Merge pull request #26 from timattox/USER-DPD_GBhacks_optimize (046900cd) · Commits · 郑智淋 / lammps

src/KOKKOS/fix_rx_kokkos.cpp

+32 −20

Original line number	Diff line number	Diff line
		@@ -79,6 +79,17 @@ FixRxKokkos<DeviceType>::~FixRxKokkos()
		{
		//printf("Inside FixRxKokkos::~FixRxKokkos copymode= %d\n", copymode);
		if (copymode) return;

		if (localTempFlag)
		memory->destroy_kokkos(k_dpdThetaLocal, dpdThetaLocal);

		memory->destroy_kokkos(k_sumWeights, sumWeights);
		//memory->destroy_kokkos(k_sumWeights);

		//delete [] scratchSpace;
		memory->destroy_kokkos(d_scratchSpace);

		memory->destroy_kokkos(k_cutsq);
		}

		/* ---------------------------------------------------------------------- */
		@@ -1433,9 +1444,12 @@ void FixRxKokkos<DeviceType>::solve_reactions(const int vflag, const bool isPreF
		{
		const int count = nlocal + (newton_pair ? nghost : 0);

		if (count > k_dpdThetaLocal.d_view.dimension_0()) {
		memory->destroy_kokkos (k_dpdThetaLocal, dpdThetaLocal);
		memory->create_kokkos (k_dpdThetaLocal, dpdThetaLocal, count, "FixRxKokkos::dpdThetaLocal");
		this->d_dpdThetaLocal = k_dpdThetaLocal.d_view;
		this->h_dpdThetaLocal = k_dpdThetaLocal.h_view;
		}

		const int neighflag = lmp->kokkos->neighflag;

		@@ -1527,7 +1541,10 @@ void FixRxKokkos<DeviceType>::solve_reactions(const int vflag, const bool isPreF
		//double scratchSpace = new double[ scratchSpaceSize nlocal ];

		//typename ArrayTypes<DeviceType>::t_double_1d d_scratchSpace("d_scratchSpace", scratchSpaceSize * nlocal);
		if (nlocal*scratchSpaceSize > d_scratchSpace.dimension_0()) {
		memory->destroy_kokkos (d_scratchSpace);
		memory->create_kokkos (d_scratchSpace, nlocal*scratchSpaceSize, "FixRxKokkos::d_scratchSpace");
		}

		#if 0
		Kokkos::parallel_reduce( nlocal, LAMMPS_LAMBDA(int i, CounterType &counter)
		@@ -1630,9 +1647,6 @@ void FixRxKokkos<DeviceType>::solve_reactions(const int vflag, const bool isPreF
		Kokkos::parallel_reduce( Kokkos::RangePolicy<DeviceType, Tag_FixRxKokkos_solveSystems<false> >(0,nlocal), *this, TotalCounters);
		#endif

		//delete [] scratchSpace;
		memory->destroy_kokkos (d_scratchSpace);

		TimerType timer_ODE = getTimeStamp();

		// Check the error flag for any failures.
		@@ -1651,9 +1665,6 @@ void FixRxKokkos<DeviceType>::solve_reactions(const int vflag, const bool isPreF

		atomKK->modified ( Host, DVECTOR_MASK );

		if (localTempFlag)
		memory->destroy_kokkos(k_dpdThetaLocal, dpdThetaLocal);

		TimerType timer_stop = getTimeStamp();

		double time_ODE = getElapsedTime(timer_localTemperature, timer_ODE);
		@@ -2012,8 +2023,11 @@ void FixRxKokkos<DeviceType>::computeLocalTemperature()
		const int ntypes = atom->ntypes;

		//memory->create_kokkos (k_cutsq, h_cutsq, ntypes+1, ntypes+1, "pair:cutsq");
		if (ntypes+1 > k_cutsq.dimension_0()) {
		memory->destroy_kokkos (k_cutsq);
		memory->create_kokkos (k_cutsq, ntypes+1, ntypes+1, "FixRxKokkos::k_cutsq");
		d_cutsq = k_cutsq.template view<DeviceType>();
		}

		for (int i = 1; i <= ntypes; ++i)
		for (int j = i; j <= ntypes; ++j)
		@@ -2030,9 +2044,12 @@ void FixRxKokkos<DeviceType>::computeLocalTemperature()
		int sumWeightsCt = nlocal + (NEWTON_PAIR ? nghost : 0);

		//memory->create_kokkos (k_sumWeights, sumWeights, sumWeightsCt, "FixRxKokkos::sumWeights");
		if (sumWeightsCt > k_sumWeights.d_view.dimension_0()) {
		memory->destroy_kokkos(k_sumWeights, sumWeights);
		memory->create_kokkos (k_sumWeights, sumWeightsCt, "FixRxKokkos::sumWeights");
		d_sumWeights = k_sumWeights.d_view;
		h_sumWeights = k_sumWeights.h_view;
		}

		// Initialize the accumulator to zero ...
		//Kokkos::parallel_for (sumWeightsCt,
		@@ -2165,11 +2182,6 @@ void FixRxKokkos<DeviceType>::computeLocalTemperature()
		Kokkos::parallel_for (Kokkos::RangePolicy<DeviceType, Tag_FixRxKokkos_2ndPairOperator<WT_FLAG, LOCAL_TEMP_FLAG> >(0, nlocal), *this);
		#endif

