From 22751a2aaef2b249f62c794c63c38c8e548ae04c Mon Sep 17 00:00:00 2001 From: sjplimp Date: Fri, 23 Aug 2013 14:46:18 +0000 Subject: [PATCH] git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@10668 f3b2605a-c512-4ea7-a41b-209d697bcdaa --- lib/gpu/Makefile.lammps.mac_ocl | 5 + lib/gpu/Makefile.lammps.opencl | 5 + lib/gpu/lal_base_three.cpp | 358 ++++++++++++++++++++++++++++++++ lib/gpu/lal_base_three.h | 221 ++++++++++++++++++++ lib/gpu/lal_beck.cpp | 152 ++++++++++++++ lib/gpu/lal_beck.h | 80 +++++++ lib/gpu/lal_beck_ext.cpp | 120 +++++++++++ lib/gpu/lal_lj_coul_msm.cpp | 200 ++++++++++++++++++ lib/gpu/lal_lj_coul_msm.h | 88 ++++++++ lib/gpu/lal_lj_coul_msm_ext.cpp | 131 ++++++++++++ lib/gpu/lal_mie.cpp | 152 ++++++++++++++ lib/gpu/lal_mie.h | 80 +++++++ lib/gpu/lal_mie_ext.cpp | 124 +++++++++++ lib/gpu/lal_soft.cpp | 145 +++++++++++++ lib/gpu/lal_soft.h | 77 +++++++ lib/gpu/lal_soft_ext.cpp | 120 +++++++++++ lib/gpu/lal_sw.cpp | 167 +++++++++++++++ lib/gpu/lal_sw.h | 73 +++++++ lib/gpu/lal_sw_ext.cpp | 128 ++++++++++++ 19 files changed, 2426 insertions(+) create mode 100644 lib/gpu/Makefile.lammps.mac_ocl create mode 100644 lib/gpu/Makefile.lammps.opencl create mode 100644 lib/gpu/lal_base_three.cpp create mode 100644 lib/gpu/lal_base_three.h create mode 100644 lib/gpu/lal_beck.cpp create mode 100644 lib/gpu/lal_beck.h create mode 100644 lib/gpu/lal_beck_ext.cpp create mode 100644 lib/gpu/lal_lj_coul_msm.cpp create mode 100644 lib/gpu/lal_lj_coul_msm.h create mode 100644 lib/gpu/lal_lj_coul_msm_ext.cpp create mode 100644 lib/gpu/lal_mie.cpp create mode 100644 lib/gpu/lal_mie.h create mode 100644 lib/gpu/lal_mie_ext.cpp create mode 100644 lib/gpu/lal_soft.cpp create mode 100644 lib/gpu/lal_soft.h create mode 100644 lib/gpu/lal_soft_ext.cpp create mode 100644 lib/gpu/lal_sw.cpp create mode 100644 lib/gpu/lal_sw.h create mode 100644 lib/gpu/lal_sw_ext.cpp diff --git a/lib/gpu/Makefile.lammps.mac_ocl b/lib/gpu/Makefile.lammps.mac_ocl new file mode 100644 index 0000000000..f6c8a36430 --- /dev/null +++ b/lib/gpu/Makefile.lammps.mac_ocl @@ -0,0 +1,5 @@ +# Settings that the LAMMPS build will import when this package library is used + +gpu_SYSINC = +gpu_SYSLIB = -framework OpenCL +gpu_SYSPATH = diff --git a/lib/gpu/Makefile.lammps.opencl b/lib/gpu/Makefile.lammps.opencl new file mode 100644 index 0000000000..413ae79210 --- /dev/null +++ b/lib/gpu/Makefile.lammps.opencl @@ -0,0 +1,5 @@ +# Settings that the LAMMPS build will import when this package library is used + +gpu_SYSINC = +gpu_SYSLIB = -lOpenCL +gpu_SYSPATH = diff --git a/lib/gpu/lal_base_three.cpp b/lib/gpu/lal_base_three.cpp new file mode 100644 index 0000000000..fe64cf44ed --- /dev/null +++ b/lib/gpu/lal_base_three.cpp @@ -0,0 +1,358 @@ +/*************************************************************************** + base_three.cpp + ------------------- + W. Michael Brown (ORNL) + + Base class for pair styles with per-particle data for position and type + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : Tue April 2, 2013 + email : brownw@ornl.gov + ***************************************************************************/ + +#include "lal_base_three.h" +using namespace LAMMPS_AL; +#define BaseThreeT BaseThree + +extern Device global_device; + +template +BaseThreeT::BaseThree() : _compiled(false), _max_bytes(0) { + device=&global_device; + ans=new Answer(); + nbor=new Neighbor(); + #ifdef THREE_CONCURRENT + ans2=new Answer(); + #endif +} + +template +BaseThreeT::~BaseThree() { + delete ans; + delete nbor; + #ifdef THREE_CONCURRENT + delete ans2; + #endif +} + +template +int BaseThreeT::bytes_per_atom_atomic(const int max_nbors) const { + int b=device->atom.bytes_per_atom()+ans->bytes_per_atom()+ + nbor->bytes_per_atom(max_nbors); + #ifdef THREE_CONCURRENT + b+=ans2->bytes_per_atom(); + #endif + return b; +} + +template +int BaseThreeT::init_three(const int nlocal, const int nall, + const int max_nbors, const int maxspecial, + const double cell_size, const double gpu_split, + FILE *_screen, const void *pair_program, + const char *k_two, const char *k_three_center, + const char *k_three_end) { + screen=_screen; + + int gpu_nbor=0; + if (device->gpu_mode()==Device::GPU_NEIGH) + gpu_nbor=1; + else if (device->gpu_mode()==Device::GPU_HYB_NEIGH) + gpu_nbor=2; + + int _gpu_host=0; + int host_nlocal=hd_balancer.first_host_count(nlocal,gpu_split,gpu_nbor); + if (host_nlocal>0) + _gpu_host=1; + + _threads_per_atom=device->threads_per_atom(); + if (_threads_per_atom>1 && gpu_nbor==0) { + nbor->packing(true); + _nbor_data=&(nbor->dev_packed); + } else + _nbor_data=&(nbor->dev_nbor); + if (_threads_per_atom*_threads_per_atom>device->warp_size()) + return -10; + + int success=device->init(*ans,false,false,nlocal,host_nlocal,nall,nbor, + maxspecial,_gpu_host,max_nbors,cell_size,false, + _threads_per_atom); + if (success!=0) + return success; + + ucl_device=device->gpu; + atom=&device->atom; + + #ifdef THREE_CONCURRENT + _end_command_queue=ucl_device->num_queues(); + ucl_device->push_command_queue(); + if (!ans2->init(ans->max_inum(),false,false,*(device->gpu))) + return -3; + ans2->cq(_end_command_queue); + #endif + + _block_pair=device->pair_block_size(); + _block_size=device->block_ellipse(); + compile_kernels(*ucl_device,pair_program,k_two,k_three_center,k_three_end); + + // Initialize host-device load balancer + hd_balancer.init(device,gpu_nbor,gpu_split); + + // Initialize timers for the selected GPU + time_pair.init(*ucl_device); + time_pair.zero(); + + pos_tex.bind_float(atom->x,4); + + _max_an_bytes=ans->gpu_bytes()+nbor->gpu_bytes(); + #ifdef THREE_CONCURRENT + _max_an_bytes+=ans2->gpu_bytes(); + #endif + + return 0; +} + +template +void BaseThreeT::estimate_gpu_overhead() { + device->estimate_gpu_overhead(1,_gpu_overhead,_driver_overhead); +} + +template +void BaseThreeT::clear_atomic() { + // Output any timing information + acc_timers(); + double avg_split=hd_balancer.all_avg_split(); + _gpu_overhead*=hd_balancer.timestep(); + _driver_overhead*=hd_balancer.timestep(); + device->output_times(time_pair,*ans,*nbor,avg_split,_max_bytes+_max_an_bytes, + _gpu_overhead,_driver_overhead,_threads_per_atom,screen); + + if (_compiled) { + k_three_center.clear(); + k_three_end.clear(); + k_three_end_vatom.clear(); + k_pair.clear(); + delete pair_program; + _compiled=false; + } + + time_pair.clear(); + hd_balancer.clear(); + + nbor->clear(); + ans->clear(); + #ifdef THREE_CONCURRENT + ans2->clear(); + assert(ucl_device->num_queues()==_end_command_queue+1); + ucl_device->pop_command_queue(); + #endif + device->clear(); +} + +// --------------------------------------------------------------------------- +// Copy neighbor list from host +// --------------------------------------------------------------------------- +template +int * BaseThreeT::reset_nbors(const int nall, const int inum, const int nlist, + int *ilist, int *numj, int **firstneigh, + bool &success) { + success=true; + + int mn=nbor->max_nbor_loop(nlist,numj,ilist); + resize_atom(inum,nall,success); + resize_local(nall,mn,success); + if (!success) + return NULL; + + nbor->get_host3(nall,nlist,ilist,numj,firstneigh,block_size()); + + double bytes=ans->gpu_bytes()+nbor->gpu_bytes(); + #ifdef THREE_CONCURRENT + bytes+=ans2->gpu_bytes(); + #endif + if (bytes>_max_an_bytes) + _max_an_bytes=bytes; + + return ilist; +} + +// --------------------------------------------------------------------------- +// Build neighbor list on device +// --------------------------------------------------------------------------- +template +inline void BaseThreeT::build_nbor_list(const int inum, const int host_inum, + const int nall, double **host_x, + int *host_type, double *sublo, + double *subhi, int *tag, + int **nspecial, int **special, + bool &success) { + success=true; + resize_atom(inum,nall,success); + resize_local(nall,host_inum,nbor->max_nbors(),success); + if (!success) + return; + atom->cast_copy_x(host_x,host_type); + + int mn; + nbor->build_nbor_list(host_x, nall, host_inum, nall, *atom, sublo, subhi, tag, + nspecial, special, success, mn); + + double bytes=ans->gpu_bytes()+nbor->gpu_bytes(); + #ifdef THREE_CONCURRENT + bytes+=ans2->gpu_bytes(); + #endif + if (bytes>_max_an_bytes) + _max_an_bytes=bytes; +} + +// --------------------------------------------------------------------------- +// Copy nbor list from host if necessary and then calculate forces, virials,.. +// --------------------------------------------------------------------------- +template +void BaseThreeT::compute(const int f_ago, const int nlocal, const int nall, + const int nlist, 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) { + acc_timers(); + if (nlist==0) { + host_start=0; + // Make sure textures are correct if realloc by a different hybrid style + resize_atom(0,nall,success); + zero_timers(); + return; + } + + int ago=hd_balancer.ago_first(f_ago); + int inum=hd_balancer.balance(ago,nlocal,cpu_time); + ans->inum(inum); + #ifdef THREE_CONCURRENT + ans2->inum(inum); + #endif + host_start=inum; + + if (ago==0) { + reset_nbors(nall, inum, nlist, ilist, numj, firstneigh, success); + if (!success) + return; + } + + atom->cast_x_data(host_x,host_type); + hd_balancer.start_timer(); + atom->add_x_data(host_x,host_type); + + int evatom=0; + if (eatom || vatom) + evatom=1; + #ifdef THREE_CONCURRENT + ucl_device->sync(); + #endif + loop(eflag,vflag,evatom); + ans->copy_answers(eflag,vflag,eatom,vatom,ilist); + device->add_ans_object(ans); + #ifdef THREE_CONCURRENT + ans2->copy_answers(eflag,vflag,eatom,vatom,ilist); + device->add_ans_object(ans2); + #endif + hd_balancer.stop_timer(); +} + +// --------------------------------------------------------------------------- +// Reneighbor on GPU if necessary and then compute forces, virials, energies +// --------------------------------------------------------------------------- +template +int ** BaseThreeT::compute(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) { + acc_timers(); + if (inum_full==0) { + host_start=0; + // Make sure textures are correct if realloc by a different hybrid style + resize_atom(0,nall,success); + zero_timers(); + return NULL; + } + + hd_balancer.balance(cpu_time); + int inum=hd_balancer.get_gpu_count(ago,inum_full); + ans->inum(inum); + #ifdef THREE_CONCURRENT + ans2->inum(inum); + #endif + host_start=inum; + + // Build neighbor list on GPU if necessary + if (ago==0) { + build_nbor_list(inum, inum_full-inum, nall, host_x, host_type, + sublo, subhi, tag, nspecial, special, success); + if (!success) + return NULL; + hd_balancer.start_timer(); + } else { + atom->cast_x_data(host_x,host_type); + hd_balancer.start_timer(); + atom->add_x_data(host_x,host_type); + } + *ilist=nbor->host_ilist.begin(); + *jnum=nbor->host_acc.begin(); + + int evatom=0; + if (eatom || vatom) + evatom=1; + #ifdef THREE_CONCURRENT + ucl_device->sync(); + #endif + loop(eflag,vflag,evatom); + ans->copy_answers(eflag,vflag,eatom,vatom); + device->add_ans_object(ans); + #ifdef THREE_CONCURRENT + ans2->copy_answers(eflag,vflag,eatom,vatom); + device->add_ans_object(ans2); + #endif + hd_balancer.stop_timer(); + + return nbor->host_jlist.begin()-host_start; +} + +template +double BaseThreeT::host_memory_usage_atomic() const { + return device->atom.host_memory_usage()+nbor->host_memory_usage()+ + 4*sizeof(numtyp)+sizeof(BaseThree); +} + +template +void BaseThreeT::compile_kernels(UCL_Device &dev, const void *pair_str, + const char *ktwo, const char *kthree_center, + const char *kthree_end) { + if (_compiled) + return; + + std::string vatom_name=std::string(kthree_end)+"_vatom"; + + pair_program=new UCL_Program(dev); + pair_program->load_string(pair_str,device->compile_string().c_str()); + k_three_center.set_function(*pair_program,kthree_center); + k_three_end.set_function(*pair_program,kthree_end); + k_three_end_vatom.set_function(*pair_program,vatom_name.c_str()); + k_pair.set_function(*pair_program,ktwo); + pos_tex.get_texture(*pair_program,"pos_tex"); + + #ifdef THREE_CONCURRENT + k_three_end.cq(ucl_device->cq(_end_command_queue)); + k_three_end_vatom.cq(ucl_device->cq(_end_command_queue)); + #endif + + _compiled=true; +} + +template class BaseThree; + diff --git a/lib/gpu/lal_base_three.h