Homec4science
Diffusion lammps (efficient_neuronet) Browse

lammps/src/USER-INTEL/intel_buffers.cpp
b50f71cd96d6efficient_neuronet

Last Change
Raw File

intel_buffers.cpp
View Options

/* ----------------------------------------------------------------------
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
This software is distributed under the GNU General Public License.
See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */
/* ----------------------------------------------------------------------
Contributing author: W. Michael Brown (Intel)
------------------------------------------------------------------------- */
#include "intel_buffers.h"
#include "force.h"
#include "memory.h"
using namespace LAMMPS_NS;
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
IntelBuffers<flt_t, acc_t>::IntelBuffers(class LAMMPS *lmp_in) :
lmp(lmp_in), _x(0), _q(0), _quat(0), _f(0), _off_threads(0),
_buf_size(0), _buf_local_size(0) {
_list_alloc_atoms = 0;
_ntypes = 0;
_off_map_maxlocal = 0;
_ccachex = 0;
_host_nmax = 0;
#ifdef _LMP_INTEL_OFFLOAD
_separate_buffers = 0;
_off_f = 0;
_off_map_ilist = 0;
_off_map_nmax = 0;
_off_map_maxhead = 0;
_off_list_alloc = false;
_off_threads = 0;
_off_ccache = 0;
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
IntelBuffers<flt_t, acc_t>::~IntelBuffers()
{
free_buffers();
free_all_nbor_buffers();
free_ccache();
set_ntypes(0);
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_buffers()
{
if (_buf_size > 0) {
atom_t * x = get_x();
flt_t * q = get_q();
quat_t * quat = get_quat();
#ifdef _LMP_INTEL_OFFLOAD
vec3_acc_t * f_start = get_off_f();
if (f_start != 0) {
acc_t * ev_global = get_ev_global();
if (ev_global != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(x:alloc_if(0) free_if(1)) \
nocopy(f_start:alloc_if(0) free_if(1)) \
nocopy(ev_global:alloc_if(0) free_if(1))
}
if (q != 0) {
#pragma offload_transfer target (mic:_cop) \
nocopy(q:alloc_if(0) free_if(1))
}
if (quat != 0) {
#pragma offload_transfer target (mic:_cop) \
nocopy(quat:alloc_if(0) free_if(1))
}
lmp->memory->destroy(f_start);
}
if (_separate_buffers) {
lmp->memory->destroy(_host_x);
if (q != 0) lmp->memory->destroy(_host_q);
if (quat != 0) lmp->memory->destroy(_host_quat);
}
#endif
lmp->memory->destroy(x);
if (q != 0) lmp->memory->destroy(q);
if (quat != 0) lmp->memory->destroy(quat);
lmp->memory->destroy(_f);
_buf_size = _buf_local_size = 0;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow(const int nall, const int nlocal,
const int nthreads,
const int offload_end)
{
free_buffers();
_buf_size = static_cast<double>(nall) * 1.1 + 1;
if (lmp->force->newton_pair)
_buf_local_size = _buf_size;
else
_buf_local_size = static_cast<double>(nlocal) * 1.1 + 1;
if (lmp->atom->torque)
_buf_local_size *= 2;
const int f_stride = get_stride(_buf_local_size);
lmp->memory->create(_x, _buf_size,"intel_x");
if (lmp->atom->q != NULL)
lmp->memory->create(_q, _buf_size, "intel_q");
if (lmp->atom->ellipsoid != NULL)
lmp->memory->create(_quat, _buf_size, "intel_quat");
lmp->memory->create(_f, f_stride * nthreads, "intel_f");
#ifdef _LMP_INTEL_OFFLOAD
if (_separate_buffers) {
lmp->memory->create(_host_x, _buf_size,"intel_host_x");
if (lmp->atom->q != NULL)
lmp->memory->create(_host_q, _buf_size, "intel_host_q");
if (lmp->atom->ellipsoid != NULL)
lmp->memory->create(_host_quat, _buf_size, "intel_host_quat");
}
if (offload_end > 0) {
lmp->memory->create(_off_f, f_stride * _off_threads, "intel_off_f");
const atom_t * const x = get_x();
const flt_t * const q = get_q();
