1 | /*! |
---|
2 | \file ep_reduce.cpp |
---|
3 | \since 2 may 2016 |
---|
4 | |
---|
5 | \brief Definitions of MPI collective function: MPI_Reduce, MPI_Allreduce |
---|
6 | */ |
---|
7 | |
---|
8 | #include "ep_lib.hpp" |
---|
9 | #include <mpi.h> |
---|
10 | #include "ep_declaration.hpp" |
---|
11 | #include "ep_mpi.hpp" |
---|
12 | |
---|
13 | using namespace std; |
---|
14 | |
---|
15 | |
---|
16 | namespace ep_lib { |
---|
17 | |
---|
18 | template<typename T> |
---|
19 | T max_op(T a, T b) |
---|
20 | { |
---|
21 | return max(a,b); |
---|
22 | } |
---|
23 | |
---|
24 | template<typename T> |
---|
25 | T min_op(T a, T b) |
---|
26 | { |
---|
27 | return min(a,b); |
---|
28 | } |
---|
29 | |
---|
30 | template<typename T> |
---|
31 | T lor_op(T a, T b) |
---|
32 | { |
---|
33 | return a||b; |
---|
34 | } |
---|
35 | |
---|
36 | template<typename T> |
---|
37 | void reduce_max(const T * buffer, T* recvbuf, int count) |
---|
38 | { |
---|
39 | transform(buffer, buffer+count, recvbuf, recvbuf, max_op<T>); |
---|
40 | } |
---|
41 | |
---|
42 | template<typename T> |
---|
43 | void reduce_min(const T * buffer, T* recvbuf, int count) |
---|
44 | { |
---|
45 | transform(buffer, buffer+count, recvbuf, recvbuf, min_op<T>); |
---|
46 | } |
---|
47 | |
---|
48 | template<typename T> |
---|
49 | void reduce_sum(const T * buffer, T* recvbuf, int count) |
---|
50 | { |
---|
51 | transform(buffer, buffer+count, recvbuf, recvbuf, std::plus<T>()); |
---|
52 | } |
---|
53 | |
---|
54 | template<typename T> |
---|
55 | void reduce_lor(const T * buffer, T* recvbuf, int count) |
---|
56 | { |
---|
57 | transform(buffer, buffer+count, recvbuf, recvbuf, lor_op<T>); |
---|
58 | } |
---|
59 | |
---|
60 | int MPI_Reduce_local(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int local_root, MPI_Comm comm) |
---|
61 | { |
---|
62 | assert(valid_type(datatype)); |
---|
63 | assert(valid_op(op)); |
---|
64 | |
---|
65 | ::MPI_Aint datasize, lb; |
---|
66 | ::MPI_Type_get_extent(to_mpi_type(datatype), &lb, &datasize); |
---|
67 | |
---|
68 | int ep_rank = comm->ep_comm_ptr->size_rank_info[0].first; |
---|
69 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
70 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
71 | |
---|
72 | #pragma omp critical (_reduce) |
---|
73 | comm->my_buffer->void_buffer[ep_rank_loc] = const_cast< void* >(sendbuf); |
---|
74 | |
---|
75 | MPI_Barrier_local(comm); |
---|
76 | |
---|
77 | if(ep_rank_loc == local_root) |
---|
78 | { |
---|
79 | |
---|
80 | memcpy(recvbuf, comm->my_buffer->void_buffer[0], datasize * count); |
---|
81 | |
---|
82 | if(op == MPI_MAX) |
---|
83 | { |
---|
84 | if(datatype == MPI_INT) |
---|
85 | { |
---|
86 | assert(datasize == sizeof(int)); |
---|
87 | for(int i=1; i<num_ep; i++) |
---|
88 | reduce_max<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
89 | } |
---|
90 | |
---|
91 | else if(datatype == MPI_FLOAT) |
---|
92 | { |
---|
93 | assert(datasize == sizeof(float)); |
---|
94 | for(int i=1; i<num_ep; i++) |
---|
95 | reduce_max<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
96 | } |
---|
97 | |
---|
98 | else if(datatype == MPI_DOUBLE) |
---|
99 | { |
---|
100 | assert(datasize == sizeof(double)); |
---|
101 | for(int i=1; i<num_ep; i++) |
---|
102 | reduce_max<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
103 | } |
---|
104 | |
---|
105 | else if(datatype == MPI_CHAR) |
---|
106 | { |
---|
107 | assert(datasize == sizeof(char)); |
---|
108 | for(int i=1; i<num_ep; i++) |
---|
109 | reduce_max<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
110 | } |
---|
111 | |
---|
112 | else if(datatype == MPI_LONG) |
---|
113 | { |
---|
114 | assert(datasize == sizeof(long)); |
---|
115 | for(int i=1; i<num_ep; i++) |
---|
116 | reduce_max<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
117 | } |
---|
118 | |
---|
