static void fill_table(t_tabledata *td,int tp,const t_forcerec *fr)
{
/* Fill the table according to the formulas in the manual.
* In principle, we only need the potential and the second
* derivative, but then we would have to do lots of calculations
* in the inner loop. By precalculating some terms (see manual)
* we get better eventual performance, despite a larger table.
*
* Since some of these higher-order terms are very small,
* we always use double precision to calculate them here, in order
* to avoid unnecessary loss of precision.
*/
#ifdef DEBUG_SWITCH
FILE *fp;
#endif
int i;
double reppow,p;
double r1,rc,r12,r13;
double r,r2,r6,rc6;
double expr,Vtab,Ftab;
/* Parameters for David's function */
double A=0,B=0,C=0,A_3=0,B_4=0;
/* Parameters for the switching function */
double ksw,swi,swi1;
/* Temporary parameters */
gmx_bool bSwitch,bShift;
double ewc=fr->ewaldcoeff;
double isp= 0.564189583547756;
bSwitch = ((tp == etabLJ6Switch) || (tp == etabLJ12Switch) ||
(tp == etabCOULSwitch) ||
(tp == etabEwaldSwitch) || (tp == etabEwaldUserSwitch));
bShift = ((tp == etabLJ6Shift) || (tp == etabLJ12Shift) ||
(tp == etabShift));
reppow = fr->reppow;
if (tprops[tp].bCoulomb) {
r1 = fr->rcoulomb_switch;
rc = fr->rcoulomb;
}
else {
r1 = fr->rvdw_switch;
rc = fr->rvdw;
}
if (bSwitch)
ksw = 1.0/(pow5(rc-r1));
else
ksw = 0.0;
if (bShift) {
if (tp == etabShift)
p = 1;
else if (tp == etabLJ6Shift)
p = 6;
else
p = reppow;
A = p * ((p+1)*r1-(p+4)*rc)/(pow(rc,p+2)*pow2(rc-r1));
B = -p * ((p+1)*r1-(p+3)*rc)/(pow(rc,p+2)*pow3(rc-r1));
C = 1.0/pow(rc,p)-A/3.0*pow3(rc-r1)-B/4.0*pow4(rc-r1);
if (tp == etabLJ6Shift) {
A=-A;
B=-B;
C=-C;
}
A_3=A/3.0;
B_4=B/4.0;
}
if (debug) { fprintf(debug,"Setting up tables\n"); fflush(debug); }
#ifdef DEBUG_SWITCH
fp=xvgropen("switch.xvg","switch","r","s");
#endif
for(i=td->nx0; (i<td->nx); i++) {
r = td->x[i];
r2 = r*r;
r6 = 1.0/(r2*r2*r2);
if (gmx_within_tol(reppow,12.0,10*GMX_DOUBLE_EPS)) {
r12 = r6*r6;
} else {
r12 = pow(r,-reppow);
}
Vtab = 0.0;
Ftab = 0.0;
if (bSwitch) {
/* swi is function, swi1 1st derivative and swi2 2nd derivative */
/* The switch function is 1 for r<r1, 0 for r>rc, and smooth for
* r1<=r<=rc. The 1st and 2nd derivatives are both zero at
* r1 and rc.
* ksw is just the constant 1/(rc-r1)^5, to save some calculations...
