| 1 | /*
|
|---|
| 2 | * vmg - a versatile multigrid solver
|
|---|
| 3 | * Copyright (C) 2012 Institute for Numerical Simulation, University of Bonn
|
|---|
| 4 | *
|
|---|
| 5 | * vmg is free software: you can redistribute it and/or modify
|
|---|
| 6 | * it under the terms of the GNU General Public License as published by
|
|---|
| 7 | * the Free Software Foundation, either version 3 of the License, or
|
|---|
| 8 | * (at your option) any later version.
|
|---|
| 9 | *
|
|---|
| 10 | * vmg is distributed in the hope that it will be useful,
|
|---|
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|---|
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|---|
| 13 | * GNU General Public License for more details.
|
|---|
| 14 | *
|
|---|
| 15 | * You should have received a copy of the GNU General Public License
|
|---|
| 16 | * along with this program. If not, see <http://www.gnu.org/licenses/>.
|
|---|
| 17 | */
|
|---|
| 18 |
|
|---|
| 19 | #ifndef DATATYPES_LOCAL_HPP_
|
|---|
| 20 | #define DATATYPES_LOCAL_HPP_
|
|---|
| 21 |
|
|---|
| 22 | #include <vector>
|
|---|
| 23 |
|
|---|
| 24 | #include "base/index.hpp"
|
|---|
| 25 | #include "comm/mpi/datatype.hpp"
|
|---|
| 26 |
|
|---|
| 27 | namespace VMG
|
|---|
| 28 | {
|
|---|
| 29 |
|
|---|
| 30 | namespace MPI
|
|---|
| 31 | {
|
|---|
| 32 |
|
|---|
| 33 | class DatatypesLocal
|
|---|
| 34 | {
|
|---|
| 35 | public:
|
|---|
| 36 | template <class T>
|
|---|
| 37 | DatatypesLocal(const T& grid, const MPI_Comm& comm, const bool& alloc)
|
|---|
| 38 | {
|
|---|
| 39 | InitDatatypesLocal(grid, comm, alloc);
|
|---|
| 40 | }
|
|---|
| 41 |
|
|---|
| 42 | ~DatatypesLocal() {}
|
|---|
| 43 |
|
|---|
| 44 | std::vector<Datatype>& Halo() {return _halo;}
|
|---|
| 45 | std::vector<Datatype>& NB() {return _nb;}
|
|---|
| 46 |
|
|---|
| 47 | const std::vector<Datatype>& Halo() const {return _halo;}
|
|---|
| 48 | const std::vector<Datatype>& NB() const {return _nb;}
|
|---|
| 49 |
|
|---|
| 50 | const std::vector<Index>& Offset() const {return _offset;}
|
|---|
| 51 |
|
|---|
| 52 | private:
|
|---|
| 53 | template <class T>
|
|---|
| 54 | void InitDatatypesLocal(const T& grid, const MPI_Comm& comm, const bool& alloc_buffer);
|
|---|
| 55 |
|
|---|
| 56 | std::vector<Datatype> _halo, _nb;
|
|---|
| 57 | std::vector<Index> _offset;
|
|---|
| 58 | };
|
|---|
| 59 |
|
|---|
| 60 | namespace
|
|---|
| 61 | {
|
|---|
| 62 |
|
|---|
| 63 | inline int to_1d(const Index& i)
|
|---|
| 64 | {
|
|---|
| 65 | return i.Z()+3*(i.Y()+3*i.X());
|
|---|
| 66 | }
|
|---|
| 67 |
|
|---|
| 68 | inline bool _is_valid(const Index& coord, const Index& dims, const Index& periods)
|
|---|
| 69 | {
|
|---|
| 70 | return (periods[0] || (coord[0] >= 0 && coord[0] < dims[0])) &&
|
|---|
| 71 | (periods[1] || (coord[1] >= 0 && coord[1] < dims[1])) &&
|
|---|
| 72 | (periods[2] || (coord[2] >= 0 && coord[2] < dims[2]));
|
|---|
| 73 | }
|
|---|
| 74 |
|
|---|
| 75 | }
|
|---|
| 76 |
|
|---|
| 77 | template <class T>
|
|---|
| 78 | void DatatypesLocal::InitDatatypesLocal(const T& grid, const MPI_Comm& comm, const bool& alloc_buffer)
|
|---|
| 79 | {
|
|---|
| 80 | if (comm != MPI_COMM_NULL) {
|
|---|
| 81 |
|
|---|
| 82 | int index, ranks[27];
|
|---|
| 83 | Index dims, periods, coords;
|
|---|
| 84 | Index sizes, subsizes, starts;
|
|---|
| 85 | Index offset, i;
|
|---|
| 86 | const LocalIndices& l = grid.Local();
|
|---|
