| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "goriacheva_k_mult_sparse_complex_matrix_ccs/stl/include/ops_stl.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <ranges> | ||
| 5 | #include <thread> | ||
| 6 | #include <vector> | ||
| 7 | |||
| 8 | #include "goriacheva_k_mult_sparse_complex_matrix_ccs/common/include/common.hpp" | ||
| 9 | |||
| 10 | namespace goriacheva_k_mult_sparse_complex_matrix_ccs { | ||
| 11 | |||
| 12 |
1/2✓ Branch 2 taken 112 times.
✗ Branch 3 not taken.
|
112 | GoriachevaKMultSparseComplexMatrixCcsSTL::GoriachevaKMultSparseComplexMatrixCcsSTL(const InType &in) { |
| 13 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 14 | GetInput() = in; | ||
| 15 | 112 | } | |
| 16 | |||
| 17 | 112 | bool GoriachevaKMultSparseComplexMatrixCcsSTL::ValidationImpl() { | |
| 18 | auto &[a, b] = GetInput(); | ||
| 19 | |||
| 20 |
1/2✓ Branch 0 taken 112 times.
✗ Branch 1 not taken.
|
112 | if (a.cols != b.rows) { |
| 21 | return false; | ||
| 22 | } | ||
| 23 | |||
| 24 |
2/4✓ Branch 0 taken 112 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 112 times.
|
112 | if (a.col_ptr.empty() || b.col_ptr.empty()) { |
| 25 | ✗ | return false; | |
| 26 | } | ||
| 27 | |||
| 28 | return true; | ||
| 29 | } | ||
| 30 | |||
| 31 | 112 | bool GoriachevaKMultSparseComplexMatrixCcsSTL::PreProcessingImpl() { | |
| 32 | 112 | GetOutput() = {}; | |
| 33 | 112 | return true; | |
| 34 | } | ||
| 35 | |||
| 36 | namespace { | ||
| 37 | 328 | void ProcessColumn(int j, const SparseMatrixCCS &a, const SparseMatrixCCS &b, std::vector<Complex> &values, | |
| 38 | std::vector<int> &rows) { | ||
| 39 | 328 | std::vector<Complex> accumulator(a.rows); | |
| 40 |
1/4✓ Branch 1 taken 328 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
328 | std::vector<int> marker(a.rows, -1); |
| 41 | 328 | std::vector<int> used_rows; | |
| 42 | |||
| 43 |
2/2✓ Branch 0 taken 224 times.
✓ Branch 1 taken 328 times.
|
552 | for (int bi = b.col_ptr[j]; bi < b.col_ptr[j + 1]; bi++) { |
| 44 | 224 | int k = b.row_ind[bi]; | |
| 45 | 224 | Complex b_val = b.values[bi]; | |
| 46 | |||
| 47 |
2/2✓ Branch 0 taken 216 times.
✓ Branch 1 taken 224 times.
|
440 | for (int ai = a.col_ptr[k]; ai < a.col_ptr[k + 1]; ai++) { |
| 48 |
1/2✓ Branch 0 taken 216 times.
✗ Branch 1 not taken.
|
216 | int i = a.row_ind[ai]; |
| 49 | |||
| 50 |
1/2✓ Branch 0 taken 216 times.
✗ Branch 1 not taken.
|
216 | if (marker[i] != j) { |
| 51 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 216 times.
|
216 | marker[i] = j; |
| 52 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 216 times.
|
216 | accumulator[i] = Complex(0.0, 0.0); |
| 53 | used_rows.push_back(i); | ||
| 54 | } | ||
| 55 | |||
| 56 | accumulator[i] += a.values[ai] * b_val; | ||
| 57 | } | ||
| 58 | } | ||
| 59 | |||
| 60 | std::ranges::sort(used_rows); | ||
| 61 | |||
| 62 |
2/2✓ Branch 0 taken 216 times.
✓ Branch 1 taken 328 times.
|
544 | for (int r : used_rows) { |
| 63 |
2/2✓ Branch 0 taken 24 times.
