jianmu
/
lammps
forked from lijiext/lammps
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
lammps/lib/gpu/lal_neighbor.cpp

621 lines
19 KiB
C++
Raw Blame History

/***************************************************************************
neighbor.cpp
-------------------
W. Michael Brown (ORNL)
Peng Wang (Nvidia)
Class for handling neighbor lists
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : brownw@ornl.gov, penwang@nvidia.com
***************************************************************************/
#include "lal_precision.h"
#include "lal_neighbor.h"
#include "lal_device.h"
#include "math.h"
using namespace LAMMPS_AL;
int Neighbor::bytes_per_atom(const int max_nbors) const {
if (_gpu_nbor==1)
return (max_nbors+2)*sizeof(int);
else if (_gpu_nbor==2)
return (max_nbors+3)*sizeof(int);
else if (_use_packing)
return ((max_nbors+2)*2)*sizeof(int);
else
return (max_nbors+3)*sizeof(int);
}
bool Neighbor::init(NeighborShared *shared, const int inum,
const int host_inum, const int max_nbors,
const int maxspecial, UCL_Device &devi,
const int gpu_nbor, const int gpu_host,
const bool pre_cut, const int block_cell_2d,
const int block_cell_id, const int block_nbor_build,
const int threads_per_atom, const int warp_size,
const bool time_device,
const std::string compile_flags) {
clear();
_threads_per_atom=threads_per_atom;
_block_cell_2d=block_cell_2d;
_block_cell_id=block_cell_id;
_max_block_nbor_build=block_nbor_build;
_block_nbor_build=block_nbor_build;
_warp_size=warp_size;
_shared=shared;
dev=&devi;
_gpu_nbor=gpu_nbor;
_time_device=time_device;
if (gpu_host==0)
_gpu_host=false;
else if (gpu_host==1)
_gpu_host=true;
else
// Not yet implemented
assert(0==1);
if (pre_cut || gpu_nbor==0)
_alloc_packed=true;
else
_alloc_packed=false;
bool success=true;
// Initialize timers for the selected GPU
_nbor_time_avail=false;
time_nbor.init(*dev);
time_kernel.init(*dev);
time_hybrid1.init(*dev);
time_hybrid2.init(*dev);
time_transpose.init(*dev);
time_nbor.zero();
time_kernel.zero();
time_hybrid1.zero();
time_hybrid2.zero();
time_transpose.zero();
_max_atoms=static_cast<int>(static_cast<double>(inum)*1.10);
if (_max_atoms==0)
_max_atoms=1000;
_max_host=static_cast<int>(static_cast<double>(host_inum)*1.10);
_max_nbors=(max_nbors/threads_per_atom+1)*threads_per_atom;
_maxspecial=maxspecial;
if (gpu_nbor==0)
_maxspecial=0;
if (gpu_nbor==0)
success=success && (host_packed.alloc(2*IJ_SIZE,*dev,
UCL_WRITE_ONLY)==UCL_SUCCESS);
alloc(success);
if (!success)
return false;
if (_use_packing==false)
_shared->compile_kernels(devi,gpu_nbor,compile_flags);
return success;
}
void Neighbor::alloc(bool &success) {
dev_nbor.clear();
host_acc.clear();
int nt=_max_atoms+_max_host;
if (_use_packing==false || _gpu_nbor>0)
success=success &&
(dev_nbor.alloc((_max_nbors+2)*_max_atoms,*dev)==UCL_SUCCESS);
else
success=success && (dev_nbor.alloc(3*_max_atoms,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
success=success && (host_acc.alloc(nt*2,*dev,
UCL_READ_WRITE)==UCL_SUCCESS);
_c_bytes=dev_nbor.row_bytes();
if (_alloc_packed) {
dev_packed.clear();
success=success && (dev_packed.alloc((_max_nbors+2)*_max_atoms,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
_c_bytes+=dev_packed.row_bytes();
}
