| 263 | |
| 264 | template <class TM, class TV_ROW, class TV_COL> |
| 265 | void SparseMatrix<TM,TV_ROW,TV_COL> :: |
| 266 | MultAdd (double s, const BaseVector & x, BaseVector & y) const |
| 267 | { |
| 268 | static Timer t("SparseMatrix::MultAdd"); RegionTimer reg(t); |
| 269 | t.AddFlops (this->NZE()*sizeof(TV_ROW)*sizeof(TV_COL)/sqr(sizeof(double))); |
| 270 | |
| 271 | ParallelForRange |
| 272 | (balance, [&] (IntRange myrange) |
| 273 | { |
| 274 | FlatVector<TVX> fx = x.FV<TVX>(); |
| 275 | FlatVector<TVY> fy = y.FV<TVY>(); |
| 276 | |
| 277 | for (auto i : myrange) |
| 278 | fy(i) += s * RowTimesVector (i, fx); |
| 279 | }); |
| 280 | |
| 281 | #ifdef OLD |
| 282 | if (task_manager) |
| 283 | { |
| 284 | FlatVector<TVX> fx = x.FV<TVX>(); |
| 285 | FlatVector<TVY> fy = y.FV<TVY>(); |
| 286 | |
| 287 | // int ntasks = task_manager->GetNumThreads(); |
| 288 | |
| 289 | task_manager -> CreateJob |
| 290 | ([&] (TaskInfo & ti) |
| 291 | { |
| 292 | int tasks_per_part = ti.ntasks / balance.Size(); |
| 293 | int mypart = ti.task_nr / tasks_per_part; |
| 294 | int num_in_part = ti.task_nr % tasks_per_part; |
| 295 | |
| 296 | auto myrange = balance[mypart].Split (num_in_part, tasks_per_part); |
| 297 | |
| 298 | for (auto row : myrange) |
| 299 | fy(row) += s * RowTimesVector (row, fx); |
| 300 | |
| 301 | }); |
| 302 | return; |
| 303 | } |
| 304 | |
| 305 | |
| 306 | FlatVector<TVX> fx = x.FV<TVX>(); |
| 307 | FlatVector<TVY> fy = y.FV<TVY>(); |
| 308 | |
| 309 | int h = this->Height(); |
| 310 | for (int i = 0; i < h; i++) |
| 311 | fy(i) += s * RowTimesVector (i, fx); |
| 312 | #endif |
| 313 | |
| 314 | |
| 315 | } |
| 316 | |
| 317 | template <class TM, class TV_ROW, class TV_COL> |
| 318 | void SparseMatrix<TM,TV_ROW,TV_COL> :: |