| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "alekseev_a_custom_reduce/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <cmath> | ||
| 7 | #include <vector> | ||
| 8 | |||
| 9 | namespace alekseev_a_custom_reduce { | ||
| 10 | |||
| 11 | namespace { | ||
| 12 | |||
| 13 | template <typename T> | ||
| 14 | 92 | void ReduceTree(const T *send_data, T *recv_data, int count, int rank, int size, int root, MPI_Comm comm, | |
| 15 | MPI_Datatype mpi_type) { | ||
| 16 | constexpr int kMessageTag = 0; | ||
| 17 | 92 | const int virtual_rank = (rank - root + size) % size; | |
| 18 | |||
| 19 | 92 | std::vector<T> accumulator(send_data, send_data + count); | |
| 20 |
1/4✓ Branch 1 taken 46 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
92 | std::vector<T> buffer(count); |
| 21 | MPI_Status status; | ||
| 22 | |||
| 23 |
2/2✓ Branch 0 taken 46 times.
✓ Branch 1 taken 23 times.
|
138 | for (int step = 1; step < size; step <<= 1) { |
| 24 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
92 | if ((virtual_rank & step) == 0) { |
| 25 | 46 | const int source_virtual = virtual_rank + step; | |
| 26 |
1/2✓ Branch 0 taken 23 times.
✗ Branch 1 not taken.
|
46 | if (source_virtual < size) { |
| 27 |
1/2✓ Branch 1 taken 23 times.
✗ Branch 2 not taken.
|
46 | const int source_rank = (source_virtual + root) % size; |
| 28 |
1/2✓ Branch 1 taken 23 times.
✗ Branch 2 not taken.
|
46 | MPI_Recv(buffer.data(), count, mpi_type, source_rank, kMessageTag, comm, &status); |
| 29 | |||
| 30 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
92 | for (int i = 0; i < count; ++i) { |
| 31 | 46 | accumulator[i] += buffer[i]; | |
| 32 | } | ||
| 33 | } | ||
| 34 | } else { | ||
| 35 | 46 | const int dest_virtual = virtual_rank - step; | |
| 36 |
1/2✓ Branch 1 taken 23 times.
✗ Branch 2 not taken.
|
46 | const int dest_rank = (dest_virtual + root) % size; |
| 37 |
1/2✓ Branch 1 taken 23 times.
✗ Branch 2 not taken.
|
46 | MPI_Send(accumulator.data(), count, mpi_type, dest_rank, kMessageTag, comm); |
| 38 | return; | ||
| 39 | } | ||
| 40 | } | ||
| 41 | |||
| 42 |
1/2✓ Branch 0 taken 23 times.
✗ Branch 1 not taken.
|
46 | if (rank == root && recv_data != nullptr) { |
| 43 | 46 | std::copy(accumulator.begin(), accumulator.end(), recv_data); | |
| 44 | } | ||
| 45 | } | ||
| 46 | |||
| 47 | 46 | int CustomReduceImpl(const void *send_data, void *recv_data, int count, MPI_Datatype datatype, MPI_Op op, int root, | |
| 48 | MPI_Comm comm) { | ||
| 49 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | if (count < 0) { |
| 50 | return MPI_ERR_COUNT; | ||
| 51 | } | ||
| 52 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | if (op != MPI_SUM) { |
| 53 | return MPI_ERR_OP; | ||
| 54 | } | ||
| 55 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 46 times.
|
46 | if (count == 0) { |
| 56 | return MPI_SUCCESS; | ||
| 57 | } | ||
| 58 | |||
| 59 | 46 | int rank = 0; | |
| 60 | 46 | int size = 0; | |
| 61 | 46 | MPI_Comm_rank(comm, &rank); | |
| 62 | 46 | MPI_Comm_size(comm, &size); | |
| 63 | |||
| 64 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | if (size <= 0) { |
| 65 | return MPI_ERR_COMM; | ||
| 66 | } | ||
| 67 |
2/4✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 46 times.
✗ Branch 3 not taken.
|
46 | if (root < 0 || root >= size) { |
| 68 | return MPI_ERR_ROOT; | ||
| 69 | } | ||
| 70 | |||
| 71 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 46 times.
|
46 | if (datatype == MPI_INT) { |
| 72 | ✗ | ReduceTree(static_cast<const int *>(send_data), static_cast<int *>(recv_data), count, rank, size, root, comm, | |
| 73 | MPI_INT); | ||
| 74 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 46 times.
|
46 | } else if (datatype == MPI_FLOAT) { |
| 75 | ✗ | ReduceTree(static_cast<const float *>(send_data), static_cast<float *>(recv_data), count, rank, size, root, comm, | |
| 76 | MPI_FLOAT); | ||
| 77 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | } else if (datatype == MPI_DOUBLE) { |
| 78 | 46 | ReduceTree(static_cast<const double *>(send_data), static_cast<double *>(recv_data), count, rank, size, root, comm, | |
| 79 | MPI_DOUBLE); | ||
| 80 | } else { | ||
| 81 | return MPI_ERR_TYPE; | ||
| 82 | } | ||
| 83 | |||
| 84 | return MPI_SUCCESS; | ||
| 85 | } | ||
| 86 | |||
| 87 | } // namespace | ||
| 88 | |||
| 89 | 46 | bool AlekseevACustomReduceMPI::ValidationImpl() { | |
| 90 | 46 | MPI_Comm_rank(MPI_COMM_WORLD, &world_rank_); | |
| 91 | 46 | MPI_Comm_size(MPI_COMM_WORLD, &world_size_); | |
| 92 | |||
| 93 | 46 | int validation_flag = 1; | |
| 94 | |||
| 95 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | if (world_rank_ == 0) { |
| 96 | const auto &input = GetInput(); | ||
| 97 | |||
| 98 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 23 times.
