Added ReductionView to Kokkos ReaxFF (54f24930) · Commits · 郑智淋 / lammps

src/KOKKOS/fix_qeq_reax_kokkos.cpp

+33 −41

Original line number	Diff line number	Diff line
		@@ -254,12 +254,18 @@ void FixQEqReaxKokkos<DeviceType>::pre_force(int vflag)
		k_t.template modify<LMPHostType>();
		k_t.template sync<DeviceType>();

		if (neighflag != FULL)
		red_o = Kokkos::Experimental::create_reduction_view<> (k_o.d_view); // allocate duplicated memory

		// 1st cg solve over b_s, s
		cg_solve1();

		// 2nd cg solve over b_t, t
		cg_solve2();

		if (neighflag != FULL)
		red_o = decltype(red_o)(); // free duplicated memory

		// calculate_Q();
		calculate_q();

		@@ -494,13 +500,9 @@ void FixQEqReaxKokkos<DeviceType>::cg_solve1()
		Kokkos::parallel_for(inum,sparse12_functor);
		if (neighflag != FULL) {
		Kokkos::parallel_for(Kokkos::RangePolicy<DeviceType,TagZeroQGhosts>(nlocal,nlocal+atom->nghost),*this);
		if (neighflag == HALF) {
		FixQEqReaxKokkosSparse13Functor<DeviceType,HALF> sparse13_functor(this);
		Kokkos::parallel_for(inum,sparse13_functor);
		} else {
		FixQEqReaxKokkosSparse13Functor<DeviceType,HALFTHREAD> sparse13_functor(this);
		Kokkos::parallel_for(inum,sparse13_functor);
		}
		Kokkos::Experimental::contribute(k_o.d_view, red_o);
		} else {
		Kokkos::parallel_for(Kokkos::TeamPolicy <DeviceType, TagSparseMatvec1> (inum, teamsize), *this);
		}
		@@ -548,18 +550,14 @@ void FixQEqReaxKokkos<DeviceType>::cg_solve1()
		Kokkos::parallel_for(inum,sparse22_functor);
		if (neighflag != FULL) {
		Kokkos::parallel_for(Kokkos::RangePolicy<DeviceType,TagZeroQGhosts>(nlocal,nlocal+atom->nghost),*this);
		if (neighflag == HALF) {
		FixQEqReaxKokkosSparse23Functor<DeviceType,HALF> sparse23_functor(this);
		Kokkos::parallel_for(inum,sparse23_functor);
		} else {
		red_o.reset_except(d_o);
		FixQEqReaxKokkosSparse23Functor<DeviceType,HALFTHREAD> sparse23_functor(this);
		Kokkos::parallel_for(inum,sparse23_functor);
		}
		Kokkos::Experimental::contribute(k_o.d_view, red_o);
		} else {
		Kokkos::parallel_for(Kokkos::TeamPolicy <DeviceType, TagSparseMatvec2> (inum, teamsize), *this);
		}


		if (neighflag != FULL) {
		k_o.template modify<DeviceType>();
		k_o.template sync<LMPHostType>();
		@@ -625,13 +623,10 @@ void FixQEqReaxKokkos<DeviceType>::cg_solve2()
		Kokkos::parallel_for(inum,sparse32_functor);
		if (neighflag != FULL) {
		Kokkos::parallel_for(Kokkos::RangePolicy<DeviceType,TagZeroQGhosts>(nlocal,nlocal+atom->nghost),*this);
		if (neighflag == HALF) {
		FixQEqReaxKokkosSparse33Functor<DeviceType,HALF> sparse33_functor(this);
		Kokkos::parallel_for(inum,sparse33_functor);
		} else {
		red_o.reset_except(d_o);
		FixQEqReaxKokkosSparse33Functor<DeviceType,HALFTHREAD> sparse33_functor(this);
		Kokkos::parallel_for(inum,sparse33_functor);
		}
		Kokkos::Experimental::contribute(k_o.d_view, red_o);
		} else {
		Kokkos::parallel_for(Kokkos::TeamPolicy <DeviceType, TagSparseMatvec3> (inum, teamsize), *this);
		}
		@@ -679,13 +674,10 @@ void FixQEqReaxKokkos<DeviceType>::cg_solve2()
		Kokkos::parallel_for(inum,sparse22_functor);
		if (neighflag != FULL) {
		Kokkos::parallel_for(Kokkos::RangePolicy<DeviceType,TagZeroQGhosts>(nlocal,nlocal+atom->nghost),*this);
		if (neighflag == HALF) {
		FixQEqReaxKokkosSparse23Functor<DeviceType,HALF> sparse23_functor(this);
		Kokkos::parallel_for(inum,sparse23_functor);
		} else {
		red_o.reset_except(d_o);
		FixQEqReaxKokkosSparse23Functor<DeviceType,HALFTHREAD> sparse23_functor(this);
		Kokkos::parallel_for(inum,sparse23_functor);
		}
		Kokkos::Experimental::contribute(k_o.d_view, red_o);
		} else {
		Kokkos::parallel_for(Kokkos::TeamPolicy <DeviceType, TagSparseMatvec2> (inum, teamsize), *this);
		}
		@@ -797,8 +789,8 @@ template<int NEIGHFLAG>
		KOKKOS_INLINE_FUNCTION
		void FixQEqReaxKokkos<DeviceType>::sparse13_item(int ii) const
		{
		// The q array is atomic for Half/Thread neighbor style
		Kokkos::View<F_FLOAT*, typename DAT::t_float_1d::array_layout,DeviceType,Kokkos::MemoryTraits<AtomicF<NEIGHFLAG>::value> > a_o = d_o;
		// The q array is duplicated for OpenMP, atomic for CUDA, and neither for Serial
		auto a_o = red_o.access();

