| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "vlasova_a_image_smoothing/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <array> | ||
| 7 | #include <cstddef> | ||
| 8 | #include <cstdint> | ||
| 9 | #include <iterator> | ||
| 10 | #include <vector> | ||
| 11 | |||
| 12 | #include "vlasova_a_image_smoothing/common/include/common.hpp" | ||
| 13 | |||
| 14 | namespace vlasova_a_image_smoothing { | ||
| 15 | |||
| 16 | namespace { | ||
| 17 | |||
| 18 | 12288 | std::uint8_t ComputePixelMedian(int col_idx, int row_idx, int overlap_start, | |
| 19 | const std::vector<std::uint8_t> &local_data, int width, int height, int window_size) { | ||
| 20 | 12288 | const int radius = window_size / 2; | |
| 21 | 12288 | std::vector<std::uint8_t> neighbors; | |
| 22 |
1/2✓ Branch 1 taken 12288 times.
✗ Branch 2 not taken.
|
12288 | neighbors.reserve(static_cast<std::size_t>(window_size) * window_size); |
| 23 | |||
| 24 |
2/2✓ Branch 0 taken 36864 times.
✓ Branch 1 taken 12288 times.
|
49152 | for (int dy = -radius; dy <= radius; ++dy) { |
| 25 |
2/2✓ Branch 0 taken 110592 times.
✓ Branch 1 taken 36864 times.
|
147456 | for (int dx = -radius; dx <= radius; ++dx) { |
| 26 | 110592 | const int neighbor_x = col_idx + dx; | |
| 27 | 110592 | const int neighbor_y = row_idx + dy; | |
| 28 | |||
| 29 |
4/4✓ Branch 0 taken 109440 times.
✓ Branch 1 taken 1152 times.
✓ Branch 2 taken 108300 times.
✓ Branch 3 taken 1140 times.
|
110592 | if (neighbor_x >= 0 && neighbor_x < width && neighbor_y >= 0 && neighbor_y < height) { |
| 30 | 108300 | const int local_row = neighbor_y - overlap_start; | |
| 31 |
1/2✓ Branch 0 taken 108300 times.
✗ Branch 1 not taken.
|
108300 | const std::size_t index = (static_cast<std::size_t>(local_row) * width) + neighbor_x; |
| 32 | neighbors.push_back(local_data[index]); | ||
| 33 | } | ||
| 34 | } | ||
| 35 | } | ||
| 36 | |||
| 37 |
1/2✓ Branch 0 taken 12288 times.
✗ Branch 1 not taken.
|
12288 | if (!neighbors.empty()) { |
| 38 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 12288 times.
|
12288 | auto middle = std::next(neighbors.begin(), static_cast<std::ptrdiff_t>(neighbors.size() / 2)); |
| 39 | 12288 | std::nth_element(neighbors.begin(), middle, neighbors.end()); | |
| 40 | 12288 | return *middle; | |
| 41 | } | ||
| 42 | |||
| 43 | ✗ | const int local_row = row_idx - overlap_start; | |
| 44 | ✗ | const std::size_t index = (static_cast<std::size_t>(local_row) * width) + col_idx; | |
| 45 | ✗ | return local_data[index]; | |
| 46 | } | ||
| 47 | |||
| 48 | 3 | void PrepareScatterData(int size, int width, int height, int radius, std::vector<int> &sendcounts, | |
| 49 | std::vector<int> &displs) { | ||
| 50 | 3 | const int base_rows = height / size; | |
| 51 | 3 | const int extra_rows = height % size; | |
| 52 | |||
| 53 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 3 times.
|
9 | for (int proc = 0; proc < size; ++proc) { |
| 54 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int proc_start = (proc * base_rows) + std::min(proc, extra_rows); |
| 55 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int proc_end = proc_start + base_rows + (proc < extra_rows ? 1 : 0); |
| 56 | |||
| 57 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | if (proc_end > proc_start) { |
| 58 | 6 | const int proc_overlap_start = std::max(0, proc_start - radius); | |
| 59 | 6 | const int proc_overlap_end = std::min(height, proc_end + radius); | |
| 60 | 6 | sendcounts[static_cast<std::size_t>(proc)] = (proc_overlap_end - proc_overlap_start) * width; | |
| 61 | 6 | displs[static_cast<std::size_t>(proc)] = proc_overlap_start * width; | |
| 62 | } else { | ||
| 63 | ✗ | sendcounts[static_cast<std::size_t>(proc)] = 0; | |
| 64 | ✗ | displs[static_cast<std::size_t>(proc)] = 0; | |
| 65 | } | ||
| 66 | } | ||
| 67 | 3 | } | |
| 68 | |||
| 69 | 3 | void PrepareGatherData(int size, int width, int height, std::vector<int> &sendcounts, std::vector<int> &displs) { | |
| 70 | 3 | const int base_rows = height / size; | |
| 71 | 3 | const int extra_rows = height % size; | |
| 72 | |||
| 73 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 3 times.
