CudaPmeRealSpaceCompute Class Reference

#include <CudaPmeSolverUtil.h>

Inheritance diagram for CudaPmeRealSpaceCompute:

Public Member Functions
	CudaPmeRealSpaceCompute (PmeGrid pmeGrid, const int jblock, const int kblock, int deviceID, cudaStream_t stream)
	~CudaPmeRealSpaceCompute ()
void	copyAtoms (const int numAtoms, const CudaAtom *atoms)
void	spreadCharge (Lattice &lattice)
void	gatherForce (Lattice &lattice, CudaForce *force)
void	gatherForceSetCallback (ComputePmeCUDADevice *devicePtr_in)
void	waitGatherForceDone ()
Public Member Functions inherited from PmeRealSpaceCompute
	PmeRealSpaceCompute (PmeGrid pmeGrid, const int jblock, const int kblock, unsigned int grid=0)
virtual	~PmeRealSpaceCompute ()
float *	getData ()
int	getDataSize ()
void	setGrid (unsigned int i)

Additional Inherited Members
Static Public Member Functions inherited from PmeRealSpaceCompute
static double	calcGridCoord (const double x, const double recip11, const int nfftx)
static void	calcGridCoord (const double x, const double y, const double z, const double recip11, const double recip22, const double recip33, const int nfftx, const int nffty, const int nfftz, double &frx, double &fry, double &frz)
static void	calcGridCoord (const float x, const float y, const float z, const float recip11, const float recip22, const float recip33, const int nfftx, const int nffty, const int nfftz, float &frx, float &fry, float &frz)
static void	calcGridCoord (const float x, const float y, const float z, const int nfftx, const int nffty, const int nfftz, float &frx, float &fry, float &frz)
static void	calcGridCoord (const double x, const double y, const double z, const int nfftx, const int nffty, const int nfftz, double &frx, double &fry, double &frz)
Protected Attributes inherited from PmeRealSpaceCompute
int	numAtoms
PmeGrid	pmeGrid
int	y0
int	z0
int	xsize
int	ysize
int	zsize
int	dataSize
float *	data
const int	jblock
const int	kblock
unsigned int	multipleGridIndex

Detailed Description

Definition at line 112 of file CudaPmeSolverUtil.h.

Constructor & Destructor Documentation

CudaPmeRealSpaceCompute()

CudaPmeRealSpaceCompute::CudaPmeRealSpaceCompute	(	PmeGrid	pmeGrid,
		const int	jblock,
		const int	kblock,
		int	deviceID,
		cudaStream_t	stream
	)

Definition at line 542 of file CudaPmeSolverUtil.C.

References cudaCheck, PmeRealSpaceCompute::data, PmeRealSpaceCompute::dataSize, NAMD_bug(), PmeRealSpaceCompute::xsize, PmeRealSpaceCompute::ysize, and PmeRealSpaceCompute::zsize.

                                                                         : 
  PmeRealSpaceCompute(pmeGrid, jblock, kblock), deviceID(deviceID), stream(stream) {
  if (dataSize < xsize*ysize*zsize)
    NAMD_bug("CudaPmeRealSpaceCompute::CudaPmeRealSpaceCompute, insufficient dataSize");
  cudaCheck(cudaSetDevice(deviceID));
  d_atomsCapacity = 0;
  d_atoms = NULL;
  d_forceCapacity = 0;
  d_force = NULL;
  #ifdef NAMD_CUDA
  tex_data = NULL;
  tex_data_len = 0;
  #else
  grid_data = NULL;
  grid_data_len = 0;
  #endif
  allocate_device<float>(&data, dataSize);
  setupGridData(data, xsize*ysize*zsize);
  cudaCheck(cudaEventCreate(&gatherForceEvent));
}

~CudaPmeRealSpaceCompute()

CudaPmeRealSpaceCompute::~CudaPmeRealSpaceCompute

(

)

Definition at line 567 of file CudaPmeSolverUtil.C.

References cudaCheck, and PmeRealSpaceCompute::data.

                                                  {
  cudaCheck(cudaSetDevice(deviceID));
  if (d_atoms != NULL) deallocate_device<CudaAtom>(&d_atoms);
  if (d_force != NULL) deallocate_device<CudaForce>(&d_force);
  // if (d_patches != NULL) deallocate_device<PatchInfo>(&d_patches);
  // deallocate_device<double>(&d_selfEnergy);
  deallocate_device<float>(&data);
  cudaCheck(cudaEventDestroy(gatherForceEvent));
}

Member Function Documentation

copyAtoms()

void CudaPmeRealSpaceCompute::copyAtoms ( const int  numAtoms, const CudaAtom *  atoms  )

virtual

Implements PmeRealSpaceCompute.

Definition at line 598 of file CudaPmeSolverUtil.C.

References cudaCheck, and PmeRealSpaceCompute::numAtoms.

