| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "liulin_y_complex_ccs/stl/include/ops_stl.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <cmath> | ||
| 5 | #include <complex> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <functional> | ||
| 8 | #include <map> | ||
| 9 | #include <thread> | ||
| 10 | #include <utility> | ||
| 11 | #include <vector> | ||
| 12 | |||
| 13 | #include "liulin_y_complex_ccs/common/include/common.hpp" | ||
| 14 | |||
| 15 | namespace liulin_y_complex_ccs { | ||
| 16 | |||
| 17 | namespace { | ||
| 18 | |||
| 19 | constexpr double kEpsilon = 1e-10; | ||
| 20 | |||
| 21 | using LocalDict = std::map<std::pair<int, int>, std::complex<double>>; | ||
| 22 | |||
| 23 | 96 | bool IsValidCCS(const CCSMatrix &mat) { | |
| 24 |
2/4✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 96 times.
✗ Branch 3 not taken.
|
96 | if (mat.count_rows <= 0 || mat.count_cols <= 0) { |
| 25 | return false; | ||
| 26 | } | ||
| 27 |
1/2✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
|
96 | if (mat.col_index.size() != static_cast<size_t>(mat.count_cols) + 1) { |
| 28 | return false; | ||
| 29 | } | ||
| 30 |
1/2✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
|
96 | if (mat.col_index[0] != 0) { |
| 31 | return false; | ||
| 32 | } | ||
| 33 |
1/2✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
|
96 | if (mat.values.size() != mat.row_index.size()) { |
| 34 | return false; | ||
| 35 | } | ||
| 36 |
1/2✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
|
96 | if (std::cmp_not_equal(mat.col_index.back(), mat.values.size())) { |
| 37 | ✗ | return false; | |
| 38 | } | ||
| 39 | return true; | ||
| 40 | } | ||
| 41 | |||
| 42 | 48 | void TransposeCCS(const CCSMatrix &mat, CCSMatrix &mat_t) { | |
| 43 | 48 | mat_t.count_rows = mat.count_cols; | |
| 44 | 48 | mat_t.count_cols = mat.count_rows; | |
| 45 | const size_t nnz = mat.values.size(); | ||
| 46 | 48 | mat_t.values.resize(nnz); | |
| 47 | 48 | mat_t.row_index.resize(nnz); | |
| 48 | 48 | mat_t.col_index.assign(static_cast<size_t>(mat_t.count_cols) + 1, 0); | |
| 49 | |||
| 50 |
2/2✓ Branch 0 taken 80 times.
✓ Branch 1 taken 48 times.
|
128 | for (size_t i = 0; i < nnz; ++i) { |
| 51 | 80 | mat_t.col_index[static_cast<size_t>(mat.row_index[i]) + 1]++; | |
| 52 | } | ||
| 53 |
2/2✓ Branch 0 taken 104 times.
✓ Branch 1 taken 48 times.
|
152 | for (int i = 0; i < mat_t.count_cols; ++i) { |
| 54 | 104 | mat_t.col_index[static_cast<size_t>(i) + 1] += mat_t.col_index[static_cast<size_t>(i)]; | |
| 55 | } | ||
| 56 | |||
| 57 | 48 | std::vector<int> current_pos(mat_t.col_index.begin(), mat_t.col_index.end()); | |
| 58 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | for (int j = 0; j < mat.count_cols; ++j) { |
| 59 |
2/2✓ Branch 0 taken 80 times.
✓ Branch 1 taken 96 times.
|
176 | for (int k = mat.col_index[static_cast<size_t>(j)]; k < mat.col_index[static_cast<size_t>(j) + 1]; ++k) { |
| 60 | 80 | const int row = mat.row_index[static_cast<size_t>(k)]; | |
| 61 | 80 | const int dest = current_pos[static_cast<size_t>(row)]++; | |
| 62 | 80 | mat_t.row_index[static_cast<size_t>(dest)] = j; | |
| 63 | 80 | mat_t.values[static_cast<size_t>(dest)] = mat.values[static_cast<size_t>(k)]; | |
| 64 | } | ||
| 65 | } | ||
| 66 | 48 | } | |
| 67 | |||
| 68 | 48 | std::vector<int> BuildColumnIndices(const CCSMatrix &mat_a) { | |
| 69 | 48 | std::vector<int> a_cols(mat_a.values.size()); | |
| 70 |
2/2✓ Branch 0 taken 104 times.
