ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
f0263bf14ccbcbaecba8e8d37d97e92ab09723be
lammps/lib/gpu/pppm_gpu_memory.cpp
W. Michael Brown f0263bf14c pppm/gpu now using all gpu kernels.
2011-03-25 11:49:09 -04:00

351 lines
11 KiB
C++
Raw Blame History

/* ----------------------------------------------------------------------
LAMMPS - Large-scale Charge/Molecular Massively Parallel Simulator
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
Copyright (2003) Sandia Corporation. Under the terms of Contract
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
certain rights in this software. This software is distributed under
the GNU General Public License.
See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */
/* ----------------------------------------------------------------------
Contributing authors: Mike Brown (ORNL), brownw@ornl.gov
------------------------------------------------------------------------- */
#ifdef USE_OPENCL
#include "pppm_gpu_cl.h"
#define MEM_THREADS 16
#else
#include "pppm_gpu_ptx.h"
#endif
#include "pppm_gpu_memory.h"
#include <cassert>
// Maximum order for spline
#define MAX_SPLINE 8
// Thread block size for all kernels (Must be >=MAX_SPLINE^2)
#define BLOCK_1D 64
// Number of threads per pencil for charge spread
//#define PENCIL_SIZE MEM_THREADS
#define PENCIL_SIZE 32
// Number of pencils per block for charge spread
//#define BLOCK_PENCILS (BLOCK_1D/PENCIL_SIZE)
#define BLOCK_PENCILS 2
#define PPPMGPUMemoryT PPPMGPUMemory<numtyp, acctyp>
extern PairGPUDevice<PRECISION,ACC_PRECISION> pair_gpu_device;
template <class numtyp, class acctyp>
PPPMGPUMemoryT::PPPMGPUMemory() : _allocated(false), _compiled(false),
_max_bytes(0) {
device=&pair_gpu_device;
ans=new PairGPUAns<numtyp,acctyp>();
}
template <class numtyp, class acctyp>
PPPMGPUMemoryT::~PPPMGPUMemory() {
clear(0.0);
delete ans;
}
template <class numtyp, class acctyp>
int PPPMGPUMemoryT::bytes_per_atom() const {
return device->atom.bytes_per_atom()+ans->bytes_per_atom()+1;
}
template <class numtyp, class acctyp>
numtyp * PPPMGPUMemoryT::init(const int nlocal, const int nall, FILE *_screen,
const int order, const int nxlo_out,
const int nylo_out, const int nzlo_out,
const int nxhi_out, const int nyhi_out,
const int nzhi_out, double **rho_coeff,
numtyp **vd_brick, bool &success) {
_max_bytes=10;
screen=_screen;
if (!device->init(*ans,nlocal,nall)) {
success=false;
return 0;
}
ucl_device=device->gpu;
atom=&device->atom;
_block_size=BLOCK_1D;
assert(BLOCK_PENCILS*PENCIL_SIZE==BLOCK_1D);
assert(MAX_SPLINE*MAX_SPLINE<=BLOCK_1D);
if (static_cast<size_t>(_block_size)>ucl_device->group_size())
_block_size=ucl_device->group_size();
compile_kernels(*ucl_device);
// Initialize timers for the selected GPU
time_in.init(*ucl_device);
time_in.zero();
time_out.init(*ucl_device);
time_out.zero();
time_map.init(*ucl_device);
time_map.zero();
time_rho.init(*ucl_device);
time_rho.zero();
time_interp.init(*ucl_device);
time_interp.zero();
pos_tex.bind_float(atom->dev_x,4);
q_tex.bind_float(atom->dev_q,1);
_allocated=true;
_max_bytes=0;
_max_an_bytes=ans->gpu_bytes();
_order=order;
_order_m_1=order-1;
_order2=_order_m_1*_order;
_nlower=-(_order-1)/2;
_nupper=order/2;
_nxlo_out=nxlo_out;
_nylo_out=nylo_out;
_nzlo_out=nzlo_out;
_nxhi_out=nxhi_out;
_nyhi_out=nyhi_out;
_nzhi_out=nzhi_out;
_max_brick_atoms=10;
