tx258/lammps-gran-kokkos
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

small kokkos fixes from upstream

Browse Source
This commit is contained in:
Axel Kohlmeyer
2022-03-24 07:18:24 -04:00
parent 8ac0ec6473
commit 87fbbd3b13
2 changed files with 49 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
src/KOKKOS/pair_tersoff_kokkos.cpp
View File

@ -149,7 +149,6 @@ void PairTersoffKokkos<DeviceType>::setup_params()
}
k_params.template modify<LMPHostType>();
}
/* ---------------------------------------------------------------------- */
@ -171,7 +170,7 @@ void PairTersoffKokkos<DeviceType>::compute(int eflag_in, int vflag_in)
memoryKK->create_kokkos(k_eatom,eatom,maxeatom,"pair:eatom");
d_eatom = k_eatom.view<DeviceType>();
}
if (vflag_atom) {
if (vflag_either) {
memoryKK->destroy_kokkos(k_vatom,vatom);
memoryKK->create_kokkos(k_vatom,vatom,maxvatom,"pair:vatom");
d_vatom = k_vatom.view<DeviceType>();
@ -271,7 +270,7 @@ void PairTersoffKokkos<DeviceType>::compute(int eflag_in, int vflag_in)
k_eatom.template sync<LMPHostType>();
}
if (vflag_atom) {
if (vflag_either) {
if (need_dup)
Kokkos::Experimental::contribute(d_vatom, dup_vatom);
k_vatom.template modify<DeviceType>();
@ -299,6 +298,7 @@ void PairTersoffKokkos<DeviceType>::operator()(TagPairTersoffComputeShortNeigh,
const X_FLOAT xtmp = x(i,0);
const X_FLOAT ytmp = x(i,1);
const X_FLOAT ztmp = x(i,2);
const F_FLOAT cutmax_sq = cutmax*cutmax;
const int jnum = d_numneigh[i];
int inside = 0;
@ -311,7 +311,7 @@ void PairTersoffKokkos<DeviceType>::operator()(TagPairTersoffComputeShortNeigh,
const X_FLOAT delz = ztmp - x(j,2);
const F_FLOAT rsq = delx*delx + dely*dely + delz*delz;
if (rsq < cutmax*cutmax) {
if (rsq < cutmax_sq) {
d_neighbors_short(i,inside) = j;
inside++;
}
@ -480,7 +480,7 @@ void PairTersoffKokkos<DeviceType>::operator()(TagPairTersoffComputeHalf<NEIGHFL
a_f(k,1) += fk[1];
a_f(k,2) += fk[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrij[3], delrik[3];
delrij[0] = -delx1; delrij[1] = -dely1; delrij[2] = -delz1;
delrik[0] = -delx2; delrik[1] = -dely2; delrik[2] = -delz2;
@ -639,7 +639,7 @@ void PairTersoffKokkos<DeviceType>::operator()(TagPairTersoffComputeFullA<NEIGHF
f_y += fi[1];
f_z += fi[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrij[3], delrik[3];
delrij[0] = -delx1; delrij[1] = -dely1; delrij[2] = -delz1;
delrik[0] = -delx2; delrik[1] = -dely2; delrik[2] = -delz2;
@ -764,7 +764,7 @@ void PairTersoffKokkos<DeviceType>::operator()(TagPairTersoffComputeFullB<NEIGHF
f_y += fj[1];
f_z += fj[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrji[3], delrjk[3];
