forked from lijiext/lammps
git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@5545 f3b2605a-c512-4ea7-a41b-209d697bcdaa
This commit is contained in:
sjplimp
parent
9e1bc37b77
commit
a499b28469
|
|
@ -0,0 +1,135 @@
|
|||
/* ----------------------------------------------------------------------
|
||||
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
|
||||
http://lammps.sandia.gov, Sandia National Laboratories
|
||||
Steve Plimpton, sjplimp@sandia.gov
|
||||
|
||||
Copyright (2003) Sandia Corporation. Under the terms of Contract
|
||||
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
|
||||
certain rights in this software. This software is distributed under
|
||||
the GNU General Public License.
|
||||
|
||||
See the README file in the top-level LAMMPS directory.
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
/* ----------------------------------------------------------------------
|
||||
Contributing authors: Mike Brown (ORNL), brownw@ornl.gov
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
#include <iostream>
|
||||
#include <cassert>
|
||||
#include <math.h>
|
||||
|
||||
#include "crml_gpu_memory.h"
|
||||
|
||||
using namespace std;
|
||||
|
||||
static CRML_GPU_Memory<PRECISION,ACC_PRECISION> CRMLMF;
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Allocate memory on host and device and copy constants to device
|
||||
// ---------------------------------------------------------------------------
|
||||
bool crml_gpu_init(const int ntypes, double cut_bothsq, double **host_lj1,
|
||||
double **host_lj2, double **host_lj3, double **host_lj4,
|
||||
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 double qqrd2e,
|
||||
const double g_ewald, const double cut_lj_innersq,
|
||||
const double denom_lj, double **epsilon,
|
||||
double **sigma, const bool mix_arithmetic) {
|
||||
CRMLMF.clear();
|
||||
gpu_mode=CRMLMF.device->gpu_mode();
|
||||
double gpu_split=CRMLMF.device->particle_split();
|
||||
int first_gpu=CRMLMF.device->first_device();
|
||||
int last_gpu=CRMLMF.device->last_device();
|
||||
int world_me=CRMLMF.device->world_me();
|
||||
int gpu_rank=CRMLMF.device->gpu_rank();
|
||||
int procs_per_gpu=CRMLMF.device->procs_per_gpu();
|
||||
|
||||
CRMLMF.device->init_message(screen,"lj/charmm/coul/long",first_gpu,last_gpu);
|
||||
|
||||
bool message=false;
|
||||
if (CRMLMF.device->replica_me()==0 && screen)
|
||||
message=true;
|
||||
|
||||
if (message) {
|
||||
fprintf(screen,"Initializing GPU and compiling on process 0...");
|
||||
fflush(screen);
|
||||
}
|
||||
|
||||
if (world_me==0) {
|
||||
bool init_ok=CRMLMF.init(ntypes, cut_bothsq, host_lj1, host_lj2, host_lj3,
|
||||
host_lj4, offset, special_lj, inum, nall, 300,
|
||||
maxspecial, cell_size, gpu_split, screen,
|
||||
host_cut_ljsq, host_cut_coulsq, host_special_coul,
|
||||
qqrd2e, g_ewald, cut_lj_innersq, denom_lj,
|
||||
epsilon,sigma,mix_arithmetic);
|
||||
if (!init_ok)
|
||||
return false;
|
||||
}
|
||||
|
||||
CRMLMF.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) {
|
||||
bool init_ok=CRMLMF.init(ntypes, cut_bothsq, host_lj1, host_lj2, host_lj3,
|
||||
host_lj4, offset, special_lj, inum, nall, 300,
|
||||
maxspecial, cell_size, gpu_split,
|
||||
screen, host_cut_ljsq, host_cut_coulsq,
|
||||
host_special_coul, qqrd2e, g_ewald,
|
||||
cut_lj_innersq, denom_lj, epsilon, sigma,
|
||||
mix_arithmetic);
|
||||
if (!init_ok)
|
||||
return false;
|
||||
}
|
||||
CRMLMF.device->gpu_barrier();
|
||||
if (message)
|
||||
fprintf(screen,"Done.\n");
|
||||
}
|
||||
if (message)
|
||||
fprintf(screen,"\n");
|
||||
return true;
|
||||
}
|
||||
|
||||
void crml_gpu_clear() {
|
||||
CRMLMF.clear();
|
||||
}
|
||||
|
||||
int * crml_gpu_compute_n(const int timestep, const int ago, const int inum_full,
|
||||
const int nall, double **host_x, int *host_type,
|
||||
double *boxlo, double *boxhi, 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, double *host_q) {
|
||||
return CRMLMF.compute(timestep, ago, inum_full, nall, host_x, host_type, boxlo,
|
||||
boxhi, tag, nspecial, special, eflag, vflag, eatom,
|
||||
vatom, host_start, cpu_time, success, host_q);
|
||||
}
|
||||
|
||||
void crml_gpu_compute(const int timestep, 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) {
|
||||
CRMLMF.compute(timestep,ago,inum_full,nall,host_x,host_type,ilist,numj,
|
||||
firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,
|
||||
host_q);
|
||||
}
|
||||
|
||||
double crml_gpu_bytes() {
|
||||
return CRMLMF.host_memory_usage();
|
||||
}
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,169 @@
|
|||
/* ----------------------------------------------------------------------
|
||||
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
|
||||
http://lammps.sandia.gov, Sandia National Laboratories
|
||||
Steve Plimpton, sjplimp@sandia.gov
|
||||
|
||||
Copyright (2003) Sandia Corporation. Under the terms of Contract
|
||||
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
|
||||
certain rights in this software. This software is distributed under
|
||||
the GNU General Public License.
