| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "alekseev_a_mult_matrix_crs/stl/include/ops_stl.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <cmath> | ||
| 5 | #include <cstddef> | ||
| 6 | #include <thread> | ||
| 7 | #include <utility> | ||
| 8 | #include <vector> | ||
| 9 | |||
| 10 | #include "alekseev_a_mult_matrix_crs/common/include/common.hpp" | ||
| 11 | |||
| 12 | namespace alekseev_a_mult_matrix_crs { | ||
| 13 | |||
| 14 |
1/2✓ Branch 2 taken 48 times.
✗ Branch 3 not taken.
|
48 | AlekseevAMultMatrixCRSSTL::AlekseevAMultMatrixCRSSTL(const InType &in) { |
| 15 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 16 | GetInput() = in; | ||
| 17 | 48 | } | |
| 18 | |||
| 19 | 48 | bool AlekseevAMultMatrixCRSSTL::ValidationImpl() { | |
| 20 | const auto &a = std::get<0>(GetInput()); | ||
| 21 | const auto &b = std::get<1>(GetInput()); | ||
| 22 |
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 a.cols == b.rows && !a.row_ptr.empty() && !b.row_ptr.empty(); |
| 23 | } | ||
| 24 | |||
| 25 | 48 | bool AlekseevAMultMatrixCRSSTL::PreProcessingImpl() { | |
| 26 | 48 | GetOutput() = {}; | |
| 27 | 48 | return true; | |
| 28 | } | ||
| 29 | |||
| 30 | 96 | void AlekseevAMultMatrixCRSSTL::ProcessRow(std::size_t i, const CRSMatrix &a, const CRSMatrix &b, | |
| 31 | std::vector<double> &temp_v, std::vector<std::size_t> &temp_c, | ||
| 32 | std::vector<double> &accum, std::vector<int> &touched_flag, | ||
| 33 | std::vector<std::size_t> &touched_cols) { | ||
| 34 |
2/2✓ Branch 0 taken 104 times.
✓ Branch 1 taken 96 times.
|
200 | for (std::size_t pos_a = a.row_ptr[i]; pos_a < a.row_ptr[i + 1]; ++pos_a) { |
| 35 | 104 | std::size_t k = a.col_indices[pos_a]; | |
| 36 | 104 | double val_a = a.values[pos_a]; | |
| 37 | |||
| 38 |
2/2✓ Branch 0 taken 160 times.
✓ Branch 1 taken 104 times.
|
264 | for (std::size_t pos_b = b.row_ptr[k]; pos_b < b.row_ptr[k + 1]; ++pos_b) { |
| 39 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 120 times.
|
160 | std::size_t j = b.col_indices[pos_b]; |
| 40 | |||
| 41 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 120 times.
|
160 | if (std::cmp_not_equal(touched_flag[j], i)) { |
| 42 |
1/2✓ Branch 0 taken 120 times.
✗ Branch 1 not taken.
|
120 | touched_flag[j] = static_cast<int>(i); |
| 43 | touched_cols.push_back(j); | ||
| 44 | 120 | accum[j] = 0.0; | |
| 45 | } | ||
| 46 | 160 | accum[j] += val_a * b.values[pos_b]; | |
| 47 | } | ||
| 48 | } | ||
| 49 | |||
| 50 | std::ranges::sort(touched_cols); | ||
| 51 | |||
| 52 |
3/4✓ Branch 0 taken 120 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 120 times.
✓ Branch 3 taken 96 times.
|
216 | for (auto col : touched_cols) { |
| 53 |
1/2✓ Branch 0 taken 120 times.