const vec3_acc_t * f_start = get_off_f();
acc_t * ev_global = get_ev_global();
if (lmp->atom->q != NULL) {
if (x != NULL && q != NULL && f_start != NULL && ev_global != NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(x,q:length(_buf_size) alloc_if(1) free_if(0)) \
nocopy(f_start:length(f_stride*_off_threads) alloc_if(1) free_if(0))\
nocopy(ev_global:length(8) alloc_if(1) free_if(0))
}
} else {
if (x != NULL && f_start != NULL && ev_global != NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(x:length(_buf_size) alloc_if(1) free_if(0)) \
nocopy(f_start:length(f_stride*_off_threads) alloc_if(1) free_if(0))\
nocopy(ev_global:length(8) alloc_if(1) free_if(0))
}
}
if (lmp->atom->ellipsoid != NULL) {
const quat_t * const quat = get_quat();
if (quat != NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(quat:length(_buf_size) alloc_if(1) free_if(0))
}
}
}
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_nmax()
{
#ifdef _LMP_INTEL_OFFLOAD
if (_off_map_nmax > 0) {
const int * tag = _off_map_tag;
const int * special = _off_map_special;
const int * nspecial = _off_map_nspecial;
const int * bins = _off_map_bins;
const int * binpacked = _binpacked;
if (tag != 0 && special != 0 && nspecial !=0 && bins != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(tag:alloc_if(0) free_if(1)) \
nocopy(special,nspecial:alloc_if(0) free_if(1)) \
nocopy(bins,binpacked:alloc_if(0) free_if(1))
}
_off_map_nmax = 0;
}
#endif
if (_host_nmax > 0) {
lmp->memory->destroy(_binpacked);
_host_nmax = 0;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow_nmax(const int offload_end)
{
free_nmax();
int size = lmp->atom->nmax;
_host_nmax = size;
lmp->memory->create(_binpacked, _host_nmax, "_binpacked");
#ifdef _LMP_INTEL_OFFLOAD
if (!offload_end) return;
int *special, *nspecial;
int tag_length, special_length, nspecial_length;
if (lmp->atom->molecular) {
special = lmp->atom->special[0];
nspecial = lmp->atom->nspecial[0];
special_length = size * lmp->atom->maxspecial;
nspecial_length = size * 3;
} else {
special = &_special_holder;
nspecial = &_nspecial_holder;
special_length = 1;
nspecial_length = 1;
}
if (_need_tag)
tag_length = size;
else
tag_length = 1;
int *tag = lmp->atom->tag;
int *bins = lmp->neighbor->bins;
int * binpacked = _binpacked;
#pragma offload_transfer target(mic:_cop) \
nocopy(bins,binpacked:length(size) alloc_if(1) free_if(0)) \
nocopy(tag:length(tag_length) alloc_if(1) free_if(0)) \
nocopy(special:length(special_length) alloc_if(1) free_if(0)) \
nocopy(nspecial:length(nspecial_length) alloc_if(1) free_if(0))
_off_map_tag = tag;
_off_map_special = special;
_off_map_nspecial = nspecial;
_off_map_nmax = size;
_off_map_bins = bins;
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_local()
{
if (_off_map_maxlocal > 0) {
int * cnumneigh = _cnumneigh;
int * atombin = _atombin;
#ifdef _LMP_INTEL_OFFLOAD
if (_off_map_ilist != NULL) {
const int * ilist = _off_map_ilist;
const int * numneigh = _off_map_numneigh;
_off_map_ilist = NULL;
if (numneigh != 0 && ilist != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(ilist,numneigh,cnumneigh,atombin:alloc_if(0) free_if(1))
}
}
#endif
lmp->memory->destroy(cnumneigh);
lmp->memory->destroy(atombin);
_off_map_maxlocal = 0;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow_local(NeighList *list,
const int offload_end)
{
free_local();
int size = list->get_maxlocal();
lmp->memory->create(_cnumneigh, size, "_cnumneigh");
lmp->memory->create(_atombin, size, "_atombin");
_off_map_maxlocal = size;
#ifdef _LMP_INTEL_OFFLOAD
if (offload_end > 0) {
