ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@12169 f3b2605a-c512-4ea7-a41b-209d697bcdaa

Browse Source
This commit is contained in:
sjplimp
2014-07-02 20:53:37 +00:00
parent 0553f4af2b
commit b5d89cd768
8 changed files with 1217 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

166
lib/gpu/lal_coul.cpp Normal file
View File

@ -0,0 +1,166 @@
/***************************************************************************
coul.cpp
-------------------
Trung Dac Nguyen
Class for acceleration of the coul/cut pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#if defined(USE_OPENCL)
#include "coul_cl.h"
#elif defined(USE_CUDART)
const char *coul=0;
#else
#include "coul_cubin.h"
#endif
#include "lal_coul.h"
#include <cassert>
using namespace LAMMPS_AL;
#define CoulT Coul<numtyp, acctyp>
extern Device<PRECISION,ACC_PRECISION> device;
template <class numtyp, class acctyp>
CoulT::Coul() : BaseCharge<numtyp,acctyp>(),
_allocated(false) {
}
template <class numtyp, class acctyp>
CoulT::~Coul() {
clear();
}
template <class numtyp, class acctyp>
int CoulT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors);
}
template <class numtyp, class acctyp>
int CoulT::init(const int ntypes, double **host_scale, double **host_cutsq,
double *host_special_coul, const int nlocal,
const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
const double gpu_split, FILE *_screen,
const double qqrd2e) {
int success;
success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
_screen,coul,"k_coul");
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;
scale.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack1(ntypes,lj_types,scale,host_write,host_scale);
cutsq.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack1(ntypes,lj_types,cutsq,host_write,host_cutsq);
sp_cl.alloc(4,*(this->ucl_device),UCL_READ_ONLY);
for (int i=0; i<4; i++) {
host_write[i]=host_special_coul[i];
}
ucl_copy(sp_cl,host_write,4,false);
_qqrd2e=qqrd2e;
_allocated=true;
this->_max_bytes=cutsq.row_bytes()+sp_cl.row_bytes();
return 0;
}
template <class numtyp, class acctyp>
void CoulT::reinit(const int ntypes, double **host_scale) {
// 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_pack1(ntypes,_lj_types,scale,host_write,host_scale);
}
template <class numtyp, class acctyp>
void CoulT::clear() {
if (!_allocated)
return;
_allocated=false;
scale.clear();
cutsq.clear();
sp_cl.clear();
this->clear_atomic();
}
template <class numtyp, class acctyp>
double CoulT::host_memory_usage() const {
return this->host_memory_usage_atomic()+sizeof(Coul<numtyp,acctyp>);
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void CoulT::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, &scale, &sp_cl,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag,
&vflag, &ainum, &nbor_pitch, &this->atom->q,
&cutsq, &_qqrd2e, &this->_threads_per_atom);
} else {
this->k_pair.set_size(GX,BX);
this->k_pair.run(&this->atom->x, &scale, &_lj_types, &sp_cl,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv,
&eflag, &vflag, &ainum, &nbor_pitch, &this->atom->q,
&cutsq, &_qqrd2e, &this->_threads_per_atom);
}
this->time_pair.stop();
}
template class Coul<PRECISION,ACC_PRECISION>;

