| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "votincev_d_matrix_mult/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <tuple> | ||
| 8 | #include <vector> | ||
| 9 | |||
| 10 | #include "votincev_d_matrix_mult/common/include/common.hpp" | ||
| 11 | |||
| 12 | namespace votincev_d_matrix_mult { | ||
| 13 | |||
| 14 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | VotincevDMatrixMultMPI::VotincevDMatrixMultMPI(const InType &in) { |
| 15 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 16 | GetInput() = in; | ||
| 17 | 20 | } | |
| 18 | |||
| 19 | // проверка входных данных | ||
| 20 | 20 | bool VotincevDMatrixMultMPI::ValidationImpl() { | |
| 21 | const auto &in = GetInput(); | ||
| 22 | 20 | int param_m = std::get<0>(in); | |
| 23 | 20 | int param_n = std::get<1>(in); | |
| 24 | 20 | int param_k = std::get<2>(in); | |
| 25 | const auto &matrix_a = std::get<3>(in); | ||
| 26 | const auto &matrix_b = std::get<4>(in); | ||
| 27 | |||
| 28 |
3/6✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 20 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 20 times.
|
20 | return (param_m > 0 && param_n > 0 && param_k > 0 && static_cast<int>(matrix_a.size()) == (param_m * param_k) && |
| 29 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
|
20 | static_cast<int>(matrix_b.size()) == (param_k * param_n)); |
| 30 | } | ||
| 31 | |||
| 32 | // препроцессинг | ||
| 33 | 20 | bool VotincevDMatrixMultMPI::PreProcessingImpl() { | |
| 34 | 20 | return true; | |
| 35 | } | ||
| 36 | |||
| 37 | // главный метод MPI | ||
| 38 | 20 | bool VotincevDMatrixMultMPI::RunImpl() { | |
| 39 | // получаю кол-во процессов | ||
| 40 | 20 | int process_n = 0; | |
| 41 | 20 | MPI_Comm_size(MPI_COMM_WORLD, &process_n); | |
| 42 | |||
| 43 | // получаю ранг текущего | ||
| 44 | 20 | int proc_rank = 0; | |
| 45 | 20 | MPI_Comm_rank(MPI_COMM_WORLD, &proc_rank); | |
| 46 | |||
| 47 | // размерности матриц получают все процессы | ||
| 48 | int m = 0; | ||
| 49 | int n = 0; | ||
| 50 | int k = 0; | ||
| 51 | const auto &in = GetInput(); | ||
| 52 | 20 | m = std::get<0>(in); | |
| 53 | 20 | n = std::get<1>(in); | |
| 54 |
2/2✓ Branch 0 taken 18 times.
✓ Branch 1 taken 2 times.
|
20 | k = std::get<2>(in); |
| 55 | |||
| 56 | // если процессов больше чем строк в матрице A - | ||
| 57 | // то активных процессов будет меньше (m) | ||
| 58 | // (потому что разедление по строкам) | ||
| 59 | 20 | process_n = std::min(process_n, m); | |
| 60 | |||
| 61 | // "лишние" процессы не работают | ||
| 62 |
2/2✓ Branch 0 taken 18 times.
✓ Branch 1 taken 2 times.
|
20 | if (proc_rank >= process_n) { |
| 63 | return true; | ||
| 64 | } | ||
| 65 | |||
| 66 | 18 | std::vector<double> matrix_a; | |
| 67 |
1/2✓ Branch 1 taken 18 times.
✗ Branch 2 not taken.
|
18 | std::vector<double> matrix_b; |
| 68 | |||
| 69 | // матрицу B получают все процессы | ||
| 70 |
1/2✓ Branch 1 taken 18 times.
✗ Branch 2 not taken.
|
18 | matrix_b = std::get<4>(in); |
| 71 | |||
| 72 | // матрицу А получит полностью только 0й процесс | ||
| 73 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 8 times.
|
18 | if (proc_rank == 0) { |
| 74 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | matrix_a = std::get<3>(in); |
| 75 | } | ||
| 76 | |||
| 77 | // если всего 1 процесс - последовательное умножение | ||
| 78 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 16 times.
|
18 | if (process_n == 1) { |
| 79 |
1/4✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
2 | GetOutput() = SeqMatrixMult(m, n, k, matrix_a, matrix_b); |
| 80 | 2 | return true; | |
| 81 | } | ||
| 82 | |||
| 83 | // какие строки каждый процесс будет перемножать | ||
| 84 | // [start0, end0, start1, end1, ...] | ||
| 85 | 16 | std::vector<int> ranges; | |
| 86 | |||
| 87 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 8 times.
|
16 | if (proc_rank == 0) { |
| 88 | 8 | auto proc_n_sizet = static_cast<size_t>(process_n); | |
| 89 |
1/2✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
|
8 | ranges.resize(proc_n_sizet * 2); |
| 90 | |||
| 91 | // минимум на обработку | ||
| 92 | 8 | int base = m / process_n; | |
| 93 | // остаток распределим | ||
| 94 | 8 | int remain = m % process_n; | |
| 95 | |||
| 96 | int start = 0; | ||
| 97 |
2/2✓ Branch 0 taken 16 times.
