CudaPmePencilY Class Reference

#include <CudaPmeSolver.h>

Inheritance diagram for CudaPmePencilY:

Public Member Functions
CudaPmePencilY_SDAG_CODE	CudaPmePencilY ()
	CudaPmePencilY (CkMigrateMessage *m)
	~CudaPmePencilY ()
void	initialize (CudaPmeXInitMsg *msg)
void	initializeDevice (InitDeviceMsg2 *msg)

Detailed Description

Definition at line 155 of file CudaPmeSolver.h.

Constructor & Destructor Documentation

CudaPmePencilY() [1/2]

CudaPmePencilY_SDAG_CODE CudaPmePencilY::CudaPmePencilY ( )

inline

Definition at line 158 of file CudaPmeSolver.h.

: numGetDeviceBufferZ(0), numGetDeviceBufferX(0), eventCreated(false) {}

CudaPmePencilY() [2/2]

CudaPmePencilY::CudaPmePencilY ( CkMigrateMessage *  m )

inline

Definition at line 159 of file CudaPmeSolver.h.

: numGetDeviceBufferZ(0), numGetDeviceBufferX(0), eventCreated(false) {}

~CudaPmePencilY()

CudaPmePencilY::~CudaPmePencilY

(

)

Definition at line 531 of file CudaPmeSolver.C.

References cudaCheck.

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

Member Function Documentation

initialize()

void CudaPmePencilY::initialize

(

CudaPmeXInitMsg *

msg

)

Definition at line 519 of file CudaPmeSolver.C.

References CudaPmeXInitMsg::pmeGrid, CudaPmeXInitMsg::pmePencilX, CudaPmeXInitMsg::pmePencilZ, CudaPmeXInitMsg::xMap, and CudaPmeXInitMsg::zMap.

                                                    {
  pmeGrid = msg->pmeGrid;
  pmePencilX = msg->pmePencilX;
  pmePencilZ = msg->pmePencilZ;
  xMap = msg->xMap;
  zMap = msg->zMap;

  delete msg;

  initBlockSizes();
}

initializeDevice()

void CudaPmePencilY::initializeDevice

(

InitDeviceMsg2 *

msg

)

Definition at line 538 of file CudaPmeSolver.C.

References cudaCheck, InitDeviceMsg2::deviceID, InitDeviceMsg2::deviceProxy, InitDeviceMsg2::mgrProxy, NUM_GRID_MAX, Perm_Y_Z_cX, and InitDeviceMsg2::stream.

                                                         {
  // Get device proxy
  CProxy_ComputePmeCUDADevice deviceProxy = msg->deviceProxy;
  deviceID = msg->deviceID;
  stream = msg->stream;
  CProxy_ComputePmeCUDAMgr mgrProxy = msg->mgrProxy;
  delete msg;
  // deviceID = deviceProxy.ckLocalBranch()->getDeviceID();
  // cudaStream_t stream = deviceProxy.ckLocalBranch()->getStream();
  // CProxy_ComputePmeCUDAMgr mgrProxy = deviceProxy.ckLocalBranch()->getMgrProxy();
  // Setup fftCompute and pmeKSpaceCompute
  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_Y_Z_cX, thisIndex.z, thisIndex.x, 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;

  deviceBuffersZ.resize(pmeGrid.yBlocks, DeviceBuffer(-1, false));
  deviceBuffersX.resize(pmeGrid.yBlocks, DeviceBuffer(-1, false));
  numDeviceBuffersZ = 0;
  numDeviceBuffersX = 0;

  for (int y=0;y < pmeGrid.yBlocks;y++) {
    int pe;
    pe = zMap.ckLocalBranch()->procNum(0, CkArrayIndex3D(thisIndex.x, y, 0));
    if (CkNodeOf(pe) == CkMyNode()) {
      int deviceID0 = mgrProxy.ckLocalBranch()->getDeviceIDPencilZ(thisIndex.x, y);
      numDeviceBuffersZ++;
      deviceBuffersZ[y] = DeviceBuffer(deviceID0, false);
      pmePencilZ(thisIndex.x, y, 0).getDeviceBuffer(thisIndex.z, (deviceID0 == deviceID), thisProxy);
    }
    pe = xMap.ckLocalBranch()->procNum(0, CkArrayIndex3D(0, y, thisIndex.z));
    if (CkNodeOf(pe) == CkMyNode()) {
      int deviceID0 = mgrProxy.ckLocalBranch()->getDeviceIDPencilX(y, thisIndex.z);
      numDeviceBuffersX++;
      deviceBuffersX[y] = DeviceBuffer(deviceID0, false);
      pmePencilX(0, y, thisIndex.z).getDeviceBuffer(thisIndex.x, (deviceID0 == deviceID), thisProxy);
    }
  }

}

---

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

• CudaPmeSolver.h
• CudaPmeSolver.C