211
lib/gpu/lal_coul.cu Normal file
View File

@ -0,0 +1,211 @@
// **************************************************************************
// coul.cu
// -------------------
// Trung Dac Nguyen
//
// Device code for acceleration of the coul/cut pair style
//
// __________________________________________________________________________
// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
// __________________________________________________________________________
//
// begin :
// email : ndtrung@umich.edu
// ***************************************************************************/
#ifdef NV_KERNEL
#include "lal_aux_fun1.h"
#ifndef _DOUBLE_DOUBLE
texture<float4> pos_tex;
texture<float> q_tex;
#else
texture<int4,1> pos_tex;
texture<int2> q_tex;
#endif
#else
#define pos_tex x_
#define q_tex q_
#endif
__kernel void k_coul(const __global numtyp4 *restrict x_,
const __global numtyp *restrict scale,
const int lj_types,
const __global numtyp *restrict sp_cl_in,
const __global int *dev_nbor,
const __global int *dev_packed,
__global acctyp4 *restrict ans,
__global acctyp *restrict engv,
const int eflag, const int vflag, const int inum,
const int nbor_pitch,
const __global numtyp *restrict q_,
const __global numtyp *restrict cutsq,
const numtyp qqrd2e, const int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
__local numtyp sp_cl[8];
sp_cl[0]=sp_cl_in[0];
sp_cl[1]=sp_cl_in[1];
sp_cl[2]=sp_cl_in[2];
sp_cl[3]=sp_cl_in[3];
acctyp energy=(acctyp)0;
acctyp e_coul=(acctyp)0;
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp virial[6];
for (int i=0; i<6; i++)
virial[i]=(acctyp)0;
if (ii<inum) {
const __global int *nbor, *list_end;
int i, numj;
__local int n_stride;
nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
n_stride,list_end,nbor);
numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
numtyp qtmp; fetch(qtmp,i,q_tex);
int itype=ix.w;
for ( ; nbor<list_end; nbor+=n_stride) {
int j=*nbor;
numtyp factor_coul;
factor_coul = sp_cl[sbmask(j)];
j &= NEIGHMASK;
numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
int jtype=jx.w;
// Compute r12
numtyp delx = ix.x-jx.x;
numtyp dely = ix.y-jx.y;
numtyp delz = ix.z-jx.z;
numtyp rsq = delx*delx+dely*dely+delz*delz;
int mtype=itype*lj_types+jtype;
if (rsq<cutsq[mtype]) {
numtyp r2inv=ucl_recip(rsq);
numtyp forcecoul, force;
fetch(forcecoul,j,q_tex);
forcecoul *= qqrd2e*scale[mtype]*qtmp*ucl_rsqrt(rsq)*factor_coul;
force = forcecoul * r2inv;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (eflag>0) {
e_coul += forcecoul;
}
if (vflag>0) {
virial[0] += delx*delx*force;
virial[1] += dely*dely*force;
virial[2] += delz*delz*force;
virial[3] += delx*dely*force;
virial[4] += delx*delz*force;
virial[5] += dely*delz*force;
}
}
} // for nbor
store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
vflag,ans,engv);
} // if ii
}
__kernel void k_coul_fast(const __global numtyp4 *restrict x_,
const __global numtyp *restrict scale,
const __global numtyp *restrict sp_cl_in,
const __global int *dev_nbor,
const __global int *dev_packed,
__global acctyp4 *restrict ans,
__global acctyp *restrict engv,
const int eflag, const int vflag, const int inum,
const int nbor_pitch,
const __global numtyp *restrict q_,
const __global numtyp *restrict _cutsq,
const numtyp qqrd2e, const int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
__local numtyp cutsq[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
__local numtyp sp_cl[4];
if (tid<4)
sp_cl[tid]=sp_cl_in[tid];
if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES) {
cutsq[tid]=_cutsq[tid];
}
acctyp energy=(acctyp)0;
acctyp e_coul=(acctyp)0;
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp virial[6];
for (int i=0; i<6; i++)
virial[i]=(acctyp)0;
__syncthreads();
if (ii<inum) {
const __global int *nbor, *list_end;
int i, numj;
__local int n_stride;
nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
n_stride,list_end,nbor);
numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
numtyp qtmp; fetch(qtmp,i,q_tex);
int iw=ix.w;
int itype=fast_mul((int)MAX_SHARED_TYPES,iw);
for ( ; nbor<list_end; nbor+=n_stride) {
int j=*nbor;
numtyp factor_coul = sp_cl[sbmask(j)];
j &= NEIGHMASK;
numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
int mtype=itype+jx.w;
// Compute r12
numtyp delx = ix.x-jx.x;
numtyp dely = ix.y-jx.y;
numtyp delz = ix.z-jx.z;
numtyp rsq = delx*delx+dely*dely+delz*delz;
if (rsq<cutsq[mtype]) {
numtyp r2inv=ucl_recip(rsq);
numtyp forcecoul, force;
fetch(forcecoul,j,q_tex);
forcecoul *= qqrd2e*scale[mtype]*qtmp*ucl_rsqrt(rsq)*factor_coul;
force = forcecoul * r2inv;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (eflag>0) {
e_coul += forcecoul;
}
if (vflag>0) {
virial[0] += delx*delx*force;
virial[1] += dely*dely*force;
virial[2] += delz*delz*force;
virial[3] += delx*dely*force;
virial[4] += delx*delz*force;
virial[5] += dely*delz*force;
}
}
} // for nbor
store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
vflag,ans,engv);
} // if ii
}

