MsmGridCutoffKernel< Vtype, Mtype > Class Template Reference

Public Member Functions
	MsmGridCutoffKernel ()
void	init ()
void	setup (MsmGridCutoffInitMsg *bmsg)
void	setupWeights (const msm::Grid< Mtype > *ptrgc, const msm::Grid< Mtype > *ptrgvc)
void	compute (GridMsg *gmsg)

Public Attributes
ComputeMsmMgr *	mgrLocal
msm::Map *	map
msm::BlockIndex	qhblockIndex
msm::BlockSend	ehblockSend
int	eia
int	eib
int	eja
int	ejb
int	eka
int	ekb
int	eni
int	enj
int	enk
int	isfold
msm::Grid< Vtype >	qh
msm::Grid< Vtype >	eh
msm::Grid< Vtype >	ehfold
const msm::Grid< Mtype > *	pgc
const msm::Grid< Mtype > *	pgvc
int	priority
int	sequence

Detailed Description

template<class Vtype, class Mtype>
class MsmGridCutoffKernel< Vtype, Mtype >

Definition at line 1790 of file ComputeMsm.C.

Constructor & Destructor Documentation

MsmGridCutoffKernel()

template<class Vtype, class Mtype>

MsmGridCutoffKernel< Vtype, Mtype >::MsmGridCutoffKernel ( )

inline

Definition at line 1806 of file ComputeMsm.C.

{ init(); }

Member Function Documentation

compute()

template<class Vtype, class Mtype>

void MsmGridCutoffKernel< Vtype, Mtype >::compute ( GridMsg *  gmsg )

inline

Definition at line 1904 of file ComputeMsm.C.

Referenced by MsmGridCutoff::compute(), and MsmC1HermiteGridCutoff::compute().

                                {
#ifdef MSM_TIMING
      double startTime, stopTime;
      startTime = CkWallTimer();
#endif
      //
      // receive block of charges
      //
      int pid;
      // qh is resized only the first time, memory allocation persists
      gmsg->get(qh, pid, sequence);
      delete gmsg;
#ifdef MSM_TIMING
      stopTime = CkWallTimer();
      mgrLocal->msmTiming[MsmTimer::COMM] += stopTime - startTime;
#endif

      //
      // grid cutoff calculation
      // this charge block -> this potential block
      //

#ifdef MSM_TIMING
      startTime = stopTime;
#endif
      // resets indexing on block
      eh.init(ehblockSend.nrange);  // (always have to re-init nrange for eh)
      eh.reset(0);
      // index range of weights
      int gia = pgc->ia();
      int gib = pgc->ib();
      int gja = pgc->ja();
      int gjb = pgc->jb();
      int gka = pgc->ka();
      int gkb = pgc->kb();
      int gni = pgc->ni();
      int gnj = pgc->nj();
      // index range of charge grid
      int qia = qh.ia();
      int qib = qh.ib();
      int qja = qh.ja();
      int qjb = qh.jb();
      int qka = qh.ka();
      int qkb = qh.kb();
      int qni = qh.ni();
      int qnj = qh.nj();
      // index range of potentials
      int ia = eh.ia();
      int ib = eh.ib();
      int ja = eh.ja();
      int jb = eh.jb();
      int ka = eh.ka();
      int kb = eh.kb();

      int index = 0;

      // access buffers directly
      const Mtype *gcbuffer = pgc->data().buffer();
      //const Mtype *gvcbuffer = pgvc->data().buffer();
      const Vtype *qhbuffer = qh.data().buffer();
      Vtype *ehbuffer = eh.data().buffer();
      //Vtype *gvsumbuffer = mgrLocal->gvsum.data().buffer();

#ifndef MSM_COMM_ONLY
      // loop over potentials
      for (int k = ka;  k <= kb;  k++) {
        // clip charges to weights along k
        int mka = ( qka >= gka + k ? qka : gka + k );
        int mkb = ( qkb <= gkb + k ? qkb : gkb + k );

        for (int j = ja;  j <= jb;  j++) {
          // clip charges to weights along j
          int mja = ( qja >= gja + j ? qja : gja + j );
          int mjb = ( qjb <= gjb + j ? qjb : gjb + j );

          for (int i = ia;  i <= ib;  i++) {
            // clip charges to weights along i
            int mia = ( qia >= gia + i ? qia : gia + i );
            int mib = ( qib <= gib + i ? qib : gib + i );

            // accumulate sum to this eh point
            Vtype ehsum = 0;

#if 0
            // loop over charge grid
            for (int qk = mka;  qk <= mkb;  qk++) {
              int qkoff = (qk - qka) * qnj;
              int gkoff = ((qk-k) - gka) * gnj;

              for (int qj = mja;  qj <= mjb;  qj++) {
                int qjkoff = (qkoff + qj - qja) * qni;
                int gjkoff = (gkoff + (qj-j) - gja) * gni;

