| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "kotelnikova_a_double_matr_mult/tbb/include/ops_tbb.hpp" | ||
| 2 | |||
| 3 | #include <tbb/blocked_range.h> | ||
| 4 | #include <tbb/parallel_for.h> | ||
| 5 | #include <tbb/parallel_scan.h> | ||
| 6 | #include <tbb/tbb.h> | ||
| 7 | |||
| 8 | #include <cmath> | ||
| 9 | #include <cstddef> | ||
| 10 | #include <vector> | ||
| 11 | |||
| 12 | #include "kotelnikova_a_double_matr_mult/common/include/common.hpp" | ||
| 13 | |||
| 14 | namespace kotelnikova_a_double_matr_mult { | ||
| 15 | |||
| 16 |
1/2✓ Branch 2 taken 20 times.
✗ Branch 3 not taken.
|
20 | KotelnikovaATaskTBB::KotelnikovaATaskTBB(const InType &in) { |
| 17 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 18 | GetInput() = in; | ||
| 19 | 20 | GetOutput() = SparseMatrixCCS(); | |
| 20 | 20 | } | |
| 21 | |||
| 22 | 40 | bool KotelnikovaATaskTBB::IsMatrixValid(const SparseMatrixCCS &matrix) { | |
| 23 |
2/4✓ Branch 0 taken 40 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 40 times.
|
40 | if (matrix.rows < 0 || matrix.cols < 0) { |
| 24 | return false; | ||
| 25 | } | ||
| 26 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
|
40 | if (matrix.col_ptrs.size() != static_cast<size_t>(matrix.cols) + 1) { |
| 27 | return false; | ||
| 28 | } | ||
| 29 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
|
40 | if (matrix.values.size() != matrix.row_indices.size()) { |
| 30 | return false; | ||
| 31 | } | ||
| 32 | |||
| 33 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 40 times.
|
40 | if (matrix.col_ptrs.empty() || matrix.col_ptrs[0] != 0) { |
| 34 | return false; | ||
| 35 | } | ||
| 36 | |||
| 37 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
|
40 | const int total_elements = static_cast<int>(matrix.values.size()); |
| 38 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 40 times.
|
40 | if (matrix.col_ptrs[matrix.cols] != total_elements) { |
| 39 | return false; | ||
| 40 | } | ||
| 41 | |||
| 42 |
2/2✓ Branch 0 taken 116 times.
✓ Branch 1 taken 40 times.
|
156 | for (size_t i = 0; i < matrix.col_ptrs.size() - 1; ++i) { |
| 43 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 116 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 116 times.
|
116 | if (matrix.col_ptrs[i] > matrix.col_ptrs[i + 1] || matrix.col_ptrs[i] < 0) { |
| 44 | return false; | ||
| 45 | } | ||
| 46 | } | ||
| 47 | |||
| 48 |
2/2✓ Branch 0 taken 144 times.
✓ Branch 1 taken 40 times.
|
184 | for (size_t i = 0; i < matrix.row_indices.size(); ++i) { |
| 49 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 144 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 144 times.
|
144 | if (matrix.row_indices[i] < 0 || matrix.row_indices[i] >= matrix.rows) { |
| 50 | return false; | ||
| 51 | } | ||
| 52 | } | ||
| 53 | |||
| 54 | return true; | ||
| 55 | } | ||
| 56 | |||
| 57 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | bool KotelnikovaATaskTBB::ValidationImpl() { |
| 58 | const auto &[a, b] = GetInput(); | ||
| 59 | |||
| 60 |
2/4✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 20 times.
✗ Branch 3 not taken.
|
20 | if (!IsMatrixValid(a) || !IsMatrixValid(b)) { |
| 61 | return false; | ||
| 62 | } | ||
| 63 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
|
20 | if (a.cols != b.rows) { |
| 64 | ✗ | return false; | |
| 65 | } | ||
| 66 | |||
| 67 | return true; | ||
| 68 | } | ||
| 69 | |||
| 70 | 20 | bool KotelnikovaATaskTBB::PreProcessingImpl() { | |
| 71 | const auto &[a, b] = GetInput(); | ||
| 72 | 20 | GetOutput() = SparseMatrixCCS(a.rows, b.cols); | |
| 73 | 20 | return true; | |
| 74 | } | ||
| 75 | |||
| 76 | namespace { | ||
| 77 | 112 | std::vector<double> ComputeColumn(const SparseMatrixCCS &a, const SparseMatrixCCS &b, int col_idx) { | |
| 78 | 112 | std::vector<double> temp(a.rows, 0.0); | |
| 79 | |||
| 80 |
2/2✓ Branch 0 taken 112 times.
✓ Branch 1 taken 112 times.
|
224 | for (int b_idx = b.col_ptrs[col_idx]; b_idx < b.col_ptrs[col_idx + 1]; ++b_idx) { |
| 81 | 112 | const int k = b.row_indices[b_idx]; | |
| 82 | 112 | const double b_val = b.values[b_idx]; | |
| 83 | |||
| 84 |
2/2✓ Branch 0 taken 168 times.
✓ Branch 1 taken 112 times.
|
280 | for (int a_idx = a.col_ptrs[k]; a_idx < a.col_ptrs[k + 1]; ++a_idx) { |
| 85 | 168 | const int i = a.row_indices[a_idx]; | |
| 86 | 168 | temp[i] += a.values[a_idx] * b_val; | |
| 87 | } | ||
| 88 | } | ||
| 89 | |||
| 90 | 112 | return temp; | |
| 91 | } | ||
| 92 | |||
| 93 | int CountNonZero(const std::vector<double> &column, double epsilon) { | ||
| 94 | int count = 0; | ||
| 95 |
4/4✓ Branch 0 taken 84 times.