// Get rho_coeff on device
int n2lo=(1-order)/2;
int numel=order*( order/2 - n2lo + 1 );
success=success && (d_rho_coeff.alloc(numel,*ucl_device,UCL_READ_ONLY)==
UCL_SUCCESS);
UCL_H_Vec<double> view;
view.view(rho_coeff[0]+n2lo,numel,*ucl_device);
ucl_copy(d_rho_coeff,view,true);
_max_bytes+=d_rho_coeff.row_bytes();
// Allocate storage for grid
_npts_x=nxhi_out-nxlo_out+1;
_npts_y=nyhi_out-nylo_out+1;
_npts_z=nzhi_out-nzlo_out+1;
_npts_yx=_npts_x*_npts_y;
success=success && (d_brick.alloc(_npts_x*_npts_y*_npts_z*4,*ucl_device)==
UCL_SUCCESS);
success=success && (h_brick.alloc(_npts_x*_npts_y*_npts_z,*ucl_device)==
UCL_SUCCESS);
success=success && (h_vd_brick.alloc(_npts_x*_npts_y*_npts_z*4,*ucl_device)==
UCL_SUCCESS);
*vd_brick=h_vd_brick.begin();
_max_bytes+=d_brick.row_bytes();
// Allocate vector with count of atoms assigned to each grid point
_nlocal_x=_npts_x+_nlower-_nupper;
_nlocal_y=_npts_y+_nlower-_nupper;
_nlocal_z=_npts_z+_nlower-_nupper;
_nlocal_yx=_nlocal_x*_nlocal_y;
_atom_stride=_nlocal_x*_nlocal_y*_nlocal_z;
success=success && (d_brick_counts.alloc(_atom_stride,*ucl_device)==
UCL_SUCCESS);
_max_bytes+=d_brick_counts.row_bytes();
// Allocate storage for atoms assigned to each grid point
success=success && (d_brick_atoms.alloc(_atom_stride*_max_brick_atoms,
*ucl_device)==UCL_SUCCESS);
_max_bytes+=d_brick_atoms.row_bytes();
// Allocate error flags for checking out of bounds atoms
success=success && (h_error_flag.alloc(1,*ucl_device)==UCL_SUCCESS);
success=success && (d_error_flag.alloc(1,*ucl_device,UCL_WRITE_ONLY)==
UCL_SUCCESS);
d_error_flag.zero();
_max_bytes+=1;
return h_brick.begin();
}
template <class numtyp, class acctyp>
void PPPMGPUMemoryT::clear(const double cpu_time) {
if (!_allocated)
return;
_allocated=false;
d_brick.clear();
h_brick.clear();
h_vd_brick.clear();
d_brick_counts.clear();
h_error_flag.clear();
d_error_flag.clear();
d_brick_atoms.clear();
acc_timers();
device->output_kspace_times(time_in,time_out,time_map,time_rho,time_interp,
*ans,_max_bytes+_max_an_bytes,cpu_time,screen);
if (_compiled) {
k_particle_map.clear();
k_make_rho.clear();
k_interp.clear();
delete pppm_program;
_compiled=false;
}
time_in.clear();
time_out.clear();
time_map.clear();
time_rho.clear();
time_interp.clear();
device->clear();
}
// ---------------------------------------------------------------------------
// Charge spreading stuff
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int PPPMGPUMemoryT::spread(const int ago, const int nlocal, const int nall,
double **host_x, int *host_type, bool &success,
double *host_q, double *boxlo,
const double delxinv, const double delyinv,
const double delzinv) {
device->stop_host_timer();
acc_timers();
if (nlocal==0) {
zero_timers();
return 0;
}
ans->inum(nlocal);
if (ago==0) {
resize_atom(nlocal,nall,success);
resize_local(nlocal,success);
if (!success)
return 0;
double bytes=ans->gpu_bytes();
if (bytes>_max_an_bytes)
_max_an_bytes=bytes;
}
atom->cast_x_data(host_x,host_type);
atom->cast_q_data(host_q);
atom->add_x_data(host_x,host_type);
atom->add_q_data();
time_map.start();
// Compute the block size and grid size to keep all cores busy
int BX=this->block_size();
int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/BX));
int _max_atoms=10;
int ainum=this->ans->inum();
// Boxlo adjusted to be upper left brick and shift for even spline order