83
lib/gpu/lal_coul.h Normal file
View File

@ -0,0 +1,83 @@
/***************************************************************************
coul.h
-------------------
Trung Dac Nguyen
Class for acceleration of the coul/cut pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#ifndef LAL_COUL_H
#define LAL_COUL_H
#include "lal_base_charge.h"
namespace LAMMPS_AL {
template <class numtyp, class acctyp>
class Coul : public BaseCharge<numtyp, acctyp> {
public:
Coul();
~Coul();
/// 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_scale,
double **host_cutsq, double *host_special_coul,
const int nlocal, const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
const double gpu_split, FILE *screen, const double qqrd2e);
/// Send updated coeffs from host to device (to be compatible with fix adapt)
void reinit(const int ntypes, double **host_scale);
/// 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 --------------------------
/// cutsq
UCL_D_Vec<numtyp> scale;
/// cutsq
UCL_D_Vec<numtyp> cutsq;
/// Special Coul values [0-3]
UCL_D_Vec<numtyp> sp_cl;
/// If atom type constants fit in shared memory, use fast kernels
bool shared_types;
/// Number of atom types
int _lj_types;
numtyp _qqrd2e;
private:
bool _allocated;
void loop(const bool _eflag, const bool _vflag);
};
}
#endif

167
lib/gpu/lal_coul_debye.cpp Normal file
View File

@ -0,0 +1,167 @@
/***************************************************************************
coul_debye.cpp
-------------------
Trung Dac Nguyen
Class for acceleration of the coul/debye pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#ifdef USE_OPENCL
#include "coul_debye_cl.h"
#elif defined(USE_CUDART)
const char *coul_debye=0;
#else
#include "coul_debye_cubin.h"
#endif
#include "lal_coul_debye.h"
#include <cassert>
using namespace LAMMPS_AL;
#define CoulDebyeT CoulDebye<numtyp, acctyp>
extern Device<PRECISION,ACC_PRECISION> device;
template <class numtyp, class acctyp>
CoulDebyeT::CoulDebye() : BaseCharge<numtyp,acctyp>(),
_allocated(false) {
}
template <class numtyp, class acctyp>
CoulDebyeT::~CoulDebye() {
clear();
}
template <class numtyp, class acctyp>
int CoulDebyeT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors);
}
template <class numtyp, class acctyp>
int CoulDebyeT::init(const int ntypes, double **host_scale,
double **host_cutsq, double *host_special_coul,
const int nlocal, const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
const double gpu_split, FILE *_screen,
const double qqrd2e, const double kappa) {
int success;
success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
_screen,coul_debye,"k_coul_debye");
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;
scale.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack1(ntypes,lj_types,scale,host_write,host_scale);
cutsq.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack1(ntypes,lj_types,cutsq,host_write,host_cutsq);
sp_cl.alloc(4,*(this->ucl_device),UCL_READ_ONLY);
for (int i=0; i<4; i++) {
host_write[i]=host_special_coul[i];
}
ucl_copy(sp_cl,host_write,4,false);
_qqrd2e=qqrd2e;
_kappa=kappa;
_allocated=true;
this->_max_bytes=cutsq.row_bytes()+scale.row_bytes()+sp_cl.row_bytes();
return 0;
}
template <class numtyp, class acctyp>
void CoulDebyeT::reinit(const int ntypes, double **host_scale) {
// 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_pack1(ntypes,_lj_types,scale,host_write,host_scale);
}
template <class numtyp, class acctyp>
void CoulDebyeT::clear() {
if (!_allocated)
return;
_allocated=false;
scale.clear();
cutsq.clear();
sp_cl.clear();
this->clear_atomic();
}
template <class numtyp, class acctyp>
double CoulDebyeT::host_memory_usage() const {
return this->host_memory_usage_atomic()+sizeof(CoulDebye<numtyp,acctyp>);
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void CoulDebyeT::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, &scale, &sp_cl,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag, &vflag,
&ainum, &nbor_pitch, &this->atom->q, &cutsq,
&_qqrd2e, &_kappa, &this->_threads_per_atom);
} else {
this->k_pair.set_size(GX,BX);
this->k_pair.run(&this->atom->x, &scale, &_lj_types, &sp_cl,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag, &vflag,
&ainum, &nbor_pitch, &this->atom->q, &cutsq,
&_qqrd2e, &_kappa, &this->_threads_per_atom);
}
this->time_pair.stop();
}
template class CoulDebye<PRECISION,ACC_PRECISION>;