119 | else if(datatype == MPI_UNSIGNED_LONG) |
---|
120 | { |
---|
121 | assert(datasize == sizeof(unsigned long)); |
---|
122 | for(int i=1; i<num_ep; i++) |
---|
123 | reduce_max<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
124 | } |
---|
125 | |
---|
126 | else if(datatype == MPI_LONG_LONG_INT) |
---|
127 | { |
---|
128 | assert(datasize == sizeof(long long)); |
---|
129 | for(int i=1; i<num_ep; i++) |
---|
130 | reduce_max<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
131 | } |
---|
132 | |
---|
133 | else |
---|
134 | { |
---|
135 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
136 | MPI_Abort(comm, 0); |
---|
137 | } |
---|
138 | |
---|
139 | } |
---|
140 | |
---|
141 | else if(op == MPI_MIN) |
---|
142 | { |
---|
143 | if(datatype ==MPI_INT) |
---|
144 | { |
---|
145 | assert(datasize == sizeof(int)); |
---|
146 | for(int i=1; i<num_ep; i++) |
---|
147 | reduce_min<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
148 | } |
---|
149 | |
---|
150 | else if(datatype == MPI_FLOAT) |
---|
151 | { |
---|
152 | assert(datasize == sizeof(float)); |
---|
153 | for(int i=1; i<num_ep; i++) |
---|
154 | reduce_min<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
155 | } |
---|
156 | |
---|
157 | else if(datatype == MPI_DOUBLE) |
---|
158 | { |
---|
159 | assert(datasize == sizeof(double)); |
---|
160 | for(int i=1; i<num_ep; i++) |
---|
161 | reduce_min<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
162 | } |
---|
163 | |
---|
164 | else if(datatype == MPI_CHAR) |
---|
165 | { |
---|
166 | assert(datasize == sizeof(char)); |
---|
167 | for(int i=1; i<num_ep; i++) |
---|
168 | reduce_min<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
169 | } |
---|
170 | |
---|
171 | else if(datatype == MPI_LONG) |
---|
172 | { |
---|
173 | assert(datasize == sizeof(long)); |
---|
174 | for(int i=1; i<num_ep; i++) |
---|
175 | reduce_min<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
176 | } |
---|
177 | |
---|
178 | else if(datatype == MPI_UNSIGNED_LONG) |
---|
179 | { |
---|
180 | assert(datasize == sizeof(unsigned long)); |
---|
181 | for(int i=1; i<num_ep; i++) |
---|
182 | reduce_min<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
183 | } |
---|
184 | |
---|
185 | else if(datatype == MPI_LONG_LONG_INT) |
---|
186 | { |
---|
187 | assert(datasize == sizeof(long long)); |
---|
188 | for(int i=1; i<num_ep; i++) |
---|
189 | reduce_min<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
190 | } |
---|
191 | |
---|
192 | else |
---|
193 | { |
---|
194 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
195 | MPI_Abort(comm, 0); |
---|
196 | } |
---|
197 | |
---|
198 | } |
---|
199 | |
---|
200 | |
---|
201 | else if(op == MPI_SUM) |
---|
202 | { |
---|
203 | if(datatype==MPI_INT) |
---|
204 | { |
---|
205 | assert(datasize == sizeof(int)); |
---|
206 | for(int i=1; i<num_ep; i++) |
---|
207 | reduce_sum<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
208 | } |
---|
209 | |
---|
210 | else if(datatype == MPI_FLOAT) |
---|
211 | { |
---|
212 | assert(datasize == sizeof(float)); |
---|
213 | for(int i=1; i<num_ep; i++) |
---|
214 | reduce_sum<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
215 | } |
---|
216 | |
---|
217 | else if(datatype == MPI_DOUBLE) |
---|
218 | { |
---|
219 | assert(datasize == sizeof(double)); |
---|
220 | for(int i=1; i<num_ep; i++) |
---|
221 | reduce_sum<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
222 | } |
---|
223 | |
---|
224 | else if(datatype == MPI_CHAR) |
---|
225 | { |
---|
226 | assert(datasize == sizeof(char)); |
---|
227 | for(int i=1; i<num_ep; i++) |
---|
228 | reduce_sum<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
229 | } |
---|
230 | |
---|
231 | else if(datatype == MPI_LONG) |
---|
232 | { |
---|
233 | assert(datasize == sizeof(long)); |
---|
234 | for(int i=1; i<num_ep; i++) |
