Added missing source files for born/coul/wolf/cs in the gpu library

/***************************************************************************

                            born_coul_wolf_cs.cpp

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

                        Trung Dac Nguyen (Northwestern)

  Class for acceleration of the born/coul/wolf/cs pair style.

 __________________________________________________________________________

    This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)

 __________________________________________________________________________

    begin                :

    email                : ndactrung@gmail.com

 ***************************************************************************/

#ifdef USE_OPENCL

#include "born_coul_wolf_cs_cl.h"

#elif defined(USE_CUDART)

const char *born_coul_wolf_cs=0;

#else

#include "born_coul_wolf_cs_cubin.h"

#endif

#include "lal_born_coul_wolf_cs.h"

#include <cassert>

using namespace LAMMPS_AL;

#define BornCoulWolfCST BornCoulWolfCS<numtyp, acctyp>

extern Device<PRECISION,ACC_PRECISION> device;

template <class numtyp, class acctyp>

int BornCoulWolfCST::init(const int ntypes, double **host_cutsq, double **host_rhoinv,

                        double **host_born1, double **host_born2, double **host_born3,

                        double **host_a, double **host_c, double **host_d,

                        double **host_sigma, double **host_offset,

                        double *host_special_lj, const int nlocal,

                        const int nall, const int max_nbors,

                        const int maxspecial, const double cell_size,

                        const double gpu_split, FILE *_screen,

                        double **host_cut_ljsq, const double host_cut_coulsq,

                        double *host_special_coul, const double qqrd2e,

                        const double alf, const double e_shift, const double f_shift) {

  int success;

  success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,

                            _screen,born_coul_wolf_cs,"k_born_coul_wolf_cs");

  if (success!=0)

    return success;

  // If atom type constants fit in shared memory use fast kernel

  int lj_types=ntypes;

  this->shared_types=false;

  int max_shared_types=this->device->max_shared_types();

  if (lj_types<=max_shared_types && this->_block_size>=max_shared_types) {

    lj_types=max_shared_types;

    this->shared_types=true;

  }

  this->_lj_types=lj_types;

  // Allocate a host write buffer for data initialization

  UCL_H_Vec<numtyp> host_write(lj_types*lj_types*32,*(this->ucl_device),

                               UCL_WRITE_ONLY);

  for (int i=0; i<lj_types*lj_types; i++)

    host_write[i]=0.0;

  this->coeff1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);

  this->atom->type_pack4(ntypes,lj_types,this->coeff1,host_write,host_rhoinv,

                         host_born1,host_born2,host_born3);

  this->coeff2.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);

  this->atom->type_pack4(ntypes,lj_types,this->coeff2,host_write,host_a,host_c,

                                     host_d,host_offset);

  this->cutsq_sigma.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);

  this->atom->type_pack4(ntypes,lj_types,this->cutsq_sigma,host_write,host_cutsq,

                         host_cut_ljsq,host_sigma);

  this->sp_lj.alloc(8,*(this->ucl_device),UCL_READ_ONLY);

  for (int i=0; i<4; i++) {

    host_write[i]=host_special_lj[i];

    host_write[i+4]=host_special_coul[i];

  }

  ucl_copy(this->sp_lj,host_write,8,false);

  this->_cut_coulsq=host_cut_coulsq;

  this->_qqrd2e=qqrd2e;

  this->_alf=alf;

  this->_e_shift=e_shift;

  this->_f_shift=f_shift;

  this->_allocated=true;

  this->_max_bytes=this->coeff1.row_bytes()+this->coeff2.row_bytes()

      +this->cutsq_sigma.row_bytes()+this->sp_lj.row_bytes();

  return 0;

}

template class BornCoulWolfCS<PRECISION,ACC_PRECISION>;

// **************************************************************************

//                            born_coul_wolf_cs.cu

//                             -------------------

//                         Trung Dac Nguyen (Northwestern)

//

//  Device code for acceleration of the born/coul/wolf/cs pair style

//

// __________________________________________________________________________

//    This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)

// __________________________________________________________________________

//

//    begin                :

//    email                : ndactrung@gmail.com

// ***************************************************************************/

#ifdef NV_KERNEL

#include "lal_aux_fun1.h"

