Diffusion lammps (efficient_neuronet) Browse

pppm_cuda.cpp
fb16c0eb301befficient_neuronet

pppm_cuda.cpp
View Options

			/* ----------------------------------------------------------------------
			LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator

			Original Version:
			http://lammps.sandia.gov, Sandia National Laboratories
			Steve Plimpton, sjplimp@sandia.gov

			See the README file in the top-level LAMMPS directory.

			-----------------------------------------------------------------------

			USER-CUDA Package and associated modifications:
			https://sourceforge.net/projects/lammpscuda/

			Christian Trott, christian.trott@tu-ilmenau.de
			Lars Winterfeld, lars.winterfeld@tu-ilmenau.de
			Theoretical Physics II, University of Technology Ilmenau, Germany

			See the README file in the USER-CUDA directory.

			This software is distributed under the GNU General Public License.
			------------------------------------------------------------------------- */

			/* ----------------------------------------------------------------------
			LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
			http://lammps.sandia.gov, Sandia National Laboratories
			Steve Plimpton, sjplimp@sandia.gov

			Copyright (2003) Sandia Corporation. Under the terms of Contract
			DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
			certain rights in this software. This software is distributed under
			the GNU General Public License.

			See the README file in the top-level LAMMPS directory.
			------------------------------------------------------------------------- */

			/* ----------------------------------------------------------------------
			Contributing authors: Roy Pollock (LLNL), Paul Crozier (SNL)
			------------------------------------------------------------------------- */


			#include "mpi.h"
			#include <cstring>
			#include <cstdio>
			#include <cstdlib>
			#include <cmath>
			#include "pppm_cuda.h"
			#include "atom.h"
			#include "comm.h"
			#include "neighbor.h"
			#include "force.h"
			#include "pair.h"
			#include "bond.h"
			#include "angle.h"
			#include "domain.h"
			#include "fft3d_wrap_cuda.h"
			#include "remap_wrap.h"
			#include "memory.h"
			#include "error.h"
			#include <ctime> //crmadd
			#include "cuda_wrapper_cu.h"
			#include "pppm_cuda_cu.h"
			#include "cuda.h"
			#include "math_const.h"

			using namespace LAMMPS_NS;
			using namespace MathConst;

			#define MAXORDER 7
			#define OFFSET 4096
			#define SMALL 0.00001
			#define LARGE 10000.0
			#define EPS_HOC 1.0e-7


			void printArray(double* data,int nx, int ny, int nz)
			{
			for(int i=0;i<nz;i++)
			for(int j=0;j<ny;j++)
			{
			printf("%i %i\n",i,j);
			for(int k=0;k<nx;k++)
			printf("%e ",data[2(inynx+jnx+k)]);
			printf("\n\n");
			}
			}
			void printArray(double*** data,int nx, int ny, int nz)
			{
			for(int i=0;i<nx;i++)
			for(int j=0;j<ny;j++)
			{
			printf("%i %i\n",i,j);
			for(int k=0;k<nz;k++)
			printf("%e ",data[i][j][k]);
			printf("\n\n");
			}
			}
			/* ---------------------------------------------------------------------- */

			PPPMCuda::PPPMCuda(LAMMPS lmp, int narg, char *arg) : PPPM(lmp, (narg==2?1:narg), arg)
			{
			cuda = lmp->cuda;
			if(cuda == NULL)
			error->all(FLERR,"You cannot use a /cuda class, without activating 'cuda' acceleration. Provide '-c on' as command-line argument to LAMMPS..");

			if ((narg > 3)\|\|(narg<1)) error->all(FLERR,"Illegal kspace_style pppm/cuda command");
			#ifndef FFT_CUFFT
			error->all(FLERR,"Using kspace_style pppm/cuda without cufft is not possible. Compile with cufft=1 to include cufft. Aborting.");
			#endif
			precision = atof(arg[0]);
			if(narg>1)
			precisionmodify=arg[1][0];
			else precisionmodify='=';

			nfactors = 3;
			factors = new int[nfactors];
			factors[0] = 2;
			factors[1] = 3;
			factors[2] = 5;

			MPI_Comm_rank(world,&me);
			MPI_Comm_size(world,&nprocs);

			density_brick = vdx_brick = vdy_brick = vdz_brick = vdx_brick_tmp = NULL;
			density_fft = NULL;
			greensfn = NULL;
			work1 = work2 = NULL;
			vg = NULL;
			fkx = fky = fkz = NULL;
			buf1 = buf2 = NULL;

			gf_b = NULL;
			rho1d = rho_coeff = NULL;

			fft1c = fft2c = NULL;
			remap = NULL;

			density_brick_int=NULL;
			density_intScale=1000000;
			cu_vdx_brick = cu_vdy_brick = cu_vdz_brick = NULL;
			cu_density_brick = NULL;
			cu_density_brick_int = NULL;
			cu_density_fft = NULL;
			cu_energy=NULL;
			cu_greensfn = NULL;
			cu_work1 = cu_work2 = cu_work3 = NULL;
			cu_vg = NULL;
			cu_fkx = cu_fky = cu_fkz = NULL;

			cu_flag = NULL;
			cu_debugdata = NULL;
			cu_rho_coeff = NULL;
			cu_virial = NULL;

			cu_gf_b = NULL;

			cu_slabbuf = NULL;
			slabbuf = NULL;

			nmax = 0;
			part2grid = NULL;
			cu_part2grid = NULL;
			adev_data_array=NULL;
			poissontime=0;
			old_nmax=0;
			cu_pppm_grid_n=NULL;
			cu_pppm_grid_ids=NULL;

			pppm_grid_nmax=0;
			pppm2partgrid=new int[3];
			pppm_grid=new int[3];
			firstpass=true;
			scale = 1.0;
			}


			/* ----------------------------------------------------------------------
			free all memory
			------------------------------------------------------------------------- */

			PPPMCuda::~PPPMCuda()
			{
			delete [] slabbuf;
			delete cu_slabbuf;

			delete [] factors;
			factors=NULL;
			deallocate();
			delete cu_part2grid;
			cu_part2grid=NULL;
			memory->destroy(part2grid);
			part2grid = NULL;
			}

			/* ----------------------------------------------------------------------
			called once before run
			------------------------------------------------------------------------- */

			void PPPMCuda::init()
			{

			cuda->shared_data.pppm.cudable_force=1;

			//if(cuda->finished_run) {PPPM::init(); return;}

			if (me == 0) {
			if (screen) fprintf(screen,"PPPMCuda initialization ...\n");
			if (logfile) fprintf(logfile,"PPPMCuda initialization ...\n");
			}

			// error check

			if (domain->triclinic)
			error->all(FLERR,"Cannot (yet) use PPPMCuda with triclinic box");
			if (domain->dimension == 2) error->all(FLERR,"Cannot use PPPMCuda with 2d simulation");

			if (!atom->q_flag) error->all(FLERR,"Kspace style requires atom attribute q");

			if (slabflag == 0 && domain->nonperiodic > 0)
			error->all(FLERR,"Cannot use nonperiodic boundaries with PPPMCuda");
			if (slabflag == 1) {
			if (domain->xperiodic != 1 \|\| domain->yperiodic != 1 \|\|
			domain->boundary[2][0] != 1 \|\| domain->boundary[2][1] != 1)
			error->all(FLERR,"Incorrect boundaries with slab PPPMCuda");
			}

			if (order > MAXORDER) {
			char str[128];
			sprintf(str,"PPPMCuda order cannot be greater than %d",MAXORDER);
			error->all(FLERR,str);
			}
			// free all arrays previously allocated

			deallocate();

			// extract short-range Coulombic cutoff from pair style

			qqrd2e = force->qqrd2e;

			if (force->pair == NULL)
			error->all(FLERR,"KSpace style is incompatible with Pair style");
			int itmp=0;
			double p_cutoff = (double ) force->pair->extract("cut_coul",itmp);
			if (p_cutoff == NULL)
			error->all(FLERR,"KSpace style is incompatible with Pair style");
			cutoff = *p_cutoff;

			// if kspace is TIP4P, extract TIP4P params from pair style

			qdist = 0.0;

			if (strcmp(force->kspace_style,"pppm/tip4p") == 0) {
			if (force->pair == NULL)
			error->all(FLERR,"KSpace style is incompatible with Pair style");
			double p_qdist = (double ) force->pair->extract("qdist",itmp);
			int p_typeO = (int ) force->pair->extract("typeO",itmp);
			int p_typeH = (int ) force->pair->extract("typeH",itmp);
			int p_typeA = (int ) force->pair->extract("typeA",itmp);
			int p_typeB = (int ) force->pair->extract("typeB",itmp);
			if (!p_qdist \|\| !p_typeO \|\| !p_typeH \|\| !p_typeA \|\| !p_typeB)
			error->all(FLERR,"KSpace style is incompatible with Pair style");
			qdist = *p_qdist;
			typeO = *p_typeO;
			typeH = *p_typeH;
			int typeA = *p_typeA;
			int typeB = *p_typeB;

			if (force->angle == NULL \|\| force->bond == NULL)
			error->all(FLERR,"Bond and angle potentials must be defined for TIP4P");
			double theta = force->angle->equilibrium_angle(typeA);
			double blen = force->bond->equilibrium_distance(typeB);
			alpha = qdist / (2.0 * cos(0.5theta) blen);
			}

