| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "badanov_a_sparse_matrix_mult_double_ccs/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <cmath> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <tuple> | ||
| 8 | #include <utility> | ||
| 9 | #include <vector> | ||
| 10 | |||
| 11 | #include "badanov_a_sparse_matrix_mult_double_ccs/common/include/common.hpp" | ||
| 12 | |||
| 13 | namespace badanov_a_sparse_matrix_mult_double_ccs { | ||
| 14 | |||
| 15 |
1/2✓ Branch 2 taken 30 times.
✗ Branch 3 not taken.
|
30 | BadanovASparseMatrixMultDoubleCcsMPI::BadanovASparseMatrixMultDoubleCcsMPI(const InType &in) { |
| 16 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 17 | GetInput() = in; | ||
| 18 | 30 | GetOutput() = {}; | |
| 19 | 30 | } | |
| 20 | |||
| 21 | 30 | bool BadanovASparseMatrixMultDoubleCcsMPI::ValidationImpl() { | |
| 22 | 30 | int world_size = 0; | |
| 23 | 30 | MPI_Comm_size(MPI_COMM_WORLD, &world_size); | |
| 24 | |||
| 25 | const auto &in = GetInput(); | ||
| 26 | |||
| 27 | const auto &values_a = std::get<0>(in); | ||
| 28 | const auto &row_indices_a = std::get<1>(in); | ||
| 29 | const auto &col_pointers_a = std::get<2>(in); | ||
| 30 | const auto &values_b = std::get<3>(in); | ||
| 31 | const auto &row_indices_b = std::get<4>(in); | ||
| 32 | const auto &col_pointers_b = std::get<5>(in); | ||
| 33 | 30 | int rows_a = std::get<6>(in); | |
| 34 | 30 | int cols_a = std::get<7>(in); | |
| 35 | 30 | int cols_b = std::get<8>(in); | |
| 36 | |||
| 37 |
2/4✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 30 times.
✗ Branch 3 not taken.
|
30 | if (rows_a <= 0 || cols_a <= 0 || cols_b <= 0) { |
| 38 | return false; | ||
| 39 | } | ||
| 40 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | if (values_a.size() != row_indices_a.size()) { |
| 41 | return false; | ||
| 42 | } | ||
| 43 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | if (col_pointers_a.size() != static_cast<size_t>(cols_a) + 1) { |
| 44 | return false; | ||
| 45 | } | ||
| 46 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | if (values_b.size() != row_indices_b.size()) { |
| 47 | return false; | ||
| 48 | } | ||
| 49 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (col_pointers_b.size() != static_cast<size_t>(cols_b) + 1) { |
| 50 | ✗ | return false; | |
| 51 | } | ||
| 52 | |||
| 53 | return true; | ||
| 54 | } | ||
| 55 | |||
| 56 | 30 | bool BadanovASparseMatrixMultDoubleCcsMPI::PreProcessingImpl() { | |
| 57 | 30 | return true; | |
| 58 | } | ||
| 59 | |||
| 60 | 30 | BadanovASparseMatrixMultDoubleCcsMPI::LocalData BadanovASparseMatrixMultDoubleCcsMPI::DistributeDataHorizontal( | |
| 61 | int world_rank, int world_size, const SparseMatrix &a, const SparseMatrix &b) { | ||
| 62 | 30 | LocalData local; | |
| 63 | 30 | local.global_rows = a.rows; | |
| 64 | 30 | local.global_inner_dim = a.cols; | |
| 65 | 30 | local.global_cols = b.cols; | |
| 66 | |||
| 67 | 30 | int rows_per_proc = a.rows / world_size; | |
| 68 | 30 | int extra_rows = a.rows % world_size; | |
| 69 | |||
| 70 | 30 | int local_start_row = world_rank * rows_per_proc; | |
| 71 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (world_rank < extra_rows) { |
| 72 | ✗ | local_start_row += world_rank; | |
| 73 | } else { | ||
| 74 | 30 | local_start_row += extra_rows; | |
| 75 | } | ||
| 76 | |||
| 77 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | int local_rows = rows_per_proc + (world_rank < extra_rows ? 1 : 0); |
| 78 | |||
| 79 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (local_rows == 0) { |
| 80 | local.a_local.rows = 0; | ||
| 81 | ✗ | local.a_local.cols = a.cols; | |
| 82 | ✗ | local.a_local.col_pointers.resize(a.cols + 1, 0); | |
| 83 | ✗ | local.b_local = b; | |
| 84 | return local; | ||
| 85 | } | ||
| 86 | |||
| 87 | 30 | std::vector<double> local_a_values; | |
| 88 | 30 | std::vector<int> local_a_row_indices; | |
| 89 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<int> local_a_col_pointers(a.cols + 1, 0); |
| 90 | 30 | int local_end_row = local_start_row + local_rows; | |
| 91 | |||
| 92 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<std::vector<double>> temp_values(a.cols); |
| 93 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | std::vector<std::vector<int>> temp_indices(a.cols); |
| 94 | |||
| 95 |
2/2✓ Branch 0 taken 10020 times.