		const int i = d_ilist[ii];
		if (mask[i] & groupbit) {
		@@ -806,9 +798,9 @@ void FixQEqReaxKokkos<DeviceType>::sparse13_item(int ii) const
		for(int jj = d_firstnbr[i]; jj < d_firstnbr[i] + d_numnbrs[i]; jj++) {
		const int j = d_jlist(jj);
		tmp += d_val(jj) * d_s[j];
		a_o[j] += d_val(jj) * d_s[i];
		a_o(j) += d_val(jj) * d_s[i];
		}
		a_o[i] += tmp;
		a_o(i) += tmp;
		}
		}

		@@ -849,8 +841,8 @@ template<int NEIGHFLAG>
		KOKKOS_INLINE_FUNCTION
		void FixQEqReaxKokkos<DeviceType>::sparse23_item(int ii) const
		{
		// The q array is atomic for Half/Thread neighbor style
		Kokkos::View<F_FLOAT*, typename DAT::t_float_1d::array_layout,DeviceType,Kokkos::MemoryTraits<AtomicF<NEIGHFLAG>::value> > a_o = d_o;
		// The q array is duplicated for OpenMP, atomic for CUDA, and neither for Serial
		auto a_o = red_o.access();

		const int i = d_ilist[ii];
		if (mask[i] & groupbit) {
		@@ -858,9 +850,9 @@ void FixQEqReaxKokkos<DeviceType>::sparse23_item(int ii) const
		for(int jj = d_firstnbr[i]; jj < d_firstnbr[i] + d_numnbrs[i]; jj++) {
		const int j = d_jlist(jj);
		tmp += d_val(jj) * d_d[j];
		a_o[j] += d_val(jj) * d_d[i];
		a_o(j) += d_val(jj) * d_d[i];
		}
		a_o[i] += tmp;
		a_o(i) += tmp;
		}
		}

		@@ -908,8 +900,8 @@ template<int NEIGHFLAG>
		KOKKOS_INLINE_FUNCTION
		void FixQEqReaxKokkos<DeviceType>::sparse33_item(int ii) const
		{
		// The q array is atomic for Half/Thread neighbor style
		Kokkos::View<F_FLOAT*, typename DAT::t_float_1d::array_layout,DeviceType,Kokkos::MemoryTraits<AtomicF<NEIGHFLAG>::value> > a_o = d_o;
		// The q array is duplicated for OpenMP, atomic for CUDA, and neither for Serial
		auto a_o = red_o.access();

		const int i = d_ilist[ii];
		if (mask[i] & groupbit) {
		@@ -917,9 +909,9 @@ void FixQEqReaxKokkos<DeviceType>::sparse33_item(int ii) const
		for(int jj = d_firstnbr[i]; jj < d_firstnbr[i] + d_numnbrs[i]; jj++) {
		const int j = d_jlist(jj);
		tmp += d_val(jj) * d_t[j];
		a_o[j] += d_val(jj) * d_t[i];
		a_o(j) += d_val(jj) * d_t[i];
		}
		a_o[i] += tmp;
		a_o(i) += tmp;
		}
		}

src/KOKKOS/fix_qeq_reax_kokkos.h

+3 −0

Original line number	Diff line number	Diff line
		@@ -22,6 +22,7 @@ FixStyle(qeq/reax/kk/host,FixQEqReaxKokkos<LMPHostType>)
		#ifndef LMP_FIX_QEQ_REAX_KOKKOS_H
		#define LMP_FIX_QEQ_REAX_KOKKOS_H

