| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "romanov_a_gauss_block/tbb/include/ops_tbb.hpp" | ||
| 2 | |||
| 3 | #include <oneapi/tbb/parallel_for.h> | ||
| 4 | #include <tbb/tbb.h> | ||
| 5 | |||
| 6 | #include <algorithm> | ||
| 7 | #include <array> | ||
| 8 | #include <cstddef> | ||
| 9 | #include <cstdint> | ||
| 10 | #include <tuple> | ||
| 11 | #include <vector> | ||
| 12 | |||
| 13 | #include "romanov_a_gauss_block/common/include/common.hpp" | ||
| 14 | |||
| 15 | namespace romanov_a_gauss_block { | ||
| 16 | |||
| 17 |
1/2✓ Branch 1 taken 32 times.
✗ Branch 2 not taken.
|
32 | RomanovAGaussBlockTBB::RomanovAGaussBlockTBB(const InType &in) { |
| 18 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 19 | GetInput() = in; | ||
| 20 | 32 | GetOutput() = std::vector<uint8_t>(); | |
| 21 | 32 | } | |
| 22 | |||
| 23 | 32 | bool RomanovAGaussBlockTBB::ValidationImpl() { | |
| 24 | 32 | return std::get<0>(GetInput()) * std::get<1>(GetInput()) * 3 == static_cast<int>(std::get<2>(GetInput()).size()); | |
| 25 | } | ||
| 26 | |||
| 27 | 32 | bool RomanovAGaussBlockTBB::PreProcessingImpl() { | |
| 28 | 32 | return true; | |
| 29 | } | ||
| 30 | |||
| 31 | namespace { | ||
| 32 | |||
| 33 | constexpr int kBlockSize = 32; | ||
| 34 | |||
| 35 | 120336 | int ApplyKernel(const std::vector<uint8_t> &img, int row, int col, int channel, int width, int height, | |
| 36 | const std::array<std::array<int, 3>, 3> &kernel) { | ||
| 37 | int sum = 0; | ||
| 38 |
2/2✓ Branch 0 taken 361008 times.
✓ Branch 1 taken 120336 times.
|
481344 | for (size_t kr = 0; kr < 3; ++kr) { |
| 39 |
2/2✓ Branch 0 taken 1083024 times.
✓ Branch 1 taken 361008 times.
|
1444032 | for (size_t kc = 0; kc < 3; ++kc) { |
| 40 | 1083024 | int nr = row + static_cast<int>(kr) - 1; | |
| 41 | 1083024 | int nc = col + static_cast<int>(kc) - 1; | |
| 42 |
4/4✓ Branch 0 taken 1074744 times.
✓ Branch 1 taken 8280 times.
✓ Branch 2 taken 1066992 times.
✓ Branch 3 taken 7752 times.
|
1083024 | if (nr >= 0 && nr < height && nc >= 0 && nc < width) { |
| 43 | 1066992 | size_t idx = (((static_cast<size_t>(nr) * width) + nc) * 3) + channel; | |
| 44 | 1066992 | sum += (static_cast<int>(img[idx]) * kernel.at(kr).at(kc)); | |
| 45 | } | ||
| 46 | } | ||
| 47 | } | ||
| 48 | 120336 | return sum; | |
| 49 | } | ||
| 50 | |||
| 51 | 36 | void ProcessFullBlock(const std::vector<uint8_t> &initial_picture, std::vector<uint8_t> &result_picture, int width, | |
| 52 | int height, int start_row, int start_col) { | ||
| 53 | static constexpr std::array<std::array<int, 3>, 3> kKernel = {{{1, 2, 1}, {2, 4, 2}, {1, 2, 1}}}; | ||
| 54 | |||
| 55 |
2/2✓ Branch 0 taken 1152 times.
✓ Branch 1 taken 36 times.
|
1188 | for (int row = start_row; row < start_row + kBlockSize; ++row) { |
| 56 |
2/2✓ Branch 0 taken 36864 times.
✓ Branch 1 taken 1152 times.
|
38016 | for (int col = start_col; col < start_col + kBlockSize; ++col) { |
| 57 |
2/2✓ Branch 0 taken 110592 times.
✓ Branch 1 taken 36864 times.
|
147456 | for (int channel = 0; channel < 3; ++channel) { |
| 58 | 110592 | int sum = ApplyKernel(initial_picture, row, col, channel, width, height, kKernel); | |
| 59 |
1/2✓ Branch 0 taken 110592 times.
✗ Branch 1 not taken.
|
110592 | int result_value = (sum + 8) / 16; |
| 60 | result_value = std::clamp(result_value, 0, 255); | ||
| 61 | 110592 | auto idx = ((static_cast<size_t>(row) * width + col) * 3) + channel; | |
| 62 | 110592 | result_picture[idx] = static_cast<uint8_t>(result_value); | |
| 63 | } | ||
| 64 | } | ||
| 65 | } | ||
| 66 | 36 | } | |
| 67 | |||
| 68 | 56 | void ProcessPartBlock(const std::vector<uint8_t> &initial_picture, std::vector<uint8_t> &result_picture, int width, | |
| 69 | int height, int start_row, int start_col) { | ||
| 70 | static constexpr std::array<std::array<int, 3>, 3> kKernel = {{{1, 2, 1}, {2, 4, 2}, {1, 2, 1}}}; | ||
| 71 | |||
| 72 | 56 | const int end_row = std::min(height, start_row + kBlockSize); | |
| 73 | 56 | const int end_col = std::min(width, start_col + kBlockSize); | |
| 74 | |||
| 75 |
2/2✓ Branch 0 taken 500 times.