b/lib/gpu/lal_base_three.h new file mode 100644 index 0000000000..7e167f4655 --- /dev/null +++ b/lib/gpu/lal_base_three.h @@ -0,0 +1,221 @@ +/*************************************************************************** + base_three.h + ------------------- + W. Michael Brown (ORNL) + + Base class for 3-body potentials + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : Tue April 2, 2013 + email : brownw@ornl.gov + ***************************************************************************/ + +#ifndef LAL_BASE_ATOMIC_H +#define LAL_BASE_ATOMIC_H + +#include "lal_device.h" +#include "lal_balance.h" +#include "mpi.h" + +#if defined(USE_OPENCL) +#include "geryon/ocl_texture.h" +#elif defined(USE_CUDART) +#include "geryon/nvc_texture.h" +#else +#include "geryon/nvd_texture.h" +#endif + +namespace LAMMPS_AL { + +template +class BaseThree { + public: + BaseThree(); + virtual ~BaseThree(); + + /// 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 + * \param k_two name for the kernel for 2-body force calculation + * \param k_three name for the kernel for 3-body force calculation + * + * 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 + * - -10 if invalid thread_per_atom setting **/ + int init_three(const int nlocal, const int nall, const int max_nbors, + const int maxspecial, const double cell_size, + const double gpu_split, FILE *screen, + const void *pair_program, const char *k_two, + const char *k_three_center, const char *k_three_end); + + /// Estimate the overhead for GPU context changes and CPU driver + void estimate_gpu_overhead(); + + /// Check if there is enough storage for atom arrays and realloc if not + /** \param success set to false if insufficient memory **/ + inline void resize_atom(const int inum, const int nall, bool &success) { + if (atom->resize(nall, success)) + pos_tex.bind_float(atom->x,4); + ans->resize(inum,success); + #ifdef THREE_CONCURRENT + ans2->resize(inum,success); + #endif + } + + /// Check if there is enough storage for neighbors and realloc if not + /** \param nlocal number of particles whose nbors must be stored on device + * \param host_inum number of particles whose nbors need to copied to host + * \param current maximum number of neighbors + * \note olist_size=total number of local particles **/ + inline void resize_local(const int inum, const int max_nbors, bool &success) { + nbor->resize(inum,max_nbors,success); + } + + /// Check if there is enough storage for neighbors and realloc if not + /** \param nlocal number of particles whose nbors must be stored on device + * \param host_inum number of particles whose nbors need to copied to host + * \param current maximum number of neighbors + * \note host_inum is 0 if the host is performing neighboring + * \note nlocal+host_inum=total number local particles + * \note olist_size=0 **/ + inline void resize_local(const int inum, const int host_inum, + const int max_nbors, bool &success) { + nbor->resize(inum,host_inum,max_nbors,success); + } + + /// Clear all host and device data + /** \note This is called at the beginning of the init() routine **/ + void clear_atomic(); + + /// Returns memory usage on device per atom + int bytes_per_atom_atomic(const int max_nbors) const; + + /// Total host memory used by library for pair style + double host_memory_usage_atomic() const; + + /// Accumulate timers + inline void acc_timers() { + if (device->time_device()) { + nbor->acc_timers(); + time_pair.add_to_total(); + atom->acc_timers(); + ans->acc_timers(); + #ifdef THREE_CONCURRENT + ans2->acc_timers(); + #endif + } + } + + /// Zero timers + inline void zero_timers() { + time_pair.zero(); + atom->zero_timers(); + ans->zero_timers(); + #ifdef THREE_CONCURRENT + ans2->zero_timers(); + #endif + } + + /// Copy neighbor list from host + int * reset_nbors(const int nall, const int inum, const int nlist, int *ilist, + int *numj, int **firstneigh, bool &success); + + /// Build neighbor list on device + void build_nbor_list(const int inum, const int host_inum, + const int nall, double **host_x, int *host_type, + double *sublo, double *subhi, int *tag, int **nspecial, + int **special, bool &success); + + /// Pair loop with host neighboring + void compute(const int f_ago, const int inum_full, const int nall, + const int nlist, 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); + + /// Pair loop with device neighboring + int * compute(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, + const double cpu_time, bool &success); + + /// Pair loop with device neighboring + int ** compute(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 **numj, const double cpu_time, bool &success); + + // -------------------------- DEVICE DATA ------------------------- + + /// Device Properties and Atom and Neighbor storage + Device *device; + + /// Geryon device + UCL_Device *ucl_device; + + /// Device Timers + UCL_Timer time_pair; + + /// Host device load balancer + Balance hd_balancer; + + /// LAMMPS pointer for screen output + FILE *screen; + + // --------------------------- ATOM DATA -------------------------- + + /// Atom Data + Atom *atom; + + // ------------------------ FORCE/ENERGY DATA ----------------------- + + Answer *ans; + #ifdef THREE_CONCURRENT + Answer *ans2; + #endif + + // --------------------------- NBOR DATA ---------------------------- + + /// Neighbor data + Neighbor *nbor; + + // ------------------------- DEVICE KERNELS ------------------------- + UCL_Program *pair_program; + UCL_Kernel k_pair, k_three_center, k_three_end, k_three_end_vatom; + inline int block_pair() { return _block_pair; } + inline int block_size() { return _block_size; } + + // --------------------------- TEXTURES ----------------------------- + UCL_Texture pos_tex; + + protected: + bool _compiled; + int _block_pair, _block_size, _threads_per_atom, _end_command_queue; + double _max_bytes, _max_an_bytes; + double _gpu_overhead, _driver_overhead; + UCL_D_Vec *_nbor_data; + + void compile_kernels(UCL_Device &dev, const void *pair_string, + const char *k_two, const char *k_three_center, + const char *k_three_end); + + virtual void loop(const bool _eflag, const bool _vflag, + const int evatom) = 0; +}; + +} + +#endif + diff --git a/lib/gpu/lal_beck.cpp b/lib/gpu/lal_beck.cpp new file mode 100644 index 0000000000..062c095957 --- /dev/null +++ b/lib/gpu/lal_beck.cpp @@ -0,0 +1,152 @@ +/*************************************************************************** + beck.