| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "kondrashova_v_gauss_filter_vertical_split/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 <vector> | ||
| 10 | |||
| 11 | #include "kondrashova_v_gauss_filter_vertical_split/common/include/common.hpp" | ||
| 12 | |||
| 13 | namespace kondrashova_v_gauss_filter_vertical_split { | ||
| 14 | |||
| 15 | const std::array<std::array<int, 3>, 3> KondrashovaVGaussFilterVerticalSplitMPI::kGaussKernel = { | ||
| 16 | {{{1, 2, 1}}, {{2, 4, 2}}, {{1, 2, 1}}}}; | ||
| 17 | const int KondrashovaVGaussFilterVerticalSplitMPI::kGaussKernelSum = 16; | ||
| 18 | |||
| 19 | ✗ | KondrashovaVGaussFilterVerticalSplitMPI::KondrashovaVGaussFilterVerticalSplitMPI(const InType &in) { | |
| 20 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 21 | GetInput() = in; | ||
| 22 | ✗ | } | |
| 23 | |||
| 24 | ✗ | uint8_t KondrashovaVGaussFilterVerticalSplitMPI::ApplyGaussToLocalPixel(const std::vector<uint8_t> &local_data, | |
| 25 | int local_width, int height, int channels, | ||
| 26 | int px, int py, int channel) { | ||
| 27 | int sum = 0; | ||
| 28 | |||
| 29 | ✗ | for (int ky = -1; ky <= 1; ++ky) { | |
| 30 | ✗ | for (int kx = -1; kx <= 1; ++kx) { | |
| 31 | ✗ | int nx = std::clamp(px + kx, 0, local_width - 1); | |
| 32 | ✗ | int ny = std::clamp(py + ky, 0, height - 1); | |
| 33 | |||
| 34 | ✗ | int idx = (((ny * local_width) + nx) * channels) + channel; | |
| 35 | ✗ | auto kernel_row = static_cast<size_t>(ky) + 1; | |
| 36 | ✗ | auto kernel_col = static_cast<size_t>(kx) + 1; | |
| 37 | ✗ | sum += local_data[idx] * kGaussKernel.at(kernel_row).at(kernel_col); | |
| 38 | } | ||
| 39 | } | ||
| 40 | |||
| 41 | ✗ | return static_cast<uint8_t>(std::clamp(sum / kGaussKernelSum, 0, 255)); | |
| 42 | } | ||
| 43 | |||
| 44 | ✗ | bool KondrashovaVGaussFilterVerticalSplitMPI::ValidationImpl() { | |
| 45 | ✗ | int rank = 0; | |
| 46 | ✗ | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 47 | |||
| 48 | ✗ | if (rank == 0) { | |
| 49 | const auto &input = GetInput(); | ||
| 50 | |||
| 51 | ✗ | auto expected_size = static_cast<size_t>(input.width) * input.height * input.channels; | |
| 52 | ✗ | return input.pixels.size() == expected_size && input.width >= 3 && input.height >= 3 && input.channels >= 1 && | |
| 53 | input.channels <= 4; | ||
| 54 | } | ||
| 55 | return true; | ||
| 56 | } | ||
| 57 | |||
| 58 | ✗ | bool KondrashovaVGaussFilterVerticalSplitMPI::PreProcessingImpl() { | |
| 59 | ✗ | int rank = 0; | |
| 60 | ✗ | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 61 | |||
| 62 | ✗ | if (rank == 0) { | |
| 63 | const auto &input = GetInput(); | ||
| 64 | auto &output = GetOutput(); | ||
| 65 | |||
| 66 | ✗ | output.width = input.width; | |
| 67 | ✗ | output.height = input.height; | |
| 68 | ✗ | output.channels = input.channels; | |
| 69 | ✗ | output.pixels.resize(input.pixels.size()); | |
| 70 | } | ||
| 71 | ✗ | return true; | |
| 72 | } | ||
| 73 | |||
| 74 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::BroadcastImageDimensions(int &width, int &height, int &channels) { | |
| 75 | ✗ | int rank = 0; | |
| 76 | ✗ | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 77 | |||
| 78 | ✗ | if (rank == 0) { | |
| 79 | ✗ | width = GetInput().width; | |
| 80 | ✗ | height = GetInput().height; | |
| 81 | ✗ | channels = GetInput().channels; | |
| 82 | } | ||
| 83 | |||
| 84 | ✗ | MPI_Bcast(&width, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 85 | ✗ | MPI_Bcast(&height, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 86 | ✗ | MPI_Bcast(&channels, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 87 | ✗ | } | |
| 88 | |||
| 89 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::CalculateColumnDistribution(int width, int size, | |
| 90 | std::vector<int> &col_counts, | ||
| 91 | std::vector<int> &col_offsets) { | ||
| 92 | ✗ | int base_cols = width / size; | |
| 93 | ✗ | int extra_cols = width % size; | |
| 94 | |||
| 95 | ✗ | col_counts.resize(size); | |
| 96 | ✗ | col_offsets.resize(size); | |