			// compute qsum & qsqsum and warn if not charge-neutral

			qsum = qsqsum = 0.0;
			for (int i = 0; i < atom->nlocal; i++) {
			qsum += atom->q[i];
			qsqsum += atom->q[i]*atom->q[i];
			}

			double tmp;
			MPI_Allreduce(&qsum,&tmp,1,MPI_DOUBLE,MPI_SUM,world);
			qsum = tmp;
			MPI_Allreduce(&qsqsum,&tmp,1,MPI_DOUBLE,MPI_SUM,world);
			qsqsum = tmp;

			if (qsqsum == 0.0)
			error->all(FLERR,"Cannot use kspace solver on system with no charge");
			if (fabs(qsum) > SMALL && me == 0) {
			char str[128];
			sprintf(str,"System is not charge neutral, net charge = %g",qsum);
			error->warning(FLERR,str);
			}

			// setup FFT grid resolution and g_ewald
			// normally one iteration thru while loop is all that is required
			// if grid stencil extends beyond neighbor proc, reduce order and try again

			int iteration = 0;

			while (order > 0) {

			if (iteration && me == 0)
			error->warning(FLERR,"Reducing PPPMCuda order b/c stencil extends "
			"beyond neighbor processor");
			iteration++;

			set_grid();

			if (nx_pppm >= OFFSET \|\| ny_pppm >= OFFSET \|\| nz_pppm >= OFFSET)
			error->all(FLERR,"PPPMCuda grid is too large");

			// global indices of PPPMCuda grid range from 0 to N-1
			// nlo_in,nhi_in = lower/upper limits of the 3d sub-brick of
			// global PPPMCuda grid that I own without ghost cells
			// for slab PPPMCuda, assign z grid as if it were not extended

			nxlo_in = comm->myloc[0]*nx_pppm / comm->procgrid[0];
			nxhi_in = (comm->myloc[0]+1)*nx_pppm / comm->procgrid[0] - 1;
			nylo_in = comm->myloc[1]*ny_pppm / comm->procgrid[1];
			nyhi_in = (comm->myloc[1]+1)*ny_pppm / comm->procgrid[1] - 1;
			nzlo_in = comm->myloc[2] *
			(static_cast<int> (nz_pppm/slab_volfactor)) / comm->procgrid[2];
			nzhi_in = (comm->myloc[2]+1) *
			(static_cast<int> (nz_pppm/slab_volfactor)) / comm->procgrid[2] - 1;

			// nlower,nupper = stencil size for mapping particles to PPPMCuda grid

			nlower = -(order-1)/2;
			nupper = order/2;

			// shift values for particle <-> grid mapping
			// add/subtract OFFSET to avoid int(-0.75) = 0 when want it to be -1

			if (order % 2) shift = OFFSET + 0.5;
			else shift = OFFSET;
			if (order % 2) shiftone = 0.0;
			else shiftone = 0.5;

			// nlo_out,nhi_out = lower/upper limits of the 3d sub-brick of
			// global PPPMCuda grid that my particles can contribute charge to
			// effectively nlo_in,nhi_in + ghost cells
			// nlo,nhi = global coords of grid pt to "lower left" of smallest/largest
			// position a particle in my box can be at
			// dist[3] = particle position bound = subbox + skin/2.0 + qdist
			// qdist = offset due to TIP4P fictitious charge
			// convert to triclinic if necessary
			// nlo_out,nhi_out = nlo,nhi + stencil size for particle mapping
			// for slab PPPMCuda, assign z grid as if it were not extended


			triclinic = domain->triclinic;
			double prd,sublo,*subhi;

			if (triclinic == 0) {
			prd = domain->prd;
			boxlo = domain->boxlo;
			sublo = domain->sublo;
			subhi = domain->subhi;
			} else {
			prd = domain->prd_lamda;
			boxlo = domain->boxlo_lamda;
			sublo = domain->sublo_lamda;
			subhi = domain->subhi_lamda;
			}

			double xprd = prd[0];
			double yprd = prd[1];
			double zprd = prd[2];
			double zprd_slab = zprd*slab_volfactor;

			double dist[3];
			double cuthalf = 0.5*neighbor->skin + qdist;
			if (triclinic == 0) dist[0] = dist[1] = dist[2] = cuthalf;
			else {
			dist[0] = cuthalf/domain->prd[0];
			dist[1] = cuthalf/domain->prd[1];
			dist[2] = cuthalf/domain->prd[2];
			}

			int nlo,nhi;

			nlo = static_cast<int> ((sublo[0]-dist[0]-boxlo[0]) *
			nx_pppm/xprd + shift) - OFFSET;
			nhi = static_cast<int> ((subhi[0]+dist[0]-boxlo[0]) *
			nx_pppm/xprd + shift) - OFFSET;
			nxlo_out = nlo + nlower;
			nxhi_out = nhi + nupper;

			nlo = static_cast<int> ((sublo[1]-dist[1]-boxlo[1]) *
			ny_pppm/yprd + shift) - OFFSET;
			nhi = static_cast<int> ((subhi[1]+dist[1]-boxlo[1]) *
			ny_pppm/yprd + shift) - OFFSET;
			nylo_out = nlo + nlower;
			nyhi_out = nhi + nupper;

			nlo = static_cast<int> ((sublo[2]-dist[2]-boxlo[2]) *
			nz_pppm/zprd_slab + shift) - OFFSET;
			nhi = static_cast<int> ((subhi[2]+dist[2]-boxlo[2]) *
			nz_pppm/zprd_slab + shift) - OFFSET;
			nzlo_out = nlo + nlower;
			nzhi_out = nhi + nupper;

			// for slab PPPMCuda, change the grid boundary for processors at +z end
			// to include the empty volume between periodically repeating slabs
			// for slab PPPMCuda, want charge data communicated from -z proc to +z proc,
			// but not vice versa, also want field data communicated from +z proc to
			// -z proc, but not vice versa
			// this is accomplished by nzhi_in = nzhi_out on +z end (no ghost cells)

			if (slabflag && ((comm->myloc[2]+1) == (comm->procgrid[2]))) {
			nzhi_in = nz_pppm - 1;
			nzhi_out = nz_pppm - 1;
			}

			// nlo_ghost,nhi_ghost = # of planes I will recv from 6 directions
			// that overlay domain I own
			// proc in that direction tells me via sendrecv()
			// if no neighbor proc, value is from self since I have ghosts regardless

			int nplanes;
			MPI_Status status;

			nplanes = nxlo_in - nxlo_out;
			if (comm->procneigh[0][0] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[0][0],0,
			&nxhi_ghost,1,MPI_INT,comm->procneigh[0][1],0,
			world,&status);
			else nxhi_ghost = nplanes;

			nplanes = nxhi_out - nxhi_in;
			if (comm->procneigh[0][1] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[0][1],0,
			&nxlo_ghost,1,MPI_INT,comm->procneigh[0][0],
			0,world,&status);
			else nxlo_ghost = nplanes;

			nplanes = nylo_in - nylo_out;
			if (comm->procneigh[1][0] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[1][0],0,
			&nyhi_ghost,1,MPI_INT,comm->procneigh[1][1],0,
			world,&status);
			else nyhi_ghost = nplanes;

			nplanes = nyhi_out - nyhi_in;
			if (comm->procneigh[1][1] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[1][1],0,
			&nylo_ghost,1,MPI_INT,comm->procneigh[1][0],0,
			world,&status);
			else nylo_ghost = nplanes;

			nplanes = nzlo_in - nzlo_out;
			if (comm->procneigh[2][0] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[2][0],0,
			&nzhi_ghost,1,MPI_INT,comm->procneigh[2][1],0,
			world,&status);
			else nzhi_ghost = nplanes;

			nplanes = nzhi_out - nzhi_in;
			if (comm->procneigh[2][1] != me)
			MPI_Sendrecv(&nplanes,1,MPI_INT,comm->procneigh[2][1],0,
			&nzlo_ghost,1,MPI_INT,comm->procneigh[2][0],0,
			world,&status);
			else nzlo_ghost = nplanes;

