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

Finishing laptop optimization of PPPM charge spread kernels.

Browse Source
This commit is contained in:
W. Michael Brown
2011-02-17 13:37:02 -05:00
parent a5163fabd2
commit e8da16ff23
6 changed files with 90 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
lib/gpu/pppm_gpu_memory.cpp
View File

@ -25,6 +25,7 @@
#define BLOCK_1D 64
#define BLOCK_X 8
#define BLOCK_Y 8
#define MAX_STENCIL 8
#define PPPMGPUMemoryT PPPMGPUMemory<numtyp, acctyp>
extern PairGPUDevice<PRECISION,ACC_PRECISION> pair_gpu_device;
@ -146,11 +147,19 @@ numtyp * PPPMGPUMemoryT::init(const int nlocal, const int nall, FILE *_screen,
d_error_flag.zero();
_max_bytes+=1;
// Used to resequence atom indices to reduce probability of collisions
// during atomic ops
_resequence_skip=ucl_device->cores();
if (_resequence_skip%_block_size!=0)
_resequence_skip=int(_resequence_skip/_block_size+1)*_block_size;
assert(_block_size%_block_x_size==0);
std::cout << "LO: " << _nxlo_out << " " << _nylo_out << " " << _nzlo_out << " " << _nlower << std::endl;
std::cout << "HI: " << _nxhi_out << " " << _nyhi_out << " " << _nzhi_out << " " << _nupper << std::endl;
std::cout << "pts: " << _npts_x << " " << _npts_y << " " << _npts_z << std::endl;
std::cout << "local: " << _nlocal_x << " " << _nlocal_y << " " << _nlocal_z << std::endl;
return h_brick.begin();
}
@ -185,7 +194,7 @@ void PPPMGPUMemoryT::clear() {
device->clear();
}
/*
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
@ -221,8 +230,11 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
// 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));
// Resequence atom indices to avoid collisions during atomic ops
int skip=_resequence_skip;
if (skip>GX*BX/8)
skip=_block_x_size;
int _max_atoms=10;
int ainum=this->ans->inum();
@ -248,7 +260,7 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
&d_brick_atoms.begin(), &f_brick_x, &f_brick_y,
&f_brick_z, &f_delxinv, &f_delyinv, &f_delzinv, &_nlocal_x,
&_nlocal_y, &_nlocal_z, &_atom_stride, &_max_brick_atoms,
&d_error_flag.begin());
&d_error_flag.begin(),&skip);
time_map.stop();
if (_order % 2)
@ -264,14 +276,17 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
int BY=block_y_size();
GX=static_cast<int>(ceil(static_cast<double>(_nlocal_x)/BX));
int GY=static_cast<int>(ceil(static_cast<double>(_nlocal_y)/BY));
int x_threads=GX*BX;
int GZ=static_cast<int>(ceil(static_cast<double>(_nlocal_z)/8));
GX*=GZ;
d_brick.zero();
k_make_rho.set_size(GX,GY,BX,BY);
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_yx, &_nlocal_x, &_nlocal_y, &_nlocal_z, &f_brick_x,
&f_brick_y, &f_brick_z, &f_delxinv, &f_delyinv, &f_delzinv,
&_order, &_order2, &f_delvolinv);
&_npts_yx, &_nlocal_x, &_nlocal_y, &_nlocal_z, &x_threads,
&f_brick_x, &f_brick_y, &f_brick_z, &f_delxinv, &f_delyinv,
&f_delzinv, &_order, &_order2, &f_delvolinv);
time_rho.stop();
time_out.start();
@ -291,7 +306,7 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
return h_error_flag[0];
}
*/
/*
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
@ -328,8 +343,11 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
// 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));
// Resequence atom indices to avoid collisions during atomic ops
int skip=_resequence_skip;
if (skip>GX*BX/8)
skip=_block_x_size;
int _max_atoms=10;
int ainum=this->ans->inum();
@ -355,7 +373,7 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
&d_brick_atoms.begin(), &f_brick_x, &f_brick_y,
&f_brick_z, &f_delxinv, &f_delyinv, &f_delzinv, &_nlocal_x,
&_nlocal_y, &_nlocal_z, &_atom_stride, &_max_brick_atoms,
&d_error_flag.begin());
&d_error_flag.begin(),&skip);
time_map.stop();
if (_order % 2)
@ -398,7 +416,7 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
return h_error_flag[0];
}
*/
/*
// ---------------------------------------------------------------------------
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
@ -434,8 +452,11 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
// 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));
// Resequence atom indices to avoid collisions during atomic ops
int skip=_resequence_skip;
if (skip>GX*BX/8)
skip=_block_x_size;
int ainum=this->ans->inum();
@ -458,13 +479,13 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
time_rho.start();
d_brick.zero();
k_make_rho.set_size(GX,BX);
k_make_rho.run(&atom->dev_x.begin(), &atom->dev_q.begin(), &ainum,
&d_brick.begin(), &d_rho_coeff.begin(), &_npts_x,
&_npts_yx, &_nlocal_x, &_nlocal_y, &_nlocal_z, &f_brick_x,
&f_brick_y, &f_brick_z, &f_delxinv, &f_delyinv, &f_delzinv,
&_order, &_order2, &f_delvolinv, &d_error_flag.begin());
&_order, &_order2, &f_delvolinv, &d_error_flag.begin(),
&skip);
time_rho.stop();
time_out.start();
@ -474,7 +495,7 @@ int PPPMGPUMemoryT::compute(const int ago, const int nlocal, const int nall,
return h_error_flag[0];
}
*/
template <class numtyp, class acctyp>
double PPPMGPUMemoryT::host_memory_usage() const {
@ -492,7 +513,7 @@ void PPPMGPUMemoryT::compile_kernels(UCL_Device &dev) {
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_rho3");
k_make_rho.set_function(*pppm_program,"make_rho");
pos_tex.get_texture(*pppm_program,"pos_tex");
q_tex.get_texture(*pppm_program,"q_tex");