| 87 |
|
|---|
| 88 | MPI_Cart_get(comm, 3, dims.vec(), periods.vec(), coords.vec());
|
|---|
| 89 |
|
|---|
| 90 | for (i.X()=-1; i.X()<=1; ++i.X())
|
|---|
| 91 | for (i.Y()=-1; i.Y()<=1; ++i.Y())
|
|---|
| 92 | for (i.Z()=-1; i.Z()<=1; ++i.Z())
|
|---|
| 93 | if (_is_valid(coords + i, dims, periods))
|
|---|
| 94 | MPI_Cart_rank(comm, (coords + i).vec(), &ranks[to_1d(i+1)]);
|
|---|
| 95 |
|
|---|
| 96 | sizes = l.SizeTotal();
|
|---|
| 97 |
|
|---|
| 98 | /* -1 0 0 */
|
|---|
| 99 | offset = Index(-1,0,0);
|
|---|
| 100 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 101 | index = to_1d(offset+1);
|
|---|
| 102 | subsizes = Index(l.HaloSize1().X(), l.Size().Y(), l.Size().Z());
|
|---|
| 103 | starts = Index(0, l.Begin().Y(), l.Begin().Z());
|
|---|
| 104 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 0, 1, alloc_buffer));
|
|---|
| 105 | starts = l.Begin();
|
|---|
| 106 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 0, 1, alloc_buffer));
|
|---|
| 107 | _offset.push_back(offset);
|
|---|
| 108 | }
|
|---|
| 109 |
|
|---|
| 110 | /* 1 0 0 */
|
|---|
| 111 | offset = Index(1,0,0);
|
|---|
| 112 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 113 | index = to_1d(offset+1);
|
|---|
| 114 | subsizes = Index(l.HaloSize2().X(), l.Size().Y(), l.Size().Z());
|
|---|
| 115 | starts = Index(l.End().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 116 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 1, 0, alloc_buffer));
|
|---|
| 117 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 118 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 1, 0, alloc_buffer));
|
|---|
| 119 | _offset.push_back(offset);
|
|---|
| 120 | }
|
|---|
| 121 |
|
|---|
| 122 | /* 0 -1 0 */
|
|---|
| 123 | offset = Index(0,-1,0);
|
|---|
| 124 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 125 | index = to_1d(offset+1);
|
|---|
| 126 | subsizes = Index(l.Size().X(), l.HaloSize1().Y(), l.Size().Z());
|
|---|
| 127 | starts = Index(l.Begin().X(), 0, l.Begin().Z());
|
|---|
| 128 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 2, 3, alloc_buffer));
|
|---|
| 129 | starts = l.Begin();
|
|---|
| 130 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 2, 3, alloc_buffer));
|
|---|
| 131 | _offset.push_back(offset);
|
|---|
| 132 | }
|
|---|
| 133 |
|
|---|
| 134 | /* 0 1 0 */
|
|---|
| 135 | offset = Index(0,1,0);
|
|---|
| 136 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 137 | index = to_1d(offset+1);
|
|---|
| 138 | subsizes = Index(l.Size().X(), l.HaloSize2().Y(), l.Size().Z());
|
|---|
| 139 | starts = Index(l.Begin().X(), l.End().Y(), l.Begin().Z());
|
|---|
| 140 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 3, 2, alloc_buffer));
|
|---|
| 141 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 142 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 3, 2, alloc_buffer));
|
|---|
| 143 | _offset.push_back(offset);
|
|---|
| 144 | }
|
|---|
| 145 |
|
|---|
| 146 | /* 0 0 -1 */
|
|---|
| 147 | offset = Index(0,0,-1);
|
|---|
| 148 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 149 | index = to_1d(offset+1);
|
|---|
| 150 | subsizes = Index(l.Size().X(), l.Size().Y(), l.HaloSize1().Z());
|
|---|
| 151 | starts = Index(l.Begin().X(), l.Begin().Y(), 0);
|
|---|
| 152 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 4, 5, alloc_buffer));
|
|---|
| 153 | starts = l.Begin();
|
|---|