int * numneigh = list->numneigh;
int * ilist = list->ilist;
int * cnumneigh = _cnumneigh;
int * atombin = _atombin;
if (cnumneigh != 0 && atombin != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(ilist:length(size) alloc_if(1) free_if(0)) \
nocopy(numneigh:length(size) alloc_if(1) free_if(0)) \
nocopy(cnumneigh:length(size) alloc_if(1) free_if(0)) \
nocopy(atombin:length(size) alloc_if(1) free_if(0))
}
_off_map_ilist = ilist;
_off_map_numneigh = numneigh;
}
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_binhead()
{
#ifdef _LMP_INTEL_OFFLOAD
if (_off_map_maxhead > 0) {
const int * binhead = _off_map_binhead;
if (binhead !=0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(binhead:alloc_if(0) free_if(1))
}
_off_map_maxhead = 0;
}
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow_binhead()
{
#ifdef _LMP_INTEL_OFFLOAD
free_binhead();
int * binhead = lmp->neighbor->binhead;
const int maxhead = lmp->neighbor->maxhead;
#pragma offload_transfer target(mic:_cop) \
nocopy(binhead:length(maxhead+1) alloc_if(1) free_if(0))
_off_map_binhead = binhead;
_off_map_maxhead = maxhead;
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_nbor_list()
{
if (_list_alloc_atoms > 0) {
#ifdef _LMP_INTEL_OFFLOAD
if (_off_list_alloc) {
int * list_alloc = _list_alloc;
int * stencil = _off_map_stencil;
if (list_alloc != 0 && stencil != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(list_alloc:alloc_if(0) free_if(1))
}
_off_list_alloc = false;
}
#endif
lmp->memory->destroy(_list_alloc);
_list_alloc_atoms = 0;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow_nbor_list(NeighList *list,
const int nlocal,
const int nthreads,
const int offload_end,
const int pack_width)
{
free_nbor_list();
_list_alloc_atoms = 1.10 * nlocal;
int nt = MAX(nthreads, _off_threads);
int list_alloc_size = (_list_alloc_atoms + nt * 2 + pack_width - 1) *
get_max_nbors();
lmp->memory->create(_list_alloc, list_alloc_size, "_list_alloc");
#ifdef _LMP_INTEL_OFFLOAD
if (offload_end > 0) {
int * list_alloc =_list_alloc;
int * stencil = list->stencil;
if (list_alloc != NULL) {
#pragma offload_transfer target(mic:_cop) \
in(stencil:length(list->maxstencil) alloc_if(1) free_if(0)) \
nocopy(list_alloc:length(list_alloc_size) alloc_if(1) free_if(0))
_off_map_stencil = stencil;
_off_list_alloc = true;
}
}
#endif
}
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::_grow_stencil(NeighList *list)
{
#ifdef _LMP_INTEL_OFFLOAD
int * stencil = _off_map_stencil;
#pragma offload_transfer target(mic:_cop) \
nocopy(stencil:alloc_if(0) free_if(1))
stencil = list->stencil;
#pragma offload_transfer target(mic:_cop) \
in(stencil:length(list->maxstencil) alloc_if(1) free_if(0))
_off_map_stencil = stencil;
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::free_ccache()
{
if (_ccachex) {
flt_t *ccachex = _ccachex;
flt_t *ccachey = _ccachey;
flt_t *ccachez = _ccachez;
flt_t *ccachew = _ccachew;
int *ccachei = _ccachei;
int *ccachej = _ccachej;
#ifdef LMP_USE_AVXCD
acc_t *ccachef = _ccachef;
#endif
#ifdef _LMP_INTEL_OFFLOAD
if (_off_ccache) {
#pragma offload_transfer target(mic:_cop) \
nocopy(ccachex,ccachey,ccachez,ccachew:alloc_if(0) free_if(1)) \
nocopy(ccachei,ccachej:alloc_if(0) free_if(1))
#ifdef LMP_USE_AVXCD
#pragma offload_transfer target(mic:_cop) \
nocopy(ccachef:alloc_if(0) free_if(1))
#endif
}
_off_ccache = 0;
#endif
lmp->memory->destroy(ccachex);
lmp->memory->destroy(ccachey);
lmp->memory->destroy(ccachez);
lmp->memory->destroy(ccachew);
lmp->memory->destroy(ccachei);
lmp->memory->destroy(ccachej);
