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

Refactored 3-body gpu styles to remove code duplication

Browse Source
This commit is contained in:
Trung Nguyen
2017-07-08 20:17:31 -05:00
parent 34fe2273f6
commit ea2b01e83b
8 changed files with 28 additions and 531 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/gpu/lal_base_three.cpp
View File

@ -180,6 +180,8 @@ int * BaseThreeT::reset_nbors(const int nall, const int inum, const int nlist,
if (!success)
return NULL;
_nall = nall;
// originally the requirement that nall == nlist was enforced
// to allow direct indexing neighbors of neighbors after re-arrangement
// nbor->get_host3(nall,nlist,ilist,numj,firstneigh,block_size());
@ -214,6 +216,8 @@ inline int BaseThreeT::build_nbor_list(const int inum, const int host_inum,
return 0;
atom->cast_copy_x(host_x,host_type);
_nall = nall;
int mn;
nbor->build_nbor_list(host_x, nall, host_inum, nall, *atom, sublo, subhi, tag,
nspecial, special, success, mn);

6
lib/gpu/lal_base_three.h
View File

@ -74,7 +74,7 @@ class BaseThree {
}
/// Check if there is enough storage for neighbors and realloc if not
/** \param nlocal number of particles whose nbors must be stored on device
/** \param inum number of particles whose nbors must be stored on device
* \param max_nbors maximum number of neighbors
* \param success set to false if insufficient memory
* \note olist_size=total number of local particles **/
@ -83,7 +83,7 @@ class BaseThree {
}
/// Check if there is enough storage for neighbors and realloc if not
/** \param nlocal number of particles whose nbors must be stored on device
/** \param inum number of particles whose nbors must be stored on device
* \param host_inum number of particles whose nbors need to copied to host
* \param max_nbors current maximum number of neighbors
* \note host_inum is 0 if the host is performing neighboring
@ -203,7 +203,7 @@ class BaseThree {
int _block_pair, _block_size, _threads_per_atom, _end_command_queue;
int _gpu_nbor;
double _max_bytes, _max_an_bytes;
int _max_nbors, _ainum;
int _max_nbors, _ainum, _nall;
double _gpu_overhead, _driver_overhead;
UCL_D_Vec<int> *_nbor_data;

152
lib/gpu/lal_tersoff.cpp
View File

@ -225,151 +225,6 @@ double TersoffT::host_memory_usage() const {
#define KTHREADS this->_threads_per_atom
#define JTHREADS this->_threads_per_atom
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void TersoffT::compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return;
}
int ago=this->hd_balancer.ago_first(f_ago);
int inum=this->hd_balancer.balance(ago,inum_full,cpu_time);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
if (ago==0) {
this->reset_nbors(nall, inum, nlist, ilist, numj, firstneigh, success);
if (!success)
return;
this->_max_nbors = this->nbor->max_nbor_loop(nlist,numj,ilist);
}
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nlist;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
}
// ---------------------------------------------------------------------------
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int ** TersoffT::compute(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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return NULL;
}
this->hd_balancer.balance(cpu_time);
int inum=this->hd_balancer.get_gpu_count(ago,inum_full);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
// Build neighbor list on GPU if necessary
if (ago==0) {
this->_max_nbors = this->build_nbor_list(inum, inum_full-inum, nall, host_x, host_type,
sublo, subhi, tag, nspecial, special, success);
if (!success)
return NULL;
this->hd_balancer.start_timer();
} else {
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
}
*ilist=this->nbor->host_ilist.begin();
*jnum=this->nbor->host_acc.begin();
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nall;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
return this->nbor->host_jlist.begin()-host_start;
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
@ -400,6 +255,13 @@ void TersoffT::loop(const bool _eflag, const bool _vflag, const int evatom) {
&this->dev_short_nbor, &ainum,
&nbor_pitch, &this->_threads_per_atom);
// re-allocate zetaij if necessary
int nall = this->_nall;
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
nbor_pitch=this->nbor->nbor_pitch();
GX=static_cast<int>(ceil(static_cast<double>(this->_ainum)/
(BX/(JTHREADS*KTHREADS))));

