Merge pull request #1735 from ndtrung81/gpu-mem-leak-init (b6313621) · Commits · 郑智淋 / lammps

lib/gpu/lal_base_atomic.cpp

+5 −8

Original line number	Original line	Diff line number	Diff line
	@@ -25,12 +25,16 @@ BaseAtomicT::BaseAtomic() : _compiled(false), _max_bytes(0) {
	device=&global_device;		device=&global_device;
	ans=new Answer<numtyp,acctyp>();		ans=new Answer<numtyp,acctyp>();
	nbor=new Neighbor();		nbor=new Neighbor();
			pair_program=NULL;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	BaseAtomicT::~BaseAtomic() {		BaseAtomicT::~BaseAtomic() {
	delete ans;		delete ans;
	delete nbor;		delete nbor;
			if (pair_program) delete pair_program;
			k_pair_fast.clear();
			k_pair.clear();
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -109,19 +113,11 @@ void BaseAtomicT::clear_atomic() {
	device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,		device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,
	_gpu_overhead,_driver_overhead,_threads_per_atom,screen);		_gpu_overhead,_driver_overhead,_threads_per_atom,screen);

	if (_compiled) {
	k_pair_fast.clear();
	k_pair.clear();
	delete pair_program;
	_compiled=false;
	}

	time_pair.clear();		time_pair.clear();
	hd_balancer.clear();		hd_balancer.clear();

	nbor->clear();		nbor->clear();
	ans->clear();		ans->clear();
	device->clear();
	}		}

	// ---------------------------------------------------------------------------		// ---------------------------------------------------------------------------
	@@ -276,6 +272,7 @@ void BaseAtomicT::compile_kernels(UCL_Device &dev, const void *pair_str,
	return;		return;

	std::string s_fast=std::string(kname)+"_fast";		std::string s_fast=std::string(kname)+"_fast";
			if (pair_program) delete pair_program;
	pair_program=new UCL_Program(dev);		pair_program=new UCL_Program(dev);
	pair_program->load_string(pair_str,device->compile_string().c_str());		pair_program->load_string(pair_str,device->compile_string().c_str());
	k_pair_fast.set_function(*pair_program,s_fast.c_str());		k_pair_fast.set_function(*pair_program,s_fast.c_str());

lib/gpu/lal_base_charge.cpp

+5 −8

Original line number	Original line	Diff line number	Diff line
	@@ -25,12 +25,16 @@ BaseChargeT::BaseCharge() : _compiled(false), _max_bytes(0) {
	device=&global_device;		device=&global_device;
	ans=new Answer<numtyp,acctyp>();		ans=new Answer<numtyp,acctyp>();
	nbor=new Neighbor();		nbor=new Neighbor();
			pair_program=NULL;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	BaseChargeT::~BaseCharge() {		BaseChargeT::~BaseCharge() {
	delete ans;		delete ans;
	delete nbor;		delete nbor;
			if (pair_program) delete pair_program;
			k_pair_fast.clear();
			k_pair.clear();
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -111,19 +115,11 @@ void BaseChargeT::clear_atomic() {
	device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,		device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,
	_gpu_overhead,_driver_overhead,_threads_per_atom,screen);		_gpu_overhead,_driver_overhead,_threads_per_atom,screen);

	if (_compiled) {
	k_pair_fast.clear();
	k_pair.clear();
	delete pair_program;
	_compiled=false;
	}

	time_pair.clear();		time_pair.clear();
	hd_balancer.clear();		hd_balancer.clear();

	nbor->clear();		nbor->clear();
	ans->clear();		ans->clear();
	device->clear();
	}		}

	// ---------------------------------------------------------------------------		// ---------------------------------------------------------------------------
	@@ -291,6 +287,7 @@ void BaseChargeT::compile_kernels(UCL_Device &dev, const void *pair_str,
	return;		return;

	std::string s_fast=std::string(kname)+"_fast";		std::string s_fast=std::string(kname)+"_fast";
			if (pair_program) delete pair_program;
	pair_program=new UCL_Program(dev);		pair_program=new UCL_Program(dev);
	pair_program->load_string(pair_str,device->compile_string().c_str());		pair_program->load_string(pair_str,device->compile_string().c_str());
	k_pair_fast.set_function(*pair_program,s_fast.c_str());		k_pair_fast.set_function(*pair_program,s_fast.c_str());