#ifdef LMP_USE_AVXCD
lmp->memory->destroy(ccachef);
#endif
_ccachex = 0;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::grow_ccache(const int off_flag,
const int nthreads,
const int width)
{
#ifdef _LMP_INTEL_OFFLOAD
if (_ccachex && off_flag && _off_ccache == 0)
free_ccache();
#endif
if (_ccachex)
return;
const int nsize = get_max_nbors() * width;
int esize = MIN(sizeof(int), sizeof(flt_t));
IP_PRE_get_stride(_ccache_stride, nsize, esize, 0);
int nt = MAX(nthreads, _off_threads);
const int vsize = _ccache_stride * nt;
lmp->memory->create(_ccachex, vsize , "_ccachex");
lmp->memory->create(_ccachey, vsize, "_ccachey");
lmp->memory->create(_ccachez, vsize, "_ccachez");
lmp->memory->create(_ccachew, vsize, "_ccachew");
lmp->memory->create(_ccachei, vsize, "_ccachei");
lmp->memory->create(_ccachej, vsize, "_ccachej");
#ifdef LMP_USE_AVXCD
IP_PRE_get_stride(_ccache_stride3, nsize * 3, sizeof(acc_t), 0);
lmp->memory->create(_ccachef, _ccache_stride3 * nt, "_ccachef");
#endif
memset(_ccachej, 0, vsize * sizeof(int));
#ifdef _LMP_INTEL_OFFLOAD
if (off_flag) {
flt_t *ccachex = _ccachex;
flt_t *ccachey = _ccachey;
flt_t *ccachez = _ccachez;
flt_t *ccachew = _ccachew;
int *ccachei = _ccachei;
int *ccachej = _ccachej;
if (ccachex != NULL && ccachey !=NULL && ccachez != NULL &&
ccachew != NULL && ccachei != NULL && ccachej !=NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(ccachex,ccachey:length(vsize) alloc_if(1) free_if(0)) \
nocopy(ccachez,ccachew:length(vsize) alloc_if(1) free_if(0)) \
nocopy(ccachei:length(vsize) alloc_if(1) free_if(0)) \
in(ccachej:length(vsize) alloc_if(1) free_if(0))
}
#ifdef LMP_USE_AVXCD
if (ccachef != NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(ccachef:length(_ccache_stride3 * nt) alloc_if(1) free_if(0))
}
#endif
_off_ccache = 1;
}
#endif
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
void IntelBuffers<flt_t, acc_t>::set_ntypes(const int ntypes)
{
if (ntypes != _ntypes) {
if (_ntypes > 0) {
#ifdef _LMP_INTEL_OFFLOAD
flt_t * cutneighsqo = _cutneighsq[0];
if (_off_threads > 0 && cutneighsqo != 0) {
#pragma offload_transfer target(mic:_cop) \
nocopy(cutneighsqo:alloc_if(0) free_if(1))
}
#endif
lmp->memory->destroy(_cutneighsq);
}
if (ntypes > 0) {
lmp->memory->create(_cutneighsq, ntypes, ntypes, "_cutneighsq");
#ifdef _LMP_INTEL_OFFLOAD
flt_t * cutneighsqo = _cutneighsq[0];
if (_off_threads > 0 && cutneighsqo != NULL) {
#pragma offload_transfer target(mic:_cop) \
nocopy(cutneighsqo:length(ntypes * ntypes) alloc_if(1) free_if(0))
}
#endif
}
_ntypes = ntypes;
}
}
/* ---------------------------------------------------------------------- */
template <class flt_t, class acc_t>
double IntelBuffers<flt_t, acc_t>::memory_usage(const int nthreads)
{
double tmem = sizeof(atom_t);
if (lmp->atom->q) tmem += sizeof(flt_t);
if (lmp->atom->torque) tmem += sizeof(quat_t);
#ifdef _LMP_INTEL_OFFLOAD
if (_separate_buffers) tmem *= 2;
#endif
tmem *= _buf_size;
const int fstride = get_stride(_buf_local_size);
tmem += fstride * nthreads * sizeof(vec3_acc_t);
#ifdef _LMP_INTEL_OFFLOAD
if (_off_f) tmem += fstride*_off_threads * sizeof(vec3_acc_t);
#endif
tmem += _off_map_maxlocal * sizeof(int) * 2;
tmem += (_list_alloc_atoms + _off_threads) * get_max_nbors() * sizeof(int);
tmem += _ntypes * _ntypes * sizeof(int);
return tmem;
}
/* ---------------------------------------------------------------------- */
template class IntelBuffers<float,float>;
template class IntelBuffers<float,double>;
template class IntelBuffers<double,double>;
c4science · Help