| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #pragma once | ||
| 2 | |||
| 3 | #include <cstdint> | ||
| 4 | #include <utility> | ||
| 5 | #include <vector> | ||
| 6 | |||
| 7 | #include "melnik_i_gauss_block_part/common/include/common.hpp" | ||
| 8 | #include "task/include/task.hpp" | ||
| 9 | |||
| 10 | namespace melnik_i_gauss_block_part { | ||
| 11 | |||
| 12 | class MelnikIGaussBlockPartMPI : public BaseTask { | ||
| 13 | public: | ||
| 14 | static constexpr ppc::task::TypeOfTask GetStaticTypeOfTask() { | ||
| 15 | return ppc::task::TypeOfTask::kMPI; | ||
| 16 | } | ||
| 17 | |||
| 18 | explicit MelnikIGaussBlockPartMPI(const InType &in); | ||
| 19 | |||
| 20 | private: | ||
| 21 | struct BlockInfo { | ||
| 22 | int start_x = 0; | ||
| 23 | int start_y = 0; | ||
| 24 | int width = 0; | ||
| 25 | int height = 0; | ||
| 26 | [[nodiscard]] bool Empty() const { | ||
| 27 |
24/48✓ Branch 0 taken 9 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 9 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 9 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 9 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 9 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 9 times.
✓ Branch 12 taken 18 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 18 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 18 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 18 times.
✗ Branch 19 not taken.
✓ Branch 20 taken 9 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 9 times.
✗ Branch 23 not taken.
✓ Branch 24 taken 9 times.
✗ Branch 25 not taken.
✗ Branch 26 not taken.
✓ Branch 27 taken 9 times.
✓ Branch 28 taken 9 times.
✗ Branch 29 not taken.
✓ Branch 30 taken 9 times.
✗ Branch 31 not taken.
✓ Branch 32 taken 18 times.
✗ Branch 33 not taken.
✓ Branch 34 taken 18 times.
✗ Branch 35 not taken.
✓ Branch 36 taken 18 times.
✗ Branch 37 not taken.
✓ Branch 38 taken 18 times.
✗ Branch 39 not taken.
✓ Branch 40 taken 18 times.
✗ Branch 41 not taken.
✓ Branch 42 taken 18 times.
✗ Branch 43 not taken.
✓ Branch 44 taken 18 times.
✗ Branch 45 not taken.
✗ Branch 46 not taken.
✓ Branch 47 taken 18 times.
|
162 | return width <= 0 || height <= 0; |
| 28 | } | ||
| 29 | }; | ||
| 30 | |||
| 31 | struct Neighbours { | ||
| 32 | int up = 0; | ||
| 33 | int down = 0; | ||
| 34 | int left = 0; | ||
| 35 | int right = 0; | ||
| 36 | int up_left = 0; | ||
| 37 | int up_right = 0; | ||
| 38 | int down_left = 0; | ||
| 39 | int down_right = 0; | ||
| 40 | }; | ||
| 41 | |||
| 42 | bool ValidationImpl() override; | ||
| 43 | bool PreProcessingImpl() override; | ||
| 44 | bool RunImpl() override; | ||
| 45 | bool PostProcessingImpl() override; | ||
| 46 | |||
| 47 | // Computes optimal 2D process grid (rows x cols) matching image aspect ratio | ||
| 48 | static std::pair<int, int> ComputeProcessGrid(int comm_size, int width, int height); | ||
| 49 | // Computes block info for a process given its rank | ||
| 50 | static BlockInfo ComputeBlockInfo(int rank, int grid_rows, int grid_cols, int width, int height); | ||
| 51 | // Computes block info for a process given its grid coordinates (process row, process column) | ||
| 52 | static BlockInfo ComputeBlockInfoByCoords(int pr, int pc, int grid_rows, int grid_cols, int width, int height); | ||
| 53 | |||
| 54 | // Clamps value v to range [low, high] | ||
| 55 | static int ClampInt(int v, int low, int high); | ||
| 56 | // Creates extended buffer (with halo borders) from local block using clamp | ||
| 57 | static void FillExtendedWithClamp(const std::vector<std::uint8_t> &local, const BlockInfo &blk, int ext_w, | ||