✓ Branch 1 taken 30 times.
|
10050 | for (int col = 0; col < a.cols; ++col) { |
| 96 |
2/2✓ Branch 0 taken 183504 times.
✓ Branch 1 taken 10020 times.
|
193524 | for (int idx = a.col_pointers[col]; idx < a.col_pointers[col + 1]; ++idx) { |
| 97 |
2/2✓ Branch 0 taken 91752 times.
✓ Branch 1 taken 91752 times.
|
183504 | int row = a.row_indices[idx]; |
| 98 | 183504 | double val = a.values[idx]; | |
| 99 | |||
| 100 |
2/2✓ Branch 0 taken 91752 times.
✓ Branch 1 taken 91752 times.
|
183504 | if (row >= local_start_row && row < local_end_row) { |
| 101 | temp_values[col].push_back(val); | ||
| 102 |
1/2✓ Branch 1 taken 91752 times.
✗ Branch 2 not taken.
|
91752 | temp_indices[col].push_back(row - local_start_row); |
| 103 | } | ||
| 104 | } | ||
| 105 | } | ||
| 106 | |||
| 107 |
2/2✓ Branch 0 taken 10020 times.
✓ Branch 1 taken 30 times.
|
10050 | for (int col = 0; col < a.cols; ++col) { |
| 108 | 10020 | local_a_col_pointers[col + 1] = local_a_col_pointers[col] + static_cast<int>(temp_values[col].size()); | |
| 109 | |||
| 110 |
2/2✓ Branch 0 taken 91752 times.
✓ Branch 1 taken 10020 times.
|
101772 | for (size_t i = 0; i < temp_values[col].size(); ++i) { |
| 111 | local_a_values.push_back(temp_values[col][i]); | ||
| 112 | local_a_row_indices.push_back(temp_indices[col][i]); | ||
| 113 | } | ||
| 114 | } | ||
| 115 | |||
| 116 | 30 | local.a_local.values = std::move(local_a_values); | |
| 117 | 30 | local.a_local.row_indices = std::move(local_a_row_indices); | |
| 118 | 30 | local.a_local.col_pointers = std::move(local_a_col_pointers); | |
| 119 | 30 | local.a_local.rows = local_rows; | |
| 120 | 30 | local.a_local.cols = a.cols; | |
| 121 | |||
| 122 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | local.b_local = b; |
| 123 | |||
| 124 | return local; | ||
| 125 | 30 | } | |
| 126 | |||
| 127 | 9220 | std::vector<double> BadanovASparseMatrixMultDoubleCcsMPI::SparseDotProduct(const SparseMatrix &a, const SparseMatrix &b, | |
| 128 | int col_b) { | ||
| 129 | 9220 | std::vector<double> result(a.rows, 0.0); | |
| 130 | |||
| 131 |
2/2✓ Branch 0 taken 183278 times.
✓ Branch 1 taken 9220 times.
|
192498 | for (int idx_b = b.col_pointers[col_b]; idx_b < b.col_pointers[col_b + 1]; ++idx_b) { |
| 132 | 183278 | int row_b = b.row_indices[idx_b]; | |
| 133 | 183278 | double val_b = b.values[idx_b]; | |
| 134 | |||
| 135 |
2/2✓ Branch 0 taken 2277491 times.
✓ Branch 1 taken 183278 times.
|
2460769 | for (int idx_a = a.col_pointers[row_b]; idx_a < a.col_pointers[row_b + 1]; ++idx_a) { |
| 136 | 2277491 | int row_a = a.row_indices[idx_a]; | |
| 137 | 2277491 | double val_a = a.values[idx_a]; | |
| 138 | 2277491 | result[row_a] += val_a * val_b; | |
| 139 | } | ||
| 140 | } | ||
| 141 | |||
| 142 | 9220 | return result; | |
| 143 | } | ||
| 144 | |||
| 145 | 30 | SparseMatrix BadanovASparseMatrixMultDoubleCcsMPI::MultiplyLocal(const LocalData &local) { | |
| 146 | 30 | std::vector<double> values_c; | |
| 147 | 30 | std::vector<int> row_indices_c; | |
| 148 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<int> col_pointers_c(local.global_cols + 1, 0); |
| 149 | |||
| 150 |
2/2✓ Branch 0 taken 9220 times.