if (_max_host>0) {
nbor_host.clear();
dev_numj_host.clear();
host_ilist.clear();
host_jlist.clear();
success=(nbor_host.alloc(_max_nbors*_max_host,*dev,UCL_READ_WRITE,
UCL_READ_WRITE)==UCL_SUCCESS) && success;
success=success && (dev_numj_host.alloc(_max_host,*dev,
UCL_READ_WRITE)==UCL_SUCCESS);
success=success && (host_ilist.alloc(nt,*dev,UCL_NOT_PINNED)==UCL_SUCCESS);
if (!success)
return;
for (int i=0; i<nt; i++)
host_ilist[i]=i;
success=success && (host_jlist.alloc(_max_host,*dev,
UCL_NOT_PINNED)==UCL_SUCCESS);
if (!success)
return;
int *ptr=nbor_host.host.begin();
for (int i=0; i<_max_host; i++) {
host_jlist[i]=ptr;
ptr+=_max_nbors;
}
_c_bytes+=nbor_host.device.row_bytes()+dev_numj_host.row_bytes();
} else {
// Some OpenCL implementations return errors for NULL pointers as args
nbor_host.device.view(dev_nbor);
dev_numj_host.view(dev_nbor);
}
if (_maxspecial>0) {
dev_nspecial.clear();
dev_special.clear();
dev_special_t.clear();
int at=_max_atoms+_max_host;
success=success && (dev_nspecial.alloc(3*at,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
success=success && (dev_special.alloc(_maxspecial*at,*dev,
UCL_READ_WRITE)==UCL_SUCCESS);
success=success && (dev_special_t.alloc(_maxspecial*at,*dev,
UCL_READ_ONLY)==UCL_SUCCESS);
_gpu_bytes+=dev_nspecial.row_bytes()+dev_special.row_bytes()+
dev_special_t.row_bytes();
}
_allocated=true;
}
void Neighbor::clear() {
_gpu_bytes=0.0;
_cell_bytes=0.0;
_c_bytes=0.0;
_bin_time=0.0;
if (_ncells>0) {
_ncells=0;
cell_counts.clear();
if (_gpu_nbor==2)
delete [] cell_iter;
}
if (_allocated) {
_allocated=false;
_nbor_time_avail=false;
host_packed.clear();
host_acc.clear();
dev_nbor.clear();
nbor_host.clear();
dev_packed.clear();
dev_numj_host.clear();
host_ilist.clear();
host_jlist.clear();
dev_nspecial.clear();
dev_special.clear();
dev_special_t.clear();
time_kernel.clear();
time_nbor.clear();
time_hybrid1.clear();
time_hybrid2.clear();
time_transpose.clear();
}
}
double Neighbor::host_memory_usage() const {
if (_gpu_nbor>0) {
if (_gpu_host)
return nbor_host.device.row_bytes()*nbor_host.rows()+
host_ilist.row_bytes()+host_jlist.row_bytes();
else
return 0;
} else
return host_packed.row_bytes()*host_packed.rows()+host_acc.row_bytes()+
sizeof(Neighbor);
}
void Neighbor::get_host(const int inum, int *ilist, int *numj,
int **firstneigh, const int block_size) {
_nbor_time_avail=true;
time_nbor.start();
UCL_H_Vec<int> ilist_view;
ilist_view.view(ilist,inum,*dev);
ucl_copy(dev_nbor,ilist_view,false);
UCL_D_Vec<int> nbor_offset;
UCL_H_Vec<int> host_offset;
int copy_count=0;
int ij_count=0;
int acc_count=0;
int dev_count=0;
int *h_ptr=host_packed.begin();
_nbor_pitch=inum;
for (int ii=0; ii<inum; ii++) {
int i=ilist[ii];
int nj=numj[i];
host_acc[ii]=nj;
host_acc[ii+inum]=acc_count;
acc_count+=nj;
int *jlist=firstneigh[i];
for (int jj=0; jj<nj; jj++) {
*h_ptr=jlist[jj];
h_ptr++;
ij_count++;
if (ij_count==IJ_SIZE) {
dev_nbor.sync();
host_offset.view_offset(IJ_SIZE*(copy_count%2),host_packed,IJ_SIZE);
nbor_offset.view_offset(dev_count,dev_packed,IJ_SIZE);
ucl_copy(nbor_offset,host_offset,true);
copy_count++;
ij_count=0;
dev_count+=IJ_SIZE;
h_ptr=host_packed.begin()+(IJ_SIZE*(copy_count%2));
}
}
}
if (ij_count!=0) {
dev_nbor.sync();