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the beck pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifdef USE_OPENCL +#include "beck_cl.h" +#elif defined(USE_CUDART) +const char *beck=0; +#else +#include "beck_cubin.h" +#endif + +#include "lal_beck.h" +#include +using namespace LAMMPS_AL; +#define BeckT Beck + +extern Device device; + +template +BeckT::Beck() : BaseAtomic(), _allocated(false) { +} + +template +BeckT::~Beck() { + clear(); +} + +template +int BeckT::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int BeckT::init(const int ntypes, + double **host_cutsq, double **host_aa, + double **host_alpha, double **host_beta, + double **host_AA, double **host_BB, + 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) { + int success; + success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split, + _screen,beck,"k_beck"); + 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,beck1,host_write,host_aa,host_alpha, + host_beta); + + beck2.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); + this->atom->type_pack4(ntypes,lj_types,beck2,host_write,host_AA,host_BB, + host_cutsq); + + 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=beck1.row_bytes()+beck2.row_bytes()+sp_lj.row_bytes(); + return 0; +} + +template +void BeckT::clear() { + if (!_allocated) + return; + _allocated=false; + + beck1.clear(); + beck2.clear(); + sp_lj.clear(); + this->clear_atomic(); +} + +template +double BeckT::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(Beck); +} + +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void BeckT::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, &beck1, &beck2, &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, &beck1, &beck2, &_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 Beck; diff --git a/lib/gpu/lal_beck.h b/lib/gpu/lal_beck.h new file mode 100644 index 0000000000..fa56db2402 --- /dev/null +++ b/lib/gpu/lal_beck.h @@ -0,0 +1,80 @@ +/*************************************************************************** + beck.h + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the beck pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifndef LAL_BECK_H +#define LAL_BECK_H + +#include "lal_base_atomic.h" + +namespace LAMMPS_AL { + +template +class Beck : public BaseAtomic { + public: + Beck(); + ~Beck(); + + /// 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_aa, double **host_alpha, + double **host_beta, double **host_AA, + double **host_BB, 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); + + /// 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 -------------------------- + + /// beck1.x = aa, beck1.y = alpha, beck1.z = beta + UCL_D_Vec beck1; + /// beck2.x = AA, beck2.y = BB, beck2.z = cutsq + UCL_D_Vec beck2; + /// Special LJ values + 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_beck_ext.cpp b/lib/gpu/lal_beck_ext.cpp new file mode 100644 index 0000000000..1a736bf02d --- /dev/null +++ b/lib/gpu/lal_beck_ext.cpp @@ -0,0 +1,120 @@ +/*************************************************************************** + beck_ext.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Functions for LAMMPS access to beck acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_beck.h" + +using namespace std; +using namespace LAMMPS_AL; + +static Beck BLMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +int beck_gpu_init(const int ntypes, double **cutsq, double **aa, + double **alpha, double **beta, double **AA, double **BB, + 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) { + BLMF.clear(); + gpu_mode=BLMF.device->gpu_mode(); + double gpu_split=BLMF.device->particle_split(); + int first_gpu=BLMF.device->first_device(); + int last_gpu=BLMF.device->last_device(); + int world_me=BLMF.device->world_me(); + int gpu_rank=BLMF.device->gpu_rank(); + int procs_per_gpu=BLMF.device->procs_per_gpu(); + + BLMF.device->init_message(screen,"beck",first_gpu,last_gpu); + + bool message=false; + if (BLMF.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) + init_ok=BLMF.init(ntypes, cutsq, aa, alpha, beta, + AA, BB, special_lj, inum, nall, 300, + maxspecial, cell_size, gpu_split, screen); + + BLMF.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) + BLMF.estimate_gpu_overhead(); + return init_ok; +} + +void beck_gpu_clear() { + BLMF.clear(); +} + +int ** beck_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 BLMF.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 beck_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) { + BLMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj, + firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success); +} + +double beck_gpu_bytes() { + return BLMF.host_memory_usage(); +} + + diff --git a/lib/gpu/lal_lj_coul_msm.cpp b/lib/gpu/lal_lj_coul_msm.cpp new file mode 100644 index 0000000000..dd045b7970 --- /dev/null +++ b/lib/gpu/lal_lj_coul_msm.cpp @@ -0,0 +1,200 @@ +/*************************************************************************** + lj_coul_msm.