| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "baranov_a_custom_allreduce/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <cmath> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <cstdlib> | ||
| 8 | #include <cstring> | ||
| 9 | #include <exception> | ||
| 10 | #include <stdexcept> | ||
| 11 | #include <variant> | ||
| 12 | #include <vector> | ||
| 13 | |||
| 14 | #include "baranov_a_custom_allreduce/common/include/common.hpp" | ||
| 15 | |||
| 16 | namespace baranov_a_custom_allreduce { | ||
| 17 | |||
| 18 | ✗ | void BaranovACustomAllreduceMPI::TreeBroadcast(void *buffer, int count, MPI_Datatype datatype, MPI_Comm comm, | |
| 19 | int root) { | ||
| 20 | ✗ | int rank = 0; | |
| 21 | ✗ | int size = 0; | |
| 22 | ✗ | MPI_Comm_rank(comm, &rank); | |
| 23 | ✗ | MPI_Comm_size(comm, &size); | |
| 24 | |||
| 25 | ✗ | if (count == 0) { | |
| 26 | ✗ | return; | |
| 27 | } | ||
| 28 | ✗ | if (rank == root) { | |
| 29 | ✗ | for (int i = 0; i < size; i++) { | |
| 30 | ✗ | if (i != root) { | |
| 31 | ✗ | MPI_Send(buffer, count, datatype, i, 0, comm); | |
| 32 | } | ||
| 33 | } | ||
| 34 | } else { | ||
| 35 | ✗ | MPI_Recv(buffer, count, datatype, root, 0, comm, MPI_STATUS_IGNORE); | |
| 36 | } | ||
| 37 | } | ||
| 38 | |||
| 39 | ✗ | void BaranovACustomAllreduceMPI::TreeReduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, | |
| 40 | MPI_Comm comm, int root) { | ||
| 41 | ✗ | int rank = 0; | |
| 42 | ✗ | int size = 0; | |
| 43 | ✗ | MPI_Comm_rank(comm, &rank); | |
| 44 | ✗ | MPI_Comm_size(comm, &size); | |
| 45 | |||
| 46 | ✗ | if (count == 0) { | |
| 47 | ✗ | return; | |
| 48 | } | ||
| 49 | |||
| 50 | ✗ | int type_size = 0; | |
| 51 | ✗ | MPI_Type_size(datatype, &type_size); | |
| 52 | |||
| 53 | ✗ | std::memcpy(recvbuf, sendbuf, static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); | |
| 54 | ✗ | if (rank == root) { | |
| 55 | ✗ | for (int i = 0; i < size; i++) { | |
| 56 | ✗ | if (i != root) { | |
| 57 | ✗ | std::vector<unsigned char> temp_buf(static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); | |
| 58 | ✗ | if (temp_buf.empty()) { | |
| 59 | ✗ | throw std::runtime_error("Memory allocation failed"); | |
| 60 | } | ||
| 61 | |||
| 62 | ✗ | MPI_Recv(temp_buf.data(), count, datatype, i, 0, comm, MPI_STATUS_IGNORE); | |
| 63 | ✗ | PerformOperation(temp_buf.data(), recvbuf, count, datatype, op); | |
| 64 | } | ||
| 65 | } | ||
| 66 | } else { | ||
| 67 | ✗ | MPI_Send(recvbuf, count, datatype, root, 0, comm); | |
| 68 | ✗ | std::memset(recvbuf, 0, static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); | |
| 69 | } | ||
| 70 | } | ||
| 71 | |||
| 72 | 13 | void BaranovACustomAllreduceMPI::PerformOperation(void *inbuf, void *inoutbuf, int count, MPI_Datatype datatype, | |
| 73 | MPI_Op op) { | ||
| 74 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 13 times.
|
13 | if (op != MPI_SUM) { |
| 75 | ✗ | throw std::runtime_error("Only MPI_SUM operation is supported"); | |
| 76 | } | ||
| 77 | |||
| 78 |
2/2✓ Branch 0 taken 1 times.
✓ Branch 1 taken 12 times.
|
13 | if (datatype == MPI_INT) { |
| 79 | auto *in = static_cast<int *>(inbuf); | ||
| 80 | auto *inout = static_cast<int *>(inoutbuf); | ||
| 81 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 1 times.
|
5 | for (int i = 0; i < count; i++) { |
| 82 | 4 | inout[i] += in[i]; | |
| 83 | } | ||
| 84 |
2/2✓ Branch 0 taken 1 times.
✓ Branch 1 taken 11 times.
|
12 | } else if (datatype == MPI_FLOAT) { |
| 85 | auto *in = static_cast<float *>(inbuf); | ||
| 86 | auto *inout = static_cast<float *>(inoutbuf); | ||
| 87 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 1 times.
|
4 | for (int i = 0; i < count; i++) { |
| 88 | 3 | inout[i] += in[i]; | |
| 89 | } | ||
| 90 |
1/2✓ Branch 0 taken 11 times.
