diff --git a/src/GPU/pair_coul_long_gpu.cpp b/src/GPU/pair_coul_long_gpu.cpp new file mode 100644 index 0000000000..468821dcec --- /dev/null +++ b/src/GPU/pair_coul_long_gpu.cpp @@ -0,0 +1,282 @@ +/* ---------------------------------------------------------------------- + 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 author: Mike Brown (SNL) +------------------------------------------------------------------------- */ + +#include "lmptype.h" +#include "math.h" +#include "stdio.h" +#include "stdlib.h" +#include "pair_coul_long_gpu.h" +#include "atom.h" +#include "atom_vec.h" +#include "comm.h" +#include "force.h" +#include "neighbor.h" +#include "neigh_list.h" +#include "integrate.h" +#include "memory.h" +#include "error.h" +#include "neigh_request.h" +#include "universe.h" +#include "update.h" +#include "domain.h" +#include "string.h" +#include "kspace.h" +#include "gpu_extra.h" + +#define MIN(a,b) ((a) < (b) ? (a) : (b)) +#define MAX(a,b) ((a) > (b) ? (a) : (b)) + +#define EWALD_F 1.12837917 +#define EWALD_P 0.3275911 +#define A1 0.254829592 +#define A2 -0.284496736 +#define A3 1.421413741 +#define A4 -1.453152027 +#define A5 1.061405429 + +// External functions from cuda library for atom decomposition + +int cl_gpu_init(const int nlocal, const int nall, const int max_nbors, + const int maxspecial, const double cell_size, int &gpu_mode, + FILE *screen, double host_cut_coulsq, double *host_special_coul, + const double qqrd2e, const double g_ewald); +void cl_gpu_clear(); +int ** cl_gpu_compute_n(const int ago, const int inum, + const int nall, double **host_x, int *host_type, + double *sublo, double *subhi, int *tag, + int **nspecial, int **special, const bool eflag, + const bool vflag, const bool eatom, const bool vatom, + int &host_start, int **ilist, int **jnum, + const double cpu_time, bool &success, double *host_q, + double *boxlo, double *prd); +void cl_gpu_compute(const int ago, const int inum, const int nall, + double **host_x, int *host_type, int *ilist, int *numj, + int **firstneigh, const bool eflag, const bool vflag, + const bool eatom, const bool vatom, int &host_start, + const double cpu_time, bool &success, double *host_q, + const int nlocal, double *boxlo, double *prd); +double cl_gpu_bytes(); + +using namespace LAMMPS_NS; + +/* ---------------------------------------------------------------------- */ + +PairCoulLongGPU::PairCoulLongGPU(LAMMPS *lmp) : + PairCoulLong(lmp), gpu_mode(GPU_PAIR) +{ + respa_enable = 0; + cpu_time = 0.0; +} + +/* ---------------------------------------------------------------------- + free all arrays +------------------------------------------------------------------------- */ + +PairCoulLongGPU::~PairCoulLongGPU() +{ + cl_gpu_clear(); +} + +/* ---------------------------------------------------------------------- */ + +void PairCoulLongGPU::compute(int eflag, int vflag) +{ + if (eflag || vflag) ev_setup(eflag,vflag); + else evflag = vflag_fdotr = 0; + + int nall = atom->nlocal + atom->nghost; + int inum, host_start; + + bool success = true; + int *ilist, *numneigh, **firstneigh; + if (gpu_mode == GPU_NEIGH) { + inum = atom->nlocal; + firstneigh = cl_gpu_compute_n(neighbor->ago, inum, nall, atom->x, + atom->type, domain->sublo, domain->subhi, + atom->tag, atom->nspecial, atom->special, + eflag, vflag, eflag_atom, vflag_atom, + host_start, &ilist, &numneigh, cpu_time, + success, atom->q, domain->boxlo, + domain->prd); + } else { + inum = list->inum; + ilist = list->ilist; + numneigh = list->numneigh; + firstneigh = list->firstneigh; + cl_gpu_compute(neighbor->ago, inum, nall, atom->x, atom->type, + ilist, numneigh, firstneigh, eflag, vflag, eflag_atom, + vflag_atom, host_start, cpu_time, success, atom->q, + atom->nlocal, domain->boxlo, domain->prd); + } + if (!success) + error->one("Out of memory on GPGPU"); + + if (host_startq_flag) + error->all("Pair style coul/long/gpu requires atom attribute q"); + if (force->newton_pair) + error->all("Cannot use newton pair with coul/long/gpu pair style"); + + // Repeat cutsq calculation because done after call to init_style + double cell_size = sqrt(cut_coul) + neighbor->skin; + + cut_coulsq = cut_coul * cut_coul; + + // insure use of KSpace long-range solver, set g_ewald + + if (force->kspace == NULL) + error->all("Pair