| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #pragma once | ||
| 2 | #include <omp.h> | ||
| 3 | |||
| 4 | #include <cmath> | ||
| 5 | #include <cstddef> | ||
| 6 | #include <vector> | ||
| 7 | |||
| 8 | #include "guseva_crs/common/include/common.hpp" | ||
| 9 | #include "guseva_crs/common/include/multiplier.hpp" | ||
| 10 | |||
| 11 | namespace guseva_crs { | ||
| 12 | |||
| 13 | 24 | class MultiplierOmp : public Multiplier { | |
| 14 | static void PerformCalculation(std::size_t k, std::size_t ind3, std::size_t ind4, const CRS &a, const CRS &bt, | ||
| 15 | double &sum, std::vector<int> &temp) { | ||
| 16 | for (k = ind3; k < ind4; k++) { | ||
| 17 | std::size_t bcol = bt.cols[k]; | ||
| 18 | int aind = temp[bcol]; | ||
| 19 | if (aind != -1) { | ||
| 20 | sum += a.values[aind] * bt.values[k]; | ||
| 21 | } | ||
| 22 | } | ||
| 23 | } | ||
| 24 | |||
| 25 | public: | ||
| 26 | 24 | [[nodiscard]] CRS Multiply(const CRS &a, const CRS &b) const override { | |
| 27 | 24 | std::size_t n = a.nrows; | |
| 28 | std::size_t i = 0; | ||
| 29 | |||
| 30 | 24 | auto bt = this->Transpose(b); | |
| 31 | |||
| 32 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | std::vector<std::vector<std::size_t>> columns(n); |
| 33 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | std::vector<std::vector<double>> values(n); |
| 34 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | std::vector<std::size_t> row_index(n + 1, 0); |
| 35 | |||
| 36 | 24 | #pragma omp parallel shared(n, a, bt, columns, values, row_index) private(i) default(none) | |
| 37 | { | ||
| 38 | std::size_t j = 0; | ||
| 39 | std::size_t k = 0; | ||
| 40 | std::vector<int> temp(n); | ||
| 41 | #pragma omp for private(j, k) | ||
| 42 | for (i = 0; i < n; i++) { | ||
| 43 | for (int &l : temp) { | ||
| 44 | l = -1; | ||
| 45 | } | ||
| 46 | std::size_t ind1 = a.row_ptrs[i]; | ||
| 47 | std::size_t ind2 = a.row_ptrs[i + 1]; | ||
| 48 | for (j = ind1; j < ind2; j++) { | ||
| 49 | std::size_t col = a.cols[j]; | ||
| 50 | temp[col] = static_cast<int>(j); | ||
| 51 | } | ||
| 52 | for (j = 0; j < n; j++) { | ||
| 53 | double sum = 0; | ||
| 54 | std::size_t ind3 = bt.row_ptrs[j]; | ||
| 55 | std::size_t ind4 = bt.row_ptrs[j + 1]; | ||
| 56 | k = 0; | ||
| 57 | PerformCalculation(k, ind3, ind4, a, bt, sum, temp); | ||
| 58 | if (std::fabs(sum) > kZERO) { | ||
| 59 | columns[i].push_back(j); | ||
| 60 | values[i].push_back(sum); | ||
| 61 | row_index[i]++; | ||
| 62 | } | ||
| 63 | } | ||
| 64 | } | ||
| 65 | temp.clear(); | ||
| 66 | } | ||
| 67 | |||
| 68 | std::size_t nz = 0; | ||
| 69 |
2/2✓ Branch 0 taken 388 times.
✓ Branch 1 taken 24 times.
|
412 | for (i = 0; i < n; i++) { |
| 70 | 388 | std::size_t tmp = row_index[i]; | |
| 71 | 388 | row_index[i] = nz; | |
| 72 | 388 | nz += tmp; | |
| 73 | } | ||
| 74 | |||
| 75 | 24 | CRS result; | |
| 76 |
2/2✓ Branch 0 taken 388 times.
✓ Branch 1 taken 24 times.
|
412 | for (i = 0; i < n; i++) { |
| 77 |
1/2✓ Branch 1 taken 388 times.
✗ Branch 2 not taken.
|
388 | result.cols.insert(result.cols.end(), columns[i].begin(), columns[i].end()); |
| 78 |
1/2✓ Branch 1 taken 388 times.
✗ Branch 2 not taken.
|
388 | result.values.insert(result.values.end(), values[i].begin(), values[i].end()); |
| 79 | } | ||
| 80 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | result.row_ptrs = row_index; |
| 81 | 24 | result.nz = result.values.size(); | |
| 82 | 24 | result.ncols = n; | |
| 83 |
1/2✓ Branch 0 taken 24 times.
✗ Branch 1 not taken.
|
24 | result.nrows = n; |
| 84 | |||
| 85 | 24 | return result; | |
| 86 | 24 | } | |
| 87 | }; | ||
| 88 | } // namespace guseva_crs | ||
| 89 |