ComputePmeCUDADevice Class Reference

#include <ComputePmeCUDAMgr.h>

Inheritance diagram for ComputePmeCUDADevice:

Public Member Functions
	ComputePmeCUDADevice ()
	ComputePmeCUDADevice (CkMigrateMessage *m)
	~ComputePmeCUDADevice ()
void	initialize (PmeGrid &pmeGrid_in, int pencilIndexY_in, int pencilIndexZ_in, int deviceID_in, int pmePencilType_in, CProxy_ComputePmeCUDAMgr mgrProxy_in, CProxy_PmeAtomFiler pmeAtomFiler_in)
int	getDeviceID ()
cudaStream_t	getStream ()
CProxy_ComputePmeCUDAMgr	getMgrProxy ()
void	setPencilProxy (CProxy_CudaPmePencilXYZ pmePencilXYZ_in)
void	setPencilProxy (CProxy_CudaPmePencilXY pmePencilXY_in)
void	setPencilProxy (CProxy_CudaPmePencilX pmePencilX_in)
void	activate_pencils ()
void	initializePatches (int numHomePatches_in)
void	registerNeighbor ()
void	recvAtoms (PmeAtomMsg *msg)
void	sendAtomsToNeighbors ()
void	sendAtomsToNeighbor (int y, int z, int atomIval)
void	recvAtomsFromNeighbor (PmeAtomPencilMsg *msg)
void	registerRecvAtomsFromNeighbor ()
void	spreadCharge ()
void	gatherForce ()
void	gatherForceDone (unsigned int iGrid)
void	sendForcesToNeighbors ()
void	recvForcesFromNeighbor (PmeForcePencilMsg *msg)
void	mergeForcesOnPatch (int homePatchIndex)
void	sendForcesToPatch (PmeForceMsg *forceMsg)
void	gatherForceDoneSubset (int first, int last)
bool	isGridEnabled (unsigned int i) const

Detailed Description

Definition at line 420 of file ComputePmeCUDAMgr.h.

Constructor & Destructor Documentation

ComputePmeCUDADevice() [1/2]

ComputePmeCUDADevice::ComputePmeCUDADevice

(

)

Definition at line 974 of file ComputePmeCUDAMgr.C.

References DebugM, endi(), and NUM_GRID_MAX.

                                           {
  // __sdag_init();
  numHomePatches = 0;
//   forceCapacity = 0;
//   force = NULL;
  DebugM(4, "ComputePmeCUDADevice::ComputePmeCUDADevice\n" << endi);
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    pmeRealSpaceComputes[iGrid] = NULL;
    enabledGrid[iGrid] = false;
    forces[iGrid] = NULL;
    forceCapacities[iGrid] = 0;
  }
//   pmeRealSpaceCompute = NULL;
  streamCreated = false;
  lock_numHomePatchesMerged = CmiCreateLock();
  lock_numPencils = CmiCreateLock();
  lock_numNeighborsRecv = CmiCreateLock();
  lock_recvAtoms = CmiCreateLock();
  numNeighborsExpected = 0;
  numStrayAtoms = 0;
  // Reset counters
  numNeighborsRecv = 0;
  numHomePatchesRecv = 0;
  numHomePatchesMerged = 0;
  atomI = 0;
  forceI = 1;
}

ComputePmeCUDADevice() [2/2]

ComputePmeCUDADevice::ComputePmeCUDADevice

(

CkMigrateMessage *

m

)

Definition at line 1002 of file ComputePmeCUDAMgr.C.

References DebugM, endi(), and NUM_GRID_MAX.

                                                              {
  // __sdag_init();
  numHomePatches = 0;
//   forceCapacity = 0;
//   force = NULL;
  DebugM(4, "ComputePmeCUDADevice::ComputePmeCUDADevice(CkMigrateMessage)\n" << endi);
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    pmeRealSpaceComputes[iGrid] = NULL;
    enabledGrid[iGrid] = false;
    forces[iGrid] = NULL;
    forceCapacities[iGrid] = 0;
  }
  streamCreated = false;
  lock_numHomePatchesMerged = CmiCreateLock();
  lock_numPencils = CmiCreateLock();
  lock_numNeighborsRecv = CmiCreateLock();
  lock_recvAtoms = CmiCreateLock();
  numNeighborsExpected = 0;
  numStrayAtoms = 0;
  // Reset counters
  numNeighborsRecv = 0;
  numHomePatchesRecv = 0;
  numHomePatchesMerged = 0;
  atomI = 0;
  forceI = 1;
}

~ComputePmeCUDADevice()

ComputePmeCUDADevice::~ComputePmeCUDADevice

(

)

Definition at line 1029 of file ComputePmeCUDAMgr.C.

References cudaCheck, and NUM_GRID_MAX.

                                            {
  if (streamCreated) {
    cudaCheck(cudaSetDevice(deviceID));
    cudaCheck(cudaStreamDestroy(stream));
  }
  for (int j=0;j < 2;j++)
    for (int i=0;i < pmeAtomStorage[j].size();i++) {
      if (pmeAtomStorageAllocatedHere[i]) delete pmeAtomStorage[j][i];
    }
//   if (force != NULL) deallocate_host<CudaForce>(&force);
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    if (pmeRealSpaceComputes[iGrid] != NULL) delete pmeRealSpaceComputes[iGrid];
    if (forces[iGrid] != NULL) deallocate_host<CudaForce>(&forces[iGrid]);
    enabledGrid[iGrid] = false;
  }
  CmiDestroyLock(lock_numHomePatchesMerged);
  CmiDestroyLock(lock_numPencils);
  CmiDestroyLock(lock_numNeighborsRecv);
  CmiDestroyLock(lock_recvAtoms);
}

Member Function Documentation

activate_pencils()

void ComputePmeCUDADevice::activate_pencils

(

)

Definition at line 1189 of file ComputePmeCUDAMgr.C.

References PmeStartMsg::dataGrid, PmeStartMsg::dataSizes, PmeStartMsg::enabledGrid, and NUM_GRID_MAX.

                                            {
  if (pmePencilType == 1) {
    PmeStartMsg* pmeStartXMsg = new PmeStartMsg();
    for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
      if (enabledGrid[iGrid] == true) {
        pmeStartXMsg->dataGrid[iGrid] = pmeRealSpaceComputes[iGrid]->getData();
        pmeStartXMsg->dataSizes[iGrid] = pmeRealSpaceComputes[iGrid]->getDataSize();
        pmeStartXMsg->enabledGrid[iGrid] = true;
      } else {
        pmeStartXMsg->dataGrid[iGrid] = NULL;
        pmeStartXMsg->dataSizes[iGrid] = 0;
        pmeStartXMsg->enabledGrid[iGrid] = false;
      }
    }
    pmePencilX(0, pencilIndexY, pencilIndexZ).start(pmeStartXMsg);
  } else if (pmePencilType == 2) {
    PmeStartMsg* pmeStartXMsg = new PmeStartMsg();
    for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
      if (enabledGrid[iGrid] == true) {
        pmeStartXMsg->dataGrid[iGrid] = pmeRealSpaceComputes[iGrid]->getData();
        pmeStartXMsg->dataSizes[iGrid] = pmeRealSpaceComputes[iGrid]->getDataSize();
        pmeStartXMsg->enabledGrid[iGrid] = true;
      } else {
        pmeStartXMsg->dataGrid[iGrid] = NULL;
        pmeStartXMsg->dataSizes[iGrid] = 0;
        pmeStartXMsg->enabledGrid[iGrid] = false;
      }
    }
    pmePencilXY(0, 0, pencilIndexZ).start(pmeStartXMsg);
  } else if (pmePencilType == 3) {
    PmeStartMsg* pmeStartMsg = new PmeStartMsg();
    for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
      if (enabledGrid[iGrid] == true) {
        pmeStartMsg->dataGrid[iGrid] = pmeRealSpaceComputes[iGrid]->getData();
        pmeStartMsg->dataSizes[iGrid] = pmeRealSpaceComputes[iGrid]->getDataSize();
        pmeStartMsg->enabledGrid[iGrid] = true;
      } else {
        pmeStartMsg->dataGrid[iGrid] = NULL;
        pmeStartMsg->dataSizes[iGrid] = 0;
        pmeStartMsg->enabledGrid[iGrid] = false;
      }
    }
    pmePencilXYZ[0].start(pmeStartMsg);
  }
}

gatherForce()

void ComputePmeCUDADevice::gatherForce

(

)

Definition at line 1770 of file ComputePmeCUDAMgr.C.