// help the vectorizer make reasonable decisions
#if defined(__INTEL_COMPILER)
#pragma vector always 
#endif
                for (int qi = mia;  qi <= mib;  qi++) {
                  int qijkoff = qjkoff + qi - qia;
                  int gijkoff = gjkoff + (qi-i) - gia;

                  ehsum += gcbuffer[gijkoff] * qhbuffer[qijkoff];
                }
              }
            } // end loop over charge grid
#else

#if 0
            // loop over charge grid
            int nn = mib - mia + 1;
            for (int qk = mka;  qk <= mkb;  qk++) {
              int qkoff = (qk - qka) * qnj;
              int gkoff = ((qk-k) - gka) * gnj;

              for (int qj = mja;  qj <= mjb;  qj++) {
                int qjkoff = (qkoff + qj - qja) * qni;
                int gjkoff = (gkoff + (qj-j) - gja) * gni;

                const Float *qbuf = qhbuffer + (qjkoff - qia + mia);
                const Float *gbuf = gcbuffer + (gjkoff - i - gia + mia);
#ifdef MSM_PROFILING
                mgrLocal->xLoopCnt[nn]++;
#endif
// help the vectorizer make reasonable decisions
#if defined(__INTEL_COMPILER)
#pragma vector always 
#endif
                for (int ii = 0;  ii < nn;  ii++) {
                  ehsum += gbuf[ii] * qbuf[ii];
                }
              }
            } // end loop over charge grid
#else
            // loop over charge grid
            int nn = mib - mia + 1;
            if (nn == 8) {  // hard coded inner loop = 8
              int qnji = qnj * qni;
              int qkoff = -qka*qnji - qja*qni - qia + mia;
              int gnji = gnj * gni;
              int gkoff = (-k-gka)*gnji + (-j-gja)*gni - i - gia + mia;

              for (int qk = mka;  qk <= mkb;  qk++) {
                int qjkoff = qkoff + qk*qnji;
                int gjkoff = gkoff + qk*gnji;

                for (int qj = mja;  qj <= mjb;  qj++) {
                  const Vtype *qbuf = qhbuffer + (qjkoff + qj*qni);
                  const Mtype *gbuf = gcbuffer + (gjkoff + qj*gni);
                  //const Mtype *gvcbuf = gvcbuffer + (gjkoff + qj*gni);
                  //Vtype *gvsumbuf = gvsumbuffer + (gjkoff + qj*gni);
#ifdef MSM_PROFILING
                  mgrLocal->xLoopCnt[nn]++;
#endif
// help the vectorizer make reasonable decisions
#if defined(__INTEL_COMPILER)
#pragma vector always 
#endif
                  for (int ii = 0;  ii < 8;  ii++) {
                    ehsum += gbuf[ii] * qbuf[ii];
                    //gvsumbuf[ii] += qbuf[ii] * qbuf[ii] * gvcbuf[ii];
                  }
                }
              } // end loop over charge grid
            }
            else {  // variable length inner loop < 8
              int qnji = qnj * qni;
              int qkoff = -qka*qnji - qja*qni - qia + mia;
              int gnji = gnj * gni;
              int gkoff = (-k-gka)*gnji + (-j-gja)*gni - i - gia + mia;

              for (int qk = mka;  qk <= mkb;  qk++) {
                int qjkoff = qkoff + qk*qnji;
                int gjkoff = gkoff + qk*gnji;

                for (int qj = mja;  qj <= mjb;  qj++) {
                  const Vtype *qbuf = qhbuffer + (qjkoff + qj*qni);
                  const Mtype *gbuf = gcbuffer + (gjkoff + qj*gni);
                  //const Mtype *gvcbuf = gvcbuffer + (gjkoff + qj*gni);
                  //Vtype *gvsumbuf = gvsumbuffer + (gjkoff + qj*gni);
#ifdef MSM_PROFILING
                  mgrLocal->xLoopCnt[nn]++;
#endif
// help the vectorizer make reasonable decisions
#if defined(__INTEL_COMPILER)
#pragma vector always 
#endif
                  for (int ii = 0;  ii < nn;  ii++) {
                    ehsum += gbuf[ii] * qbuf[ii];
                    //gvsumbuf[ii] += qbuf[ii] * qbuf[ii] * gvcbuf[ii];
                  }
                }
              } // end loop over charge grid
            }
#endif // 0

#endif // 0

            ehbuffer[index] = ehsum;
            index++;
          }
        }
      } // end loop over potentials
#endif // !MSM_COMM_ONLY

