jianmu
/
lammps
forked from lijiext/lammps
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
lammps/lib/gpu/lal_table.cpp

341 lines
13 KiB
C++
Raw Blame History

/***************************************************************************
lal_table.cpp
-------------------
Trung Dac Nguyen (ORNL)
Class for acceleration of the table pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : nguyentd@ornl.gov
***************************************************************************/
#if defined(USE_OPENCL)
#include "table_cl.h"
#elif defined(USE_CUDART)
const char *table=0;
#else
#include "table_cubin.h"
#endif
#include "lal_table.h"
#include <cassert>
namespace LAMMPS_AL {
#define TableT Table<numtyp, acctyp>
#define LOOKUP 0
#define LINEAR 1
#define SPLINE 2
#define BITMAP 3
extern Device<PRECISION,ACC_PRECISION> device;
template <class numtyp, class acctyp>
TableT::Table() : BaseAtomic<numtyp,acctyp>(),
_allocated(false), _compiled_styles(false) {
}
template <class numtyp, class acctyp>
TableT::~Table() {
clear();
}
template <class numtyp, class acctyp>
int TableT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors);
}
template <class numtyp, class acctyp>
int TableT::init(const int ntypes,
double **host_cutsq, double ***host_table_coeffs,
double **host_table_data,
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 tabstyle, int ntables, int tablength) {
int success;
success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,
gpu_split,_screen,table,"k_table");
if (success!=0)
return success;
k_pair_linear.set_function(*(this->pair_program),"k_table_linear");
k_pair_linear_fast.set_function(*(this->pair_program),"k_table_linear_fast");
k_pair_spline.set_function(*(this->pair_program),"k_table_spline");
k_pair_spline_fast.set_function(*(this->pair_program),"k_table_spline_fast");
k_pair_bitmap.set_function(*(this->pair_program),"k_table_bitmap");
k_pair_bitmap_fast.set_function(*(this->pair_program),"k_table_bitmap_fast");
_compiled_styles = true;
// 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;
_tabstyle = tabstyle;
_ntables = ntables;
if (tabstyle != BITMAP) _tablength = tablength;
else _tablength = 1 << tablength;
// Allocate a host write buffer for data initialization
UCL_H_Vec<int> host_write_int(lj_types*lj_types,*(this->ucl_device),
UCL_WRITE_ONLY);
for (int i=0; i<lj_types*lj_types; i++)
host_write_int[i] = 0;
tabindex.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
nshiftbits.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
nmask.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
for (int ix=1; ix<ntypes; ix++)
for (int iy=1; iy<ntypes; iy++)
host_write_int[ix*lj_types+iy] = (int)host_table_coeffs[ix][iy][0]; // tabindex
ucl_copy(tabindex,host_write_int,false);
for (int ix=1; ix<ntypes; ix++)
for (int iy=1; iy<ntypes; iy++)
host_write_int[ix*lj_types+iy] = (int)host_table_coeffs[ix][iy][1]; // nshiftbits
ucl_copy(nshiftbits,host_write_int,false);
for (int ix=1; ix<ntypes; ix++)
for (int iy=1; iy<ntypes; iy++)
host_write_int[ix*lj_types+iy] = (int)host_table_coeffs[ix][iy][2]; // nmask
ucl_copy(nmask,host_write_int,false);
UCL_H_Vec<numtyp4> host_write(lj_types*lj_types,*(this->ucl_device),
UCL_WRITE_ONLY);
coeff2.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
for (int ix=1; ix<ntypes; ix++)
for (int iy=1; iy<ntypes; iy++) {
host_write[ix*lj_types+iy].x = host_table_coeffs[ix][iy][3]; // innersq
host_write[ix*lj_types+iy].y = host_table_coeffs[ix][iy][4]; // invdelta
host_write[ix*lj_types+iy].z = host_table_coeffs[ix][iy][5]; // deltasq6
host_write[ix*lj_types+iy].w = (numtyp)0.0;
}
ucl_copy(coeff2,host_write,false);
// Allocate tablength arrays
UCL_H_Vec<numtyp4> host_write2(_ntables*_tablength,*(this->ucl_device),
UCL_WRITE_ONLY);
for (int i=0; i<_ntables*_tablength; i++) {
host_write2[i].x = 0.0;
host_write2[i].y = 0.0;
host_write2[i].z = 0.0;
host_write2[i].w = 0.0;
}
coeff3.alloc(_ntables*_tablength,*(this->ucl_device),UCL_READ_ONLY);
for (int n=0; n<_ntables; n++) {
if (tabstyle == LOOKUP) {
for (int k=0; k<_tablength-1; k++) {
host_write2[n*_tablength+k].x = (numtyp)0;
host_write2[n*_tablength+k].y = host_table_data[n][6*k+1]; // e
host_write2[n*_tablength+k].z = host_table_data[n][6*k+2]; // f
host_write2[n*_tablength+k].w = (numtyp)0;
}
} else if (tabstyle == LINEAR || tabstyle == SPLINE || tabstyle == BITMAP) {
for (int k=0; k<_tablength; k++) {
host_write2[n*_tablength+k].x = host_table_data[n][6*k+0]; // rsq
host_write2[n*_tablength+k].y = host_table_data[n][6*k+1]; // e
