void check_symmetry(m2sim *m2, char *msg)
{
#if DEBUG_SYMMETRY
#define NOTZERO(a) (fabs(a)>1e-12)
int *L = m2->local_grid_size;
int i,j,k;
for (i=0; i<L[1]-1; ++i) {
for (j=1; j<L[2]-1; ++j) {
/*for (k=1; k<L[3]-1; ++k) {*/
for (k=0; k<1; ++k) {
m2vol *V0 = M2_VOL(i, j, k);
m2vol *V1 = M2_VOL(i, L[2] - j, k);
if (NOTZERO(V0->emf2 - V1->emf2) ||
NOTZERO(V0->prim.d - V1->prim.d) ||
NOTZERO(V0->prim.p - V1->prim.p) ||
NOTZERO(V0->flux1[B33] - V1->flux1[B33]) ||
NOTZERO(V0->Bflux1A - V1->Bflux1A)) {
printf("%s [%d %d %d]: "
"p = (%+14.12e %+14.12e) "
"d = (%+14.12e %+14.12e) "
"flux1[B3] = (%+14.12e %+14.12e) "
"emf1 = (%+14.12e %+14.12e) "
"emf2 = (%+14.12e %+14.12e)\n",
msg, V0->global_index[1], V0->global_index[2], V0->global_index[3],
V0->flux1[B33], V1->flux1[B33],
V0->prim.p, V1->prim.p,
V0->prim.d, V1->prim.d,
V0->emf2, V1->emf2,
V0->emf3, V1->emf3);
}
}
}
}
for (i=0; i<L[1]; ++i) {
for (j=0; j<L[2]; ++j) {
m2vol *V0 = M2_VOL(i, j , 0);
m2vol *V1 = M2_VOL(i, L[2] - j-1, 0);
if (NOTZERO(V0->emf1 + V1->emf1)) {
printf("%s [%d %d] / [%d %d]: emf1 = (%+14.12e %+14.12e)\n",
msg,
V0->global_index[1], V0->global_index[2],
V1->global_index[1], V1->global_index[2],
V0->emf1, V1->emf1);
}
}
}
#undef NOTZERO
#endif
}
// pre-contracted grid AO evaluator
// contracted factors = \sum c_{i} exp(-a_i*r_i**2)
void GTOshell_eval_grid_cart(double *gto, double *ri, double *exps,
double *coord, double *alpha, double *coeff,
int l, int np, int nc, int blksize)
{
int lx, ly, lz, i, k;
double ce[3];
double xpows[8*blksize];
double ypows[8*blksize];
double zpows[8*blksize];
double *gridx = coord;
double *gridy = coord+blksize;
double *gridz = coord+blksize*2;
double *gto1, *gto2, *gto3, *gto4,
*gto5, *gto6, *gto7, *gto8, *gto9;
switch (l) {
case 0:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
gto[i] = exps[i];
}
exps += blksize;
gto += blksize;
}
break;
case 1:
for (k = 0; k < nc; k++) {
gto1 = gto + blksize*1;
gto2 = gto + blksize*2;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
gto [i] = gridx[i] * exps[i];
gto1[i] = gridy[i] * exps[i];
gto2[i] = gridz[i] * exps[i];
} else {
gto [i] = 0;
gto1[i] = 0;
gto2[i] = 0;
}
}
exps += blksize;
gto += blksize * 3;
}
break;
case 2:
for (k = 0; k < nc; k++) {
gto1 = gto + blksize*1;
gto2 = gto + blksize*2;
gto3 = gto + blksize*3;
gto4 = gto + blksize*4;
gto5 = gto + blksize*5;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
ce[0] = gridx[i] * exps[i];
ce[1] = gridy[i] * exps[i];
ce[2] = gridz[i] * exps[i];
gto [i] = ce[0] * gridx[i]; // xx
gto1[i] = ce[0] * gridy[i]; // xy
gto2[i] = ce[0] * gridz[i]; // xz
gto3[i] = ce[1] * gridy[i]; // yy
