| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "krymova_k_scatter/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <cmath> | ||
| 7 | #include <cstdint> | ||
| 8 | #include <cstring> | ||
| 9 | #include <utility> | ||
| 10 | #include <vector> | ||
| 11 | |||
| 12 | #include "krymova_k_scatter/common/include/common.hpp" | ||
| 13 | |||
| 14 | namespace krymova_k_scatter { | ||
| 15 | |||
| 16 | 30 | KrymovaKScatterMPI::KrymovaKScatterMPI(const InType &in) { | |
| 17 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 18 | 30 | GetInput() = in; | |
| 19 | 30 | } | |
| 20 | |||
| 21 | 30 | bool KrymovaKScatterMPI::ValidationImpl() { | |
| 22 | const auto &args = GetInput(); | ||
| 23 | |||
| 24 |
2/4✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 30 times.
|
30 | bool is_invalid_counts = (args.send_count <= 0) || (args.send_count != args.recv_count); |
| 25 | 30 | bool is_invalid_root = (args.root_rank < 0); | |
| 26 | |||
| 27 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (is_invalid_counts || is_invalid_root) { |
| 28 | return false; | ||
| 29 | } | ||
| 30 | |||
| 31 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (args.send_type != args.recv_type) { |
| 32 | return false; | ||
| 33 | } | ||
| 34 | |||
| 35 |
5/6✓ Branch 0 taken 20 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 10 times.
|
30 | auto is_supported = [](MPI_Datatype t) { return (t == MPI_INT || t == MPI_FLOAT || t == MPI_DOUBLE); }; |
| 36 | |||
| 37 | return is_supported(args.send_type); | ||
| 38 | } | ||
| 39 | |||
| 40 | 30 | bool KrymovaKScatterMPI::PreProcessingImpl() { | |
| 41 | auto &args = GetInput(); | ||
| 42 | 30 | int comm_size = 0; | |
| 43 | 30 | MPI_Comm_size(args.comm, &comm_size); | |
| 44 | |||
| 45 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 20 times.
|
30 | if (args.root_rank >= comm_size) { |
| 46 | 10 | args.root_rank %= comm_size; | |
| 47 | } | ||
| 48 | |||
| 49 | 30 | return true; | |
| 50 | } | ||
| 51 | |||
| 52 | namespace { | ||
| 53 | size_t GetTypeByteSize(MPI_Datatype type) { | ||
| 54 | 30 | MPI_Aint lb = 0; | |
| 55 | 30 | MPI_Aint extent = 0; | |
| 56 | 30 | MPI_Type_get_extent(type, &lb, &extent); | |
| 57 | 30 | return static_cast<size_t>(extent); | |
| 58 | } | ||
| 59 | |||
| 60 | int MapVirtToReal(int virt_rank, int root, int size) { | ||
| 61 | 30 | return (virt_rank + root) % size; | |
| 62 | } | ||
| 63 | |||
| 64 | 15 | std::vector<uint8_t> PrepareRootData(const void *src, int size, int root, int count, size_t type_size) { | |
| 65 | 15 | size_t chunk_bytes = static_cast<size_t>(count) * type_size; | |
| 66 | 15 | size_t total_bytes = static_cast<size_t>(size) * chunk_bytes; | |
| 67 | |||
| 68 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
15 | std::vector<uint8_t> buffer(total_bytes); |
| 69 | |||
| 70 | const auto *src_bytes = static_cast<const uint8_t *>(src); | ||
| 71 | uint8_t *dst_bytes = buffer.data(); | ||
| 72 | |||
| 73 | 15 | size_t offset = root * chunk_bytes; | |
| 74 | 15 | size_t tail_size = total_bytes - offset; | |
| 75 | |||
| 76 |
1/2✓ Branch 0 taken 15 times.
✗ Branch 1 not taken.
|
15 | std::copy(src_bytes + offset, src_bytes + total_bytes, dst_bytes); |
| 77 | |||
| 78 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 9 times.
|
15 | std::copy(src_bytes, src_bytes + offset, dst_bytes + tail_size); |
| 79 | |||
| 80 | 15 | return buffer; | |
| 81 | } | ||
| 82 | |||
| 83 | 30 | void ExecScatterCycle(int size, int root, int rank, int count, MPI_Datatype type, size_t type_size, MPI_Comm comm, | |
| 84 | const uint8_t *&active_ptr, std::vector<uint8_t> &buffer) { | ||
| 85 | 30 | int relative_rank = (rank - root + size) % size; | |
| 86 | |||
| 87 | int start_stride = 1; | ||
| 88 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 30 times.
|
60 | while (start_stride < size) { |
| 89 | 30 | start_stride <<= 1; | |
| 90 | } | ||
| 91 | 30 | start_stride >>= 1; | |
| 92 | |||
| 93 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 30 times.
|
60 | for (int stride = start_stride; stride > 0; stride >>= 1) { |
| 94 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (relative_rank % stride != 0) { |
| 95 | ✗ | continue; | |
| 96 | } | ||
| 97 | |||
| 98 | 30 | bool is_sender = (relative_rank % (stride << 1) == 0); | |
| 99 | |||
| 100 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (is_sender) { |
| 101 | 15 | int virt_dest = relative_rank + stride; | |
| 102 | |||
| 103 |
1/2✓ Branch 0 taken 15 times.