                                                                                 {
  cudaCheck(cudaSetDevice(deviceID));
  this->numAtoms = numAtoms;

  // Reallocate device arrays as neccessary
  reallocate_device<CudaAtom>(&d_atoms, &d_atomsCapacity, numAtoms, 1.5f);

  // Copy atom data to device
  copy_HtoD<CudaAtom>(atoms, d_atoms, numAtoms, stream);
}

gatherForce()

void CudaPmeRealSpaceCompute::gatherForce ( Lattice &  lattice, CudaForce *  force  )

virtual

Implements PmeRealSpaceCompute.

Definition at line 763 of file CudaPmeSolverUtil.C.

References cudaCheck, PmeRealSpaceCompute::data, gather_force(), PmeGrid::K1, PmeGrid::K2, PmeGrid::K3, NAMD_EVENT_START, NAMD_EVENT_STOP, PmeRealSpaceCompute::numAtoms, PmeGrid::order, PmeRealSpaceCompute::pmeGrid, PmeRealSpaceCompute::xsize, PmeRealSpaceCompute::y0, PmeGrid::yBlocks, PmeRealSpaceCompute::ysize, PmeRealSpaceCompute::z0, PmeGrid::zBlocks, and PmeRealSpaceCompute::zsize.

                                                                            {
  cudaCheck(cudaSetDevice(deviceID));

  NAMD_EVENT_START(1, NamdProfileEvent::GATHER_FORCE);

  // Re-allocate force array if needed
  reallocate_device<CudaForce>(&d_force, &d_forceCapacity, numAtoms, 1.5f);

#ifdef TESTPID
  if (1) {
    fprintf(stderr, "AP gather force arguments\n");
    fprintf(stderr, "numAtoms = %d\n", numAtoms);
    fprintf(stderr, "pmeGrid.K1 = %d\n", pmeGrid.K1);
    fprintf(stderr, "pmeGrid.K2 = %d\n", pmeGrid.K2);
    fprintf(stderr, "pmeGrid.K3 = %d\n", pmeGrid.K3);
    fprintf(stderr, "xsize = %d\n", xsize);
    fprintf(stderr, "ysize = %d\n", ysize);
    fprintf(stderr, "zsize = %d\n", zsize);
    fprintf(stderr, "y0 = %d\n", y0);
    fprintf(stderr, "z0 = %d\n", z0);
    fprintf(stderr, "(pmeGrid.yBlocks == 1) = %d\n", (pmeGrid.yBlocks == 1));
    fprintf(stderr, "(pmeGrid.zBlocks == 1) = %d\n", (pmeGrid.zBlocks == 1));
    fprintf(stderr, "pmeGrid.order = %d\n", pmeGrid.order);
    fprintf(stderr, "gridTexObj = %p\n", gridTexObj);
  }
#endif
  // The patch-level PME kernels are only used for the GPU-resident code path. The default constructor
  // of PatchLevelPmeData will initialize the compatibility variables to false, so the patch-level kernels
  // won't be used here.
  PatchLevelPmeData patchLevelPmeData;  
  gather_force(patchLevelPmeData,
    (const float4*)d_atoms, numAtoms,
    pmeGrid.K1, pmeGrid.K2, pmeGrid.K3,
    xsize, ysize, zsize, xsize, y0, z0, (pmeGrid.yBlocks == 1), (pmeGrid.zBlocks == 1),
    data, pmeGrid.order, (float3*)d_force, 
#ifdef NAMD_CUDA
    gridTexObj,
#endif
    stream);
#ifdef TESTPID
  if (1) {
    cudaCheck(cudaStreamSynchronize(stream));
    fprintf(stderr, "AP GATHER FORCE\n");
    fprintf(stderr, "COPY DEVICE ARRAYS BACK TO HOST\n");
    float *xyz;
    int natoms = numAtoms;
    allocate_host<float>(&xyz, 3*natoms);
    copy_DtoH<float>((float*)d_force, xyz, 3*natoms, stream);
    cudaCheck(cudaStreamSynchronize(stream));
    TestArray_write<float>("gather_force_good.bin",
        "gather force good", xyz, 3*natoms);
    deallocate_host<float>(&xyz);
  }
#endif

  copy_DtoH<CudaForce>(d_force, force, numAtoms, stream);

  cudaCheck(cudaEventRecord(gatherForceEvent, stream));

  NAMD_EVENT_STOP(1, NamdProfileEvent::GATHER_FORCE);
}

gatherForceSetCallback()

void CudaPmeRealSpaceCompute::gatherForceSetCallback

(

ComputePmeCUDADevice *

devicePtr_in

)

Definition at line 715 of file CudaPmeSolverUtil.C.

References CcdCallBacksReset(), and cudaCheck.

                                                                                       {
  cudaCheck(cudaSetDevice(deviceID));
  devicePtr = devicePtr_in;
  checkCount = 0;
  CcdCallBacksReset(0, CmiWallTimer());
  // Set the call back at 0.1ms
  CcdCallFnAfter(cuda_gatherforce_check, this, 0.1);
}

spreadCharge()

void CudaPmeRealSpaceCompute::spreadCharge ( Lattice &  lattice )

virtual

Implements PmeRealSpaceCompute.