References CUDA_PME_SPREADCHARGE_EVENT, NUM_GRID_MAX, Node::Object(), Node::simParameters, and simParams.

                                       {
  traceUserBracketEvent(CUDA_PME_SPREADCHARGE_EVENT, beforeWalltime, CmiWallTimer());
  beforeWalltime = CmiWallTimer();
  // gather (i.e. un-grid) forces
  SimParameters *simParams = Node::Object()->simParameters;
//   if (simParameters->alchOn) {
//     pmeRealSpaceComputes[1]->gatherForce(lattice, force);
//     ((CudaPmeRealSpaceCompute*)(pmeRealSpaceComputes[1]))->gatherForceSetCallback(this);
//   } else {
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    if (enabledGrid[iGrid]) {
//       fprintf(stdout, "gatherForce at grid %u\n", iGrid);
      pmeRealSpaceComputes[iGrid]->gatherForce(lattice, forces[iGrid]);
      // Set callback that will call gatherForceDone() once gatherForce is done
      ((CudaPmeRealSpaceCompute*)(pmeRealSpaceComputes[iGrid]))->gatherForceSetCallback(this);
    }
  }
  // ((CudaPmeRealSpaceCompute*)pmeRealSpaceCompute)->waitGatherForceDone();
    // gatherForceDone();
//   }
}

gatherForceDone()

void ComputePmeCUDADevice::gatherForceDone

(

unsigned int

iGrid

)

Definition at line 1816 of file ComputePmeCUDAMgr.C.

References CUDA_PME_GATHERFORCE_EVENT, gatherForceDoneLoop(), NUM_GRID_MAX, Node::Object(), sendForcesToNeighbors(), Node::simParameters, and SimParameters::useCkLoop.

                                                             {
  // CHC: prevent race condition when there are multiple pmeRealSpaceCompute objects
  forceReady[iGrid] = 1;
  bool all_force_ready = true;
//   fprintf(stdout, "gatherForceDone at grid %u\n", iGrid);
  // loop over forceReady to check if all forces are gathered
  for (unsigned int i = 0; i < NUM_GRID_MAX; ++i) {
    if (forceReady[i] == -1) continue;
    if (forceReady[i] == 0) all_force_ready = false;
  }
  if (all_force_ready) {
    for (unsigned int i = 0; i < NUM_GRID_MAX; ++i) {
      if (forceReady[i] == -1) continue;
      if (forceReady[i] == 1) forceReady[i] = 0;
    }
//     fprintf(stdout, "all force ready\n");
    // Primary pencil has the forces

    traceUserBracketEvent(CUDA_PME_GATHERFORCE_EVENT, beforeWalltime, CmiWallTimer());

    // Send forces to neighbors
    sendForcesToNeighbors();

#if CMK_SMP && USE_CKLOOP
    int useCkLoop = Node::Object()->simParameters->useCkLoop;
    if (useCkLoop >= 1) {
      CkLoop_Parallelize(gatherForceDoneLoop, 1, (void *)this, CkMyNodeSize(), 0, numHomePatches-1);
    } else
#endif

    {
      // Loop through home patches and mark the primary pencil as "done"
      for (int homePatchIndex=0;homePatchIndex < numHomePatches;homePatchIndex++) {
        bool done = false;
        // ----------------------------- lock start ---------------------------
        // NOTE: We use node-wide lock here for the entire numPencils[] array, while
        //       we really would only need to each element but this would required
        //       numHomePatches number of locks.
        if (pmePencilType != 3) CmiLock(lock_numPencils);
        numPencils[forceI][homePatchIndex]--;
        if (numPencils[forceI][homePatchIndex] == 0) done = true;
        if (pmePencilType != 3) CmiUnlock(lock_numPencils);
        // ----------------------------- lock end  ---------------------------
        if (done) {
          // This home patch is done, launch force merging
          thisProxy[CkMyNode()].mergeForcesOnPatch(homePatchIndex);
        }
      }
    }

    // In case we have no home patches, clear the primary pencil storage here
    if (numHomePatches == 0) {
      int pp0 = 0-ylo + (0-zlo)*yNBlocks;
      pmeAtomStorage[forceI][pp0]->clear();
    }
  }
}

gatherForceDoneSubset()

void ComputePmeCUDADevice::gatherForceDoneSubset	(	int	first,
		int	last
	)

Definition at line 1797 of file ComputePmeCUDAMgr.C.

References mergeForcesOnPatch().

Referenced by gatherForceDoneLoop().

                                                                    {
  for (int homePatchIndex=first;homePatchIndex <= last;homePatchIndex++) {
    bool done = false;
    // ----------------------------- lock start ---------------------------
    // NOTE: We use node-wide lock here for the entire numPencils[] array, while
    //       we really would only need to each element but this would required
    //       numHomePatches number of locks.
    if (pmePencilType != 3) CmiLock(lock_numPencils);
    numPencils[forceI][homePatchIndex]--;
    if (numPencils[forceI][homePatchIndex] == 0) done = true;
    if (pmePencilType != 3) CmiUnlock(lock_numPencils);
    // ----------------------------- lock end  ---------------------------
    if (done) {
      // This home patch is done, launch force merging
      mergeForcesOnPatch(homePatchIndex);
    }
  }
}

getDeviceID()

int ComputePmeCUDADevice::getDeviceID

(

)

Definition at line 1159 of file ComputePmeCUDAMgr.C.

                                      {
  return deviceID;
}

getMgrProxy()

CProxy_ComputePmeCUDAMgr ComputePmeCUDADevice::getMgrProxy

(

)

Definition at line 1163 of file ComputePmeCUDAMgr.C.

                                                           {
  return mgrProxy;
}

getStream()

cudaStream_t ComputePmeCUDADevice::getStream

(

)

Definition at line 1155 of file ComputePmeCUDAMgr.C.

                                             {
  return stream;
}

initialize()

void ComputePmeCUDADevice::initialize	(	PmeGrid &	pmeGrid_in,
		int	pencilIndexY_in,
		int	pencilIndexZ_in,
		int	deviceID_in,
		int	pmePencilType_in,
		CProxy_ComputePmeCUDAMgr	mgrProxy_in,
		CProxy_PmeAtomFiler	pmeAtomFiler_in
	)

Definition at line 1050 of file ComputePmeCUDAMgr.C.

References createStream(), cudaCheck, DebugM, endi(), NUM_GRID_MAX, Node::Object(), Node::simParameters, simParams, PmeGrid::yBlocks, and PmeGrid::zBlocks.

