eschnett/cactusamrex - Change T35FSMI7CA2CFZSRLTCZQBT2KDSEPDZD5UAPNNGR4YYAJ67LT4PAC

CarpetX: Use multi-threading for all levels simultaneously

Created by Erik Schnetter on June 2, 2020

T35FSMI7CA2CFZSRLTCZQBT2KDSEPDZD5UAPNNGR4YYAJ67LT4PAC

Dependencies

In channels

master

main

Change contents

Replacement in CarpetX/src/schedule.cxx at line 62 [5.531]
B:BD[4.10002] → [4.10002:10020]
```
  MFIter *mfiter;
```
[4.10002]
[4.10020]
```
  const MFPointer *restrict mfpointer;
```

Replacement in CarpetX/src/schedule.cxx at line 137 [5.531]

B:BD[6.11948] → [6.11948:12023]

GridDesc::GridDesc(const GHExt::LevelData &leveldata, const MFIter &mfi) {

[6.11948]

[3.0]

GridDesc::GridDesc(const GHExt::LevelData &leveldata, const MFPointer &mfp) {

Replacement in CarpetX/src/schedule.cxx at line 140 [5.531]

∅:D[3.28] → [6.12023:12296]

B:BD[6.12023] → [6.12023:12296]

  const Box &fbx = mfi.fabbox();   // allocated array
  const Box &vbx = mfi.validbox(); // interior region (without ghosts)
  // const Box &bx = mfi.tilebox();       // current region (without ghosts)
  const Box &gbx = mfi.growntilebox(); // current region (with ghosts)

[3.28]

[6.12296]

  const Box &fbx = mfp.fabbox();   // allocated array
  const Box &vbx = mfp.validbox(); // interior region (without ghosts)
  const Box &gbx = mfp.growntilebox(); // current region (with ghosts)

Replacement in CarpetX/src/schedule.cxx at line 147 [5.531]
B:BD[6.12447] → [6.12447:12506]
```
    nghostzones[d] = mfi.theFabArrayBase().nGrowVect()[d];
```
[6.12447]
[6.12506]
```
    nghostzones[d] = mfp.nGrowVect()[d];
```

Replacement in CarpetX/src/schedule.cxx at line 231 [5.531]

B:BD[6.13888] → [6.13888:13967]

B:BD[6.13967] → [4.10357:10442]

GridPtrDesc::GridPtrDesc(const GHExt::LevelData &leveldata, const MFIter &mfi)
    : GridDesc(leveldata, mfi) {
  const Box &fbx = mfi.fabbox(); // allocated array

[6.13888]

[4.10442]

GridPtrDesc::GridPtrDesc(const GHExt::LevelData &leveldata,
                         const MFPointer &mfp)
    : GridDesc(leveldata, mfp) {
  const Box &fbx = mfp.fabbox(); // allocated array

Replacement in CarpetX/src/schedule.cxx at line 240 [5.531]

B:BD[4.10611] → [4.10611:10657]

∅:D[4.10657] → [6.13967:14052]

B:BD[6.13967] → [6.13967:14052]

                           const MFIter &mfi)
    : GridDesc(leveldata, mfi) {
  const Box &fbx = mfi.fabbox(); // allocated array

[4.10611]

[6.14052]

                           const MFPointer &mfp)
    : GridDesc(leveldata, mfp) {
  const Box &fbx = mfp.fabbox(); // allocated array

Replacement in CarpetX/src/schedule.cxx at line 954 [5.531]

B:BD[6.14470] → [6.14470:14513]

                      const MFIter &mfi) {

[6.14470]

[7.1078]

                      const MFPointer &mfp) {

Replacement in CarpetX/src/schedule.cxx at line 956 [5.531]
B:BD[7.1111] → [2.10776:10818]
```
  const GridPtrDesc grid(leveldata, mfi);
```
[7.1111]
[5.3311]
```
  const GridPtrDesc grid(leveldata, mfp);
```

Replacement in CarpetX/src/schedule.cxx at line 986 [5.531]