#ifndef _DOUBLE_DOUBLE

texture<float4> pos_tex;

texture<float> q_tex;

#else

texture<int4,1> pos_tex;

texture<int2> q_tex;

#endif

#else

#define pos_tex x_

#define q_tex q_

#endif

#define EPSILON (acctyp)(1.0e-20)

#define MY_PIS (acctyp)1.77245385090551602729

__kernel void k_born_coul_wolf_cs(const __global numtyp4 *restrict x_,

                          const __global numtyp4 *restrict coeff1,

                          const __global numtyp4 *restrict coeff2,

                          const int lj_types,

                          const __global numtyp *restrict sp_lj_in,

                          const __global int *dev_nbor,

                          const __global int *dev_packed,

                          __global acctyp4 *restrict ans,

                          __global acctyp *restrict engv,

                          const int eflag, const int vflag, const int inum,

                          const int nbor_pitch,

                          const __global numtyp *restrict q_,

                          const __global numtyp4 *restrict cutsq_sigma,

                          const numtyp cut_coulsq, const numtyp qqrd2e,

                          const numtyp alf, const numtyp e_shift,

                          const numtyp f_shift, const int t_per_atom) {

  int tid, ii, offset;

  atom_info(t_per_atom,ii,tid,offset);

  __local numtyp sp_lj[8];

  sp_lj[0]=sp_lj_in[0];

  sp_lj[1]=sp_lj_in[1];

  sp_lj[2]=sp_lj_in[2];

  sp_lj[3]=sp_lj_in[3];

  sp_lj[4]=sp_lj_in[4];

  sp_lj[5]=sp_lj_in[5];

  sp_lj[6]=sp_lj_in[6];

  sp_lj[7]=sp_lj_in[7];

  acctyp energy=(acctyp)0;

  acctyp e_coul=(acctyp)0;

  acctyp4 f;

  f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;

  acctyp virial[6];

  for (int i=0; i<6; i++)

    virial[i]=(acctyp)0;

  if (ii<inum) {

    int nbor, nbor_end;

    int i, numj;

    __local int n_stride;

    nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,

              n_stride,nbor_end,nbor);

    numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];

    numtyp qtmp; fetch(qtmp,i,q_tex);

    int itype=ix.w;

    if (eflag>0) {

      acctyp e_self = -((acctyp)0.5*e_shift + alf/MY_PIS) *

        qtmp*qtmp*qqrd2e/(acctyp)t_per_atom;

      e_coul += (acctyp)2.0*e_self;

    }

    for ( ; nbor<nbor_end; nbor+=n_stride) {

      int j=dev_packed[nbor];

      numtyp factor_lj, factor_coul;

      factor_lj = sp_lj[sbmask(j)];

      factor_coul = sp_lj[sbmask(j)+4];

      j &= NEIGHMASK;

      numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];

      int jtype=jx.w;

      // Compute r12

      numtyp delx = ix.x-jx.x;

      numtyp dely = ix.y-jx.y;

      numtyp delz = ix.z-jx.z;

      numtyp rsq = delx*delx+dely*dely+delz*delz;

      int mtype=itype*lj_types+jtype;

      if (rsq<cutsq_sigma[mtype].x) { // cutsq

        rsq += EPSILON; // Add Epsilon for case: r = 0; Interaction must be removed by special bond;

        acctyp r2inv = ucl_recip(rsq);

        numtyp forcecoul,forceborn,force,prefactor,rexp;

        acctyp v_sh,r6inv;

        if (rsq < cutsq_sigma[mtype].y) { // cut_ljsq

          numtyp r = ucl_sqrt(rsq);

          rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);

          r6inv = r2inv*r2inv*r2inv;

          forceborn = (coeff1[mtype].y*r*rexp - coeff1[mtype].z*r6inv

            + coeff1[mtype].w*r2inv*r6inv)*factor_lj;

        } else forceborn = (numtyp)0.0;

        if (rsq < cut_coulsq) {

          numtyp r = ucl_rsqrt(r2inv);

          acctyp arij = alf * r;

          acctyp erfcd = ucl_exp(-arij*arij);

          fetch(prefactor,j,q_tex);

          prefactor *= qqrd2e * qtmp/r;

          const acctyp erfcc = erfc(arij);

          v_sh = (erfcc - e_shift*r)*prefactor;

          acctyp dvdrr = (erfcc/rsq + (numtyp)2.0*alf/MY_PIS * erfcd/r) + f_shift;

          forcecoul = prefactor * dvdrr*rsq;

          if (factor_coul < (numtyp)1.0) forcecoul -= ((numtyp)1.0-factor_coul)*prefactor;