lib/gpu/lal_base_dipole.cpp

+5 −8

Original line number	Original line	Diff line number	Diff line
	@@ -25,12 +25,16 @@ BaseDipoleT::BaseDipole() : _compiled(false), _max_bytes(0) {
	device=&global_device;		device=&global_device;
	ans=new Answer<numtyp,acctyp>();		ans=new Answer<numtyp,acctyp>();
	nbor=new Neighbor();		nbor=new Neighbor();
			pair_program=NULL;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	BaseDipoleT::~BaseDipole() {		BaseDipoleT::~BaseDipole() {
	delete ans;		delete ans;
	delete nbor;		delete nbor;
			if (pair_program) delete pair_program;
			k_pair_fast.clear();
			k_pair.clear();
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -113,19 +117,11 @@ void BaseDipoleT::clear_atomic() {
	device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,		device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,
	_gpu_overhead,_driver_overhead,_threads_per_atom,screen);		_gpu_overhead,_driver_overhead,_threads_per_atom,screen);

	if (_compiled) {
	k_pair_fast.clear();
	k_pair.clear();
	delete pair_program;
	_compiled=false;
	}

	time_pair.clear();		time_pair.clear();
	hd_balancer.clear();		hd_balancer.clear();

	nbor->clear();		nbor->clear();
	ans->clear();		ans->clear();
	device->clear();
	}		}

	// ---------------------------------------------------------------------------		// ---------------------------------------------------------------------------
	@@ -299,6 +295,7 @@ void BaseDipoleT::compile_kernels(UCL_Device &dev, const void *pair_str,
	return;		return;

	std::string s_fast=std::string(kname)+"_fast";		std::string s_fast=std::string(kname)+"_fast";
			if (pair_program) delete pair_program;
	pair_program=new UCL_Program(dev);		pair_program=new UCL_Program(dev);
	pair_program->load_string(pair_str,device->compile_string().c_str());		pair_program->load_string(pair_str,device->compile_string().c_str());
	k_pair_fast.set_function(*pair_program,s_fast.c_str());		k_pair_fast.set_function(*pair_program,s_fast.c_str());

lib/gpu/lal_base_dpd.cpp

+5 −8

Original line number	Original line	Diff line number	Diff line
	@@ -25,12 +25,16 @@ BaseDPDT::BaseDPD() : _compiled(false), _max_bytes(0) {
	device=&global_device;		device=&global_device;
	ans=new Answer<numtyp,acctyp>();		ans=new Answer<numtyp,acctyp>();
	nbor=new Neighbor();		nbor=new Neighbor();
			pair_program=NULL;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	BaseDPDT::~BaseDPD() {		BaseDPDT::~BaseDPD() {
	delete ans;		delete ans;
	delete nbor;		delete nbor;
			if (pair_program) delete pair_program;
			k_pair_fast.clear();
			k_pair.clear();
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -112,19 +116,11 @@ void BaseDPDT::clear_atomic() {
	device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,		device->output_times(time_pair,ans,nbor,avg_split,_max_bytes+_max_an_bytes,
	_gpu_overhead,_driver_overhead,_threads_per_atom,screen);		_gpu_overhead,_driver_overhead,_threads_per_atom,screen);

	if (_compiled) {
	k_pair_fast.clear();
	k_pair.clear();
	delete pair_program;
	_compiled=false;
	}

	time_pair.clear();		time_pair.clear();
	hd_balancer.clear();		hd_balancer.clear();

	nbor->clear();		nbor->clear();
	ans->clear();		ans->clear();
	device->clear();
	}		}

	// ---------------------------------------------------------------------------		// ---------------------------------------------------------------------------
	@@ -297,6 +293,7 @@ void BaseDPDT::compile_kernels(UCL_Device &dev, const void *pair_str,
	return;		return;

	std::string s_fast=std::string(kname)+"_fast";		std::string s_fast=std::string(kname)+"_fast";
			if (pair_program) delete pair_program;
	pair_program=new UCL_Program(dev);		pair_program=new UCL_Program(dev);
	pair_program->load_string(pair_str,device->compile_string().c_str());		pair_program->load_string(pair_str,device->compile_string().c_str());
	k_pair_fast.set_function(*pair_program,s_fast.c_str());		k_pair_fast.set_function(*pair_program,s_fast.c_str());