15
lib/gpu/lal_tersoff.h
View File

@ -47,21 +47,6 @@ class Tersoff : public BaseThree<numtyp, acctyp> {
const double* h, const double* gamma, const double* beta,
const double* powern, const double* cutsq);
/// Pair loop with host neighboring
void compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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);
/// Pair loop with device neighboring
int ** compute(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 **numj, const double cpu_time, bool &success);
/// Clear all host and device data
/** \note This is called at the beginning of the init() routine **/
void clear();

176
lib/gpu/lal_tersoff_mod.cpp
View File

@ -225,175 +225,6 @@ double TersoffMT::host_memory_usage() const {
#define KTHREADS this->_threads_per_atom
#define JTHREADS this->_threads_per_atom
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void TersoffMT::compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return;
}
int ago=this->hd_balancer.ago_first(f_ago);
int inum=this->hd_balancer.balance(ago,inum_full,cpu_time);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
if (ago==0) {
this->reset_nbors(nall, inum, nlist, ilist, numj, firstneigh, success);
if (!success)
return;
this->_max_nbors = this->nbor->max_nbor_loop(nlist,numj,ilist);
}
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nlist;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int ainum=nlist;
int nbor_pitch=this->nbor->nbor_pitch();
int BX=this->block_pair();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/
(BX/(JTHREADS*KTHREADS))));
this->k_zeta.set_size(GX,BX);
this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &cutsq,
&map, &elem2param, &_nelements, &_nparams, &_zetaij,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&_eflag, &ainum, &nbor_pitch, &this->_threads_per_atom);
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
}
// ---------------------------------------------------------------------------
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int ** TersoffMT::compute(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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return NULL;
}
this->hd_balancer.balance(cpu_time);
int inum=this->hd_balancer.get_gpu_count(ago,inum_full);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
// Build neighbor list on GPU if necessary
if (ago==0) {
this->_max_nbors = this->build_nbor_list(inum, inum_full-inum, nall, host_x, host_type,
sublo, subhi, tag, nspecial, special, success);
if (!success)
return NULL;
this->hd_balancer.start_timer();
} else {
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
}
*ilist=this->nbor->host_ilist.begin();
*jnum=this->nbor->host_acc.begin();
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nall;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int ainum=nall;
int nbor_pitch=this->nbor->nbor_pitch();
int BX=this->block_pair();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/
(BX/(JTHREADS*KTHREADS))));
this->k_zeta.set_size(GX,BX);
this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &cutsq,
&map, &elem2param, &_nelements, &_nparams, &_zetaij,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&_eflag, &ainum, &nbor_pitch, &this->_threads_per_atom);
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
return this->nbor->host_jlist.begin()-host_start;
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
@ -424,6 +255,13 @@ void TersoffMT::loop(const bool _eflag, const bool _vflag, const int evatom) {
&this->dev_short_nbor, &ainum,
&nbor_pitch, &this->_threads_per_atom);
// re-allocate zetaij if necessary
int nall = this->_nall;
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
nbor_pitch=this->nbor->nbor_pitch();
GX=static_cast<int>(ceil(static_cast<double>(this->_ainum)/
(BX/(JTHREADS*KTHREADS))));

15
lib/gpu/lal_tersoff_mod.h
View File

@ -47,21 +47,6 @@ class TersoffMod : public BaseThree<numtyp, acctyp> {
const double* h, const double* beta, const double* powern,
const double* powern_del, const double* ca1, const double* cutsq);
/// Pair loop with host neighboring
void compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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);
/// Pair loop with device neighboring
int ** compute(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 **numj, const double cpu_time, bool &success);
/// Clear all host and device data
/** \note This is called at the beginning of the init() routine **/
void clear();