			// test that ghost overlap is not bigger than my sub-domain

			int flag = 0;
			if (nxlo_ghost > nxhi_in-nxlo_in+1) flag = 1;
			if (nxhi_ghost > nxhi_in-nxlo_in+1) flag = 1;
			if (nylo_ghost > nyhi_in-nylo_in+1) flag = 1;
			if (nyhi_ghost > nyhi_in-nylo_in+1) flag = 1;
			if (nzlo_ghost > nzhi_in-nzlo_in+1) flag = 1;
			if (nzhi_ghost > nzhi_in-nzlo_in+1) flag = 1;

			int flag_all;
			MPI_Allreduce(&flag,&flag_all,1,MPI_INT,MPI_SUM,world);

			if (flag_all == 0) break;
			order--;
			}

			if (order == 0) error->all(FLERR,"PPPMCuda order has been reduced to 0");
			//printf("PPPMCuda: order is %i\n");



			// decomposition of FFT mesh
			// global indices range from 0 to N-1
			// proc owns entire x-dimension, clump of columns in y,z dimensions
			// npey_fft,npez_fft = # of procs in y,z dims
			// if nprocs is small enough, proc can own 1 or more entire xy planes,
			// else proc owns 2d sub-blocks of yz plane
			// me_y,me_z = which proc (0-npe_fft-1) I am in y,z dimensions
			// nlo_fft,nhi_fft = lower/upper limit of the section
			// of the global FFT mesh that I own

			int npey_fft,npez_fft;
			if (nz_pppm >= nprocs) {
			npey_fft = 1;
			npez_fft = nprocs;
			} else procs2grid2d(nprocs,ny_pppm,nz_pppm,&npey_fft,&npez_fft);

			int me_y = me % npey_fft;
			int me_z = me / npey_fft;

			nxlo_fft = 0;
			nxhi_fft = nx_pppm - 1;
			nylo_fft = me_y*ny_pppm/npey_fft;
			nyhi_fft = (me_y+1)*ny_pppm/npey_fft - 1;
			nzlo_fft = me_z*nz_pppm/npez_fft;
			nzhi_fft = (me_z+1)*nz_pppm/npez_fft - 1;

			// PPPMCuda grid for this proc, including ghosts

			ngrid = (nxhi_out-nxlo_out+1) * (nyhi_out-nylo_out+1) *
			(nzhi_out-nzlo_out+1);

			// FFT arrays on this proc, without ghosts
			// nfft = FFT points in FFT decomposition on this proc
			// nfft_brick = FFT points in 3d brick-decomposition on this proc
			// nfft_both = greater of 2 values

			nfft = (nxhi_fft-nxlo_fft+1) * (nyhi_fft-nylo_fft+1) *
			(nzhi_fft-nzlo_fft+1);
			int nfft_brick = (nxhi_in-nxlo_in+1) * (nyhi_in-nylo_in+1) *
			(nzhi_in-nzlo_in+1);
			nfft_both = MAX(nfft,nfft_brick);

			// buffer space for use in brick2fft and fillbrick
			// idel = max # of ghost planes to send or recv in +/- dir of each dim
			// nx,ny,nz = owned planes (including ghosts) in each dim
			// nxx,nyy,nzz = max # of grid cells to send in each dim
			// nbuf = max in any dim, augment by 3x for components of vd_xyz in fillbrick

			int idelx,idely,idelz,nx,ny,nz,nxx,nyy,nzz;

			idelx = MAX(nxlo_ghost,nxhi_ghost);
			idelx = MAX(idelx,nxhi_out-nxhi_in);
			idelx = MAX(idelx,nxlo_in-nxlo_out);

			idely = MAX(nylo_ghost,nyhi_ghost);
			idely = MAX(idely,nyhi_out-nyhi_in);
			idely = MAX(idely,nylo_in-nylo_out);

			idelz = MAX(nzlo_ghost,nzhi_ghost);
			idelz = MAX(idelz,nzhi_out-nzhi_in);
			idelz = MAX(idelz,nzlo_in-nzlo_out);

			nx = nxhi_out - nxlo_out + 1;
			ny = nyhi_out - nylo_out + 1;
			nz = nzhi_out - nzlo_out + 1;

			nxx = idelx * ny * nz;
			nyy = idely * nx * nz;
			nzz = idelz * nx * ny;

			nbuf = MAX(nxx,nyy);
			nbuf = MAX(nbuf,nzz);
			nbuf *= 3;

			// print stats

			int ngrid_max,nfft_both_max,nbuf_max;
			MPI_Allreduce(&ngrid,&ngrid_max,1,MPI_INT,MPI_MAX,world);
			MPI_Allreduce(&nfft_both,&nfft_both_max,1,MPI_INT,MPI_MAX,world);
			MPI_Allreduce(&nbuf,&nbuf_max,1,MPI_INT,MPI_MAX,world);

			if (me == 0) {
			if (screen) fprintf(screen," brick FFT buffer size/proc = %d %d %d\n",
			ngrid_max,nfft_both_max,nbuf_max);
			if (logfile) fprintf(logfile," brick FFT buffer size/proc = %d %d %d\n",
			ngrid_max,nfft_both_max,nbuf_max);
			}
			cuda_shared_pppm* ap=&(cuda->shared_data.pppm);

			ap->density_intScale=density_intScale;
			ap->nxlo_in=nxlo_in;
			ap->nxhi_in=nxhi_in;
			ap->nxlo_out=nxlo_out;
			ap->nxhi_out=nxhi_out;
			ap->nylo_in=nylo_in;
			ap->nyhi_in=nyhi_in;
			ap->nylo_out=nylo_out;
			ap->nyhi_out=nyhi_out;
			ap->nzlo_in=nzlo_in;
			ap->nzhi_in=nzhi_in;
			ap->nzlo_out=nzlo_out;
			ap->nzhi_out=nzhi_out;
			ap->nxlo_in=nxlo_fft;
			ap->nxhi_in=nxhi_fft;
			ap->nylo_in=nylo_fft;
			ap->nyhi_in=nyhi_fft;
			ap->nzlo_in=nzlo_fft;
			ap->nzhi_in=nzhi_fft;
			ap->nx_pppm=nx_pppm;
			ap->ny_pppm=ny_pppm;
			ap->nz_pppm=nz_pppm;
			ap->qqrd2e=qqrd2e;
			ap->order=order;
			ap->nmax=nmax;
			ap->nlocal=atom->nlocal;
			ap->delxinv=delxinv;
			ap->delyinv=delyinv;
			ap->delzinv=delzinv;
			ap->nlower=nlower;
			ap->nupper=nupper;
			ap->shiftone=shiftone;

			// allocate K-space dependent memory


			allocate();

			// pre-compute Green's function denomiator expansion
			// pre-compute 1d charge distribution coefficients

			compute_gf_denom();
			compute_rho_coeff();
			}

			/* ----------------------------------------------------------------------
			adjust PPPMCuda coeffs, called initially and whenever volume has changed
			------------------------------------------------------------------------- */

			void PPPMCuda::setup()
			{
			int i,j,k,l,m,n;
			double *prd;
			cu_gf_b->upload();
			// volume-dependent factors
			// adjust z dimension for 2d slab PPPMCuda
			// z dimension for 3d PPPMCuda is zprd since slab_volfactor = 1.0

			if (triclinic == 0) prd = domain->prd;
			else prd = domain->prd_lamda;

			double xprd = prd[0];
			double yprd = prd[1];
			double zprd = prd[2];
			double zprd_slab = zprd*slab_volfactor;
			volume = xprd * yprd * zprd_slab;

			delxinv = nx_pppm/xprd;
			delyinv = ny_pppm/yprd;
			delzinv = nz_pppm/zprd_slab;

			delvolinv = delxinvdelyinvdelzinv;

			double unitkx = (2.0*MY_PI/xprd);
			double unitky = (2.0*MY_PI/yprd);
			double unitkz = (2.0*MY_PI/zprd_slab);

			// fkx,fky,fkz for my FFT grid pts
			Cuda_PPPM_Setup_fkxyz_vg(nx_pppm, ny_pppm,nz_pppm,unitkx,unitky,unitkz,g_ewald);


			/* cu_vg->download();
			int offset=8100-2;//10(nxhi_fft-nxlo_fft+1)(nyhi_fft-nylo_fft+1)+10*(nyhi_fft-nylo_fft+1);
			for (int i=nxlo_fft; i <= nxhi_fft+1;i++) printf("%e ",vg[i-nxlo_fft+offset][0]);
			printf("\n\n");
			double per;

			#ifndef FFT_CUFFT
			for (i = nxlo_fft; i <= nxhi_fft; i++) {
			per = i - nx_pppm(2i/nx_pppm);
			fkx[i] = unitkx*per;
			}

			for (i = nylo_fft; i <= nyhi_fft; i++) {
			per = i - ny_pppm(2i/ny_pppm);
			fky[i] = unitky*per;
			}

			for (i = nzlo_fft; i <= nzhi_fft; i++) {
			per = i - nz_pppm(2i/nz_pppm);
			fkz[i] = unitkz*per;
			}
			#endif
			#ifdef FFT_CUFFT
			for (i = 0; i < nx_pppm; i++) {
			per = i - nx_pppm(2i/nx_pppm);
			fkx[i] = unitkx*per;
			}

			for (i = 0; i < ny_pppm; i++) {
			per = i - ny_pppm(2i/ny_pppm);
			fky[i] = unitky*per;
			}

			for (i = 0; i < nz_pppm; i++) {
			per = i - nz_pppm(2i/nz_pppm);
			fkz[i] = unitkz*per;
			}
			#endif