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the lj/cut/coul/msm pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : brownw@ornl.gov + ***************************************************************************/ + +#if defined(USE_OPENCL) +#include "lj_coul_msm_cl.h" +#elif defined(USE_CUDART) +const char *lj_coul_msm=0; +#else +#include "lj_coul_msm_cubin.h" +#endif + +#include "lal_lj_coul_msm.h" +#include +using namespace LAMMPS_AL; +#define LJCoulMSMT LJCoulMSM + +extern Device device; + +template +LJCoulMSMT::LJCoulMSM() : BaseCharge(), + _allocated(false) { +} + +template +LJCoulMSMT::~LJCoulMSM() { + clear(); +} + +template +int LJCoulMSMT::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int LJCoulMSMT::init(const int ntypes, + double **host_cutsq, double **host_lj1, + double **host_lj2, double **host_lj3, + double **host_lj4, double **host_gcons, + double **host_dgcons, 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 int order, + const double qqrd2e) { + int success; + success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split, + _screen,lj_coul_msm,"k_lj_coul_msm"); + 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, host_cut_ljsq); + + 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, + host_offset); + + // pack gcons and dgcons + int nrows, ncols; + nrows = 7; + ncols = 7; + UCL_H_Vec dview_gcons(nrows*ncols,*(this->ucl_device), + UCL_WRITE_ONLY); + + for (int ix=0; ixucl_device),UCL_READ_ONLY); + ucl_copy(gcons,dview_gcons,false); + gcons_tex.get_texture(*(this->pair_program),"gcons_tex"); + gcons_tex.bind_float(gcons,1); + + nrows = 7; + ncols = 6; + UCL_H_Vec dview_dgcons(nrows*ncols,*(this->ucl_device), + UCL_WRITE_ONLY); + + for (int ix=0; ixucl_device),UCL_READ_ONLY); + ucl_copy(dgcons,dview_dgcons,false); + dgcons_tex.get_texture(*(this->pair_program),"dgcons_tex"); + dgcons_tex.bind_float(dgcons,1); + + 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(sp_lj,host_write,8,false); + + _cut_coulsq=host_cut_coulsq; + _qqrd2e=qqrd2e; + _order=order; + + _allocated=true; + this->_max_bytes=lj1.row_bytes()+lj3.row_bytes()+ + gcons.row_bytes()+dgcons.row_bytes()+sp_lj.row_bytes(); + return 0; +} + +template +void LJCoulMSMT::clear() { + if (!_allocated) + return; + _allocated=false; + + lj1.clear(); + lj3.clear(); + gcons.clear(); + dgcons.clear(); + sp_lj.clear(); + this->clear_atomic(); +} + +template +double LJCoulMSMT::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(LJCoulMSM); +} + +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void LJCoulMSMT::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, &gcons, &dgcons, &sp_lj, + &this->nbor->dev_nbor, &this->_nbor_data->begin(), + &this->ans->force, &this->ans->engv, &eflag, + &vflag, &ainum, &nbor_pitch, &this->atom->q, + &_cut_coulsq, &_qqrd2e, &_order, + &this->_threads_per_atom); + } else { + this->k_pair.set_size(GX,BX); + this->k_pair.run(&this->atom->x, &lj1, &lj3, &gcons, &dgcons, + &_lj_types, &sp_lj, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), &this->ans->force, + &this->ans->engv, &eflag, &vflag, &ainum, + &nbor_pitch, &this->atom->q, &_cut_coulsq, + &_qqrd2e, &_order, &this->_threads_per_atom); + } + this->time_pair.stop(); +} + +template class LJCoulMSM; diff --git a/lib/gpu/lal_lj_coul_msm.h b/lib/gpu/lal_lj_coul_msm.h new file mode 100644 index 0000000000..f1ef71cc2b --- /dev/null +++ b/lib/gpu/lal_lj_coul_msm.h @@ -0,0 +1,88 @@ +/*************************************************************************** + lj_coul_msm.h + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the lj/cut/coul/msm pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifndef LAL_LJ_COUL_MSM_H +#define LAL_LJ_COUL_MSM_H + +#include "lal_base_charge.h" + +namespace LAMMPS_AL { + +template +class LJCoulMSM : public BaseCharge { + public: + LJCoulMSM(); + ~LJCoulMSM(); + + /// 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_gcons, double **host_dgcons, + 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 int order, const double qqrd2e); + + /// 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 = cutsq_vdw + UCL_D_Vec lj1; + /// lj3.x = lj3, lj3.y = lj4, lj3.z = offset + UCL_D_Vec lj3; + /// Special LJ values [0-3] and Special Coul values [4-7] + UCL_D_Vec sp_lj; + + UCL_D_Vec gcons, dgcons; + UCL_Texture gcons_tex, dgcons_tex; + + /// If atom type constants fit in shared memory, use fast kernels + bool shared_types; + + /// Number of atom types + int _lj_types; + + numtyp _cut_coulsq, _qqrd2e; + int _order; + + private: + bool _allocated; + void loop(const bool _eflag, const bool _vflag); +}; + +} + +#endif diff --git a/lib/gpu/lal_lj_coul_msm_ext.cpp b/lib/gpu/lal_lj_coul_msm_ext.cpp new file mode 100644 index 0000000000..3480b14f3e --- /dev/null +++ b/lib/gpu/lal_lj_coul_msm_ext.cpp @@ -0,0 +1,131 @@ +/*************************************************************************** + lj_coul_msm_ext.cpp + ------------------- + W. Michael Brown (ORNL) + + Functions for LAMMPS access to lj/cut/coul/msm acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : brownw@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_lj_coul_msm.h" + +using namespace std; +using namespace LAMMPS_AL; + +static LJCoulMSM LJCMLMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +int ljcm_gpu_init(const int ntypes, double **cutsq, double **host_lj1, + double **host_lj2, double **host_lj3, double **host_lj4, + double **host_gcons, double **host_dgcons, + 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 int order, const double qqrd2e) { + LJCMLMF.clear(); + gpu_mode=LJCMLMF.device->gpu_mode(); + double gpu_split=LJCMLMF.device->particle_split(); + int first_gpu=LJCMLMF.device->first_device(); + int last_gpu=LJCMLMF.device->last_device(); + int world_me=LJCMLMF.device->world_me(); + int gpu_rank=LJCMLMF.device->gpu_rank(); + int procs_per_gpu=LJCMLMF.device->procs_per_gpu(); + + LJCMLMF.device->init_message(screen,"lj/cut/coul/msm",first_gpu,last_gpu); + + bool message=false; + if (LJCMLMF.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) + init_ok=LJCMLMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4, + host_gcons, host_dgcons, offset, + special_lj, inum, nall, 300, maxspecial, + cell_size, gpu_split, screen, host_cut_ljsq, + host_cut_coulsq, host_special_coul, order, qqrd2e); + + LJCMLMF.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) + LJCMLMF.estimate_gpu_overhead(); + return init_ok; +} + +void ljcm_gpu_clear() { + LJCMLMF.clear(); +} + +int** ljcm_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, double *host_q, double *boxlo, + double *prd) { + return LJCMLMF.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 ljcm_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) { + LJCMLMF.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 ljcm_gpu_bytes() { + return LJCMLMF.host_memory_usage(); +} + + diff --git a/lib/gpu/lal_mie.cpp b/lib/gpu/lal_mie.cpp new file mode 100644 index 0000000000..2ab7cb8d14 --- /dev/null +++ b/lib/gpu/lal_mie.cpp @@ -0,0 +1,152 @@ +/*************************************************************************** + mie.