220
lib/gpu/lal_coul_debye.cu Normal file
View File

@ -0,0 +1,220 @@
// **************************************************************************
// coul_debye.cu
// -------------------
// Trung Dac Nguyen
//
// Device code for acceleration of the coul/debye pair style
//
// __________________________________________________________________________
// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
// __________________________________________________________________________
//
// begin :
// email : ndtrung@umich.edu
// ***************************************************************************/
#ifdef NV_KERNEL
#include "lal_aux_fun1.h"
#ifndef _DOUBLE_DOUBLE
texture<float4> pos_tex;
texture<float> q_tex;
#else
texture<int4,1> pos_tex;
texture<int2> q_tex;
#endif
#else
#define pos_tex x_
#define q_tex q_
#endif
__kernel void k_coul_debye(const __global numtyp4 *restrict x_,
const __global numtyp *restrict scale,
const int lj_types,
const __global numtyp *restrict sp_cl_in,
const __global int *dev_nbor,
const __global int *dev_packed,
__global acctyp4 *restrict ans,
__global acctyp *restrict engv,
const int eflag, const int vflag, const int inum,
const int nbor_pitch,
const __global numtyp *restrict q_ ,
const __global numtyp *restrict cutsq,
const numtyp qqrd2e, const numtyp kappa,
const int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
__local numtyp sp_cl[4];
sp_cl[0]=sp_cl_in[0];
sp_cl[1]=sp_cl_in[1];
sp_cl[2]=sp_cl_in[2];
sp_cl[3]=sp_cl_in[3];
acctyp energy=(acctyp)0;
acctyp e_coul=(acctyp)0;
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp virial[6];
for (int i=0; i<6; i++)
virial[i]=(acctyp)0;
if (ii<inum) {
const __global int *nbor, *list_end;
int i, numj;
__local int n_stride;
nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
n_stride,list_end,nbor);
numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
numtyp qtmp; fetch(qtmp,i,q_tex);
int itype=ix.w;
for ( ; nbor<list_end; nbor+=n_stride) {
int j=*nbor;
numtyp factor_coul = sp_cl[sbmask(j)];
j &= NEIGHMASK;
numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
int jtype=jx.w;
// Compute r12
numtyp delx = ix.x-jx.x;
numtyp dely = ix.y-jx.y;
numtyp delz = ix.z-jx.z;
numtyp rsq = delx*delx+dely*dely+delz*delz;
int mtype=itype*lj_types+jtype;
if (rsq<cutsq[mtype]) {
numtyp r2inv=ucl_recip(rsq);
numtyp forcecoul, force, r, rinv, screening;
r = ucl_sqrt(rsq);
rinv = ucl_recip(r);
fetch(screening,j,q_tex);
screening *= ucl_exp(-kappa*r);
forcecoul = qqrd2e*qtmp*scale[mtype]*(kappa+rinv)*screening*factor_coul;
force = forcecoul * r2inv;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (eflag>0) {
e_coul+=qqrd2e*scale[mtype]*qtmp*rinv*screening*factor_coul;
}
if (vflag>0) {
virial[0] += delx*delx*force;
virial[1] += dely*dely*force;
virial[2] += delz*delz*force;
virial[3] += delx*dely*force;
virial[4] += delx*delz*force;
virial[5] += dely*delz*force;
}
}
} // for nbor
store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
vflag,ans,engv);
} // if ii
}
__kernel void k_coul_debye_fast(const __global numtyp4 *restrict x_,
const __global numtyp *restrict scale_in,
const __global numtyp *restrict sp_cl_in,
const __global int *dev_nbor,
const __global int *dev_packed,
__global acctyp4 *restrict ans,
__global acctyp *restrict engv,
const int eflag, const int vflag, const int inum,
const int nbor_pitch,
const __global numtyp *restrict q_,