✓ Branch 1 taken 8 times.
|
24 | for (size_t i = 0; i < proc_n_sizet; i++) { |
| 98 | int part = base; | ||
| 99 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 10 times.
|
16 | if (remain > 0) { |
| 100 | 6 | part++; | |
| 101 | 6 | remain--; | |
| 102 | } | ||
| 103 | |||
| 104 | // i+i ~~ size_t+size_t .... i*2 ~~ size_t*int (clang-tidy: size_t*int - is bad) | ||
| 105 | 16 | ranges[i + i] = start; | |
| 106 | 16 | ranges[i + i + 1] = start + part; // end (не включительно) | |
| 107 | |||
| 108 | start += part; | ||
| 109 | } | ||
| 110 | } | ||
| 111 | |||
| 112 | // my_range получит [start, end] | ||
| 113 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<int> my_range = {0, 0}; |
| 114 | |||
| 115 | // local_matrix — локальный блок матрицы A данного процесса | ||
| 116 | 16 | std::vector<double> local_matrix; | |
| 117 | |||
| 118 | // отправляю всем часть матрицы A в local_matrix | ||
| 119 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | SendData(k, proc_rank, process_n, my_range, ranges, local_matrix, matrix_a); |
| 120 | |||
| 121 | // теперь каждый владеет своим куском (local_matrix) | ||
| 122 | // вызываем обычное перемножение, результат умножение кладется в local_matrix | ||
| 123 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | MatrixPartMult(k, n, local_matrix, matrix_b); |
| 124 | |||
| 125 | // сбор результатов назад к 0му | ||
| 126 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 8 times.
|
16 | if (proc_rank == 0) { |
| 127 |
2/6✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
8 | std::vector<double> final_result(static_cast<size_t>(m * n)); |
| 128 |
1/2✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
|
8 | int my_rows = ranges[1] - ranges[0]; // сколько строк |
| 129 | // копирую откуда до куда куда | ||
| 130 | std::ranges::copy(local_matrix.begin(), local_matrix.end(), final_result.begin()); | ||
| 131 | |||
| 132 | // смещение; изначально равно количеству уже записанных значений | ||
| 133 | 8 | int offset = my_rows * n; | |
| 134 | 8 | auto proc_n_sizet = static_cast<size_t>(process_n); | |
| 135 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 8 times.
|
16 | for (size_t i = 1; i < proc_n_sizet; ++i) { |
| 136 | 8 | int start_i = ranges[i + i]; | |
| 137 | 8 | int end_i = ranges[i + i + 1]; | |
| 138 | 8 | int rows = end_i - start_i; | |
| 139 | |||
| 140 | // сколько в данной пачке элементов | ||
| 141 | 8 | int count = rows * n; | |
| 142 | |||
| 143 |
1/2✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
|
8 | MPI_Recv(final_result.data() + static_cast<size_t>(offset), count, MPI_DOUBLE, static_cast<int>(i), 0, |
| 144 | MPI_COMM_WORLD, MPI_STATUS_IGNORE); | ||
| 145 | |||
| 146 | 8 | offset += count; | |
| 147 | } | ||
| 148 | // 0й процесс собрал всё от других процессов | ||
| 149 | |||
| 150 |
1/2✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
|
8 | GetOutput() = final_result; |
| 151 | |||
| 152 | } else { | ||
| 153 | // другие процессы посылают свои результаты основному 0му процессу | ||
| 154 | 8 | int rows = my_range[1] - my_range[0]; | |
| 155 |
1/2✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
|
8 | MPI_Send(local_matrix.data(), rows * n, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD); |
| 156 | } | ||
| 157 | |||
| 158 | return true; | ||
| 159 | } | ||
| 160 | |||
| 161 | // =============================== | ||
| 162 | // ==== дополнительные функции === | ||
| 163 | // =============================== | ||
| 164 | |||
| 165 | 16 | void VotincevDMatrixMultMPI::SendData(int k, int proc_rank, int process_n, std::vector<int> &my_range, | |
| 166 | std::vector<int> &ranges, std::vector<double> &local_matrix, | ||
| 167 | std::vector<double> &matrix_a) { | ||
| 168 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 8 times.
|
16 | if (proc_rank == 0) { |
| 169 | // заполняю данные для себя — свои строки матрицы А | ||
| 170 | 8 | int my_start = ranges[0]; | |
| 171 | 8 | int my_end = ranges[1]; | |
| 172 | 8 | int my_rows = my_end - my_start; | |
| 173 | |||
| 174 | 8 | auto local_matrix_size = static_cast<size_t>(my_rows) * static_cast<size_t>(k); | |
| 175 | 8 | local_matrix.resize(local_matrix_size); | |
| 176 |
2/2✓ Branch 0 taken 218 times.