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the mie pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifdef USE_OPENCL +#include "mie_cl.h" +#elif defined(USE_CUDART) +const char *mie=0; +#else +#include "mie_cubin.h" +#endif + +#include "lal_mie.h" +#include +using namespace LAMMPS_AL; +#define MieT Mie + +extern Device device; + +template +MieT::Mie() : BaseAtomic(), _allocated(false) { +} + +template +MieT::~Mie() { + clear(); +} + +template +int MieT::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int MieT::init(const int ntypes, double **host_cutsq, + double **host_mie1, double **host_mie2, + double **host_mie3, double **host_mie4, + double **host_gamA, double **host_gamR, + 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) { + int success; + success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split, + _screen,mie,"k_mie"); + 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,mie1,host_write,host_mie1,host_mie2, + host_gamA,host_gamR); + + mie3.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); + this->atom->type_pack4(ntypes,lj_types,mie3,host_write,host_mie3,host_mie4, + host_offset,host_cutsq); + + 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=mie1.row_bytes()+mie3.row_bytes()+sp_lj.row_bytes(); + return 0; +} + +template +void MieT::clear() { + if (!_allocated) + return; + _allocated=false; + + mie1.clear(); + mie3.clear(); + sp_lj.clear(); + this->clear_atomic(); +} + +template +double MieT::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(Mie); +} + +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void MieT::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, &mie1, &mie3, &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, &mie1, &mie3, &_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 Mie; diff --git a/lib/gpu/lal_mie.h b/lib/gpu/lal_mie.h new file mode 100644 index 0000000000..e6b8efebf0 --- /dev/null +++ b/lib/gpu/lal_mie.h @@ -0,0 +1,80 @@ +/*************************************************************************** + mie.h + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the mie/cut pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifndef LAL_MIE_H +#define LAL_MIE_H + +#include "lal_base_atomic.h" + +namespace LAMMPS_AL { + +template +class Mie : public BaseAtomic { + public: + Mie(); + ~Mie(); + + /// 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_mie1, double **host_mie2, double **host_mie3, + double **host_mie4, double **host_gamA, double **host_gamR, + 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); + + /// 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 -------------------------- + + /// mie1.x = mie1, mie1.y = mie2, mie1.z = gamA, mie1.w = gamR + UCL_D_Vec mie1; + /// mie3.x = mie3, mie3.y = mie4, mie3.z = offset, mie3.w = cutsq + UCL_D_Vec mie3; + /// Special Mie values + 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_mie_ext.cpp b/lib/gpu/lal_mie_ext.cpp new file mode 100644 index 0000000000..52668cf52d --- /dev/null +++ b/lib/gpu/lal_mie_ext.cpp @@ -0,0 +1,124 @@ +/*************************************************************************** + mie_ext.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Functions for LAMMPS access to mie acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_mie.h" + +using namespace std; +using namespace LAMMPS_AL; + +static Mie MLMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +int mie_gpu_init(const int ntypes, double **cutsq, double **host_mie1, + double **host_mie2, double **host_mie3, double **host_mie4, + double **host_gamA, double **host_gamR, + 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) { + MLMF.clear(); + gpu_mode=MLMF.device->gpu_mode(); + double gpu_split=MLMF.device->particle_split(); + int first_gpu=MLMF.device->first_device(); + int last_gpu=MLMF.device->last_device(); + int world_me=MLMF.device->world_me(); + int gpu_rank=MLMF.device->gpu_rank(); + int procs_per_gpu=MLMF.device->procs_per_gpu(); + + MLMF.device->init_message(screen,"mie",first_gpu,last_gpu); + + bool message=false; + if (MLMF.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) + init_ok=MLMF.init(ntypes, cutsq, host_mie1, host_mie2, + host_mie3, host_mie4, host_gamA, host_gamR, + offset, special_lj, inum, nall, 300, + maxspecial, cell_size, gpu_split, screen); + + MLMF.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) + MLMF.estimate_gpu_overhead(); + return init_ok; +} + +void mie_gpu_clear() { + MLMF.clear(); +} + +int ** mie_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 MLMF.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 mie_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) { + MLMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj, + firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success); +} + +double mie_gpu_bytes() { + return MLMF.host_memory_usage(); +} + + diff --git a/lib/gpu/lal_soft.cpp b/lib/gpu/lal_soft.cpp new file mode 100644 index 0000000000..8db144080e --- /dev/null +++ b/lib/gpu/lal_soft.cpp @@ -0,0 +1,145 @@ +/*************************************************************************** + soft.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the soft pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifdef USE_OPENCL +#include "soft_cl.h" +#elif defined(USE_CUDART) +const char *soft=0; +#else +#include "soft_cubin.h" +#endif + +#include "lal_soft.h" +#include +using namespace LAMMPS_AL; +#define SoftT Soft + +extern Device device; + +template +SoftT::Soft() : BaseAtomic(), _allocated(false) { +} + +template +SoftT::~Soft() { + clear(); +} + +template +int SoftT::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int SoftT::init(const int ntypes, double **host_cutsq, + double **host_prefactor, double **host_cut, + 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) { + int success; + success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split, + _screen,soft,"k_soft"); + 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,coeff,host_write,host_prefactor, + host_cut,host_cutsq); + + 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=coeff.row_bytes()+sp_lj.row_bytes(); + return 0; +} + +template +void SoftT::clear() { + if (!