		// Clean up the local kokkos data.
		//memory->destroy_kokkos(k_cutsq, h_cutsq);
		memory->destroy_kokkos(k_cutsq);
		//memory->destroy_kokkos(k_sumWeights, sumWeights);
		memory->destroy_kokkos(k_sumWeights);
		}

		/* ---------------------------------------------------------------------- */

src/KOKKOS/pair_multi_lucy_rx_kokkos.cpp

+6 −4

Original line number	Diff line number	Diff line
		@@ -180,10 +180,12 @@ void PairMultiLucyRXKokkos<DeviceType>::compute_style(int eflag_in, int vflag_in

		{
		const int ntotal = nlocal + nghost;
		if (ntotal > d_mixWtSite1.dimension_0()) {
		d_mixWtSite1old = typename AT::t_float_1d("PairMultiLucyRX::mixWtSite1old",ntotal);
		d_mixWtSite2old = typename AT::t_float_1d("PairMultiLucyRX::mixWtSite2old",ntotal);
		d_mixWtSite1 = typename AT::t_float_1d("PairMultiLucyRX::mixWtSite1",ntotal);
		d_mixWtSite2 = typename AT::t_float_1d("PairMultiLucyRX::mixWtSite2",ntotal);
		}

		Kokkos::parallel_for(Kokkos::RangePolicy<DeviceType, TagPairMultiLucyRXgetMixingWeights>(0,ntotal),*this);
		}

src/KOKKOS/pair_table_rx_kokkos.cpp

+14 −12

Original line number	Diff line number	Diff line
		@@ -534,10 +534,10 @@ static void compute_all_items(
		typename ArrayTypes<DeviceType>::t_int_1d_const d_numneigh,
		typename ArrayTypes<DeviceType>::t_x_array_randomread x,
		typename ArrayTypes<DeviceType>::t_int_1d_randomread type,
		Kokkos::View<double*, DeviceType> mixWtSite1old,
		Kokkos::View<double*, DeviceType> mixWtSite2old,
		Kokkos::View<double*, DeviceType> mixWtSite1,
		Kokkos::View<double*, DeviceType> mixWtSite2,
		Kokkos::View<double*, DeviceType> const& mixWtSite1old,
		Kokkos::View<double*, DeviceType> const& mixWtSite2old,
		Kokkos::View<double*, DeviceType> const& mixWtSite1,
		Kokkos::View<double*, DeviceType> const& mixWtSite2,
		Few<int, 4> special_lj,
		Few<Few<F_FLOAT, MAX_TYPES_STACKPARAMS+1>, MAX_TYPES_STACKPARAMS+1> m_cutsq,
		typename ArrayTypes<DeviceType>::t_ffloat_2d d_cutsq,
		@@ -597,10 +597,10 @@ static void getAllMixingWeights(
		int nspecies,
		int isite1, int isite2,
		bool fractionalWeighting,
		Kokkos::View<double*, DeviceType> mixWtSite1old,
		Kokkos::View<double*, DeviceType> mixWtSite2old,
		Kokkos::View<double*, DeviceType> mixWtSite1,
		Kokkos::View<double*, DeviceType> mixWtSite2) {
		Kokkos::View<double*, DeviceType> const& mixWtSite1old,
		Kokkos::View<double*, DeviceType> const& mixWtSite2old,
		Kokkos::View<double*, DeviceType> const& mixWtSite1,
		Kokkos::View<double*, DeviceType> const& mixWtSite2) {
		Kokkos::parallel_for(ntotal,
		LAMMPS_LAMBDA(int i) {
		getMixingWeights<DeviceType>(dvector,nspecies,isite1,isite2,fractionalWeighting,
		@@ -651,10 +651,12 @@ void PairTableRXKokkos<DeviceType>::compute_style(int eflag_in, int vflag_in)
		// loop over neighbors of my atoms

		const int ntotal = atom->nlocal + atom->nghost;
		auto mixWtSite1old = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite1old", ntotal);
		auto mixWtSite2old = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite2old", ntotal);
		auto mixWtSite1 = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite1", ntotal);
		auto mixWtSite2 = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite2", ntotal);
		if (ntotal > mixWtSite1.dimension_0()) {
		mixWtSite1old = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite1old", ntotal);
		mixWtSite2old = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite2old", ntotal);
		mixWtSite1 = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite1", ntotal);
		mixWtSite2 = Kokkos::View<double*, DeviceType>("PairTableRXKokkos::mixWtSite2", ntotal);
		}

		getAllMixingWeights(ntotal, atomKK->k_dvector.template view<DeviceType>(),
		nspecies, isite1, isite2, fractionalWeighting,

src/KOKKOS/pair_table_rx_kokkos.h

+5 −0

Original line number	Diff line number	Diff line
		@@ -96,6 +96,11 @@ class PairTableRXKokkos : public PairTable {

		/* PairTableRX members */

		Kokkos::View<double*, DeviceType> mixWtSite1old;
		Kokkos::View<double*, DeviceType> mixWtSite2old;
		Kokkos::View<double*, DeviceType> mixWtSite1;
		Kokkos::View<double*, DeviceType> mixWtSite2;

		int nspecies;
		char site1, site2;
		int isite1, isite2;

Admin message