✗ Branch 1 not taken.
|
15 | if (virt_dest < size) { |
| 104 | 15 | int limit = std::min(virt_dest + stride, size); | |
| 105 | 15 | int send_amt = (limit - virt_dest) * count; | |
| 106 | |||
| 107 | 15 | size_t byte_shift = static_cast<size_t>(virt_dest - relative_rank) * count * type_size; | |
| 108 | int real_dest = MapVirtToReal(virt_dest, root, size); | ||
| 109 | |||
| 110 | 15 | MPI_Send(active_ptr + byte_shift, send_amt, type, real_dest, 0, comm); | |
| 111 | } | ||
| 112 | } else { | ||
| 113 | 15 | int virt_src = relative_rank - stride; | |
| 114 | int real_src = MapVirtToReal(virt_src, root, size); | ||
| 115 | |||
| 116 | 15 | int limit = std::min(relative_rank + stride, size); | |
| 117 | 15 | int recv_amt = (limit - relative_rank) * count; | |
| 118 | |||
| 119 | 15 | size_t required_bytes = static_cast<size_t>(recv_amt) * type_size; | |
| 120 | 15 | buffer.resize(required_bytes); | |
| 121 | |||
| 122 | 15 | MPI_Recv(buffer.data(), recv_amt, type, real_src, 0, comm, MPI_STATUS_IGNORE); | |
| 123 | |||
| 124 | 15 | active_ptr = buffer.data(); | |
| 125 | } | ||
| 126 | } | ||
| 127 | 30 | } | |
| 128 | |||
| 129 | } // namespace | ||
| 130 | |||
| 131 | 30 | bool KrymovaKScatterMPI::RunImpl() { | |
| 132 | auto &args = GetInput(); | ||
| 133 | |||
| 134 | 30 | int rank = 0; | |
| 135 | 30 | int size = 0; | |
| 136 | 30 | MPI_Comm_rank(args.comm, &rank); | |
| 137 | 30 | MPI_Comm_size(args.comm, &size); | |
| 138 | |||
| 139 | 30 | size_t type_size = GetTypeByteSize(args.recv_type); | |
| 140 | |||
| 141 | 30 | std::vector<uint8_t> internal_buf; | |
| 142 | 30 | const uint8_t *active_data_ptr = nullptr; | |
| 143 | |||
| 144 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
|
30 | if (rank == args.root_rank) { |
| 145 |
1/2✓ Branch 1 taken 15 times.
✗ Branch 2 not taken.
|
30 | internal_buf = PrepareRootData(args.src_buffer, size, args.root_rank, args.recv_count, type_size); |
| 146 | 15 | active_data_ptr = internal_buf.data(); | |
| 147 | } | ||
| 148 | |||
| 149 |
1/2✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
|
30 | ExecScatterCycle(size, args.root_rank, rank, args.recv_count, args.recv_type, type_size, args.comm, active_data_ptr, |
| 150 | internal_buf); | ||
| 151 | |||
| 152 |
2/4✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 30 times.
✗ Branch 3 not taken.
|
30 | if (args.dst_buffer != MPI_IN_PLACE && active_data_ptr != nullptr) { |
| 153 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | size_t bytes_to_copy = args.recv_count * type_size; |
| 154 | auto *user_dst = static_cast<uint8_t *>(args.dst_buffer); | ||
| 155 | std::copy(active_data_ptr, active_data_ptr + bytes_to_copy, user_dst); | ||
| 156 | } | ||
| 157 | |||
| 158 | 30 | size_t result_bytes = args.recv_count * type_size; | |
| 159 |
1/4✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
30 | std::vector<uint8_t> result(result_bytes); |
| 160 | |||
| 161 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | const void *final_src = (args.dst_buffer != nullptr) ? args.dst_buffer : active_data_ptr; |
| 162 |
1/2✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
|
30 | if (final_src != nullptr) { |
| 163 | const auto *ptr = static_cast<const uint8_t *>(final_src); | ||
| 164 | 30 | std::copy(ptr, ptr + result_bytes, result.begin()); | |
| 165 | } | ||
| 166 | |||
| 167 | GetOutput() = std::move(result); | ||
| 168 | 30 | return true; | |
| 169 | } | ||
| 170 | |||
| 171 | 30 | bool KrymovaKScatterMPI::PostProcessingImpl() { | |
| 172 | 30 | return true; | |
| 173 | } | ||
| 174 | |||
| 175 | } // namespace krymova_k_scatter | ||
| 176 |