			// virial coefficients

			double sqk,vterm;
			int save_n=0;
			int s_i,s_j,s_k;
			double max=0.0;
			n = 0;
			for (k = nzlo_fft; k <= nzhi_fft; k++) {
			for (j = nylo_fft; j <= nyhi_fft; j++) {
			for (i = nxlo_fft; i <= nxhi_fft; i++) {
			sqk = fkx[i]fkx[i] + fky[j]fky[j] + fkz[k]*fkz[k];
			if(n==8100) printf("%lf\n",sqk);
			if (sqk == 0.0) {
			vg[n][0] = 0.0;
			vg[n][1] = 0.0;
			vg[n][2] = 0.0;
			vg[n][3] = 0.0;
			vg[n][4] = 0.0;
			vg[n][5] = 0.0;
			} else {
			vterm = -2.0 * (1.0/sqk + 0.25/(g_ewald*g_ewald));
			double tmp=vg[n][0];
			vg[n][0] = 1.0 + vtermfkx[i]fkx[i];
			if(((vg[n][0]-tmp)*(vg[n][0]-tmp)>1e-6)&&(save_n==0)) {save_n=n;s_k=k;s_j=j;s_i=i;}
			vg[n][1] = 1.0 + vtermfky[j]fky[j];
			vg[n][2] = 1.0 + vtermfkz[k]fkz[k];
			vg[n][3] = vtermfkx[i]fky[j];
			vg[n][4] = vtermfkx[i]fkz[k];
			vg[n][5] = vtermfky[j]fkz[k];
			//if(vg[n][0]>max) {max=vg[n][0]; save_n=n;}
			}
			n++;
			}
			}
			}
			printf("%lf %i %i %i %i\n",max,save_n,s_k,s_j,s_i);
			for (int i=nxlo_fft; i <= nxhi_fft;i++) printf("%e ",vg[i-nxlo_fft+offset][0]);
			printf("\n\n");

			//cu_fkx->upload();
			//cu_fky->upload();
			// cu_fkz->upload();
			//cu_vg->upload(); */
			// modified (Hockney-Eastwood) Coulomb Green's function

			double sqk;
			int nx,ny,nz,kper,lper,mper;
			double snx,sny,snz,snx2,sny2,snz2;
			double argx,argy,argz,wx,wy,wz,sx,sy,sz,qx,qy,qz;
			double sum1,dot1,dot2;
			double numerator,denominator;

			int nbx = static_cast<int> ((g_ewaldxprd/(MY_PInx_pppm)) *
			pow(-log(EPS_HOC),0.25));
			int nby = static_cast<int> ((g_ewaldyprd/(MY_PIny_pppm)) *
			pow(-log(EPS_HOC),0.25));
			int nbz = static_cast<int> ((g_ewaldzprd_slab/(MY_PInz_pppm)) *
			pow(-log(EPS_HOC),0.25));
			Cuda_PPPM_setup_greensfn(nx_pppm,ny_pppm,nz_pppm,unitkx,unitky,unitkz,g_ewald,
			nbx,nby,nbz,xprd,yprd,zprd_slab);
			/*
			double form = 1.0;

			n = 0;
			#ifndef FFT_CUFFT
			for (m = nzlo_fft; m <= nzhi_fft; m++) {
			#endif
			#ifdef FFT_CUFFT
			for (m = 0; m < nz_pppm; m++) {
			#endif
			mper = m - nz_pppm(2m/nz_pppm);
			snz = sin(0.5unitkzmper*zprd_slab/nz_pppm);
			snz2 = snz*snz;

			#ifndef FFT_CUFFT
			for (l = nylo_fft; l <= nyhi_fft; l++) {
			#endif
			#ifdef FFT_CUFFT
			for (l = 0; l < ny_pppm; l++) {
			#endif
			lper = l - ny_pppm(2l/ny_pppm);
			sny = sin(0.5unitkylper*yprd/ny_pppm);
			sny2 = sny*sny;

			#ifndef FFT_CUFFT
			for (k = nxlo_fft; k <= nxhi_fft; k++) {
			#endif
			#ifdef FFT_CUFFT
			for (k = 0; k < nx_pppm; k++) {
			#endif
			kper = k - nx_pppm(2k/nx_pppm);
			snx = sin(0.5unitkxkper*xprd/nx_pppm);
			snx2 = snx*snx;

			sqk = pow(unitkxkper,2.0) + pow(unitkylper,2.0) +
			pow(unitkz*mper,2.0);

			if (sqk != 0.0) {
			numerator = form*12.5663706/sqk;
			denominator = gf_denom(snx2,sny2,snz2);
			sum1 = 0.0;
			for (nx = -nbx; nx <= nbx; nx++) {
			qx = unitkx(kper+nx_pppmnx);
			sx = exp(-.25*pow(qx/g_ewald,2.0));
			wx = 1.0;
			argx = 0.5qxxprd/nx_pppm;
			if (argx != 0.0) wx = pow(sin(argx)/argx,order);
			for (ny = -nby; ny <= nby; ny++) {
			qy = unitky(lper+ny_pppmny);
			sy = exp(-.25*pow(qy/g_ewald,2.0));
			wy = 1.0;
			argy = 0.5qyyprd/ny_pppm;
			if (argy != 0.0) wy = pow(sin(argy)/argy,order);
			for (nz = -nbz; nz <= nbz; nz++) {
			qz = unitkz(mper+nz_pppmnz);
			sz = exp(-.25*pow(qz/g_ewald,2.0));
			wz = 1.0;
			argz = 0.5qzzprd_slab/nz_pppm;
			if (argz != 0.0) wz = pow(sin(argz)/argz,order);

			dot1 = unitkxkperqx + unitkylperqy + unitkzmperqz;
			dot2 = qxqx+qyqy+qz*qz;
			sum1 += (dot1/dot2) * sxsysz * pow(wxwywz,2.0);
			}
			}
			}
			greensfn[n++] = numerator*sum1/denominator;
			} else greensfn[n++] = 0.0;
			}
			}
			}*/

			#ifdef FFT_CUFFT
			//cu_greensfn->upload();
			//cu_fkx->upload();
			//cu_fky->upload();
			//cu_fkz->upload();
			//cu_vg->upload();
			cu_vdx_brick->upload();
			cu_vdy_brick->upload();
			cu_vdz_brick->upload();

			#endif
			cu_rho_coeff->upload();
			cu_density_brick->memset_device(0);
			pppm_device_init_setup(&cuda->shared_data,shiftone,delxinv,delyinv,delzinv,nlower,nupper);
			}

			/* ----------------------------------------------------------------------
			compute the PPPMCuda long-range force, energy, virial
			------------------------------------------------------------------------- */

			void PPPMCuda::compute(int eflag, int vflag)
			{

			// printf("PPPMCuda::compute START\n");
			cuda_shared_atom* cu_atom = & cuda->shared_data.atom;

			int i;
			timespec starttime;
			timespec endtime;
			timespec starttotal;
			timespec endtotal;
			// convert atoms from box to lamda coords

			if (triclinic == 0) boxlo = domain->boxlo;
			else {
			boxlo = domain->boxlo_lamda;
			domain->x2lamda(atom->nlocal);
			}

			// extend size of per-atom arrays if necessary

			if ((cu_atom->update_nmax)\|\|(old_nmax==0)) {
			memory->destroy(part2grid);
			nmax = atom->nmax;
			memory->create(part2grid,nmax,3,"pppm:part2grid");
			delete cu_part2grid;
			delete [] adev_data_array;
			adev_data_array=new dev_array[1];
			cu_part2grid = new cCudaData<int , int , yx > ((int*)part2grid,adev_data_array, nmax,3);

			pppm_device_update(&cuda->shared_data,cu_part2grid->dev_data(),atom->nlocal,atom->nmax);
			old_nmax=nmax;
			}
			if(cu_atom->update_nlocal) {pppm_update_nlocal(cu_atom->nlocal);}

			energy = 0.0;
			if (vflag)
			{
			for (i = 0; i < 6; i++) virial[i] = 0.0;
			cu_virial->memset_device(0);
			}
			if(eflag) cu_energy->memset_device(0);
			clock_gettime(CLOCK_REALTIME,&starttotal);