gto4[i] = ce[1] * gridz[i]; // yz
gto5[i] = ce[2] * gridz[i]; // zz
} else {
gto [i] = 0;
gto1[i] = 0;
gto2[i] = 0;
gto3[i] = 0;
gto4[i] = 0;
gto5[i] = 0;
}
}
exps += blksize;
gto += blksize * 6;
}
break;
case 3:
for (k = 0; k < nc; k++) {
gto1 = gto + blksize*1;
gto2 = gto + blksize*2;
gto3 = gto + blksize*3;
gto4 = gto + blksize*4;
gto5 = gto + blksize*5;
gto6 = gto + blksize*6;
gto7 = gto + blksize*7;
gto8 = gto + blksize*8;
gto9 = gto + blksize*9;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
ce[0] = gridx[i] * gridx[i] * exps[i];
ce[1] = gridy[i] * gridy[i] * exps[i];
ce[2] = gridz[i] * gridz[i] * exps[i];
gto [i] = ce[0] * gridx[i]; // xxx
gto1[i] = ce[0] * gridy[i]; // xxy
gto2[i] = ce[0] * gridz[i]; // xxz
gto3[i] = gridx[i] * ce[1]; // xyy
gto4[i] = gridx[i]*gridy[i]*gridz[i] * exps[i]; // xyz
gto5[i] = gridx[i] * ce[2]; // xzz
gto6[i] = ce[1] * gridy[i]; // yyy
gto7[i] = ce[1] * gridz[i]; // yyz
gto8[i] = gridy[i] * ce[2]; // yzz
gto9[i] = gridz[i] * ce[2]; // zzz
} else {
gto [i] = 0;
gto1[i] = 0;
gto2[i] = 0;
gto3[i] = 0;
gto4[i] = 0;
gto5[i] = 0;
gto6[i] = 0;
gto7[i] = 0;
gto8[i] = 0;
gto9[i] = 0;
}
}
exps += blksize;
gto += blksize * 10;
}
break;
default:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
xpows[i*8+0] = 1;
ypows[i*8+0] = 1;
zpows[i*8+0] = 1;
for (lx = 1; lx < l+1; lx++) {
xpows[i*8+lx] = xpows[i*8+lx-1] * gridx[i];
ypows[i*8+lx] = ypows[i*8+lx-1] * gridy[i];
zpows[i*8+lx] = zpows[i*8+lx-1] * gridz[i];
}
}
}
for (lx = l; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--) {
lz = l - lx - ly;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
gto[i] = xpows[i*8+lx]
* ypows[i*8+ly]
* zpows[i*8+lz]*exps[i];
} else {
gto[i] = 0;
}
}
gto += blksize;
} }
exps += blksize;
}
}
}
void GTOshell_eval_grid_ip_cart(double *gto, double *ri, double *exps,
double *coord, double *alpha, double *coeff,
int l, int np, int nc, int blksize)
{
const int degen = _len_cart[l];
const int gtosize = nc*degen*blksize;
int lx, ly, lz, i, k, n;
double ax, ay, az, tmp;
double ce[6];
double rre[10];
double xpows_1less_in_power[64];
double ypows_1less_in_power[64];
double zpows_1less_in_power[64];
double *xpows = xpows_1less_in_power + 1;
double *ypows = ypows_1less_in_power + 1;
double *zpows = zpows_1less_in_power + 1;
double *gridx = coord;
double *gridy = coord+blksize;
double *gridz = coord+blksize*2;
double *gtox = gto;
double *gtoy = gto + gtosize;
double *gtoz = gto + gtosize * 2;
double *exps_2a = exps + NPRIMAX*blksize;
double *gtox1, *gtox2, *gtox3, *gtox4,
*gtox5, *gtox6, *gtox7, *gtox8, *gtox9;
double *gtoy1, *gtoy2, *gtoy3, *gtoy4,