_allocated) + return; + _allocated=false; + + coeff.clear(); + sp_lj.clear(); + this->clear_atomic(); +} + +template +double SoftT::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(Soft); +} + +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void SoftT::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, &coeff, &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, &coeff, &_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 Soft; diff --git a/lib/gpu/lal_soft.h b/lib/gpu/lal_soft.h new file mode 100644 index 0000000000..40e568e847 --- /dev/null +++ b/lib/gpu/lal_soft.h @@ -0,0 +1,77 @@ +/*************************************************************************** + soft.h + ------------------- + Trung Dac Nguyen (ORNL) + + Class for acceleration of the soft pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#ifndef LAL_GAUSS_H +#define LAL_GAYSS_H + +#include "lal_base_atomic.h" + +namespace LAMMPS_AL { + +template +class Soft : public BaseAtomic { + public: + Soft(); + ~Soft(); + + /// 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_prefactor, double **host_cut, + 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); + + /// 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 -------------------------- + + /// coeff.x = prefactor, coeff.y = cut, coeff.z = cutsq + UCL_D_Vec coeff; + /// Special LJ values + UCL_D_Vec sp_lj; + + /// If atom type constants fit in shared memory, use fast kßernels + 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_soft_ext.cpp b/lib/gpu/lal_soft_ext.cpp new file mode 100644 index 0000000000..3efeece427 --- /dev/null +++ b/lib/gpu/lal_soft_ext.cpp @@ -0,0 +1,120 @@ +/*************************************************************************** + soft_ext.cpp + ------------------- + Trung Dac Nguyen (ORNL) + + Functions for LAMMPS access to soft acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : + email : nguyentd@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_soft.h" + +using namespace std; +using namespace LAMMPS_AL; + +static Soft SLMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +int soft_gpu_init(const int ntypes, double **cutsq, double **host_prefactor, + double **host_cut, 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) { + SLMF.clear(); + gpu_mode=SLMF.device->gpu_mode(); + double gpu_split=SLMF.device->particle_split(); + int first_gpu=SLMF.device->first_device(); + int last_gpu=SLMF.device->last_device(); + int world_me=SLMF.device->world_me(); + int gpu_rank=SLMF.device->gpu_rank(); + int procs_per_gpu=SLMF.device->procs_per_gpu(); + + SLMF.device->init_message(screen,"soft",first_gpu,last_gpu); + + bool message=false; + if (SLMF.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) + init_ok=SLMF.init(ntypes, cutsq, host_prefactor, host_cut, + special_lj, inum, nall, 300, + maxspecial, cell_size, gpu_split, screen); + + SLMF.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) + SLMF.estimate_gpu_overhead(); + return init_ok; +} + +void soft_gpu_clear() { + SLMF.clear(); +} + +int ** soft_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 SLMF.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 soft_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) { + SLMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj, + firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success); +} + +double soft_gpu_bytes() { + return SLMF.host_memory_usage(); +} + + diff --git a/lib/gpu/lal_sw.cpp b/lib/gpu/lal_sw.cpp new file mode 100644 index 0000000000..bf6f820b69 --- /dev/null +++ b/lib/gpu/lal_sw.cpp @@ -0,0 +1,167 @@ +/*************************************************************************** + sw.cpp + ------------------- + W. Michael Brown (ORNL) + + Class for acceleration of the sw pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : Tue March 26, 2013 + email : brownw@ornl.gov + ***************************************************************************/ + +#if defined(USE_OPENCL) +#include "sw_cl.h" +#elif defined(USE_CUDART) +const char *lj=0; +#else +#include "sw_cubin.h" +#endif + +#include "lal_sw.h" +#include +using namespace LAMMPS_AL; +#define SWT SW + +extern Device device; + +template +SWT::SW() : BaseThree(), _allocated(false) { +} + +template +SWT::~SW() { + clear(); +} + +template +int SWT::bytes_per_atom(const int max_nbors) const { + return this->bytes_per_atom_atomic(max_nbors); +} + +template +int SWT::init(const int nlocal, const int nall, const int max_nbors, + const double cell_size, const double gpu_split, FILE *_screen, + const double epsilon, const double sigma, + const double lambda, const double gamma, + const double costheta, const double biga, + const double bigb, const double powerp, + const double powerq, const double cut, const double cutsq) { + + sw_epsilon=static_cast(epsilon); + sw_sigma=static_cast(sigma); + sw_lambda=static_cast(lambda); + sw_gamma=static_cast(gamma); + sw_costheta=static_cast(costheta); + sw_biga=static_cast(biga); + sw_bigb=static_cast(bigb); + sw_powerp=static_cast(powerp); + sw_powerq=static_cast(powerq); + sw_cut=static_cast(cut); + sw_cutsq=static_cast(cutsq); + if (sw_cutsq>=sw_cut*sw_cut) + sw_cutsq=sw_cut*sw_cut-1e-4; + + int success; + success=this->init_three(nlocal,nall,max_nbors,0,cell_size,gpu_split, + _screen,sw,"k_sw","k_sw_three_center", + "k_sw_three_end"); + if (success!=0) + return success; + + // If atom type constants fit in shared memory use fast kernel + shared_types=true; + + _allocated=true; + this->_max_bytes=0; + return 0; +} + +template +void SWT::clear() { + if (!