#ifdef MSM_PROFILING
      mgrLocal->doneProfiling();
#endif

      //
      // send block of potentials
      //

#ifdef MSM_FOLD_FACTOR
      // if "fold factor" is active for this level,
      // need to sum unfolded potential grid back into periodic grid
      if (isfold) {
        // copy unfolded grid
        ehfold = eh;
        // reset eh indexing to correctly folded size
        eh.set(eia, eni, eja, enj, eka, enk);
        eh.reset(0);
#ifdef DEBUG_MSM_GRID
        printf("level=%d   ehfold:  [%d..%d] x [%d..%d] x [%d..%d]  "
            "(%d x %d x %d)\n"
                "              eh:  [%d..%d] x [%d..%d] x [%d..%d]  "
            "(%d x %d x %d)\n"
               "         eh lower:  %d %d %d\n",
            qhblockIndex.level,
            ehfold.ia(), ehfold.ib(), 
            ehfold.ja(), ehfold.jb(),
            ehfold.ka(), ehfold.kb(),
            ehfold.ni(), ehfold.nj(), ehfold.nk(),
            eh.ia(), eh.ib(), 
            eh.ja(), eh.jb(),
            eh.ka(), eh.kb(),
            eh.ni(), eh.nj(), eh.nk(),
            ehblockSend.nrange_wrap.lower().i,
            ehblockSend.nrange_wrap.lower().j,
            ehblockSend.nrange_wrap.lower().k
            );
#endif
        const Vtype *ehfoldbuf = ehfold.data().buffer();
        Vtype *ehbuf = eh.data().buffer();
        // now we "fold" eh by calculating the
        // wrap around sum of ehfold into correctly sized eh
        int index = 0;
        for (int k = ka;  k <= kb;  k++) {
          int kk = k;
          if      (kk < eka)  do { kk += enk; } while (kk < eka);
          else if (kk > ekb)  do { kk -= enk; } while (kk > ekb);
          int koff = (kk - eka) * enj;
          for (int j = ja;  j <= jb;  j++) {
            int jj = j;
            if      (jj < eja)  do { jj += enj; } while (jj < eja);
            else if (jj > ejb)  do { jj -= enj; } while (jj > ejb);
            int jkoff = (koff + (jj - eja)) * eni;
            for (int i = ia;  i <= ib;  i++, index++) {
              int ii = i;
              if      (ii < eia)  do { ii += eni; } while (ii < eia);
              else if (ii > eib)  do { ii -= eni; } while (ii > eib);
              int ijkoff = jkoff + (ii - eia);
              ehbuf[ijkoff] += ehfoldbuf[index];
            }
          }
        }
      }
      else {
        // shift grid index range to its true (wrapped) values
        eh.updateLower( ehblockSend.nrange_wrap.lower() );
      }
#else    // !MSM_FOLD_FACTOR
      // shift grid index range to its true (wrapped) values
      eh.updateLower( ehblockSend.nrange_wrap.lower() );
#endif   // MSM_FOLD_FACTOR

#ifdef MSM_TIMING
      stopTime = CkWallTimer();
      mgrLocal->msmTiming[MsmTimer::GRIDCUTOFF] += stopTime - startTime;
#endif
    } // compute()

init()

template<class Vtype, class Mtype>

void MsmGridCutoffKernel< Vtype, Mtype >::init ( )

inline

Definition at line 1808 of file ComputeMsm.C.

Referenced by MsmGridCutoff::init(), MsmC1HermiteGridCutoff::init(), and MsmGridCutoffKernel< Float, Float >::MsmGridCutoffKernel().

                {
      isfold = 0;
      mgrLocal = CProxy_ComputeMsmMgr::ckLocalBranch(
          CkpvAccess(BOCclass_group).computeMsmMgr);
      map = &(mgrLocal->mapData());
      mgrLocal->addVirialContrib();
#ifdef MSM_TIMING
      mgrLocal->addTiming();
#endif
#ifdef MSM_PROFILING
      mgrLocal->addProfiling();
#endif
    }

setup()

template<class Vtype, class Mtype>

void MsmGridCutoffKernel< Vtype, Mtype >::setup ( MsmGridCutoffInitMsg *  bmsg )

inline

Definition at line 1838 of file ComputeMsm.C.

Referenced by MsmGridCutoff::setup(), and MsmC1HermiteGridCutoff::setup().