host_offset.view_offset(IJ_SIZE*(copy_count%2),host_packed,ij_count);
nbor_offset.view_offset(dev_count,dev_packed,ij_count);
ucl_copy(nbor_offset,host_offset,true);
}
UCL_D_Vec<int> acc_view;
acc_view.view_offset(inum,dev_nbor,inum*2);
ucl_copy(acc_view,host_acc,true);
time_nbor.stop();
if (_use_packing==false) {
time_kernel.start();
int GX=static_cast<int>(ceil(static_cast<double>(inum)*_threads_per_atom/
block_size));
_shared->k_nbor.set_size(GX,block_size);
_shared->k_nbor.run(&dev_nbor, &dev_packed, &inum, &_threads_per_atom);
time_kernel.stop();
}
}
// This is the same as get host, but the requirement that ilist[i]=i and
// inum=nlocal is forced to be true to allow direct indexing of neighbors of
// neighbors
void Neighbor::get_host3(const int inum, const int nlist, int *ilist, int *numj,
int **firstneigh, const int block_size) {
_nbor_time_avail=true;
time_nbor.start();
UCL_H_Vec<int> ilist_view;
ilist_view.view(ilist,inum,*dev);
ucl_copy(dev_nbor,ilist_view,false);
UCL_D_Vec<int> nbor_offset;
UCL_H_Vec<int> host_offset;
int copy_count=0;
int ij_count=0;
int acc_count=0;
int dev_count=0;
int *h_ptr=host_packed.begin();
_nbor_pitch=inum;
if (nlist!=inum)
host_acc.zero(inum);
for (int ii=0; ii<nlist; ii++) {
int i=ilist[ii];
int nj=numj[i];
host_acc[i]=nj;
host_acc[i+inum]=acc_count;
acc_count+=nj;
}
for (int i=0; i<inum; i++) {
int nj=host_acc[i];
int *jlist=firstneigh[i];
for (int jj=0; jj<nj; jj++) {
*h_ptr=jlist[jj];
h_ptr++;
ij_count++;
if (ij_count==IJ_SIZE) {
dev_nbor.sync();
host_offset.view_offset(IJ_SIZE*(copy_count%2),host_packed,IJ_SIZE);
nbor_offset.view_offset(dev_count,dev_packed,IJ_SIZE);
ucl_copy(nbor_offset,host_offset,true);
copy_count++;
ij_count=0;
dev_count+=IJ_SIZE;
h_ptr=host_packed.begin()+(IJ_SIZE*(copy_count%2));
}
}
}
if (ij_count!=0) {
dev_nbor.sync();
host_offset.view_offset(IJ_SIZE*(copy_count%2),host_packed,ij_count);
nbor_offset.view_offset(dev_count,dev_packed,ij_count);
ucl_copy(nbor_offset,host_offset,true);
}
UCL_D_Vec<int> acc_view;
acc_view.view_offset(inum,dev_nbor,inum*2);
ucl_copy(acc_view,host_acc,true);
time_nbor.stop();
if (_use_packing==false) {
time_kernel.start();
int GX=static_cast<int>(ceil(static_cast<double>(inum)*_threads_per_atom/
block_size));
_shared->k_nbor.set_size(GX,block_size);
_shared->k_nbor.run(&dev_nbor, &dev_packed, &inum, &_threads_per_atom);
time_kernel.stop();
}
}
template <class numtyp, class acctyp>
void Neighbor::resize_max_neighbors(const int maxn, bool &success) {
if (maxn>_max_nbors) {
int mn=static_cast<int>(static_cast<double>(maxn)*1.10);
mn=(mn/_threads_per_atom+1)*_threads_per_atom;
success=success && (dev_nbor.resize((mn+1)*_max_atoms)==UCL_SUCCESS);
_gpu_bytes=dev_nbor.row_bytes();
if (_max_host>0) {
success=success && (nbor_host.resize(mn*_max_host)==UCL_SUCCESS);
int *ptr=nbor_host.host.begin();
for (int i=0; i<_max_host; i++) {
host_jlist[i]=ptr;
ptr+=mn;
}
_gpu_bytes+=nbor_host.row_bytes();
} else {
nbor_host.device.view(dev_nbor);
dev_numj_host.view(dev_nbor);
}
if (_alloc_packed) {
success=success && (dev_packed.resize((mn+2)*_max_atoms)==UCL_SUCCESS);
_gpu_bytes+=dev_packed.row_bytes();
}
_max_nbors=mn;
}
}
template <class numtyp, class acctyp>
void Neighbor::build_nbor_list(double **x, const int inum, const int host_inum,