host_write2[n*_tablength+k].z = host_table_data[n][6*k+2]; // f
host_write2[n*_tablength+k].w = (numtyp)0;
}
}
}
ucl_copy(coeff3,host_write2,false);
coeff4.alloc(_ntables*_tablength,*(this->ucl_device),UCL_READ_ONLY);
for (int i=0; i<_ntables*_tablength; i++) {
host_write2[i].x = 0.0;
host_write2[i].y = 0.0;
host_write2[i].z = 0.0;
host_write2[i].w = 0.0;
}
for (int n=0; n<_ntables; n++) {
if (tabstyle == LINEAR) {
for (int k=0; k<_tablength-1; k++) {
host_write2[n*_tablength+k].x = (numtyp)0;
host_write2[n*_tablength+k].y = host_table_data[n][6*k+3]; // de
host_write2[n*_tablength+k].z = host_table_data[n][6*k+4]; // df
host_write2[n*_tablength+k].w = (numtyp)0;
}
} else if (tabstyle == SPLINE) {
for (int k=0; k<_tablength; k++) {
host_write2[n*_tablength+k].x = (numtyp)0;
host_write2[n*_tablength+k].y = host_table_data[n][6*k+3]; // e2
host_write2[n*_tablength+k].z = host_table_data[n][6*k+4]; // f2
host_write2[n*_tablength+k].w = (numtyp)0;
}
} else if (tabstyle == BITMAP) {
for (int k=0; k<_tablength; k++) {
host_write2[n*_tablength+k].x = (numtyp)0;
host_write2[n*_tablength+k].y = host_table_data[n][6*k+3]; // de
host_write2[n*_tablength+k].z = host_table_data[n][6*k+4]; // df
host_write2[n*_tablength+k].w = host_table_data[n][6*k+5]; // drsq
}
}
}
ucl_copy(coeff4,host_write2,false);
UCL_H_Vec<numtyp> host_rsq(lj_types*lj_types,*(this->ucl_device),
UCL_WRITE_ONLY);
cutsq.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack1(ntypes,lj_types,cutsq,host_rsq,host_cutsq);
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=tabindex.row_bytes()+nshiftbits.row_bytes()
+nmask.row_bytes()+coeff2.row_bytes()
+coeff3.row_bytes()+coeff4.row_bytes()+cutsq.row_bytes()
+sp_lj.row_bytes();
return 0;
}
template <class numtyp, class acctyp>
void TableT::clear() {
if (!_allocated)
return;
_allocated=false;
tabindex.clear();
nshiftbits.clear();
nmask.clear();
coeff2.clear();
coeff3.clear();
coeff4.clear();
sp_lj.clear();
if (_compiled_styles) {
k_pair_linear_fast.clear();
k_pair_linear.clear();
k_pair_spline_fast.clear();
k_pair_spline.clear();
k_pair_bitmap_fast.clear();
k_pair_bitmap.clear();
_compiled_styles=false;
}
this->clear_atomic();
}
template <class numtyp, class acctyp>
double TableT::host_memory_usage() const {
return this->host_memory_usage_atomic()+sizeof(Table<numtyp,acctyp>);
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void TableT::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) {
if (_tabstyle == LOOKUP) {
this->k_pair_fast.set_size(GX,BX);
this->k_pair_fast.run(&this->atom->x, &tabindex, &coeff2, &coeff3,
&coeff4, &cutsq, &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, &_tablength);
} else if (_tabstyle == LINEAR) {
this->k_pair_linear_fast.set_size(GX,BX);
this->k_pair_linear_fast.run(&this->atom->x, &tabindex, &coeff2,
&coeff3, &coeff4, &cutsq, &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, &_tablength);
} else if (_tabstyle == SPLINE) {
this->k_pair_spline_fast.set_size(GX,BX);
this->k_pair_spline_fast.run(&this->atom->x, &tabindex, &coeff2,
&coeff3, &coeff4, &cutsq, &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, &_tablength);
} else if (_tabstyle == BITMAP) {
this->k_pair_bitmap_fast.set_size(GX,BX);
this->k_pair_bitmap_fast.run(&this->atom->x, &tabindex, &nshiftbits,
&nmask, &coeff2, &coeff3, &coeff4, &cutsq,
&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, &_tablength);
}
} else {
if (_tabstyle == LOOKUP) {
this->k_pair.set_size(GX,BX);
this->k_pair.run(&this->atom->x, &tabindex, &coeff2, &coeff3,
&coeff4, &_lj_types, &cutsq, &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,
&_tablength);
} else if (_tabstyle == LINEAR) {
this->k_pair_linear.set_size(GX,BX);
this->k_pair_linear.run(&this->atom->x, &tabindex, &coeff2, &coeff3,
&coeff4, &_lj_types, &cutsq, &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, &_tablength);
} else if (_tabstyle == SPLINE) {
this->k_pair_spline.set_size(GX,BX);
this->k_pair_spline.run(&this->atom->x, &tabindex, &coeff2, &coeff3,
&coeff4, &_lj_types, &cutsq, &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, &_tablength);
} else if (_tabstyle == BITMAP) {
this->k_pair_bitmap.set_size(GX,BX);
this->k_pair_bitmap.run(&this->atom->x, &tabindex, &nshiftbits,
&nmask, &coeff2, &coeff3, &coeff4, &_lj_types,
&cutsq, &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,
&_tablength);
}
}
this->time_pair.stop();
}
template class Table<PRECISION,ACC_PRECISION>;
}
Reference in New Issue View Git Blame Copy Permalink