|
23 | if (input.data.empty()) { |
| 99 | ✗ | validation_flag = 0; | |
| 100 | } | ||
| 101 | |||
| 102 |
2/4✓ Branch 0 taken 23 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 23 times.
|
23 | if (input.root < 0 || input.root >= world_size_) { |
| 103 | ✗ | validation_flag = 0; | |
| 104 | } | ||
| 105 | } | ||
| 106 | |||
| 107 | 46 | MPI_Bcast(&validation_flag, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 108 | 46 | return validation_flag == 1; | |
| 109 | } | ||
| 110 | |||
| 111 | 46 | bool AlekseevACustomReduceMPI::PreProcessingImpl() { | |
| 112 | 46 | MPI_Comm_rank(MPI_COMM_WORLD, &world_rank_); | |
| 113 | 46 | MPI_Comm_size(MPI_COMM_WORLD, &world_size_); | |
| 114 | |||
| 115 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | if (world_rank_ == 0) { |
| 116 | 23 | root_ = GetInput().root; | |
| 117 | } | ||
| 118 | 46 | MPI_Bcast(&root_, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 119 | |||
| 120 | 46 | int global_size = 0; | |
| 121 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | if (world_rank_ == 0) { |
| 122 | 23 | global_size = static_cast<int>(GetInput().data.size()); | |
| 123 | } | ||
| 124 | 46 | MPI_Bcast(&global_size, 1, MPI_INT, 0, MPI_COMM_WORLD); | |
| 125 | |||
| 126 | 46 | std::vector<int> counts(world_size_); | |
| 127 |
1/4✓ Branch 1 taken 46 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
46 | std::vector<int> displacements(world_size_, 0); |
| 128 | |||
| 129 | 46 | const int base_chunk = global_size / world_size_; | |
| 130 | 46 | const int remainder = global_size % world_size_; | |
| 131 | |||
| 132 |
2/2✓ Branch 0 taken 92 times.
✓ Branch 1 taken 46 times.
|
138 | for (int i = 0; i < world_size_; ++i) { |
| 133 |
4/4✓ Branch 0 taken 70 times.
✓ Branch 1 taken 22 times.
✓ Branch 2 taken 46 times.
✓ Branch 3 taken 46 times.
|
162 | counts[i] = base_chunk + (i < remainder ? 1 : 0); |
| 134 |
2/2✓ Branch 0 taken 46 times.
✓ Branch 1 taken 46 times.
|
92 | if (i > 0) { |
| 135 | 46 | displacements[i] = displacements[i - 1] + counts[i - 1]; | |
| 136 | } | ||
| 137 | } | ||
| 138 | |||
| 139 |
1/2✓ Branch 1 taken 46 times.
✗ Branch 2 not taken.
|
46 | local_size_ = counts[world_rank_]; |
| 140 |
1/2✓ Branch 1 taken 46 times.
✗ Branch 2 not taken.
|
46 | local_data_.resize(local_size_); |
| 141 | |||
| 142 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | const double *global_data = world_rank_ == 0 ? GetInput().data.data() : nullptr; |
| 143 | |||
| 144 |
1/2✓ Branch 1 taken 46 times.
✗ Branch 2 not taken.
|
46 | MPI_Scatterv(global_data, counts.data(), displacements.data(), MPI_DOUBLE, local_data_.data(), local_size_, |
| 145 | MPI_DOUBLE, 0, MPI_COMM_WORLD); | ||
| 146 | |||
| 147 | 46 | local_sum_ = 0.0; | |
| 148 |
2/2✓ Branch 0 taken 139 times.
✓ Branch 1 taken 46 times.
|
185 | for (double value : local_data_) { |
| 149 | 139 | local_sum_ += value; | |
| 150 | } | ||
| 151 | |||
| 152 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | GetOutput() = 0.0; |
| 153 | |||
| 154 | 46 | return true; | |
| 155 | } | ||
| 156 | |||
| 157 | 46 | bool AlekseevACustomReduceMPI::RunImpl() { | |
| 158 | 46 | double global_sum = 0.0; | |
| 159 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | double *recv_buffer = world_rank_ == root_ ? &global_sum : nullptr; |
| 160 | |||
| 161 | 46 | const int reduce_result = CustomReduceImpl(&local_sum_, recv_buffer, 1, MPI_DOUBLE, MPI_SUM, root_, MPI_COMM_WORLD); | |
| 162 | |||
| 163 |
1/2✓ Branch 0 taken 46 times.
✗ Branch 1 not taken.
|
46 | if (reduce_result != MPI_SUCCESS) { |
| 164 | return false; | ||
| 165 | } | ||
| 166 | |||
| 167 |
2/2✓ Branch 0 taken 23 times.
✓ Branch 1 taken 23 times.
|
46 | if (world_rank_ == root_) { |
| 168 | 23 | GetOutput() = global_sum; | |
| 169 | } | ||
| 170 | |||
| 171 | return true; | ||
| 172 | } | ||
| 173 | |||
| 174 | 46 | bool AlekseevACustomReduceMPI::PostProcessingImpl() { | |
| 175 | 46 | return true; | |
| 176 | } | ||
| 177 | |||
| 178 | } // namespace alekseev_a_custom_reduce | ||
| 179 |