*/
if(r<=r1) {
swi = 1.0;
swi1 = 0.0;
} else if (r>=rc) {
swi = 0.0;
swi1 = 0.0;
} else {
swi = 1 - 10*pow3(r-r1)*ksw*pow2(rc-r1)
+ 15*pow4(r-r1)*ksw*(rc-r1) - 6*pow5(r-r1)*ksw;
swi1 = -30*pow2(r-r1)*ksw*pow2(rc-r1)
+ 60*pow3(r-r1)*ksw*(rc-r1) - 30*pow4(r-r1)*ksw;
}
}
else { /* not really needed, but avoids compiler warnings... */
swi = 1.0;
swi1 = 0.0;
}
#ifdef DEBUG_SWITCH
fprintf(fp,"%10g %10g %10g %10g\n",r,swi,swi1,swi2);
#endif
rc6 = rc*rc*rc;
rc6 = 1.0/(rc6*rc6);
switch (tp) {
case etabLJ6:
/* Dispersion */
Vtab = -r6;
Ftab = 6.0*Vtab/r;
break;
case etabLJ6Switch:
case etabLJ6Shift:
/* Dispersion */
if (r < rc) {
Vtab = -r6;
Ftab = 6.0*Vtab/r;
}
break;
case etabLJ12:
/* Repulsion */
Vtab = r12;
Ftab = reppow*Vtab/r;
break;
case etabLJ12Switch:
case etabLJ12Shift:
/* Repulsion */
if (r < rc) {
Vtab = r12;
Ftab = reppow*Vtab/r;
}
break;
case etabLJ6Encad:
if(r < rc) {
Vtab = -(r6-6.0*(rc-r)*rc6/rc-rc6);
Ftab = -(6.0*r6/r-6.0*rc6/rc);
} else { /* r>rc */
Vtab = 0;
Ftab = 0;
}
break;
case etabLJ12Encad:
if(r < rc) {
Vtab = r12-12.0*(rc-r)*rc6*rc6/rc-1.0*rc6*rc6;
Ftab = 12.0*r12/r-12.0*rc6*rc6/rc;
} else { /* r>rc */
Vtab = 0;
Ftab = 0;
}
break;
case etabCOUL:
Vtab = 1.0/r;
Ftab = 1.0/r2;
break;
case etabCOULSwitch:
case etabShift:
if (r < rc) {
Vtab = 1.0/r;
Ftab = 1.0/r2;
}
break;
case etabEwald:
case etabEwaldSwitch:
Vtab = gmx_erfc(ewc*r)/r;
Ftab = gmx_erfc(ewc*r)/r2+2*exp(-(ewc*ewc*r2))*ewc*isp/r;
break;
case etabEwaldUser:
case etabEwaldUserSwitch:
/* Only calculate minus the reciprocal space contribution */
Vtab = -gmx_erf(ewc*r)/r;
Ftab = -gmx_erf(ewc*r)/r2+2*exp(-(ewc*ewc*r2))*ewc*isp/r;
break;
case etabRF:
case etabRF_ZERO:
Vtab = 1.0/r + fr->k_rf*r2 - fr->c_rf;
Ftab = 1.0/r2 - 2*fr->k_rf*r;
if (tp == etabRF_ZERO && r >= rc) {
Vtab = 0;
Ftab = 0;
}
break;
case etabEXPMIN:
expr = exp(-r);
Vtab = expr;
Ftab = expr;
break;
case etabCOULEncad:
if(r < rc) {
Vtab = 1.0/r-(rc-r)/(rc*rc)-1.0/rc;
Ftab = 1.0/r2-1.0/(rc*rc);
} else { /* r>rc */
Vtab = 0;
Ftab = 0;
}
break;
default:
gmx_fatal(FARGS,"Table type %d not implemented yet. (%s,%d)",
tp,__FILE__,__LINE__);
}
if (bShift) {
/* Normal coulomb with cut-off correction for potential */
if (r < rc) {
Vtab -= C;
/* If in Shifting range add something to it */
if (r > r1) {
r12 = (r-r1)*(r-r1);
r13 = (r-r1)*r12;
Vtab += - A_3*r13 - B_4*r12*r12;
Ftab += A*r12 + B*r13;
}
}
}
if (ETAB_USER(tp)) {
Vtab += td->v[i];
Ftab += td->f[i];
}
if ((r > r1) && bSwitch) {
Ftab = Ftab*swi - Vtab*swi1;
Vtab = Vtab*swi;
}
/* Convert to single precision when we store to mem */
td->v[i] = Vtab;
td->f[i] = Ftab;
}
/* Continue the table linearly from nx0 to 0.
* These values are only required for energy minimization with overlap or TPI.