const __global numtyp *restrict _cutsq,
const numtyp qqrd2e, const numtyp kappa,
const int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
__local numtyp scale[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
__local numtyp cutsq[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
__local numtyp sp_cl[4];
if (tid<4)
sp_cl[tid]=sp_cl_in[tid];
if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES) {
scale[tid]=scale_in[tid];
cutsq[tid]=_cutsq[tid];
}
acctyp energy=(acctyp)0;
acctyp e_coul=(acctyp)0;
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp virial[6];
for (int i=0; i<6; i++)
virial[i]=(acctyp)0;
__syncthreads();
if (ii<inum) {
const __global int *nbor, *list_end;
int i, numj;
__local int n_stride;
nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
n_stride,list_end,nbor);
numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
numtyp qtmp; fetch(qtmp,i,q_tex);
int iw=ix.w;
int itype=fast_mul((int)MAX_SHARED_TYPES,iw);
for ( ; nbor<list_end; nbor+=n_stride) {
int j=*nbor;
numtyp factor_coul = sp_cl[sbmask(j)];
j &= NEIGHMASK;
numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
int mtype=itype+jx.w;
// Compute r12
numtyp delx = ix.x-jx.x;
numtyp dely = ix.y-jx.y;
numtyp delz = ix.z-jx.z;
numtyp rsq = delx*delx+dely*dely+delz*delz;
if (rsq<cutsq[mtype]) {
numtyp r2inv=ucl_recip(rsq);
numtyp forcecoul, force, r, rinv, screening;
r = ucl_sqrt(rsq);
rinv = ucl_recip(r);
fetch(screening,j,q_tex);
screening *= ucl_exp(-kappa*r);
forcecoul = qqrd2e*scale[mtype]*qtmp*(kappa+rinv)*screening*factor_coul;
force = forcecoul * r2inv;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (eflag>0) {
e_coul+=qqrd2e*scale[mtype]*qtmp*rinv*screening*factor_coul;
}
if (vflag>0) {
virial[0] += delx*delx*force;
virial[1] += dely*dely*force;
virial[2] += delz*delz*force;
virial[3] += delx*dely*force;
virial[4] += delx*delz*force;
virial[5] += dely*delz*force;
}
}
} // for nbor
store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
vflag,ans,engv);
} // if ii
}

84
lib/gpu/lal_coul_debye.h Normal file
View File

@ -0,0 +1,84 @@
/***************************************************************************
coul_debye.h
-------------------
Trung Dac Nguyen
Class for acceleration of the coul/debye pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#ifndef LAL_COUL_DEBYE_H
#define LAL_COUL_DEBYE_H
#include "lal_base_charge.h"
namespace LAMMPS_AL {
template <class numtyp, class acctyp>
class CoulDebye : public BaseCharge<numtyp, acctyp> {
public:
CoulDebye();
~CoulDebye();
/// 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_scale,
double **host_cutsq, double *host_special_coul,
const int nlocal, const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
const double gpu_split, FILE *screen,
const double qqrd2e, const double kappa);
/// Send updated coeffs from host to device (to be compatible with fix adapt)
void reinit(const int ntypes, double **host_scale);
/// 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 --------------------------
/// scale
UCL_D_Vec<numtyp> scale;
/// cutsq
UCL_D_Vec<numtyp> cutsq;
/// Special Coul values [0-3]
UCL_D_Vec<numtyp> sp_cl;
/// If atom type constants fit in shared memory, use fast kernels
bool shared_types;
/// Number of atom types
int _lj_types;
numtyp _qqrd2e,_kappa;
private:
bool _allocated;
void loop(const bool _eflag, const bool _vflag);
};
}
#endif

