ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reax: Preprocessing optimizations to ComputeAngular,Torsion. Modularity boosts + memory reductions for BuildLists

Browse Source
This commit is contained in:
Evan Weinberg
2022-03-30 08:23:41 -07:00
parent b92d117524
commit 14f54aae40
3 changed files with 1425 additions and 243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/KOKKOS/kokkos_type.h
View File

@ -1214,6 +1214,15 @@ struct params_lj_coul {
F_FLOAT cut_ljsq,cut_coulsq,lj1,lj2,lj3,lj4,offset;
};
#ifdef OPT_ANGULAR_TORSION
// ReaxFF
struct alignas(4 * sizeof(int)) reax_int4 {
int i0, i1, i2, i3;
};
#endif
// Pair SNAP
#define SNAP_KOKKOS_REAL double

1543
src/KOKKOS/pair_reaxff_kokkos.cpp
View File

File diff suppressed because it is too large Load Diff

116
src/KOKKOS/pair_reaxff_kokkos.h
View File

@ -90,6 +90,26 @@ struct TagPairReaxComputeMulti1{};
template<int NEIGHFLAG, int EFLAG>
struct TagPairReaxComputeMulti2{};
#ifdef OPT_ANGULAR_TORSION
#ifdef OPT_SPLIT_COUNT_ANGULAR_TORSION
template<bool POPULATE>
struct TagPairReaxCountAngular{};
template<bool POPULATE>
struct TagPairReaxCountTorsion{};
#else
template<bool POPULATE>
struct TagPairReaxCountAngularTorsion{};
#endif
template<int NEIGHFLAG, int EVFLAG>
struct TagPairReaxComputeAngularPreprocessed{};
template<int NEIGHFLAG, int EVFLAG>
struct TagPairReaxComputeTorsionPreprocessed{};
#else
template<int NEIGHFLAG, int EVFLAG>
struct TagPairReaxComputeAngular{};
@ -98,6 +118,8 @@ struct TagPairReaxComputeTorsionPreview{};
template<int NEIGHFLAG, int EVFLAG>
struct TagPairReaxComputeTorsion{};
#endif
template<int NEIGHFLAG, int EVFLAG>
struct TagPairReaxComputeHydrogen{};
@ -120,7 +142,7 @@ class PairReaxFFKokkos : public PairReaxFF {
// "Blocking" factors to reduce thread divergence within some kernels
using blocking_t = unsigned short int;
// "PairReaxFFComputeTorsionBlocking"
// "PairReaxFFComputeTorsion"
static constexpr int compute_torsion_blocksize = 8;
// "PairReaxBuildListsHalfBlocking"
@ -176,9 +198,28 @@ class PairReaxFFKokkos : public PairReaxFF {
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxBuildListsHalfPreview<NEIGHFLAG>, const int&) const;
// Isolated function that builds the hbond list, reused across
// TagPairReaxBuildListsHalfBlocking, HalfBlockingPreview, HalfPreview
template<int NEIGHFLAG>
KOKKOS_INLINE_FUNCTION
void build_hb_list(F_FLOAT, int, int, int, int, int) const;
// Isolated function that builds the bond order list, reused across
// TagPairReaxBuildListsHalfBlocking, HalfBlockingPreview, HalfPreview
// Returns if we need to populate d_d* functions or not
template<int NEIGHFLAG>
KOKKOS_INLINE_FUNCTION
bool build_bo_list(int, int, int, int, int, int&, int&) const;
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxBuildListsFull, const int&) const;
// Isolated function that computes bond order parameters
// Returns BO_s, BO_pi, BO_pi2, C12, C34, C56 by reference
KOKKOS_INLINE_FUNCTION
void compute_bo(F_FLOAT, int, int, F_FLOAT, F_FLOAT, F_FLOAT,
F_FLOAT&, F_FLOAT&, F_FLOAT&, F_FLOAT&, F_FLOAT&, F_FLOAT&) const;
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxZero, const int&) const;
@ -222,6 +263,53 @@ class PairReaxFFKokkos : public PairReaxFF {
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeMulti2<NEIGHFLAG,EFLAG>, const int&) const;
#ifdef OPT_ANGULAR_TORSION
#ifdef OPT_SPLIT_COUNT_ANGULAR_TORSION
template<bool POPULATE>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxCountAngular<POPULATE>, const int&) const;
template<bool POPULATE>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxCountTorsion<POPULATE>, const int&) const;