			// find grid points for all my particles
			// map my particle charge onto my local 3d density grid


			clock_gettime(CLOCK_REALTIME,&starttime);

			particle_map();

			clock_gettime(CLOCK_REALTIME,&endtime);
			cuda->shared_data.cuda_timings.pppm_particle_map+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			//cu_part2grid->download();
			clock_gettime(CLOCK_REALTIME,&starttime);
			make_rho();
			clock_gettime(CLOCK_REALTIME,&endtime);
			cuda->shared_data.cuda_timings.pppm_make_rho+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			// all procs communicate density values from their ghost cells
			// to fully sum contribution in their 3d bricks
			// remap from 3d decomposition to FFT decomposition

			int nprocs=comm->nprocs;

			clock_gettime(CLOCK_REALTIME,&starttime);

			if(nprocs>1)
			{
			cu_density_brick->download();
			brick2fft();
			}
			else
			{
			#ifdef FFT_CUFFT
			pppm_initfftdata(&cuda->shared_data,(PPPM_FLOAT)cu_density_brick->dev_data(),(FFT_FLOAT)cu_work2->dev_data());
			#endif
			}

			clock_gettime(CLOCK_REALTIME,&endtime);
			cuda->shared_data.cuda_timings.pppm_brick2fft+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			// compute potential gradient on my FFT grid and
			// portion of e_long on this proc's FFT grid
			// return gradients (electric fields) in 3d brick decomposition

			clock_gettime(CLOCK_REALTIME,&starttime);
			poisson(eflag,vflag);
			clock_gettime(CLOCK_REALTIME,&endtime);
			cuda->shared_data.cuda_timings.pppm_poisson+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			// all procs communicate E-field values to fill ghost cells
			// surrounding their 3d bricks

			// not necessary since all the calculations are done on one proc

			// calculate the force on my particles
			//cu_vdx_brick->download();
			//cu_vdy_brick->download();
			//cu_vdz_brick->download();

			clock_gettime(CLOCK_REALTIME,&starttime);
			fieldforce();
			clock_gettime(CLOCK_REALTIME,&endtime);
			cuda->shared_data.cuda_timings.pppm_fieldforce+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			// sum energy across procs and add in volume-dependent term

			clock_gettime(CLOCK_REALTIME,&endtotal);
			cuda->shared_data.cuda_timings.pppm_compute+=(endtotal.tv_sec-starttotal.tv_sec+1.0*(endtotal.tv_nsec-starttotal.tv_nsec)/1000000000);

			if (eflag) {
			double energy_all;
			MPI_Allreduce(&energy,&energy_all,1,MPI_DOUBLE,MPI_SUM,world);
			energy = energy_all;

			energy = 0.5volume;
			energy -= g_ewald*qsqsum/1.772453851 +
			MY_PI2qsumqsum / (g_ewaldg_ewaldvolume);
			energy *= qqrd2e;
			}

			// sum virial across procs

			if (vflag) {
			double virial_all[6];
			MPI_Allreduce(virial,virial_all,6,MPI_DOUBLE,MPI_SUM,world);
			for (i = 0; i < 6; i++) virial[i] = 0.5qqrd2evolume*virial_all[i];
			}

			// 2d slab correction

			if (slabflag) slabcorr(eflag);

			// convert atoms back from lamda to box coords

			if (triclinic) domain->lamda2x(atom->nlocal);

			if(firstpass) firstpass=false;
			}


			/* ----------------------------------------------------------------------
			allocate memory that depends on # of K-vectors and order
			------------------------------------------------------------------------- */


			void PPPMCuda::allocate()
			{
			//printf("PPPMCuda::allocate START Mem: %i\n",CudaWrapper_CheckMemUseage());
			/*if(sizeof(CUDA_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision\n");

			#ifdef PPPM_PRECISION
			if(sizeof(PPPM_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision for pppm core\n");
			if(sizeof(PPPM_FLOAT)==sizeof(double)) printf("PPPMCuda: Using double precision for pppm core\n");
			#endif
			#ifdef ENERGY_PRECISION
			if(sizeof(ENERGY_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision for energy\n");
			if(sizeof(ENERGY_FLOAT)==sizeof(double)) printf("PPPMCuda: Using double precision for energy\n");
			#endif
			#ifdef ENERGY_PRECISION
			if(sizeof(FFT_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision for fft\n");
			if(sizeof(FFT_FLOAT)==sizeof(double)) printf("PPPMCuda: Using double precision for fft\n");
			#endif
			#ifdef X_PRECISION
			if(sizeof(X_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision for positions\n");
			if(sizeof(X_FLOAT)==sizeof(double)) printf("PPPMCuda: Using double precision for positions\n");
			#endif
			#ifdef F_PRECISION
			if(sizeof(F_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision for forces\n");
			if(sizeof(F_FLOAT)==sizeof(double)) printf("PPPMCuda: Using double precision for forces\n");
			#endif*/

			//if(sizeof(PPPM_FLOAT)==sizeof(float)) printf("PPPMCuda: Using single precision\n");

			struct dev_array* dev_tmp=new struct dev_array[20];
			int n_cudata=0;


			memory->create3d_offset(density_brick,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:density_brick");
			memory->create3d_offset(density_brick_int,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:density_brick_int");


			cu_density_brick = new cCudaData<double, PPPM_FLOAT, x> ((double*) &(density_brick[nzlo_out][nylo_out][nxlo_out]), & (dev_tmp[n_cudata++]),
			(nzhi_out-nzlo_out+1)(nyhi_out-nylo_out+1)(nxhi_out-nxlo_out+1));

			cu_density_brick_int = new cCudaData<int, int, x> ((int*) &(density_brick_int[nzlo_out][nylo_out][nxlo_out]), & (dev_tmp[n_cudata++]),
			(nzhi_out-nzlo_out+1)(nyhi_out-nylo_out+1)(nxhi_out-nxlo_out+1));

			memory->create3d_offset(vdx_brick,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:vdx_brick");
			memory->create3d_offset(vdx_brick_tmp,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:vdx_brick_tmp");

			cu_vdx_brick = new cCudaData<double, PPPM_FLOAT, x> ((double*) &(vdx_brick[nzlo_out][nylo_out][nxlo_out]), & (dev_tmp[n_cudata++]),
			(nzhi_out-nzlo_out+1)(nyhi_out-nylo_out+1)(nxhi_out-nxlo_out+1));

			memory->create3d_offset(vdy_brick,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:vdy_brick");
			cu_vdy_brick = new cCudaData<double, PPPM_FLOAT, x> ((double*) &(vdy_brick[nzlo_out][nylo_out][nxlo_out]), & (dev_tmp[n_cudata++]),
			(nzhi_out-nzlo_out+1)(nyhi_out-nylo_out+1)(nxhi_out-nxlo_out+1));

			memory->create3d_offset(vdz_brick,nzlo_out,nzhi_out,nylo_out,nyhi_out,
			nxlo_out,nxhi_out,"pppm:vdz_brick");
			cu_vdz_brick = new cCudaData<double, PPPM_FLOAT, x> ((double*) &(vdz_brick[nzlo_out][nylo_out][nxlo_out]), & (dev_tmp[n_cudata++]),
			(nzhi_out-nzlo_out+1)(nyhi_out-nylo_out+1)(nxhi_out-nxlo_out+1));

			memory->create(density_fft,nfft_both,"pppm:density_fft");

			cu_density_fft = new cCudaData<double, PPPM_FLOAT, x> (density_fft, & (dev_tmp[n_cudata++]),nfft_both);

			cu_energy = new cCudaData<double, ENERGY_FLOAT, x> (NULL, &(dev_tmp[n_cudata++]),ny_pppm*nz_pppm);
			cu_virial = new cCudaData<double, ENERGY_FLOAT, x> (NULL, &(dev_tmp[n_cudata++]),ny_pppmnz_pppm6);

			memory->create(greensfn,nfft_both,"pppm:greensfn");
			cu_greensfn = new cCudaData<double, PPPM_FLOAT, x> (greensfn, & (dev_tmp[n_cudata++]) , nx_pppmny_pppmnz_pppm);

			memory->create(work1,2nx_pppmny_pppm*nz_pppm,"pppm:work1");
			memory->create(work2,2nx_pppmny_pppm*nz_pppm,"pppm:work2");
			memory->create(work3,2nx_pppmny_pppm*nz_pppm,"pppm:work3");

			cu_work1 = new cCudaData<double, FFT_FLOAT, x> (work1, & (dev_tmp[n_cudata++]) , 2nx_pppmny_pppm*nz_pppm);
			cu_work2 = new cCudaData<double, FFT_FLOAT, x> (work2, & (dev_tmp[n_cudata++]) , 2nx_pppmny_pppm*nz_pppm);
			cu_work3 = new cCudaData<double, FFT_FLOAT, x> (work3, & (dev_tmp[n_cudata++]) , 2nx_pppmny_pppm*nz_pppm);


			memory->create(fkx,nx_pppm,"pppmcuda:fkx");
			cu_fkx = new cCudaData<double, PPPM_FLOAT, x> (fkx, & (dev_tmp[n_cudata++]) , nx_pppm);
			memory->create(fky,ny_pppm,"pppmcuda:fky");
			cu_fky = new cCudaData<double, PPPM_FLOAT, x> (fky, & (dev_tmp[n_cudata++]) , ny_pppm);
			memory->create(fkz,nz_pppm,"pppmcuda:fkz");
			cu_fkz = new cCudaData<double, PPPM_FLOAT, x> (fkz, & (dev_tmp[n_cudata++]) , nz_pppm);

			memory->create(vg,nfft_both,6,"pppm:vg");

			cu_vg = new cCudaData<double, PPPM_FLOAT, xy> ((double*)vg, & (dev_tmp[n_cudata++]) , nfft_both,6);

			memory->create(buf1,nbuf,"pppm:buf1");
			memory->create(buf2,nbuf,"pppm:buf2");


