MsmC1HermiteGridCutoff Class Reference

Inheritance diagram for MsmC1HermiteGridCutoff:

Public Member Functions
	MsmC1HermiteGridCutoff ()
	MsmC1HermiteGridCutoff (CkMigrateMessage *m)
void	init ()
void	setup (MsmGridCutoffInitMsg *bmsg)
void	setupSections (MsmC1HermiteGridCutoffSetupMsg *msg)
void	compute (GridMsg *gmsg)
void	compute_specialized (GridMsg *gmsg)
Public Member Functions inherited from MsmGridCutoffKernel< C1Vector, C1Matrix >
	MsmGridCutoffKernel ()
void	init ()
void	setup (MsmGridCutoffInitMsg *bmsg)
void	setupWeights (const msm::Grid< C1Matrix > *ptrgc, const msm::Grid< C1Matrix > *ptrgvc)
void	compute (GridMsg *gmsg)

Public Attributes
CProxyElement_MsmC1HermiteBlock	msmBlockElementProxy
CkSectionInfo	cookie
msm::Grid< C1Vector >	ehfull
Public Attributes inherited from MsmGridCutoffKernel< C1Vector, C1Matrix >
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< C1Vector >	qh
msm::Grid< C1Vector >	eh
msm::Grid< C1Vector >	ehfold
const msm::Grid< C1Matrix > *	pgc
const msm::Grid< C1Matrix > *	pgvc
int	priority
int	sequence

Detailed Description

Definition at line 2327 of file ComputeMsm.C.

Constructor & Destructor Documentation

MsmC1HermiteGridCutoff() [1/2]

MsmC1HermiteGridCutoff::MsmC1HermiteGridCutoff ( )

inline

Definition at line 2338 of file ComputeMsm.C.

{ }

MsmC1HermiteGridCutoff() [2/2]

MsmC1HermiteGridCutoff::MsmC1HermiteGridCutoff ( CkMigrateMessage *  m )

inline

Definition at line 2340 of file ComputeMsm.C.

    { }

Member Function Documentation

compute()

void MsmC1HermiteGridCutoff::compute ( GridMsg *  gmsg )

inline

Definition at line 2407 of file ComputeMsm.C.

References msm::Array< T >::buffer(), MsmTimer::COMM, MsmGridCutoffKernel< Vtype, Mtype >::compute(), compute_specialized(), cookie, msm::Grid< T >::data(), MsmGridCutoffKernel< C1Vector, C1Matrix >::eh, MsmGridCutoffKernel< C1Vector, C1Matrix >::ehblockSend, ehfull, msm::Ivec::i, msm::Ivec::j, msm::Ivec::k, msm::Array< T >::len(), msm::BlockIndex::level, MsmGridCutoffKernel< C1Vector, C1Matrix >::mgrLocal, msmBlockElementProxy, ComputeMsmMgr::msmC1HermiteBlock, msm::BlockIndex::n, msm::BlockSend::nblock_wrap, msm::IndexRange::nn(), MsmGridCutoffKernel< C1Vector, C1Matrix >::priority, GridMsg::put(), msm::Grid< T >::reset(), MsmGridCutoffKernel< C1Vector, C1Matrix >::sequence, and SET_PRIORITY.

                                {
#ifdef DEBUG_MSM_GRID
      printf("MsmC1HermiteGridCutoff %d:  compute()\n", thisIndex);
#endif
#if 0
      // base class consumes this grid message
      MsmGridCutoffKernel<C1Vector,C1Matrix>::compute(gmsg);
#else
      compute_specialized(gmsg);
#endif

#ifdef MSM_TIMING
      double startTime, stopTime;
      startTime = CkWallTimer();
#endif
#ifdef MSM_REDUCE_GRID

      // perform section reduction over potential grids
      CProxy_CkMulticastMgr mcastProxy =
        CkpvAccess(BOCclass_group).multicastMgr;
      CkMulticastMgr *mcastPtr =
        CProxy_CkMulticastMgr(mcastProxy).ckLocalBranch();
      CkCallback cb(CkIndex_MsmC1HermiteBlock::sumReducedPotential(NULL),
          msmBlockElementProxy);
      // sum into "full" sized buffer needed for contribute
      ehfull.reset(0);
      ehfull += eh;
      mcastPtr->contribute(
          ehfull.nn() * sizeof(C1Vector), ehfull.data().buffer(), 
          CkReduction::sum_float, cookie, cb);

