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

update code for ufm/gpu to match changes in base class where uf4 is removed

Browse Source
This commit is contained in:
Axel Kohlmeyer
2020-08-03 08:12:09 -04:00
parent d271ee65e8
commit 7e5e741ea1
4 changed files with 46 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
lib/gpu/lal_ufm.cpp
View File

@ -10,7 +10,7 @@
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________ __________________________________________________________________________
begin : begin :
email : pl.rodolfo@gmail.com email : pl.rodolfo@gmail.com
dekoning@ifi.unicamp.br dekoning@ifi.unicamp.br
***************************************************************************/ ***************************************************************************/
@ -38,7 +38,7 @@ template <class numtyp, class acctyp>
UFMT::~UFM() { UFMT::~UFM() {
clear(); clear();
} }
template <class numtyp, class acctyp> template <class numtyp, class acctyp>
int UFMT::bytes_per_atom(const int max_nbors) const { int UFMT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors); return this->bytes_per_atom_atomic(max_nbors);
@ -46,9 +46,9 @@ int UFMT::bytes_per_atom(const int max_nbors) const {
template <class numtyp, class acctyp> template <class numtyp, class acctyp>
int UFMT::init(const int ntypes, int UFMT::init(const int ntypes,
double **host_cutsq, double **host_uf1, double **host_cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf2, double **host_uf3,
double **host_uf4, double **host_offset, double **host_offset,
double *host_special_lj, const int nlocal, double *host_special_lj, const int nlocal,
const int nall, const int max_nbors, const int nall, const int max_nbors,
const int maxspecial, const double cell_size, const int maxspecial, const double cell_size,
@ -78,11 +78,11 @@ int UFMT::init(const int ntypes,
uf1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); uf1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,uf1,host_write,host_uf1,host_uf2, this->atom->type_pack4(ntypes,lj_types,uf1,host_write,host_uf1,host_uf2,
host_cutsq); host_cutsq);
uf3.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY); uf3.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,uf3,host_write,host_uf3,host_uf4, this->atom->type_pack4(ntypes,lj_types,uf3,host_write,host_uf3,host_uf2,
host_offset); host_offset);
UCL_H_Vec<double> dview; UCL_H_Vec<double> dview;
sp_lj.alloc(4,*(this->ucl_device),UCL_READ_ONLY); sp_lj.alloc(4,*(this->ucl_device),UCL_READ_ONLY);
@ -96,18 +96,17 @@ int UFMT::init(const int ntypes,
template <class numtyp, class acctyp> template <class numtyp, class acctyp>
void UFMT::reinit(const int ntypes, double **host_cutsq, double **host_uf1, void UFMT::reinit(const int ntypes, double **host_cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf2, double **host_uf3, double **host_offset) {
double **host_uf4, double **host_offset) {
// Allocate a host write buffer for data initialization // Allocate a host write buffer for data initialization
UCL_H_Vec<numtyp> host_write(_lj_types*_lj_types*32,*(this->ucl_device), UCL_H_Vec<numtyp> host_write(_lj_types*_lj_types*32,*(this->ucl_device),
UCL_WRITE_ONLY); UCL_WRITE_ONLY);
for (int i=0; i<_lj_types*_lj_types; i++) for (int i=0; i<_lj_types*_lj_types; i++)
host_write[i]=0.0; host_write[i]=0.0;
this->atom->type_pack4(ntypes,_lj_types,uf1,host_write,host_uf1,host_uf2, this->atom->type_pack4(ntypes,_lj_types,uf1,host_write,host_uf1,host_uf2,
host_cutsq); host_cutsq);
this->atom->type_pack4(ntypes,_lj_types,uf3,host_write,host_uf3,host_uf4, this->atom->type_pack4(ntypes,_lj_types,uf3,host_write,host_uf3,host_uf2,
host_offset); host_offset);
} }
@ -145,7 +144,7 @@ void UFMT::loop(const bool _eflag, const bool _vflag) {
vflag=1; vflag=1;
else else
vflag=0; vflag=0;
int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/ int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
(BX/this->_threads_per_atom))); (BX/this->_threads_per_atom)));
@ -157,12 +156,12 @@ void UFMT::loop(const bool _eflag, const bool _vflag) {
this->k_pair_fast.run(&this->atom->x, &uf1, &uf3, &sp_lj, this->k_pair_fast.run(&this->atom->x, &uf1, &uf3, &sp_lj,
&this->nbor->dev_nbor, &this->_nbor_data->begin(), &this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag, &this->ans->force, &this->ans->engv, &eflag,
&vflag, &ainum, &nbor_pitch, &vflag, &ainum, &nbor_pitch,
&this->_threads_per_atom); &this->_threads_per_atom);
} else { } else {
this->k_pair.set_size(GX,BX); this->k_pair.set_size(GX,BX);
this->k_pair.run(&this->atom->x, &uf1, &uf3, &_lj_types, &sp_lj, this->k_pair.run(&this->atom->x, &uf1, &uf3, &_lj_types, &sp_lj,
&this->nbor->dev_nbor, &this->_nbor_data->begin(), &this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag, &vflag, &this->ans->force, &this->ans->engv, &eflag, &vflag,
&ainum, &nbor_pitch, &this->_threads_per_atom); &ainum, &nbor_pitch, &this->_threads_per_atom);
} }