✓ Branch 1 taken 8 times.
|
226 | for (size_t i = 0; i < local_matrix_size; i++) { |
| 177 | 218 | local_matrix[i] = matrix_a[i]; | |
| 178 | } | ||
| 179 | |||
| 180 | // рассылаю остальным | ||
| 181 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 8 times.
|
16 | for (int i = 1; i < process_n; i++) { |
| 182 | 8 | int start_i = ranges[i + i]; | |
| 183 | 8 | int end_i = ranges[i + i + 1]; | |
| 184 | |||
| 185 | 8 | MPI_Send(&start_i, 1, MPI_INT, i, 0, MPI_COMM_WORLD); | |
| 186 | 8 | MPI_Send(&end_i, 1, MPI_INT, i, 0, MPI_COMM_WORLD); | |
| 187 | |||
| 188 | 8 | int elem_count = (end_i - start_i) * k; | |
| 189 | |||
| 190 | 8 | auto offset = static_cast<size_t>(start_i) * static_cast<size_t>(k); | |
| 191 | 8 | MPI_Send(matrix_a.data() + offset, elem_count, MPI_DOUBLE, i, 0, MPI_COMM_WORLD); | |
| 192 | } | ||
| 193 | |||
| 194 | } else { | ||
| 195 | // получаю диапазон | ||
| 196 | 8 | MPI_Recv(my_range.data(), 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); | |
| 197 | 8 | MPI_Recv(&my_range[1], 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); | |
| 198 | |||
| 199 | 8 | int start_i = my_range[0]; | |
| 200 | 8 | int end_i = my_range[1]; | |
| 201 | 8 | int elem_count = (end_i - start_i) * k; | |
| 202 | |||
| 203 | 8 | local_matrix.resize(elem_count); | |
| 204 | |||
| 205 | // получаю матрицу (часть) | ||
| 206 | 8 | MPI_Recv(local_matrix.data(), elem_count, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); | |
| 207 | } | ||
| 208 | 16 | } | |
| 209 | |||
| 210 | // простое последовательное умножение (если кол-во_процессов == 1) | ||
| 211 | 2 | std::vector<double> VotincevDMatrixMultMPI::SeqMatrixMult(int param_m, int param_n, int param_k, | |
| 212 | std::vector<double> &matrix_a, | ||
| 213 | std::vector<double> &matrix_b) { | ||
| 214 | 2 | std::vector<double> matrix_res; | |
| 215 |
1/4✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
2 | matrix_res.assign(static_cast<size_t>(param_m) * static_cast<size_t>(param_n), 0.0); |
| 216 | |||
| 217 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
|
4 | for (int i = 0; i < param_m; ++i) { |
| 218 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 2 times.
|
12 | for (int j = 0; j < param_n; ++j) { |
| 219 | double sum = 0.0; | ||
| 220 |
2/2✓ Branch 0 taken 82 times.
✓ Branch 1 taken 10 times.
|
92 | for (int k = 0; k < param_k; ++k) { |
| 221 | 82 | sum += matrix_a[(i * param_k) + k] * matrix_b[(k * param_n) + j]; | |
| 222 | } | ||
| 223 | 10 | matrix_res[(i * param_n) + j] = sum; | |
| 224 | } | ||
| 225 | } | ||
| 226 | 2 | return matrix_res; | |
| 227 | } | ||
| 228 | |||
| 229 | // умножение части матрицы A на всю матрицу B | ||
| 230 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | void VotincevDMatrixMultMPI::MatrixPartMult(int param_k, int param_n, std::vector<double> &local_matrix, |
| 231 | const std::vector<double> &matrix_b) { | ||
| 232 | 16 | size_t str_count = local_matrix.size() / param_k; | |
| 233 | |||
| 234 | 16 | std::vector<double> result; | |
| 235 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | result.resize(str_count * param_n); |
| 236 | |||
| 237 |
2/2✓ Branch 0 taken 54 times.
✓ Branch 1 taken 16 times.
|
70 | for (size_t i = 0; i < str_count; i++) { |
| 238 |
2/2✓ Branch 0 taken 390 times.
✓ Branch 1 taken 54 times.
|
444 | for (int j = 0; j < param_n; j++) { |
| 239 | double sum = 0.0; | ||
| 240 |
2/2✓ Branch 0 taken 3323 times.
✓ Branch 1 taken 390 times.
|
3713 | for (int k = 0; k < param_k; k++) { |
| 241 | 3323 | sum += local_matrix[(i * param_k) + k] * matrix_b[(k * param_n) + j]; | |
| 242 | } | ||
| 243 | 390 | result[(i * param_n) + j] = sum; | |
| 244 | } | ||
| 245 | } | ||
| 246 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | local_matrix = result; |
| 247 | 16 | } | |
| 248 | |||
| 249 | 20 | bool VotincevDMatrixMultMPI::PostProcessingImpl() { | |
| 250 | 20 | return true; | |
| 251 | } | ||
| 252 | |||
| 253 | } // namespace votincev_d_matrix_mult | ||
| 254 |