|
9 | for (int proc = 0; proc < size; ++proc) { |
| 74 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int proc_start = (proc * base_rows) + std::min(proc, extra_rows); |
| 75 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int proc_end = proc_start + base_rows + (proc < extra_rows ? 1 : 0); |
| 76 | 6 | sendcounts[static_cast<std::size_t>(proc)] = (proc_end - proc_start) * width; | |
| 77 | 6 | displs[static_cast<std::size_t>(proc)] = proc_start * width; | |
| 78 | } | ||
| 79 | 3 | } | |
| 80 | |||
| 81 | } // namespace | ||
| 82 | |||
| 83 |
1/2✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
|
6 | VlasovaAImageSmoothingMPI::VlasovaAImageSmoothingMPI(const InType &in) { |
| 84 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 85 | GetInput() = in; | ||
| 86 | 6 | } | |
| 87 | |||
| 88 | 6 | bool VlasovaAImageSmoothingMPI::ValidationImpl() { | |
| 89 | 6 | int rank = 0; | |
| 90 | 6 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 91 | |||
| 92 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 93 | const auto &input = GetInput(); | ||
| 94 | |||
| 95 |
2/4✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
✗ Branch 3 not taken.
|
3 | if (input.width <= 0 || input.height <= 0) { |
| 96 | return false; | ||
| 97 | } | ||
| 98 | |||
| 99 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | if (input.data.empty()) { |
| 100 | return false; | ||
| 101 | } | ||
| 102 | |||
| 103 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | const std::size_t expected_size = static_cast<std::size_t>(input.width) * input.height; |
| 104 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | if (input.data.size() != expected_size) { |
| 105 | return false; | ||
| 106 | } | ||
| 107 | } | ||
| 108 | |||
| 109 | 6 | int is_valid = 1; | |
| 110 | 6 | MPI_Bcast(&is_valid, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 111 | |||
| 112 | 6 | return is_valid == 1; | |
| 113 | } | ||
| 114 | |||
| 115 | 6 | bool VlasovaAImageSmoothingMPI::PreProcessingImpl() { | |
| 116 | 6 | int rank = 0; | |
| 117 | 6 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 118 | |||
| 119 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 120 | const auto &input = GetInput(); | ||
| 121 | 3 | width_ = input.width; | |
| 122 | 3 | height_ = input.height; | |
| 123 | } | ||
| 124 | |||
| 125 | 6 | std::array<int, 2> dimensions = {width_, height_}; | |
| 126 | 6 | MPI_Bcast(dimensions.data(), 2, MPI_INT, 0, MPI_COMM_WORLD); | |
| 127 | 6 | MPI_Bcast(&window_size_, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 128 | |||
| 129 | 6 | width_ = dimensions[0]; | |
| 130 | 6 | height_ = dimensions[1]; | |
| 131 | |||
| 132 | 6 | return true; | |
| 133 | } | ||
| 134 | |||
| 135 | 6 | bool VlasovaAImageSmoothingMPI::RunImpl() { | |
| 136 | 6 | int rank = 0; | |
| 137 | 6 | int size = 1; | |
| 138 | 6 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 139 | 6 | MPI_Comm_size(MPI_COMM_WORLD, &size); | |
| 140 | |||
| 141 | 6 | const int radius = window_size_ / 2; | |
| 142 | 6 | const int base_rows = height_ / size; | |
| 143 | 6 | const int extra_rows = height_ % size; | |
| 144 | |||
| 145 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int start_row = (rank * base_rows) + std::min(rank, extra_rows); |
| 146 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | const int end_row = start_row + base_rows + (rank < extra_rows ? 1 : 0); |
| 147 | 6 | const int local_height = end_row - start_row; | |
| 148 | |||
| 149 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | if (local_height <= 0) { |
| 150 | ✗ | GetOutput().width = width_; | |
| 151 | ✗ | GetOutput().height = height_; | |
| 152 | GetOutput().data.clear(); | ||
| 153 | ✗ | return true; | |
| 154 | } | ||
| 155 | |||
| 156 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | const int overlap_start = std::max(0, start_row - radius); |
| 157 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | const int overlap_end = std::min(height_, end_row + radius); |
| 158 | 6 | const int overlap_height = overlap_end - overlap_start; | |
| 159 | |||
| 160 | 6 | std::vector<std::uint8_t> full_image; | |
| 161 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 162 |
1/2✓ Branch 1 taken 3 times.