143
lib/gpu/lal_coul_debye_ext.cpp Normal file
View File

@ -0,0 +1,143 @@
/***************************************************************************
coul_debye_ext.cpp
-------------------
Trung Dac Nguyen
Functions for LAMMPS access to coul/debye acceleration routines.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#include <iostream>
#include <cassert>
#include <math.h>
#include "lal_coul_debye.h"
using namespace std;
using namespace LAMMPS_AL;
static CoulDebye<PRECISION,ACC_PRECISION> CDEMF;
// ---------------------------------------------------------------------------
// Allocate memory on host and device and copy constants to device
// ---------------------------------------------------------------------------
int cdebye_gpu_init(const int ntypes, double **host_scale, double **cutsq,
double *host_special_coul, const int inum,
const int nall, const int max_nbors, const int maxspecial,
const double cell_size, int &gpu_mode, FILE *screen,
const double qqrd2e, const double kappa) {
CDEMF.clear();
gpu_mode=CDEMF.device->gpu_mode();
double gpu_split=CDEMF.device->particle_split();
int first_gpu=CDEMF.device->first_device();
int last_gpu=CDEMF.device->last_device();
int world_me=CDEMF.device->world_me();
int gpu_rank=CDEMF.device->gpu_rank();
int procs_per_gpu=CDEMF.device->procs_per_gpu();
CDEMF.device->init_message(screen,"coul/debye",first_gpu,last_gpu);
bool message=false;
if (CDEMF.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=CDEMF.init(ntypes, host_scale, cutsq, host_special_coul, inum, nall, 300,
maxspecial, cell_size, gpu_split, screen, qqrd2e, kappa);
CDEMF.device->world_barrier();
if (message)
fprintf(screen,"Done.\n");
for (int i=0; i<procs_per_gpu; i++) {
if (message) {
if (last_gpu-first_gpu==0)
fprintf(screen,"Initializing GPU %d on core %d...",first_gpu,i);
else
fprintf(screen,"Initializing GPUs %d-%d on core %d...",first_gpu,
last_gpu,i);
fflush(screen);
}
if (gpu_rank==i && world_me!=0)
init_ok=CDEMF.init(ntypes, host_scale, cutsq, host_special_coul, inum, nall, 300,
maxspecial, cell_size, gpu_split, screen, qqrd2e, kappa);
CDEMF.device->gpu_barrier();
if (message)
fprintf(screen,"Done.\n");
}
if (message)
fprintf(screen,"\n");
if (init_ok==0)
CDEMF.estimate_gpu_overhead();
return init_ok;
}
// ---------------------------------------------------------------------------
// Copy updated constants to device
// ---------------------------------------------------------------------------
void cdebye_gpu_reinit(const int ntypes, double **host_scale) {
int world_me=CDEMF.device->world_me();
int gpu_rank=CDEMF.device->gpu_rank();
int procs_per_gpu=CDEMF.device->procs_per_gpu();
if (world_me==0)
CDEMF.reinit(ntypes, host_scale);
CDEMF.device->world_barrier();
for (int i=0; i<procs_per_gpu; i++) {
if (gpu_rank==i && world_me!=0)
CDEMF.reinit(ntypes, host_scale);
CDEMF.device->gpu_barrier();
}
}
void cdebye_gpu_clear() {
CDEMF.clear();
}
int** cdebye_gpu_compute_n(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
double *sublo, double *subhi, tagint *tag, int **nspecial,
tagint **special, const bool eflag, const bool vflag,
const bool eatom, const bool vatom, int &host_start,
int **ilist, int **jnum, const double cpu_time,
bool &success, double *host_q, double *boxlo,
double *prd) {
return CDEMF.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 cdebye_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) {
CDEMF.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 cdebye_gpu_bytes() {
return CDEMF.host_memory_usage();
}

