Cleaned up 3-body gpu styles, and fixed a bug for tersoff/zbl/gpu. (3d1d0c58) · Commits · 郑智淋 / lammps

lib/gpu/lal_base_three.cpp

+15 −26

Original line number	Diff line number	Diff line
		@@ -20,7 +20,7 @@ using namespace LAMMPS_AL;
		extern Device<PRECISION,ACC_PRECISION> global_device;

		template <class numtyp, class acctyp>
		BaseThreeT::BaseThree() : _compiled(false), _max_bytes(0), _short_nbor(false) {
		BaseThreeT::BaseThree() : _compiled(false), _max_bytes(0) {
		device=&global_device;
		ans=new Answer<numtyp,acctyp>();
		nbor=new Neighbor();
		@@ -73,6 +73,7 @@ int BaseThreeT::init_three(const int nlocal, const int nall,
		if (_threads_per_atom>1 && gpu_nbor==0) { // neigh no and tpa > 1
		nbor->packing(true);
		_nbor_data=&(nbor->dev_packed);
		_threads_per_atom = 1; // enforce tpa = 1 for now
		} else // neigh yes or tpa == 1
		_nbor_data=&(nbor->dev_nbor);
		if (_threads_per_atom*_threads_per_atom>device->warp_size())
		@@ -113,14 +114,10 @@ int BaseThreeT::init_three(const int nlocal, const int nall,
		_max_an_bytes+=ans2->gpu_bytes();
		#endif

		// if short neighbor list is supported
		if (short_nbor) {
		_short_nbor = true;
		int ef_nall=nall;
		if (ef_nall==0)
		ef_nall=2000;
		dev_short_nbor.alloc(ef_nall(2+max_nbors),(this->ucl_device),UCL_READ_WRITE);
		}

		return 0;
		}
		@@ -269,15 +266,11 @@ void BaseThreeT::compute(const int f_ago, const int inum_full, const int nall,
		hd_balancer.start_timer();
		atom->add_x_data(host_x,host_type);

		// if short neighbor list is supported
		if (_short_nbor) {

		// re-allocate dev_short_nbor if necessary
		if (nall*(2+_max_nbors) > dev_short_nbor.cols()) {
		int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
		dev_short_nbor.resize((2+_max_nbors)*_nmax);
		}
		}

		// _ainum to be used in loop() for short neighbor list build
		_ainum = nlist;
		@@ -342,15 +335,11 @@ int ** BaseThreeT::compute(const int ago, const int inum_full,
		*ilist=nbor->host_ilist.begin();
		*jnum=nbor->host_acc.begin();

		// if short neighbor list is supported
		if (_short_nbor) {

		// re-allocate dev_short_nbor if necessary
		if (nall*(2+_max_nbors) > dev_short_nbor.cols()) {
		int _nmax=static_cast<int>(static_cast<double>(nall)*1.10);
		dev_short_nbor.resize((2+_max_nbors)*_nmax);
		}
		}

		// _ainum to be used in loop() for short neighbor list build
		_ainum = nall;
		@@ -394,7 +383,7 @@ void BaseThreeT::compile_kernels(UCL_Device &dev, const void *pair_str,
		k_three_end.set_function(*pair_program,three_end);
		k_three_end_vatom.set_function(*pair_program,vatom_name.c_str());
		k_pair.set_function(*pair_program,two);
		if (short_nbor) k_short_nbor.set_function(*pair_program,short_nbor);
		k_short_nbor.set_function(*pair_program,short_nbor);
		pos_tex.get_texture(*pair_program,"pos_tex");

		#ifdef THREE_CONCURRENT

lib/gpu/lal_base_three.h

+1 −1

Original line number	Diff line number	Diff line
		@@ -199,7 +199,7 @@ class BaseThree {
		UCL_Texture pos_tex;

		protected:
		bool _compiled,_short_nbor;
		bool _compiled;
		int _block_pair, _block_size, _threads_per_atom, _end_command_queue;
		int _gpu_nbor;
		double _max_bytes, _max_an_bytes;

lib/gpu/lal_tersoff_zbl.cpp

+1 −1

Original line number	Diff line number	Diff line
		@@ -293,7 +293,7 @@ void TersoffZT::loop(const bool _eflag, const bool _vflag, const int evatom) {
		(BX/(JTHREADS*KTHREADS))));

		this->k_zeta.set_size(GX,BX);
		this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &cutsq,
		this->k_zeta.run(&this->atom->x, &ts1, &ts2, &ts3, &ts4, &ts5, &ts6, &cutsq,
		&map, &elem2param, &_nelements, &_nparams, &_zetaij,
		&this->nbor->dev_nbor, &this->_nbor_data->begin(),
		&this->dev_short_nbor,

lib/gpu/lal_tersoff_mod.cu

+1 −1

File changed.

Contains only whitespace changes.

lib/gpu/lal_tersoff_zbl.cu

+1 −1

File changed.

Contains only whitespace changes.

Admin message