✓ Branch 1 taken 84 times.
✓ Branch 2 taken 168 times.
✓ Branch 3 taken 56 times.
|
224 | for (double val : column) { |
| 96 |
2/2✓ Branch 0 taken 84 times.
✓ Branch 1 taken 84 times.
|
168 | if (std::abs(val) > epsilon) { |
| 97 | 84 | ++count; | |
| 98 | } | ||
| 99 | } | ||
| 100 | return count; | ||
| 101 | } | ||
| 102 | |||
| 103 | void FillColumn(const std::vector<double> &column, double epsilon, std::vector<int> &row_indices, | ||
| 104 | std::vector<double> &values, int start_pos) { | ||
| 105 | int pos = start_pos; | ||
| 106 |
2/2✓ Branch 0 taken 168 times.
✓ Branch 1 taken 56 times.
|
224 | for (size_t i = 0; i < column.size(); ++i) { |
| 107 |
2/2✓ Branch 0 taken 84 times.
✓ Branch 1 taken 84 times.
|
168 | if (std::abs(column[i]) > epsilon) { |
| 108 | 84 | row_indices[pos] = static_cast<int>(i); | |
| 109 | 84 | values[pos] = column[i]; | |
| 110 | 84 | ++pos; | |
| 111 | } | ||
| 112 | } | ||
| 113 | } | ||
| 114 | |||
| 115 | } // namespace | ||
| 116 | |||
| 117 | // Оптимизированная версия с grain size и partitioner | ||
| 118 | 20 | SparseMatrixCCS KotelnikovaATaskTBB::MultiplyMatrices(const SparseMatrixCCS &a, const SparseMatrixCCS &b) { | |
| 119 | 20 | SparseMatrixCCS result(a.rows, b.cols); | |
| 120 | |||
| 121 | 20 | const double epsilon = 1e-10; | |
| 122 | const int grain_size = 8; // Аналог chunk size в OpenMP | ||
| 123 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | std::vector<int> col_start(b.cols, 0); |
| 124 | |||
| 125 | // Первый проход: подсчет ненулевых элементов с grain size и simple_partitioner | ||
| 126 | 40 | tbb::parallel_for(tbb::blocked_range<int>(0, b.cols, grain_size), | |
| 127 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
40 | [&](const tbb::blocked_range<int> &range) { |
| 128 |
2/2✓ Branch 0 taken 56 times.
✓ Branch 1 taken 20 times.
|
76 | for (int j = range.begin(); j < range.end(); ++j) { |
| 129 | 56 | std::vector<double> column = ComputeColumn(a, b, j); | |
| 130 |
1/2✓ Branch 0 taken 56 times.
✗ Branch 1 not taken.
|
112 | col_start[j] = CountNonZero(column, epsilon); |
| 131 | } | ||
| 132 | 20 | }, | |
| 133 | 20 | tbb::simple_partitioner() // Минимизирует overhead для небольших задач | |
| 134 | ); | ||
| 135 | |||
| 136 | // Префиксная сумма для построения col_ptrs | ||
| 137 |
1/4✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
20 | std::vector<int> col_ptr(b.cols + 1, 0); |
| 138 |
2/2✓ Branch 0 taken 56 times.
✓ Branch 1 taken 20 times.
|
76 | for (int j = 0; j < b.cols; ++j) { |
| 139 | 56 | col_ptr[j + 1] = col_ptr[j] + col_start[j]; | |
| 140 | } | ||
| 141 | |||
| 142 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | const int total_nnz = col_ptr[b.cols]; |
| 143 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | result.values.resize(total_nnz); |
| 144 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | result.row_indices.resize(total_nnz); |
| 145 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | result.col_ptrs = col_ptr; |
| 146 | |||
| 147 | // Второй проход: заполнение данных с grain size и simple_partitioner | ||
| 148 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
40 | tbb::parallel_for(tbb::blocked_range<int>(0, b.cols, grain_size), [&](const tbb::blocked_range<int> &range) { |
| 149 |
2/2✓ Branch 0 taken 56 times.
✓ Branch 1 taken 20 times.
|
76 | for (int j = range.begin(); j < range.end(); ++j) { |
| 150 | 56 | std::vector<double> column = ComputeColumn(a, b, j); | |
| 151 | 56 | FillColumn(column, epsilon, result.row_indices, result.values, col_ptr[j]); | |
| 152 | } | ||
| 153 |
1/4✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
20 | }, tbb::simple_partitioner()); |
| 154 | |||
| 155 | 20 | return result; | |
| 156 | ✗ | } | |
| 157 | |||
| 158 | 20 | bool KotelnikovaATaskTBB::RunImpl() { | |
| 159 | const auto &[a, b] = GetInput(); | ||
| 160 | 20 | GetOutput() = MultiplyMatrices(a, b); | |
| 161 | 20 | return true; | |
| 162 | } | ||
| 163 | |||
| 164 | 20 | bool KotelnikovaATaskTBB::PostProcessingImpl() { | |
| 165 | 20 | return true; | |
| 166 | } | ||
| 167 | |||
| 168 | } // namespace kotelnikova_a_double_matr_mult | ||
| 169 |