*gtoy5, *gtoy6, *gtoy7, *gtoy8, *gtoy9;
double *gtoz1, *gtoz2, *gtoz3, *gtoz4,
*gtoz5, *gtoz6, *gtoz7, *gtoz8, *gtoz9;
switch (l) {
case 0:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
gtox[i] = exps_2a[i] * gridx[i];
gtoy[i] = exps_2a[i] * gridy[i];
gtoz[i] = exps_2a[i] * gridz[i];
} else {
gtox[i] = 0;
gtoy[i] = 0;
gtoz[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize;
gtoy += blksize;
gtoz += blksize;
}
break;
case 1:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
ax = exps_2a[i] * gridx[i];
ay = exps_2a[i] * gridy[i];
az = exps_2a[i] * gridz[i];
gtox [i] = ax * gridx[i] + exps[i];
gtox1[i] = ax * gridy[i];
gtox2[i] = ax * gridz[i];
gtoy [i] = ay * gridx[i];
gtoy1[i] = ay * gridy[i] + exps[i];
gtoy2[i] = ay * gridz[i];
gtoz [i] = az * gridx[i];
gtoz1[i] = az * gridy[i];
gtoz2[i] = az * gridz[i] + exps[i];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 3;
gtoy += blksize * 3;
gtoz += blksize * 3;
}
break;
case 2:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
gtox3 = gtox + blksize*3;
gtoy3 = gtoy + blksize*3;
gtoz3 = gtoz + blksize*3;
gtox4 = gtox + blksize*4;
gtoy4 = gtoy + blksize*4;
gtoz4 = gtoz + blksize*4;
gtox5 = gtox + blksize*5;
gtoy5 = gtoy + blksize*5;
gtoz5 = gtoz + blksize*5;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
tmp = exps_2a[i]/(exps[i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * exps[i];
ce[1] = gridy[i] * exps[i];
ce[2] = gridz[i] * exps[i];
rre[0] = gridx[i] * ce[0]; // xx
rre[1] = gridx[i] * ce[1]; // xy
rre[2] = gridx[i] * ce[2]; // xz
rre[3] = gridy[i] * ce[1]; // yy
rre[4] = gridy[i] * ce[2]; // yz
rre[5] = gridz[i] * ce[2]; // zz
gtox [i] = ax * rre[0] + 2 * ce[0];
gtox1[i] = ax * rre[1] + ce[1];
gtox2[i] = ax * rre[2] + ce[2];
gtox3[i] = ax * rre[3];
gtox4[i] = ax * rre[4];
gtox5[i] = ax * rre[5];
gtoy [i] = ay * rre[0];
gtoy1[i] = ay * rre[1] + ce[0];
gtoy2[i] = ay * rre[2];
gtoy3[i] = ay * rre[3] + 2 * ce[1];
gtoy4[i] = ay * rre[4] + ce[2];
gtoy5[i] = ay * rre[5];
gtoz [i] = az * rre[0];
gtoz1[i] = az * rre[1];
gtoz2[i] = az * rre[2] + ce[0];
gtoz3[i] = az * rre[3];
gtoz4[i] = az * rre[4] + ce[1];
gtoz5[i] = az * rre[5] + 2 * ce[2];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtox3[i] = 0;
gtox4[i] = 0;
gtox5[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoy3[i] = 0;
gtoy4[i] = 0;
gtoy5[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
gtoz3[i] = 0;
gtoz4[i] = 0;
gtoz5[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 6;
gtoy += blksize * 6;
gtoz += blksize * 6;
}
break;
case 3:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
gtox3 = gtox + blksize*3;
