| 1 | #ifdef HAVE_CONFIG_H
|
|---|
| 2 | #include <config.h>
|
|---|
| 3 | #endif
|
|---|
| 4 |
|
|---|
| 5 | #ifdef HAVE_MPI
|
|---|
| 6 | #include <mpi.h>
|
|---|
| 7 | #else
|
|---|
| 8 | #error MPI is needed to compile VMG::MPI::DatatypesLocal
|
|---|
| 9 | #endif
|
|---|
| 10 |
|
|---|
| 11 | #include <map>
|
|---|
| 12 |
|
|---|
| 13 | #include "base/index.hpp"
|
|---|
| 14 | #include "comm/mpi/datatypes_local.hpp"
|
|---|
| 15 | #include "comm/mpi/key.hpp"
|
|---|
| 16 | #include "grid/grid.hpp"
|
|---|
| 17 |
|
|---|
| 18 | using namespace VMG;
|
|---|
| 19 |
|
|---|
| 20 | static inline int to_1d(const int& x, const int& y, const int& z)
|
|---|
| 21 | {
|
|---|
| 22 | return z+3*(y+3*x);
|
|---|
| 23 | }
|
|---|
| 24 |
|
|---|
| 25 | VMG::MPI::DatatypesLocal::DatatypesLocal(const Grid& grid, const MPI_Comm& comm)
|
|---|
| 26 | {
|
|---|
| 27 | if (comm != MPI_COMM_NULL) {
|
|---|
| 28 |
|
|---|
| 29 | int index, my_rank, ranks[27];
|
|---|
| 30 | Index my_coords;
|
|---|
| 31 | Index sizes, subsizes, starts;
|
|---|
| 32 | const LocalIndices& l = grid.Local();
|
|---|
| 33 |
|
|---|
| 34 | MPI_Comm_rank(comm, &my_rank);
|
|---|
| 35 | MPI_Cart_coords(comm, my_rank, 3, my_coords.vec());
|
|---|
| 36 |
|
|---|
| 37 | for (int i=-1; i<=1; ++i)
|
|---|
| 38 | for (int j=-1; j<=1; ++j)
|
|---|
| 39 | for (int k=-1; k<=1; ++k)
|
|---|
| 40 | MPI_Cart_rank(comm, (my_coords+Index(i,j,k)).vec(), &ranks[to_1d(i+1,j+1,k+1)]);
|
|---|
| 41 |
|
|---|
| 42 | sizes = l.SizeTotal();
|
|---|
| 43 |
|
|---|
| 44 | /* -1 0 0 */
|
|---|
| 45 | index = to_1d(0,1,1);
|
|---|
| 46 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.Size().Y(), l.Size().Z());
|
|---|
| 47 | starts = Index(0, l.Begin().Y(), l.Begin().Z());
|
|---|
| 48 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 0, 1));
|
|---|
| 49 | starts = l.Begin();
|
|---|
| 50 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 0, 1));
|
|---|
| 51 |
|
|---|
| 52 | /* 1 0 0 */
|
|---|
| 53 | index = to_1d(2,1,1);
|
|---|
| 54 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.Size().Y(), l.Size().Z());
|
|---|
| 55 | starts = Index(l.End().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 56 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 1, 0));
|
|---|
| 57 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 58 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 1, 0));
|
|---|
| 59 |
|
|---|
| 60 | /* 0 -1 0 */
|
|---|
| 61 | index = to_1d(1,0,1);
|
|---|
| 62 | subsizes = Index(l.Size().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.Size().Z());
|
|---|
| 63 | starts = Index(l.Begin().X(), 0, l.Begin().Z());
|
|---|
| 64 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 2, 3));
|
|---|
| 65 | starts = l.Begin();
|
|---|
| 66 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 2, 3));
|
|---|
| 67 |
|
|---|
| 68 | /* 0 1 0 */
|
|---|
| 69 | index = to_1d(1,2,1);
|
|---|
| 70 | subsizes = Index(l.Size().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.Size().Z());
|
|---|
| 71 | starts = Index(l.Begin().X(), l.End().Y(), l.Begin().Z());
|
|---|
| 72 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 3, 2));
|
|---|
| 73 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 74 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 3, 2));
|
|---|
| 75 |
|
|---|
| 76 | /* 0 0 -1 */
|
|---|
| 77 | index = to_1d(1,1,0);
|
|---|
| 78 | subsizes = Index(l.Size().X(), l.Size().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 79 | starts = Index(l.Begin().X(), l.Begin().Y(), 0);
|
|---|
| 80 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 4, 5));
|
|---|
| 81 | starts = l.Begin();
|
|---|
| 82 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 4, 5));
|
|---|
| 83 |
|
|---|
| 84 | /* 0 0 1 */
|
|---|
| 85 | index = to_1d(1,1,2);
|
|---|
| 86 | subsizes = Index(l.Size().X(), l.Size().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 87 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z());
|
|---|
| 88 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 5, 4));
|
|---|
| 89 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 90 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 5, 4));
|
|---|
| 91 |
|
|---|
| 92 | /* -1 -1 0 */
|
|---|
| 93 | index = to_1d(0,0,1);
|
|---|
| 94 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.Size().Z());
|
|---|
| 95 | starts = Index(0, 0, l.Begin().Z());
|
|---|
| 96 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 6, 7));