✓ Branch 1 taken 48 times.
|
152 | for (int j = 0; j < mat_a.count_cols; ++j) { |
| 71 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 104 times.
|
224 | for (int k = mat_a.col_index[static_cast<size_t>(j)]; k < mat_a.col_index[static_cast<size_t>(j) + 1]; ++k) { |
| 72 | 120 | a_cols[static_cast<size_t>(k)] = j; | |
| 73 | } | ||
| 74 | } | ||
| 75 | 48 | return a_cols; | |
| 76 | } | ||
| 77 | |||
| 78 | 120 | void ProcessChunk(int start_idx, int end_idx, const CCSMatrix &mat_a, const std::vector<int> &a_cols, | |
| 79 | const CCSMatrix &mat_b_t, LocalDict &local_map) { | ||
| 80 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 120 times.
|
240 | for (int k = start_idx; k < end_idx; ++k) { |
| 81 | 120 | const int row_left = mat_a.row_index[static_cast<size_t>(k)]; | |
| 82 | 120 | const int col_left = a_cols[static_cast<size_t>(k)]; | |
| 83 | 120 | const std::complex<double> val_left = mat_a.values[static_cast<size_t>(k)]; | |
| 84 | |||
| 85 | 120 | const int b_start = mat_b_t.col_index[static_cast<size_t>(col_left)]; | |
| 86 | 120 | const int b_end = mat_b_t.col_index[static_cast<size_t>(col_left) + 1]; | |
| 87 | |||
| 88 |
2/2✓ Branch 0 taken 80 times.
✓ Branch 1 taken 120 times.
|
200 | for (int idx_p = b_start; idx_p < b_end; ++idx_p) { |
| 89 | 80 | const int col_right = mat_b_t.row_index[static_cast<size_t>(idx_p)]; | |
| 90 | 80 | const std::complex<double> val_right = mat_b_t.values[static_cast<size_t>(idx_p)]; | |
| 91 | |||
| 92 | 80 | local_map[{col_right, row_left}] += val_left * val_right; | |
| 93 | } | ||
| 94 | } | ||
| 95 | 120 | } | |
| 96 | |||
| 97 | 48 | void AggregateLocalMaps(const std::vector<LocalDict> &local_maps, LocalDict &global_map) { | |
| 98 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 48 times.
|
168 | for (const auto &local_map : local_maps) { |
| 99 |
2/2✓ Branch 0 taken 80 times.
✓ Branch 1 taken 120 times.
|
200 | for (const auto &[key, value] : local_map) { |
| 100 | 80 | global_map[key] += value; | |
| 101 | } | ||
| 102 | } | ||
| 103 | 48 | } | |
| 104 | |||
| 105 | 48 | void ConstructResultMatrix(const LocalDict &global_map, int rows, int cols, CCSMatrix &mat_res) { | |
| 106 | 48 | mat_res.count_rows = rows; | |
| 107 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | mat_res.count_cols = cols; |
| 108 | mat_res.values.clear(); | ||
| 109 | mat_res.row_index.clear(); | ||
| 110 | 48 | mat_res.col_index.assign(static_cast<size_t>(cols) + 1, 0); | |
| 111 | |||
| 112 |
2/2✓ Branch 0 taken 72 times.
✓ Branch 1 taken 48 times.
|
120 | for (const auto &[key, value] : global_map) { |
| 113 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 72 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
72 | if (std::abs(value.real()) > kEpsilon || std::abs(value.imag()) > kEpsilon) { |
| 114 | 72 | const int col = key.first; | |
| 115 | 72 | const int row = key.second; | |
| 116 | |||
| 117 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 72 times.
|
72 | mat_res.values.push_back(value); |
| 118 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 72 times.
|
72 | mat_res.row_index.push_back(row); |
| 119 | 72 | mat_res.col_index[static_cast<size_t>(col) + 1]++; | |
| 120 | } | ||
| 121 | } | ||
| 122 | |||
| 123 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | for (int i = 0; i < cols; ++i) { |
| 124 | 96 | mat_res.col_index[static_cast<size_t>(i) + 1] += mat_res.col_index[static_cast<size_t>(i)]; | |
| 125 | } | ||
| 126 | 48 | } | |
| 127 | |||
| 128 | } // namespace | ||
| 129 | |||
| 130 |
1/2✓ Branch 2 taken 48 times.