                                       {

  deviceID = deviceID_in;
  DebugM(4, "ComputePmeCUDADevice::initialize deviceID "<< deviceID <<"\n"<< endi);
  cudaCheck(cudaSetDevice(deviceID));
  pmePencilType = pmePencilType_in;
  pmeGrid = pmeGrid_in;
#ifdef DEBUGM
  //  pmeGrid.print();
#endif
  pencilIndexY = pencilIndexY_in;
  pencilIndexZ = pencilIndexZ_in;
  mgrProxy = mgrProxy_in;
  pmeAtomFiler = pmeAtomFiler_in;
  // Size of the neighboring pencil grid, max 3x3
  yNBlocks = std::min(pmeGrid.yBlocks, 3);
  zNBlocks = std::min(pmeGrid.zBlocks, 3);
  // Local pencil is at y=0,z=0
  if (yNBlocks == 1) {
    ylo = 0;
    yhi = 0;
  } else if (yNBlocks == 2) {
    ylo = -1;
    yhi = 0;
  } else {
    ylo = -1;
    yhi = 1;
  }
  if (zNBlocks == 1) {
    zlo = 0;
    zhi = 0;
  } else if (zNBlocks == 2) {
    zlo = -1;
    zhi = 0;
  } else {
    zlo = -1;
    zhi = 1;
  }
  
  neighborForcePencilMsgs.resize(yNBlocks*zNBlocks, NULL);
  // neighborForcePencils.resize(yNBlocks*zNBlocks);
  for (int j=0;j < 2;j++)
    homePatchIndexList[j].resize(yNBlocks*zNBlocks);
  neighborPatchIndex.resize(yNBlocks*zNBlocks);

  pmeAtomStorageAllocatedHere.resize(yNBlocks*zNBlocks, false);
  SimParameters *simParams = Node::Object()->simParameters;
  for (int j=0;j < 2;j++) {
    pmeAtomStorage[j].resize(yNBlocks*zNBlocks, NULL);
    for (int z=zlo;z <= zhi;z++) {
      for (int y=ylo;y <= yhi;y++) {
        int pp = y-ylo + (z-zlo)*yNBlocks;
        int yt = (pencilIndexY + y + pmeGrid.yBlocks) % pmeGrid.yBlocks;
        int zt = (pencilIndexZ + z + pmeGrid.zBlocks) % pmeGrid.zBlocks;
        if (y == 0 && z == 0) {
          // Primary pencil
          pmeAtomStorage[j][pp] = new CudaPmeAtomStorage(pmePencilType != 3);
          pmeAtomStorage[j][pp]->setupAlch(*simParams);
        } else {
          pmeAtomStorage[j][pp] = new CpuPmeAtomStorage(pmePencilType != 3);
          pmeAtomStorage[j][pp]->setupAlch(*simParams);
        }
        pmeAtomStorageAllocatedHere[pp] = true;
      }
    }
  }

  // Create stream for this device
  createStream(stream);
  streamCreated = true;
  // CHC: enable at least 1 grid
  // CHC: do we need a different stream?
  pmeRealSpaceComputes[0] = new CudaPmeRealSpaceCompute(pmeGrid, pencilIndexY, pencilIndexZ, deviceID, stream);
  pmeRealSpaceComputes[0]->setGrid(0);
  enabledGrid[0] = true;
  if (simParams->alchOn) {
    pmeRealSpaceComputes[1] = new CudaPmeRealSpaceCompute(pmeGrid, pencilIndexY, pencilIndexZ, deviceID, stream);
    pmeRealSpaceComputes[1]->setGrid(1);
    // at least two grids are required for alchemical transformation
    enabledGrid[1] = true;
    if (simParams->alchDecouple) {
      pmeRealSpaceComputes[2] = new CudaPmeRealSpaceCompute(pmeGrid, pencilIndexY, pencilIndexZ, deviceID, stream);
      pmeRealSpaceComputes[2]->setGrid(2);
      enabledGrid[2] = true;
      pmeRealSpaceComputes[3] = new CudaPmeRealSpaceCompute(pmeGrid, pencilIndexY, pencilIndexZ, deviceID, stream);
      pmeRealSpaceComputes[3]->setGrid(3);
      enabledGrid[3] = true;
    }
    if (simParams->alchElecLambdaStart || simParams->alchThermIntOn) {
      pmeRealSpaceComputes[4] = new CudaPmeRealSpaceCompute(pmeGrid, pencilIndexY, pencilIndexZ, deviceID, stream);
      pmeRealSpaceComputes[4]->setGrid(4);
      enabledGrid[4] = true;
    }
  }
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    if (enabledGrid[iGrid]) {
      forceReady[iGrid] = 0;
    } else {
      forceReady[iGrid] = -1;
    }
  }
}

initializePatches()

void ComputePmeCUDADevice::initializePatches

(

int

numHomePatches_in

)

Definition at line 1235 of file ComputePmeCUDAMgr.C.

References PmeGrid::yBlocks, and PmeGrid::zBlocks.

                                                                  {
  numHomePatches = numHomePatches_in;
  for (int j=0;j < 2;j++)
    numPencils[j].resize(numHomePatches);
  for (int j=0;j < 2;j++)
    plList[j].resize(numHomePatches);
  for (int j=0;j < 2;j++)
    homePatchForceMsgs[j].resize(numHomePatches);
  // for (int j=0;j < 2;j++)
  //   numHomeAtoms[j].resize(numHomePatches);
  // If we have home patches, register this pencil with the neighbors and with self
  if (numHomePatches > 0) {
    for (int z=zlo;z <= zhi;z++) {
      for (int y=ylo;y <= yhi;y++) {
        int yt = (pencilIndexY + y + pmeGrid.yBlocks) % pmeGrid.yBlocks;
        int zt = (pencilIndexZ + z + pmeGrid.zBlocks) % pmeGrid.zBlocks;
        int node = mgrProxy.ckLocalBranch()->getNode(yt, zt);
        mgrProxy[node].registerNeighbor(yt, zt);
      }
    }
  }
}

isGridEnabled()

bool ComputePmeCUDADevice::isGridEnabled

(

unsigned int

i

)

const

Definition at line 1167 of file ComputePmeCUDAMgr.C.

                                                             {
  return enabledGrid[i];
}

mergeForcesOnPatch()

void ComputePmeCUDADevice::mergeForcesOnPatch

(

int

homePatchIndex

)

Definition at line 1982 of file ComputePmeCUDAMgr.C.

References Node::Object(), sendForcesToPatch(), Node::simParameters, simParams, CudaForce::x, CudaForce::y, and CudaForce::z.

Referenced by gatherForceDoneSubset().

                                                                {
  // We have all the forces for this patch => merge on a single Pe

  int pp0 = 0-ylo + (0-zlo)*yNBlocks;

  // Message that goes out to the compute
  PmeForceMsg *forceMsg = homePatchForceMsgs[forceI][homePatchIndex];

  SimParameters* simParams = Node::Object()->simParameters;
  if (pmePencilType == 3) {
    // 3D box => simple memory copy will do
    // Location of forces in the force[] array
    int* patchPos = pmeAtomStorage[forceI][pp0]->getPatchPos();
    // plList[homePatchIndex] array tells you the location of pencils that are sharing this home patch
    int pencilPatchIndex = plList[forceI][homePatchIndex][0].pencilPatchIndex;
    int atomStart = (pencilPatchIndex == 0) ? 0 : patchPos[pencilPatchIndex-1];
    int atomEnd   = patchPos[pencilPatchIndex];
    int numAtoms = atomEnd-atomStart;
    if (forceMsg->zeroCopy) {
      // Zero-copy, just pass the pointer
      forceMsg->force = forces[0]+atomStart;
      if (simParams->alchOn) {
        forceMsg->force2 = forces[1]+atomStart;
        if (simParams->alchDecouple) {
          forceMsg->force3 = forces[2]+atomStart;
          forceMsg->force4 = forces[3]+atomStart;
        }
        if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
          forceMsg->force5 = forces[4]+atomStart;
        }
      }
    } else {
      memcpy(forceMsg->force, forces[0]+atomStart, numAtoms*sizeof(CudaForce));
      if (simParams->alchOn) {
        memcpy(forceMsg->force2, forces[1]+atomStart, numAtoms*sizeof(CudaForce));
        if (simParams->alchDecouple) {
          memcpy(forceMsg->force3, forces[2]+atomStart, numAtoms*sizeof(CudaForce));
          memcpy(forceMsg->force4, forces[3]+atomStart, numAtoms*sizeof(CudaForce));
        }
        if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
          memcpy(forceMsg->force5, forces[4]+atomStart, numAtoms*sizeof(CudaForce));
        }
      }
    }
  } else {