| 97 | |||
| 98 | ✗ | for (int i = 0; i < size; ++i) { | |
| 99 | ✗ | col_counts[i] = base_cols + (i < extra_cols ? 1 : 0); | |
| 100 | ✗ | col_offsets[i] = (i == 0) ? 0 : col_offsets[i - 1] + col_counts[i - 1]; | |
| 101 | } | ||
| 102 | ✗ | } | |
| 103 | |||
| 104 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::CopyPixelsToBuffer(const std::vector<uint8_t> &src, | |
| 105 | std::vector<uint8_t> &dst, int src_width, | ||
| 106 | int dst_width, int height, int channels, | ||
| 107 | int src_start_col) { | ||
| 108 | ✗ | for (int row = 0; row < height; ++row) { | |
| 109 | ✗ | for (int col = 0; col < dst_width; ++col) { | |
| 110 | ✗ | for (int ch = 0; ch < channels; ++ch) { | |
| 111 | ✗ | int src_idx = (((row * src_width) + (src_start_col + col)) * channels) + ch; | |
| 112 | ✗ | int dst_idx = (((row * dst_width) + col) * channels) + ch; | |
| 113 | ✗ | dst[dst_idx] = src[src_idx]; | |
| 114 | } | ||
| 115 | } | ||
| 116 | } | ||
| 117 | ✗ | } | |
| 118 | |||
| 119 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::CopyBufferToOutput(const std::vector<uint8_t> &src, | |
| 120 | std::vector<uint8_t> &dst, int src_width, | ||
| 121 | int dst_width, int height, int channels, | ||
| 122 | int dst_start_col) { | ||
| 123 | ✗ | for (int row = 0; row < height; ++row) { | |
| 124 | ✗ | for (int col = 0; col < src_width; ++col) { | |
| 125 | ✗ | for (int ch = 0; ch < channels; ++ch) { | |
| 126 | ✗ | int src_idx = (((row * src_width) + col) * channels) + ch; | |
| 127 | ✗ | int dst_idx = (((row * dst_width) + (dst_start_col + col)) * channels) + ch; | |
| 128 | ✗ | dst[dst_idx] = src[src_idx]; | |
| 129 | } | ||
| 130 | } | ||
| 131 | } | ||
| 132 | ✗ | } | |
| 133 | |||
| 134 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::DistributeImageData(int rank, int size, int width, int height, | |
| 135 | int channels, const std::vector<int> &col_counts, | ||
| 136 | const std::vector<int> &col_offsets, | ||
| 137 | std::vector<uint8_t> &local_data, int extended_cols) { | ||
| 138 | ✗ | local_data.resize(static_cast<size_t>(extended_cols) * height * channels); | |
| 139 | |||
| 140 | ✗ | if (rank == 0) { | |
| 141 | ✗ | const auto &input_pixels = GetInput().pixels; | |
| 142 | |||
| 143 | ✗ | for (int proc = 0; proc < size; ++proc) { | |
| 144 | ✗ | int p_start = std::max(0, col_offsets[proc] - 1); | |
| 145 | ✗ | int p_end = std::min(width, col_offsets[proc] + col_counts[proc] + 1); | |
| 146 | ✗ | int p_cols = p_end - p_start; | |
| 147 | |||
| 148 | ✗ | std::vector<uint8_t> send_data(static_cast<size_t>(p_cols) * height * channels); | |
| 149 | ✗ | CopyPixelsToBuffer(input_pixels, send_data, width, p_cols, height, channels, p_start); | |
| 150 | |||
| 151 | ✗ | if (proc == 0) { | |
| 152 | ✗ | local_data = send_data; | |
| 153 | } else { | ||
| 154 | ✗ | MPI_Send(send_data.data(), static_cast<int>(send_data.size()), MPI_BYTE, proc, 0, MPI_COMM_WORLD); | |
| 155 | } | ||
| 156 | } | ||
| 157 | } else { | ||
| 158 | MPI_Status status; | ||
| 159 | ✗ | MPI_Recv(local_data.data(), static_cast<int>(local_data.size()), MPI_BYTE, 0, 0, MPI_COMM_WORLD, &status); | |
| 160 | } | ||
| 161 | ✗ | } | |
| 162 | |||
| 163 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::ApplyGaussFilterToLocalData(const std::vector<uint8_t> &local_data, | |
| 164 | std::vector<uint8_t> &local_result, | ||
| 165 | int extended_cols, int local_cols, int height, | ||
| 166 | int channels, int offset_in_extended) { | ||
| 167 | ✗ | local_result.resize(static_cast<size_t>(local_cols) * height * channels); | |
| 168 | |||
| 169 | ✗ | for (int row = 0; row < height; ++row) { | |
| 170 | ✗ | for (int lx = 0; lx < local_cols; ++lx) { | |
| 171 | ✗ | int col = offset_in_extended + lx; | |
| 172 | ✗ | for (int ch = 0; ch < channels; ++ch) { | |
| 173 | ✗ | int result_idx = (((row * local_cols) + lx) * channels) + ch; | |
| 174 | ✗ | local_result[result_idx] = ApplyGaussToLocalPixel(local_data, extended_cols, height, channels, col, row, ch); | |