_allocated) + return; + _allocated=false; + + this->clear_atomic(); +} + +template +double SWT::host_memory_usage() const { + return this->host_memory_usage_atomic()+sizeof(SW); +} + +#define KTHREADS this->_threads_per_atom +#define JTHREADS this->_threads_per_atom +// --------------------------------------------------------------------------- +// Calculate energies, forces, and torques +// --------------------------------------------------------------------------- +template +void SWT::loop(const bool _eflag, const bool _vflag, const int evatom) { + // Compute the block size and grid size to keep all cores busy + int BX=this->block_pair(); + 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(); + this->k_pair.set_size(GX,BX); + this->k_pair.run(&this->atom->x, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), &this->ans->force, + &this->ans->engv, &eflag, &vflag, &ainum, &nbor_pitch, + &this->_threads_per_atom, &sw_cut, &sw_epsilon, &sw_sigma, + &sw_biga, &sw_bigb, &sw_powerp, &sw_powerq, &sw_cutsq); + + BX=this->block_size(); + GX=static_cast(ceil(static_cast(this->ans->inum())/ + (BX/(KTHREADS*JTHREADS)))); + this->k_three_center.set_size(GX,BX); + this->k_three_center.run(&this->atom->x, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), &this->ans->force, + &this->ans->engv, &eflag, &vflag, &ainum, + &nbor_pitch, &this->_threads_per_atom, &evatom, + &sw_cut, &sw_epsilon, &sw_sigma, &sw_lambda, &sw_gamma, + &sw_costheta, &sw_cutsq); + Answer *end_ans; + #ifdef THREE_CONCURRENT + end_ans=this->ans2; + #else + end_ans=this->ans; + #endif + if (evatom!=0) { + this->k_three_end_vatom.set_size(GX,BX); + this->k_three_end_vatom.run(&this->atom->x, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), &end_ans->force, + &end_ans->engv, &eflag, &vflag, &ainum, + &nbor_pitch, &this->_threads_per_atom, &sw_cut, + &sw_epsilon, &sw_sigma, &sw_lambda, &sw_gamma, + &sw_costheta, &sw_cutsq); + } else { + this->k_three_end.set_size(GX,BX); + this->k_three_end.run(&this->atom->x, &this->nbor->dev_nbor, + &this->_nbor_data->begin(), &end_ans->force, + &end_ans->engv, &eflag, &vflag, &ainum, + &nbor_pitch, &this->_threads_per_atom, &sw_cut, + &sw_epsilon, &sw_sigma, &sw_lambda, &sw_gamma, + &sw_costheta, &sw_cutsq); + } + this->time_pair.stop(); +} + +template class SW; + diff --git a/lib/gpu/lal_sw.h b/lib/gpu/lal_sw.h new file mode 100644 index 0000000000..82fc7d24e0 --- /dev/null +++ b/lib/gpu/lal_sw.h @@ -0,0 +1,73 @@ +/*************************************************************************** + sw.h + ------------------- + W. Michael Brown (ORNL) + + Class for acceleration of the sw pair style. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : Tue March 26, 2013 + email : brownw@ornl.gov + ***************************************************************************/ + +#ifndef LAL_SW_H +#define LAL_SW_H + +#include "lal_base_three.h" + +namespace LAMMPS_AL { + +template +class SW : public BaseThree { + public: + SW(); + ~SW(); + + /// 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 nlocal, const int nall, const int max_nbors, + const double cell_size, const double gpu_split, FILE *screen, + const double epsilon, const double sigma, + const double lambda, const double gamma, + const double costheta, const double biga, + const double bigb, const double powerp, + const double powerq, const double cut, const double cutsq); + + /// 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 -------------------------- + + /// If atom type constants fit in shared memory, use fast kernels + bool shared_types; + + private: + bool _allocated; + void loop(const bool _eflag, const bool _vflag, const int evatom); + numtyp sw_epsilon, sw_sigma, sw_lambda, sw_gamma, sw_costheta; + numtyp sw_biga, sw_bigb, sw_powerp, sw_powerq, sw_cut, sw_cutsq; +}; + +} + +#endif + diff --git a/lib/gpu/lal_sw_ext.cpp b/lib/gpu/lal_sw_ext.cpp new file mode 100644 index 0000000000..33abf073d6 --- /dev/null +++ b/lib/gpu/lal_sw_ext.cpp @@ -0,0 +1,128 @@ +/*************************************************************************** + sw_ext.cpp + ------------------- + W. Michael Brown (ORNL) + + Functions for LAMMPS access to sw acceleration routines. + + __________________________________________________________________________ + This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) + __________________________________________________________________________ + + begin : Tue March 26, 2013 + email : brownw@ornl.gov + ***************************************************************************/ + +#include +#include +#include + +#include "lal_sw.h" + +using namespace std; +using namespace LAMMPS_AL; + +static SW SWMF; + +// --------------------------------------------------------------------------- +// Allocate memory on host and device and copy constants to device +// --------------------------------------------------------------------------- +int sw_gpu_init(const int inum, const int nall, const int max_nbors, + const double cell_size, int &gpu_mode, FILE *screen, + 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) { + SWMF.clear(); + gpu_mode=SWMF.device->gpu_mode(); + double gpu_split=SWMF.device->particle_split(); + int first_gpu=SWMF.device->first_device(); + int last_gpu=SWMF.device->last_device(); + int world_me=SWMF.device->world_me(); + int gpu_rank=SWMF.device->gpu_rank(); + int procs_per_gpu=SWMF.device->procs_per_gpu(); + + // disable host/device split for now + if (gpu_split != 1.0) + return -8; + + SWMF.device->init_message(screen,"sw/gpu",first_gpu,last_gpu); + + bool message=false; + if (SWMF.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) + init_ok=SWMF.init(inum, nall, 300, cell_size, gpu_split, screen, + sw_epsilon, sw_sigma, sw_lambda, sw_gamma, sw_costheta, + sw_biga, sw_bigb, sw_powerp, sw_powerq, sw_cut, sw_cutsq); + + SWMF.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) + SWMF.estimate_gpu_overhead(); + return init_ok; +} + +void sw_gpu_clear() { + SWMF.clear(); +} + +int ** sw_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 SWMF.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 sw_gpu_compute(const int ago, const int nlocal, const int nall, + const int nlist, 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) { + SWMF.compute(ago,nlocal,nall,nlist,host_x,host_type,ilist,numj, + firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success); +} + +double sw_gpu_bytes() { + return SWMF.host_memory_usage(); +} + +