23
lib/gpu/lal_ufm.h
View File

@ -10,7 +10,7 @@
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________ __________________________________________________________________________
begin : begin :
email : pl.rodolfo@gmail.com email : pl.rodolfo@gmail.com
dekoning@ifi.unicamp.br dekoning@ifi.unicamp.br
***************************************************************************/ ***************************************************************************/
@ -32,7 +32,7 @@ class UFM : public BaseAtomic<numtyp, acctyp> {
/** \param max_nbors initial number of rows in the neighbor matrix /** \param max_nbors initial number of rows in the neighbor matrix
* \param cell_size cutoff + skin * \param cell_size cutoff + skin
* \param gpu_split fraction of particles handled by device * \param gpu_split fraction of particles handled by device
* *
* Returns: * Returns:
* - 0 if successful * - 0 if successful
* - -1 if fix gpu not found * - -1 if fix gpu not found
@ -41,16 +41,15 @@ class UFM : public BaseAtomic<numtyp, acctyp> {
* - -5 Double precision is not supported on card **/ * - -5 Double precision is not supported on card **/
int init(const int ntypes, double **host_cutsq, int init(const int ntypes, double **host_cutsq,
double **host_uf1, double **host_uf2, double **host_uf3, double **host_uf1, double **host_uf2, double **host_uf3,
double **host_uf4, double **host_offset, double *host_special_lj, double **host_offset, double *host_special_lj,
const int nlocal, const int nall, const int max_nbors, const int nlocal, const int nall, const int max_nbors,
const int maxspecial, const double cell_size, const int maxspecial, const double cell_size,
const double gpu_split, FILE *screen); const double gpu_split, FILE *screen);
/// Send updated coeffs from host to device (to be compatible with fix adapt) /// Send updated coeffs from host to device (to be compatible with fix adapt)
void reinit(const int ntypes, double **host_cutsq, void reinit(const int ntypes, double **host_cutsq, double **host_uf1,
double **host_uf1, double **host_uf2, double **host_uf3, double **host_uf2, double **host_uf3, double **host_offset);
double **host_uf4, double **host_offset);
/// Clear all host and device data /// Clear all host and device data
/** \note This is called at the beginning of the init() routine **/ /** \note This is called at the beginning of the init() routine **/
void clear(); void clear();
@ -65,7 +64,7 @@ class UFM : public BaseAtomic<numtyp, acctyp> {
/// uf1.x = uf1, uf1.y = uf2, uf1.z = cutsq /// uf1.x = uf1, uf1.y = uf2, uf1.z = cutsq
UCL_D_Vec<numtyp4> uf1; UCL_D_Vec<numtyp4> uf1;
/// uf3.x = uf3, uf3.y = uf4, uf3.z = offset /// uf3.x = uf3, uf3.y = uf2, uf3.z = offset
UCL_D_Vec<numtyp4> uf3; UCL_D_Vec<numtyp4> uf3;
/// Special LJ values /// Special LJ values
UCL_D_Vec<numtyp> sp_lj; UCL_D_Vec<numtyp> sp_lj;
@ -73,7 +72,7 @@ class UFM : public BaseAtomic<numtyp, acctyp> {
/// If atom type constants fit in shared memory, use fast kernels /// If atom type constants fit in shared memory, use fast kernels
bool shared_types; bool shared_types;
/// Number of atom types /// Number of atom types
int _lj_types; int _lj_types;
private: private:

33
lib/gpu/lal_ufm_ext.cpp
View File

@ -10,7 +10,7 @@
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL) This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________ __________________________________________________________________________
begin : begin :
email : pl.rodolfo@gmail.com email : pl.rodolfo@gmail.com
dekoning@ifi.unicamp.br dekoning@ifi.unicamp.br
***************************************************************************/ ***************************************************************************/
@ -30,10 +30,10 @@ static UFM<PRECISION,ACC_PRECISION> UFMLMF;
// Allocate memory on host and device and copy constants to device // Allocate memory on host and device and copy constants to device
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1, int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf4, double **host_uf2, double **host_uf3, double **offset,
double **offset, double *special_lj, const int inum, const int nall, double *special_lj, const int inum, const int nall,
const int max_nbors, const int maxspecial, const double cell_size, const int max_nbors, const int maxspecial, const double cell_size,
int &gpu_mode, FILE *screen) { int &gpu_mode, FILE *screen) {
UFMLMF.clear(); UFMLMF.clear();
gpu_mode=UFMLMF.device->gpu_mode(); gpu_mode=UFMLMF.device->gpu_mode();
double gpu_split=UFMLMF.device->particle_split(); double gpu_split=UFMLMF.device->particle_split();
@ -57,8 +57,8 @@ int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1,
int init_ok=0; int init_ok=0;
if (world_me==0) if (world_me==0)
init_ok=UFMLMF.init(ntypes, cutsq, host_uf1, host_uf2, host_uf3, init_ok=UFMLMF.init(ntypes, cutsq, host_uf1, host_uf2, host_uf3,
host_uf4, offset, special_lj, inum, nall, 300, offset, special_lj, inum, nall, 300,
maxspecial, cell_size, gpu_split, screen); maxspecial, cell_size, gpu_split, screen);
UFMLMF.device->world_barrier(); UFMLMF.device->world_barrier();
if (message) if (message)
@ -74,12 +74,12 @@ int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1,
fflush(screen); fflush(screen);
} }
if (gpu_rank==i && world_me!=0) if (gpu_rank==i && world_me!=0)
init_ok=UFMLMF.init(ntypes, cutsq, host_uf1, host_uf2, host_uf3, host_uf4, init_ok=UFMLMF.init(ntypes, cutsq, host_uf1, host_uf2, host_uf3,
offset, special_lj, inum, nall, 300, maxspecial, offset, special_lj, inum, nall, 300, maxspecial,
cell_size, gpu_split, screen); cell_size, gpu_split, screen);
UFMLMF.device->gpu_barrier(); UFMLMF.device->gpu_barrier();
if (message) if (message)
fprintf(screen,"Done.\n"); fprintf(screen,"Done.\n");
} }
if (message) if (message)
@ -94,19 +94,18 @@ int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1,
// Copy updated coeffs from host to device // Copy updated coeffs from host to device
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
void ufml_gpu_reinit(const int ntypes, double **cutsq, double **host_uf1, void ufml_gpu_reinit(const int ntypes, double **cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf4, double **host_uf2, double **host_uf3, double **offset) {
double **offset) {
int world_me=UFMLMF.device->world_me(); int world_me=UFMLMF.device->world_me();
int gpu_rank=UFMLMF.device->gpu_rank(); int gpu_rank=UFMLMF.device->gpu_rank();
int procs_per_gpu=UFMLMF.device->procs_per_gpu(); int procs_per_gpu=UFMLMF.device->procs_per_gpu();
if (world_me==0) if (world_me==0)
UFMLMF.reinit(ntypes, cutsq, host_uf1, host_uf2, host_uf3, host_uf4, offset); UFMLMF.reinit(ntypes, cutsq, host_uf1, host_uf2, host_uf3, offset);
UFMLMF.device->world_barrier(); UFMLMF.device->world_barrier();
for (int i=0; i<procs_per_gpu; i++) { for (int i=0; i<procs_per_gpu; i++) {
if (gpu_rank==i && world_me!=0) if (gpu_rank==i && world_me!=0)
UFMLMF.reinit(ntypes, cutsq, host_uf1, host_uf2, host_uf3, host_uf4, offset); UFMLMF.reinit(ntypes, cutsq, host_uf1, host_uf2, host_uf3, offset);
UFMLMF.device->gpu_barrier(); UFMLMF.device->gpu_barrier();
} }
} }
@ -125,8 +124,8 @@ int ** ufml_gpu_compute_n(const int ago, const int inum_full,
return UFMLMF.compute(ago, inum_full, nall, host_x, host_type, sublo, return UFMLMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
subhi, tag, nspecial, special, eflag, vflag, eatom, subhi, tag, nspecial, special, eflag, vflag, eatom,
vatom, host_start, ilist, jnum, cpu_time, success); vatom, host_start, ilist, jnum, cpu_time, success);
} }
void ufml_gpu_compute(const int ago, const int inum_full, const int nall, void ufml_gpu_compute(const int ago, const int inum_full, const int nall,
double **host_x, int *host_type, int *ilist, int *numj, double **host_x, int *host_type, int *ilist, int *numj,
int **firstneigh, const bool eflag, const bool vflag, int **firstneigh, const bool eflag, const bool vflag,