✓ Branch 1 taken 30 times.
|
9250 | for (int col_b = 0; col_b < local.global_cols; ++col_b) { |
| 151 |
1/2✓ Branch 1 taken 9220 times.
✗ Branch 2 not taken.
|
9220 | std::vector<double> local_col = SparseDotProduct(local.a_local, local.b_local, col_b); |
| 152 | |||
| 153 |
2/2✓ Branch 0 taken 2392600 times.
✓ Branch 1 taken 9220 times.
|
2401820 | for (int row = 0; row < local.a_local.rows; ++row) { |
| 154 |
2/2✓ Branch 0 taken 1221186 times.
✓ Branch 1 taken 1171414 times.
|
2392600 | if (std::abs(local_col[row]) > 1e-10) { |
| 155 | values_c.push_back(local_col[row]); | ||
| 156 | row_indices_c.push_back(row); | ||
| 157 | 1221186 | col_pointers_c[col_b + 1]++; | |
| 158 | } | ||
| 159 | } | ||
| 160 | } | ||
| 161 | |||
| 162 |
2/2✓ Branch 0 taken 9220 times.
✓ Branch 1 taken 30 times.
|
9250 | for (int col = 0; col < local.global_cols; ++col) { |
| 163 | 9220 | col_pointers_c[col + 1] += col_pointers_c[col]; | |
| 164 | } | ||
| 165 | |||
| 166 | 30 | SparseMatrix c_local; | |
| 167 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | c_local.values = values_c; |
| 168 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | c_local.row_indices = row_indices_c; |
| 169 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | c_local.col_pointers = col_pointers_c; |
| 170 | 30 | c_local.rows = local.a_local.rows; | |
| 171 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | c_local.cols = local.global_cols; |
| 172 | |||
| 173 | 30 | return c_local; | |
| 174 | ✗ | } | |
| 175 | |||
| 176 | 30 | void BadanovASparseMatrixMultDoubleCcsMPI::GatherResults(int world_rank, int world_size, const SparseMatrix &local_c, | |
| 177 | SparseMatrix &global_c) { | ||
| 178 | 30 | std::vector<int> local_nnz_per_col(local_c.cols, 0); | |
| 179 |
2/2✓ Branch 0 taken 9220 times.
✓ Branch 1 taken 30 times.
|
9250 | for (int col = 0; col < local_c.cols; ++col) { |
| 180 | 9220 | local_nnz_per_col[col] = local_c.col_pointers[col + 1] - local_c.col_pointers[col]; | |
| 181 | } | ||
| 182 | |||
| 183 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<int> global_nnz_per_col(local_c.cols, 0); |
| 184 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Allreduce(local_nnz_per_col.data(), global_nnz_per_col.data(), local_c.cols, MPI_INT, MPI_SUM, MPI_COMM_WORLD); |
| 185 | |||
| 186 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | global_c.col_pointers.resize(local_c.cols + 1, 0); |
| 187 |
2/2✓ Branch 0 taken 9220 times.
✓ Branch 1 taken 30 times.
|
9250 | for (int col = 0; col < local_c.cols; ++col) { |
| 188 | 9220 | global_c.col_pointers[col + 1] = global_c.col_pointers[col] + global_nnz_per_col[col]; | |
| 189 | } | ||
| 190 | |||
| 191 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | int total_nnz = global_c.col_pointers[local_c.cols]; |
| 192 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | global_c.values.resize(total_nnz); |
| 193 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | global_c.row_indices.resize(total_nnz); |
| 194 | 30 | global_c.rows = local_c.rows * world_size; | |
| 195 | 30 | global_c.cols = local_c.cols; | |
| 196 | |||
| 197 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<int> displs(world_size, 0); |
| 198 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<int> recvcounts(world_size, 0); |
| 199 | |||
| 200 |
2/2✓ Branch 0 taken 9220 times.
✓ Branch 1 taken 30 times.
|
9250 | for (int col = 0; col < local_c.cols; ++col) { |
| 201 |
1/2✓ Branch 1 taken 9220 times.