| 58 | std::vector<std::uint8_t> &ext); | ||
| 59 | |||
| 60 | // Exchanges halo regions (rows, columns, corners) with neighboring processes | ||
| 61 | static void ExchangeHalos(const BlockInfo &blk, int grid_rows, int grid_cols, int rank, | ||
| 62 | const std::vector<BlockInfo> &all_blocks, std::vector<std::uint8_t> &ext); | ||
| 63 | |||
| 64 | // Computes ranks of all 8 neighboring processes (up, down, left, right, and 4 diagonals) | ||
| 65 | static Neighbours ComputeNeighbours(const BlockInfo &blk, int grid_rows, int grid_cols, int rank, | ||
| 66 | const std::vector<BlockInfo> &all_blocks); | ||
| 67 | // Exchanges top and bottom row halos with vertical neighbors | ||
| 68 | static void ExchangeRowHalos(const BlockInfo &blk, const Neighbours &nbh, int ext_w, std::vector<std::uint8_t> &ext); | ||
| 69 | // Exchanges left and right column halos with horizontal neighbors | ||
| 70 | static void ExchangeColHalos(const BlockInfo &blk, const Neighbours &nbh, int ext_w, std::vector<std::uint8_t> &ext); | ||
| 71 | // Exchanges corner pixel values with diagonal neighbors | ||
| 72 | static void ExchangeCornerHalos(const BlockInfo &blk, const Neighbours &nbh, int ext_w, | ||
| 73 | std::vector<std::uint8_t> &ext); | ||
| 74 | // Fixes corner halo values when diagonal neighbor is absent, using adjacent edge halos | ||
| 75 | static void FixCornersWithoutDiagonal(const BlockInfo &blk, const Neighbours &nbh, int ext_w, | ||
| 76 | std::vector<std::uint8_t> &ext); | ||
| 77 | |||
| 78 | // Applies 3x3 Gaussian convolution to extended buffer, producing local output block | ||
| 79 | static void ApplyGaussianFromExtended(const BlockInfo &blk, const std::vector<std::uint8_t> &ext, | ||
| 80 | std::vector<std::uint8_t> &local_out); | ||
| 81 | |||
| 82 | // Broadcasts image dimensions (width, height) from rank 0 to all processes | ||
| 83 | static void BroadcastImageSize(int rank, int &width, int &height); | ||
| 84 | // Builds block information for all processes in the communicator | ||
| 85 | static std::vector<BlockInfo> BuildAllBlocks(int comm_size, int grid_rows, int grid_cols, int width, int height); | ||
| 86 | // Sends data blocks from root process to all other processes | ||
| 87 | static void SendBlocksToOthers(int comm_size, int width, int height, const std::vector<BlockInfo> &blocks, | ||
| 88 | const std::vector<std::uint8_t> &root_data); | ||
| 89 | // Scatters input image blocks | ||
| 90 | static std::vector<std::uint8_t> ScatterBlock(int rank, int comm_size, int width, int height, | ||
| 91 | const std::vector<BlockInfo> &blocks, const BlockInfo &my_blk, | ||
| 92 | const std::vector<std::uint8_t> &root_data); | ||
| 93 | // Computes local convolution result | ||
| 94 | static std::vector<std::uint8_t> ComputeLocal(const BlockInfo &my_blk, int grid_rows, int grid_cols, int rank, | ||
| 95 | const std::vector<BlockInfo> &blocks, | ||
| 96 | const std::vector<std::uint8_t> &local_data); | ||
| 97 | // Gathers local results from all processes into global output array on rank 0 | ||
| 98 | static std::vector<std::uint8_t> GatherGlobal(int rank, int comm_size, int width, int height, | ||
| 99 | const std::vector<BlockInfo> &blocks, const BlockInfo &my_blk, | ||
| 100 | const std::vector<std::uint8_t> &local_out); | ||
| 101 | // Finalizes output: broadcast | ||
| 102 | void FinalizeOutput(int rank, int width, int height, std::vector<std::uint8_t> &global_out); | ||
| 103 | }; | ||
| 104 | |||
| 105 | } // namespace melnik_i_gauss_block_part | ||
| 106 |