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

recover compilation of tersoff kernels with CUDA

Browse Source
This commit is contained in:
Axel Kohlmeyer
2023-02-08 11:16:46 -05:00
parent 1882dc2e8c
commit 912f046cd7
4 changed files with 70 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
lib/gpu/lal_preprocessor.h
View File

@ -101,6 +101,7 @@
#if defined(NV_KERNEL) || defined(USE_HIP) #if defined(NV_KERNEL) || defined(USE_HIP)
#include "lal_pre_cuda_hip.h" #include "lal_pre_cuda_hip.h"
#define ucl_pow pow
#endif #endif
// ------------------------------------------------------------------------- // -------------------------------------------------------------------------

50
lib/gpu/lal_tersoff_extra.h
View File

@ -90,10 +90,10 @@ ucl_inline void costheta_d(const numtyp cos_theta,
numtyp drk[3]) numtyp drk[3])
{ {
// first element is derivative wrt Ri, second wrt Rj, third wrt Rk // first element is derivative wrt Ri, second wrt Rj, third wrt Rk
vec3_scaleadd(-cos_theta,rij_hat,rik_hat,drj); vec3_scaleadd(-cos_theta,(numtyp *)rij_hat,(numtyp *)rik_hat,drj);
vec3_scale(ucl_recip(rij),drj,drj); vec3_scale(ucl_recip(rij),(numtyp *)drj,drj);
vec3_scaleadd(-cos_theta,rik_hat,rij_hat,drk); vec3_scaleadd(-cos_theta,(numtyp *)rik_hat,(numtyp *)rij_hat,drk);
vec3_scale(ucl_recip(rik),drk,drk); vec3_scale(ucl_recip(rik),(numtyp *)drk,drk);
vec3_add(drj,drk,dri); vec3_add(drj,drk,dri);
vec3_scale((numtyp)-1.0,dri,dri); vec3_scale((numtyp)-1.0,dri,dri);
} }
@ -240,10 +240,10 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
// compute the derivative wrt Rj // compute the derivative wrt Rj
@ -251,7 +251,7 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
// compute the derivative wrt Rk // compute the derivative wrt Rk
@ -259,9 +259,9 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,(numtyp *)dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -306,10 +306,10 @@ ucl_inline void ters_zetaterm_d_fi(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
} }
@ -354,7 +354,7 @@ ucl_inline void ters_zetaterm_d_fj(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
} }
@ -399,9 +399,9 @@ ucl_inline void ters_zetaterm_d_fk(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -511,8 +511,8 @@ ucl_inline void attractive(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fi, fj, fk); param_c, param_d, param_h, param_gamma, fi, fj, fk);
@ -536,8 +536,8 @@ ucl_inline void attractive_fi(const numtyp param_bigr,
numtyp fi[3]) numtyp fi[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fi); param_c, param_d, param_h, param_gamma, fi);
@ -561,8 +561,8 @@ ucl_inline void attractive_fj(const numtyp param_bigr,
numtyp fj[3]) numtyp fj[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fj); param_c, param_d, param_h, param_gamma, fj);
@ -586,8 +586,8 @@ ucl_inline void attractive_fk(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fk); param_c, param_d, param_h, param_gamma, fk);

44
lib/gpu/lal_tersoff_mod_extra.h
View File

@ -94,9 +94,9 @@ ucl_inline void costheta_d(const numtyp rij_hat[3],
numtyp cos_theta = vec3_dot(rij_hat,rik_hat); numtyp cos_theta = vec3_dot(rij_hat,rik_hat);
vec3_scaleadd(-cos_theta,rij_hat,rik_hat,drj); vec3_scaleadd(-cos_theta,(numtyp *)rij_hat,(numtyp *)rik_hat,drj);
vec3_scale(ucl_recip(rij),drj,drj); vec3_scale(ucl_recip(rij),drj,drj);
vec3_scaleadd(-cos_theta,rik_hat,rij_hat,drk); vec3_scaleadd(-cos_theta,(numtyp *)rik_hat,rij_hat,drk);
vec3_scale(ucl_recip(rik),drk,drk); vec3_scale(ucl_recip(rik),drk,drk);
vec3_add(drj,drk,dri); vec3_add(drj,drk,dri);
vec3_scale((numtyp)-1.0,dri,dri); vec3_scale((numtyp)-1.0,dri,dri);
@ -240,10 +240,10 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
// compute the derivative wrt Rj // compute the derivative wrt Rj
@ -251,7 +251,7 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
// compute the derivative wrt Rk // compute the derivative wrt Rk
@ -259,9 +259,9 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -310,10 +310,10 @@ ucl_inline void ters_zetaterm_d_fi(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
} }
@ -361,7 +361,7 @@ ucl_inline void ters_zetaterm_d_fj(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
} }
@ -410,9 +410,9 @@ ucl_inline void ters_zetaterm_d_fk(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -531,8 +531,8 @@ ucl_inline void attractive(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_h, param_c1, param_c2, param_c3, param_c4, param_c5, param_h, param_c1, param_c2, param_c3, param_c4, param_c5,
@ -559,8 +559,8 @@ ucl_inline void attractive_fi(const numtyp param_bigr,
numtyp fi[3]) numtyp fi[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_h, param_c1, param_c2, param_c3, param_c4, param_c5, param_h, param_c1, param_c2, param_c3, param_c4, param_c5,
@ -587,8 +587,8 @@ ucl_inline void attractive_fj(const numtyp param_bigr,
numtyp fj[3]) numtyp fj[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_h, param_c1, param_c2, param_c3, param_c4, param_c5, param_h, param_c1, param_c2, param_c3, param_c4, param_c5,
@ -615,8 +615,8 @@ ucl_inline void attractive_fk(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_h, param_c1, param_c2, param_c3, param_c4, param_c5, param_h, param_c1, param_c2, param_c3, param_c4, param_c5,