    // Zero force array
    // memset(forceMsg->force, 0, numHomeAtoms[forceI][homePatchIndex]*sizeof(CudaForce));
    memset(forceMsg->force, 0, forceMsg->numAtoms*sizeof(CudaForce));
    if (simParams->alchOn) {
      memset(forceMsg->force2, 0, forceMsg->numAtoms*sizeof(CudaForce));
      if (simParams->alchDecouple) {
        memset(forceMsg->force3, 0, forceMsg->numAtoms*sizeof(CudaForce));
        memset(forceMsg->force4, 0, forceMsg->numAtoms*sizeof(CudaForce));
      }
      if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
        memset(forceMsg->force5, 0, forceMsg->numAtoms*sizeof(CudaForce));
      }
    }

    // Store forces from primary pencil
    {
      int* patchPos = pmeAtomStorage[forceI][pp0]->getPatchPos();
      int* index = pmeAtomStorage[forceI][pp0]->getAtomIndex();
      int pencilPatchIndex = plList[forceI][homePatchIndex][0].pencilPatchIndex;
      int atomStart = (pencilPatchIndex == 0) ? 0 : patchPos[pencilPatchIndex-1];
      int atomEnd   = patchPos[pencilPatchIndex];
      int numAtoms = atomEnd-atomStart;

      // Copy in local forces that are stored in the force[] array
      for (int i=0;i < numAtoms;i++) {
        forceMsg->force[index[atomStart + i]] = forces[0][atomStart + i];
        if (simParams->alchOn) {
          forceMsg->force2[index[atomStart + i]] = forces[1][atomStart + i];
          if (simParams->alchDecouple) {
            forceMsg->force3[index[atomStart + i]] = forces[2][atomStart + i];
            forceMsg->force4[index[atomStart + i]] = forces[3][atomStart + i];
          }
          if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
            forceMsg->force5[index[atomStart + i]] = forces[4][atomStart + i];
          }
        }
      }

    }

    // Add forces from neighboring pencils
    for (int j=1;j < plList[forceI][homePatchIndex].size();j++) {
      int pp               = plList[forceI][homePatchIndex][j].pp;
      int pencilPatchIndex = plList[forceI][homePatchIndex][j].pencilPatchIndex;

      int* patchPos = pmeAtomStorage[forceI][pp]->getPatchPos();
      int* index = pmeAtomStorage[forceI][pp]->getAtomIndex();
      int atomStart = (pencilPatchIndex == 0) ? 0 : patchPos[pencilPatchIndex-1];
      int atomEnd   = patchPos[pencilPatchIndex];
      int numAtoms = atomEnd-atomStart;
      CudaForce *dstForce = forceMsg->force;
      // CudaForce *srcForce = neighborForcePencils[pp].force;
      CudaForce *dstForce2 = forceMsg->force2;
      CudaForce *dstForce3 = forceMsg->force3;
      CudaForce *dstForce4 = forceMsg->force4;
      CudaForce *dstForce5 = forceMsg->force5;
      CudaForce *srcForce = neighborForcePencilMsgs[pp]->force;
      CudaForce *srcForce2 = neighborForcePencilMsgs[pp]->force2;
      CudaForce *srcForce3 = neighborForcePencilMsgs[pp]->force3;
      CudaForce *srcForce4 = neighborForcePencilMsgs[pp]->force4;
      CudaForce *srcForce5 = neighborForcePencilMsgs[pp]->force5;

      for (int i=0;i < numAtoms;i++) {
        dstForce[index[atomStart + i]].x += srcForce[atomStart + i].x;
        dstForce[index[atomStart + i]].y += srcForce[atomStart + i].y;
        dstForce[index[atomStart + i]].z += srcForce[atomStart + i].z;
        if (simParams->alchOn) {
          dstForce2[index[atomStart + i]].x += srcForce2[atomStart + i].x;
          dstForce2[index[atomStart + i]].y += srcForce2[atomStart + i].y;
          dstForce2[index[atomStart + i]].z += srcForce2[atomStart + i].z;
          if (simParams->alchDecouple) {
            dstForce3[index[atomStart + i]].x += srcForce3[atomStart + i].x;
            dstForce3[index[atomStart + i]].y += srcForce3[atomStart + i].y;
            dstForce3[index[atomStart + i]].z += srcForce3[atomStart + i].z;
            dstForce4[index[atomStart + i]].x += srcForce4[atomStart + i].x;
            dstForce4[index[atomStart + i]].y += srcForce4[atomStart + i].y;
            dstForce4[index[atomStart + i]].z += srcForce4[atomStart + i].z;
          }
          if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
            dstForce5[index[atomStart + i]].x += srcForce5[atomStart + i].x;
            dstForce5[index[atomStart + i]].y += srcForce5[atomStart + i].y;
            dstForce5[index[atomStart + i]].z += srcForce5[atomStart + i].z;
          }
        }
      }

    }
  }

  // Clear storage
  plList[forceI][homePatchIndex].clear();

  // ----------------------------- lock start ---------------------------
  // bool done = false;
  CmiLock(lock_numHomePatchesMerged);
  numHomePatchesMerged++;
  if (numHomePatchesMerged == numHomePatches) {
    // Reset counter
    numHomePatchesMerged = 0;

    // Delete messages
    for (int i=0;i < neighborForcePencilMsgs.size();i++) {
      if (neighborForcePencilMsgs[i] != NULL) {
        delete neighborForcePencilMsgs[i];
        neighborForcePencilMsgs[i] = NULL;
      }
    }

    // Done merging and sending forces => clear storage
    for (int pp=0;pp < homePatchIndexList[forceI].size();pp++)
      homePatchIndexList[forceI][pp].clear();
    for (int pp=0;pp < pmeAtomStorage[forceI].size();pp++)
      pmeAtomStorage[forceI][pp]->clear();

  }
  CmiUnlock(lock_numHomePatchesMerged);
  // ----------------------------- lock end  ---------------------------

  // Patch is done => send over to the node that contains the ComputePmeCUDA compute,
  // this node will then rely the message to the Pe that originally sent the atoms
  int pe = forceMsg->pe;
  if (CkNodeOf(pe) != CkMyNode())
    thisProxy[CkNodeOf(pe)].sendForcesToPatch(forceMsg);
  else
    sendForcesToPatch(forceMsg);

}

recvAtoms()

void ComputePmeCUDADevice::recvAtoms

(

PmeAtomMsg *

msg

)

Definition at line 1267 of file ComputePmeCUDAMgr.C.

References PmeAtomMsg::atoms, PmeAtomMsg::compute, PmeForceMsg::compute, PmeAtomMsg::doEnergy, PmeAtomMsg::doVirial, PmeAtomFiler::fileAtoms(), PmeAtomFiler::getAtomIndex(), PmeAtomFiler::getNumAtoms(), PmeAtomMsg::lattice, NAMD_bug(), PmeAtomMsg::numAtoms, PmeForceMsg::numAtoms, PmeForceMsg::numStrayAtoms, Node::Object(), PmeAtomMsg::pe, PmeForceMsg::pe, PRIORITY_SIZE, sendAtomsToNeighbors(), Node::simParameters, simParams, PmeAtomMsg::simulationStep, CudaAtom::x, CudaAtom::y, PmeGrid::yBlocks, CudaAtom::z, PmeGrid::zBlocks, and PmeForceMsg::zeroCopy.

