CudaPmePencilXY Class Reference

#include <CudaPmeSolver.h>

Inheritance diagram for CudaPmePencilXY:

Public Member Functions
CudaPmePencilXY_SDAG_CODE	CudaPmePencilXY ()
	CudaPmePencilXY (CkMigrateMessage *m)
	~CudaPmePencilXY ()
void	initialize (CudaPmeXYInitMsg *msg)
void	initializeDevice (InitDeviceMsg *msg)

Detailed Description

Definition at line 84 of file CudaPmeSolver.h.

Constructor & Destructor Documentation

CudaPmePencilXY_SDAG_CODE CudaPmePencilXY::CudaPmePencilXY ( )

inline

Definition at line 87 of file CudaPmeSolver.h.

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

CudaPmePencilXY::CudaPmePencilXY ( CkMigrateMessage *  m )

inline

Definition at line 88 of file CudaPmeSolver.h.

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

CudaPmePencilXY::~CudaPmePencilXY

(

)

Definition at line 66 of file CudaPmeSolver.C.

References cudaCheck.

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

Member Function Documentation

void CudaPmePencilXY::initialize

(

CudaPmeXYInitMsg *

msg

)

Definition at line 56 of file CudaPmeSolver.C.

References CudaPmeXYInitMsg::pmeGrid, CudaPmeXYInitMsg::pmePencilZ, and CudaPmeXYInitMsg::zMap.

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

  delete msg;

  initBlockSizes();
}

void CudaPmePencilXY::initializeDevice

(

InitDeviceMsg *

msg

)

Definition at line 73 of file CudaPmeSolver.C.

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

                                                         {
  // 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);
  pmeTranspose = new CudaPmeTranspose(pmeGrid, Perm_cX_Y_Z, 0, thisIndex.z, deviceID, stream);  

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

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

/*
  bool useMultiGPUfft = true;
  bool allDeviceOnSameNode = true;
  for (int x=0;x < pmeGrid.xBlocks;x++) {
    int pe = zMap.ckLocalBranch()->procNum(0, CkArrayIndex3D(x,0,0));
    allDeviceOnSameNode &= (CkNodeOf(pe) == CkMyNode());
  }

  if (useMultiGPUfft && allDeviceOnSameNode && pmeGrid.xBlocks > 1) {



  } else {
*/

  for (int x=0;x < pmeGrid.xBlocks;x++) {
    int pe = zMap.ckLocalBranch()->procNum(0, CkArrayIndex3D(x,0,0));
    if (CkNodeOf(pe) == CkMyNode()) {
      // Get device ID on a device on this node
      int deviceID0 = mgrProxy.ckLocalBranch()->getDeviceIDPencilZ(x, 0);
      // Check for Peer-to-Peer access
      int canAccessPeer = 0;
      if (deviceID != deviceID0) {
        cudaCheck(cudaSetDevice(deviceID));
        cudaCheck(cudaDeviceCanAccessPeer(&canAccessPeer, deviceID, deviceID0));
#ifdef DISABLE_P2P
        canAccessPeer = 0;
#endif
        if (canAccessPeer) {
          unsigned int flags = 0;
          cudaCheck(cudaDeviceEnablePeerAccess(deviceID0, flags));
          // fprintf(stderr, "device %d can access device %d\n", deviceID, deviceID0);
        }
      }
      numDeviceBuffers++;
      deviceBuffers[x] = DeviceBuffer(deviceID0, canAccessPeer, NULL);
      pmePencilZ(x,0,0).getDeviceBuffer(thisIndex.z, (deviceID0 == deviceID) || canAccessPeer, thisProxy);
    }
  }

  // }

}

---

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

• CudaPmeSolver.h
• CudaPmeSolver.C