143
lib/gpu/lal_coul_ext.cpp Normal file
View File

@ -0,0 +1,143 @@
/***************************************************************************
coul_ext.cpp
-------------------
Trung Dac Nguyen
Functions for LAMMPS access to coul/cut acceleration routines.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : ndtrung@umich.edu
***************************************************************************/
#include <iostream>
#include <cassert>
#include <math.h>
#include "lal_coul.h"
using namespace std;
using namespace LAMMPS_AL;
static Coul<PRECISION,ACC_PRECISION> COULMF;
// ---------------------------------------------------------------------------
// Allocate memory on host and device and copy constants to device
// ---------------------------------------------------------------------------
int coul_gpu_init(const int ntypes, double **host_scale,
double **cutsq, double *special_coul,
const int inum, const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
int &gpu_mode, FILE *screen, const double qqrd2e) {
COULMF.clear();
gpu_mode=COULMF.device->gpu_mode();
double gpu_split=COULMF.device->particle_split();
int first_gpu=COULMF.device->first_device();
int last_gpu=COULMF.device->last_device();
int world_me=COULMF.device->world_me();
int gpu_rank=COULMF.device->gpu_rank();
int procs_per_gpu=COULMF.device->procs_per_gpu();
COULMF.device->init_message(screen,"coul/cut",first_gpu,last_gpu);
bool message=false;
if (COULMF.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=COULMF.init(ntypes, host_scale, cutsq, special_coul, inum, nall, 300,
maxspecial, cell_size, gpu_split, screen, qqrd2e);
COULMF.device->world_barrier();
if (message)
fprintf(screen,"Done.\n");
for (int i=0; i<procs_per_gpu; i++) {
if (message) {
if (last_gpu-first_gpu==0)
fprintf(screen,"Initializing GPU %d on core %d...",first_gpu,i);
else
fprintf(screen,"Initializing GPUs %d-%d on core %d...",first_gpu,
last_gpu,i);
fflush(screen);
}
if (gpu_rank==i && world_me!=0)
init_ok=COULMF.init(ntypes, host_scale, cutsq, special_coul, inum, nall, 300,
maxspecial, cell_size, gpu_split, screen, qqrd2e);
COULMF.device->gpu_barrier();
if (message)
fprintf(screen,"Done.\n");
}
if (message)
fprintf(screen,"\n");
if (init_ok==0)
COULMF.estimate_gpu_overhead();
return init_ok;
}
// ---------------------------------------------------------------------------
// Copy updated constants to device
// ---------------------------------------------------------------------------
void coul_gpu_reinit(const int ntypes, double **host_scale) {
int world_me=COULMF.device->world_me();
int gpu_rank=COULMF.device->gpu_rank();
int procs_per_gpu=COULMF.device->procs_per_gpu();
if (world_me==0)
COULMF.reinit(ntypes, host_scale);
COULMF.device->world_barrier();
for (int i=0; i<procs_per_gpu; i++) {
if (gpu_rank==i && world_me!=0)
COULMF.reinit(ntypes, host_scale);
COULMF.device->gpu_barrier();
}
}
void coul_gpu_clear() {
COULMF.clear();
}
int** coul_gpu_compute_n(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
double *sublo, double *subhi, tagint *tag, int **nspecial,
tagint **special, const bool eflag, const bool vflag,
const bool eatom, const bool vatom, int &host_start,
int **ilist, int **jnum, const double cpu_time,
bool &success, double *host_q, double *boxlo,
double *prd) {
return COULMF.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 coul_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) {
COULMF.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 coul_gpu_bytes() {
return COULMF.host_memory_usage();
}