			// summation coeffs


			gf_b = new double[order];
			cu_gf_b = new cCudaData<double,PPPM_FLOAT,x> (gf_b, &(dev_tmp[n_cudata++]) , order);
			memory->create2d_offset(rho1d,3,-order/2,order/2,"pppm:rho1d");
			memory->create2d_offset(rho_coeff,order,(1-order)/2,order/2,"pppm:rho_coeff");

			cu_rho_coeff = new cCudaData<double, PPPM_FLOAT, x> ((double) &(rho_coeff[0][(1-order)/2]), & (dev_tmp[n_cudata++]) , order(order/2-(1-order)/2+1));

			debugdata=new PPPM_FLOAT[100];
			cu_debugdata = new cCudaData<PPPM_FLOAT, PPPM_FLOAT, x> (debugdata,& (dev_tmp[n_cudata++]),100);
			cu_flag = new cCudaData<int, int, x> (&global_flag,& (dev_tmp[n_cudata++]),3);

			// create 2 FFTs and a Remap
			// 1st FFT keeps data in FFT decompostion
			// 2nd FFT returns data in 3d brick decomposition
			// remap takes data from 3d brick to FFT decomposition

			int tmp;




			fft1c = new FFT3dCuda(lmp,world,nx_pppm,ny_pppm,nz_pppm,
			nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,
			nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,
			0,0,&tmp,true);

			fft2c = new FFT3dCuda(lmp,world,nx_pppm,ny_pppm,nz_pppm,
			nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,
			nxlo_in,nxhi_in,nylo_in,nyhi_in,nzlo_in,nzhi_in,
			0,0,&tmp,false);


			#ifdef FFT_CUFFT
			fft1c->set_cudata(cu_work2->dev_data(),cu_work1->dev_data());
			fft2c->set_cudata(cu_work2->dev_data(),cu_work3->dev_data());
			#endif

			remap = new Remap(lmp,world,
			nxlo_in,nxhi_in,nylo_in,nyhi_in,nzlo_in,nzhi_in,
			nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,
			1,0,0,2);


			pppm_device_init(cu_density_brick->dev_data(), cu_vdx_brick->dev_data(), cu_vdy_brick->dev_data(), cu_vdz_brick->dev_data(), cu_density_fft->dev_data(),cu_energy->dev_data(),cu_virial->dev_data()
			, cu_work1->dev_data(), cu_work2->dev_data(), cu_work3->dev_data(), cu_greensfn->dev_data(), cu_fkx->dev_data(), cu_fky->dev_data(), cu_fkz->dev_data(), cu_vg->dev_data()
			,nxlo_in,nxhi_in,nylo_in,nyhi_in,nzlo_in,nzhi_in,nxlo_out,nxhi_out,nylo_out,nyhi_out,nzlo_out,nzhi_out,nx_pppm,ny_pppm,nz_pppm
			,nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,cu_gf_b->dev_data()
			,qqrd2e,order,cu_rho_coeff->dev_data(),cu_debugdata->dev_data(),cu_density_brick_int->dev_data(),slabflag
			);
			}



			/* ----------------------------------------------------------------------
			deallocate memory that depends on # of K-vectors and order
			---------------------------------------------------------------------- */

			void PPPMCuda::deallocate()
			{
			memory->destroy3d_offset(density_brick,nzlo_out,nylo_out,nxlo_out);
			memory->destroy3d_offset(vdx_brick,nzlo_out,nylo_out,nxlo_out);
			memory->destroy3d_offset(vdy_brick,nzlo_out,nylo_out,nxlo_out);
			memory->destroy3d_offset(vdz_brick,nzlo_out,nylo_out,nxlo_out);

			density_brick = vdx_brick = vdy_brick = vdz_brick = NULL;

			memory->destroy(density_fft);
			memory->destroy(greensfn);
			memory->destroy(work1);
			memory->destroy(work2);
			memory->destroy(vg);

			density_fft = NULL;
			greensfn = NULL;
			work1 = NULL;
			work2 = NULL;
			vg = NULL;

			memory->destroy(fkx);
			memory->destroy(fky);
			memory->destroy(fkz);

			fkx = NULL;
			fky = NULL;
			fkz = NULL;

			delete cu_density_brick;
			delete cu_density_brick_int;
			delete cu_vdx_brick;
			delete cu_vdy_brick;
			delete cu_vdz_brick;
			delete cu_density_fft;
			delete cu_energy;
			delete cu_virial;
			#ifdef FFT_CUFFT
			delete cu_greensfn;
			delete cu_gf_b;
			delete cu_vg;
			delete cu_work1;
			delete cu_work2;
			delete cu_work3;
			delete cu_fkx;
			delete cu_fky;
			delete cu_fkz;
			#endif

			delete cu_flag;
			delete cu_debugdata;
			delete cu_rho_coeff;


			cu_vdx_brick = cu_vdy_brick = cu_vdz_brick = NULL;
			cu_density_brick = NULL;
			cu_density_brick_int = NULL;
			cu_density_fft = NULL;
			cu_energy=NULL;
			cu_virial=NULL;
			#ifdef FFT_CUFFT
			cu_greensfn = NULL;
			cu_gf_b = NULL;
			cu_work1 = cu_work2 = cu_work3 = NULL;
			cu_vg = NULL;
			cu_fkx = cu_fky = cu_fkz = NULL;
			#endif

			cu_flag = NULL;
			cu_debugdata = NULL;
			cu_rho_coeff = NULL;
			cu_part2grid = NULL;

			memory->destroy(buf1);
			memory->destroy(buf2);

			delete [] gf_b;
			gf_b = NULL;
			memory->destroy2d_offset(rho1d,-order/2); rho1d = NULL;
			memory->destroy2d_offset(rho_coeff,(1-order)/2); rho_coeff = NULL;

			delete fft1c;
			fft1c = NULL;
			double end=CudaWrapper_CheckMemUseage()/1024/1024;
			delete fft2c;
			fft2c = NULL;
			delete remap;
			remap = NULL;
			buf1 = NULL;
			buf2 = NULL;
			}

			/* ----------------------------------------------------------------------
			set size of FFT grid (nx,ny,nz_pppm) and g_ewald
			-------------------------------------------------------------------------*/

			void PPPMCuda::set_grid()
			{
			// see JCP 109, pg 7698 for derivation of coefficients
			// higher order coefficients may be computed if needed

			double **acons;
			memory->create(acons,8,7,"pppm:acons");

			acons[1][0] = 2.0 / 3.0;
			acons[2][0] = 1.0 / 50.0;
			acons[2][1] = 5.0 / 294.0;
			acons[3][0] = 1.0 / 588.0;
			acons[3][1] = 7.0 / 1440.0;
			acons[3][2] = 21.0 / 3872.0;
			acons[4][0] = 1.0 / 4320.0;
			acons[4][1] = 3.0 / 1936.0;
			acons[4][2] = 7601.0 / 2271360.0;
			acons[4][3] = 143.0 / 28800.0;
			acons[5][0] = 1.0 / 23232.0;
			acons[5][1] = 7601.0 / 13628160.0;
			acons[5][2] = 143.0 / 69120.0;
			acons[5][3] = 517231.0 / 106536960.0;
			acons[5][4] = 106640677.0 / 11737571328.0;
			acons[6][0] = 691.0 / 68140800.0;
			acons[6][1] = 13.0 / 57600.0;
			acons[6][2] = 47021.0 / 35512320.0;
			acons[6][3] = 9694607.0 / 2095994880.0;
			acons[6][4] = 733191589.0 / 59609088000.0;
			acons[6][5] = 326190917.0 / 11700633600.0;
			acons[7][0] = 1.0 / 345600.0;
			acons[7][1] = 3617.0 / 35512320.0;
			acons[7][2] = 745739.0 / 838397952.0;
			acons[7][3] = 56399353.0 / 12773376000.0;
			acons[7][4] = 25091609.0 / 1560084480.0;
			acons[7][5] = 1755948832039.0 / 36229939200000.0;
			acons[7][6] = 4887769399.0 / 37838389248.0;

			double q2 = qsqsum / force->dielectric;
			bigint natoms = atom->natoms;