44
lib/gpu/lal_tersoff_zbl_extra.h
View File

@ -95,9 +95,9 @@ ucl_inline void costheta_d(const numtyp rij_hat[3],
numtyp cos_theta = vec3_dot(rij_hat,rik_hat); numtyp cos_theta = vec3_dot(rij_hat,rik_hat);
vec3_scaleadd(-cos_theta,rij_hat,rik_hat,drj); vec3_scaleadd(-cos_theta,(numtyp *)rij_hat,(numtyp *)rik_hat,drj);
vec3_scale(ucl_recip(rij),drj,drj); vec3_scale(ucl_recip(rij),drj,drj);
vec3_scaleadd(-cos_theta,rik_hat,rij_hat,drk); vec3_scaleadd(-cos_theta,(numtyp *)rik_hat,(numtyp *)rij_hat,drk);
vec3_scale(ucl_recip(rik),drk,drk); vec3_scale(ucl_recip(rik),drk,drk);
vec3_add(drj,drk,dri); vec3_add(drj,drk,dri);
vec3_scale((numtyp)-1.0,dri,dri); vec3_scale((numtyp)-1.0,dri,dri);
@ -273,10 +273,10 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
// compute the derivative wrt Rj // compute the derivative wrt Rj
@ -284,7 +284,7 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
// compute the derivative wrt Rk // compute the derivative wrt Rk
@ -292,9 +292,9 @@ ucl_inline void ters_zetaterm_d(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -341,10 +341,10 @@ ucl_inline void ters_zetaterm_d_fi(const numtyp prefactor,
// dri += fc*gijk_d*ex_delr*dcosdri; // dri += fc*gijk_d*ex_delr*dcosdri;
// dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat); // dri += fc*gijk*ex_delr_d*(rik_hat - rij_hat);
vec3_scale(-dfc*gijk*ex_delr,rik_hat,dri); vec3_scale(-dfc*gijk*ex_delr,(numtyp *)rik_hat,dri);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdri,dri,dri);
vec3_scaleadd(fc*gijk*ex_delr_d,rik_hat,dri,dri); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rik_hat,dri,dri);
vec3_scaleadd(-fc*gijk*ex_delr_d,rij_hat,dri,dri); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rij_hat,dri,dri);
vec3_scale(prefactor,dri,dri); vec3_scale(prefactor,dri,dri);
} }
@ -390,7 +390,7 @@ ucl_inline void ters_zetaterm_d_fj(const numtyp prefactor,
// drj += fc*gijk*ex_delr_d*rij_hat; // drj += fc*gijk*ex_delr_d*rij_hat;
vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj); vec3_scale(fc*gijk_d*ex_delr,dcosdrj,drj);
vec3_scaleadd(fc*gijk*ex_delr_d,rij_hat,drj,drj); vec3_scaleadd(fc*gijk*ex_delr_d,(numtyp *)rij_hat,drj,drj);
vec3_scale(prefactor,drj,drj); vec3_scale(prefactor,drj,drj);
} }
@ -437,9 +437,9 @@ ucl_inline void ters_zetaterm_d_fk(const numtyp prefactor,
// drk += fc*gijk_d*ex_delr*dcosdrk; // drk += fc*gijk_d*ex_delr*dcosdrk;
// drk += -fc*gijk*ex_delr_d*rik_hat; // drk += -fc*gijk*ex_delr_d*rik_hat;
vec3_scale(dfc*gijk*ex_delr,rik_hat,drk); vec3_scale(dfc*gijk*ex_delr,(numtyp *)rik_hat,drk);
vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk); vec3_scaleadd(fc*gijk_d*ex_delr,dcosdrk,drk,drk);
vec3_scaleadd(-fc*gijk*ex_delr_d,rik_hat,drk,drk); vec3_scaleadd(-fc*gijk*ex_delr_d,(numtyp *)rik_hat,drk,drk);
vec3_scale(prefactor,drk,drk); vec3_scale(prefactor,drk,drk);
} }
@ -604,8 +604,8 @@ ucl_inline void attractive(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fi, fj, fk); param_c, param_d, param_h, param_gamma, fi, fj, fk);
@ -629,8 +629,8 @@ ucl_inline void attractive_fi(const numtyp param_bigr,
numtyp fi[3]) numtyp fi[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fi(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fi); param_c, param_d, param_h, param_gamma, fi);
@ -654,8 +654,8 @@ ucl_inline void attractive_fj(const numtyp param_bigr,
numtyp fj[3]) numtyp fj[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fj(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fj); param_c, param_d, param_h, param_gamma, fj);
@ -679,8 +679,8 @@ ucl_inline void attractive_fk(const numtyp param_bigr,
numtyp fk[3]) numtyp fk[3])
{ {
numtyp rij_hat[3],rik_hat[3]; numtyp rij_hat[3],rik_hat[3];
vec3_scale(rijinv,delrij,rij_hat); vec3_scale(rijinv,(numtyp *)delrij,rij_hat);
vec3_scale(rikinv,delrik,rik_hat); vec3_scale(rikinv,(numtyp *)delrik,rik_hat);
ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik, ters_zetaterm_d_fk(prefactor,rij_hat,rij,rik_hat,rik,
param_bigr, param_bigd, param_powermint, param_lam3, param_bigr, param_bigd, param_powermint, param_lam3,
param_c, param_d, param_h, param_gamma, fk); param_c, param_d, param_h, param_gamma, fk);