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@6053 f3b2605a-c512-4ea7-a41b-209d697bcdaa

Browse Source
This commit is contained in:
sjplimp
2011-05-02 15:02:52 +00:00
parent 2be078632d
commit 5f799182b3
70 changed files with 4489 additions and 2253 deletions
Download Patch File Download Diff File Expand all files Collapse all files

360
lib/gpu/pair_gpu_atom.cpp
View File

@ -29,9 +29,8 @@ __win_sort _win_sort;
#endif
template <class numtyp, class acctyp>
PairGPUAtomT::PairGPUAtom() : _compiled(false),_allocated(false),_eflag(false),
_vflag(false),_inum(0),_ilist(NULL),
_newton(false) {
PairGPUAtomT::PairGPUAtom() : _compiled(false),_allocated(false),
_max_gpu_bytes(0) {
#ifndef USE_OPENCL
sort_config.op = CUDPP_ADD;
sort_config.datatype = CUDPP_UINT;
@ -56,28 +55,20 @@ int PairGPUAtomT::bytes_per_atom() const {
int id_space=0;
if (_gpu_nbor)
id_space=2;
int bytes=4*sizeof(numtyp)+11*sizeof(acctyp)+id_space;
int bytes=4*sizeof(numtyp)+id_space;
if (_rot)
bytes+=4*sizeof(numtyp)+4*sizeof(acctyp);
bytes+=4*sizeof(numtyp);
if (_charge)
bytes+=sizeof(numtyp);
return bytes;
}
template <class numtyp, class acctyp>
bool PairGPUAtomT::alloc(const int inum, const int nall) {
bool PairGPUAtomT::alloc(const int nall) {
_max_atoms=static_cast<int>(static_cast<double>(nall)*1.10);
if (_newton)
_max_local=_max_atoms;
else
_max_local=static_cast<int>(static_cast<double>(inum)*1.10);
bool success=true;
int ans_elements=4;
if (_rot)
ans_elements+=4;
// Ignore host/device transfers?
bool cpuview=false;
if (dev->device_type()==UCL_CPU)
@ -107,8 +98,6 @@ bool PairGPUAtomT::alloc(const int inum, const int nall) {
success=success && (host_x.alloc(_max_atoms*4,*dev,
UCL_WRITE_OPTIMIZED)==UCL_SUCCESS);
#endif
success=success &&(host_ans.alloc(ans_elements*_max_local,*dev)==UCL_SUCCESS);
success=success &&(host_engv.alloc(_ev_fields*_max_local,*dev)==UCL_SUCCESS);
// Buffer for casting only if different precisions
if (_charge)
success=success && (host_q.alloc(_max_atoms,*dev,
@ -120,15 +109,13 @@ bool PairGPUAtomT::alloc(const int inum, const int nall) {
// --------------------------- Device allocations
_gpu_bytes=0;
int gpu_bytes=0;
if (cpuview) {
#ifdef GPU_CAST
assert(0==1);
#else
dev_x.view(host_x);
#endif
dev_engv.view(host_engv);
dev_ans.view(host_ans);
if (_rot)
dev_quat.view(host_quat);
if (_charge)
@ -140,49 +127,80 @@ bool PairGPUAtomT::alloc(const int inum, const int nall) {
dev_x_cast.alloc(_max_atoms*3,*dev,UCL_READ_ONLY));
success=success && (UCL_SUCCESS==
dev_type_cast.alloc(_max_atoms,*dev,UCL_READ_ONLY));
_gpu_bytes+=dev_x_cast.row_bytes()+dev_type_cast.row_bytes();
gpu_bytes+=dev_x_cast.row_bytes()+dev_type_cast.row_bytes();
#else
success=success && (UCL_SUCCESS==
dev_x.alloc(_max_atoms*4,*dev,UCL_READ_ONLY));
#endif
success=success && (dev_engv.alloc(_ev_fields*_max_local,*dev,
UCL_WRITE_ONLY)==UCL_SUCCESS);
success=success && (dev_ans.alloc(ans_elements*_max_local,
*dev,UCL_WRITE_ONLY)==UCL_SUCCESS);
if (_charge) {
success=success && (dev_q.alloc(_max_atoms,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
_gpu_bytes+=dev_q.row_bytes();
gpu_bytes+=dev_q.row_bytes();
}
if (_rot) {
success=success && (dev_quat.alloc(_max_atoms*4,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
_gpu_bytes+=dev_quat.row_bytes();
gpu_bytes+=dev_quat.row_bytes();
}
}
if (_gpu_nbor) {
success=success && (dev_cell_id.alloc(_max_atoms,*dev)==UCL_SUCCESS);
success=success && (dev_particle_id.alloc(_max_atoms,*dev)==UCL_SUCCESS);