const int nall, Atom<numtyp,acctyp> &atom,
double *sublo, double *subhi, tagint *tag,
int **nspecial, tagint **special, bool &success,
int &mn) {
_nbor_time_avail=true;
const int nt=inum+host_inum;
// Calculate number of cells and allocate storage for binning as necessary
int ncellx, ncelly, ncellz, ncell_3d;
int ghost_cells=2*_cells_in_cutoff;
ncellx = static_cast<int>(ceil((subhi[0]-sublo[0])/_cell_size))+ghost_cells;
ncelly = static_cast<int>(ceil((subhi[1]-sublo[1])/_cell_size))+ghost_cells;
ncellz = static_cast<int>(ceil((subhi[2]-sublo[2])/_cell_size))+ghost_cells;
ncell_3d = ncellx * ncelly * ncellz;
if (ncell_3d+1>_ncells) {
cell_counts.clear();
if (_gpu_nbor==2) {
if (_ncells>0)
delete [] cell_iter;
cell_iter = new int[ncell_3d+1];
cell_counts.alloc(ncell_3d+1,dev_nbor,UCL_READ_WRITE,UCL_READ_ONLY);
} else {
cell_counts.device.clear();
cell_counts.device.alloc(ncell_3d+1,dev_nbor);
}
_ncells=ncell_3d+1;
_cell_bytes=cell_counts.device.row_bytes();
}
const numtyp cutoff_cast=static_cast<numtyp>(_cutoff);
if (_maxspecial>0) {
time_nbor.start();
UCL_H_Vec<int> view_nspecial;
UCL_H_Vec<tagint> view_special, view_tag;
view_nspecial.view(nspecial[0],nt*3,*dev);
view_special.view(special[0],nt*_maxspecial,*dev);
view_tag.view(tag,nall,*dev);
ucl_copy(dev_nspecial,view_nspecial,nt*3,false);
ucl_copy(dev_special_t,view_special,nt*_maxspecial,false);
ucl_copy(atom.dev_tag,view_tag,nall,false);
time_nbor.stop();
if (_time_device)
time_nbor.add_to_total();
time_transpose.start();
const int b2x=_block_cell_2d;
const int b2y=_block_cell_2d;
const int g2x=static_cast<int>(ceil(static_cast<double>(_maxspecial)/b2x));
const int g2y=static_cast<int>(ceil(static_cast<double>(nt)/b2y));
_shared->k_transpose.set_size(g2x,g2y,b2x,b2y);
_shared->k_transpose.run(&dev_special,&dev_special_t,&_maxspecial,&nt);
time_transpose.stop();
}
// If binning on CPU, do this now
if (_gpu_nbor==2) {
double stime = MPI_Wtime();
int *cell_id=atom.host_cell_id.begin();
int *particle_id=atom.host_particle_id.begin();
// Build cell list on CPU
cell_counts.host.zero();
double i_cell_size=1.0/_cell_size;
int offset_hi=_cells_in_cutoff+1;
for (int i=0; i<nt; i++) {
double px, py, pz;
px=x[i][0]-sublo[0];
py=x[i][1]-sublo[1];
pz=x[i][2]-sublo[2];
int ix = static_cast<int>(px*i_cell_size+1);
ix = std::max(ix,_cells_in_cutoff);
ix = std::min(ix,ncellx-offset_hi);
int iy = static_cast<int>(py*i_cell_size+1);
iy = std::max(iy,_cells_in_cutoff);
iy = std::min(iy,ncelly-offset_hi);
int iz = static_cast<int>(pz*i_cell_size+1);
iz = std::max(iz,_cells_in_cutoff);
iz = std::min(iz,ncellz-offset_hi);
int id = ix+iy*ncellx+iz*ncellx*ncelly;
cell_id[i] = id;
cell_counts[id+1]++;
}
for (int i=nt; i<nall; i++) {
double px, py, pz;
px=x[i][0]-sublo[0];
py=x[i][1]-sublo[1];
pz=x[i][2]-sublo[2];
int ix = static_cast<int>(px*i_cell_size+1);
ix = std::max(ix,0);
ix = std::min(ix,ncellx-1);
int iy = static_cast<int>(py*i_cell_size+1);
iy = std::max(iy,0);
iy = std::min(iy,ncelly-1);
int iz = static_cast<int>(pz*i_cell_size+1);
iz = std::max(iz,0);
iz = std::min(iz,ncellz-1);
int id = ix+iy*ncellx+iz*ncellx*ncelly;
cell_id[i] = id;
cell_counts[id+1]++;
}
mn=0;
for (int i=0; i<_ncells; i++)
mn=std::max(mn,cell_counts[i]);
mn*=8;