#else
      // place eh into message
      const msm::BlockIndex& bindex = ehblockSend.nblock_wrap;
      int msgsz = eh.data().len() * sizeof(C1Vector);
      GridMsg *gm = new(msgsz, sizeof(int)) GridMsg;
      SET_PRIORITY(gm, sequence, priority);
      gm->put(eh, bindex.level, sequence);
      // lookup in ComputeMsmMgr proxy array by level
      mgrLocal->msmC1HermiteBlock[bindex.level](
          bindex.n.i, bindex.n.j, bindex.n.k).addPotential(gm);

#endif // MSM_REDUCE_GRID

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

compute_specialized()

void MsmC1HermiteGridCutoff::compute_specialized

(

GridMsg *

gmsg

)

Definition at line 2463 of file ComputeMsm.C.

References msm::Array< T >::buffer(), C1_VECTOR_SIZE, MsmTimer::COMM, msm::Grid< T >::data(), MsmGridCutoffKernel< C1Vector, C1Matrix >::eh, MsmGridCutoffKernel< C1Vector, C1Matrix >::ehblockSend, MsmGridCutoffKernel< C1Vector, C1Matrix >::ehfold, MsmGridCutoffKernel< C1Vector, C1Matrix >::eia, MsmGridCutoffKernel< C1Vector, C1Matrix >::eib, MsmGridCutoffKernel< C1Vector, C1Matrix >::eja, MsmGridCutoffKernel< C1Vector, C1Matrix >::ejb, MsmGridCutoffKernel< C1Vector, C1Matrix >::eka, MsmGridCutoffKernel< C1Vector, C1Matrix >::ekb, MsmGridCutoffKernel< C1Vector, C1Matrix >::eni, MsmGridCutoffKernel< C1Vector, C1Matrix >::enj, MsmGridCutoffKernel< C1Vector, C1Matrix >::enk, GridMsg::get(), MsmTimer::GRIDCUTOFF, msm::Ivec::i, msm::IndexRange::ia(), msm::IndexRange::ib(), msm::Grid< T >::init(), MsmGridCutoffKernel< C1Vector, C1Matrix >::isfold, msm::Ivec::j, msm::IndexRange::ja(), msm::IndexRange::jb(), msm::Ivec::k, msm::IndexRange::ka(), msm::IndexRange::kb(), msm::BlockIndex::level, msm::IndexRange::lower(), MsmGridCutoffKernel< C1Vector, C1Matrix >::mgrLocal, msm::IndexRange::ni(), msm::IndexRange::nj(), msm::IndexRange::nk(), msm::BlockSend::nrange, msm::BlockSend::nrange_wrap, MsmGridCutoffKernel< C1Vector, C1Matrix >::pgc, MsmGridCutoffKernel< C1Vector, C1Matrix >::qh, MsmGridCutoffKernel< C1Vector, C1Matrix >::qhblockIndex, msm::Grid< T >::reset(), MsmGridCutoffKernel< C1Vector, C1Matrix >::sequence, msm::Grid< T >::set(), msm::Grid< T >::updateLower(), and C1Vector::velem.

Referenced by 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 C1Matrix *gcbuffer = pgc->data().buffer();
      const C1Vector *qhbuffer = qh.data().buffer();
      C1Vector *ehbuffer = eh.data().buffer();
#ifdef DEBUG_MEMORY_ALIGNMENT
      printf("gcbuffer mem:  addr=%p  div32=%lu  mod32=%lu\n",
          gcbuffer,
          (unsigned long)(gcbuffer)/32,
          (unsigned long)(gcbuffer)%32);
      printf("qhbuffer mem:  addr=%p  div32=%lu  mod32=%lu\n",
          qhbuffer,
          (unsigned long)(qhbuffer)/32,
          (unsigned long)(qhbuffer)%32);
      printf("ehbuffer mem:  addr=%p  div32=%lu  mod32=%lu\n",
          ehbuffer,
          (unsigned long)(ehbuffer)/32,
          (unsigned long)(ehbuffer)%32);
#endif

#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
            C1Vector ehsum = 0;

            // loop over charge grid
            int nn = mib - mia + 1;

            {
              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 C1Vector *qbuf = qhbuffer + (qjkoff + qj*qni);
                  const C1Matrix *gbuf = gcbuffer + (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++) {

#if 0
                    ehsum += gbuf[ii] * qbuf[ii];
#else
                    // skip matvec when matrix is 0
                    // first matrix element tells us if this is the case
                    if ( *((int *)(gbuf)) != 0) {