| 154 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 4, 5, alloc_buffer));
|
|---|
| 155 | _offset.push_back(offset);
|
|---|
| 156 | }
|
|---|
| 157 |
|
|---|
| 158 | /* 0 0 1 */
|
|---|
| 159 | offset = Index(0,0,1);
|
|---|
| 160 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 161 | index = to_1d(offset+1);
|
|---|
| 162 | subsizes = Index(l.Size().X(), l.Size().Y(), l.HaloSize2().Z());
|
|---|
| 163 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z());
|
|---|
| 164 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 5, 4, alloc_buffer));
|
|---|
| 165 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 166 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 5, 4, alloc_buffer));
|
|---|
| 167 | _offset.push_back(offset);
|
|---|
| 168 | }
|
|---|
| 169 |
|
|---|
| 170 | /* -1 -1 0 */
|
|---|
| 171 | offset = Index(-1,-1,0);
|
|---|
| 172 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 173 | index = to_1d(offset+1);
|
|---|
| 174 | subsizes = Index(l.HaloSize1().X(), l.HaloSize1().Y(), l.Size().Z());
|
|---|
| 175 | starts = Index(0, 0, l.Begin().Z());
|
|---|
| 176 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 6, 7, alloc_buffer));
|
|---|
| 177 | starts = l.Begin();
|
|---|
| 178 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 6, 7, alloc_buffer));
|
|---|
| 179 | _offset.push_back(offset);
|
|---|
| 180 | }
|
|---|
| 181 |
|
|---|
| 182 | /* -1 1 0 */
|
|---|
| 183 | offset = Index(-1,1,0);
|
|---|
| 184 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 185 | index = to_1d(offset+1);
|
|---|
| 186 | subsizes = Index(l.HaloSize1().X(), l.HaloSize2().Y(), l.Size().Z());
|
|---|
| 187 | starts = Index(0, l.End().Y(), l.Begin().Z());
|
|---|
| 188 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 8, 9, alloc_buffer));
|
|---|
| 189 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 190 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 8, 9, alloc_buffer));
|
|---|
| 191 | _offset.push_back(offset);
|
|---|
| 192 | }
|
|---|
| 193 |
|
|---|
| 194 | /* 1 -1 0 */
|
|---|
| 195 | offset = Index(1,-1,0);
|
|---|
| 196 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 197 | index = to_1d(offset+1);
|
|---|
| 198 | subsizes = Index(l.HaloSize2().X(), l.HaloSize1().Y(), l.Size().Z());
|
|---|
| 199 | starts = Index(l.End().X(), 0, l.Begin().Z());
|
|---|
| 200 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 9, 8, alloc_buffer));
|
|---|
| 201 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 202 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 9, 8, alloc_buffer));
|
|---|
| 203 | _offset.push_back(offset);
|
|---|
| 204 | }
|
|---|
| 205 |
|
|---|
| 206 | /* 1 1 0 */
|
|---|
| 207 | offset = Index(1,1,0);
|
|---|
| 208 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 209 | index = to_1d(offset+1);
|
|---|
| 210 | subsizes = Index(l.HaloSize2().X(), l.HaloSize2().Y(), l.Size().Z());
|
|---|
| 211 | starts = Index(l.End().X(), l.End().Y(), l.Begin().Z());
|
|---|
| 212 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 7, 6, alloc_buffer));
|
|---|
| 213 | starts = Index(l.End().X()-l.HaloSize2().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 214 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 7, 6, alloc_buffer));
|
|---|
| 215 | _offset.push_back(offset);
|
|---|
| 216 | }
|
|---|
| 217 |
|
|---|
| 218 | /* -1 0 -1 */
|
|---|
| 219 | offset = Index(-1,0,-1);
|
|---|