✓ Branch 1 taken 56 times.
|
556 | for (int row = start_row; row < end_row; ++row) { |
| 76 |
2/2✓ Branch 0 taken 3248 times.
✓ Branch 1 taken 500 times.
|
3748 | for (int col = start_col; col < end_col; ++col) { |
| 77 |
2/2✓ Branch 0 taken 9744 times.
✓ Branch 1 taken 3248 times.
|
12992 | for (int channel = 0; channel < 3; ++channel) { |
| 78 | 9744 | int sum = ApplyKernel(initial_picture, row, col, channel, width, height, kKernel); | |
| 79 |
1/2✓ Branch 0 taken 9744 times.
✗ Branch 1 not taken.
|
9744 | int result_value = (sum + 8) / 16; |
| 80 | result_value = std::clamp(result_value, 0, 255); | ||
| 81 | 9744 | auto idx = ((static_cast<size_t>(row) * width + col) * 3) + channel; | |
| 82 | 9744 | result_picture[idx] = static_cast<uint8_t>(result_value); | |
| 83 | } | ||
| 84 | } | ||
| 85 | } | ||
| 86 | 56 | } | |
| 87 | |||
| 88 | } // namespace | ||
| 89 | |||
| 90 | 32 | bool RomanovAGaussBlockTBB::RunImpl() { | |
| 91 | 32 | const int width = std::get<0>(GetInput()); | |
| 92 | 32 | const int height = std::get<1>(GetInput()); | |
| 93 | |||
| 94 | const std::vector<uint8_t> &initial_picture = std::get<2>(GetInput()); | ||
| 95 | 32 | std::vector<uint8_t> result_picture(static_cast<size_t>(height * width * 3)); | |
| 96 | |||
| 97 | 32 | const int num_row_blocks = height / kBlockSize; | |
| 98 | 32 | const int num_col_blocks = width / kBlockSize; | |
| 99 | |||
| 100 |
1/2✓ Branch 1 taken 32 times.
✗ Branch 2 not taken.
|
44 | tbb::parallel_for(tbb::blocked_range<int>(0, num_row_blocks, 1), [&](const tbb::blocked_range<int> &r) { |
| 101 |
2/2✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
|
24 | for (int bi = r.begin(); bi != r.end(); ++bi) { |
| 102 |
2/2✓ Branch 0 taken 36 times.
✓ Branch 1 taken 12 times.
|
48 | for (int bj = 0; bj < num_col_blocks; ++bj) { |
| 103 | 36 | ProcessFullBlock(initial_picture, result_picture, width, height, bi * kBlockSize, bj * kBlockSize); | |
| 104 | } | ||
| 105 | } | ||
| 106 | 12 | }); | |
| 107 | |||
| 108 |
1/2✓ Branch 1 taken 32 times.
✗ Branch 2 not taken.
|
44 | tbb::parallel_for(tbb::blocked_range<int>(0, num_row_blocks, 1), [&](const tbb::blocked_range<int> &r) { |
| 109 |
2/2✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
|
24 | for (int bi = r.begin(); bi != r.end(); ++bi) { |
| 110 | 12 | ProcessPartBlock(initial_picture, result_picture, width, height, bi * kBlockSize, width - (width % kBlockSize)); | |
| 111 | } | ||
| 112 | 12 | }); | |
| 113 | |||
| 114 |
1/4✓ Branch 1 taken 32 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
44 | tbb::parallel_for(tbb::blocked_range<int>(0, num_col_blocks, 1), [&](const tbb::blocked_range<int> &r) { |
| 115 |
2/2✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
|
24 | for (int bj = r.begin(); bj != r.end(); ++bj) { |
| 116 | 12 | ProcessPartBlock(initial_picture, result_picture, width, height, height - (height % kBlockSize), bj * kBlockSize); | |
| 117 | } | ||
| 118 | 12 | }); | |
| 119 | |||
| 120 | 32 | ProcessPartBlock(initial_picture, result_picture, width, height, height - (height % kBlockSize), | |
| 121 | 32 | width - (width % kBlockSize)); | |
| 122 | |||
| 123 |
1/2✓ Branch 1 taken 32 times.
✗ Branch 2 not taken.
|
32 | GetOutput() = result_picture; |
| 124 | 32 | return true; | |
| 125 | } | ||
| 126 | |||
| 127 | 32 | bool RomanovAGaussBlockTBB::PostProcessingImpl() { | |
| 128 | 32 | return true; | |
| 129 | } | ||
| 130 | |||
| 131 | } // namespace romanov_a_gauss_block | ||
| 132 |