✗ Branch 1 not taken.
|
11 | } else if (datatype == MPI_DOUBLE) { |
| 91 | auto *in = static_cast<double *>(inbuf); | ||
| 92 | auto *inout = static_cast<double *>(inoutbuf); | ||
| 93 |
2/2✓ Branch 0 taken 1037 times.
✓ Branch 1 taken 11 times.
|
1048 | for (int i = 0; i < count; i++) { |
| 94 | 1037 | inout[i] += in[i]; | |
| 95 | } | ||
| 96 | } else { | ||
| 97 | ✗ | throw std::runtime_error("Unsupported datatype"); | |
| 98 | } | ||
| 99 | 13 | } | |
| 100 | |||
| 101 | namespace { | ||
| 102 | 13 | void ProcessRootReceive(std::vector<unsigned char> &temp_buf, int count, MPI_Datatype datatype, MPI_Op op, | |
| 103 | MPI_Comm comm, int root, int size, int type_size) { | ||
| 104 |
2/2✓ Branch 0 taken 26 times.
✓ Branch 1 taken 13 times.
|
39 | for (int i = 0; i < size; i++) { |
| 105 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 13 times.
|
26 | if (i != root) { |
| 106 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 13 times.
|
13 | std::vector<unsigned char> recv_buf(static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); |
| 107 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 13 times.
|
13 | if (recv_buf.empty()) { |
| 108 | ✗ | throw std::runtime_error("Memory allocation failed"); | |
| 109 | } | ||
| 110 | |||
| 111 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | MPI_Recv(recv_buf.data(), count, datatype, i, 0, comm, MPI_STATUS_IGNORE); |
| 112 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | BaranovACustomAllreduceMPI::PerformOperation(recv_buf.data(), temp_buf.data(), count, datatype, op); |
| 113 | } | ||
| 114 | } | ||
| 115 | 13 | } | |
| 116 | |||
| 117 | 13 | void ProcessRootSend(std::vector<unsigned char> &temp_buf, void *recvbuf, int count, MPI_Datatype datatype, | |
| 118 | MPI_Comm comm, int root, int size, int type_size) { | ||
| 119 |
2/2✓ Branch 0 taken 26 times.
✓ Branch 1 taken 13 times.
|
39 | for (int i = 0; i < size; i++) { |
| 120 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 13 times.
|
26 | if (i != root) { |
| 121 | 13 | MPI_Send(temp_buf.data(), count, datatype, i, 1, comm); | |
| 122 | } | ||
| 123 | } | ||
| 124 | 13 | std::memcpy(recvbuf, temp_buf.data(), static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); | |
| 125 | 13 | } | |
| 126 | |||
| 127 | } // namespace | ||
| 128 | |||
| 129 | 26 | void BaranovACustomAllreduceMPI::CustomAllreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, | |
| 130 | MPI_Op op, MPI_Comm comm, int root) { | ||
| 131 | 26 | int rank = 0; | |
| 132 | 26 | int size = 0; | |
| 133 | 26 | MPI_Comm_rank(comm, &rank); | |
| 134 | 26 | MPI_Comm_size(comm, &size); | |
| 135 | |||
| 136 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 26 times.
|
26 | if (count == 0) { |
| 137 | ✗ | return; | |
| 138 | } | ||
| 139 | |||
| 140 | 26 | int type_size = 0; | |
| 141 | 26 | MPI_Type_size(datatype, &type_size); | |
| 142 | |||
| 143 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 26 times.
|
26 | std::vector<unsigned char> temp_buf(static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); |
| 144 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 26 times.
|
26 | if (temp_buf.empty()) { |
| 145 | ✗ | throw std::runtime_error("Memory allocation failed"); | |
| 146 | } | ||
| 147 | |||
| 148 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 13 times.
|
26 | std::memcpy(temp_buf.data(), sendbuf, static_cast<std::size_t>(count) * static_cast<std::size_t>(type_size)); |
| 149 | |||
| 150 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 13 times.
|
26 | if (rank == root) { |
| 151 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | ProcessRootReceive(temp_buf, count, datatype, op, comm, root, size, type_size); |
| 152 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | ProcessRootSend(temp_buf, recvbuf, count, datatype, comm, root, size, type_size); |
| 153 | } else { | ||
| 154 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | MPI_Send(sendbuf, count, datatype, root, 0, comm); |
| 155 |
1/2✓ Branch 1 taken 13 times.
✗ Branch 2 not taken.
|
13 | MPI_Recv(recvbuf, count, datatype, root, 1, comm, MPI_STATUS_IGNORE); |
| 156 | } | ||
| 157 | } | ||
| 158 | |||
| 159 | template <typename T> | ||
| 160 | ✗ | std::vector<T> BaranovACustomAllreduceMPI::GetVectorFromVariant(const InTypeVariant &variant) { | |
| 161 | try { | ||
| 162 | ✗ | return std::get<std::vector<T>>(variant); | |
| 163 | ✗ | } catch (const std::bad_variant_access &) { | |
| 164 | ✗ | throw std::runtime_error("Wrong variant type accessed"); | |
| 165 | } | ||
| 166 | } | ||
| 167 | |||
| 168 | template std::vector<double> BaranovACustomAllreduceMPI::GetVectorFromVariant<double>(const InTypeVariant &variant); | ||
| 169 | |||
| 170 |
1/2✓ Branch 1 taken 28 times.
