | displs.assign(world_size, 0); | |
| 71 | |||
| 72 | 20 | int base_cols = cols / world_size; | |
| 73 | 20 | int remainder = cols % world_size; | |
| 74 | |||
| 75 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 20 times.
|
60 | for (int proc = 0; proc < world_size; ++proc) { |
| 76 |
2/2✓ Branch 0 taken 26 times.
✓ Branch 1 taken 14 times.
|
40 | int local_cols = base_cols + (proc < remainder ? 1 : 0); |
| 77 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | sendcounts[proc] = local_cols * rows; |
| 78 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | if (proc > 0) { |
| 79 | 20 | displs[proc] = displs[proc - 1] + sendcounts[proc - 1]; | |
| 80 | } | ||
| 81 | } | ||
| 82 | 20 | } | |
| 83 | |||
| 84 | 10 | void FillFlatMatrix(const InType &in, int rows, int cols, std::vector<int> &flat_matrix) { | |
| 85 | 10 | flat_matrix.resize(static_cast<std::size_t>(rows) * static_cast<std::size_t>(cols)); | |
| 86 |
2/2✓ Branch 0 taken 49 times.
✓ Branch 1 taken 10 times.
|
59 | for (int col = 0; col < cols; ++col) { |
| 87 |
2/2✓ Branch 0 taken 571 times.
✓ Branch 1 taken 49 times.
|
620 | for (int row = 0; row < rows; ++row) { |
| 88 | 571 | flat_matrix[(col * rows) + row] = in[row][col]; | |
| 89 | } | ||
| 90 | } | ||
| 91 | 10 | } | |
| 92 | |||
| 93 | 20 | std::vector<int> ComputeLocalMaxes(const std::vector<int> &local_data, int rows, int local_cols) { | |
| 94 | 20 | std::vector<int> local_maxes(local_cols, std::numeric_limits<int>::min()); | |
| 95 |
2/2✓ Branch 0 taken 49 times.
✓ Branch 1 taken 20 times.
|
69 | for (int col_idx = 0; col_idx < local_cols; ++col_idx) { |
| 96 |
1/4✓ Branch 1 taken 49 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
49 | std::vector<int> column(rows); |
| 97 |
2/2✓ Branch 0 taken 571 times.
✓ Branch 1 taken 49 times.
|
620 | for (int row = 0; row < rows; ++row) { |
| 98 | 571 | column[row] = local_data[(col_idx * rows) + row]; | |
| 99 | } | ||
| 100 |
2/4✓ Branch 1 taken 49 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 49 times.
✗ Branch 4 not taken.
|
49 | local_maxes[col_idx] = LiulinYMatrixMaxColumnMPI::TournamentMax(column); |
| 101 | } | ||
| 102 | 20 | return local_maxes; | |
| 103 | } | ||
| 104 | } // namespace | ||
| 105 | |||
| 106 | 24 | bool LiulinYMatrixMaxColumnMPI::RunImpl() { | |
| 107 | const auto &in = GetInput(); | ||
| 108 | auto &out = GetOutput(); | ||
| 109 | |||
| 110 | 24 | int world_size = 0; | |
| 111 | 24 | int world_rank = 0; | |
| 112 | 24 | MPI_Comm_size(MPI_COMM_WORLD, &world_size); | |
| 113 | 24 | MPI_Comm_rank(MPI_COMM_WORLD, &world_rank); | |
| 114 | |||
| 115 | 24 | int rows = 0; | |
| 116 | 24 | int cols = 0; | |
| 117 |
6/6✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 11 times.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 10 times.
✓ Branch 5 taken 1 times.
|
24 | if (world_rank == 0 && !in.empty() && !in[0].empty()) { |
| 118 | 10 | rows = static_cast<int>(in.size()); | |
| 119 | 10 | cols = static_cast<int>(in[0].size()); | |
| 120 | } | ||
| 121 | |||
| 122 | 24 | MPI_Bcast(&rows, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 123 | 24 | MPI_Bcast(&cols, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 124 | |||
| 125 |
3/4✓ Branch 0 taken 20 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 20 times.
|
24 | if (rows <= 0 || cols <= 0) { |
| 126 | out.clear(); | ||
| 127 | 4 | return true; | |
| 128 | } | ||
| 129 | |||
| 130 | 20 | out.assign(cols, std::numeric_limits<int>::min()); | |
| 131 | |||
| 132 | 20 | std::vector<int> sendcounts; | |
| 133 | 20 | std::vector<int> displs; | |
| 134 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | PrepareCounts(rows, cols, world_size, sendcounts, displs); |
| 135 | |||
| 136 | 20 | std::vector<int> flat_matrix; | |
| 137 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
20 | if (world_rank == 0) { |
| 138 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | FillFlatMatrix(in, rows, cols, flat_matrix); |
| 139 | } | ||
| 140 | |||
| 141 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | int local_cols = sendcounts[world_rank] / rows; |
| 142 | 20 | int local_elements = local_cols * rows; | |
| 143 |
3/6✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 10 times.
✓ Branch 4 taken 10 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
20 | std::vector<int> local_data(local_elements); |
| 144 | |||
| 145 |
3/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 3 taken 20 times.
✗ Branch 4 not taken.
|
30 | MPI_Scatterv(world_rank == 0 ? flat_matrix.data() : nullptr, sendcounts.data(), displs.data(), MPI_INT, |
| 146 | local_data.data(), local_elements, MPI_INT, 0, MPI_COMM_WORLD); | ||
| 147 | |||
| 148 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | std::vector<int> local_maxes = ComputeLocalMaxes(local_data, rows, local_cols); |
| 149 | |||
| 150 |
1/4✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
20 | std::vector<int> recvcounts(world_size); |
| 151 |
1/4✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
20 | std::vector<int> displs_gather(world_size, 0); |
| 152 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 20 times.
|
60 | for (int proc = 0; proc < world_size; ++proc) { |
| 153 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | recvcounts[proc] = sendcounts[proc] / rows; |
| 154 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | if (proc > 0) { |
| 155 | 20 | displs_gather[proc] = displs_gather[proc - 1] + recvcounts[proc - 1]; | |
| 156 | } | ||
| 157 | } | ||
| 158 | |||
| 159 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | MPI_Gatherv(local_maxes.data(), local_cols, MPI_INT, out.data(), recvcounts.data(), displs_gather.data(), MPI_INT, 0, |
| 160 | MPI_COMM_WORLD); | ||
| 161 | |||
| 162 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | MPI_Bcast(out.data(), cols, MPI_INT, 0, MPI_COMM_WORLD); |
| 163 | |||
| 164 | return true; | ||
| 165 | } | ||
| 166 | |||
| 167 | 24 | bool LiulinYMatrixMaxColumnMPI::PostProcessingImpl() { | |
| 168 | 24 | return true; | |
| 169 | } | ||
| 170 | |||
| 171 | } // namespace liulin_y_matrix_max_column | ||
| 172 |