_gpu_bytes+=dev_cell_id.row_bytes()+dev_particle_id.row_bytes();
gpu_bytes+=dev_cell_id.row_bytes()+dev_particle_id.row_bytes();
if (_bonds) {
success=success && (dev_tag.alloc(_max_atoms,*dev)==UCL_SUCCESS);
_gpu_bytes+=dev_tag.row_bytes();
gpu_bytes+=dev_tag.row_bytes();
}
}
_gpu_bytes+=dev_x.row_bytes()+dev_engv.row_bytes()+dev_ans.row_bytes();
gpu_bytes+=dev_x.row_bytes();
if (gpu_bytes>_max_gpu_bytes)
_max_gpu_bytes=gpu_bytes;
_allocated=true;
return success;
}
template <class numtyp, class acctyp>
bool PairGPUAtomT::init(const int inum, const int nall, const bool charge,
const bool rot, UCL_Device &devi, const bool gpu_nbor,
bool PairGPUAtomT::add_fields(const bool charge, const bool rot,
const bool gpu_nbor, const bool bonds) {
bool realloc=false;
if (charge && _charge==false) {
_charge=true;
realloc=true;
}
if (rot && _rot==false) {
_rot=true;
realloc=true;
}
if (gpu_nbor && _gpu_nbor==false) {
_gpu_nbor=true;
realloc=true;
}
if (bonds && _bonds==false) {
_bonds=true;
realloc=true;
}
if (realloc) {
_other=_charge || _rot;
int max_atoms=_max_atoms;
clear_resize();
return alloc(max_atoms);
}
return true;
}
template <class numtyp, class acctyp>
bool PairGPUAtomT::init(const int nall, const bool charge, const bool rot,
UCL_Device &devi, const bool gpu_nbor,
const bool bonds) {
clear();
bool success=true;
_x_avail=false;
_q_avail=false;
_quat_avail=false;
_resized=false;
_gpu_nbor=gpu_nbor;
_bonds=bonds;
_charge=charge;
@ -190,33 +208,25 @@ bool PairGPUAtomT::init(const int inum, const int nall, const bool charge,
_other=_charge || _rot;
dev=&devi;
_e_fields=1;
if (_charge)
_e_fields++;
_ev_fields=6+_e_fields;
// Initialize atom and nbor data
int ef_inum=inum;
if (ef_inum==0)
ef_inum=1000;
int ef_nall=nall;
if (ef_nall<=ef_inum)
ef_nall=ef_inum*2;
if (ef_nall==0)
ef_nall=2000;
// Initialize timers for the selected device
time_pos.init(*dev);
time_other.init(*dev);
time_answer.init(*dev);
time_q.init(*dev);
time_quat.init(*dev);
time_pos.zero();
time_other.zero();
time_answer.zero();
time_q.zero();
time_quat.zero();
_time_cast=0.0;
#ifdef GPU_CAST
compile_kernels(*dev);
#endif
return success && alloc(ef_inum,ef_nall);
return success && alloc(ef_nall);
}
template <class numtyp, class acctyp>
@ -234,16 +244,12 @@ void PairGPUAtomT::clear_resize() {
dev_quat.clear();
host_quat.clear();
}
dev_ans.clear();
dev_engv.clear();
#ifndef GPU_CAST
host_x.clear();
#else
host_x_cast.clear();
host_type_cast.clear();
#endif
host_ans.clear();
host_engv.clear();
dev_cell_id.clear();
dev_particle_id.clear();
dev_tag.clear();
@ -261,17 +267,14 @@ void PairGPUAtomT::clear_resize() {
template <class numtyp, class acctyp>
void PairGPUAtomT::clear() {
_gpu_bytes=0;
_max_gpu_bytes=0;
if (!_allocated)
return;
time_pos.clear();
time_other.clear();
time_answer.clear();
time_q.clear();
time_quat.clear();
clear_resize();
_inum=0;
_eflag=false;
_vflag=false;
#ifdef GPU_CAST
if (_compiled) {
@ -289,255 +292,10 @@ double PairGPUAtomT::host_memory_usage() const {
atom_bytes+=1;
if (_rot)
atom_bytes+=4;
int ans_bytes=atom_bytes+_ev_fields;
return _max_atoms*atom_bytes*sizeof(numtyp)+
ans_bytes*(_max_local)*sizeof(acctyp)+
sizeof(PairGPUAtom<numtyp,acctyp>);
}
template <class numtyp, class acctyp>
void PairGPUAtomT::copy_answers(const bool eflag, const bool vflag,
const bool ef_atom, const bool vf_atom) {
time_answer.start();
_eflag=eflag;
_vflag=vflag;
_ef_atom=ef_atom;
_vf_atom=vf_atom;
int csize=_ev_fields;
if (!eflag)
csize-=_e_fields;
if (!vflag)
csize-=6;
