Page Menu
Home
c4science
Search
Configure Global Search
Log In
Files
F88028745
lal_gauss.cpp
No One
Temporary
Actions
Download File
Edit File
Delete File
View Transforms
Subscribe
Mute Notifications
Award Token
Subscribers
None
File Metadata
Details
File Info
Storage
Attached
Created
Wed, Oct 16, 09:52
Size
5 KB
Mime Type
text/x-c++
Expires
Fri, Oct 18, 09:52 (2 d)
Engine
blob
Format
Raw Data
Handle
21703443
Attached To
rLAMMPS lammps
lal_gauss.cpp
View Options
/***************************************************************************
gauss.cpp
-------------------
Trung Dac Nguyen (ORNL)
Class for acceleration of the gauss pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : nguyentd@ornl.gov
***************************************************************************/
#ifdef USE_OPENCL
#include "gauss_cl.h"
#elif defined(USE_CUDART)
const char *gauss=0;
#else
#include "gauss_cubin.h"
#endif
#include "lal_gauss.h"
#include <cassert>
using namespace LAMMPS_AL;
#define GaussT Gauss<numtyp, acctyp>
extern Device<PRECISION,ACC_PRECISION> device;
template <class numtyp, class acctyp>
GaussT::Gauss() : BaseAtomic<numtyp,acctyp>(), _allocated(false) {
}
template <class numtyp, class acctyp>
GaussT::~Gauss() {
clear();
}
template <class numtyp, class acctyp>
int GaussT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors);
}
template <class numtyp, class acctyp>
int GaussT::init(const int ntypes,
double **host_cutsq, double **host_a,
double **host_b, 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,gauss,"k_gauss");
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<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;
gauss1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,gauss1,host_write,host_a,host_b,
host_cutsq,host_offset);
UCL_H_Vec<double> 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=gauss1.row_bytes()+sp_lj.row_bytes();
return 0;
}
template <class numtyp, class acctyp>
void GaussT::reinit(const int ntypes, double **host_cutsq, double **host_a,
double **host_b, double **host_offset) {
// 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->atom->type_pack4(ntypes,_lj_types,gauss1,host_write,host_a,host_b,
host_cutsq,host_offset);
}
template <class numtyp, class acctyp>
void GaussT::clear() {
if (!_allocated)
return;
_allocated=false;
gauss1.clear();
sp_lj.clear();
this->clear_atomic();
}
template <class numtyp, class acctyp>
double GaussT::host_memory_usage() const {
return this->host_memory_usage_atomic()+sizeof(Gauss<numtyp,acctyp>);
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void GaussT::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<int>(ceil(static_cast<double>(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, &gauss1, &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, &gauss1, &_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 Gauss<PRECISION,ACC_PRECISION>;
Event Timeline
Log In to Comment