diff --git a/lib/gpu/lal_lj_gromacs.cpp b/lib/gpu/lal_lj_gromacs.cpp new file mode 100644 index 0000000000..75b2345378 --- /dev/null +++ b/lib/gpu/lal_lj_gromacs.cpp @@ -0,0 +1,164 @@ +/*************************************************************************** + lj_gromacs.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the lj/gromacs pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#if defined(USE_OPENCL) +#include "lj_gromacs_cl.h" +#elif defined(USE_CUDART) +const char *lj_gromacs=0; +#else +#include "lj_gromacs_cubin.h" +#endif + +#include "lal_lj_gromacs.h" +#include +using namespace LAMMPS_AL; +#define LJGROMACST LJGROMACS + +extern Device device; + +template +LJGROMACST::LJGROMACS() : BaseAtomic(), + _allocated(false) { +} + +template +LJGROMACST::~LJGROMACS() { + clear(); +} + +template +int LJGROMACST::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int LJGROMACST::init(const int ntypes, double **host_cutsq, + double **host_lj1, double **host_lj2, double **host_lj3, + double **host_lj4, 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_ljsw1, double **host_ljsw2, double **host_ljsw3, + double **host_ljsw4, double **host_ljsw5, + double **cut_inner, double **cut_inner_sq) { + int success; + success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split, + _screen,lj_gromacs,"k_lj_gromacs"); + if (success!=0) + return success; + + // If atom type constants fit in shared memory use fast kernel + int lj_types=ntypes; + 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; + shared_types=true; + } + _lj_types=lj_types; + + // Allocate a host write buffer for data initialization + UCL_H_Vec host_write(lj_types*lj_types*32,*(this->ucl_device), + UCL_WRITE_ONLY); + + for (int i=0; iucl_device),UCL_READ_ONLY); + this->atom->type_pack4(ntypes,lj_types,lj1,host_write,host_lj1,host_lj2, + host_cutsq,cut_inner_sq); + + lj3.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); + this->atom->type_pack4(ntypes,lj_types,lj3,host_write,host_lj3,host_lj4, + cut_inner,host_ljsw5); + + ljsw.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); + this->atom->type_pack4(ntypes,lj_types,ljsw,host_write,host_ljsw1,host_ljsw2, + host_ljsw3,host_ljsw4); + + UCL_H_Vec dview; + sp_lj.alloc(4,*(this->ucl_device),UCL_READ_ONLY); + dview.view(host_special_lj,4,*(this->ucl_device)); + ucl_copy(sp_lj,dview,false); + + _allocated=true; + this->_max_bytes=lj1.row_bytes()+lj3.row_bytes() + +ljsw.row_bytes()+sp_lj.row_bytes(); + return 0; +} + +template +void LJGROMACST::clear() { + if (!_allocated) + return; + _allocated=false; + + lj1.clear(); + lj3.clear(); + ljsw.clear(); + sp_lj.clear(); + this->clear_atomic(); +} + +template +double LJGROMACST::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(LJGROMACS); +} + +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void LJGROMACST::loop(const bool _eflag, const bool _vflag) { + // Compute the block size and grid size to keep all cores busy + const int BX=this->block_size(); + int eflag, vflag; + if (_eflag) + eflag=1; + else + eflag=0; + + if (_vflag) + vflag=1; + else + vflag=0; + + int GX=static_cast(ceil(static_cast(this->ans->inum())/ + (BX/this->_threads_per_atom))); + + int ainum=this->ans->inum(); + int nbor_pitch=this->nbor->nbor_pitch(); + this->time_pair.start(); + if (shared_types) { + this->k_pair_fast.set_size(GX,BX); + this->k_pair_fast.run(&this->atom->x, &lj1, &lj3, &ljsw, + &sp_lj, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), + &this->ans->force, &this->ans->engv, + &eflag, &vflag, &ainum, &nbor_pitch, + &this->_threads_per_atom); + } else { + this->k_pair.set_size(GX,BX); + this->k_pair.run(&this->atom->x, &lj1, &lj3, &ljsw, &_lj_types, + &sp_lj, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), + &this->ans->force, &this->ans->engv, + &eflag, &vflag, &ainum, &nbor_pitch, + &this->_threads_per_atom); + } + this->time_pair.stop(); +} + +template