                                                    {

  PmeAtomFiler *pmeAtomFilerPtr = pmeAtomFiler[CkMyPe()].ckLocalBranch();
  // Store "virial" and "energy" flags
  doVirial = msg->doVirial;
  doEnergy = msg->doEnergy;
  simulationStep = msg->simulationStep;
  // Store lattice
  SimParameters *simParams = Node::Object()->simParameters;
  lattice = msg->lattice;

  // Primary pencil index
  int pp0 = 0-ylo + (0-zlo)*yNBlocks;
  int p0 = 0;
  int pencilPatchIndex[9];
  int numStrayAtomsPatch = 0;
  if (pmePencilType == 3) {
    // 3D box => store atoms directly without index
    // NOTE: We don't check for stray atoms here!
    if (simParams->alchOn) {
      if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
        // only FEP, no alchDecouple and alchElecLambdaStart == 0, use 2 grids
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2});
      }
      if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
        // FEP with alchDecouple, use 4 grids
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4});
      }
      if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
        // FEP with alchDecouple and alchElecLambdaStart > 0, use 5 grids
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
      }
      if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
        // FEP without alchDecouple and alchElecLambdaStart > 0, use 3 grids (1,2,5)
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, NULL, NULL, msg->chargeFactors5});
      }
      if (simParams->alchThermIntOn && !simParams->alchDecouple) {
        // TI estimator without alchDecouple, use 3 grids (1,2,5)
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, NULL, NULL, msg->chargeFactors5});
      }
      if (simParams->alchThermIntOn && simParams->alchDecouple) {
        // TI estimator with alchDecouple, use all grids
        pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
      }
    } else {
      // no alchemistry
      pencilPatchIndex[p0] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{});
    }
  } else {

    // File atoms
    pmeAtomFilerPtr->fileAtoms(msg->numAtoms, msg->atoms, lattice, pmeGrid,
      pencilIndexY, pencilIndexZ, ylo, yhi, zlo, zhi);

    // Loop through pencils and add atoms to pencil atom lists
    // NOTE: we only store to neighboring pencil if there are atoms to store
    int numAtomsCheck = 0;
    for (int p=0;p < 9;p++) {

      int y = (p % 3);
      int z = (p / 3);

      int pp = y + z*yNBlocks;
      int numAtoms = pmeAtomFilerPtr->getNumAtoms(p);
      if (pp == pp0) p0 = p;
      if (pp == pp0 || numAtoms > 0) {
        if (pmeGrid.yBlocks == 1 && pmeGrid.zBlocks == 1 && (y != 0 || z != 0))
          NAMD_bug("ComputePmeCUDADevice::recvAtoms, problem with atom filing");
        int* index = pmeAtomFilerPtr->getAtomIndex(p);
        if (simParams->alchOn) {
          if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
            // only FEP, no alchDecouple and alchElecLambdaStart == 0, use 2 grids
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2});
          }
          if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
            // FEP with alchDecouple, use 4 grids
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4});
          }
          if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
            // FEP without alchDecouple and alchElecLambdaStart > 0, use 3 grids (1,2,5)
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, NULL, NULL, msg->chargeFactors5});
          }
          if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
            // FEP with alchDecouple and alchElecLambdaStart > 0, use 5 grids
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
          }
          if (simParams->alchThermIntOn && !simParams->alchDecouple) {
            // TI estimator without alchDecouple, use 3 grids (1,2,5)
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, NULL, NULL, msg->chargeFactors5});
          }
          if (simParams->alchThermIntOn && simParams->alchDecouple) {
            // TI estimator with alchDecouple, use all grids
            pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
          }
        } else {
          // no alchemistry
          pencilPatchIndex[p] = pmeAtomStorage[atomI][pp]->addAtomsWithIndex(numAtoms, msg->atoms, index, std::vector<float*>{});
        }
        // Number of patches in this storage tells you how many home patches contributed and
        // homePatchIndex (pe) tells you which patch contributed
        numAtomsCheck += numAtoms;
      }
    }

    // Deal with stray atoms
    numStrayAtomsPatch = pmeAtomFilerPtr->getNumAtoms(9);
    if (numStrayAtomsPatch > 0) {
      int* index = pmeAtomFilerPtr->getAtomIndex(9);
      CkPrintf("%d stray charges detected. Up to 10 listed below (index in patch, x, y, z):\n", numStrayAtomsPatch);
      for (int i=0;i < std::min(numStrayAtomsPatch, 10);i++) {
        int j = index[i];
        CkPrintf("%d %f %f %f\n", j, msg->atoms[j].x, msg->atoms[j].y, msg->atoms[j].z);
      }
    }

    if (numAtomsCheck + numStrayAtomsPatch < msg->numAtoms)
      NAMD_bug("ComputePmeCUDADevice::recvAtoms, missing atoms");
  }

  // Create storage for home patch forces
  PmeForceMsg *forceMsg;
  if (pmePencilType == 3 && CkNodeOf(msg->pe) == CkMyNode()) {
    // 3D FFT and compute resides on the same node => use zero-copy forces
    // CHC: forces are zero-copy so do we need this for alchDecouple?
    forceMsg = new (0, 0, 0, 0, 0, PRIORITY_SIZE) PmeForceMsg();
    forceMsg->zeroCopy = true;
  } else {
    const int alchGrid = simParams->alchOn ? 1 : 0;
    const int alchDecoupleGrid = simParams->alchDecouple ? 1: 0;
    const int alchSoftCoreOrTI = (simParams->alchElecLambdaStart > 0 || simParams->alchThermIntOn) ? 1 : 0;
    forceMsg = new (msg->numAtoms, alchGrid * msg->numAtoms,
                    alchDecoupleGrid * msg->numAtoms, alchDecoupleGrid * msg->numAtoms,
                    alchSoftCoreOrTI * msg->numAtoms, PRIORITY_SIZE) PmeForceMsg();
    forceMsg->zeroCopy = false;
  }
  forceMsg->numAtoms = msg->numAtoms;
  forceMsg->pe = msg->pe;
  forceMsg->compute = msg->compute;
  forceMsg->numStrayAtoms = numStrayAtomsPatch;

  bool done = false;
  // ----------------------------- lock start ---------------------------
  // Only after writing has finished, we get homePatchIndex
  // This quarantees that for whatever thread that receives "done=true", writing has finished on
  // ALL threads.
  CmiLock(lock_recvAtoms);
  numStrayAtoms += numStrayAtomsPatch;
  // Secure homePatchIndex. All writes after this must be inside lock-region
  int homePatchIndex = numHomePatchesRecv;
  // Store primary pencil first
  plList[atomI][homePatchIndex].push_back(PencilLocation(pp0, pencilPatchIndex[p0]));
  if (pmePencilType != 3) {
    // Go back to through neighboring pencils and store "homePatchIndex"
    for (int p=0;p < 9;p++) {

      int y = (p % 3);
      int z = (p / 3);

      int pp = y + z*yNBlocks;
      int numAtoms = pmeAtomFilerPtr->getNumAtoms(p);
      if (pp != pp0 && numAtoms > 0) {
        homePatchIndexList[atomI][pp].push_back(homePatchIndex);
        // plList[0...numHomePatches-1] = for each home patch stores the location of pencils that are
        //                                sharing it
        // plList[homePatchIndex].size() tells the number of pencils that the home patch is shared with
        plList[atomI][homePatchIndex].push_back(PencilLocation(pp, pencilPatchIndex[p]));
      }
    }
  }
  homePatchForceMsgs[atomI][homePatchIndex] = forceMsg;
  // numHomeAtoms[atomI][homePatchIndex] = msg->numAtoms;
  // Set the number of pencils contributing to this home patch
  numPencils[atomI][homePatchIndex] = plList[atomI][homePatchIndex].size();
  //
  numHomePatchesRecv++;
  if (numHomePatchesRecv == numHomePatches) {
    // Reset counter
    numHomePatchesRecv = 0;
    done = true;
  }
  CmiUnlock(lock_recvAtoms);
  // ----------------------------- lock end  ---------------------------

  // plList[atomI][homePatchIndex] array tells you the location of pencils that are sharing this home patch

  delete msg;

  if (done) {
    // Pencil has received all home patches and writing to memory is done => send atoms to neighbors
    sendAtomsToNeighbors();
  }
}

recvAtomsFromNeighbor()

void ComputePmeCUDADevice::recvAtomsFromNeighbor

(

PmeAtomPencilMsg *

msg

)

Definition at line 1574 of file ComputePmeCUDAMgr.C.

References PmeAtomPencilMsg::atoms, PmeAtomPencilMsg::doEnergy, PmeAtomPencilMsg::doVirial, PmeAtomPencilMsg::lattice, NAMD_bug(), PmeAtomPencilMsg::numAtoms, Node::Object(), registerRecvAtomsFromNeighbor(), Node::simParameters, simParams, PmeAtomPencilMsg::simulationStep, PmeAtomPencilMsg::srcY, PmeAtomPencilMsg::srcZ, PmeGrid::yBlocks, and PmeGrid::zBlocks.