lib/gpu/lal_base_ellipsoid.cpp

+17 −15

Original line number	Original line	Diff line number	Diff line
	@@ -33,12 +33,25 @@ BaseEllipsoidT::BaseEllipsoid() : _compiled(false), _max_bytes(0) {
	device=&global_device;		device=&global_device;
	ans=new Answer<numtyp,acctyp>();		ans=new Answer<numtyp,acctyp>();
	nbor=new Neighbor();		nbor=new Neighbor();
			nbor_program=NULL;
			ellipsoid_program=NULL;
			lj_program=NULL;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	BaseEllipsoidT::~BaseEllipsoid() {		BaseEllipsoidT::~BaseEllipsoid() {
	delete ans;		delete ans;
	delete nbor;		delete nbor;
			k_nbor_fast.clear();
			k_nbor.clear();
			k_ellipsoid.clear();
			k_ellipsoid_sphere.clear();
			k_sphere_ellipsoid.clear();
			k_lj_fast.clear();
			k_lj.clear();
			if (nbor_program) delete nbor_program;
			if (ellipsoid_program) delete ellipsoid_program;
			if (lj_program) delete lj_program;
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -146,20 +159,6 @@ void BaseEllipsoidT::clear_base() {
	output_times();		output_times();
	host_olist.clear();		host_olist.clear();

	if (_compiled) {
	k_nbor_fast.clear();
	k_nbor.clear();
	k_ellipsoid.clear();
	k_ellipsoid_sphere.clear();
	k_sphere_ellipsoid.clear();
	k_lj_fast.clear();
	k_lj.clear();
	delete nbor_program;
	delete ellipsoid_program;
	delete lj_program;
	_compiled=false;
	}

	time_nbor1.clear();		time_nbor1.clear();
	time_ellipsoid.clear();		time_ellipsoid.clear();
	time_nbor2.clear();		time_nbor2.clear();
	@@ -171,7 +170,6 @@ void BaseEllipsoidT::clear_base() {

	nbor->clear();		nbor->clear();
	ans->clear();		ans->clear();
	device->clear();
	}		}

	template <class numtyp, class acctyp>		template <class numtyp, class acctyp>
	@@ -437,6 +435,7 @@ int** BaseEllipsoidT::compute(const int ago, const int inum_full, const int nall
	ans->copy_answers(eflag,vflag,eatom,vatom);		ans->copy_answers(eflag,vflag,eatom,vatom);
	device->add_ans_object(ans);		device->add_ans_object(ans);
	hd_balancer.stop_timer();		hd_balancer.stop_timer();

	return nbor->host_jlist.begin()-host_start;		return nbor->host_jlist.begin()-host_start;
	}		}

	@@ -462,18 +461,21 @@ void BaseEllipsoidT::compile_kernels(UCL_Device &dev,

	std::string flags=device->compile_string();		std::string flags=device->compile_string();

			if (nbor_program) delete nbor_program;
	nbor_program=new UCL_Program(dev);		nbor_program=new UCL_Program(dev);
	nbor_program->load_string(ellipsoid_nbor,flags.c_str());		nbor_program->load_string(ellipsoid_nbor,flags.c_str());
	k_nbor_fast.set_function(*nbor_program,"kernel_nbor_fast");		k_nbor_fast.set_function(*nbor_program,"kernel_nbor_fast");
	k_nbor.set_function(*nbor_program,"kernel_nbor");		k_nbor.set_function(*nbor_program,"kernel_nbor");
	neigh_tex.get_texture(*nbor_program,"pos_tex");		neigh_tex.get_texture(*nbor_program,"pos_tex");

			if (ellipsoid_program) delete ellipsoid_program;
	ellipsoid_program=new UCL_Program(dev);		ellipsoid_program=new UCL_Program(dev);
	ellipsoid_program->load_string(ellipsoid_string,flags.c_str());		ellipsoid_program->load_string(ellipsoid_string,flags.c_str());
	k_ellipsoid.set_function(*ellipsoid_program,kname);		k_ellipsoid.set_function(*ellipsoid_program,kname);
	pos_tex.get_texture(*ellipsoid_program,"pos_tex");		pos_tex.get_texture(*ellipsoid_program,"pos_tex");
	quat_tex.get_texture(*ellipsoid_program,"quat_tex");		quat_tex.get_texture(*ellipsoid_program,"quat_tex");

			if (lj_program) delete lj_program;
	lj_program=new UCL_Program(dev);		lj_program=new UCL_Program(dev);
	lj_program->load_string(lj_string,flags.c_str());		lj_program->load_string(lj_string,flags.c_str());
	k_sphere_ellipsoid.set_function(*lj_program,s_sphere_ellipsoid.c_str());		k_sphere_ellipsoid.set_function(*lj_program,s_sphere_ellipsoid.c_str());

Admin message