class LJGROMACS; diff --git a/lib/gpu/lal_lj_gromacs.cu b/lib/gpu/lal_lj_gromacs.cu new file mode 100644 index 0000000000..b1e1d135c4 --- /dev/null +++ b/lib/gpu/lal_lj_gromacs.cu @@ -0,0 +1,233 @@ +// ************************************************************************** +// lj_gromacs.cu +// ------------------- +// Trung Dac Nguyen (ORNL) +// +// Device code for acceleration of the gromacs/coul/long pair style +// +// __________________________________________________________________________ +// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) +// __________________________________________________________________________ +// +// begin : +// email : nguyentd@ornl.gov +// ***************************************************************************/ + +#ifdef NV_KERNEL + +#include "lal_aux_fun1.h" +#ifndef _DOUBLE_DOUBLE +texture pos_tex; +#else +texture pos_tex; +#endif + +#else +#define pos_tex x_ +#endif + +__kernel void k_lj_gromacs(const __global numtyp4 *restrict x_, + const __global numtyp4 *restrict lj1, + const __global numtyp4 *restrict lj3, + const __global numtyp4 *restrict ljsw, + 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 int t_per_atom) { + int tid, ii, offset; + atom_info(t_per_atom,ii,tid,offset); + + __local numtyp sp_lj[4]; + 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]; + + acctyp energy=(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 lj1[mtype].w) { + numtyp r = ucl_sqrt(rsq); + t = r - lj3[mtype].z; + numtyp fswitch = r*t*t*(ljsw[mtype].x + ljsw[mtype].y*t); + force_lj += fswitch; + } + + force = factor_lj*force_lj * r2inv; + + f.x+=delx*force; + f.y+=dely*force; + f.z+=delz*force; + + if (eflag>0) { + numtyp e=r6inv*(lj3[mtype].x*r6inv-lj3[mtype].y); + e += lj3[mtype].w; + if (rsq > lj1[mtype].w) { + numtyp eswitch = t*t*t*(ljsw[mtype].z + ljsw[mtype].w*t); + e += eswitch; + } + energy+=factor_lj*e; + } + 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(f,energy,virial,ii,inum,tid,t_per_atom,offset,eflag, + vflag,ans,engv); + } // if ii +} + +__kernel void k_lj_gromacs_fast(const __global numtyp4 *restrict x_, + const __global numtyp4 *restrict lj1_in, + const __global numtyp4 *restrict lj3_in, + const __global numtyp4 *restrict ljsw_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 int t_per_atom) { + int tid, ii, offset; + atom_info(t_per_atom,ii,tid,offset); + + __local numtyp4 lj1[MAX_SHARED_TYPES*MAX_SHARED_TYPES]; + __local numtyp4 lj3[MAX_SHARED_TYPES*MAX_SHARED_TYPES]; + __local numtyp4 ljsw[MAX_SHARED_TYPES*MAX_SHARED_TYPES]; + __local numtyp sp_lj[4]; + if (tid<4) + sp_lj[tid]=sp_lj_in[tid]; + if (tid lj1[mtype].w) { + numtyp r = ucl_sqrt(rsq); + t = r - lj3[mtype].z; + numtyp fswitch = r*t*t*(ljsw[mtype].x + ljsw[mtype].y*t); + force_lj += fswitch; + } + + force = factor_lj*force_lj * r2inv; + + f.x+=delx*force; + f.y+=dely*force; + f.z+=delz*force; + + if (eflag>0) { + numtyp e=r6inv*(lj3[mtype].x*r6inv-lj3[mtype].y); + e += lj3[mtype].w; + if (rsq > lj1[mtype].w) { + numtyp eswitch = t*t*t*(ljsw[mtype].z + ljsw[mtype].w*t); + e += eswitch; + } + energy+=factor_lj*e; + } + 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(f,energy,virial,ii,inum,tid,t_per_atom,offset,eflag, + vflag,ans,engv); + } // if ii +} + diff --git a/lib/gpu/lal_lj_gromacs.h b/lib/gpu/lal_lj_gromacs.h new file mode 100644 index 0000000000..dc949be4a9 --- /dev/null +++ b/lib/gpu/lal_lj_gromacs.h @@ -0,0 +1,84 @@ +/*************************************************************************** + lj_gromacs.h + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the lj/gromacs pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifndef LAL_LJ_GROMACS_H +#define LAL_LJ_GROMACS_H + +#include "lal_base_atomic.h" + +namespace LAMMPS_AL { + +template +class LJGROMACS : public BaseAtomic { + public: + LJGROMACS(); + ~LJGROMACS(); + + /// 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_lj1, double **host_lj2, double **host_lj3, + double **host_lj4, 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_ljsw1, double **host_ljsw2, double **host_ljsw3, + double **host_ljsw4, double **host_ljsw5, + double **cut_inner, double **cut_inner_sq); + + /// Clear all host and device data + /** \note This is called at the beginning of the init() routine **/ + void clear(); + + /// Returns memory usage on device per atom + int bytes_per_atom(const int max_nbors) const; + + /// Total host memory used by library for pair style + double host_memory_usage() const; + + // --------------------------- TYPE DATA -------------------------- + + /// lj1.x = lj1, lj1.y = lj2, lj1.z = cutsq, lj1.w = cut_inner_sq + UCL_D_Vec lj1; + /// lj3.x = lj3, lj3.y = lj4, lj3.z = cut_inner, lj3.w = ljsw5 + UCL_D_Vec lj3; + /// ljsw.x = ljsw1, ljsw.y = ljsw2, ljsw.z = ljsw3, ljsw.w = ljsw4 + UCL_D_Vec ljsw; + /// Special LJ values [0-3] and Special Coul values [4-7] + UCL_D_Vec sp_lj; + + /// If atom type constants fit in shared memory, use fast kernels + bool shared_types; + + /// Number of atom types + int _lj_types; + + private: + bool _allocated; + void loop(const bool _eflag, const bool _vflag); +}; + +} + +#endif diff --git a/lib/gpu/lal_lj_gromacs_ext.cpp b/lib/gpu/lal_lj_gromacs_ext.cpp new file mode 100644 index 0000000000..c2db048fe6 --- /dev/null +++ b/lib/gpu/lal_lj_gromacs_ext.cpp @@ -0,0 +1,126 @@ +/*************************************************************************** + lj_gromacs_ext.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Functions for LAMMPS access to lj/gromacs acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_lj_gromacs.h" + +using namespace std; +using namespace LAMMPS_AL; + +static LJGROMACS LJGRMMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +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_inner_sq) { + LJGRMMF.clear(); + gpu_mode=LJGRMMF.device->gpu_mode(); + double gpu_split=LJGRMMF.device->particle_split(); + int first_gpu=LJGRMMF.device->first_device(); + int last_gpu=LJGRMMF.device->last_device(); + int world_me=LJGRMMF.device->world_me(); + int gpu_rank=LJGRMMF.device->gpu_rank(); + int procs_per_gpu=LJGRMMF.device->procs_per_gpu(); + + LJGRMMF.device->init_message(screen,"lj/gromacs",first_gpu,last_gpu); + + bool message=false; + if (LJGRMMF.device->replica_me()==0 && screen) + message=true; + + if (message) { + fprintf(screen,"Initializing GPU and compiling on process 0..."); + fflush(screen); + } + + int init_ok=0; + if (world_me==0) + LJGRMMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4, + special_lj, inum, nall, 300, maxspecial, cell_size, + gpu_split, screen, host_ljsw1, host_ljsw2, host_ljsw3, + host_ljsw4, host_ljsw5, cut_inner, cut_inner_sq); + + LJGRMMF.device->world_barrier(); + if (message) + fprintf(screen,"Done.\n"); + + for (int i=0; igpu_barrier(); + if (message) + fprintf(screen,"Done.\n"); + } + if (message) + fprintf(screen,"\n"); + + if (init_ok==0) + LJGRMMF.estimate_gpu_overhead(); + return init_ok; +} + +void ljgrm_gpu_clear() { + LJGRMMF.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) { + return LJGRMMF.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); +} + +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) { + LJGRMMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj, + firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success); +} + + +double ljgrm_gpu_bytes() { + return LJGRMMF.host_memory_usage(); +} + +