                                                                      {
  // Store into primary pencil
  int pp0 = 0-ylo + (0-zlo)*yNBlocks;
  // Compute pencil index relative to primary pencil
  int y = msg->srcY - pencilIndexY;
  if (y < ylo) y += pmeGrid.yBlocks;
  if (y > yhi) y -= pmeGrid.yBlocks;
  int z = msg->srcZ - pencilIndexZ;
  if (z < zlo) z += pmeGrid.zBlocks;
  if (z > zhi) z -= pmeGrid.zBlocks;
  if (y < ylo || y > yhi || z < zlo || z > zhi || (y == 0 && z == 0)) {
    NAMD_bug("ComputePmeCUDADevice::recvAtomsFromNeighbor, pencil index outside bounds");
  }
  // Read energy and virial flags
  doEnergy = msg->doEnergy;
  doVirial = msg->doVirial;
  simulationStep = msg->simulationStep;
  // Read lattice
  lattice = msg->lattice;
  // Pencil index where atoms came from
  int pp = y-ylo + (z-zlo)*yNBlocks;
  // Store atoms and mark down the patch index where these atoms were added
  SimParameters *simParams = Node::Object()->simParameters;
  if (simParams->alchOn) {
    if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2});
    }
    if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4});
    }
    if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
    }
    if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, 0, 0, msg->chargeFactors5});
    }
    if (simParams->alchThermIntOn && !simParams->alchDecouple) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, 0, 0, msg->chargeFactors5});
    }
    if (simParams->alchThermIntOn && simParams->alchDecouple) {
      neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{msg->chargeFactors1, msg->chargeFactors2, msg->chargeFactors3, msg->chargeFactors4, msg->chargeFactors5});
    }
  } else {
    neighborPatchIndex[pp] = pmeAtomStorage[atomI][pp0]->addAtoms(msg->numAtoms, msg->atoms, std::vector<float*>{});
  }

  delete msg;

  registerRecvAtomsFromNeighbor();
}

recvForcesFromNeighbor()

void ComputePmeCUDADevice::recvForcesFromNeighbor

(

PmeForcePencilMsg *

msg

)

Definition at line 1927 of file ComputePmeCUDAMgr.C.

References NAMD_bug(), PmeForcePencilMsg::srcY, PmeForcePencilMsg::srcZ, PmeGrid::yBlocks, and PmeGrid::zBlocks.

                                                                        {

  // Source pencil index
  int y = msg->srcY - pencilIndexY;
  if (y < ylo) y += pmeGrid.yBlocks;
  if (y > yhi) y -= pmeGrid.yBlocks;
  int z = msg->srcZ - pencilIndexZ;
  if (z < zlo) z += pmeGrid.zBlocks;
  if (z > zhi) z -= pmeGrid.zBlocks;

  if (y < ylo || y > yhi || z < zlo || z > zhi || (y == 0 && z == 0)) {
    NAMD_bug("ComputePmeCUDADevice::recvForcesFromNeighbor, pencil index outside bounds");
  }

  // Source pencil
  int pp = y-ylo + (z-zlo)*yNBlocks;

  // Store message (deleted in mergeForcesOnPatch)
  neighborForcePencilMsgs[pp] = msg;

  // neighborForcePencils[pp].force = new CudaForce[msg->numAtoms];
  // memcpy(neighborForcePencils[pp].force, msg->force, sizeof(CudaForce)*msg->numAtoms);
  // neighborForcePencils[pp].numAtoms = msg->numAtoms;
  // neighborForcePencils[pp].y = msg->y;
  // neighborForcePencils[pp].z = msg->z;
  // neighborForcePencils[pp].srcY = msg->srcY;
  // neighborForcePencils[pp].srcZ = msg->srcZ;
  // delete msg;

  // numPatches = number of home patches this pencil has
  int numPatches = pmeAtomStorage[forceI][pp]->getNumPatches();
  if (numPatches != homePatchIndexList[forceI][pp].size()) {
    NAMD_bug("ComputePmeCUDADevice::recvForcesFromNeighbor, numPatches incorrect");
  }
  for (int i=0;i < numPatches;i++) {
    // this pencil contributed to home patch with index "homePatchIndex"
    int homePatchIndex = homePatchIndexList[forceI][pp][i];
    // ----------------------------- lock start ---------------------------
    // NOTE: We use node-wide lock here for the entire numPencils[] array, while
    //       we really would only need to each element but this would required
    //       numHomePatches number of locks.
    bool done = false;
    CmiLock(lock_numPencils);
    numPencils[forceI][homePatchIndex]--;
    if (numPencils[forceI][homePatchIndex] == 0) done = true;
    CmiUnlock(lock_numPencils);
    // ----------------------------- lock end  ---------------------------
    if (done) {
      // This home patch is done, launch force merging
      thisProxy[CkMyNode()].mergeForcesOnPatch(homePatchIndex);
    }
  }

}

registerNeighbor()

void ComputePmeCUDADevice::registerNeighbor

(

)

Definition at line 1258 of file ComputePmeCUDAMgr.C.

                                            {
  CmiLock(lock_numHomePatchesMerged);
  numNeighborsExpected++;
  CmiUnlock(lock_numHomePatchesMerged);
}

registerRecvAtomsFromNeighbor()

void ComputePmeCUDADevice::registerRecvAtomsFromNeighbor

(

)

Definition at line 1625 of file ComputePmeCUDAMgr.C.

References spreadCharge().

Referenced by recvAtomsFromNeighbor(), and sendAtomsToNeighbors().

                                                         {
  // Primary pencil
  int pp0 = 0-ylo + (0-zlo)*yNBlocks;

  bool done = false;
  // ----------------------------- lock start ---------------------------
  CmiLock(lock_numNeighborsRecv);
  numNeighborsRecv++;
  if (numNeighborsRecv == numNeighborsExpected) {
    // Reset counter
    numNeighborsRecv = 0;
    done = true;
  }
  CmiUnlock(lock_numNeighborsRecv);
  // ----------------------------- lock end  ---------------------------

  if (done) {
    // Primary pencil has received all atoms and writing has finished => spread charge
    spreadCharge();
  }  
}

sendAtomsToNeighbor()

void ComputePmeCUDADevice::sendAtomsToNeighbor	(	int	y,
		int	z,
		int	atomIval
	)

Definition at line 1478 of file ComputePmeCUDAMgr.C.

References PmeAtomPencilMsg::atoms, PmeAtomPencilMsg::chargeFactors1, PmeAtomPencilMsg::chargeFactors2, PmeAtomPencilMsg::chargeFactors3, PmeAtomPencilMsg::chargeFactors4, PmeAtomPencilMsg::chargeFactors5, PmeAtomPencilMsg::doEnergy, PmeAtomPencilMsg::doVirial, PmeAtomPencilMsg::lattice, PmeAtomPencilMsg::numAtoms, Node::Object(), PRIORITY_SIZE, Node::simParameters, simParams, PmeAtomPencilMsg::simulationStep, PmeAtomPencilMsg::srcY, PmeAtomPencilMsg::srcZ, PmeAtomPencilMsg::y, PmeGrid::yBlocks, PmeAtomPencilMsg::z, and PmeGrid::zBlocks.