set_nbor_block_size(mn/2);
resize_max_neighbors<numtyp,acctyp>(mn,success);
if (!success)
return;
_total_atoms=nt;
cell_iter[0]=0;
for (int i=1; i<_ncells; i++) {
cell_counts[i]+=cell_counts[i-1];
cell_iter[i]=cell_counts[i];
}
time_hybrid1.start();
cell_counts.update_device(true);
time_hybrid1.stop();
for (int i=0; i<nall; i++) {
int celli=cell_id[i];
int ploc=cell_iter[celli];
cell_iter[celli]++;
particle_id[ploc]=i;
}
time_hybrid2.start();
ucl_copy(atom.dev_particle_id,atom.host_particle_id,true);
time_hybrid2.stop();
_bin_time+=MPI_Wtime()-stime;
}
time_kernel.start();
_nbor_pitch=inum;
_shared->neigh_tex.bind_float(atom.x,4);
// If binning on GPU, do this now
if (_gpu_nbor==1) {
const numtyp i_cell_size=static_cast<numtyp>(1.0/_cell_size);
const int neigh_block=_block_cell_id;
const int GX=(int)ceil((float)nall/neigh_block);
const numtyp sublo0=static_cast<numtyp>(sublo[0]);
const numtyp sublo1=static_cast<numtyp>(sublo[1]);
const numtyp sublo2=static_cast<numtyp>(sublo[2]);
_shared->k_cell_id.set_size(GX,neigh_block);
_shared->k_cell_id.run(&atom.x, &atom.dev_cell_id,
&atom.dev_particle_id, &sublo0, &sublo1,
&sublo2, &i_cell_size, &ncellx, &ncelly, &ncellz,
&nt, &nall, &_cells_in_cutoff);
atom.sort_neighbor(nall);
/* calculate cell count */
_shared->k_cell_counts.set_size(GX,neigh_block);
_shared->k_cell_counts.run(&atom.dev_cell_id, &cell_counts, &nall,
&ncell_3d);
}
/* build the neighbor list */
const int cell_block=_block_nbor_build;
_shared->k_build_nbor.set_size(ncellx-ghost_cells,(ncelly-ghost_cells)*
(ncellz-ghost_cells),cell_block,1);
_shared->k_build_nbor.run(&atom.x, &atom.dev_particle_id,
&cell_counts, &dev_nbor, &nbor_host,
&dev_numj_host, &_max_nbors, &cutoff_cast, &ncellx,
&ncelly, &ncellz, &inum, &nt, &nall,
&_threads_per_atom, &_cells_in_cutoff);
/* Get the maximum number of nbors and realloc if necessary */
UCL_D_Vec<int> numj;
numj.view_offset(inum,dev_nbor,inum);
ucl_copy(host_acc,numj,inum,true);
if (nt>inum) {
UCL_H_Vec<int> host_offset;
host_offset.view_offset(inum,host_acc,nt-inum);
ucl_copy(host_offset,dev_numj_host,nt-inum,true);
}
if (_gpu_nbor!=2) {
host_acc.sync();
mn=host_acc[0];
for (int i=1; i<nt; i++)
mn=std::max(mn,host_acc[i]);
set_nbor_block_size(mn);
if (mn>_max_nbors) {
resize_max_neighbors<numtyp,acctyp>(mn,success);
if (!success)
return;
time_kernel.stop();
if (_time_device)
time_kernel.add_to_total();
build_nbor_list(x, inum, host_inum, nall, atom, sublo, subhi, tag,
nspecial, special, success, mn);
return;
}
}
if (_maxspecial>0) {
const int GX2=static_cast<int>(ceil(static_cast<double>
(nt*_threads_per_atom)/cell_block));
_shared->k_special.set_size(GX2,cell_block);
_shared->k_special.run(&dev_nbor, &nbor_host, &dev_numj_host,
&atom.dev_tag, &dev_nspecial, &dev_special,
&inum, &nt, &_max_nbors, &_threads_per_atom);
}
time_kernel.stop();
time_nbor.start();
if (inum<nt) {
nbor_host.update_host(true);
nbor_host.sync();
}
time_nbor.stop();
}
template void Neighbor::build_nbor_list<PRECISION,ACC_PRECISION>
(double **x, const int inum, const int host_inum, const int nall,
Atom<PRECISION,ACC_PRECISION> &atom, double *sublo, double *subhi,
tagint *, int **, tagint **, bool &success, int &mn);
Reference in New Issue View Git Blame Copy Permalink