*/
for(i=td->nx0-1; i>=0; i--) {
td->v[i] = td->v[i+1] + td->f[i+1]*(td->x[i+1] - td->x[i]);
td->f[i] = td->f[i+1];
}
#ifdef DEBUG_SWITCH
gmx_fio_fclose(fp);
#endif
}
t_forcetable make_tables(FILE *out,const output_env_t oenv,
const t_forcerec *fr,
gmx_bool bVerbose,const char *fn,
real rtab,int flags)
{
const char *fns[3] = { "ctab.xvg", "dtab.xvg", "rtab.xvg" };
const char *fns14[3] = { "ctab14.xvg", "dtab14.xvg", "rtab14.xvg" };
FILE *fp;
t_tabledata *td;
gmx_bool b14only,bReadTab,bGenTab;
real x0,y0,yp;
int i,j,k,nx,nx0,tabsel[etiNR];
t_forcetable table;
b14only = (flags & GMX_MAKETABLES_14ONLY);
if (flags & GMX_MAKETABLES_FORCEUSER) {
tabsel[etiCOUL] = etabUSER;
tabsel[etiLJ6] = etabUSER;
tabsel[etiLJ12] = etabUSER;
} else {
set_table_type(tabsel,fr,b14only);
}
snew(td,etiNR);
table.r = rtab;
table.scale = 0;
table.n = 0;
table.scale_exp = 0;
nx0 = 10;
nx = 0;
/* Check whether we have to read or generate */
bReadTab = FALSE;
bGenTab = FALSE;
for(i=0; (i<etiNR); i++) {
if (ETAB_USER(tabsel[i]))
bReadTab = TRUE;
if (tabsel[i] != etabUSER)
bGenTab = TRUE;
}
if (bReadTab) {
read_tables(out,fn,etiNR,0,td);
if (rtab == 0 || (flags & GMX_MAKETABLES_14ONLY)) {
rtab = td[0].x[td[0].nx-1];
table.n = td[0].nx;
nx = table.n;
} else {
if (td[0].x[td[0].nx-1] < rtab)
gmx_fatal(FARGS,"Tables in file %s not long enough for cut-off:\n"
"\tshould be at least %f nm\n",fn,rtab);
nx = table.n = (int)(rtab*td[0].tabscale + 0.5);
}
table.scale = td[0].tabscale;
nx0 = td[0].nx0;
}
if (bGenTab) {
if (!bReadTab) {
#ifdef GMX_DOUBLE
table.scale = 2000.0;
#else
table.scale = 500.0;
#endif
nx = table.n = rtab*table.scale;
}
}
if (fr->bBHAM) {
if(fr->bham_b_max!=0)
table.scale_exp = table.scale/fr->bham_b_max;
else
table.scale_exp = table.scale;
}
/* Each table type (e.g. coul,lj6,lj12) requires four
* numbers per nx+1 data points. For performance reasons we want
* the table data to be aligned to 16-byte.
*/
snew_aligned(table.tab, 12*(nx+1)*sizeof(real),16);
for(k=0; (k<etiNR); k++) {
if (tabsel[k] != etabUSER) {
init_table(out,nx,nx0,
(tabsel[k] == etabEXPMIN) ? table.scale_exp : table.scale,
&(td[k]),!bReadTab);
fill_table(&(td[k]),tabsel[k],fr);
if (out)
fprintf(out,"%s table with %d data points for %s%s.\n"
"Tabscale = %g points/nm\n",
ETAB_USER(tabsel[k]) ? "Modified" : "Generated",
td[k].nx,b14only?"1-4 ":"",tprops[tabsel[k]].name,
td[k].tabscale);
}
copy2table(table.n,k*4,12,td[k].x,td[k].v,td[k].f,table.tab);
if (bDebugMode() && bVerbose) {
if (b14only)
fp=xvgropen(fns14[k],fns14[k],"r","V",oenv);
else
fp=xvgropen(fns[k],fns[k],"r","V",oenv);
/* plot the output 5 times denser than the table data */
for(i=5*((nx0+1)/2); i<5*table.n; i++) {
x0 = i*table.r/(5*(table.n-1));
evaluate_table(table.tab,4*k,12,table.scale,x0,&y0,&yp);
fprintf(fp,"%15.10e %15.10e %15.10e\n",x0,y0,yp);
}
gmx_fio_fclose(fp);
}
done_tabledata(&(td[k]));
}
sfree(td);
return table;
}
t_forcetable make_tables(FILE *out,const output_env_t oenv,
const t_forcerec *fr,
gmx_bool bVerbose,const char *fn,
real rtab,int flags)
{
const char *fns[3] = { "ctab.xvg", "dtab.xvg", "rtab.xvg" };
const char *fns14[3] = { "ctab14.xvg", "dtab14.xvg", "rtab14.xvg" };
FILE *fp;
t_tabledata *td;