✗ Branch 2 not taken.
|
9220 | int local_start = local_c.col_pointers[col]; |
| 202 |
1/2✓ Branch 1 taken 9220 times.
✗ Branch 2 not taken.
|
9220 | int local_count = local_c.col_pointers[col + 1] - local_start; |
| 203 | |||
| 204 |
1/2✓ Branch 1 taken 9220 times.
✗ Branch 2 not taken.
|
9220 | MPI_Gather(&local_count, 1, MPI_INT, recvcounts.data(), 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 205 | |||
| 206 |
2/2✓ Branch 0 taken 4610 times.
✓ Branch 1 taken 4610 times.
|
9220 | if (world_rank == 0) { |
| 207 | 4610 | displs[0] = global_c.col_pointers[col]; | |
| 208 |
2/2✓ Branch 0 taken 4610 times.
✓ Branch 1 taken 4610 times.
|
9220 | for (int i = 1; i < world_size; ++i) { |
| 209 | 4610 | displs[i] = displs[i - 1] + recvcounts[i - 1]; | |
| 210 | } | ||
| 211 | } | ||
| 212 | |||
| 213 |
2/2✓ Branch 0 taken 9147 times.
✓ Branch 1 taken 73 times.
|
9220 | if (local_count > 0) { |
| 214 |
1/2✓ Branch 1 taken 9147 times.
✗ Branch 2 not taken.
|
9147 | MPI_Gatherv(&local_c.values[local_start], local_count, MPI_DOUBLE, global_c.values.data(), recvcounts.data(), |
| 215 | displs.data(), MPI_DOUBLE, 0, MPI_COMM_WORLD); | ||
| 216 | |||
| 217 |
1/4✓ Branch 1 taken 9147 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
9147 | std::vector<int> adjusted_rows(local_count); |
| 218 | 9147 | int row_offset = world_rank * local_c.rows; | |
| 219 |
2/2✓ Branch 0 taken 1221186 times.
✓ Branch 1 taken 9147 times.
|
1230333 | for (int i = 0; i < local_count; ++i) { |
| 220 | 1221186 | adjusted_rows[i] = local_c.row_indices[local_start + i] + row_offset; | |
| 221 | } | ||
| 222 | |||
| 223 |
1/2✓ Branch 1 taken 9147 times.
✗ Branch 2 not taken.
|
9147 | MPI_Gatherv(adjusted_rows.data(), local_count, MPI_INT, global_c.row_indices.data(), recvcounts.data(), |
| 224 | displs.data(), MPI_INT, 0, MPI_COMM_WORLD); | ||
| 225 | } else { | ||
| 226 |
1/2✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
|
73 | MPI_Gatherv(nullptr, 0, MPI_DOUBLE, global_c.values.data(), recvcounts.data(), displs.data(), MPI_DOUBLE, 0, |
| 227 | MPI_COMM_WORLD); | ||
| 228 | |||
| 229 |
1/2✓ Branch 1 taken 73 times.
✗ Branch 2 not taken.
|
73 | MPI_Gatherv(nullptr, 0, MPI_INT, global_c.row_indices.data(), recvcounts.data(), displs.data(), MPI_INT, 0, |
| 230 | MPI_COMM_WORLD); | ||
| 231 | } | ||
| 232 | } | ||
| 233 | |||
| 234 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (world_rank != 0) { |
| 235 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | global_c.values.resize(total_nnz); |
| 236 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | global_c.row_indices.resize(total_nnz); |
| 237 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | global_c.col_pointers.resize(local_c.cols + 1); |
| 238 | } | ||
| 239 | |||
| 240 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(global_c.values.data(), total_nnz, MPI_DOUBLE, 0, MPI_COMM_WORLD); |
| 241 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(global_c.row_indices.data(), total_nnz, MPI_INT, 0, MPI_COMM_WORLD); |
| 242 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(global_c.col_pointers.data(), local_c.cols + 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 243 | 30 | } | |
| 244 | |||
| 245 | 30 | bool BadanovASparseMatrixMultDoubleCcsMPI::RunImpl() { | |
| 246 | 30 | int world_rank = 0; | |
| 247 | 30 | int world_size = 0; | |
| 248 | 30 | MPI_Comm_rank(MPI_COMM_WORLD, &world_rank); | |
| 249 | 30 | MPI_Comm_size(MPI_COMM_WORLD, &world_size); | |
| 250 | |||
| 251 | 30 | std::vector<double> values_a; | |
| 252 | 30 | std::vector<double> values_b; | |
| 253 | 30 | std::vector<int> row_indices_a; | |
| 254 | 30 | std::vector<int> row_indices_b; | |
| 255 | 30 | std::vector<int> col_pointers_a; | |
| 256 | 30 | std::vector<int> col_pointers_b; | |
| 257 | 30 | int rows_a = 0; | |
| 258 | 30 | int cols_a = 0; | |
| 259 | 30 | int cols_b = 0; | |
| 260 | |||
| 261 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (world_rank == 0) { |
| 262 | const auto &in = GetInput(); | ||
| 263 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | values_a = std::get<0>(in); |
| 264 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | row_indices_a = std::get<1>(in); |
| 265 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | col_pointers_a = std::get<2>(in); |
| 266 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | values_b = std::get<3>(in); |