double shift=0.0;
if (_order % 2)
shift=0.5;
_brick_x=boxlo[0]+(_nxlo_out-_nlower-shift)/delxinv;
_brick_y=boxlo[1]+(_nylo_out-_nlower-shift)/delyinv;
_brick_z=boxlo[2]+(_nzlo_out-_nlower-shift)/delzinv;
_delxinv=delxinv;
_delyinv=delyinv;
_delzinv=delzinv;
double delvolinv = delxinv*delyinv*delzinv;
numtyp f_delvolinv = delvolinv;
d_brick_counts.zero();
k_particle_map.set_size(GX,BX);
k_particle_map.run(&atom->dev_x.begin(), &atom->dev_q.begin(), &f_delvolinv,
&ainum, &d_brick_counts.begin(), &d_brick_atoms.begin(),
&_brick_x, &_brick_y, &_brick_z, &_delxinv, &_delyinv,
&_delzinv, &_nlocal_x, &_nlocal_y, &_nlocal_z,
&_atom_stride, &_max_brick_atoms, &d_error_flag.begin());
time_map.stop();
time_rho.start();
BX=block_size();
GX=static_cast<int>(ceil(static_cast<double>(_npts_y*_npts_z)/BLOCK_PENCILS));
k_make_rho.set_size(GX,BX);
k_make_rho.run(&atom->dev_x.begin(), &atom->dev_q.begin(),
&d_brick_counts.begin(), &d_brick_atoms.begin(),
&d_brick.begin(), &d_rho_coeff.begin(), &_atom_stride, &_npts_x,
&_npts_y, &_npts_z, &_nlocal_x, &_nlocal_y, &_nlocal_z,
&_order_m_1, &_order, &_order2);
time_rho.stop();
time_out.start();
ucl_copy(h_brick,d_brick,_npts_yx*_npts_z,true);
ucl_copy(h_error_flag,d_error_flag,false);
time_out.stop();
if (h_error_flag[0]==2) {
// Not enough storage for atoms on the brick
_max_brick_atoms*=2;
d_error_flag.zero();
d_brick_atoms.clear();
d_brick_atoms.alloc(_atom_stride*_max_atoms,*ucl_device);
_max_bytes+=d_brick_atoms.row_bytes();
return spread(ago,nlocal,nall,host_x,host_type,success,host_q,boxlo,
delxinv,delyinv,delzinv);
}
return h_error_flag[0];
}
// ---------------------------------------------------------------------------
// Charge spreading stuff
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void PPPMGPUMemoryT::interp(const numtyp qqrd2e_scale) {
time_in.start();
ucl_copy(d_brick,h_vd_brick,true);
time_in.stop();
time_interp.start();
// Compute the block size and grid size to keep all cores busy
int BX=this->block_size();
int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/BX));
int ainum=this->ans->inum();
k_interp.set_size(GX,BX);
k_interp.run(&atom->dev_x.begin(), &atom->dev_q.begin(), &ainum,
&d_brick.begin(), &d_rho_coeff.begin(), &_npts_x, &_npts_yx,
&_brick_x, &_brick_y, &_brick_z, &_delxinv, &_delyinv, &_delzinv,
&_order, &_order2, &qqrd2e_scale, &ans->dev_ans.begin());
time_interp.stop();
ans->copy_answers(false,false,false,false);
device->add_ans_object(ans);
}
template <class numtyp, class acctyp>
double PPPMGPUMemoryT::host_memory_usage() const {
return device->atom.host_memory_usage()+sizeof(PPPMGPUMemory<numtyp,acctyp>);
}
template <class numtyp, class acctyp>
void PPPMGPUMemoryT::compile_kernels(UCL_Device &dev) {
if (_compiled)
return;
std::string flags="-cl-fast-relaxed-math -cl-mad-enable "+
std::string(OCL_PRECISION_COMPILE);
pppm_program=new UCL_Program(dev);
pppm_program->load_string(pppm_gpu_kernel,flags.c_str());
k_particle_map.set_function(*pppm_program,"particle_map");
k_make_rho.set_function(*pppm_program,"make_rho");
k_interp.set_function(*pppm_program,"interp");
pos_tex.get_texture(*pppm_program,"pos_tex");
q_tex.get_texture(*pppm_program,"q_tex");
_compiled=true;
}
template class PPPMGPUMemory<PRECISION,ACC_PRECISION>;
Reference in New Issue View Git Blame Copy Permalink