| 175 | } | ||
| 176 | } | ||
| 177 | } | ||
| 178 | ✗ | } | |
| 179 | |||
| 180 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::GatherResults(int rank, int size, int width, int height, int channels, | |
| 181 | const std::vector<int> &col_counts, | ||
| 182 | const std::vector<int> &col_offsets, int local_start_col, | ||
| 183 | int local_cols, const std::vector<uint8_t> &local_result) { | ||
| 184 | ✗ | if (rank == 0) { | |
| 185 | ✗ | auto &output_pixels = GetOutput().pixels; | |
| 186 | |||
| 187 | ✗ | CopyBufferToOutput(local_result, output_pixels, local_cols, width, height, channels, local_start_col); | |
| 188 | |||
| 189 | ✗ | for (int proc = 1; proc < size; ++proc) { | |
| 190 | ✗ | int p_cols = col_counts[proc]; | |
| 191 | ✗ | std::vector<uint8_t> recv_data(static_cast<size_t>(p_cols) * height * channels); | |
| 192 | MPI_Status status; | ||
| 193 | ✗ | MPI_Recv(recv_data.data(), static_cast<int>(recv_data.size()), MPI_BYTE, proc, 1, MPI_COMM_WORLD, &status); | |
| 194 | |||
| 195 | ✗ | CopyBufferToOutput(recv_data, output_pixels, p_cols, width, height, channels, col_offsets[proc]); | |
| 196 | } | ||
| 197 | } else { | ||
| 198 | ✗ | MPI_Send(local_result.data(), static_cast<int>(local_result.size()), MPI_BYTE, 0, 1, MPI_COMM_WORLD); | |
| 199 | } | ||
| 200 | ✗ | } | |
| 201 | |||
| 202 | ✗ | void KondrashovaVGaussFilterVerticalSplitMPI::BroadcastResultToAllProcesses(int width, int height, int channels) { | |
| 203 | ✗ | int rank = 0; | |
| 204 | ✗ | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 205 | |||
| 206 | ✗ | if (rank != 0) { | |
| 207 | auto &output = GetOutput(); | ||
| 208 | ✗ | output.width = width; | |
| 209 | ✗ | output.height = height; | |
| 210 | ✗ | output.channels = channels; | |
| 211 | ✗ | output.pixels.resize(static_cast<size_t>(width) * height * channels); | |
| 212 | } | ||
| 213 | |||
| 214 | ✗ | MPI_Bcast(GetOutput().pixels.data(), static_cast<int>(GetOutput().pixels.size()), MPI_BYTE, 0, MPI_COMM_WORLD); | |
| 215 | ✗ | } | |
| 216 | |||
| 217 | ✗ | bool KondrashovaVGaussFilterVerticalSplitMPI::RunImpl() { | |
| 218 | ✗ | int rank = 0; | |
| 219 | ✗ | int size = 0; | |
| 220 | ✗ | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 221 | ✗ | MPI_Comm_size(MPI_COMM_WORLD, &size); | |
| 222 | |||
| 223 | ✗ | int width = 0; | |
| 224 | ✗ | int height = 0; | |
| 225 | ✗ | int channels = 0; | |
| 226 | ✗ | BroadcastImageDimensions(width, height, channels); | |
| 227 | |||
| 228 | ✗ | std::vector<int> col_counts; | |
| 229 | ✗ | std::vector<int> col_offsets; | |
| 230 | ✗ | CalculateColumnDistribution(width, size, col_counts, col_offsets); | |
| 231 | |||
| 232 | ✗ | int local_start_col = col_offsets[rank]; | |
| 233 | ✗ | int local_cols = col_counts[rank]; | |
| 234 | |||
| 235 | ✗ | int extended_start = std::max(0, local_start_col - 1); | |
| 236 | ✗ | int extended_end = std::min(width, local_start_col + local_cols + 1); | |
| 237 | ✗ | int extended_cols = extended_end - extended_start; | |
| 238 | ✗ | int offset_in_extended = local_start_col - extended_start; | |
| 239 | |||
| 240 | ✗ | std::vector<uint8_t> local_data; | |
| 241 | ✗ | DistributeImageData(rank, size, width, height, channels, col_counts, col_offsets, local_data, extended_cols); | |
| 242 | |||
| 243 | ✗ | std::vector<uint8_t> local_result; | |
| 244 | ✗ | ApplyGaussFilterToLocalData(local_data, local_result, extended_cols, local_cols, height, channels, | |
| 245 | offset_in_extended); | ||
| 246 | |||
| 247 | ✗ | GatherResults(rank, size, width, height, channels, col_counts, col_offsets, local_start_col, local_cols, | |
| 248 | local_result); | ||
| 249 | |||
| 250 | ✗ | BroadcastResultToAllProcesses(width, height, channels); | |
| 251 | |||
| 252 | ✗ | return true; | |
| 253 | } | ||
| 254 | |||
| 255 | ✗ | bool KondrashovaVGaussFilterVerticalSplitMPI::PostProcessingImpl() { | |
| 256 | ✗ | return true; | |
| 257 | } | ||
| 258 | |||
| 259 | } // namespace kondrashova_v_gauss_filter_vertical_split | ||
| 260 |