        } else forcecoul = (numtyp)0.0;

        force = (forceborn + forcecoul) * r2inv;

        f.x+=delx*force;

        f.y+=dely*force;

        f.z+=delz*force;

        if (eflag>0) {

          if (rsq < cut_coulsq) {

            acctyp e=v_sh;

            if (factor_coul < (numtyp)1.0) e -= ((numtyp)1.0-factor_coul)*prefactor;

            e_coul += e;

          }

          if (rsq < cutsq_sigma[mtype].y) {

            numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv

              + coeff2[mtype].z*r2inv*r6inv;

            energy+=factor_lj*(e-coeff2[mtype].w);

          }

        }

        if (vflag>0) {

          virial[0] += delx*delx*force;

          virial[1] += dely*dely*force;

          virial[2] += delz*delz*force;

          virial[3] += delx*dely*force;

          virial[4] += delx*delz*force;

          virial[5] += dely*delz*force;

        }

      }

    } // for nbor

    store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,

                    vflag,ans,engv);

  } // if ii

}

__kernel void k_born_coul_wolf_cs_fast(const __global numtyp4 *restrict x_,

                               const __global numtyp4 *restrict coeff1_in,

                               const __global numtyp4 *restrict coeff2_in,

                               const __global numtyp *restrict sp_lj_in,

                               const __global int *dev_nbor,

                               const __global int *dev_packed,

                               __global acctyp4 *restrict ans,

                               __global acctyp *restrict engv,

                               const int eflag, const int vflag, const int inum,

                               const int nbor_pitch,

                               const __global numtyp *restrict q_,

                               const __global numtyp4 *restrict cutsq_sigma,

                               const numtyp cut_coulsq, const numtyp qqrd2e,

                               const numtyp alf, const numtyp e_shift,

                               const numtyp f_shift, const int t_per_atom) {

  int tid, ii, offset;

  atom_info(t_per_atom,ii,tid,offset);

  __local numtyp4 coeff1[MAX_SHARED_TYPES*MAX_SHARED_TYPES];

  __local numtyp4 coeff2[MAX_SHARED_TYPES*MAX_SHARED_TYPES];

  __local numtyp sp_lj[8];

  if (tid<8)

    sp_lj[tid]=sp_lj_in[tid];

  if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES) {

    coeff1[tid]=coeff1_in[tid];

    if (eflag>0)

      coeff2[tid]=coeff2_in[tid];

  }

  acctyp energy=(acctyp)0;

  acctyp e_coul=(acctyp)0;

  acctyp4 f;

  f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;

  acctyp virial[6];

  for (int i=0; i<6; i++)

    virial[i]=(acctyp)0;

  __syncthreads();

  if (ii<inum) {

    int nbor, nbor_end;

    int i, numj;

    __local int n_stride;

    nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,

              n_stride,nbor_end,nbor);

    numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];

    numtyp qtmp; fetch(qtmp,i,q_tex);

    int iw=ix.w;

    int itype=fast_mul((int)MAX_SHARED_TYPES,iw);

    if (eflag>0) {

      acctyp e_self = -((acctyp)0.5*e_shift + alf/MY_PIS) *

        qtmp*qtmp*qqrd2e/(acctyp)t_per_atom;

      e_coul += (acctyp)2.0*e_self;

    }

    for ( ; nbor<nbor_end; nbor+=n_stride) {

      int j=dev_packed[nbor];

      numtyp factor_lj, factor_coul;

      factor_lj = sp_lj[sbmask(j)];

      factor_coul = sp_lj[sbmask(j)+4];

      j &= NEIGHMASK;

      numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];

      int mtype=itype+jx.w;