✓ Branch 1 taken 192 times.
|
216 | if (accumulator[r] != Complex(0.0, 0.0)) { |
| 64 | rows.push_back(r); | ||
| 65 | values.push_back(accumulator[r]); | ||
| 66 | } | ||
| 67 | } | ||
| 68 | 328 | } | |
| 69 | } // namespace | ||
| 70 | |||
| 71 | 112 | bool GoriachevaKMultSparseComplexMatrixCcsSTL::RunImpl() { | |
| 72 | auto &a = std::get<0>(GetInput()); | ||
| 73 | auto &b = std::get<1>(GetInput()); | ||
| 74 | auto &c = GetOutput(); | ||
| 75 | |||
| 76 | 112 | c.rows = a.rows; | |
| 77 | 112 | c.cols = b.cols; | |
| 78 | 112 | c.col_ptr.resize(c.cols + 1); | |
| 79 | |||
| 80 | 112 | std::vector<std::vector<Complex>> local_values(c.cols); | |
| 81 |
1/2✓ Branch 1 taken 112 times.
✗ Branch 2 not taken.
|
112 | std::vector<std::vector<int>> local_rows(c.cols); |
| 82 | |||
| 83 | 112 | int num_threads = static_cast<int>(std::thread::hardware_concurrency()); | |
| 84 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 112 times.
|
112 | if (num_threads == 0) { |
| 85 | num_threads = 1; | ||
| 86 | } | ||
| 87 | |||
| 88 | 112 | std::vector<std::thread> threads; | |
| 89 | |||
| 90 | 328 | auto worker = [&](int start, int end) { | |
| 91 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 328 times.
|
656 | for (int j = start; j < end; j++) { |
| 92 | 328 | ProcessColumn(j, a, b, local_values[j], local_rows[j]); | |
| 93 | } | ||
| 94 | 440 | }; | |
| 95 | |||
| 96 | 112 | int cols_per_thread = (b.cols + num_threads - 1) / num_threads; | |
| 97 | |||
| 98 |
2/2✓ Branch 0 taken 408 times.
✓ Branch 1 taken 32 times.
|
440 | for (int thread_idx = 0; thread_idx < num_threads; thread_idx++) { |
| 99 | 408 | int start = thread_idx * cols_per_thread; | |
| 100 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 80 times.
|
408 | int end = std::min(start + cols_per_thread, b.cols); |
| 101 | |||
| 102 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 80 times.
|
408 | if (start >= end) { |
| 103 | break; | ||
| 104 | } | ||
| 105 | |||
| 106 |
1/2✓ Branch 1 taken 328 times.
✗ Branch 2 not taken.
|
328 | threads.emplace_back(worker, start, end); |
| 107 | } | ||
| 108 | |||
| 109 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 112 times.
|
440 | for (auto &th : threads) { |
| 110 |
1/2✓ Branch 1 taken 328 times.
✗ Branch 2 not taken.
|
328 | th.join(); |
| 111 | } | ||
| 112 | |||
| 113 | int nnz = 0; | ||
| 114 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 112 times.
|
440 | for (int j = 0; j < c.cols; j++) { |
| 115 | 328 | c.col_ptr[j] = nnz; | |
| 116 | 328 | nnz += static_cast<int>(local_values[j].size()); | |
| 117 | } | ||
| 118 |
1/2✓ Branch 1 taken 112 times.
✗ Branch 2 not taken.
|
112 | c.col_ptr[c.cols] = nnz; |
| 119 | |||
| 120 |
1/2✓ Branch 1 taken 112 times.
✗ Branch 2 not taken.
|
112 | c.values.reserve(nnz); |
| 121 |
1/2✓ Branch 1 taken 112 times.
✗ Branch 2 not taken.
|
112 | c.row_ind.reserve(nnz); |
| 122 | |||
| 123 |
2/2✓ Branch 0 taken 328 times.
✓ Branch 1 taken 112 times.
|
440 | for (int j = 0; j < c.cols; j++) { |
| 124 |
2/4✓ Branch 1 taken 328 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 328 times.
✗ Branch 5 not taken.
|
328 | c.values.insert(c.values.end(), local_values[j].begin(), local_values[j].end()); |
| 125 | 328 | c.row_ind.insert(c.row_ind.end(), local_rows[j].begin(), local_rows[j].end()); | |
| 126 | } | ||
| 127 | |||
| 128 | 112 | return true; | |
| 129 | 112 | } | |
| 130 | |||
| 131 | 112 | bool GoriachevaKMultSparseComplexMatrixCcsSTL::PostProcessingImpl() { | |
| 132 | 112 | return true; | |
| 133 | } | ||
| 134 | |||
| 135 | } // namespace goriacheva_k_mult_sparse_complex_matrix_ccs | ||
| 136 |