| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "../../tbb/include/ops_tbb.hpp" | ||
| 2 | |||
| 3 | #include <tbb/tbb.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <complex> | ||
| 7 | #include <cstddef> | ||
| 8 | #include <utility> | ||
| 9 | #include <vector> | ||
| 10 | |||
| 11 | #include "oneapi/tbb/blocked_range.h" | ||
| 12 | #include "oneapi/tbb/parallel_for.h" | ||
| 13 | #include "shvetsova_k_mult_matrix_complex_col/common/include/common.hpp" | ||
| 14 | |||
| 15 | namespace shvetsova_k_mult_matrix_complex_col { | ||
| 16 | |||
| 17 | struct SparseColumn { | ||
| 18 | std::vector<int> rows; | ||
| 19 | std::vector<std::complex<double>> vals; | ||
| 20 | }; | ||
| 21 | |||
| 22 |
1/2✓ Branch 2 taken 20 times.
✗ Branch 3 not taken.
|
20 | ShvetsovaKMultMatrixComplexTBB::ShvetsovaKMultMatrixComplexTBB(const InType &in) { |
| 23 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 24 | GetInput() = in; | ||
| 25 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | GetOutput() = MatrixCCS(0, 0, std::vector<int>{0}, std::vector<int>{}, std::vector<std::complex<double>>{}); |
| 26 | 20 | } | |
| 27 | |||
| 28 | 20 | bool ShvetsovaKMultMatrixComplexTBB::ValidationImpl() { | |
| 29 | 20 | return true; | |
| 30 | } | ||
| 31 | |||
| 32 | 20 | bool ShvetsovaKMultMatrixComplexTBB::PreProcessingImpl() { | |
| 33 | const auto &matrix_a = std::get<0>(GetInput()); | ||
| 34 | const auto &matrix_b = std::get<1>(GetInput()); | ||
| 35 | |||
| 36 | auto &matrix_c = GetOutput(); | ||
| 37 | 20 | matrix_c.rows = matrix_a.rows; | |
| 38 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
|
20 | matrix_c.cols = matrix_b.cols; |
| 39 | matrix_c.row_ind.clear(); | ||
| 40 | matrix_c.values.clear(); | ||
| 41 | matrix_c.col_ptr.clear(); | ||
| 42 | 20 | matrix_c.col_ptr.push_back(0); | |
| 43 | 20 | return true; | |
| 44 | } | ||
| 45 | |||
| 46 | 20 | bool ShvetsovaKMultMatrixComplexTBB::RunImpl() { | |
| 47 | const MatrixCCS &matrix_a = std::get<0>(GetInput()); | ||
| 48 | const MatrixCCS &matrix_b = std::get<1>(GetInput()); | ||
| 49 | |||
| 50 | auto &matrix_c = GetOutput(); | ||
| 51 | |||
| 52 | 20 | std::vector<SparseColumn> columns_c(matrix_b.cols); | |
| 53 | |||
| 54 |
1/2✓ Branch 0 taken 36 times.
✗ Branch 1 not taken.
|
36 | auto compute_column = [&](int i, std::vector<std::complex<double>> &column_c) { |
| 55 | std::ranges::fill(column_c, std::complex<double>{0.0, 0.0}); | ||
| 56 | |||
| 57 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 36 times.
|
76 | for (int j = matrix_b.col_ptr[i]; j < matrix_b.col_ptr[i + 1]; j++) { |
| 58 | 40 | int tmp_ind = matrix_b.row_ind[j]; | |
| 59 | 40 | std::complex<double> tmp_val = matrix_b.values[j]; | |
| 60 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 40 times.
|
80 | for (int ind = matrix_a.col_ptr[tmp_ind]; ind < matrix_a.col_ptr[tmp_ind + 1]; ind++) { |
| 61 | 40 | int row = matrix_a.row_ind[ind]; | |
| 62 | 40 | std::complex<double> val_a = matrix_a.values[ind]; | |
| 63 | 40 | column_c[row] += tmp_val * val_a; | |
| 64 | } | ||
| 65 | } | ||
| 66 |
1/2✓ Branch 0 taken 104 times.
✗ Branch 1 not taken.
|
104 | for (int index = 0; std::cmp_less(index, column_c.size()); ++index) { |
| 67 |
3/4✗ Branch 0 not taken.
✓ Branch 1 taken 32 times.
✓ Branch 2 taken 32 times.
✓ Branch 3 taken 36 times.
|
68 | if (column_c[index].real() != 0.0 || column_c[index].imag() != 0.0) { |
| 68 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 36 times.
|
36 | columns_c[i].rows.push_back(index); |
| 69 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 36 times.
|
36 | columns_c[i].vals.push_back(column_c[index]); |
| 70 | } | ||
| 71 | } | ||
| 72 | 56 | }; | |
| 73 | |||
| 74 | size_t grain_size = 100; | ||
| 75 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
40 | tbb::parallel_for(tbb::blocked_range<int>(0, matrix_b.cols, grain_size), [&](const tbb::blocked_range<int> &r) { |
| 76 | 20 | std::vector<std::complex<double>> column_c_local(matrix_a.rows, {0.0, 0.0}); | |
| 77 |
2/2✓ Branch 0 taken 36 times.
✓ Branch 1 taken 20 times.
|
56 | for (int i = r.begin(); i != r.end(); ++i) { |
| 78 |
1/2✓ Branch 1 taken 36 times.
✗ Branch 2 not taken.
|
36 | compute_column(i, column_c_local); |
| 79 | } | ||
| 80 | 20 | }, tbb::static_partitioner()); | |
| 81 | |||
| 82 |
2/2✓ Branch 0 taken 36 times.
✓ Branch 1 taken 20 times.
|
56 | for (int i = 0; i < matrix_b.cols; i++) { |
| 83 |
1/2✓ Branch 1 taken 36 times.
✗ Branch 2 not taken.
|
36 | matrix_c.row_ind.insert(matrix_c.row_ind.end(), columns_c[i].rows.begin(), columns_c[i].rows.end()); |
| 84 |
1/2✓ Branch 1 taken 36 times.
✗ Branch 2 not taken.
|
36 | matrix_c.values.insert(matrix_c.values.end(), columns_c[i].vals.begin(), columns_c[i].vals.end()); |
| 85 |
1/2✓ Branch 1 taken 36 times.
✗ Branch 2 not taken.
|
36 | matrix_c.col_ptr.push_back(static_cast<int>(matrix_c.row_ind.size())); |
| 86 | } | ||
| 87 | |||
| 88 | 20 | return true; | |
| 89 | 20 | } | |
| 90 | |||
| 91 | 20 | bool ShvetsovaKMultMatrixComplexTBB::PostProcessingImpl() { | |
| 92 | 20 | return true; | |
| 93 | } | ||
| 94 | |||
| 95 | } // namespace shvetsova_k_mult_matrix_complex_col | ||
| 96 |