      // Compute r12

      numtyp delx = ix.x-jx.x;

      numtyp dely = ix.y-jx.y;

      numtyp delz = ix.z-jx.z;

      acctyp rsq = delx*delx+dely*dely+delz*delz;

      if (rsq<cutsq_sigma[mtype].x) {

        rsq += EPSILON; // Add Epsilon for case: r = 0; Interaction must be removed by special bond;

        acctyp r2inv = ucl_recip(rsq);

        numtyp forcecoul,forceborn,force,prefactor,rexp;

        acctyp v_sh,r6inv;

        if (rsq < cutsq_sigma[mtype].y) {

          r6inv = r2inv*r2inv*r2inv;

          numtyp r = ucl_sqrt(rsq);

          rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);

          forceborn = (coeff1[mtype].y*r*rexp - coeff1[mtype].z*r6inv

            + coeff1[mtype].w*r2inv*r6inv)*factor_lj;

        } else forceborn = (numtyp)0.0;

       if (rsq < cut_coulsq) {

          numtyp r = ucl_sqrt(rsq);

          acctyp arij = alf * r;

          acctyp erfcd = ucl_exp(-arij*arij);

          fetch(prefactor,j,q_tex);

          prefactor *= qqrd2e * qtmp/r;

          const acctyp erfcc = erfc(arij);

          v_sh = (erfcc - e_shift*r)*prefactor;

          acctyp dvdrr = (erfcc/rsq + (numtyp)2.0*alf/MY_PIS * erfcd/r) + f_shift;

          forcecoul = prefactor * dvdrr*rsq;

          if (factor_coul < (numtyp)1.0) forcecoul -= ((numtyp)1.0-factor_coul)*prefactor;

        } else forcecoul = (numtyp)0.0;

        force = (forceborn + forcecoul) * r2inv;

        f.x+=delx*force;

        f.y+=dely*force;

        f.z+=delz*force;

        if (eflag>0) {

          if (rsq < cut_coulsq) {

            acctyp e=v_sh;

            if (factor_coul < (numtyp)1.0) e -= ((numtyp)1.0-factor_coul)*prefactor;

            e_coul += e;

          }

          if (rsq < cutsq_sigma[mtype].y) {

            numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv

              + coeff2[mtype].z*r2inv*r6inv;

            energy+=factor_lj*(e-coeff2[mtype].w);

          }

        }

        if (vflag>0) {

          virial[0] += delx*delx*force;

          virial[1] += dely*dely*force;

          virial[2] += delz*delz*force;

          virial[3] += delx*dely*force;

          virial[4] += delx*delz*force;

          virial[5] += dely*delz*force;

        }

      }

    } // for nbor

    store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,

                    vflag,ans,engv);

  } // if ii

}

/***************************************************************************

                              born_coul_wolf_cs.h

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

                           Trung Dac Nguyen (Northwestern)

  Class for acceleration of the born/coul/wolf/cs pair style.

 __________________________________________________________________________

    This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)

 __________________________________________________________________________

    begin                :

    email                : ndactrung@gmail.com

 ***************************************************************************/

#ifndef LAL_BORN_COUL_WOLF_CS_H

#define LAL_BORN_COUL_WOLF_CS_H

#include "lal_born_coul_wolf.h"

namespace LAMMPS_AL {

template <class numtyp, class acctyp>

class BornCoulWolfCS : public BornCoulWolf<numtyp, acctyp> {

 public:

  BornCoulWolfCS() {}

  ~BornCoulWolfCS() {}

  /// Clear any previous data and set up for a new LAMMPS run

  /** \param max_nbors initial number of rows in the neighbor matrix

    * \param cell_size cutoff + skin

    * \param gpu_split fraction of particles handled by device

    *

    * Returns:

    * -  0 if successfull

    * - -1 if fix gpu not found

    * - -3 if there is an out of memory error

    * - -4 if the GPU library was not compiled for GPU

    * - -5 Double precision is not supported on card **/

  int init(const int ntypes, double **host_cutsq, double **host_rhoinv,

           double **host_born1, double **host_born2, double **host_born3,

           double **host_a, double **host_c, double **host_d,

           double **host_sigma, double **host_offset, double *host_special_lj,

           const int nlocal, const int nall, const int max_nbors,

           const int maxspecial, const double cell_size,

           const double gpu_split, FILE *screen, double **host_cut_ljsq,

           const double host_cut_coulsq, double *host_special_coul,

           const double qqrd2e, const double alf, const double e_shift,

           const double f_shift);

};

}

#endif

/***************************************************************************

                           born_coul_wolf_cs_ext.cpp

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

                           Trung Dac Nguyen (Northwestern)

  Functions for LAMMPS access to born/coul/wolf/cs acceleration routines.