delrji[0] = -delx1; delrji[1] = -dely1; delrji[2] = -delz1;
delrjk[0] = -delx2; delrjk[1] = -dely2; delrjk[2] = -delz2;
@ -924,10 +924,13 @@ double PairTersoffKokkos<DeviceType>::ters_dbij(const int &i, const int &j,
const int &k, const F_FLOAT &bo) const
{
const F_FLOAT tmp = paramskk(i,j,k).beta * bo;
if (tmp > paramskk(i,j,k).c1) return paramskk(i,j,k).beta * -0.5/sqrt(tmp*tmp);//*pow(tmp,-1.5);
const F_FLOAT factor = -0.5/sqrt(tmp*tmp*tmp); //pow(tmp,-1.5)
if (tmp > paramskk(i,j,k).c1) return paramskk(i,j,k).beta * factor;
if (tmp > paramskk(i,j,k).c2)
return paramskk(i,j,k).beta * (-0.5/sqrt(tmp*tmp) * //*pow(tmp,-1.5) *
(1.0 - 0.5*(1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
return paramskk(i,j,k).beta * (factor *
// error in negligible 2nd term fixed 2/21/2022
// (1.0 - 0.5*(1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
(1.0 - (1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
pow(tmp,-paramskk(i,j,k).powern)));
if (tmp < paramskk(i,j,k).c4) return 0.0;
if (tmp < paramskk(i,j,k).c3)
@ -1127,7 +1130,6 @@ void PairTersoffKokkos<DeviceType>::ters_dthbk(
vec3_scaleadd(fc*dgijk*ex_delr,dcosfk,fk,fk);
vec3_scaleadd(-fc*gijk*dex_delr,rik_hat,fk,fk);
vec3_scale(prefactor,fk,fk);
}
/* ---------------------------------------------------------------------- */
@ -1216,12 +1218,12 @@ void PairTersoffKokkos<DeviceType>::v_tally3(EV_FLOAT &ev, const int &i, const i
F_FLOAT v[6];
v[0] = THIRD * (drij[0]*fj[0] + drik[0]*fk[0]);
v[1] = THIRD * (drij[1]*fj[1] + drik[1]*fk[1]);
v[2] = THIRD * (drij[2]*fj[2] + drik[2]*fk[2]);
v[3] = THIRD * (drij[0]*fj[1] + drik[0]*fk[1]);
v[4] = THIRD * (drij[0]*fj[2] + drik[0]*fk[2]);
v[5] = THIRD * (drij[1]*fj[2] + drik[1]*fk[2]);
v[0] = (drij[0]*fj[0] + drik[0]*fk[0]);
v[1] = (drij[1]*fj[1] + drik[1]*fk[1]);
v[2] = (drij[2]*fj[2] + drik[2]*fk[2]);
v[3] = (drij[0]*fj[1] + drik[0]*fk[1]);
v[4] = (drij[0]*fj[2] + drik[0]*fk[2]);
v[5] = (drij[1]*fj[2] + drik[1]*fk[2]);
if (vflag_global) {
ev.v[0] += v[0];
@ -1233,6 +1235,13 @@ void PairTersoffKokkos<DeviceType>::v_tally3(EV_FLOAT &ev, const int &i, const i
}
if (vflag_atom) {
v[0] *= THIRD;
v[1] *= THIRD;
v[2] *= THIRD;
v[3] *= THIRD;
v[4] *= THIRD;
v[5] *= THIRD;
a_vatom(i,0) += v[0]; a_vatom(i,1) += v[1]; a_vatom(i,2) += v[2];
a_vatom(i,3) += v[3]; a_vatom(i,4) += v[4]; a_vatom(i,5) += v[5];
if (NEIGHFLAG != FULL) {
@ -1242,7 +1251,6 @@ void PairTersoffKokkos<DeviceType>::v_tally3(EV_FLOAT &ev, const int &i, const i
a_vatom(k,3) += v[3]; a_vatom(k,4) += v[4]; a_vatom(k,5) += v[5];
}
}
}
/* ---------------------------------------------------------------------- */