			// use xprd,yprd,zprd even if triclinic so grid size is the same
			// adjust z dimension for 2d slab PPPMCuda
			// 3d PPPMCuda just uses zprd since slab_volfactor = 1.0

			double xprd = domain->xprd;
			double yprd = domain->yprd;
			double zprd = domain->zprd;
			double zprd_slab = zprd*slab_volfactor;

			// make initial g_ewald estimate
			// based on desired error and real space cutoff
			// fluid-occupied volume used to estimate real-space error
			// zprd used rather than zprd_slab

			double h_x,h_y,h_z;

			if (!gewaldflag)
			g_ewald = sqrt(-log(precisionsqrt(natomscutoffxprdyprd*zprd) /
			(2.0*q2))) / cutoff;

			// set optimal nx_pppm,ny_pppm,nz_pppm based on order and precision
			// nz_pppm uses extended zprd_slab instead of zprd
			// h = 1/g_ewald is upper bound on h such that h*g_ewald <= 1
			// reduce it until precision target is met

			if (!gridflag) {
			double err;
			h_x = h_y = h_z = 1/g_ewald;

			nx_pppm = static_cast<int> (xprd/h_x + 1);
			ny_pppm = static_cast<int> (yprd/h_y + 1);
			nz_pppm = static_cast<int> (zprd_slab/h_z + 1);

			err = rms(h_x,xprd,natoms,q2,acons);
			while (err > precision) {
			err = rms(h_x,xprd,natoms,q2,acons);
			nx_pppm++;
			h_x = xprd/nx_pppm;
			}

			err = rms(h_y,yprd,natoms,q2,acons);
			while (err > precision) {
			err = rms(h_y,yprd,natoms,q2,acons);
			ny_pppm++;
			h_y = yprd/ny_pppm;
			}

			err = rms(h_z,zprd_slab,natoms,q2,acons);
			while (err > precision) {
			err = rms(h_z,zprd_slab,natoms,q2,acons);
			nz_pppm++;
			h_z = zprd_slab/nz_pppm;
			}
			}

			// boost grid size until it is factorable

			while (!factorable(nx_pppm)) nx_pppm++;
			while (!factorable(ny_pppm)) ny_pppm++;
			while (!factorable(nz_pppm)) nz_pppm++;

			// if allowed try to change grid size until it is a power of a single prime factor
			if(precisionmodify!='=')
			{
			if (me == 0) {
			if (screen) {
			fprintf(screen,"Initial grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			}
			if (logfile) {
			fprintf(logfile,"Initial grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			}
			}
			make_power_of_prime(&nx_pppm);
			make_power_of_prime(&ny_pppm);
			make_power_of_prime(&nz_pppm);
			if (me == 0) {
			if (screen) {
			fprintf(screen,"Modified grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			}
			if (logfile) {
			fprintf(logfile,"Modified grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			}
			}
			}

			// adjust g_ewald for new grid size

			h_x = xprd/nx_pppm;
			h_y = yprd/ny_pppm;
			h_z = zprd_slab/nz_pppm;

			if (!gewaldflag) {
			double gew1,gew2,dgew,f,fmid,hmin,rtb;
			int ncount;

			gew1 = 0.0;
			g_ewald = gew1;
			f = diffpr(h_x,h_y,h_z,q2,acons);

			hmin = MIN(h_x,MIN(h_y,h_z));
			gew2 = 10/hmin;
			g_ewald = gew2;
			fmid = diffpr(h_x,h_y,h_z,q2,acons);

			if (f*fmid >= 0.0) error->all(FLERR,"Cannot compute PPPMCuda G");
			rtb = f < 0.0 ? (dgew=gew2-gew1,gew1) : (dgew=gew1-gew2,gew2);
			ncount = 0;
			while (fabs(dgew) > SMALL && fmid != 0.0) {
			dgew *= 0.5;
			g_ewald = rtb + dgew;
			fmid = diffpr(h_x,h_y,h_z,q2,acons);
			if (fmid <= 0.0) rtb = g_ewald;
			ncount++;
			if (ncount > LARGE) error->all(FLERR,"Cannot compute PPPMCuda G");
			}
			}

			// final RMS precision

			double lprx = rms(h_x,xprd,natoms,q2,acons);
			double lpry = rms(h_y,yprd,natoms,q2,acons);
			double lprz = rms(h_z,zprd_slab,natoms,q2,acons);
			double lpr = sqrt(lprxlprx + lprylpry + lprz*lprz) / sqrt(3.0);
			double spr = 2.0q2 exp(-g_ewaldg_ewaldcutoff*cutoff) /
			sqrt(natomscutoffxprdyprdzprd_slab);

			// free local memory

			memory->destroy(acons);

			// print info

			if (me == 0) {
			if (screen) {
			fprintf(screen," G vector = %g\n",g_ewald);
			fprintf(screen," grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			fprintf(screen," stencil order = %d\n",order);
			fprintf(screen," RMS precision = %g\n",MAX(lpr,spr));
			}
			if (logfile) {
			fprintf(logfile," G vector = %g\n",g_ewald);
			fprintf(logfile," grid = %d %d %d\n",nx_pppm,ny_pppm,nz_pppm);
			fprintf(logfile," stencil order = %d\n",order);
			fprintf(logfile," RMS precision = %g\n",MAX(lpr,spr));
			}
			}
			}


			/* ----------------------------------------------------------------------
			check if all factors of n are prime
			return 1 if yes, 0 if no
			-------------------------------------------------------------------------*/

			void PPPMCuda::make_power_of_prime(int* n)
			{

			if((precisionmodify!='+')&&(precisionmodify!='-')&&(precisionmodify!='c'))
			{error->all(FLERR,"Unknown Option for PPPMCuda, assumeing '='");return;}
			int oldn=*n;
			int* primelist=new int[1000];
			int count=0;

			int prime=1;
			while(prime<2000) primelist[count++]=prime*=2;
			prime=1;
			while(prime<2000) primelist[count++]=prime*=3;
			prime=1;
			while(prime<2000) primelist[count++]=prime*=5;
			prime=1;
			while(prime<2000) primelist[count++]=prime*=7;

			for(int i=0;i<count-1;i++)
			for(int j=0;j<count-i-1;j++)
			{
			if(primelist[j]>primelist[j+1])
			{
			int a=primelist[j+1];
			primelist[j+1]=primelist[j];
			primelist[j]=a;
			}
			}

			int nextsmaller=0;
			while((primelist[nextsmaller+1]<*n)&&(nextsmaller+1<count)) nextsmaller++;

			int nextlarger=count-1;
			while((primelist[nextlarger-1]>*n)&&(nextlarger>0)) nextlarger--;

			if(precisionmodify=='-')
			*n=primelist[nextsmaller];
			if((precisionmodify=='+')&&
			(primelist[nextlarger]primelist[nextlarger]primelist[nextlarger]<2(n)(n)(n)))
			*n=primelist[nextlarger];
			if(precisionmodify=='c')
			{
			double factorsmaller=1.0(n)(n)(n)/(primelist[nextsmaller]primelist[nextsmaller]primelist[nextsmaller]);
			double factorlarger=1.0(primelist[nextlarger]primelist[nextlarger]primelist[nextlarger])/((n)(n)(n));
			if((factorlarger<factorsmaller)&&(factorlarger<2)) *n=primelist[nextlarger];
			else *n=primelist[nextsmaller];

			}
			delete [] primelist;
			primelist = NULL;
			if(n<0.75oldn)
			if (me == 0) {
			if (screen) {
			fprintf(screen,"\n\npppm/cuda WARNING: \t\tSignificantly lower gridsize than requested.\n\t\t\t\t\tYou should most likely use '=' or '+' as precision modify option\n");
			}
			if (logfile) {
			fprintf(logfile,"\n\npppm/cuda WARNING: \t\tSignificantly lower gridsize than requested.\n\t\t\t\t\tYou should most likely use '=' or '+' as precision modify option\n.");
			}
			}
			}