| 220 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 221 | index = to_1d(offset+1);
|
|---|
| 222 | subsizes = Index(l.HaloSize1().X(), l.Size().Y(), l.HaloSize1().Z());
|
|---|
| 223 | starts = Index(0, l.Begin().Y(), 0);
|
|---|
| 224 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 10, 11, alloc_buffer));
|
|---|
| 225 | starts = l.Begin();
|
|---|
| 226 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 10, 11, alloc_buffer));
|
|---|
| 227 | _offset.push_back(offset);
|
|---|
| 228 | }
|
|---|
| 229 |
|
|---|
| 230 | /* -1 0 1 */
|
|---|
| 231 | offset = Index(-1,0,1);
|
|---|
| 232 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 233 | index = to_1d(offset+1);
|
|---|
| 234 | subsizes = Index(l.HaloSize1().X(), l.Size().Y(), l.HaloSize2().Z());
|
|---|
| 235 | starts = Index(0, l.Begin().Y(), l.End().Z());
|
|---|
| 236 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 12, 13, alloc_buffer));
|
|---|
| 237 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 238 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 12, 13, alloc_buffer));
|
|---|
| 239 | _offset.push_back(offset);
|
|---|
| 240 | }
|
|---|
| 241 |
|
|---|
| 242 | /* 1 0 -1 */
|
|---|
| 243 | offset = Index(1,0,-1);
|
|---|
| 244 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 245 | index = to_1d(offset+1);
|
|---|
| 246 | subsizes = Index(l.HaloSize2().X(), l.Size().Y(), l.HaloSize1().Z());
|
|---|
| 247 | starts = Index(l.End().X(), l.Begin().Y(), 0);
|
|---|
| 248 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 13, 12, alloc_buffer));
|
|---|
| 249 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 250 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 13, 12, alloc_buffer));
|
|---|
| 251 | _offset.push_back(offset);
|
|---|
| 252 | }
|
|---|
| 253 |
|
|---|
| 254 | /* 1 0 1 */
|
|---|
| 255 | offset = Index(1,0,1);
|
|---|
| 256 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 257 | index = to_1d(offset+1);
|
|---|
| 258 | subsizes = Index(l.HaloSize2().X(), l.Size().Y(), l.HaloSize2().Z());
|
|---|
| 259 | starts = Index(l.End().X(), l.Begin().Y(), l.End().Z());
|
|---|
| 260 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 11, 10, alloc_buffer));
|
|---|
| 261 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 262 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 11, 10, alloc_buffer));
|
|---|
| 263 | _offset.push_back(offset);
|
|---|
| 264 | }
|
|---|
| 265 |
|
|---|
| 266 | /* 0 -1 -1 */
|
|---|
| 267 | offset = Index(0,-1,-1);
|
|---|
| 268 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 269 | index = to_1d(offset+1);
|
|---|
| 270 | subsizes = Index(l.Size().X(), l.HaloSize1().Y(), l.HaloSize1().Z());
|
|---|
| 271 | starts = Index(l.Begin().X(), 0, 0);
|
|---|
| 272 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 14, 15, alloc_buffer));
|
|---|
| 273 | starts = l.Begin();
|
|---|
| 274 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 14, 15, alloc_buffer));
|
|---|
| 275 | _offset.push_back(offset);
|
|---|
| 276 | }
|
|---|
| 277 |
|
|---|
| 278 | /* 0 -1 1 */
|
|---|
| 279 | offset = Index(0,-1,1);
|
|---|
| 280 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 281 | index = to_1d(offset+1);
|
|---|
| 282 | subsizes = Index(l.Size().X(), l.HaloSize1().Y(), l.HaloSize2().Z());
|
|---|
| 283 | starts = Index(l.Begin().X(), 0, l.End().Z());
|
|---|