                                           {
      qhblockIndex = bmsg->qhBlockIndex;
      ehblockSend = bmsg->ehBlockSend;
      delete bmsg;

      // set message priority
      priority = mgrLocal->nlevels
        + 2*(mgrLocal->nlevels - ehblockSend.nblock_wrap.level) - 1;
      // allocate qh buffer
      qh.init(map->blockLevel[qhblockIndex.level](qhblockIndex.n).nrange);
      // allocate eh buffer
      eh.init(ehblockSend.nrange);
      // preprocess "fold factor" if active for this level
      if (map->foldfactor[qhblockIndex.level].active) {
        // allocate ehfold buffer
        ehfold = eh;
        // set index range of potentials
        eia = eh.ia();
        eib = eh.ib();
        eja = eh.ja();
        ejb = eh.jb();
        eka = eh.ka();
        ekb = eh.kb();
        eni = eh.ni();
        enj = eh.nj();
        enk = eh.nk();
        if (map->blockLevel[qhblockIndex.level].nn() == 1) {
          if (map->ispx) { eia = qh.ia();  eib = qh.ib();  eni = qh.ni(); }
          if (map->ispy) { eja = qh.ja();  ejb = qh.jb();  enj = qh.nj(); }
          if (map->ispz) { eka = qh.ka();  ekb = qh.kb();  enk = qh.nk(); }
        }
        else {
          // find destination block index
          int level = qhblockIndex.level;
          msm::BlockIndex bn = map->blockOfGridIndex(
              ehblockSend.nrange_wrap.lower(), level);
          map->wrapBlockIndex(bn);
          if (map->ispx) {
            eia = bn.n.i * map->bsx[level];
            eib = eia + qh.ni() - 1;
            eni = qh.ni();
          }
          if (map->ispy) {
            eja = bn.n.j * map->bsy[level];
            ejb = eja + qh.nj() - 1;
            enj = qh.nj();
          }
          if (map->ispz) {
            eka = bn.n.k * map->bsz[level];
            ekb = eka + qh.nk() - 1;
            enk = qh.nk();
          }
        }
        isfold = 1;
      } // if fold factor
    } // setup()

setupWeights()

template<class Vtype, class Mtype>

void MsmGridCutoffKernel< Vtype, Mtype >::setupWeights ( const msm::Grid< Mtype > *  ptrgc, const msm::Grid< Mtype > *  ptrgvc  )

inline

Definition at line 1895 of file ComputeMsm.C.

Referenced by MsmGridCutoff::setup(), and MsmC1HermiteGridCutoff::setup().

          {
      pgc = ptrgc;
      pgvc = ptrgvc;
    } // setupWeights()

Member Data Documentation

eh

template<class Vtype, class Mtype>

msm::Grid<Vtype> MsmGridCutoffKernel< Vtype, Mtype >::eh

Definition at line 1799 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

ehblockSend

template<class Vtype, class Mtype>

msm::BlockSend MsmGridCutoffKernel< Vtype, Mtype >::ehblockSend

Definition at line 1795 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

ehfold

template<class Vtype, class Mtype>

msm::Grid<Vtype> MsmGridCutoffKernel< Vtype, Mtype >::ehfold

Definition at line 1800 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

eia

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::eia

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

eib

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::eib

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

eja

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::eja

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

ejb

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::ejb

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

eka

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::eka

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

ekb

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::ekb

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

eni

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::eni

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

enj

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::enj

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

enk

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::enk

Definition at line 1796 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

isfold

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::isfold

Definition at line 1797 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), MsmGridCutoffKernel< Float, Float >::init(), and MsmGridCutoffKernel< Float, Float >::setup().

map

template<class Vtype, class Mtype>

msm::Map* MsmGridCutoffKernel< Vtype, Mtype >::map

Definition at line 1793 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::init(), and MsmGridCutoffKernel< Float, Float >::setup().

mgrLocal

template<class Vtype, class Mtype>

ComputeMsmMgr* MsmGridCutoffKernel< Vtype, Mtype >::mgrLocal

Definition at line 1792 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), MsmGridCutoffKernel< Float, Float >::init(), and MsmGridCutoffKernel< Float, Float >::setup().

pgc

template<class Vtype, class Mtype>

const msm::Grid<Mtype>* MsmGridCutoffKernel< Vtype, Mtype >::pgc

Definition at line 1801 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setupWeights().

pgvc

template<class Vtype, class Mtype>

const msm::Grid<Mtype>* MsmGridCutoffKernel< Vtype, Mtype >::pgvc

Definition at line 1802 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::setupWeights().

priority

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::priority

Definition at line 1803 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::setup().

qh

template<class Vtype, class Mtype>

msm::Grid<Vtype> MsmGridCutoffKernel< Vtype, Mtype >::qh

Definition at line 1798 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

qhblockIndex

template<class Vtype, class Mtype>

msm::BlockIndex MsmGridCutoffKernel< Vtype, Mtype >::qhblockIndex

Definition at line 1794 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute(), and MsmGridCutoffKernel< Float, Float >::setup().

sequence

template<class Vtype, class Mtype>

int MsmGridCutoffKernel< Vtype, Mtype >::sequence

Definition at line 1804 of file ComputeMsm.C.

Referenced by MsmGridCutoffKernel< Float, Float >::compute().

---

The documentation for this class was generated from the following file:

• ComputeMsm.C