35
src/KOKKOS/pair_tersoff_zbl_kokkos.cpp
View File

@ -31,6 +31,7 @@
#include "memory_kokkos.h"
#include "error.h"
#include "atom_masks.h"
#include "suffix.h"
using namespace LAMMPS_NS;
using namespace MathConst;
@ -44,6 +45,7 @@ template<class DeviceType>
PairTersoffZBLKokkos<DeviceType>::PairTersoffZBLKokkos(LAMMPS *lmp) : PairTersoffZBL(lmp)
{
respa_enable = 0;
suffix_flag |= Suffix::KOKKOS;
kokkosable = 1;
atomKK = (AtomKokkos *) atom;
@ -185,7 +187,7 @@ void PairTersoffZBLKokkos<DeviceType>::compute(int eflag_in, int vflag_in)
memoryKK->create_kokkos(k_eatom,eatom,maxeatom,"pair:eatom");
d_eatom = k_eatom.view<DeviceType>();
}
if (vflag_atom) {
if (vflag_either) {
memoryKK->destroy_kokkos(k_vatom,vatom);
memoryKK->create_kokkos(k_vatom,vatom,maxvatom,"pair:vatom");
d_vatom = k_vatom.view<DeviceType>();
@ -285,7 +287,7 @@ void PairTersoffZBLKokkos<DeviceType>::compute(int eflag_in, int vflag_in)
k_eatom.template sync<LMPHostType>();
}
if (vflag_atom) {
if (vflag_either) {
if (need_dup)
Kokkos::Experimental::contribute(d_vatom, dup_vatom);
k_vatom.template modify<DeviceType>();
@ -524,7 +526,7 @@ void PairTersoffZBLKokkos<DeviceType>::operator()(TagPairTersoffZBLComputeHalf<N
a_f(k,1) += fk[1];
a_f(k,2) += fk[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrij[3], delrik[3];
delrij[0] = -delx1; delrij[1] = -dely1; delrij[2] = -delz1;
delrik[0] = -delx2; delrik[1] = -dely2; delrik[2] = -delz2;
@ -713,7 +715,7 @@ void PairTersoffZBLKokkos<DeviceType>::operator()(TagPairTersoffZBLComputeFullA<
f_y += fi[1];
f_z += fi[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrij[3], delrik[3];
delrij[0] = -delx1; delrij[1] = -dely1; delrij[2] = -delz1;
delrik[0] = -delx2; delrik[1] = -dely2; delrik[2] = -delz2;
@ -838,7 +840,7 @@ void PairTersoffZBLKokkos<DeviceType>::operator()(TagPairTersoffZBLComputeFullB<
f_y += fj[1];
f_z += fj[2];
if (vflag_atom) {
if (vflag_either) {
F_FLOAT delrji[3], delrjk[3];
delrji[0] = -delx1; delrji[1] = -dely1; delrji[2] = -delz1;
delrjk[0] = -delx2; delrjk[1] = -dely2; delrjk[2] = -delz2;
@ -1003,7 +1005,9 @@ double PairTersoffZBLKokkos<DeviceType>::ters_dbij(const int &i, const int &j,
if (tmp > paramskk(i,j,k).c1) return paramskk(i,j,k).beta * -0.5*pow(tmp,-1.5);
if (tmp > paramskk(i,j,k).c2)
return paramskk(i,j,k).beta * (-0.5*pow(tmp,-1.5) *
(1.0 - 0.5*(1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
// error in negligible 2nd term fixed 2/21/2022
//(1.0 - 0.5*(1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
(1.0 - (1.0 + 1.0/(2.0*paramskk(i,j,k).powern)) *
pow(tmp,-paramskk(i,j,k).powern)));
if (tmp < paramskk(i,j,k).c4) return 0.0;
if (tmp < paramskk(i,j,k).c3)
@ -1313,12 +1317,12 @@ void PairTersoffZBLKokkos<DeviceType>::v_tally3(EV_FLOAT &ev, const int &i, cons
F_FLOAT v[6];
v[0] = THIRD * (drij[0]*fj[0] + drik[0]*fk[0]);
v[1] = THIRD * (drij[1]*fj[1] + drik[1]*fk[1]);
v[2] = THIRD * (drij[2]*fj[2] + drik[2]*fk[2]);
v[3] = THIRD * (drij[0]*fj[1] + drik[0]*fk[1]);
v[4] = THIRD * (drij[0]*fj[2] + drik[0]*fk[2]);
v[5] = THIRD * (drij[1]*fj[2] + drik[1]*fk[2]);
v[0] = (drij[0]*fj[0] + drik[0]*fk[0]);
v[1] = (drij[1]*fj[1] + drik[1]*fk[1]);
v[2] = (drij[2]*fj[2] + drik[2]*fk[2]);
v[3] = (drij[0]*fj[1] + drik[0]*fk[1]);
v[4] = (drij[0]*fj[2] + drik[0]*fk[2]);
v[5] = (drij[1]*fj[2] + drik[1]*fk[2]);
if (vflag_global) {
ev.v[0] += v[0];
@ -1330,6 +1334,13 @@ void PairTersoffZBLKokkos<DeviceType>::v_tally3(EV_FLOAT &ev, const int &i, cons
}
if (vflag_atom) {
v[0] *= THIRD;
v[1] *= THIRD;
v[2] *= THIRD;
v[3] *= THIRD;
v[4] *= THIRD;
v[5] *= THIRD;
a_vatom(i,0) += v[0]; a_vatom(i,1) += v[1]; a_vatom(i,2) += v[2];
a_vatom(i,3) += v[3]; a_vatom(i,4) += v[4]; a_vatom(i,5) += v[5];
if (NEIGHFLAG != FULL) {