| 267 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | row_indices_b = std::get<4>(in); |
| 268 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | col_pointers_b = std::get<5>(in); |
| 269 | 15 | rows_a = std::get<6>(in); | |
| 270 | 15 | cols_a = std::get<7>(in); | |
| 271 | 15 | cols_b = std::get<8>(in); | |
| 272 | } | ||
| 273 | |||
| 274 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(&rows_a, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 275 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(&cols_a, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 276 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(&cols_b, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 277 | |||
| 278 | 30 | int nnz_a = 0; | |
| 279 | 30 | int nnz_b = 0; | |
| 280 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (world_rank == 0) { |
| 281 | 15 | nnz_a = static_cast<int>(values_a.size()); | |
| 282 | 15 | nnz_b = static_cast<int>(values_b.size()); | |
| 283 | } | ||
| 284 | |||
| 285 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(&nnz_a, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 286 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(&nnz_b, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 287 | |||
| 288 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (world_rank != 0) { |
| 289 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | values_a.resize(nnz_a); |
| 290 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | row_indices_a.resize(nnz_a); |
| 291 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | col_pointers_a.resize(cols_a + 1); |
| 292 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | values_b.resize(nnz_b); |
| 293 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | row_indices_b.resize(nnz_b); |
| 294 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | col_pointers_b.resize(cols_b + 1); |
| 295 | } | ||
| 296 | |||
| 297 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(values_a.data(), nnz_a, MPI_DOUBLE, 0, MPI_COMM_WORLD); |
| 298 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(row_indices_a.data(), nnz_a, MPI_INT, 0, MPI_COMM_WORLD); |
| 299 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(col_pointers_a.data(), cols_a + 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 300 | |||
| 301 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(values_b.data(), nnz_b, MPI_DOUBLE, 0, MPI_COMM_WORLD); |
| 302 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(row_indices_b.data(), nnz_b, MPI_INT, 0, MPI_COMM_WORLD); |
| 303 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | MPI_Bcast(col_pointers_b.data(), cols_b + 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 304 | |||
| 305 | 30 | SparseMatrix a_global; | |
| 306 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | a_global.values = values_a; |
| 307 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | a_global.row_indices = row_indices_a; |
| 308 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | a_global.col_pointers = col_pointers_a; |
| 309 | 30 | a_global.rows = rows_a; | |
| 310 | 30 | a_global.cols = cols_a; | |
| 311 | |||
| 312 | 30 | SparseMatrix b_global; | |
| 313 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | b_global.values = values_b; |
| 314 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | b_global.row_indices = row_indices_b; |
| 315 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | b_global.col_pointers = col_pointers_b; |
| 316 | 30 | b_global.rows = cols_a; | |
| 317 | 30 | b_global.cols = cols_b; | |
| 318 | |||
| 319 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | LocalData local_data = DistributeDataHorizontal(world_rank, world_size, a_global, b_global); |
| 320 | |||
| 321 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | SparseMatrix local_c = MultiplyLocal(local_data); |
| 322 | |||
| 323 | 30 | SparseMatrix global_c; | |
| 324 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | GatherResults(world_rank, world_size, local_c, global_c); |
| 325 | |||
| 326 | 30 | GetOutput() = std::make_tuple(global_c.values, global_c.row_indices, global_c.col_pointers); | |
| 327 | |||
| 328 | 30 | return true; | |
| 329 | 60 | } | |
| 330 | |||
| 331 | 30 | bool BadanovASparseMatrixMultDoubleCcsMPI::PostProcessingImpl() { | |
| 332 | 30 | return true; | |
| 333 | } | ||
| 334 | |||
| 335 | } // namespace badanov_a_sparse_matrix_mult_double_ccs | ||
| 336 |