if (csize>0)
ucl_copy(host_engv,dev_engv,_inum*csize,true);
if (_rot)
ucl_copy(host_ans,dev_ans,_inum*4*2,true);
else
ucl_copy(host_ans,dev_ans,_inum*4,true);
time_answer.stop();
}
template <class numtyp, class acctyp>
void PairGPUAtomT::copy_answers(const bool eflag, const bool vflag,
const bool ef_atom, const bool vf_atom,
int *ilist) {
_ilist=ilist;
copy_answers(eflag,vflag,ef_atom,vf_atom);
}
template <class numtyp, class acctyp>
double PairGPUAtomT::energy_virial(double *eatom, double **vatom,
double *virial) {
if (_eflag==false && _vflag==false)
return 0.0;
double evdwl=0.0;
if (_gpu_nbor) {
for (int i=0; i<_inum; i++) {
acctyp *ap=host_engv.begin()+i;
if (_eflag) {
if (_ef_atom) {
evdwl+=*ap;
eatom[i]+=*ap*0.5;
ap+=_inum;
} else {
evdwl+=*ap;
ap+=_inum;
}
}
if (_vflag) {
if (_vf_atom) {
for (int j=0; j<6; j++) {
vatom[i][j]+=*ap*0.5;
virial[j]+=*ap;
ap+=_inum;
}
} else {
for (int j=0; j<6; j++) {
virial[j]+=*ap;
ap+=_inum;
}
}
}
}
for (int j=0; j<6; j++)
virial[j]*=0.5;
} else {
for (int i=0; i<_inum; i++) {
acctyp *ap=host_engv.begin()+i;
int ii=_ilist[i];
if (_eflag) {
if (_ef_atom) {
evdwl+=*ap;
eatom[ii]+=*ap*0.5;
ap+=_inum;
} else {
evdwl+=*ap;
ap+=_inum;
}
}
if (_vflag) {
if (_vf_atom) {
for (int j=0; j<6; j++) {
vatom[ii][j]+=*ap*0.5;
virial[j]+=*ap;
ap+=_inum;
}
} else {
for (int j=0; j<6; j++) {
virial[j]+=*ap;
ap+=_inum;
}
}
}
}
for (int j=0; j<6; j++)
virial[j]*=0.5;
}
evdwl*=0.5;
return evdwl;
}
template <class numtyp, class acctyp>
double PairGPUAtomT::energy_virial(double *eatom, double **vatom,
double *virial, double &ecoul) {
if (_eflag==false && _vflag==false) {
ecoul=0.0;
return 0.0;
}
if (_charge==false)
return energy_virial(eatom,vatom,virial);
double evdwl=0.0;
double _ecoul=0.0;
if (_gpu_nbor) {
for (int i=0; i<_inum; i++) {
acctyp *ap=host_engv.begin()+i;
if (_eflag) {
if (_ef_atom) {
evdwl+=*ap;
eatom[i]+=*ap*0.5;
ap+=_inum;
_ecoul+=*ap;
eatom[i]+=*ap*0.5;
ap+=_inum;
} else {
evdwl+=*ap;
ap+=_inum;
_ecoul+=*ap;
ap+=_inum;
}
}
if (_vflag) {
if (_vf_atom) {
for (int j=0; j<6; j++) {
vatom[i][j]+=*ap*0.5;
virial[j]+=*ap;
ap+=_inum;
}
} else {
for (int j=0; j<6; j++) {
virial[j]+=*ap;
ap+=_inum;
}
}
}
}
for (int j=0; j<6; j++)
virial[j]*=0.5;
} else {
for (int i=0; i<_inum; i++) {
acctyp *ap=host_engv.begin()+i;
int ii=_ilist[i];
if (_eflag) {
if (_ef_atom) {
evdwl+=*ap;
eatom[ii]+=*ap*0.5;
ap+=_inum;
_ecoul+=*ap;
eatom[ii]+=*ap*0.5;
ap+=_inum;
} else {
evdwl+=*ap;
ap+=_inum;
_ecoul+=*ap;
ap+=_inum;
}
}
if (_vflag) {
if (_vf_atom) {
for (int j=0; j<6; j++) {
vatom[ii][j]+=*ap*0.5;
virial[j]+=*ap;
ap+=_inum;
}
} else {
for (int j=0; j<6; j++) {
virial[j]+=*ap;
ap+=_inum;
}
}
}
}
for (int j=0; j<6; j++)
virial[j]*=0.5;
}
evdwl*=0.5;
ecoul+=_ecoul*0.5;
return evdwl;
}
template <class numtyp, class acctyp>
void PairGPUAtomT::get_answers(double **f, double **tor) {
acctyp *ap=host_ans.begin();
if (_gpu_nbor) {
for (int i=0; i<_inum; i++) {
f[i][0]+=*ap;
ap++;
f[i][1]+=*ap;
ap++;
f[i][2]+=*ap;
ap+=2;
}
if (_rot) {
for (int i=0; i<_inum; i++) {
tor[i][0]+=*ap;
ap++;
tor[i][1]+=*ap;
ap++;
tor[i][2]+=*ap;
ap+=2;
}
}
} else {
for (int i=0; i<_inum; i++) {
int ii=_ilist[i];
f[ii][0]+=*ap;
ap++;
f[ii][1]+=*ap;
ap++;
f[ii][2]+=*ap;
ap+=2;
}
if (_rot) {
for (int i=0; i<_inum; i++) {
int ii=_ilist[i];
tor[ii][0]+=*ap;
ap++;
tor[ii][1]+=*ap;
ap++;
tor[ii][2]+=*ap;
ap+=2;
}
}
}
}
// Sort arrays for neighbor list calculation
template <class numtyp, class acctyp>
void PairGPUAtomT::sort_neighbor(const int num_atoms) {