1 | #ifdef _usingEP |
---|
2 | /*! |
---|
3 | \file ep_gather.cpp |
---|
4 | \since 2 may 2016 |
---|
5 | |
---|
6 | \brief Definitions of MPI collective function: MPI_Gather, MPI_Allgather |
---|
7 | */ |
---|
8 | |
---|
9 | #include "ep_lib.hpp" |
---|
10 | #include <mpi.h> |
---|
11 | #include "ep_declaration.hpp" |
---|
12 | #include "ep_mpi.hpp" |
---|
13 | |
---|
14 | using namespace std; |
---|
15 | |
---|
16 | namespace ep_lib |
---|
17 | { |
---|
18 | |
---|
19 | int MPI_Gather_local(const void *sendbuf, int count, MPI_Datatype datatype, void *recvbuf, int local_root, MPI_Comm comm) |
---|
20 | { |
---|
21 | assert(valid_type(datatype)); |
---|
22 | |
---|
23 | ::MPI_Aint datasize, lb; |
---|
24 | ::MPI_Type_get_extent(to_mpi_type(datatype), &lb, &datasize); |
---|
25 | |
---|
26 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
27 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
28 | |
---|
29 | #pragma omp critical (_gather) |
---|
30 | comm->my_buffer->void_buffer[ep_rank_loc] = const_cast< void* >(sendbuf); |
---|
31 | |
---|
32 | MPI_Barrier_local(comm); |
---|
33 | |
---|
34 | if(ep_rank_loc == local_root) |
---|
35 | { |
---|
36 | for(int i=0; i<num_ep; i++) |
---|
37 | memcpy(recvbuf + datasize * i * count, comm->my_buffer->void_buffer[i], datasize * count); |
---|
38 | } |
---|
39 | |
---|
40 | MPI_Barrier_local(comm); |
---|
41 | } |
---|
42 | |
---|
43 | int MPI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) |
---|
44 | { |
---|
45 | if(!comm->is_ep) return ::MPI_Gather(const_cast<void*>(sendbuf), sendcount, to_mpi_type(sendtype), recvbuf, recvcount, to_mpi_type(recvtype), |
---|
46 | root, to_mpi_comm(comm->mpi_comm)); |
---|
47 | if(comm->is_intercomm) return MPI_Gather_intercomm(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm); |
---|
48 | |
---|
49 | assert(sendcount == recvcount && sendtype == recvtype); |
---|
50 | |
---|
51 | int ep_rank = comm->ep_comm_ptr->size_rank_info[0].first; |
---|
52 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
53 | int mpi_rank = comm->ep_comm_ptr->size_rank_info[2].first; |
---|
54 | int ep_size = comm->ep_comm_ptr->size_rank_info[0].second; |
---|
55 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
56 | int mpi_size = comm->ep_comm_ptr->size_rank_info[2].second; |
---|
57 | |
---|
58 | int root_mpi_rank = comm->ep_rank_map->at(root).second; |
---|
59 | int root_ep_loc = comm->ep_rank_map->at(root).first; |
---|
60 | |
---|
61 | ::MPI_Aint datasize, lb; |
---|
62 | ::MPI_Type_get_extent(to_mpi_type(sendtype), &lb, &datasize); |
---|
63 | |
---|
64 | bool is_master = (ep_rank_loc==0 && mpi_rank != root_mpi_rank ) || ep_rank == root; |
---|
65 | bool is_root = ep_rank == root; |
---|
66 | |
---|
67 | void* local_recvbuf; |
---|
68 | |
---|
69 | if(is_master) |
---|
70 | { |
---|
71 | local_recvbuf = new void*[datasize * num_ep * sendcount]; |
---|
72 | } |
---|
73 | |
---|
74 | void* tmp_recvbuf; |
---|
75 | if(is_root) tmp_recvbuf = new void*[datasize * recvcount * ep_size]; |
---|
76 | |
---|
77 | |
---|
78 | if(mpi_rank == root_mpi_rank) MPI_Gather_local(sendbuf, sendcount, sendtype, local_recvbuf, root_ep_loc, comm); |
---|
79 | else MPI_Gather_local(sendbuf, sendcount, sendtype, local_recvbuf, 0, comm); |
---|
80 | |
---|
81 | std::vector<int> recvcounts(mpi_size, 0); |
---|
82 | std::vector<int> displs(mpi_size, 0); |
---|
83 | |
---|
84 | |
---|
85 | if(is_master) |
---|
86 | { |
---|
87 | for(int i=0; i<ep_size; i++) |
---|
88 | { |
---|
89 | recvcounts[comm->ep_rank_map->at(i).second]+=sendcount; |
---|
90 | } |
---|
91 | |
---|
92 | for(int i=1; i<mpi_size; i++) |
---|
93 | displs[i] = displs[i-1] + recvcounts[i-1]; |
---|
94 | |
---|
95 | ::MPI_Gatherv(local_recvbuf, sendcount*num_ep, to_mpi_type(sendtype), tmp_recvbuf, recvcounts.data(), displs.data(), to_mpi_type(recvtype), root_mpi_rank, to_mpi_comm(comm->mpi_comm)); |
---|
96 | } |
---|
97 | |
---|
98 | |
---|
99 | // reorder data |
---|
100 | if(is_root) |
---|
101 | { |
---|
102 | int offset; |
---|
103 | for(int i=0; i<ep_size; i++) |
---|
104 | { |
---|
105 | offset = displs[comm->ep_rank_map->at(i).second] + comm->ep_rank_map->at(i).first * sendcount; |
---|
106 | memcpy(recvbuf + i*sendcount*datasize, tmp_recvbuf+offset*datasize, sendcount*datasize); |
---|
107 | |
---|
108 | |
---|
109 | } |
---|
110 | |
---|
111 | } |
---|
112 | |
---|
113 | |
---|
114 | if(is_master) |
---|
115 | { |
---|
116 | delete[] local_recvbuf; |
---|
117 | } |
---|
118 | if(is_root) delete[] tmp_recvbuf; |
---|
119 | |
---|
120 | } |
---|
121 | |
---|
122 | |
---|
123 | int MPI_Gather_intercomm(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) |
---|
124 | { |
---|
125 | printf("MPI_Gather_intercomm not yet implemented\n"); |
---|
126 | MPI_Abort(comm, 0); |
---|
127 | } |
---|
128 | } |
---|
129 | #endif |
---|