gtoy3 = gtoy + blksize*3;
gtoz3 = gtoz + blksize*3;
gtox4 = gtox + blksize*4;
gtoy4 = gtoy + blksize*4;
gtoz4 = gtoz + blksize*4;
gtox5 = gtox + blksize*5;
gtoy5 = gtoy + blksize*5;
gtoz5 = gtoz + blksize*5;
gtox6 = gtox + blksize*6;
gtoy6 = gtoy + blksize*6;
gtoz6 = gtoz + blksize*6;
gtox7 = gtox + blksize*7;
gtoy7 = gtoy + blksize*7;
gtoz7 = gtoz + blksize*7;
gtox8 = gtox + blksize*8;
gtoy8 = gtoy + blksize*8;
gtoz8 = gtoz + blksize*8;
gtox9 = gtox + blksize*9;
gtoy9 = gtoy + blksize*9;
gtoz9 = gtoz + blksize*9;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
tmp = exps_2a[i]/(exps[i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * gridx[i] * exps[i];
ce[1] = gridx[i] * gridy[i] * exps[i];
ce[2] = gridx[i] * gridz[i] * exps[i];
ce[3] = gridy[i] * gridy[i] * exps[i];
ce[4] = gridy[i] * gridz[i] * exps[i];
ce[5] = gridz[i] * gridz[i] * exps[i];
rre[0] = gridx[i] * ce[0]; // xxx
rre[1] = gridx[i] * ce[1]; // xxy
rre[2] = gridx[i] * ce[2]; // xxz
rre[3] = gridx[i] * ce[3]; // xyy
rre[4] = gridx[i] * ce[4]; // xyz
rre[5] = gridx[i] * ce[5]; // xzz
rre[6] = gridy[i] * ce[3]; // yyy
rre[7] = gridy[i] * ce[4]; // yyz
rre[8] = gridy[i] * ce[5]; // yzz
rre[9] = gridz[i] * ce[5]; // zzz
gtox [i] = ax * rre[0] + 3 * ce[0];
gtox1[i] = ax * rre[1] + 2 * ce[1];
gtox2[i] = ax * rre[2] + 2 * ce[2];
gtox3[i] = ax * rre[3] + ce[3];
gtox4[i] = ax * rre[4] + ce[4];
gtox5[i] = ax * rre[5] + ce[5];
gtox6[i] = ax * rre[6];
gtox7[i] = ax * rre[7];
gtox8[i] = ax * rre[8];
gtox9[i] = ax * rre[9];
gtoy [i] = ay * rre[0];
gtoy1[i] = ay * rre[1] + ce[0];
gtoy2[i] = ay * rre[2];
gtoy3[i] = ay * rre[3] + 2 * ce[1];
gtoy4[i] = ay * rre[4] + ce[2];
gtoy5[i] = ay * rre[5];
gtoy6[i] = ay * rre[6] + 3 * ce[3];
gtoy7[i] = ay * rre[7] + 2 * ce[4];
gtoy8[i] = ay * rre[8] + ce[5];
gtoy9[i] = ay * rre[9];
gtoz [i] = az * rre[0];
gtoz1[i] = az * rre[1];
gtoz2[i] = az * rre[2] + ce[0];
gtoz3[i] = az * rre[3];
gtoz4[i] = az * rre[4] + ce[1];
gtoz5[i] = az * rre[5] + 2 * ce[2];
gtoz6[i] = az * rre[6];
gtoz7[i] = az * rre[7] + ce[3];
gtoz8[i] = az * rre[8] + 2 * ce[4];
gtoz9[i] = az * rre[9] + 3 * ce[5];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtox3[i] = 0;
gtox4[i] = 0;
gtox5[i] = 0;
gtox6[i] = 0;
gtox7[i] = 0;
gtox8[i] = 0;
gtox9[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoy3[i] = 0;
gtoy4[i] = 0;
gtoy5[i] = 0;
gtoy6[i] = 0;
gtoy7[i] = 0;
gtoy8[i] = 0;
gtoy9[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
gtoz3[i] = 0;
gtoz4[i] = 0;
gtoz5[i] = 0;
gtoz6[i] = 0;
gtoz7[i] = 0;
gtoz8[i] = 0;
gtoz9[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 10;
gtoy += blksize * 10;