                                                                         {
  // Pencil index  
  int pp = y-ylo + (z-zlo)*yNBlocks;
  // This neighbor pencil is done, finish it up before accessing it
  pmeAtomStorage[atomIval][pp]->finish();
  // Compute destination neighbor pencil index (yt,zt)
  int yt = (pencilIndexY + y + pmeGrid.yBlocks) % pmeGrid.yBlocks;
  int zt = (pencilIndexZ + z + pmeGrid.zBlocks) % pmeGrid.zBlocks;
  int numAtoms = pmeAtomStorage[atomIval][pp]->getNumAtoms();
  CudaAtom* atoms = pmeAtomStorage[atomIval][pp]->getAtoms();
  SimParameters *simParams = Node::Object()->simParameters;
  PmeAtomPencilMsg* msgPencil;
  if (simParams->alchOn) {
    if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, 0, 0, 0, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
    }
    if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == false)) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, numAtoms, numAtoms, 0, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      float* chargeFactors3 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(2);
      float* chargeFactors4 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(3);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors3, chargeFactors3, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors4, chargeFactors4, numAtoms*sizeof(float));
    }
    if (simParams->alchFepOn && simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, numAtoms, numAtoms, numAtoms, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      float* chargeFactors3 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(2);
      float* chargeFactors4 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(3);
      float* chargeFactors5 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(4);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors3, chargeFactors3, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors4, chargeFactors4, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors5, chargeFactors5, numAtoms*sizeof(float));
    }
    if (simParams->alchFepOn && !simParams->alchDecouple && (bool(simParams->alchElecLambdaStart) == true)) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, 0, 0, numAtoms, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      float* chargeFactors5 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(4);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors5, chargeFactors5, numAtoms*sizeof(float));
    }
    if (simParams->alchThermIntOn && !simParams->alchDecouple) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, 0, 0, numAtoms, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      float* chargeFactors5 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(4);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors5, chargeFactors5, numAtoms*sizeof(float));
    }
    if (simParams->alchThermIntOn && simParams->alchDecouple) {
      msgPencil = new (numAtoms, numAtoms, numAtoms, numAtoms, numAtoms, numAtoms, PRIORITY_SIZE) PmeAtomPencilMsg;
      float* chargeFactors1 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(0);
      float* chargeFactors2 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(1);
      float* chargeFactors3 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(2);
      float* chargeFactors4 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(3);
      float* chargeFactors5 = pmeAtomStorage[atomIval][pp]->getAtomElecFactors(4);
      memcpy(msgPencil->chargeFactors1, chargeFactors1, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors2, chargeFactors2, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors3, chargeFactors3, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors4, chargeFactors4, numAtoms*sizeof(float));
      memcpy(msgPencil->chargeFactors5, chargeFactors5, numAtoms*sizeof(float));
    }
  } else {
    msgPencil = new (numAtoms, 0, 0, 0, 0, 0, PRIORITY_SIZE) PmeAtomPencilMsg;
  }
  memcpy(msgPencil->atoms, atoms, numAtoms*sizeof(CudaAtom));
  msgPencil->numAtoms = numAtoms;
  // Store destination pencil index
  msgPencil->y = yt;
  msgPencil->z = zt;
  // Store source pencil index
  msgPencil->srcY = pencilIndexY;
  msgPencil->srcZ = pencilIndexZ;
  // Store energy and virial flags
  msgPencil->doEnergy = doEnergy;
  msgPencil->doVirial = doVirial;
  msgPencil->simulationStep = simulationStep;
  // Store lattice
  msgPencil->lattice = lattice;
  int node = mgrProxy.ckLocalBranch()->getNode(yt, zt);
  mgrProxy[node].recvAtomsFromNeighbor(msgPencil);
}

sendAtomsToNeighbors()

void ComputePmeCUDADevice::sendAtomsToNeighbors

(

)

Definition at line 1463 of file ComputePmeCUDAMgr.C.

References registerRecvAtomsFromNeighbor().

Referenced by recvAtoms().

                                                {
  for (int z=zlo;z <= zhi;z++) {
    for (int y=ylo;y <= yhi;y++) {
      // Only send to neighbors, not self
      if (y != 0 || z != 0) {
        // NOTE: Must send atomI -value since this will change in spreadCharge(), which might occur
        // before these sends have been performed
        thisProxy[CkMyNode()].sendAtomsToNeighbor(y, z, atomI);
      }
    }
  }
  // Register primary pencil
  registerRecvAtomsFromNeighbor();
}

sendForcesToNeighbors()

void ComputePmeCUDADevice::sendForcesToNeighbors

(

)

Definition at line 1877 of file ComputePmeCUDAMgr.C.

References PmeForcePencilMsg::force, PmeForcePencilMsg::force2, PmeForcePencilMsg::force3, PmeForcePencilMsg::force4, PmeForcePencilMsg::force5, PmeForcePencilMsg::numAtoms, Node::Object(), PRIORITY_SIZE, Node::simParameters, simParams, PmeForcePencilMsg::srcY, PmeForcePencilMsg::srcZ, PmeForcePencilMsg::y, PmeGrid::yBlocks, PmeForcePencilMsg::z, and PmeGrid::zBlocks.

Referenced by gatherForceDone().

                                                 {
  // Primary pencil has the forces
  int pp0 = 0-ylo + (0-zlo)*yNBlocks;
  int* patchPos = pmeAtomStorage[forceI][pp0]->getPatchPos();
  SimParameters *simParams = Node::Object()->simParameters;
  const int alchGrid = simParams->alchOn ? 1 : 0;
  const int alchDecoupleGrid = simParams->alchDecouple ? 1: 0;
  const int alchSoftCoreOrTI = (simParams->alchElecLambdaStart > 0 || simParams->alchThermIntOn) ? 1 : 0;
  // Loop through neighboring pencils
  for (int z=zlo;z <= zhi;z++) {
    for (int y=ylo;y <= yhi;y++) {
      // Only send to neighbors, not self
      if (y != 0 || z != 0) {
        int pp = y-ylo + (z-zlo)*yNBlocks;
        int patchIndex = neighborPatchIndex[pp];
        int atomStart = (patchIndex == 0) ? 0 : patchPos[patchIndex-1];
        int atomEnd   = patchPos[patchIndex];
        int natom = atomEnd-atomStart;
        // copy forces
        PmeForcePencilMsg *msg;
        msg = new (natom, alchGrid * natom, alchDecoupleGrid * natom, 
                   alchDecoupleGrid * natom, alchSoftCoreOrTI * natom,
                   PRIORITY_SIZE) PmeForcePencilMsg;
        msg->numAtoms = natom;
        memcpy(msg->force, forces[0]+atomStart, natom*sizeof(CudaForce));
        if (simParams->alchOn) {
          memcpy(msg->force2, forces[1]+atomStart, natom*sizeof(CudaForce));
          if (simParams->alchDecouple) {
            memcpy(msg->force3, forces[2]+atomStart, natom*sizeof(CudaForce));
            memcpy(msg->force4, forces[3]+atomStart, natom*sizeof(CudaForce));
          }
          if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
            memcpy(msg->force5, forces[4]+atomStart, natom*sizeof(CudaForce));
          }
        }
        // Calculate destination pencil index (dstY, dstZ) for this neighbor
        int dstY = (pencilIndexY + y + pmeGrid.yBlocks) % pmeGrid.yBlocks;
        int dstZ = (pencilIndexZ + z + pmeGrid.zBlocks) % pmeGrid.zBlocks;
        int node = mgrProxy.ckLocalBranch()->getNode(dstY, dstZ);
        msg->y = dstY;
        msg->z = dstZ;
        // Store source pencil index
        msg->srcY = pencilIndexY;
        msg->srcZ = pencilIndexZ;
        mgrProxy[node].recvForcesFromNeighbor(msg);
      }
    }
  }
}

sendForcesToPatch()

void ComputePmeCUDADevice::sendForcesToPatch

(

PmeForceMsg *

forceMsg

)

Definition at line 2156 of file ComputePmeCUDAMgr.C.

References PmeForceMsg::compute, Compute::localWorkMsg, PmeForceMsg::pe, and ComputePmeCUDA::storePmeForceMsg().

Referenced by mergeForcesOnPatch().

                                                                  {
  // Now we're on the node that has Pe, hence "compute" -pointer is valid
  int pe                  = forceMsg->pe;
  ComputePmeCUDA *compute = forceMsg->compute;

  // Store message for use in ComputePmeCUDA, where it'll also be deleted.
  if (compute->storePmeForceMsg(forceMsg)) {
    // Enqueue on the pe that sent the atoms in the first place
    LocalWorkMsg *lmsg = compute->localWorkMsg;
    CProxy_WorkDistrib wdProxy(CkpvAccess(BOCclass_group).workDistrib);
    wdProxy[pe].enqueuePme(lmsg);
  }
}

setPencilProxy() [1/3]

void ComputePmeCUDADevice::setPencilProxy

(

CProxy_CudaPmePencilXYZ

pmePencilXYZ_in

)

Definition at line 1171 of file ComputePmeCUDAMgr.C.