✗ Branch 1 not taken.
|
120 | if (std::abs(accum[col]) > 1e-15) { |
| 54 | temp_v.push_back(accum[col]); | ||
| 55 | temp_c.push_back(col); | ||
| 56 | } | ||
| 57 | } | ||
| 58 | touched_cols.clear(); | ||
| 59 | 96 | } | |
| 60 | |||
| 61 | 48 | bool AlekseevAMultMatrixCRSSTL::RunImpl() { | |
| 62 | const auto &a = std::get<0>(GetInput()); | ||
| 63 | const auto &b = std::get<1>(GetInput()); | ||
| 64 | auto &c = GetOutput(); | ||
| 65 | |||
| 66 | 48 | c.rows = a.rows; | |
| 67 | 48 | c.cols = b.cols; | |
| 68 | 48 | c.row_ptr.assign(c.rows + 1, 0); | |
| 69 | |||
| 70 | 48 | std::vector<std::vector<double>> temp_values(c.rows); | |
| 71 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | std::vector<std::vector<std::size_t>> temp_cols(c.rows); |
| 72 | |||
| 73 | 48 | unsigned int num_threads = std::thread::hardware_concurrency(); | |
| 74 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | if (num_threads == 0) { |
| 75 | num_threads = 2; | ||
| 76 | } | ||
| 77 | |||
| 78 | 48 | std::vector<std::thread> threads; | |
| 79 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | threads.reserve(num_threads); |
| 80 | 48 | std::size_t chunk_size = (c.rows + num_threads - 1) / num_threads; | |
| 81 | |||
| 82 |
1/2✓ Branch 0 taken 144 times.
✗ Branch 1 not taken.
|
144 | for (unsigned int thread_idx = 0; thread_idx < num_threads; ++thread_idx) { |
| 83 | 144 | std::size_t start_row = thread_idx * chunk_size; | |
| 84 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | std::size_t end_row = std::min(start_row + chunk_size, c.rows); |
| 85 | |||
| 86 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | if (start_row >= end_row) { |
| 87 | break; | ||
| 88 | } | ||
| 89 | |||
| 90 |
1/2✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
|
96 | threads.emplace_back([&, start_row, end_row]() { |
| 91 | 96 | std::vector<double> accum(c.cols, 0.0); | |
| 92 |
1/4✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
96 | std::vector<int> touched_flag(c.cols, -1); |
| 93 | 96 | std::vector<std::size_t> touched_cols; | |
| 94 |
1/2✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
|
96 | touched_cols.reserve(c.cols); |
| 95 | |||
| 96 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 96 times.
|
192 | for (std::size_t i = start_row; i < end_row; ++i) { |
| 97 |
1/2✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
|
96 | ProcessRow(i, a, b, temp_values[i], temp_cols[i], accum, touched_flag, touched_cols); |
| 98 | } | ||
| 99 | 96 | }); | |
| 100 | } | ||
| 101 | |||
| 102 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | for (auto &thread : threads) { |
| 103 |
1/2✓ Branch 0 taken 96 times.
✗ Branch 1 not taken.
|
96 | if (thread.joinable()) { |
| 104 |
1/2✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
|
96 | thread.join(); |
| 105 | } | ||
| 106 | } | ||
| 107 | |||
| 108 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | for (std::size_t i = 0; i < c.rows; ++i) { |
| 109 | 96 | c.row_ptr[i + 1] = c.row_ptr[i] + temp_values[i].size(); | |
| 110 | } | ||
| 111 | |||
| 112 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | c.values.reserve(c.row_ptr[c.rows]); |
| 113 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | c.col_indices.reserve(c.row_ptr[c.rows]); |
| 114 | |||
| 115 |
2/2✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
|
144 | for (std::size_t i = 0; i < c.rows; ++i) { |
| 116 |
1/2✓ Branch 1 taken 96 times.
✗ Branch 2 not taken.
|
96 | c.values.insert(c.values.end(), temp_values[i].begin(), temp_values[i].end()); |
| 117 | 96 | c.col_indices.insert(c.col_indices.end(), temp_cols[i].begin(), temp_cols[i].end()); | |
| 118 | } | ||
| 119 | |||
| 120 | 48 | return true; | |
| 121 | 48 | } | |
| 122 | |||
| 123 | 48 | bool AlekseevAMultMatrixCRSSTL::PostProcessingImpl() { | |
| 124 | 48 | return true; | |
| 125 | } | ||
| 126 | |||
| 127 | } // namespace alekseev_a_mult_matrix_crs | ||
| 128 |