✗ Branch 2 not taken.
|
28 | BaranovACustomAllreduceMPI::BaranovACustomAllreduceMPI(const InType &in) { |
| 171 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 172 | GetInput() = in; | ||
| 173 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 26 times.
|
28 | if (std::holds_alternative<std::vector<int>>(in)) { |
| 174 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto vec = std::get<std::vector<int>>(in); |
| 175 |
2/6✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
4 | GetOutput() = InTypeVariant{std::vector<int>(vec.size(), 0)}; |
| 176 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 24 times.
|
26 | } else if (std::holds_alternative<std::vector<float>>(in)) { |
| 177 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto vec = std::get<std::vector<float>>(in); |
| 178 |
2/6✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
4 | GetOutput() = InTypeVariant{std::vector<float>(vec.size(), 0.0F)}; |
| 179 | } else { | ||
| 180 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | auto vec = std::get<std::vector<double>>(in); |
| 181 |
3/6✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 22 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
48 | GetOutput() = InTypeVariant{std::vector<double>(vec.size(), 0.0)}; |
| 182 | } | ||
| 183 | 28 | } | |
| 184 | |||
| 185 | 28 | bool BaranovACustomAllreduceMPI::ValidationImpl() { | |
| 186 | 28 | return true; | |
| 187 | } | ||
| 188 | |||
| 189 | 28 | bool BaranovACustomAllreduceMPI::PreProcessingImpl() { | |
| 190 | 28 | return true; | |
| 191 | } | ||
| 192 | |||
| 193 |
1/2✓ Branch 1 taken 28 times.
✗ Branch 2 not taken.
|
28 | bool BaranovACustomAllreduceMPI::RunImpl() { |
| 194 | try { | ||
| 195 | auto input = GetInput(); | ||
| 196 | auto output = GetOutput(); | ||
| 197 | |||
| 198 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 26 times.
|
28 | if (std::holds_alternative<std::vector<int>>(input)) { |
| 199 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto data = std::get<std::vector<int>>(input); |
| 200 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
2 | if (data.empty()) { |
| 201 | ✗ | GetOutput() = InTypeVariant{std::vector<int>{}}; | |
| 202 | return true; | ||
| 203 | } | ||
| 204 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto result_data = std::get<std::vector<int>>(output); |
| 205 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | CustomAllreduce(data.data(), result_data.data(), static_cast<int>(data.size()), MPI_INT, MPI_SUM, MPI_COMM_WORLD, |
| 206 | 0); | ||
| 207 | |||
| 208 |
1/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2 | GetOutput() = InTypeVariant{result_data}; |
| 209 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 24 times.
|
26 | } else if (std::holds_alternative<std::vector<float>>(input)) { |
| 210 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto data = std::get<std::vector<float>>(input); |
| 211 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
2 | if (data.empty()) { |
| 212 | ✗ | GetOutput() = InTypeVariant{std::vector<float>{}}; | |
| 213 | return true; | ||
| 214 | } | ||
| 215 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | auto result_data = std::get<std::vector<float>>(output); |
| 216 |
1/2✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
|
2 | CustomAllreduce(data.data(), result_data.data(), static_cast<int>(data.size()), MPI_FLOAT, MPI_SUM, |
| 217 | MPI_COMM_WORLD, 0); | ||
| 218 |
1/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2 | GetOutput() = InTypeVariant{result_data}; |
| 219 |
1/2✓ Branch 0 taken 24 times.
✗ Branch 1 not taken.
|
24 | } else if (std::holds_alternative<std::vector<double>>(input)) { |
| 220 |
1/2✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
|
24 | auto data = std::get<std::vector<double>>(input); |
| 221 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 22 times.
|
24 | if (data.empty()) { |
| 222 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
2 | GetOutput() = InTypeVariant{std::vector<double>{}}; |
| 223 | return true; | ||
| 224 | } | ||
| 225 |
1/2✓ Branch 1 taken 22 times.
✗ Branch 2 not taken.
|
22 | auto result_data = std::get<std::vector<double>>(output); |
| 226 |
1/2✓ Branch 1 taken 22 times.
✗ Branch 2 not taken.
|
22 | CustomAllreduce(data.data(), result_data.data(), static_cast<int>(data.size()), MPI_DOUBLE, MPI_SUM, |
| 227 | MPI_COMM_WORLD, 0); | ||
| 228 |
1/4✓ Branch 0 taken 22 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
22 | GetOutput() = InTypeVariant{result_data}; |
| 229 | } | ||
| 230 | return true; | ||
| 231 | ✗ | } catch (const std::exception &) { | |
| 232 | return false; | ||
| 233 | ✗ | } | |
| 234 | } | ||
| 235 | |||
| 236 | 28 | bool BaranovACustomAllreduceMPI::PostProcessingImpl() { | |
| 237 | 28 | return true; | |
| 238 | } | ||
| 239 | |||
| 240 | } // namespace baranov_a_custom_allreduce | ||
| 241 |