| 284 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 16, 17, alloc_buffer));
|
|---|
| 285 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 286 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 16, 17, alloc_buffer));
|
|---|
| 287 | _offset.push_back(offset);
|
|---|
| 288 | }
|
|---|
| 289 |
|
|---|
| 290 | /* 0 1 -1 */
|
|---|
| 291 | offset = Index(0,1,-1);
|
|---|
| 292 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 293 | index = to_1d(offset+1);
|
|---|
| 294 | subsizes = Index(l.Size().X(), l.HaloSize2().Y(), l.HaloSize1().Z());
|
|---|
| 295 | starts = Index(l.Begin().X(), l.End().Y(), 0);
|
|---|
| 296 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 17, 16, alloc_buffer));
|
|---|
| 297 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 298 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 17, 16, alloc_buffer));
|
|---|
| 299 | _offset.push_back(offset);
|
|---|
| 300 | }
|
|---|
| 301 |
|
|---|
| 302 | /* 0 1 1 */
|
|---|
| 303 | offset = Index(0,1,1);
|
|---|
| 304 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 305 | index = to_1d(offset+1);
|
|---|
| 306 | subsizes = Index(l.Size().X(), l.HaloSize2().Y(), l.HaloSize2().Z());
|
|---|
| 307 | starts = Index(l.Begin().X(), l.End().Y(), l.End().Z());
|
|---|
| 308 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 15, 14, alloc_buffer));
|
|---|
| 309 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 310 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 15, 14, alloc_buffer));
|
|---|
| 311 | _offset.push_back(offset);
|
|---|
| 312 | }
|
|---|
| 313 |
|
|---|
| 314 | /* -1 -1 -1 */
|
|---|
| 315 | offset = Index(-1,-1,-1);
|
|---|
| 316 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 317 | index = to_1d(offset+1);
|
|---|
| 318 | subsizes = l.HaloSize1();
|
|---|
| 319 | starts = 0;
|
|---|
| 320 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 18, 19, alloc_buffer));
|
|---|
| 321 | starts = l.Begin();
|
|---|
| 322 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 18, 19, alloc_buffer));
|
|---|
| 323 | _offset.push_back(offset);
|
|---|
| 324 | }
|
|---|
| 325 |
|
|---|
| 326 | /* -1 -1 1 */
|
|---|
| 327 | offset = Index(-1,-1,1);
|
|---|
| 328 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 329 | index = to_1d(offset+1);
|
|---|
| 330 | subsizes = Index(l.HaloSize1().X(), l.HaloSize1().Y(), l.HaloSize2().Z());
|
|---|
| 331 | starts = Index(0, 0, l.End().Z());;
|
|---|
| 332 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 20, 21, alloc_buffer));
|
|---|
| 333 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 334 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 20, 21, alloc_buffer));
|
|---|
| 335 | _offset.push_back(offset);
|
|---|
| 336 | }
|
|---|
| 337 |
|
|---|
| 338 | /* -1 1 -1 */
|
|---|
| 339 | offset = Index(-1,1,-1);
|
|---|
| 340 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 341 | index = to_1d(offset+1);
|
|---|
| 342 | subsizes = Index(l.HaloSize1().X(), l.HaloSize2().Y(), l.HaloSize1().Z());
|
|---|
| 343 | starts = Index(0, l.End().Y(), 0);
|
|---|
| 344 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 22, 23, alloc_buffer));
|
|---|
| 345 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 346 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 22, 23, alloc_buffer));
|
|---|
| 347 | _offset.push_back(offset);
|
|---|
| 348 | }
|
|---|
| 349 |
|
|---|
| 350 | /* 1 -1 -1 */
|
|---|