176
lib/gpu/lal_tersoff_zbl.cpp
View File

@ -250,175 +250,6 @@ double TersoffZT::host_memory_usage() const {
#define KTHREADS this->_threads_per_atom
#define JTHREADS this->_threads_per_atom
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void TersoffZT::compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return;
}
int ago=this->hd_balancer.ago_first(f_ago);
int inum=this->hd_balancer.balance(ago,inum_full,cpu_time);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
if (ago==0) {
this->reset_nbors(nall, inum, nlist, ilist, numj, firstneigh, success);
if (!success)
return;
this->_max_nbors = this->nbor->max_nbor_loop(nlist,numj,ilist);
}
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nlist;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int ainum=nlist;
int nbor_pitch=this->nbor->nbor_pitch();
int BX=this->block_pair();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/
(BX/(JTHREADS*KTHREADS))));
this->k_zeta.set_size(GX,BX);
this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &ts6, &cutsq,
&map, &elem2param, &_nelements, &_nparams, &_zetaij,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&_eflag, &ainum, &nbor_pitch, &this->_threads_per_atom);
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom,ilist);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
}
// ---------------------------------------------------------------------------
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int ** TersoffZT::compute(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) {
this->acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
this->resize_atom(0,nall,success);
this->zero_timers();
return NULL;
}
this->hd_balancer.balance(cpu_time);
int inum=this->hd_balancer.get_gpu_count(ago,inum_full);
this->ans->inum(inum);
#ifdef THREE_CONCURRENT
this->ans2->inum(inum);
#endif
host_start=inum;
// Build neighbor list on GPU if necessary
if (ago==0) {
this->_max_nbors = this->build_nbor_list(inum, inum_full-inum, nall, host_x, host_type,
sublo, subhi, tag, nspecial, special, success);
if (!success)
return NULL;
this->hd_balancer.start_timer();
} else {
this->atom->cast_x_data(host_x,host_type);
this->hd_balancer.start_timer();
this->atom->add_x_data(host_x,host_type);
}
*ilist=this->nbor->host_ilist.begin();
*jnum=this->nbor->host_acc.begin();
// re-allocate zetaij if necessary
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
this->_ainum=nall;
int _eflag;
if (eflag)
_eflag=1;
else
_eflag=0;
int ainum=nall;
int nbor_pitch=this->nbor->nbor_pitch();
int BX=this->block_pair();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/
(BX/(JTHREADS*KTHREADS))));
this->k_zeta.set_size(GX,BX);
this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &ts6, &cutsq,
&map, &elem2param, &_nelements, &_nparams, &_zetaij,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&_eflag, &ainum, &nbor_pitch, &this->_threads_per_atom);
int evatom=0;
if (eatom || vatom)
evatom=1;
#ifdef THREE_CONCURRENT
this->ucl_device->sync();
#endif
loop(eflag,vflag,evatom);
this->ans->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans);
#ifdef THREE_CONCURRENT
this->ans2->copy_answers(eflag,vflag,eatom,vatom);
this->device->add_ans_object(this->ans2);
#endif
this->hd_balancer.stop_timer();
return this->nbor->host_jlist.begin()-host_start;
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
@ -449,6 +280,13 @@ void TersoffZT::loop(const bool _eflag, const bool _vflag, const int evatom) {
&this->dev_short_nbor, &ainum,
&nbor_pitch, &this->_threads_per_atom);
// re-allocate zetaij if necessary
int nall = this->_nall;
if (nall*this->_max_nbors > _zetaij.cols()) {
int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
_zetaij.resize(this->_max_nbors*_nmax);
}
nbor_pitch=this->nbor->nbor_pitch();
GX=static_cast<int>(ceil(static_cast<double>(this->_ainum)/
(BX/(JTHREADS*KTHREADS))));

15
lib/gpu/lal_tersoff_zbl.h
View File

@ -49,21 +49,6 @@ class TersoffZBL : public BaseThree<numtyp, acctyp> {
const double* ZBLcut, const double* ZBLexpscale, const double global_e,
const double global_a_0, const double global_epsilon_0, const double* cutsq);
/// Pair loop with host neighboring
void compute(const int f_ago, const int inum_full, const int nall,
const int nlist, 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);
/// Pair loop with device neighboring
int ** compute(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 **numj, const double cpu_time, bool &success);
/// Clear all host and device data
/** \note This is called at the beginning of the init() routine **/
void clear();