		#include <Kokkos_ReductionView.hpp>
		#include "fix_qeq_reax.h"
		#include "kokkos_type.h"
		#include "neigh_list.h"
		@@ -192,6 +193,8 @@ class FixQEqReaxKokkos : public FixQEqReax {
		HAT::t_ffloat_2d h_s_hist, h_t_hist;
		typename AT::t_ffloat_2d_randomread r_s_hist, r_t_hist;

		Kokkos::Experimental::ReductionView<F_FLOAT*, typename AT::t_ffloat_1d::array_layout, DeviceType> red_o;

		void init_shielding_k();
		void init_hist();
		void allocate_matrix();

src/KOKKOS/pair_kokkos.h

+14 −1

Original line number	Diff line number	Diff line
		@@ -19,9 +19,10 @@
		#define LMP_PAIR_KOKKOS_H

		#include "Kokkos_Macros.hpp"
		#include "Kokkos_Vectorization.hpp"
		#include "Kokkos_ReductionView.hpp"
		#include "pair.h"
		#include "neigh_list_kokkos.h"
		#include "Kokkos_Vectorization.hpp"

		namespace LAMMPS_NS {

		@@ -58,6 +59,18 @@ struct AtomicF<HALFTHREAD> {
		enum {value = Kokkos::Atomic\|Kokkos::Unmanaged};
		};

		// Determine memory traits for force array
		// Do atomic trait when running HALFTHREAD neighbor list style
		template<int NEIGHFLAG>
		struct AtomicDup {
		enum {value = Kokkos::Experimental::ReductionNonAtomic};
		};

		template<>
		struct AtomicDup<HALFTHREAD> {
		enum {value = Kokkos::Experimental::ReductionAtomic};
		};

		//Specialisation for Neighborlist types Half, HalfThread, Full
		template <class PairStyle, int NEIGHFLAG, bool STACKPARAMS, class Specialisation = void>
		struct PairComputeFunctor {

src/KOKKOS/pair_reaxc_kokkos.cpp

+138 −31

File changed.

Preview size limit exceeded, changes collapsed.

src/KOKKOS/pair_reaxc_kokkos.h

+11 −0

Original line number	Diff line number	Diff line
		@@ -24,6 +24,7 @@ PairStyle(reax/c/kk/host,PairReaxCKokkos<LMPHostType>)
		#define LMP_PAIR_REAXC_KOKKOS_H

		#include <stdio.h>
		#include <Kokkos_ReductionView.hpp>
		#include "pair_kokkos.h"
		#include "pair_reaxc.h"
		#include "neigh_list_kokkos.h"
		@@ -411,6 +412,16 @@ class PairReaxCKokkos : public PairReaxC {
		typename AT::t_ffloat_2d_dl d_C1dbopi2, d_C2dbopi2, d_C3dbopi2, d_C4dbopi2;
		typename AT::t_ffloat_2d_dl d_Cdbo, d_Cdbopi, d_Cdbopi2, d_dDeltap_self;

		Kokkos::Experimental::ReductionView<F_FLOAT*, typename DAT::t_float_1d::array_layout,DeviceType> dup_total_bo;
		Kokkos::Experimental::ReductionView<F_FLOAT*, typename DAT::t_float_1d::array_layout,DeviceType> dup_CdDelta;
		Kokkos::Experimental::ReductionView<E_FLOAT*, typename DAT::t_efloat_1d::array_layout,DeviceType> dup_eatom;
		Kokkos::Experimental::ReductionView<F_FLOAT*[3], typename DAT::t_f_array::array_layout,DeviceType> dup_f;
		Kokkos::Experimental::ReductionView<F_FLOAT*[6], typename DAT::t_virial_array::array_layout,DeviceType> dup_vatom;
		Kokkos::Experimental::ReductionView<F_FLOAT**, typename DAT::t_ffloat_2d_dl::array_layout,DeviceType> dup_dDeltap_self;
		Kokkos::Experimental::ReductionView<F_FLOAT**, typename DAT::t_ffloat_2d_dl::array_layout,DeviceType> dup_Cdbo;
		Kokkos::Experimental::ReductionView<F_FLOAT**, typename DAT::t_ffloat_2d_dl::array_layout,DeviceType> dup_Cdbopi;
		Kokkos::Experimental::ReductionView<F_FLOAT**, typename DAT::t_ffloat_2d_dl::array_layout,DeviceType> dup_Cdbopi2;

		typedef Kokkos::DualView<F_FLOAT**[7],typename DeviceType::array_layout,DeviceType> tdual_ffloat_2d_n7;
		typedef typename tdual_ffloat_2d_n7::t_dev_const_randomread t_ffloat_2d_n7_randomread;
		typedef typename tdual_ffloat_2d_n7::t_host t_host_ffloat_2d_n7;

Admin message