			/* ----------------------------------------------------------------------
			find center grid pt for each of my particles
			check that full stencil for the particle will fit in my 3d brick
			store central grid pt indices in part2grid array
			------------------------------------------------------------------------- */


			void PPPMCuda::particle_map()
			{
			MYDBG(printf("# CUDA PPPMCuda::particle_map() ... start\n");)
			int flag = 0;

			cu_flag->memset_device(0);
			flag=cuda_particle_map(&cuda->shared_data,cu_flag->dev_data());
			if(flag)
			{
			cu_debugdata->download();
			printf("Out of range atom: ");
			printf("ID: %i ",atom->tag[int(debugdata[0])]);
			printf("x: %e ",debugdata[7]);
			printf("y: %e ",debugdata[8]);
			printf("z: %e ",debugdata[9]);
			printf("nx: %e ",debugdata[4]);
			printf("ny: %e ",debugdata[5]);

			printf("\n");
			//printf("debugdata: cpu: %e %e %e %i\n",boxlo[0],boxlo[1],boxlo[2],atom->nlocal);
			cuda->cu_x->download();
			int nx,ny,nz;

			double **x = atom->x;
			int nlocal = atom->nlocal;
			for (int i = 0; i < nlocal; i++) {
			nx = static_cast<int> ((x[i][0]-boxlo[0])*delxinv+shift) - OFFSET;
			ny = static_cast<int> ((x[i][1]-boxlo[1])*delyinv+shift) - OFFSET;
			nz = static_cast<int> ((x[i][2]-boxlo[2])*delzinv+shift) - OFFSET;

			if(i==1203)printf("Outside Atom: %i %e %e %e (%i %i %i)\n",i,x[i][0],x[i][1],x[i][2],nx,ny,nz);
			if (nx+nlower < nxlo_out \|\| nx+nupper > nxhi_out \|\|
			ny+nlower < nylo_out \|\| ny+nupper > nyhi_out \|\|
			nz+nlower < nzlo_out \|\| nz+nupper > nzhi_out \|\| i==1203) {printf("Outside Atom: %i %e %e %e (%i %i %i)\n",i,x[i][0],x[i][1],x[i][2],nx,ny,nz); }
			}

			}

			int flag_all;
			MPI_Allreduce(&flag,&flag_all,1,MPI_INT,MPI_SUM,world);
			if (flag_all) error->all(FLERR,"Out of range atoms - cannot compute PPPMCuda!");
			}

			/* ----------------------------------------------------------------------
			create discretized "density" on section of global grid due to my particles
			density(x,y,z) = charge "density" at grid points of my 3d brick
			(nxlo:nxhi,nylo:nyhi,nzlo:nzhi) is extent of my brick (including ghosts)
			in global grid
			------------------------------------------------------------------------- */


			void PPPMCuda::make_rho()
			{
			cuda_make_rho(&cuda->shared_data,cu_flag->dev_data(),&density_intScale,nxhi_out,nxlo_out,nyhi_out,nylo_out,nzhi_out,nzlo_out,cu_density_brick->dev_data(),cu_density_brick_int->dev_data());
			}


			/* ----------------------------------------------------------------------
			FFT-based Poisson solver
			------------------------------------------------------------------------- */
			void PPPMCuda::poisson(int eflag, int vflag)
			{

			#ifndef FFT_CUFFT
			PPPM::poisson(eflag,vflag);
			return;
			#endif
			#ifdef FFT_CUFFT
			timespec starttime,starttime2;
			timespec endtime,endtime2;

			int nprocs=comm->nprocs;
			clock_gettime(CLOCK_REALTIME,&starttime);
			fft1c->compute(density_fft,work1,1);

			clock_gettime(CLOCK_REALTIME,&endtime);
			poissontime+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);



			if (eflag \|\| vflag) {
			poisson_energy(nxlo_fft,nxhi_fft,nylo_fft,nyhi_fft,nzlo_fft,nzhi_fft,vflag);
			ENERGY_FLOAT gpuvirial[6];
			energy+=sum_energy(cu_virial->dev_data(),cu_energy->dev_data(),nx_pppm,ny_pppm,nz_pppm,vflag,gpuvirial);
			if(vflag)
			{
			for(int j=0;j<6;j++) virial[j]+=gpuvirial[j];
			}
			}


			// scale by 1/total-grid-pts to get rho(k)
			// multiply by Green's function to get V(k)

			poisson_scale(nx_pppm,ny_pppm,nz_pppm);

			// compute gradients of V(r) in each of 3 dims by transformimg -ik*V(k)
			// FFT leaves data in 3d brick decomposition
			// copy it into inner portion of vdx,vdy,vdz arrays

			// x direction gradient


			poisson_xgrad(nx_pppm,ny_pppm,nz_pppm);


			clock_gettime(CLOCK_REALTIME,&starttime);
			fft2c->compute(work2,work2,-1);
			clock_gettime(CLOCK_REALTIME,&endtime);
			poissontime+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			poisson_vdx_brick(nxhi_out,nxlo_out,nyhi_out,nylo_out,nzhi_out,nzlo_out,nx_pppm,ny_pppm,nz_pppm);


			// y direction gradient

			poisson_ygrad(nx_pppm,ny_pppm,nz_pppm);

			clock_gettime(CLOCK_REALTIME,&starttime);
			fft2c->compute(work2,work2,-1);
			clock_gettime(CLOCK_REALTIME,&endtime);
			poissontime+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			poisson_vdy_brick(nxhi_out,nxlo_out,nyhi_out,nylo_out,nzhi_out,nzlo_out,nx_pppm,ny_pppm,nz_pppm);

			// z direction gradient

			poisson_zgrad(nx_pppm,ny_pppm,nz_pppm);

			clock_gettime(CLOCK_REALTIME,&starttime);
			fft2c->compute(work2,work2,-1);
			clock_gettime(CLOCK_REALTIME,&endtime);
			poissontime+=(endtime.tv_sec-starttime.tv_sec+1.0*(endtime.tv_nsec-starttime.tv_nsec)/1000000000);

			poisson_vdz_brick(nxhi_out,nxlo_out,nyhi_out,nylo_out,nzhi_out,nzlo_out,nx_pppm,ny_pppm,nz_pppm);
			#endif
			}

			/*----------------------------------------------------------------------
			interpolate from grid to get electric field & force on my particles
			-------------------------------------------------------------------------*/

			void PPPMCuda::fieldforce()
			{
			cuda_fieldforce(& cuda->shared_data,cu_flag);
			return;
			}




			/* ----------------------------------------------------------------------
			perform and time the 4 FFTs required for N timesteps
			------------------------------------------------------------------------- */

			void PPPMCuda::timing(int n, double &time3d, double &time1d)
			{

			double time1,time2;

			for (int i = 0; i < 2*nfft_both; i++) work1[i] = 0.0;

			MPI_Barrier(world);
			time1 = MPI_Wtime();

			for (int i = 0; i < n; i++) {
			fft1c->compute(work1,work1,1);
			fft2c->compute(work1,work1,-1);
			fft2c->compute(work1,work1,-1);
			fft2c->compute(work1,work1,-1);
			}

			MPI_Barrier(world);
			time2 = MPI_Wtime();
			time3d = time2 - time1;

			MPI_Barrier(world);
			/*time1 = MPI_Wtime();

			for (int i = 0; i < n; i++) {
			fft1c->timing1d(work1,nfft_both,1);
			fft2c->timing1d(work1,nfft_both,-1);
			fft2c->timing1d(work1,nfft_both,-1);
			fft2c->timing1d(work1,nfft_both,-1);
			}

			MPI_Barrier(world);
			time2 = MPI_Wtime();
			time1d = time2 - time1;*/

			}

			void PPPMCuda::slabcorr(int eflag)
			{
			// compute local contribution to global dipole moment
			if(slabbuf==NULL)
			{
			slabbuf=new ENERGY_FLOAT[(atom->nmax+31)/32];
			cu_slabbuf = new cCudaData<ENERGY_FLOAT,ENERGY_FLOAT, x> (slabbuf, (atom->nmax+31)/32);
			}
			if((atom->nlocal+31)/32*sizeof(ENERGY_FLOAT)>=cu_slabbuf->dev_size())
			{
			delete [] slabbuf;
			delete cu_slabbuf;
			slabbuf=new ENERGY_FLOAT[(atom->nmax+31)/32];
			cu_slabbuf = new cCudaData<ENERGY_FLOAT,ENERGY_FLOAT, x> (slabbuf, (atom->nmax+31)/32);
			}


			double dipole = cuda_slabcorr_energy(&cuda->shared_data,slabbuf,(ENERGY_FLOAT*) cu_slabbuf->dev_data());

			double dipole_all;
			MPI_Allreduce(&dipole,&dipole_all,1,MPI_DOUBLE,MPI_SUM,world);

			// compute corrections

			double e_slabcorr = 2.0MY_PIdipole_all*dipole_all/volume;

			if (eflag) energy += qqrd2escale e_slabcorr;

			double ffact = -4.0MY_PIdipole_all/volume;

			cuda_slabcorr_force(&cuda->shared_data,ffact);
			}

lammps/src/USER-CUDA/pppm_cuda.cppfb16c0eb301befficient_neuronet

pppm_cuda.cppView Options

lammps/src/USER-CUDA/pppm_cuda.cpp
fb16c0eb301befficient_neuronet

pppm_cuda.cpp
View Options