✗ Branch 2 not taken.
|
3 | full_image = GetInput().data; |
| 163 | } | ||
| 164 | |||
| 165 |
1/4✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
6 | std::vector<int> sendcounts(static_cast<std::size_t>(size)); |
| 166 |
1/4✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
6 | std::vector<int> displs(static_cast<std::size_t>(size)); |
| 167 | |||
| 168 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 169 | 3 | PrepareScatterData(size, width_, height_, radius, sendcounts, displs); | |
| 170 | } | ||
| 171 | |||
| 172 |
1/4✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
6 | std::vector<std::uint8_t> local_image(static_cast<std::size_t>(overlap_height) * width_); |
| 173 |
1/2✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
|
6 | MPI_Scatterv((rank == 0 ? full_image.data() : nullptr), sendcounts.data(), displs.data(), MPI_UNSIGNED_CHAR, |
| 174 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | local_image.data(), overlap_height * width_, MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD); |
| 175 | |||
| 176 |
1/4✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
6 | std::vector<std::uint8_t> local_result(static_cast<std::size_t>(local_height) * width_); |
| 177 |
2/2✓ Branch 0 taken 192 times.
✓ Branch 1 taken 6 times.
|
198 | for (int local_row = 0; local_row < local_height; ++local_row) { |
| 178 | 192 | const int global_row = start_row + local_row; | |
| 179 | |||
| 180 |
2/2✓ Branch 0 taken 12288 times.
✓ Branch 1 taken 192 times.
|
12480 | for (int col_idx = 0; col_idx < width_; ++col_idx) { |
| 181 | 12288 | const std::size_t index = (static_cast<std::size_t>(local_row) * width_) + col_idx; | |
| 182 | 12288 | local_result[index] = | |
| 183 |
1/2✓ Branch 1 taken 12288 times.
✗ Branch 2 not taken.
|
12288 | ComputePixelMedian(col_idx, global_row, overlap_start, local_image, width_, height_, window_size_); |
| 184 | } | ||
| 185 | } | ||
| 186 | |||
| 187 | 6 | std::vector<std::uint8_t> result_image; | |
| 188 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 189 |
1/2✓ Branch 1 taken 3 times.
✗ Branch 2 not taken.
|
3 | result_image.resize(static_cast<std::size_t>(width_) * height_); |
| 190 | } | ||
| 191 | |||
| 192 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 193 | 3 | PrepareGatherData(size, width_, height_, sendcounts, displs); | |
| 194 | } | ||
| 195 | |||
| 196 |
1/2✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
|
6 | MPI_Gatherv(local_result.data(), local_height * width_, MPI_UNSIGNED_CHAR, |
| 197 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | (rank == 0 ? result_image.data() : nullptr), sendcounts.data(), displs.data(), MPI_UNSIGNED_CHAR, 0, |
| 198 | MPI_COMM_WORLD); | ||
| 199 | |||
| 200 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank == 0) { |
| 201 | 3 | GetOutput().width = width_; | |
| 202 | 3 | GetOutput().height = height_; | |
| 203 |
1/2✓ Branch 1 taken 3 times.
✗ Branch 2 not taken.
|
3 | GetOutput().data = result_image; |
| 204 | } else { | ||
| 205 | 3 | GetOutput().width = width_; | |
| 206 | 3 | GetOutput().height = height_; | |
| 207 | } | ||
| 208 | |||
| 209 | 6 | int data_size = static_cast<int>(GetOutput().data.size()); | |
| 210 |
1/2✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
|
6 | MPI_Bcast(&data_size, 1, MPI_INT, 0, MPI_COMM_WORLD); |
| 211 | |||
| 212 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 3 times.
|
6 | if (rank != 0) { |
| 213 |
1/2✓ Branch 1 taken 3 times.
✗ Branch 2 not taken.
|
3 | GetOutput().data.resize(static_cast<std::size_t>(data_size)); |
| 214 | } | ||
| 215 | |||
| 216 |
1/2✓ Branch 1 taken 6 times.
✗ Branch 2 not taken.
|
6 | MPI_Bcast(GetOutput().data.data(), data_size, MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD); |
| 217 | return true; | ||
| 218 | } | ||
| 219 | |||
| 220 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | bool VlasovaAImageSmoothingMPI::PostProcessingImpl() { |
| 221 | const auto &output = GetOutput(); | ||
| 222 | |||
| 223 |
1/2✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
|
6 | if (output.data.empty()) { |
| 224 | return false; | ||
| 225 | } | ||
| 226 | |||
| 227 | 6 | const std::size_t expected_size = static_cast<std::size_t>(width_) * height_; | |
| 228 | 6 | return output.data.size() == expected_size; | |
| 229 | } | ||
| 230 | |||
| 231 | } // namespace vlasova_a_image_smoothing | ||
| 232 |