|
||||
|
||||
See the README file in the top-level LAMMPS directory.
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
/* ----------------------------------------------------------------------
|
||||
Contributing authors: Mike Brown (ORNL), brownw@ornl.gov
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
#ifdef USE_OPENCL
|
||||
#include "crml_gpu_cl.h"
|
||||
#else
|
||||
#include "crml_gpu_ptx.h"
|
||||
#endif
|
||||
|
||||
#include "crml_gpu_memory.h"
|
||||
#include <cassert>
|
||||
#define CRML_GPU_MemoryT CRML_GPU_Memory<numtyp, acctyp>
|
||||
|
||||
extern PairGPUDevice<PRECISION,ACC_PRECISION> pair_gpu_device;
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
CRML_GPU_MemoryT::CRML_GPU_Memory() : ChargeGPUMemory<numtyp,acctyp>(),
|
||||
_allocated(false) {
|
||||
}
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
CRML_GPU_MemoryT::~CRML_GPU_Memory() {
|
||||
clear();
|
||||
}
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
int CRML_GPU_MemoryT::bytes_per_atom(const int max_nbors) const {
|
||||
return this->bytes_per_atom_atomic(max_nbors);
|
||||
}
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
bool CRML_GPU_MemoryT::init(const int ntypes,
|
||||
double host_cut_bothsq, double **host_lj1,
|
||||
double **host_lj2, double **host_lj3,
|
||||
double **host_lj4, 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 double qqrd2e,
|
||||
const double g_ewald, const double cut_lj_innersq,
|
||||
const double denom_lj, double **epsilon,
|
||||
double **sigma, const bool mix_arithmetic) {
|
||||
this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
|
||||
_screen,crml_gpu_kernel);
|
||||
|
||||
// If atom type constants fit in shared memory use fast kernel
|
||||
int lj_types=ntypes;
|
||||
shared_types=false;
|
||||
if (this->_block_size>=64 && mix_arithmetic)
|
||||
shared_types=true;
|
||||
_lj_types=lj_types;
|
||||
|
||||
// Allocate a host write buffer for data initialization
|
||||
int h_size=lj_types*lj_types;
|
||||
if (h_size<MAX_BIO_SHARED_TYPES)
|
||||
h_size=MAX_BIO_SHARED_TYPES;
|
||||
UCL_H_Vec<numtyp> host_write(h_size*32,*(this->ucl_device),
|
||||
UCL_WRITE_OPTIMIZED);
|
||||
for (int i=0; i<h_size*32; i++)
|
||||
host_write[i]=0.0;
|
||||
|
||||
lj1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
|
||||
this->atom->type_pack4(ntypes,lj_types,lj1,host_write,host_lj1,host_lj2,
|
||||
host_lj3,host_lj4);
|
||||
|
||||
ljd.alloc(MAX_BIO_SHARED_TYPES,*(this->ucl_device),UCL_READ_ONLY);
|
||||
this->atom->self_pack2(ntypes,ljd,host_write,epsilon,sigma);
|
||||
|
||||
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_bothsq = host_cut_bothsq;
|
||||
_cut_coulsq = host_cut_coulsq;
|
||||
_cut_ljsq = host_cut_ljsq;
|
||||
_cut_lj_innersq = cut_lj_innersq;
|
||||
_qqrd2e=qqrd2e;
|
||||
_g_ewald=g_ewald;
|
||||
_denom_lj=denom_lj;
|
||||
|
||||
_allocated=true;
|
||||
this->_max_bytes=lj1.row_bytes()+ljd.row_bytes()+sp_lj.row_bytes();
|
||||
return true;
|
||||
}
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
void CRML_GPU_MemoryT::clear() {
|
||||
if (!_allocated)
|
||||
return;
|
||||
_allocated=false;
|
||||
|
||||
lj1.clear();
|
||||
ljd.clear();
|
||||
sp_lj.clear();
|
||||
this->clear_atomic();
|
||||
}
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
double CRML_GPU_MemoryT::host_memory_usage() const {
|
||||
return this->host_memory_usage_atomic()+sizeof(CRML_GPU_Memory<numtyp,acctyp>);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Calculate energies, forces, and torques
|
||||
// ---------------------------------------------------------------------------
|
||||
template <class numtyp, class acctyp>
