lal_charmm_long.cpp
No OneTemporary
Actions

Subscribers

None

File Metadata

Created: Tue, Jun 18, 22:26

lal_charmm_long.cpp
View Options

	/***************************************************************************
	charmm_long.cpp
	-------------------
	W. Michael Brown (ORNL)

	Class for acceleration of the charmm/coul/long pair style.

	__________________________________________________________________________
	This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
	__________________________________________________________________________

	begin :
	email : brownw@ornl.gov
	***************************************************************************/

	#if defined(USE_OPENCL)
	#include "charmm_long_cl.h"
	#elif defined(USE_CUDART)
	const char *charmm_long=0;
	#else
	#include "charmm_long_cubin.h"
	#endif

	#include "lal_charmm_long.h"
	#include <cassert>
	using namespace LAMMPS_AL;
	#define CHARMMLongT CHARMMLong<numtyp, acctyp>

	extern Device<PRECISION,ACC_PRECISION> device;

	template <class numtyp, class acctyp>
	CHARMMLongT::CHARMMLong() : BaseCharge<numtyp,acctyp>(),
	_allocated(false) {
	}

	template <class numtyp, class acctyp>
	CHARMMLongT::~CHARMMLong() {
	clear();
	}

	template <class numtyp, class acctyp>
	int CHARMMLongT::bytes_per_atom(const int max_nbors) const {
	return this->bytes_per_atom_atomic(max_nbors);
	}

	template <class numtyp, class acctyp>
	int CHARMMLongT::init(const int ntypes,
	double host_cut_bothsq, double **host_lj1,
	double host_lj2, double host_lj3,
	double host_lj4, double host_offset,
	double *host_special_lj, const int nlocal,
	const int nall, const int max_nbors,
	const int maxspecial, const double cell_size,
	const double gpu_split, FILE *_screen,
	double host_cut_ljsq, const double host_cut_coulsq,
	double *host_special_coul, const double qqrd2e,
	const double g_ewald, const double cut_lj_innersq,
	const double denom_lj, double **epsilon,
	double **sigma, const bool mix_arithmetic) {
	int success;
	success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
	_screen,charmm_long,"k_charmm_long");
	if (success!=0)
	return success;

	// If atom type constants fit in shared memory use fast kernel
	int lj_types=ntypes;
	shared_types=false;
	int max_bio_shared_types=this->device->max_bio_shared_types();
	if (this->_block_bio_size>=64 && mix_arithmetic &&
	lj_types<=max_bio_shared_types)
	shared_types=true;
	_lj_types=lj_types;

	// Allocate a host write buffer for data initialization
	int h_size=lj_types*lj_types;
	if (h_size<max_bio_shared_types)
	h_size=max_bio_shared_types;
	UCL_H_Vec<numtyp> host_write(h_size32,(this->ucl_device),
	UCL_WRITE_ONLY);
	for (int i=0; i<h_size*32; i++)
	host_write[i]=0.0;

	lj1.alloc(lj_typeslj_types,(this->ucl_device),UCL_READ_ONLY);
	this->atom->type_pack4(ntypes,lj_types,lj1,host_write,host_lj1,host_lj2,
	host_lj3,host_lj4);

	if (shared_types) {
	ljd.alloc(max_bio_shared_types,*(this->ucl_device),UCL_READ_ONLY);
	this->atom->self_pack2(ntypes,ljd,host_write,epsilon,sigma);
	}

	sp_lj.alloc(8,*(this->ucl_device),UCL_READ_ONLY);
	for (int i=0; i<4; i++) {
	host_write[i]=host_special_lj[i];
	host_write[i+4]=host_special_coul[i];
	}
	ucl_copy(sp_lj,host_write,8,false);

	_cut_bothsq = host_cut_bothsq;
	_cut_coulsq = host_cut_coulsq;
	_cut_ljsq = host_cut_ljsq;
	_cut_lj_innersq = cut_lj_innersq;
	_qqrd2e=qqrd2e;
	_g_ewald=g_ewald;
	_denom_lj=denom_lj;

	_allocated=true;
	this->_max_bytes=lj1.row_bytes()+ljd.row_bytes()+sp_lj.row_bytes();
	return 0;
	}

	template <class numtyp, class acctyp>
	void CHARMMLongT::clear() {
	if (!_allocated)
	return;
	_allocated=false;

	lj1.clear();
	ljd.clear();
	sp_lj.clear();
	this->clear_atomic();
	}

	template <class numtyp, class acctyp>
	double CHARMMLongT::host_memory_usage() const {
	return this->host_memory_usage_atomic()+sizeof(CHARMMLong<numtyp,acctyp>);
	}

	// ---------------------------------------------------------------------------
	// Calculate energies, forces, and torques
	// ---------------------------------------------------------------------------
	template <class numtyp, class acctyp>
	void CHARMMLongT::loop(const bool _eflag, const bool _vflag) {
	// Compute the block size and grid size to keep all cores busy
	const int BX=this->_block_bio_size;
	int eflag, vflag;
	if (_eflag)
	eflag=1;
	else
	eflag=0;

	if (_vflag)
	vflag=1;
	else
	vflag=0;

	int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
	(BX/this->_threads_per_atom)));

	int ainum=this->ans->inum();
	int nbor_pitch=this->nbor->nbor_pitch();
	this->time_pair.start();
	if (shared_types) {
	this->k_pair_fast.set_size(GX,BX);
	this->k_pair_fast.run(&this->atom->x, &ljd, &sp_lj,
	&this->nbor->dev_nbor, &this->_nbor_data->begin(),
	&this->ans->force, &this->ans->engv, &eflag,
	&vflag, &ainum, &nbor_pitch, &this->atom->q,
	&_cut_coulsq, &_qqrd2e, &_g_ewald, &_denom_lj,
	&_cut_bothsq, &_cut_ljsq, &_cut_lj_innersq,
	&this->_threads_per_atom);
	} else {
	this->k_pair.set_size(GX,BX);
	this->k_pair.run(&this->atom->x, &lj1, &_lj_types, &sp_lj,
	&this->nbor->dev_nbor, &this->_nbor_data->begin(),
	&this->ans->force, &this->ans->engv, &eflag, &vflag,
	&ainum, &nbor_pitch, &this->atom->q,
	&_cut_coulsq, &_qqrd2e, &_g_ewald, &_denom_lj,
	&_cut_bothsq, &_cut_ljsq, &_cut_lj_innersq,
	&this->_threads_per_atom);
	}
	this->time_pair.stop();
	}

	template class CHARMMLong<PRECISION,ACC_PRECISION>;

lal_charmm_long.cppNo OneTemporaryActions

File Metadata

lal_charmm_long.cppView Options

Event Timeline

lal_charmm_long.cpp
No OneTemporary
Actions

lal_charmm_long.cpp
View Options