tx258/lammps-gran-kokkos
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Adding API for the repulsion term to hippo/gpu

Browse Source
This commit is contained in:
Trung Nguyen
2021-09-28 23:42:04 -05:00
parent b95508125b
commit 17edd797a7
6 changed files with 104 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/gpu/lal_hippo.cpp
View File

@ -173,7 +173,7 @@ int** HippoT::compute_repulsion(const int ago, const int inum_full,
int &host_start, int **ilist, int **jnum,
const double cpu_time, bool &success,
const double aewald, const double off2_repulse,
double *host_q, double *boxlo, double *prd) {
double *host_q, double *boxlo, double *prd, void **tep_ptr) {
this->acc_timers();
int eflag, vflag;
if (eatom) eflag=2;
@ -210,6 +210,14 @@ int** HippoT::compute_repulsion(const int ago, const int inum_full,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
// ------------------- Resize _tep array ------------------------
if (inum_full>this->_max_tep_size) {
this->_max_tep_size=static_cast<int>(static_cast<double>(inum_full)*1.10);
this->_tep.resize(this->_max_tep_size*4);
}
*tep_ptr=this->_tep.host.begin();
this->_off2_repulse = off2_repulse;
this->_aewald = aewald;
const int red_blocks=repulsion(eflag,vflag);
@ -222,6 +230,10 @@ int** HippoT::compute_repulsion(const int ago, const int inum_full,
this->hd_balancer.stop_timer();
// copy tep from device to host
this->_tep.update_host(this->_max_tep_size*4,false);
return firstneigh; // nbor->host_jlist.begin()-host_start;
}
@ -257,7 +269,7 @@ int HippoT::repulsion(const int eflag, const int vflag) {
&coeff_amtype, &coeff_amclass, &sp_nonpolar,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->dev_short_nbor,
&this->ans->force, &this->ans->engv,
&this->ans->force, &this->ans->engv, &this->_tep,
&eflag, &vflag, &ainum, &_nall, &nbor_pitch,
&this->_threads_per_atom, &this->_aewald,
&this->_off2_repulse);

2
lib/gpu/lal_hippo.h
View File

@ -67,7 +67,7 @@ class Hippo : public BaseAmoeba<numtyp, acctyp> {
int &host_start, int **ilist, int **jnum,
const double cpu_time, bool &success,
const double aewald, const double off2_repulse,
double *host_q, double *boxlo, double *prd);
double *host_q, double *boxlo, double *prd, void** tep_ptr);
/// Compute dispersion real-space with device neighboring
int** compute_dispersion_real(const int ago, const int inum_full, const int nall,

17
lib/gpu/lal_hippo_ext.cpp
View File

@ -120,6 +120,23 @@ void hippo_gpu_clear() {
HIPPOMF.clear();
}
int** hippo_gpu_compute_repulsion(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
int *host_amtype, int *host_amgroup, double **host_rpole,
double *sublo, double *subhi, tagint *tag, int **nspecial,
tagint **special, int *nspecial15, tagint** special15,
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, const double aewald, const double off2,
double *host_q, double *boxlo, double *prd, void **tep_ptr) {
return HIPPOMF.compute_repulsion(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, sublo, subhi,
tag, nspecial, special, nspecial15, special15,
eflag, vflag, eatom, vatom, host_start, ilist, jnum,
cpu_time, success, aewald, off2, host_q, boxlo, prd, tep_ptr);
}
int** hippo_gpu_compute_dispersion_real(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
int *host_amtype, int *host_amgroup, double **host_rpole,

2
src/AMOEBA/pair_amoeba.h
View File

@ -343,7 +343,7 @@ class PairAmoeba : public Pair {
void hal();
void repulsion();
virtual void repulsion();
void damprep(double, double, double, double, double, double, double, double,
int, double, double, double *);

70
src/GPU/pair_hippo_gpu.cpp
View File

@ -66,6 +66,17 @@ int hippo_gpu_init(const int ntypes, const int max_amtype, const int max_amclass
const double polar_dscale, const double polar_uscale, int& tq_size);
void hippo_gpu_clear();
int** hippo_gpu_compute_repulsion(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
int *host_amtype, int *host_amgroup, double **host_rpole,
double *sublo, double *subhi, tagint *tag, int **nspecial,
tagint **special, int *nspecial15, tagint** special15,
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, const double aewald, const double off2,
double *host_q, double *boxlo, double *prd, void **tep_ptr);
int** hippo_gpu_compute_dispersion_real(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
int *host_amtype, int *host_amgroup, double **host_rpole,
@ -209,6 +220,65 @@ void PairHippoGPU::init_style()
/* ---------------------------------------------------------------------- */
void PairHippoGPU::repulsion()
{
if (!gpu_repulsion_ready) {
PairAmoeba::repulsion();
return;
}
int eflag=1, vflag=1;
int nall = atom->nlocal + atom->nghost;
int inum, host_start;
bool success = true;
int *ilist, *numneigh, **firstneigh;
double sublo[3],subhi[3];
if (domain->triclinic == 0) {
sublo[0] = domain->sublo[0];
sublo[1] = domain->sublo[1];
sublo[2] = domain->sublo[2];
subhi[0] = domain->subhi[0];
subhi[1] = domain->subhi[1];
subhi[2] = domain->subhi[2];
} else {
domain->bbox(domain->sublo_lamda,domain->subhi_lamda,sublo,subhi);
}
inum = atom->nlocal;
// select the correct cutoff for the term
choose(REPULSE);
// set the energy unit conversion factor for multipolar real-space calculation
firstneigh = hippo_gpu_compute_repulsion(neighbor->ago, inum, nall, atom->x,
atom->type, amtype, amgroup, rpole,
sublo, subhi, atom->tag,
atom->nspecial, atom->special,
atom->nspecial15, atom->special15,
eflag, vflag, eflag_atom, vflag_atom,
host_start, &ilist, &numneigh, cpu_time,
success, aewald, off2, atom->q,
domain->boxlo, domain->prd, &tq_pinned);
if (!success)
error->one(FLERR,"Insufficient memory on accelerator");
// reference to the tep array from GPU lib
if (tq_single) {
float *tq_ptr = (float *)tq_pinned;
compute_force_from_torque<float>(tq_ptr, frepulse, virrepulse);
} else {
double *tq_ptr = (double *)tq_pinned;
compute_force_from_torque<double>(tq_ptr, frepulse, virrepulse);
}
}
/* ---------------------------------------------------------------------- */
void PairHippoGPU::dispersion_real()
{
if (!gpu_dispersion_real_ready) {

1
src/GPU/pair_hippo_gpu.h
View File

@ -35,6 +35,7 @@ class PairHippoGPU : public PairAmoeba {
virtual void induce();
virtual void repulsion();
virtual void dispersion_real();
virtual void multipole_real();
virtual void udirect2b(double **, double **);