gmx_bool b14only,bReadTab,bGenTab;
real x0,y0,yp;
int i,j,k,nx,nx0,tabsel[etiNR];
real scalefactor;
t_forcetable table;
b14only = (flags & GMX_MAKETABLES_14ONLY);
if (flags & GMX_MAKETABLES_FORCEUSER) {
tabsel[etiCOUL] = etabUSER;
tabsel[etiLJ6] = etabUSER;
tabsel[etiLJ12] = etabUSER;
} else {
set_table_type(tabsel,fr,b14only);
}
snew(td,etiNR);
table.r = rtab;
table.scale = 0;
table.n = 0;
table.scale_exp = 0;
nx0 = 10;
nx = 0;
table.interaction = GMX_TABLE_INTERACTION_ELEC_VDWREP_VDWDISP;
table.format = GMX_TABLE_FORMAT_CUBICSPLINE_YFGH;
table.formatsize = 4;
table.ninteractions = 3;
table.stride = table.formatsize*table.ninteractions;
/* Check whether we have to read or generate */
bReadTab = FALSE;
bGenTab = FALSE;
for(i=0; (i<etiNR); i++) {
if (ETAB_USER(tabsel[i]))
bReadTab = TRUE;
if (tabsel[i] != etabUSER)
bGenTab = TRUE;
}
if (bReadTab) {
read_tables(out,fn,etiNR,0,td);
if (rtab == 0 || (flags & GMX_MAKETABLES_14ONLY)) {
rtab = td[0].x[td[0].nx-1];
table.n = td[0].nx;
nx = table.n;
} else {
if (td[0].x[td[0].nx-1] < rtab)
gmx_fatal(FARGS,"Tables in file %s not long enough for cut-off:\n"
"\tshould be at least %f nm\n",fn,rtab);
nx = table.n = (int)(rtab*td[0].tabscale + 0.5);
}
table.scale = td[0].tabscale;
nx0 = td[0].nx0;
}
if (bGenTab) {
if (!bReadTab) {
#ifdef GMX_DOUBLE
table.scale = 2000.0;
#else
table.scale = 500.0;
#endif
nx = table.n = rtab*table.scale;
}
}
if (fr->bBHAM) {
if(fr->bham_b_max!=0)
table.scale_exp = table.scale/fr->bham_b_max;
else
table.scale_exp = table.scale;
}
/* Each table type (e.g. coul,lj6,lj12) requires four
* numbers per nx+1 data points. For performance reasons we want
* the table data to be aligned to 16-byte.
*/
snew_aligned(table.data, 12*(nx+1)*sizeof(real),16);
for(k=0; (k<etiNR); k++) {
if (tabsel[k] != etabUSER) {
init_table(out,nx,nx0,
(tabsel[k] == etabEXPMIN) ? table.scale_exp : table.scale,
&(td[k]),!bReadTab);
fill_table(&(td[k]),tabsel[k],fr);
if (out)
fprintf(out,"%s table with %d data points for %s%s.\n"
"Tabscale = %g points/nm\n",
ETAB_USER(tabsel[k]) ? "Modified" : "Generated",
td[k].nx,b14only?"1-4 ":"",tprops[tabsel[k]].name,
td[k].tabscale);
}
/* Set scalefactor for c6/c12 tables. This is because we save flops in the non-table kernels
* by including the derivative constants (6.0 or 12.0) in the parameters, since
* we no longer calculate force in most steps. This means the c6/c12 parameters
* have been scaled up, so we need to scale down the table interactions too.
* It comes here since we need to scale user tables too.
*/
if(k==etiLJ6)
{
scalefactor = 1.0/6.0;
}
else if(k==etiLJ12 && tabsel[k]!=etabEXPMIN)
{
scalefactor = 1.0/12.0;
}
else
{
scalefactor = 1.0;
}
copy2table(table.n,k*4,12,td[k].x,td[k].v,td[k].f,scalefactor,table.data);
if (bDebugMode() && bVerbose) {
if (b14only)
fp=xvgropen(fns14[k],fns14[k],"r","V",oenv);
else
fp=xvgropen(fns[k],fns[k],"r","V",oenv);
/* plot the output 5 times denser than the table data */
for(i=5*((nx0+1)/2); i<5*table.n; i++) {
x0 = i*table.r/(5*(table.n-1));
evaluate_table(table.data,4*k,12,table.scale,x0,&y0,&yp);
fprintf(fp,"%15.10e %15.10e %15.10e\n",x0,y0,yp);
}
gmx_fio_fclose(fp);
}
done_tabledata(&(td[k]));
}
sfree(td);
return table;
}
