diff --git a/src/GPU/Install.sh b/src/GPU/Install.sh index 89a342496b..54b0d2a6a1 100644 --- a/src/GPU/Install.sh +++ b/src/GPU/Install.sh @@ -83,6 +83,8 @@ action pair_lj_cut_gpu.cpp action pair_lj_cut_gpu.h action pair_lj_expand_gpu.cpp action pair_lj_expand_gpu.h +action pair_lj_gromacs_gpu.cpp +action pair_lj_gromacs_gpu.h action pair_lj_sdk_coul_long_gpu.cpp pair_lj_sdk_coul_long.cpp action pair_lj_sdk_coul_long_gpu.h pair_lj_sdk_coul_long.cpp action pair_lj_sdk_gpu.cpp pair_lj_sdk.cpp diff --git a/src/GPU/pair_lj_gromacs_gpu.cpp b/src/GPU/pair_lj_gromacs_gpu.cpp new file mode 100644 index 0000000000..97d2f53884 --- /dev/null +++ b/src/GPU/pair_lj_gromacs_gpu.cpp @@ -0,0 +1,247 @@ +/* ---------------------------------------------------------------------- + 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: Trung Dac Nguyen (ORNL) +------------------------------------------------------------------------- */ + +#include "lmptype.h" +#include "math.h" +#include "stdio.h" +#include "stdlib.h" +#include "pair_lj_gromacs_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" + +// External functions from cuda library for atom decomposition + +int ljgrm_gpu_init(const int ntypes, double **cutsq, double **host_lj1, + double **host_lj2, double **host_lj3, double **host_lj4, + double *special_lj, const int inum, + const int nall, const int max_nbors, const int maxspecial, + const double cell_size, int &gpu_mode, FILE *screen, + double **host_ljsw1, double **host_ljsw2, double **host_ljsw3, + double **host_ljsw4, double **host_ljsw5, + double **cut_inner, double **cut_innersq); +void ljgrm_gpu_clear(); +int ** ljgrm_gpu_compute_n(const int ago, const int inum_full, + 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); +void ljgrm_gpu_compute(const int ago, const int inum_full, 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 ljgrm_gpu_bytes(); + + +using namespace LAMMPS_NS; + +/* ---------------------------------------------------------------------- */ + +PairLJGromacsGPU::PairLJGromacsGPU(LAMMPS *lmp) : + PairLJGromacs(lmp), gpu_mode(GPU_FORCE) +{ + respa_enable = 0; + cpu_time = 0.0; + GPU_EXTRA::gpu_ready(lmp->modify, lmp->error); +} + +/* ---------------------------------------------------------------------- + free all arrays +------------------------------------------------------------------------- */ + +PairLJGromacsGPU::~PairLJGromacsGPU() +{ + ljgrm_gpu_clear(); +} + +/* ---------------------------------------------------------------------- */ + +void PairLJGromacsGPU::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_FORCE) { + inum = atom->nlocal; + firstneigh = ljgrm_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); + } else { + inum = list->inum; + ilist = list->ilist; + numneigh = list->numneigh; + firstneigh = list->firstneigh; + ljgrm_gpu_compute(neighbor->ago, inum, nall, atom->x, atom->type, + ilist, numneigh, firstneigh, eflag, vflag, eflag_atom, + vflag_atom, host_start, cpu_time, success); + } + if (!success) + error->one(FLERR,"Insufficient memory on accelerator"); + + if (host_startnewton_pair) + error->all(FLERR,"Cannot use newton pair with lj/gromacs/gpu pair style"); + + // Repeat cutsq calculation because done after call to init_style + double maxcut = -1.0; + double mcut; + for (int i = 1; i <= atom->ntypes; i++) { + for (int j = i; j <= atom->ntypes; j++) { + if (setflag[i][j] != 0 || (setflag[i][i] != 0 && setflag[j][j] != 0)) { + mcut = init_one(i,j); + mcut *= mcut; + if (mcut > maxcut) + maxcut = mcut; + cutsq[i][j] = cutsq[j][i] = mcut; + } else + cutsq[i][j] = cutsq[j][i] = 0.0; + } + } + double