Definition at line 612 of file CudaPmeSolverUtil.C.

References cudaCheck, PmeRealSpaceCompute::data, PmeGrid::K1, PmeGrid::K2, PmeGrid::K3, NAMD_EVENT_START, NAMD_EVENT_STOP, PmeRealSpaceCompute::numAtoms, PmeGrid::order, PmeRealSpaceCompute::pmeGrid, spread_charge(), PmeRealSpaceCompute::xsize, PmeRealSpaceCompute::y0, PmeGrid::yBlocks, PmeRealSpaceCompute::ysize, PmeRealSpaceCompute::z0, PmeGrid::zBlocks, and PmeRealSpaceCompute::zsize.

                                                           {
  cudaCheck(cudaSetDevice(deviceID));
#if 0
  if (1) {
    static int step = 0;
    float *xyzq;
    int natoms = numAtoms;
    allocate_host<float>(&xyzq, 4*natoms);
    copy_DtoH<float>((float *)d_atoms, xyzq, 4*natoms, stream);
    cudaCheck(cudaStreamSynchronize(stream));
    char fname[64], remark[64];
    sprintf(fname, "pme_atoms_xyzq_soa_%d.bin", step);
    sprintf(remark, "SOA PME atoms xyzq, step %d\n", step);
    TestArray_write<float>(fname, remark, xyzq, 4*natoms);
    deallocate_host<float>(&xyzq);
    step += 2;
  }
#endif

  NAMD_EVENT_START(1, NamdProfileEvent::SPREAD_CHARGE);

  // Clear grid
  clear_device_array<float>(data, xsize*ysize*zsize, stream);

#if defined(TESTPID)
  fprintf(stderr, "Calling spread_charge with parameters:\n");
  fprintf(stderr, "numAtoms = %d\n", numAtoms);
  fprintf(stderr, "pmeGrid.K1 = %d\n", pmeGrid.K1);
  fprintf(stderr, "pmeGrid.K2 = %d\n", pmeGrid.K2);
  fprintf(stderr, "pmeGrid.K3 = %d\n", pmeGrid.K3);
  fprintf(stderr, "xsize = %d\n", xsize);
  fprintf(stderr, "ysize = %d\n", ysize);
  fprintf(stderr, "zsize = %d\n", zsize);
  fprintf(stderr, "y0 = %d\n", y0);
  fprintf(stderr, "z0 = %d\n", z0);
  fprintf(stderr, "(pmeGrid.yBlocks == 1) = %d\n", (pmeGrid.yBlocks == 1));
  fprintf(stderr, "(pmeGrid.zBlocks == 1) = %d\n", (pmeGrid.zBlocks == 1));
  fprintf(stderr, "pmeGrid.order = %d\n", pmeGrid.order);
#endif
  spread_charge((const float4*)d_atoms, numAtoms,
    pmeGrid.K1, pmeGrid.K2, pmeGrid.K3, xsize, ysize, zsize,
    xsize, y0, z0, (pmeGrid.yBlocks == 1), (pmeGrid.zBlocks == 1),
    data, pmeGrid.order, stream);
#ifdef TESTPID
  if (1) {
    cudaCheck(cudaStreamSynchronize(stream));
    fprintf(stderr, "AP SPREAD CHARGES\n");
    fprintf(stderr, "COPY DEVICE ARRAYS BACK TO HOST\n");
    float *xyzq;
    allocate_host<float>(&xyzq, 4*numAtoms);
    copy_DtoH<float>((float *)d_atoms, xyzq, 4*numAtoms, stream);
    int gridlen = pmeGrid.K1 * pmeGrid.K2 * pmeGrid.K3;
    float *grid;
    allocate_host<float>(&grid, gridlen);
    copy_DtoH<float>(data, grid, gridlen, stream);
    cudaCheck(cudaStreamSynchronize(stream));
    TestArray_write<float>("xyzq_good.bin", "xyzq good", xyzq, 4*numAtoms);
    TestArray_write<float>("charge_grid_good.bin", "charge grid good",
        grid, gridlen);
    deallocate_host<float>(&xyzq);
    deallocate_host<float>(&grid);
  }
#endif

  // ncall++;

  // if (ncall == 1) writeRealToDisk(data, xsize*ysize*zsize, "data.txt");
  NAMD_EVENT_STOP(1, NamdProfileEvent::SPREAD_CHARGE);
}

waitGatherForceDone()

void CudaPmeRealSpaceCompute::waitGatherForceDone

(

)

Definition at line 724 of file CudaPmeSolverUtil.C.

References cudaCheck.

                                                  {
  cudaCheck(cudaSetDevice(deviceID));
  cudaCheck(cudaEventSynchronize(gatherForceEvent));
}

---

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

• CudaPmeSolverUtil.h
• CudaPmeSolverUtil.C