|
||||
void CRML_GPU_MemoryT::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->atom->inum())/BX));
|
||||
|
||||
int ainum=this->atom->inum();
|
||||
int anall=this->atom->nall();
|
||||
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->dev_x.begin(), &ljd.begin(),
|
||||
&sp_lj.begin(), &this->nbor->dev_nbor.begin(),
|
||||
&this->atom->dev_ans.begin(),
|
||||
&this->atom->dev_engv.begin(), &eflag, &vflag,
|
||||
&ainum, &anall, &nbor_pitch,
|
||||
&this->atom->dev_q.begin(), &_cut_coulsq,
|
||||
&_qqrd2e, &_g_ewald, &_denom_lj, &_cut_bothsq,
|
||||
&_cut_ljsq, &_cut_lj_innersq);
|
||||
} else {
|
||||
this->k_pair.set_size(GX,BX);
|
||||
this->k_pair.run(&this->atom->dev_x.begin(), &lj1.begin(),
|
||||
&_lj_types, &sp_lj.begin(), &this->nbor->dev_nbor.begin(),
|
||||
&this->atom->dev_ans.begin(),
|
||||
&this->atom->dev_engv.begin(), &eflag, &vflag, &ainum,
|
||||
&anall, &nbor_pitch, &this->atom->dev_q.begin(),
|
||||
&_cut_coulsq, &_qqrd2e, &_g_ewald, &_denom_lj,
|
||||
&_cut_bothsq, &_cut_ljsq, &_cut_lj_innersq);
|
||||
}
|
||||
this->time_pair.stop();
|
||||
}
|
||||
|
||||
template class CRML_GPU_Memory<PRECISION,ACC_PRECISION>;
|
||||
|
|
@ -0,0 +1,79 @@
|
|||
/* ----------------------------------------------------------------------
|
||||
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
|
||||
http://lammps.sandia.gov, Sandia National Laboratories
|
||||
Steve Plimpton, sjplimp@sandia.gov
|
||||
|
||||
Copyright (2003) Sandia Corporation. Under the terms of Contract
|
||||
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
|
||||
certain rights in this software. This software is distributed under
|
||||
the GNU General Public License.
|
||||
|
||||
See the README file in the top-level LAMMPS directory.
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
/* ----------------------------------------------------------------------
|
||||
Contributing authors: Mike Brown (ORNL), brownw@ornl.gov
|
||||
------------------------------------------------------------------------- */
|
||||
|
||||
#ifndef CRML_GPU_MEMORY_H
|
||||
#define CRML_GPU_MEMORY_H
|
||||
|
||||
#include "charge_gpu_memory.h"
|
||||
|
||||
template <class numtyp, class acctyp>
|
||||
class CRML_GPU_Memory : public ChargeGPUMemory<numtyp, acctyp> {
|
||||
public:
|
||||
CRML_GPU_Memory();
|
||||
~CRML_GPU_Memory();
|
||||
|
||||
/// 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 **/
|
||||
bool init(const int ntypes, double host_cut_bothsq,
|
||||
double **host_lj1, double **host_lj2, double **host_lj3,
|
||||
double **host_lj4, 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 double qqrd2e, const double g_ewald,
|
||||
const double cut_lj_innersq, const double denom_lj,
|
||||
double **epsilon, double **sigma, const bool mix_arithmetic);
|
||||
|
||||
/// 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 --------------------------
|
||||
|
||||
/// x = lj1, y = lj2, z = lj3, w = lj4
|
||||
UCL_D_Vec<numtyp4> lj1;
|
||||
/// x = epsilon, y = sigma
|
||||
UCL_D_Vec<numtyp2> ljd;
|
||||
/// Special LJ values [0-3] and Special Coul values [4-7]
|
||||
UCL_D_Vec<numtyp> sp_lj;
|
||||
|
||||
/// If atom type constants fit in shared memory, use fast kernels
|
||||
bool shared_types;
|
||||
|
||||
/// Number of atom types
|
||||
int _lj_types;
|
||||
|
||||
numtyp _qqrd2e, _g_ewald, _denom_lj;
|
||||
|
||||
numtyp _cut_coulsq, _cut_bothsq, _cut_ljsq, _cut_lj_innersq;
|
||||
|
||||
private:
|
||||
bool _allocated;
|
||||
void loop(const bool _eflag, const bool _vflag);
|
||||
};
|
||||
|
||||
#endif
|
||||
|
||||
Loading…
Reference in New Issue