---|
235 | reduce_sum<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
236 | } |
---|
237 | |
---|
238 | else if(datatype ==MPI_UNSIGNED_LONG) |
---|
239 | { |
---|
240 | assert(datasize == sizeof(unsigned long)); |
---|
241 | for(int i=1; i<num_ep; i++) |
---|
242 | reduce_sum<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
243 | } |
---|
244 | |
---|
245 | else if(datatype ==MPI_LONG_LONG_INT) |
---|
246 | { |
---|
247 | assert(datasize == sizeof(long long)); |
---|
248 | for(int i=1; i<num_ep; i++) |
---|
249 | reduce_sum<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
250 | } |
---|
251 | |
---|
252 | else |
---|
253 | { |
---|
254 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
255 | MPI_Abort(comm, 0); |
---|
256 | } |
---|
257 | |
---|
258 | } |
---|
259 | |
---|
260 | else if(op == MPI_LOR) |
---|
261 | { |
---|
262 | if(datatype != MPI_INT) |
---|
263 | printf("datatype Error, must be MPI_INT\n"); |
---|
264 | else |
---|
265 | { |
---|
266 | assert(datasize == sizeof(int)); |
---|
267 | for(int i=1; i<num_ep; i++) |
---|
268 | reduce_lor<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
269 | } |
---|
270 | } |
---|
271 | |
---|
272 | else |
---|
273 | { |
---|
274 | printf("op type Error in ep_reduce : MPI_MAX, MPI_MIN, MPI_SUM, MPI_LOR\n"); |
---|
275 | MPI_Abort(comm, 0); |
---|
276 | } |
---|
277 | } |
---|
278 | |
---|
279 | MPI_Barrier_local(comm); |
---|
280 | |
---|
281 | } |
---|
282 | |
---|
283 | |
---|
284 | int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) |
---|
285 | { |
---|
286 | |
---|
287 | if(!comm->is_ep) return ::MPI_Reduce(sendbuf, recvbuf, count, to_mpi_type(datatype), to_mpi_op(op), root, to_mpi_comm(comm->mpi_comm)); |
---|
288 | if(comm->is_intercomm) return MPI_Reduce_intercomm(sendbuf, recvbuf, count, datatype, op, root, comm); |
---|
289 | |
---|
290 | |
---|
291 | |
---|
292 | int ep_rank = comm->ep_comm_ptr->size_rank_info[0].first; |
---|
293 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
294 | int mpi_rank = comm->ep_comm_ptr->size_rank_info[2].first; |
---|
295 | int ep_size = comm->ep_comm_ptr->size_rank_info[0].second; |
---|
296 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
297 | int mpi_size = comm->ep_comm_ptr->size_rank_info[2].second; |
---|
298 | |
---|
299 | int root_mpi_rank = comm->ep_rank_map->at(root).second; |
---|
300 | int root_ep_loc = comm->ep_rank_map->at(root).first; |
---|
301 | |
---|
302 | ::MPI_Aint datasize, lb; |
---|
303 | |
---|
304 | ::MPI_Type_get_extent(*(static_cast< ::MPI_Datatype*>(datatype)), &lb, &datasize); |
---|
305 | |
---|
306 | bool is_master = (ep_rank_loc==0 && mpi_rank != root_mpi_rank ) || ep_rank == root; |
---|
307 | bool is_root = ep_rank == root; |
---|
308 | |
---|
309 | void* local_recvbuf; |
---|
310 | |
---|
311 | if(is_master) |
---|
312 | { |
---|
313 | local_recvbuf = new void*[datasize * count]; |
---|
314 | } |
---|
315 | |
---|
316 | if(mpi_rank == root_mpi_rank) MPI_Reduce_local(sendbuf, local_recvbuf, count, datatype, op, root_ep_loc, comm); |
---|
317 | else MPI_Reduce_local(sendbuf, local_recvbuf, count, datatype, op, 0, comm); |
---|
318 | |
---|
319 | |
---|
320 | |
---|
321 | if(is_master) |
---|
322 | { |
---|
323 | ::MPI_Reduce(local_recvbuf, recvbuf, count, to_mpi_type(datatype), to_mpi_op(op), root_mpi_rank, to_mpi_comm(comm->mpi_comm)); |
---|
324 | |
---|
325 | } |
---|
326 | |
---|
327 | if(is_master) |
---|
328 | { |
---|
329 | delete[] local_recvbuf; |
---|
330 | } |
---|
331 | |
---|
332 | MPI_Barrier_local(comm); |
---|
333 | } |
---|
334 | |
---|
335 | |
---|
336 | int MPI_Reduce_intercomm(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) |
---|
337 | { |
---|
338 | printf("MPI_Reduce_intercomm not yet implemented\n"); |
---|
339 | MPI_Abort(comm, 0); |
---|
340 | } |
---|
341 | } |
---|
342 | |
---|