8
src/GPU/pair_ufm_gpu.cpp
View File

@ -43,13 +43,13 @@ using namespace LAMMPS_NS;
// External functions from cuda library for atom decomposition // External functions from cuda library for atom decomposition
int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1, int ufml_gpu_init(const int ntypes, double **cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf4, double **host_uf2, double **host_uf3,
double **offset, double *special_lj, const int nlocal, double **offset, double *special_lj, const int nlocal,
const int nall, const int max_nbors, const int maxspecial, const int nall, const int max_nbors, const int maxspecial,
const double cell_size, int &gpu_mode, FILE *screen); const double cell_size, int &gpu_mode, FILE *screen);
int ufml_gpu_reinit(const int ntypes, double **cutsq, double **host_uf1, int ufml_gpu_reinit(const int ntypes, double **cutsq, double **host_uf1,
double **host_uf2, double **host_uf3, double **host_uf4, double **host_uf2, double **host_uf3,
double **offset); double **offset);
void ufml_gpu_clear(); void ufml_gpu_clear();
@ -166,7 +166,7 @@ void PairUFMGPU::init_style()
int maxspecial=0; int maxspecial=0;
if (atom->molecular) if (atom->molecular)
maxspecial=atom->maxspecial; maxspecial=atom->maxspecial;
int success = ufml_gpu_init(atom->ntypes+1, cutsq, uf1, uf2, uf3, uf4, int success = ufml_gpu_init(atom->ntypes+1, cutsq, uf1, uf2, uf3,
offset, force->special_lj, atom->nlocal, offset, force->special_lj, atom->nlocal,
atom->nlocal+atom->nghost, 300, maxspecial, atom->nlocal+atom->nghost, 300, maxspecial,
cell_size, gpu_mode, screen); cell_size, gpu_mode, screen);
@ -185,7 +185,7 @@ void PairUFMGPU::reinit()
{ {
Pair::reinit(); Pair::reinit();
ufml_gpu_reinit(atom->ntypes+1, cutsq, uf1, uf2, uf3, uf4, offset); ufml_gpu_reinit(atom->ntypes+1, cutsq, uf1, uf2, uf3, offset);
} }
/* ---------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- */