| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "makoveeva_matmul_double/seq/include/ops_seq.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <cmath> | ||
| 5 | #include <cstddef> | ||
| 6 | #include <vector> | ||
| 7 | |||
| 8 | #include "makoveeva_matmul_double/common/include/common.hpp" | ||
| 9 | |||
| 10 | namespace makoveeva_matmul_double_seq { | ||
| 11 | namespace { | ||
| 12 | |||
| 13 | // Вспомогательная функция для выбора размера блока | ||
| 14 | 64 | int ChooseBlockSize(int n) { | |
| 15 | 64 | int block_size = static_cast<int>(std::sqrt(static_cast<double>(n))); | |
| 16 | 64 | block_size = std::max(1, block_size); | |
| 17 | |||
| 18 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 64 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
64 | while ((n % block_size != 0) && (block_size > 1)) { |
| 19 | ✗ | --block_size; | |
| 20 | } | ||
| 21 | |||
| 22 | 64 | return block_size; | |
| 23 | } | ||
| 24 | |||
| 25 | // Вспомогательная функция для умножения блоков | ||
| 26 | 1808 | void MultiplyBlocks(const std::vector<double> &a, const std::vector<double> &b, std::vector<double> &c, int n, | |
| 27 | int row_start, int row_end, int col_start, int col_end, int k_start, int k_end) { | ||
| 28 | 1808 | const auto n_size = static_cast<size_t>(n); | |
| 29 | |||
| 30 |
2/2✓ Branch 0 taken 4056 times.
✓ Branch 1 taken 1808 times.
|
5864 | for (int row = row_start; row < row_end; ++row) { |
| 31 | 4056 | const auto row_idx = static_cast<size_t>(row); | |
| 32 | 4056 | const auto row_offset = row_idx * n_size; | |
| 33 | |||
| 34 |
2/2✓ Branch 0 taken 10008 times.
✓ Branch 1 taken 4056 times.
|
14064 | for (int col = col_start; col < col_end; ++col) { |
| 35 | 10008 | const auto col_idx = static_cast<size_t>(col); | |
| 36 | double sum = 0.0; | ||
| 37 | |||
| 38 |
2/2✓ Branch 0 taken 26280 times.
✓ Branch 1 taken 10008 times.
|
36288 | for (int k = k_start; k < k_end; ++k) { |
| 39 | 26280 | const auto k_idx = static_cast<size_t>(k); | |
| 40 | 26280 | const auto a_idx = row_offset + k_idx; | |
| 41 | 26280 | const auto b_idx = (k_idx * n_size) + col_idx; | |
| 42 | 26280 | sum += a[a_idx] * b[b_idx]; | |
| 43 | } | ||
| 44 | |||
| 45 | 10008 | const auto c_idx = row_offset + col_idx; | |
| 46 | 10008 | c[c_idx] += sum; | |
| 47 | } | ||
| 48 | } | ||
| 49 | 1808 | } | |
| 50 | |||
| 51 | } // namespace | ||
| 52 | |||
| 53 |
1/2✓ Branch 1 taken 64 times.
✗ Branch 2 not taken.
|
64 | MatmulDoubleSeqTask::MatmulDoubleSeqTask(const InType &in) |
| 54 |
4/10✓ Branch 1 taken 64 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 64 times.
✗ Branch 5 not taken.
✓ Branch 7 taken 64 times.
✗ Branch 8 not taken.
✓ Branch 10 taken 64 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
|
64 | : n_(std::get<0>(in)), A_(std::get<1>(in)), B_(std::get<2>(in)), C_(n_ * n_, 0.0) { |
| 55 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 56 |
1/2✓ Branch 1 taken 64 times.
✗ Branch 2 not taken.
|
64 | GetOutput() = C_; |
| 57 | 64 | } | |
| 58 | |||
| 59 | 64 | bool MatmulDoubleSeqTask::ValidationImpl() { | |
| 60 | 64 | const auto expected_size = n_ * n_; | |
| 61 |
3/6✓ Branch 0 taken 64 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 64 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 64 times.
|
64 | const auto is_valid = (n_ > 0) && (A_.size() == expected_size) && (B_.size() == expected_size); |
| 62 | 64 | return is_valid; | |
| 63 | } | ||
| 64 | |||
| 65 | 64 | bool MatmulDoubleSeqTask::PreProcessingImpl() { | |
| 66 | 64 | return true; | |
| 67 | } | ||
| 68 | |||
| 69 | 64 | bool MatmulDoubleSeqTask::RunImpl() { | |
| 70 |
1/2✓ Branch 0 taken 64 times.
✗ Branch 1 not taken.
|
64 | if (n_ <= 0) { |
| 71 | return false; | ||
| 72 | } | ||
| 73 | |||
| 74 | 64 | const auto n_int = static_cast<int>(n_); | |
| 75 | 64 | auto block_size = ChooseBlockSize(n_int); | |
| 76 | |||
| 77 | if (block_size == 1 && n_int > 1 && (n_int % block_size != 0)) { | ||
| 78 | block_size = n_int; | ||
| 79 | } | ||
| 80 | |||
| 81 | 64 | const auto total_size = n_ * n_; | |
| 82 | 64 | C_.assign(total_size, 0.0); | |
| 83 | |||
| 84 | 64 | const auto grid_size = n_int / block_size; | |
| 85 | |||
| 86 |
2/2✓ Branch 0 taken 176 times.
✓ Branch 1 taken 64 times.
|
240 | for (int stage = 0; stage < grid_size; ++stage) { |
| 87 |
2/2✓ Branch 0 taken 544 times.
✓ Branch 1 taken 176 times.
|
720 | for (int i_block = 0; i_block < grid_size; ++i_block) { |
| 88 |
2/2✓ Branch 0 taken 1808 times.
✓ Branch 1 taken 544 times.
|
2352 | for (int j_block = 0; j_block < grid_size; ++j_block) { |
| 89 | 1808 | const auto root_block = (i_block + stage) % grid_size; | |
| 90 | |||
| 91 | 1808 | const auto row_start = i_block * block_size; | |
| 92 | 1808 | const auto row_end = row_start + block_size; | |
| 93 | 1808 | const auto col_start = j_block * block_size; | |
| 94 | 1808 | const auto col_end = col_start + block_size; | |
| 95 | 1808 | const auto k_start = root_block * block_size; | |
| 96 | 1808 | const auto k_end = k_start + block_size; | |
| 97 | |||
| 98 | 1808 | MultiplyBlocks(A_, B_, C_, n_int, row_start, row_end, col_start, col_end, k_start, k_end); | |
| 99 | } | ||
| 100 | } | ||
| 101 | } | ||
| 102 | |||
| 103 | 64 | GetOutput() = C_; | |
| 104 | 64 | return true; | |
| 105 | } | ||
| 106 | |||
| 107 | 64 | bool MatmulDoubleSeqTask::PostProcessingImpl() { | |
| 108 | 64 | return true; | |
| 109 | } | ||
| 110 | |||
| 111 | } // namespace makoveeva_matmul_double_seq | ||
| 112 |