style is incompatible with KSpace style"); + g_ewald = force->kspace->g_ewald; + + // setup force tables + + if (ncoultablebits) init_tables(); + + int maxspecial=0; + if (atom->molecular) + maxspecial=atom->maxspecial; + int success = cl_gpu_init(atom->nlocal, atom->nlocal+atom->nghost, 300, + maxspecial, cell_size, gpu_mode, screen, cut_coulsq, + force->special_coul, force->qqrd2e, g_ewald); + + GPU_EXTRA::check_flag(success,error,world); + + if (gpu_mode != GPU_NEIGH) { + int irequest = neighbor->request(this); + neighbor->requests[irequest]->half = 0; + neighbor->requests[irequest]->full = 1; + } +} + +/* ---------------------------------------------------------------------- */ + +double PairCoulLongGPU::memory_usage() +{ + double bytes = Pair::memory_usage(); + return bytes + cl_gpu_bytes(); +} + +/* ---------------------------------------------------------------------- */ + +void PairCoulLongGPU::cpu_compute(int start, int inum, int eflag, + int vflag, int *ilist, int *numneigh, + int **firstneigh) +{ + int i,j,ii,jj,jnum,itype,jtype,itable; + double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,evdwl,ecoul,fpair; + double fraction,table; + double r,r2inv,r6inv,forcecoul,forcelj,factor_coul,factor_lj; + double grij,expm2,prefactor,t,erfc; + int *jlist; + double rsq; + + evdwl = ecoul = 0.0; + + double **x = atom->x; + double **f = atom->f; + double *q = atom->q; + int *type = atom->type; + double *special_coul = force->special_coul; + double qqrd2e = force->qqrd2e; + + // loop over neighbors of my atoms + + for (ii = start; ii < inum; ii++) { + i = ilist[ii]; + qtmp = q[i]; + xtmp = x[i][0]; + ytmp = x[i][1]; + ztmp = x[i][2]; + itype = type[i]; + jlist = firstneigh[i]; + jnum = numneigh[i]; + + for (jj = 0; jj < jnum; jj++) { + j = jlist[jj]; + factor_coul = special_coul[sbmask(j)]; + j &= NEIGHMASK; + + delx = xtmp - x[j][0]; + dely = ytmp - x[j][1]; + delz = ztmp - x[j][2]; + rsq = delx*delx + dely*dely + delz*delz; + jtype = type[j]; + + r2inv = 1.0/rsq; + + if (rsq < cut_coulsq) { + if (!ncoultablebits || rsq <= tabinnersq) { + r = sqrt(rsq); + grij = g_ewald * r; + expm2 = exp(-grij*grij); + t = 1.0 / (1.0 + EWALD_P*grij); + erfc = t * (A1+t*(A2+t*(A3+t*(A4+t*A5)))) * expm2; + prefactor = qqrd2e * qtmp*q[j]/r; + forcecoul = prefactor * (erfc + EWALD_F*grij*expm2); + if (factor_coul < 1.0) forcecoul -= (1.0-factor_coul)*prefactor; + } else { + union_int_float_t rsq_lookup; + rsq_lookup.f = rsq; + itable = rsq_lookup.i & ncoulmask; + itable >>= ncoulshiftbits; + fraction = (rsq_lookup.f - rtable[itable]) * drtable[itable]; + table = ftable[itable] + fraction*dftable[itable]; + forcecoul = qtmp*q[j] * table; + if (factor_coul < 1.0) { + table = ctable[itable] + fraction*dctable[itable]; + prefactor = qtmp*q[j] * table; + forcecoul -= (1.0-factor_coul)*prefactor; + } + } + + fpair = forcecoul * r2inv; + + f[i][0] += delx*fpair; + f[i][1] += dely*fpair; + f[i][2] += delz*fpair; + + if (eflag) { + if (rsq < cut_coulsq) { + if (!ncoultablebits || rsq <= tabinnersq) + ecoul = prefactor*erfc; + else { + table = etable[itable] + fraction*detable[itable]; + ecoul = qtmp*q[j] * table; + } + if (factor_coul < 1.0) ecoul -= (1.0-factor_coul)*prefactor; + } else ecoul = 0.0; + } + + if (evflag) ev_tally_full(i,0.0,ecoul,fpair,delx,dely,delz); + } + } + } +} diff --git a/src/GPU/pair_coul_long_gpu.h b/src/GPU/pair_coul_long_gpu.h new file mode 100644 index 0000000000..a84b281447 --- /dev/null +++ b/src/GPU/pair_coul_long_gpu.h @@ -0,0 +1,47 @@ +/* ---------------------------------------------------------------------- + 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. +------------------------------------------------------------------------- */ + +#ifdef PAIR_CLASS + +PairStyle(coul/long/gpu,PairCoulLongGPU) + +#else + +#ifndef LMP_PAIR_COUL_LONG_GPU_H +#define LMP_PAIR_COUL_LONG_GPU_H + +#include "pair_coul_long.h" + +namespace LAMMPS_NS { + +class PairCoulLongGPU : public PairCoulLong { + public: + PairCoulLongGPU(LAMMPS *lmp); + ~PairCoulLongGPU(); + void cpu_compute(int, int, int, int, int *, int *, int **); + void compute(int, int); + void init_style(); + double memory_usage(); + + enum { GPU_PAIR, GPU_NEIGH }; + + private: + int gpu_mode; + double cpu_time; + int *gpulist; +}; + +} +#endif +#endif +