CudaPmePencilX Class Reference

#include <CudaPmeSolver.h>

Inheritance diagram for CudaPmePencilX:

Public Member Functions
CudaPmePencilX_SDAG_CODE	CudaPmePencilX ()
	CudaPmePencilX (CkMigrateMessage *m)
	~CudaPmePencilX ()
void	initialize (CudaPmeXInitMsg *msg)
void	initializeDevice (InitDeviceMsg *msg)

Detailed Description

Definition at line 127 of file CudaPmeSolver.h.

Constructor & Destructor Documentation

CudaPmePencilX() [1/2]

CudaPmePencilX_SDAG_CODE CudaPmePencilX::CudaPmePencilX ( )

inline

Definition at line 130 of file CudaPmeSolver.h.

: numGetDeviceBuffer(0), eventCreated(false) {}

CudaPmePencilX() [2/2]

CudaPmePencilX::CudaPmePencilX ( CkMigrateMessage *  m )

inline

Definition at line 131 of file CudaPmeSolver.h.

: numGetDeviceBuffer(0), eventCreated(false) {}

~CudaPmePencilX()

CudaPmePencilX::~CudaPmePencilX

(

)

Definition at line 320 of file CudaPmeSolver.C.

References cudaCheck.

                                {
  if (eventCreated) cudaCheck(cudaEventDestroy(event));
}

Member Function Documentation

initialize()

void CudaPmePencilX::initialize

(

CudaPmeXInitMsg *

msg

)

Definition at line 309 of file CudaPmeSolver.C.

References CudaPmeXInitMsg::pmeGrid, CudaPmeXInitMsg::pmePencilY, and CudaPmeXInitMsg::yMap.

                                                    {
  pmeGrid = msg->pmeGrid;
  pmePencilY = msg->pmePencilY;
  yMap = msg->yMap;

  delete msg;

  initBlockSizes();

}

initializeDevice()

void CudaPmePencilX::initializeDevice

(

InitDeviceMsg *

msg

)

Definition at line 327 of file CudaPmeSolver.C.

References cudaCheck, InitDeviceMsg::deviceProxy, NUM_GRID_MAX, and Perm_cX_Y_Z.

                                                        {
  // Store device proxy
  deviceProxy = msg->deviceProxy;
  delete msg;
  deviceID = deviceProxy.ckLocalBranch()->getDeviceID();
  stream = deviceProxy.ckLocalBranch()->getStream();
  CProxy_ComputePmeCUDAMgr mgrProxy = deviceProxy.ckLocalBranch()->getMgrProxy();
  // Setup fftCompute and pmeKSpaceCompute
//   fftCompute = new CudaFFTCompute(deviceID, stream);
//   fftCompute2 = new CudaFFTCompute(deviceID, stream);
//   pmeTranspose = new CudaPmeTranspose(pmeGrid, Perm_cX_Y_Z, thisIndex.y, thisIndex.z, deviceID, stream);
//   pmeTranspose2 = new CudaPmeTranspose(pmeGrid, Perm_cX_Y_Z, thisIndex.y, thisIndex.z, deviceID, stream);
  for (unsigned int iGrid = 0; iGrid < NUM_GRID_MAX; ++iGrid) {
    if (deviceProxy.ckLocalBranch()->isGridEnabled(iGrid) == true) {
      fftComputes[iGrid] = new CudaFFTCompute(deviceID, stream);
      pmeTransposes[iGrid] = new CudaPmeTranspose(pmeGrid, Perm_cX_Y_Z, thisIndex.y, thisIndex.z, deviceID, stream);
    } else {
      fftComputes[iGrid] = NULL;
      pmeTransposes[iGrid] = NULL;
    }
  }

  // Create event. NOTE: Events are tied to devices, hence the cudaSetDevice() here
  cudaCheck(cudaSetDevice(deviceID));
  cudaCheck(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
  eventCreated = true;

  deviceBuffers.resize(pmeGrid.xBlocks, DeviceBuffer(-1, false));
  numDeviceBuffers = 0;

  for (int x=0;x < pmeGrid.xBlocks;x++) {
    int pe = yMap.ckLocalBranch()->procNum(0, CkArrayIndex3D(x,0,thisIndex.z));
    if (CkNodeOf(pe) == CkMyNode()) {
      int deviceID0 = mgrProxy.ckLocalBranch()->getDeviceIDPencilY(x, thisIndex.z);
      numDeviceBuffers++;
      deviceBuffers[x] = DeviceBuffer(deviceID0, false);
      pmePencilY(x,0,thisIndex.z).getDeviceBuffer(thisIndex.y, (deviceID0 == deviceID), thisProxy);
    }
  }

}

---

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

• CudaPmeSolver.h
• CudaPmeSolver.C