 __________________________________________________________________________

    This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)

 __________________________________________________________________________

    begin                :

    email                : ndactrung@gmail.com

 ***************************************************************************/

#include <iostream>

#include <cassert>

#include <cmath>

#include "lal_born_coul_wolf_cs.h"

using namespace std;

using namespace LAMMPS_AL;

static BornCoulWolfCS<PRECISION,ACC_PRECISION> BornCWCST;

// ---------------------------------------------------------------------------

// Allocate memory on host and device and copy constants to device

// ---------------------------------------------------------------------------

int borncwcs_gpu_init(const int ntypes, double **cutsq, double **host_rhoinv,

                    double **host_born1, double **host_born2, double **host_born3,

                    double **host_a, double **host_c, double **host_d,

                    double **sigma, double **offset, 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_cut_ljsq, double host_cut_coulsq,

                    double *host_special_coul, const double qqrd2e,

                    const double alf, const double e_shift, const double f_shift) {

  BornCWCST.clear();

  gpu_mode=BornCWCST.device->gpu_mode();

  double gpu_split=BornCWCST.device->particle_split();

  int first_gpu=BornCWCST.device->first_device();

  int last_gpu=BornCWCST.device->last_device();

  int world_me=BornCWCST.device->world_me();

  int gpu_rank=BornCWCST.device->gpu_rank();

  int procs_per_gpu=BornCWCST.device->procs_per_gpu();

  BornCWCST.device->init_message(screen,"born/coul/wolf/cs",first_gpu,last_gpu);

  bool message=false;

  if (BornCWCST.device->replica_me()==0 && screen)

    message=true;

  if (message) {

    fprintf(screen,"Initializing Device and compiling on process 0...");

    fflush(screen);

  }

  int init_ok=0;

  if (world_me==0)

    init_ok=BornCWCST.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,

                          host_born3, host_a, host_c, host_d, sigma,

                          offset, special_lj, inum, nall, 300,

                          maxspecial, cell_size, gpu_split, screen, host_cut_ljsq,

                          host_cut_coulsq, host_special_coul, qqrd2e,

                          alf, e_shift, f_shift);

  BornCWCST.device->world_barrier();

  if (message)

    fprintf(screen,"Done.\n");

  for (int i=0; i<procs_per_gpu; i++) {

    if (message) {

      if (last_gpu-first_gpu==0)

        fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);

      else

        fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,

                last_gpu,i);

      fflush(screen);

    }

    if (gpu_rank==i && world_me!=0)

      init_ok=BornCWCST.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,

                            host_born3, host_a, host_c, host_d, sigma,

                            offset, special_lj, inum, nall, 300,

                            maxspecial, cell_size, gpu_split, screen, host_cut_ljsq,

                            host_cut_coulsq, host_special_coul, qqrd2e,

                            alf, e_shift, f_shift);

    BornCWCST.device->gpu_barrier();

    if (message)

      fprintf(screen,"Done.\n");

  }

  if (message)

    fprintf(screen,"\n");

  if (init_ok==0)

    BornCWCST.estimate_gpu_overhead();

  return init_ok;

}

void borncwcs_gpu_clear() {

  BornCWCST.clear();

}

int** borncwcs_gpu_compute_n(const int ago, const int inum_full,

                           const int nall, double **host_x, int *host_type,

                           double *sublo, double *subhi, tagint *tag, int **nspecial,

                           tagint **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) {

  return BornCWCST.compute(ago, inum_full, nall, host_x, host_type, sublo,

                          subhi, tag, nspecial, special, eflag, vflag, eatom,

                          vatom, host_start, ilist, jnum, cpu_time, success,

                          host_q, boxlo, prd);

}

void borncwcs_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 *host_q,

                        const int nlocal, double *boxlo, double *prd) {

  BornCWCST.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,

                   firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,

                   host_q,nlocal,boxlo,prd);

}

double borncwcs_gpu_bytes() {

  return BornCWCST.host_memory_usage();

}