References NAMD_bug().

                                                                                 {
  if (pmePencilType != 3)
    NAMD_bug("ComputePmeCUDADevice::setPencilProxy(1), invalid pmePencilType");
  pmePencilXYZ = pmePencilXYZ_in;
}

setPencilProxy() [2/3]

void ComputePmeCUDADevice::setPencilProxy

(

CProxy_CudaPmePencilXY

pmePencilXY_in

)

Definition at line 1177 of file ComputePmeCUDAMgr.C.

References NAMD_bug().

                                                                               {
  if (pmePencilType != 2)
    NAMD_bug("ComputePmeCUDADevice::setPencilProxy(2), invalid pmePencilType");
  pmePencilXY = pmePencilXY_in;
}

setPencilProxy() [3/3]

void ComputePmeCUDADevice::setPencilProxy

(

CProxy_CudaPmePencilX

pmePencilX_in

)

Definition at line 1183 of file ComputePmeCUDAMgr.C.

References NAMD_bug().

                                                                             {
  if (pmePencilType != 1)
    NAMD_bug("ComputePmeCUDADevice::setPencilProxy(3), invalid pmePencilType");
  pmePencilX = pmePencilX_in;
}

spreadCharge()

void ComputePmeCUDADevice::spreadCharge

(

)

Definition at line 1647 of file ComputePmeCUDAMgr.C.

References PmeRunMsg::doEnergy, PmeRunMsg::doVirial, PmeRunMsg::lattice, NUM_GRID_MAX, PmeRunMsg::numStrayAtoms, Node::Object(), CudaAtom::q, Node::simParameters, simParams, PmeRunMsg::simulationStep, msm::swap(), CudaAtom::x, CudaAtom::y, and CudaAtom::z.

Referenced by registerRecvAtomsFromNeighbor().

                                        {
  // Spread charges in primary pencil
  int pp0 = 0-ylo + (0-zlo)*yNBlocks;
  // Primary pencil is done, finish it up before accessing it
  // (clearing is done in mergeForcesOnPatch)
  pmeAtomStorage[atomI][pp0]->finish();
  // Get the number of atoms and pointer to atoms
  int numAtoms = pmeAtomStorage[atomI][pp0]->getNumAtoms();
  CudaAtom* atoms = pmeAtomStorage[atomI][pp0]->getAtoms();
  SimParameters *simParams = Node::Object()->simParameters;
  CudaAtom* atoms2 = NULL;
  CudaAtom* atoms3 = NULL;
  CudaAtom* atoms4 = NULL;
  CudaAtom* atoms5 = NULL;
  float* chargeFactors1 = NULL;
  float* chargeFactors2 = NULL;
  float* chargeFactors3 = NULL;
  float* chargeFactors4 = NULL;
  float* chargeFactors5 = NULL;
  if (simParams->alchOn) {
    chargeFactors1 = pmeAtomStorage[atomI][pp0]->getAtomElecFactors(0);
    chargeFactors2 = pmeAtomStorage[atomI][pp0]->getAtomElecFactors(1);
    allocate_host<CudaAtom>(&atoms2, numAtoms);
    if (simParams->alchDecouple) {
      chargeFactors3 = pmeAtomStorage[atomI][pp0]->getAtomElecFactors(2);
      chargeFactors4 = pmeAtomStorage[atomI][pp0]->getAtomElecFactors(3);
      allocate_host<CudaAtom>(&atoms3, numAtoms);
      allocate_host<CudaAtom>(&atoms4, numAtoms);
    }
    if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
      chargeFactors5 = pmeAtomStorage[atomI][pp0]->getAtomElecFactors(4);
      allocate_host<CudaAtom>(&atoms5, numAtoms);
    }
  }
  // Flip atomI <-> forceI
  std::swap(atomI, forceI);
  // Re-allocate force buffer if needed
  reallocate_host<CudaForce>(&forces[0], &forceCapacities[0], numAtoms, 1.5f);
  if (simParams->alchOn) {
    reallocate_host<CudaForce>(&forces[1], &forceCapacities[1], numAtoms, 1.5f);
    if (simParams->alchDecouple) {
      reallocate_host<CudaForce>(&forces[2], &forceCapacities[2], numAtoms, 1.5f);
      reallocate_host<CudaForce>(&forces[3], &forceCapacities[3], numAtoms, 1.5f);
    }
    if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
      reallocate_host<CudaForce>(&forces[4], &forceCapacities[4], numAtoms, 1.5f);
    }
  }
  // Setup patches and atoms
  // Lattice lattice = simParams->lattice;
  if (simParams->alchOn) {
    for (int i = 0; i < numAtoms; ++i) {
      // copy atoms and scale the charges with factors
      atoms2[i].x = atoms[i].x;
      atoms2[i].y = atoms[i].y;
      atoms2[i].z = atoms[i].z;
      atoms2[i].q = atoms[i].q * chargeFactors2[i];
      if (simParams->alchDecouple) {
        atoms3[i].x = atoms[i].x;
        atoms3[i].y = atoms[i].y;
        atoms3[i].z = atoms[i].z;
        atoms3[i].q = atoms[i].q * chargeFactors3[i];
        atoms4[i].x = atoms[i].x;
        atoms4[i].y = atoms[i].y;
        atoms4[i].z = atoms[i].z;
        atoms4[i].q = atoms[i].q * chargeFactors4[i];
      }
      if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
        atoms5[i].x = atoms[i].x;
        atoms5[i].y = atoms[i].y;
        atoms5[i].z = atoms[i].z;
        atoms5[i].q = atoms[i].q * chargeFactors5[i];
      }
      atoms[i].q *= chargeFactors1[i];
    }
    pmeRealSpaceComputes[0]->copyAtoms(numAtoms, atoms);
    pmeRealSpaceComputes[1]->copyAtoms(numAtoms, atoms2);
    if (simParams->alchDecouple) {
      pmeRealSpaceComputes[2]->copyAtoms(numAtoms, atoms3);
      pmeRealSpaceComputes[3]->copyAtoms(numAtoms, atoms4);
      deallocate_host<CudaAtom>(&atoms4);
      deallocate_host<CudaAtom>(&atoms3);
    }
    if (bool(simParams->alchElecLambdaStart) == true || simParams->alchThermIntOn) {
      pmeRealSpaceComputes[4]->copyAtoms(numAtoms, atoms5);
      deallocate_host<CudaAtom>(&atoms5);
    }
    deallocate_host<CudaAtom>(&atoms2);
  } else {
    pmeRealSpaceComputes[0]->copyAtoms(numAtoms, atoms);
  }
  // Spread charge
  beforeWalltime = CmiWallTimer();
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    if (enabledGrid[iGrid] == true) {
      pmeRealSpaceComputes[iGrid]->spreadCharge(lattice);
    }
  }
  // Send "charge grid ready to PME solver"
  PmeRunMsg *pmeRunMsg = new PmeRunMsg();
  pmeRunMsg->doVirial = doVirial;
  pmeRunMsg->doEnergy = doEnergy;
  pmeRunMsg->simulationStep = simulationStep;
  pmeRunMsg->lattice = lattice;
  pmeRunMsg->numStrayAtoms = numStrayAtoms;
  // Reset stray atom counter
  numStrayAtoms = 0;
  switch(pmePencilType) {
    case 1:
    pmePencilX(0, pencilIndexY, pencilIndexZ).chargeGridReady(pmeRunMsg);
    break;
    case 2:
    pmePencilXY(0, 0, pencilIndexZ).chargeGridReady(pmeRunMsg);
    break;
    case 3:
    pmePencilXYZ[0].chargeGridReady(pmeRunMsg);
    break;
  }
}

---

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

• ComputePmeCUDAMgr.h
• ComputePmeCUDAMgr.C