                      // expand matrix-vector multiply
#if defined(__INTEL_COMPILER)
#pragma vector always
#endif
                      for (int km=0, jm=0;  jm < C1_VECTOR_SIZE;  jm++) {
                        for (int im=0;  im < C1_VECTOR_SIZE;  im++, km++) {
                          ehsum.velem[jm] += gbuf->melem[km] * qbuf->velem[im];
                        }
                      }
                    } // if
                    gbuf++;
                    qbuf++;
#endif
                  }
                }
              } // end loop over charge grid

            }

            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 C1Vector *ehfoldbuf = ehfold.data().buffer();
        C1Vector *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
} // MsmC1HermiteGridCutoff::compute_specialized()

init()

void MsmC1HermiteGridCutoff::init ( )

inline

Definition at line 2367 of file ComputeMsm.C.

References MsmGridCutoffKernel< Vtype, Mtype >::init().

                {
      MsmGridCutoffKernel<C1Vector,C1Matrix>::init();
    }

setup()

void MsmC1HermiteGridCutoff::setup ( MsmGridCutoffInitMsg *  bmsg )

inline

Definition at line 2371 of file ComputeMsm.C.

References msm::Map::blockLevel, MsmGridCutoffKernel< C1Vector, C1Matrix >::ehblockSend, ehfull, msm::Map::gc_c1hermite, msm::Ivec::i, msm::Grid< T >::init(), msm::Ivec::j, msm::Ivec::k, msm::BlockIndex::level, MsmGridCutoffKernel< C1Vector, C1Matrix >::map, msm::BlockIndex::n, msm::BlockSend::nblock_wrap, MsmGridCutoffKernel< Vtype, Mtype >::setup(), and MsmGridCutoffKernel< Vtype, Mtype >::setupWeights().

                                           {
      // base class consumes this init proxy  message
      MsmGridCutoffKernel<C1Vector,C1Matrix>::setup(bmsg);
      // access type dependent constants from map
      MsmGridCutoffKernel<C1Vector,C1Matrix>::setupWeights(
          &(map->gc_c1hermite[ehblockSend.nblock_wrap.level]),
          NULL
          );
#ifdef DEBUG_MSM_GRID
      printf("MsmC1HermiteGridCutoff[%d]:  setup()"
          " send to level=%d block=(%d,%d,%d)\n",
          thisIndex, ehblockSend.nblock_wrap.level,
          ehblockSend.nblock_wrap.n.i,
          ehblockSend.nblock_wrap.n.j,
          ehblockSend.nblock_wrap.n.k);
#endif
#ifdef MSM_REDUCE_GRID
      // allocate full buffer space needed for section reduction
      int level = ehblockSend.nblock_wrap.level;
      int i = ehblockSend.nblock_wrap.n.i;
      int j = ehblockSend.nblock_wrap.n.j;
      int k = ehblockSend.nblock_wrap.n.k;
      ehfull.init( map->blockLevel[level](i,j,k).nrange );
#endif // MSM_REDUCE_GRID
    }

setupSections()

void MsmC1HermiteGridCutoff::setupSections ( MsmC1HermiteGridCutoffSetupMsg *  msg )

inline

Definition at line 2397 of file ComputeMsm.C.

References cookie, MsmC1HermiteGridCutoffSetupMsg::get(), and msmBlockElementProxy.

                                                            {
#ifdef DEBUG_MSM_GRID
      CkPrintf("MSM C1 HERMITE GRID CUTOFF %d setup section on PE %d\n",
          thisIndex, CkMyPe());
#endif
      CkGetSectionInfo(cookie, msg);  // init the cookie
      msg->get(&msmBlockElementProxy);  // get proxy to MsmC1HermiteBlock
      delete msg;
    }

Member Data Documentation

cookie

CkSectionInfo MsmC1HermiteGridCutoff::cookie

Definition at line 2333 of file ComputeMsm.C.

Referenced by compute(), and setupSections().

ehfull

msm::Grid<C1Vector> MsmC1HermiteGridCutoff::ehfull

Definition at line 2335 of file ComputeMsm.C.

Referenced by compute(), and setup().

msmBlockElementProxy

CProxyElement_MsmC1HermiteBlock MsmC1HermiteGridCutoff::msmBlockElementProxy

Definition at line 2332 of file ComputeMsm.C.

Referenced by compute(), and setupSections().

---

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

• ComputeMsm.C