B:BD[8.1847] → [2.10819:10876]

    const GridPtrDesc1 grid1(leveldata, groupdata, mfi);

[8.1847]

[9.1400]

    const GridPtrDesc1 grid1(leveldata, groupdata, mfp);

Replacement in CarpetX/src/schedule.cxx at line 988 [5.531]

B:BD[9.1462] → [9.1462:1535]

      const Array4<CCTK_REAL> &vars = groupdata.mfab.at(tl)->array(mfi);

[9.1462]

[4.12277]

      const Array4<CCTK_REAL> vars = groupdata.mfab.at(tl)->array(mfp.index());

Replacement in CarpetX/src/schedule.cxx at line 1022 [5.531]
∅:D[10.2597] → [11.2946:2989]
B:BD[11.2946] → [11.2946:2989]
```
                      const MFIter &mfi) {
```
[10.2597]
[7.1112]
```
                      const MFPointer &mfp) {
```
Replacement in CarpetX/src/schedule.cxx at line 1199 [5.531]
B:BD[12.1988] → [4.12742:12789]
```
    thread_local_info.at(n)->mfiter = nullptr;
```
[12.1988]
[13.941]
```
    thread_local_info.at(n)->mfpointer = nullptr;
```
Insertion in CarpetX/src/schedule.cxx at line 1780 [5.531]
[4.13219]
[5.5238]
```
#if 0
```

Replacement in CarpetX/src/schedule.cxx at line 1790 [5.531]

B:BD[12.2229] → [4.13368:13430]

      TileBox &restrict thread_tilebox = thread_info.tilebox;

[12.2229]

[14.6107]

       TileBox &restrict thread_tilebox = thread_info.tilebox;

Replacement in CarpetX/src/schedule.cxx at line 1792 [5.531]

B:BD[14.6108] → [6.17249:17316]

      // Loop over all levels
      // TODO: parallelize this loop

[14.6108]

[15.10244]

                   // Loop over all levels
                   // TODO: parallelize this loop

Replacement in CarpetX/src/schedule.cxx at line 1800 [5.531]

B:BD[16.17068] → [4.13431:13468]

∅:D[4.13468] → [6.17375:17445]

∅:D[16.17068] → [6.17375:17445]

B:BD[11.3778] → [6.17375:17445]

          thread_info.mfiter = &mfi;
          enter_local_mode(threadGH, thread_tilebox, leveldata, mfi);

[16.17068]

[11.3828]

          cout << "level=" << level << " mfi.currentIndex=" << mfi.tileIndex()
               << " mfi.length=" << mfi.length() << "\n";
          MFPointer mfp(mfi);
          thread_info.mfpointer = &mfp;
          enter_local_mode(threadGH, thread_tilebox, leveldata, mfp);

Replacement in CarpetX/src/schedule.cxx at line 1806 [5.531]

B:BD[11.3888] → [6.17446:17516]

B:BD[6.17516] → [4.13469:13509]

          leave_local_mode(threadGH, thread_tilebox, leveldata, mfi);
          thread_info.mfiter = nullptr;

[11.3888]

[11.3938]

          leave_local_mode(threadGH, thread_tilebox, leveldata, mfp);
          thread_info.mfpointer = nullptr;

Insertion in CarpetX/src/schedule.cxx at line 1812 [5.531]

[5.5564]

[4.13510]