cell_size = sqrt(maxcut) + neighbor->skin; + + int maxspecial=0; + if (atom->molecular) + maxspecial=atom->maxspecial; + + int success = ljgrm_gpu_init(atom->ntypes+1, cutsq, lj1, lj2, lj3, lj4, + force->special_lj, atom->nlocal, + atom->nlocal+atom->nghost, 300, maxspecial, + cell_size, gpu_mode, screen, ljsw1, ljsw2, + ljsw3, ljsw4, ljsw5, cut_inner, cut_inner_sq); + GPU_EXTRA::check_flag(success,error,world); + + if (gpu_mode == GPU_FORCE) { + int irequest = neighbor->request(this); + neighbor->requests[irequest]->half = 0; + neighbor->requests[irequest]->full = 1; + } +} + +/* ---------------------------------------------------------------------- */ + +double PairLJGromacsGPU::memory_usage() +{ + double bytes = Pair::memory_usage(); + return bytes + ljgrm_gpu_bytes(); +} + +/* ---------------------------------------------------------------------- */ + +void PairLJGromacsGPU::cpu_compute(int start, int inum, int eflag, + int vflag, int *ilist, + int *numneigh, int **firstneigh) +{ + int i,j,ii,jj,jnum,itype,jtype; + double xtmp,ytmp,ztmp,delx,dely,delz,evdwl,fpair; + double rsq,r2inv,r6inv,forcelj,factor_lj; + double r,t,fswitch,eswitch; + int *jlist; + + double **x = atom->x; + double **f = atom->f; + int *type = atom->type; + double *special_lj = force->special_lj; + + // loop over neighbors of my atoms + + for (ii = start; ii < inum; ii++) { + i = ilist[ii]; + 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_lj = special_lj[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]; + + if (rsq < cutsq[itype][jtype]) { + r2inv = 1.0/rsq; + r6inv = r2inv*r2inv*r2inv; + forcelj = r6inv * (lj1[itype][jtype]*r6inv - lj2[itype][jtype]); + if (rsq > cut_inner_sq[itype][jtype]) { + r = sqrt(rsq); + t = r - cut_inner[itype][jtype]; + fswitch = r*t*t*(ljsw1[itype][jtype] + ljsw2[itype][jtype]*t); + forcelj += fswitch; + } + fpair = factor_lj*forcelj * r2inv; + + f[i][0] += delx*fpair; + f[i][1] += dely*fpair; + f[i][2] += delz*fpair; + + if (eflag) { + evdwl = r6inv * (lj3[itype][jtype]*r6inv - lj4[itype][jtype]); + evdwl += ljsw5[itype][jtype]; + if (rsq > cut_inner_sq[itype][jtype]) { + eswitch = t*t*t*(ljsw3[itype][jtype] + ljsw4[itype][jtype]*t); + evdwl += eswitch; + } + evdwl *= factor_lj; + } + + if (evflag) ev_tally_full(i,evdwl,0.0,fpair,delx,dely,delz); + } + } + } +} diff --git a/src/GPU/pair_lj_gromacs_gpu.h b/src/GPU/pair_lj_gromacs_gpu.h new file mode 100644 index 0000000000..c45e91beb3 --- /dev/null +++ b/src/GPU/pair_lj_gromacs_gpu.h @@ -0,0 +1,64 @@ +/* ---------------------------------------------------------------------- + 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(lj/gromacs/gpu,PairLJGromacsGPU) + +#else + +#ifndef LMP_PAIR_LJ_GROMACS_GPU_H +#define LMP_PAIR_LJ_GROMACS_GPU_H + +#include "pair_lj_gromacs.h" + +namespace LAMMPS_NS { + +class PairLJGromacsGPU : public PairLJGromacs { + public: + PairLJGromacsGPU(LAMMPS *lmp); + ~PairLJGromacsGPU(); + void cpu_compute(int, int, int, int, int *, int *, int **); + void compute(int, int); + void init_style(); + double memory_usage(); + + enum { GPU_FORCE, GPU_NEIGH, GPU_HYB_NEIGH }; + + private: + int gpu_mode; + double cpu_time; + int *gpulist; +}; + +} +#endif +#endif + +/* ERROR/WARNING messages: + +E: Insufficient memory on accelerator + +There is insufficient memory on one of the devices specified for the gpu +package + +E: Cannot use newton pair with lj/gromacs/gpu pair style + +Self-explanatory. + +E: Pair style is incompatible with KSpace style + +If a pair style with a long-range Coulombic component is selected, +then a kspace style must also be used. + +*/ diff --git