|
|---|
| 97 | starts = l.Begin();
|
|---|
| 98 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 6, 7));
|
|---|
| 99 |
|
|---|
| 100 | /* -1 1 0 */
|
|---|
| 101 | index = to_1d(0,2,1);
|
|---|
| 102 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.Size().Z());
|
|---|
| 103 | starts = Index(0, l.End().Y(), l.Begin().Z());
|
|---|
| 104 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 8, 9));
|
|---|
| 105 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 106 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 8, 9));
|
|---|
| 107 |
|
|---|
| 108 | /* 1 -1 0 */
|
|---|
| 109 | index = to_1d(2,0,1);
|
|---|
| 110 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.Size().Z());
|
|---|
| 111 | starts = Index(l.End().X(), 0, l.Begin().Z());
|
|---|
| 112 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 9, 8));
|
|---|
| 113 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 114 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 9, 8));
|
|---|
| 115 |
|
|---|
| 116 | /* 1 1 0 */
|
|---|
| 117 | index = to_1d(2,2,1);
|
|---|
| 118 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.Size().Z());
|
|---|
| 119 | starts = Index(l.End().X(), l.End().Y(), l.Begin().Z());
|
|---|
| 120 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 7, 6));
|
|---|
| 121 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 122 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 7, 6));
|
|---|
| 123 |
|
|---|
| 124 | /* -1 0 -1 */
|
|---|
| 125 | index = to_1d(0,1,0);
|
|---|
| 126 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.Size().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 127 | starts = Index(0, l.Begin().Y(), 0);
|
|---|
| 128 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 10, 11));
|
|---|
| 129 | starts = l.Begin();
|
|---|
| 130 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 10, 11));
|
|---|
| 131 |
|
|---|
| 132 | /* -1 0 1 */
|
|---|
| 133 | index = to_1d(0,1,2);
|
|---|
| 134 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.Size().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 135 | starts = Index(0, l.Begin().Y(), l.End().Z());
|
|---|
| 136 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 12, 13));
|
|---|
| 137 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 138 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 12, 13));
|
|---|
| 139 |
|
|---|
| 140 | /* 1 0 -1 */
|
|---|
| 141 | index = to_1d(2,1,0);
|
|---|
| 142 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.Size().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 143 | starts = Index(l.End().X(), l.Begin().Y(), 0);
|
|---|
| 144 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 13, 12));
|
|---|
| 145 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 146 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 13, 12));
|
|---|
| 147 |
|
|---|
| 148 | /* 1 0 1 */
|
|---|
| 149 | index = to_1d(2,1,2);
|
|---|
| 150 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.Size().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 151 | starts = Index(l.End().X(), l.Begin().Y(), l.End().Z());
|
|---|
| 152 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 11, 10));
|
|---|
| 153 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 154 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 11, 10));
|
|---|
| 155 |
|
|---|
| 156 | /* 0 -1 -1 */
|
|---|
| 157 | index = to_1d(1,0,0);
|
|---|
| 158 | subsizes = Index(l.Size().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 159 | starts = Index(l.Begin().X(), 0, 0);
|
|---|
| 160 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 14, 15));
|
|---|
| 161 | starts = l.Begin();
|
|---|
| 162 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 14, 15));
|
|---|
| 163 |
|
|---|
| 164 | /* 0 -1 1 */
|
|---|
| 165 | index = to_1d(1,0,2);
|
|---|
| 166 | subsizes = Index(l.Size().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 167 | starts = Index(l.Begin().X(), 0, l.End().Z());
|
|---|
| 168 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 16, 17));
|
|---|
| 169 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 170 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 16, 17));
|
|---|
| 171 |
|
|---|
| 172 | /* 0 1 -1 */
|
|---|
| 173 | index = to_1d(1,2,0);
|
|---|
| 174 | subsizes = Index(l.Size().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 175 | starts = Index(l.Begin().X(), l.End().Y(), 0);
|
|---|