#else
template<bool POPULATE>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxCountAngularTorsion<POPULATE>, const int&) const;
#endif
// Abstraction for computing SBSO2, CSBO2, dSBO1, dsBO2
KOKKOS_INLINE_FUNCTION
void compute_angular_sbo(int, int, int, int) const;
// Abstraction for counting and populating angular intermediates
template<bool POPULATE>
KOKKOS_INLINE_FUNCTION
int preprocess_angular(int, int, int, int, int) const;
// Abstraction for counting and populating torsion intermediated
template<bool POPULATE>
KOKKOS_INLINE_FUNCTION
int preprocess_torsion(int, int, int, F_FLOAT, F_FLOAT, F_FLOAT, int, int, int) const;
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeAngularPreprocessed<NEIGHFLAG,EVFLAG>, const int&, EV_FLOAT_REAX&) const;
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeAngularPreprocessed<NEIGHFLAG,EVFLAG>, const int&) const;
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeTorsionPreprocessed<NEIGHFLAG,EVFLAG>, const int&, EV_FLOAT_REAX&) const;
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeTorsionPreprocessed<NEIGHFLAG,EVFLAG>, const int&) const;
#else
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeAngular<NEIGHFLAG,EVFLAG>, const int&, EV_FLOAT_REAX&) const;
@ -241,6 +329,8 @@ class PairReaxFFKokkos : public PairReaxFF {
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeTorsion<NEIGHFLAG,EVFLAG>, const int&) const;
#endif
template<int NEIGHFLAG, int EVFLAG>
KOKKOS_INLINE_FUNCTION
void operator()(TagPairReaxComputeHydrogen<NEIGHFLAG,EVFLAG>, const int&, EV_FLOAT_REAX&) const;
@ -395,9 +485,13 @@ class PairReaxFFKokkos : public PairReaxFF {
typename AT::t_float_1d d_bo_rij, d_hb_rsq, d_Deltap, d_Deltap_boc, d_total_bo, d_s;
typename AT::t_float_1d d_Delta, d_Delta_boc, d_Delta_lp, d_dDelta_lp, d_Delta_lp_temp, d_CdDelta;
typename AT::t_ffloat_2d_dl d_BO, d_BO_s, d_BO_pi, d_BO_pi2, d_dBOp;
typename AT::t_ffloat_2d_dl d_BO, d_BO_s, d_BO_pi, d_BO_pi2;
#ifdef OPT_REDUCE_DXDYDZ
typename AT::t_ffloat_2d_dl d_dln_BOp_pi, d_dln_BOp_pi2;
#else
typename AT::t_ffloat_2d_dl d_dln_BOp_pix, d_dln_BOp_piy, d_dln_BOp_piz;
typename AT::t_ffloat_2d_dl d_dln_BOp_pi2x, d_dln_BOp_pi2y, d_dln_BOp_pi2z;
#endif
typename AT::t_ffloat_2d_dl d_C1dbo, d_C2dbo, d_C3dbo;
typename AT::t_ffloat_2d_dl d_C1dbopi, d_C2dbopi, d_C3dbopi, d_C4dbopi;
typename AT::t_ffloat_2d_dl d_C1dbopi2, d_C2dbopi2, d_C3dbopi2, d_C4dbopi2;
@ -447,7 +541,11 @@ class PairReaxFFKokkos : public PairReaxFF {
typename AT::t_int_scalar d_resize_bo, d_resize_hb;
typename AT::t_ffloat_2d_dl d_sum_ovun;
#ifdef OPT_REDUCE_DXDYDZ
typename AT::t_ffloat_2d_dl d_dBOp;
#else
typename AT::t_ffloat_2d_dl d_dBOpx, d_dBOpy, d_dBOpz;
#endif
int neighflag, newton_pair, maxnumneigh, maxhb, maxbo;
int nlocal,nn,NN,eflag,vflag,acks2_flag;
@ -480,6 +578,19 @@ class PairReaxFFKokkos : public PairReaxFF {
typename AT::t_ffloat_1d d_buf;
DAT::tdual_int_scalar k_nbuf_local;
#ifdef OPT_ANGULAR_TORSION
typedef Kokkos::View<reax_int4**, LMPDeviceType::array_layout, DeviceType> t_reax_int4_2d;
t_reax_int4_2d d_angular_pack, d_torsion_pack;
typename AT::t_ffloat_2d d_angular_intermediates;
typename AT::tdual_int_1d k_count_angular_torsion;
typename AT::t_int_1d d_count_angular_torsion;
#else
// for fast ComputeTorsion preprocessor kernel
typedef Kokkos::View<int*, LMPPinnedHostType> t_hostpinned_int_1d;
@ -489,6 +600,7 @@ class PairReaxFFKokkos : public PairReaxFF {
t_hostpinned_int_1d counters_jj_max;
t_hostpinned_int_1d counters_kk_min;
t_hostpinned_int_1d counters_kk_max;
#endif
};
template <class DeviceType>