| 351 | offset = Index(1,-1,-1);
|
|---|
| 352 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 353 | index = to_1d(offset+1);
|
|---|
| 354 | subsizes = Index(l.HaloSize2().X(), l.HaloSize1().Y(), l.HaloSize1().Z());
|
|---|
| 355 | starts = Index(l.End().X(), 0, 0);
|
|---|
| 356 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 24, 25, alloc_buffer));
|
|---|
| 357 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 358 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 24, 25, alloc_buffer));
|
|---|
| 359 | _offset.push_back(offset);
|
|---|
| 360 | }
|
|---|
| 361 |
|
|---|
| 362 | /* -1 1 1 */
|
|---|
| 363 | offset = Index(-1,1,1);
|
|---|
| 364 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 365 | index = to_1d(offset+1);
|
|---|
| 366 | subsizes = Index(l.HaloSize1().X(), l.HaloSize2().Y(), l.HaloSize2().Z());
|
|---|
| 367 | starts = Index(0, l.End().Y(), l.End().Z());
|
|---|
| 368 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 25, 24, alloc_buffer));
|
|---|
| 369 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloSize2().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 370 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 25, 24, alloc_buffer));
|
|---|
| 371 | _offset.push_back(offset);
|
|---|
| 372 | }
|
|---|
| 373 |
|
|---|
| 374 | /* 1 -1 1 */
|
|---|
| 375 | offset = Index(1,-1,1);
|
|---|
| 376 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 377 | index = to_1d(offset+1);
|
|---|
| 378 | subsizes = Index(l.HaloSize2().X(), l.HaloSize1().Y(), l.HaloSize2().Z());
|
|---|
| 379 | starts = Index(l.End().X(), 0, l.End().Z());
|
|---|
| 380 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 23, 22, alloc_buffer));
|
|---|
| 381 | starts = Index(l.End().X()-l.HaloSize2().X(), l.Begin().Y(), l.End().Z()-l.HaloSize2().Z());
|
|---|
| 382 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 23, 22, alloc_buffer));
|
|---|
| 383 | _offset.push_back(offset);
|
|---|
| 384 | }
|
|---|
| 385 |
|
|---|
| 386 | /* 1 1 -1 */
|
|---|
| 387 | offset = Index(1,1,-1);
|
|---|
| 388 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 389 | index = to_1d(offset+1);
|
|---|
| 390 | subsizes = Index(l.HaloSize2().X(), l.HaloSize2().Y(), l.HaloSize1().Z());
|
|---|
| 391 | starts = Index(l.End().X(), l.End().Y(), 0);
|
|---|
| 392 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 21, 20, alloc_buffer));
|
|---|
| 393 | starts = Index(l.End().X()-l.HaloSize2().X(), l.End().Y()-l.HaloSize2().Y(), l.Begin().Z());
|
|---|
| 394 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 21, 20, alloc_buffer));
|
|---|
| 395 | _offset.push_back(offset);
|
|---|
| 396 | }
|
|---|
| 397 |
|
|---|
| 398 | /* 1 1 1 */
|
|---|
| 399 | offset = Index(1,1,1);
|
|---|
| 400 | if (_is_valid(coords + offset, dims, periods)) {
|
|---|
| 401 | index = to_1d(offset+1);
|
|---|
| 402 | subsizes = l.HaloSize2();
|
|---|
| 403 | starts = l.End();
|
|---|
| 404 | _halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 19, 18, alloc_buffer));
|
|---|
| 405 | starts = l.End()-l.HaloSize2();
|
|---|
| 406 | _nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 19, 18, alloc_buffer));
|
|---|
| 407 | _offset.push_back(offset);
|
|---|
| 408 | }
|
|---|
| 409 |
|
|---|
| 410 | }
|
|---|
| 411 | }
|
|---|
| 412 |
|
|---|
| 413 | }
|
|---|
| 414 |
|
|---|
| 415 | }
|
|---|
| 416 |
|
|---|
| 417 | #endif /* DATATYPES_LOCAL_HPP_ */
|
|---|