a/src/GPU/pair_sw_gpu.cpp b/src/GPU/pair_sw_gpu.cpp index 672b3e517e..db7a1f6617 100644 --- a/src/GPU/pair_sw_gpu.cpp +++ b/src/GPU/pair_sw_gpu.cpp @@ -35,11 +35,15 @@ // External functions from cuda library for atom decomposition -int sw_gpu_init(const int nlocal, const int nall, const int max_nbors, +int sw_gpu_init(const int ntypes, const int inum, const int nall, const int max_nbors, const double cell_size, int &gpu_mode, FILE *screen, - const double, const double, const double, const double, - const double, const double, const double, const double, - const double, const double, const double); + int* host_map, const int nelements, int*** host_elem2param, const int nparams, + const double* sw_epsilon, const double* sw_sigma, + const double* sw_lambda, const double* sw_gamma, + const double* sw_costheta, const double* sw_biga, + const double* sw_bigb, const double* sw_powerp, + const double* sw_powerq, const double* sw_cut, + const double* sw_cutsq); void sw_gpu_clear(); int ** sw_gpu_compute_n(const int ago, const int inum, const int nall, double **host_x, int *host_type, @@ -135,21 +139,64 @@ void PairSWGPU::allocate() void PairSWGPU::init_style() { - double cell_size = sqrt(params[0].cutsq) + neighbor->skin; + double cell_size = cutmax + neighbor->skin; if (atom->tag_enable == 0) error->all(FLERR,"Pair style sw/gpu requires atom IDs"); if (force->newton_pair != 0) error->all(FLERR,"Pair style sw/gpu requires newton pair off"); - if (nparams > 1) - error->all(FLERR,"Pair style sw/gpu is currently limited to one element."); - int success = sw_gpu_init(atom->nlocal, atom->nlocal+atom->nghost, 300, - cell_size, gpu_mode, screen,params[0].epsilon, - params[0].sigma, params[0].lambda, params[0].gamma, - params[0].costheta, params[0].biga, params[0].bigb, - params[0].powerp, params[0].powerq, params[0].cut, - params[0].cutsq); + double *epsilon, *sigma, *lambda, *gamma; + double *biga, *bigb, *powerp, *powerq; + double *_cut, *_cutsq, *costheta; + epsilon = sigma = lambda = gamma = NULL; + biga = bigb = powerp = powerq = NULL; + _cut = _cutsq = costheta = NULL; + + memory->create(epsilon,nparams,"pair:epsilon"); + memory->create(sigma,nparams,"pair:sigma"); + memory->create(lambda,nparams,"pair:lambda"); + memory->create(gamma,nparams,"pair:gamma"); + memory->create(biga,nparams,"pair:biga"); + memory->create(bigb,nparams,"pair:bigb"); + memory->create(powerp,nparams,"pair:powerp"); + memory->create(powerq,nparams,"pair:powerq"); + memory->create(_cut,nparams,"pair:_cut"); + memory->create(_cutsq,nparams,"pair:_cutsq"); + memory->create(costheta,nparams,"pair:costheta"); + + for (int i = 0; i < nparams; i++) { + epsilon[i] = params[i].epsilon; + sigma[i] = params[i].sigma; + lambda[i] = params[i].lambda; + gamma[i] = params[i].gamma; + biga[i] = params[i].biga; + bigb[i] = params[i].bigb; + powerp[i] = params[i].powerp; + powerq[i] = params[i].powerq; + _cut[i] = params[i].cut; + _cutsq[i] = params[i].cutsq; + costheta[i] = params[i].costheta; + } + + int success = sw_gpu_init(atom->ntypes+1, atom->nlocal, atom->nlocal+atom->nghost, 300, + cell_size, gpu_mode, screen, map, nelements, + elem2param, nparams, epsilon, + sigma, lambda, gamma, costheta, biga, bigb, + powerp, powerq, _cut, _cutsq); + + memory->destroy(epsilon); + memory->destroy(sigma); + memory->destroy(lambda); + memory->destroy(gamma); + memory->destroy(biga); + memory->destroy(bigb); + memory->destroy(powerp); + memory->destroy(powerq); + memory->destroy(_cut); + memory->destroy(_cutsq); + memory->destroy(costheta); + GPU_EXTRA::check_flag(success,error,world); if (gpu_mode == GPU_FORCE) {