#else
    vector<std::function<void()> > tasks;
    for (int level = min_level; level < max_level; ++level) {
      const auto &restrict leveldata = ghext->leveldata.at(level);
      const auto mfitinfo = MFItInfo().EnableTiling(
          {max_tile_size_x, max_tile_size_y, max_tile_size_z});
      // Note: The MFIter uses global variables and OpenMP barriers
      for (MFIter mfi(*leveldata.mfab0, mfitinfo); mfi.isValid(); ++mfi) {
        const MFPointer mfp(mfi);
        assert(leveldata.mfab0->array(mfi).dataPtr() ==
               leveldata.mfab0->array(mfp.index()).dataPtr());
        const auto task{[level, mfp, function, attribute] {
          const int thread_num = omp_get_thread_num();
          thread_local_info_t &restrict thread_info =
              *thread_local_info.at(thread_num);
          cGH *restrict const threadGH = &thread_info.cctkGH;
          TileBox &restrict thread_tilebox = thread_info.tilebox;
          const auto &restrict leveldata = ghext->leveldata.at(level);
          thread_info.mfpointer = &mfp;
          enter_level_mode(threadGH, leveldata);
          enter_local_mode(threadGH, thread_tilebox, leveldata, mfp);
          CCTK_CallFunction(function, attribute, threadGH);
          leave_local_mode(threadGH, thread_tilebox, leveldata, mfp);
          leave_level_mode(threadGH, leveldata);
          thread_info.mfpointer = nullptr;
        }};
        tasks.push_back(task);
      }
    }
#pragma omp parallel
    {
      // Initialize thread-local state variables
      const int thread_num = omp_get_thread_num();
      thread_local_info_t &restrict thread_info =
          *thread_local_info.at(thread_num);
      cGH *restrict const threadGH = &thread_info.cctkGH;
      update_cctkGH(threadGH, cctkGH);
      // run all tasks
#pragma omp for schedule(dynamic)
      for (size_t i = 0; i < tasks.size(); ++i)
        tasks[i]();
    }
#endif

Replacement in CarpetX/src/schedule.cxx at line 2108 [5.531]

B:BD[17.752] → [18.10830:10957]

        // Refined level: Prolongate boundaries from next coarser
        // level, and copy from adjacent boxes on same level

[17.752]

[17.873]

        // Refined level: Prolongate boundaries from next coarser level, and
        // copy from adjacent boxes on same level

Insertion in CarpetX/src/schedule.hxx at line 54 [5.5992]

[15.13385]


// Like an MFIter, but does not support iteration, instead it can be copied
struct MFPointer {
  int m_index;
  Box m_fabbox;
  Box m_growntilebox;
  Box m_validbox;
  IntVect m_nGrowVect;
  MFPointer() = delete;
  MFPointer(const MFPointer &) = default;
  MFPointer(MFPointer &&) = default;
  MFPointer &operator=(const MFPointer &) = default;
  MFPointer &operator=(MFPointer &&) = default;
  MFPointer(const MFIter &mfi)
      : m_index(mfi.index()), m_fabbox(mfi.fabbox()),
        m_growntilebox(mfi.growntilebox()), m_validbox(mfi.validbox()),
        m_nGrowVect(mfi.theFabArrayBase().nGrowVect()) {}
  constexpr int index() const noexcept { return m_index; }
  constexpr Box fabbox() const noexcept { return m_fabbox; }
  constexpr Box growntilebox() const noexcept { return m_growntilebox; }
  constexpr Box validbox() const noexcept { return m_validbox; }
  constexpr IntVect nGrowVect() const noexcept { return m_nGrowVect; }
};

Replacement in CarpetX/src/schedule.hxx at line 83 [5.5992]

B:BD[6.18121] → [6.18121:18187]

  GridDesc(const GHExt::LevelData &leveldata, const MFIter &mfi);

[6.18121]

[6.18187]

  GridDesc(const GHExt::LevelData &leveldata, const MFPointer &mfp);

Replacement in CarpetX/src/schedule.hxx at line 92 [5.5992]

B:BD[6.18328] → [6.18328:18397]

  GridPtrDesc(const GHExt::LevelData &leveldata, const MFIter &mfi);

[6.18328]

[6.18397]

  GridPtrDesc(const GHExt::LevelData &leveldata, const MFPointer &mfp);

Replacement in CarpetX/src/schedule.hxx at line 114 [5.5992]

B:BD[4.14513] → [4.14513:14594]

               const GHExt::LevelData::GroupData &groupdata, const MFIter &mfi);

[4.14513]

[4.14594]

               const GHExt::LevelData::GroupData &groupdata,
               const MFPointer &mfp);