✗ Branch 3 not taken.
|
48 | LiulinYComplexCcsStl::LiulinYComplexCcsStl(const InType &in) : BaseTask() { |
| 131 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 132 | GetInput() = in; | ||
| 133 | 48 | } | |
| 134 | |||
| 135 | 48 | bool LiulinYComplexCcsStl::ValidationImpl() { | |
| 136 | 48 | const auto &mat_a = GetInput().first; | |
| 137 | 48 | const auto &mat_b = GetInput().second; | |
| 138 |
3/6✓ Branch 0 taken 48 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 48 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 48 times.
|
48 | return IsValidCCS(mat_a) && IsValidCCS(mat_b) && (mat_a.count_cols == mat_b.count_rows); |
| 139 | } | ||
| 140 | |||
| 141 | 48 | bool LiulinYComplexCcsStl::PreProcessingImpl() { | |
| 142 | 48 | return true; | |
| 143 | } | ||
| 144 | |||
| 145 | 48 | bool LiulinYComplexCcsStl::RunImpl() { | |
| 146 | 48 | const auto &mat_a = GetInput().first; | |
| 147 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | const auto &mat_b = GetInput().second; |
| 148 | auto &mat_res = GetOutput(); | ||
| 149 | |||
| 150 | 48 | const int nnz_a = static_cast<int>(mat_a.values.size()); | |
| 151 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | if (nnz_a == 0) { |
| 152 | ✗ | mat_res.count_rows = mat_a.count_rows; | |
| 153 | ✗ | mat_res.count_cols = mat_b.count_cols; | |
| 154 | ✗ | mat_res.col_index.assign(static_cast<size_t>(mat_res.count_cols) + 1, 0); | |
| 155 | ✗ | return true; | |
| 156 | } | ||
| 157 | |||
| 158 | 48 | CCSMatrix mat_b_t; | |
| 159 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | TransposeCCS(mat_b, mat_b_t); |
| 160 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | std::vector<int> a_cols = BuildColumnIndices(mat_a); |
| 161 | |||
| 162 | 48 | unsigned int hw = std::thread::hardware_concurrency(); | |
| 163 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | int num_threads = std::min(nnz_a, (hw == 0) ? 1 : static_cast<int>(hw)); |
| 164 | |||
| 165 | 48 | std::vector<std::thread> threads; | |
| 166 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | std::vector<LocalDict> local_maps(static_cast<size_t>(num_threads)); |
| 167 | 48 | int chunk = nnz_a / num_threads; | |
| 168 | 48 | int rem = nnz_a % num_threads; | |
| 169 | |||
| 170 | 48 | int start = 0; | |
| 171 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 48 times.
|
168 | for (int i = 0; i < num_threads; ++i) { |
| 172 |
1/2✓ Branch 0 taken 120 times.
✗ Branch 1 not taken.
|
120 | int end = start + chunk + (i < rem ? 1 : 0); |
| 173 |
1/2✓ Branch 1 taken 120 times.
✗ Branch 2 not taken.
|
120 | threads.emplace_back(ProcessChunk, start, end, std::ref(mat_a), std::ref(a_cols), std::ref(mat_b_t), |
| 174 |
1/2✓ Branch 1 taken 120 times.
✗ Branch 2 not taken.
|
120 | std::ref(local_maps[static_cast<size_t>(i)])); |
| 175 | 120 | start = end; | |
| 176 | } | ||
| 177 | |||
| 178 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 48 times.
|
168 | for (auto &t : threads) { |
| 179 |
1/2✓ Branch 0 taken 120 times.
✗ Branch 1 not taken.
|
120 | if (t.joinable()) { |
| 180 |
1/2✓ Branch 1 taken 120 times.
✗ Branch 2 not taken.
|
120 | t.join(); |
| 181 | } | ||
| 182 | } | ||
| 183 | |||
| 184 | LocalDict global_map; | ||
| 185 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | AggregateLocalMaps(local_maps, global_map); |
| 186 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | ConstructResultMatrix(global_map, mat_a.count_rows, mat_b.count_cols, mat_res); |
| 187 | |||
| 188 | return true; | ||
| 189 | 96 | } | |
| 190 | |||
| 191 | 48 | bool LiulinYComplexCcsStl::PostProcessingImpl() { | |
| 192 | 48 | return true; | |
| 193 | } | ||
| 194 | |||
| 195 | } // namespace liulin_y_complex_ccs | ||
| 196 |