| 176 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 17, 16));
|
|---|
| 177 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 178 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 17, 16));
|
|---|
| 179 |
|
|---|
| 180 | /* 0 1 1 */
|
|---|
| 181 | index = to_1d(1,2,2);
|
|---|
| 182 | subsizes = Index(l.Size().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 183 | starts = Index(l.Begin().X(), l.End().Y(), l.End().Z());
|
|---|
| 184 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 15, 14));
|
|---|
| 185 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 186 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 15, 14));
|
|---|
| 187 |
|
|---|
| 188 | /* -1 -1 -1 */
|
|---|
| 189 | index = to_1d(0,0,0);
|
|---|
| 190 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 191 | starts = 0;
|
|---|
| 192 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 18, 19));
|
|---|
| 193 | starts = l.Begin();
|
|---|
| 194 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 18, 19));
|
|---|
| 195 |
|
|---|
| 196 | /* -1 -1 1 */
|
|---|
| 197 | index = to_1d(0,0,2);
|
|---|
| 198 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 199 | starts = Index(0, 0, l.End().Z());;
|
|---|
| 200 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 20, 21));
|
|---|
| 201 | starts = Index(l.Begin().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 202 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 20, 21));
|
|---|
| 203 |
|
|---|
| 204 | /* -1 1 -1 */
|
|---|
| 205 | index = to_1d(0,2,0);
|
|---|
| 206 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 207 | starts = Index(0, l.End().Y(), 0);
|
|---|
| 208 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 22, 23));
|
|---|
| 209 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 210 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 22, 23));
|
|---|
| 211 |
|
|---|
| 212 | /* 1 -1 -1 */
|
|---|
| 213 | index = to_1d(2,0,0);
|
|---|
| 214 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 215 | starts = Index(l.End().X(), 0, 0);
|
|---|
| 216 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 24, 25));
|
|---|
| 217 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.Begin().Z());
|
|---|
| 218 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 24, 25));
|
|---|
| 219 |
|
|---|
| 220 | /* -1 1 1 */
|
|---|
| 221 | index = to_1d(0,2,2);
|
|---|
| 222 | subsizes = Index(l.HaloEnd1().X()-l.HaloBegin1().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 223 | starts = Index(0, l.End().Y(), l.End().Z());
|
|---|
| 224 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 25, 24));
|
|---|
| 225 | starts = Index(l.Begin().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 226 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 25, 24));
|
|---|
| 227 |
|
|---|
| 228 | /* 1 -1 1 */
|
|---|
| 229 | index = to_1d(2,0,2);
|
|---|
| 230 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd1().Y()-l.HaloBegin1().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 231 | starts = Index(l.End().X(), 0, l.End().Z());
|
|---|
| 232 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 23, 22));
|
|---|
| 233 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.Begin().Y(), l.End().Z()-l.HaloEnd2().Z()+l.HaloBegin2().Z());
|
|---|
| 234 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 23, 22));
|
|---|
| 235 |
|
|---|
| 236 | /* 1 1 -1 */
|
|---|
| 237 | index = to_1d(2,2,0);
|
|---|
| 238 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd1().Z()-l.HaloBegin1().Z());
|
|---|
| 239 | starts = Index(l.End().X(), l.End().Y(), 0);
|
|---|
| 240 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 21, 20));
|
|---|
| 241 | starts = Index(l.End().X()-l.HaloEnd2().X()+l.HaloBegin2().X(), l.End().Y()-l.HaloEnd2().Y()+l.HaloBegin2().Y(), l.Begin().Z());
|
|---|
| 242 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 21, 20));
|
|---|
| 243 |
|
|---|
| 244 | /* 1 1 1 */
|
|---|
| 245 | index = to_1d(2,2,2);
|
|---|
| 246 | subsizes = Index(l.HaloEnd2().X()-l.HaloBegin2().X(), l.HaloEnd2().Y()-l.HaloBegin2().Y(), l.HaloEnd2().Z()-l.HaloBegin2().Z());
|
|---|
| 247 | starts = l.End();
|
|---|
| 248 | halo.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 19, 18));
|
|---|
| 249 | starts = l.End()-l.HaloEnd2()+l.HaloBegin2();
|
|---|
| 250 | nb.push_back(VMG::MPI::Datatype(sizes, subsizes, starts, ranks[index], 19, 18));
|
|---|
| 251 |
|
|---|
| 252 | }
|
|---|
| 253 | }
|
|---|