gtoz += blksize * 10;
}
break;
default:
xpows_1less_in_power[0] = 0;
ypows_1less_in_power[0] = 0;
zpows_1less_in_power[0] = 0;
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
xpows[0] = 1;
ypows[0] = 1;
zpows[0] = 1;
for (lx = 1; lx <= l; lx++) {
xpows[lx] = xpows[lx-1] *gridx[i];
ypows[lx] = ypows[lx-1] *gridy[i];
zpows[lx] = zpows[lx-1] *gridz[i];
}
for (lx = l, n = 0; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--, n++) {
lz = l - lx - ly;
tmp = xpows[lx] * ypows[ly] * zpows[lz];
gtox[n*blksize+i] = exps_2a[i] * gridx[i] * tmp;
gtoy[n*blksize+i] = exps_2a[i] * gridy[i] * tmp;
gtoz[n*blksize+i] = exps_2a[i] * gridz[i] * tmp;
gtox[n*blksize+i] += exps[i] * lx * xpows[lx-1] * ypows[ly] * zpows[lz];
gtoy[n*blksize+i] += exps[i] * ly * xpows[lx] * ypows[ly-1] * zpows[lz];
gtoz[n*blksize+i] += exps[i] * lz * xpows[lx] * ypows[ly] * zpows[lz-1];
} }
} else {
for (n = 0; n < degen; n++) {
gtox[n*blksize+i] = 0;
gtoy[n*blksize+i] = 0;
gtoz[n*blksize+i] = 0;
}
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * degen;
gtoy += blksize * degen;
gtoz += blksize * degen;
}
}
}
// pre-contracted grid AO evaluator
// contracted factors = \sum c_{i} exp(-a_i*r_i**2)
void GTOshell_eval_grid_cart(double *gto, double *ri, double *exps,
double *coord, double *alpha, double *coeff,
double *env, int l, int np, int nc,
size_t nao, size_t ngrids, size_t blksize)
{
int lx, ly, lz;
size_t i, k;
double ce[3];
double xpows[8*blksize];
double ypows[8*blksize];
double zpows[8*blksize];
double *gridx = coord;
double *gridy = coord+BLKSIZE;
double *gridz = coord+BLKSIZE*2;
switch (l) {
case 0:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
gto[k*ngrids+i] = exps[k*BLKSIZE+i];
}
}
break;
case 1:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
gto[ i] = gridx[i] * exps[k*BLKSIZE+i];
gto[1*ngrids+i] = gridy[i] * exps[k*BLKSIZE+i];
gto[2*ngrids+i] = gridz[i] * exps[k*BLKSIZE+i];
}
gto += ngrids * 3;
}
break;
case 2:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
ce[0] = gridx[i] * exps[k*BLKSIZE+i];
ce[1] = gridy[i] * exps[k*BLKSIZE+i];
ce[2] = gridz[i] * exps[k*BLKSIZE+i];
gto[ i] = ce[0] * gridx[i]; // xx
gto[1*ngrids+i] = ce[0] * gridy[i]; // xy
gto[2*ngrids+i] = ce[0] * gridz[i]; // xz
gto[3*ngrids+i] = ce[1] * gridy[i]; // yy
gto[4*ngrids+i] = ce[1] * gridz[i]; // yz
gto[5*ngrids+i] = ce[2] * gridz[i]; // zz
} else {
gto[ i] = 0;
gto[1*ngrids+i] = 0;
gto[2*ngrids+i] = 0;
gto[3*ngrids+i] = 0;
gto[4*ngrids+i] = 0;
gto[5*ngrids+i] = 0;
}
}
gto += ngrids * 6;
}
break;
case 3:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
ce[0] = gridx[i] * gridx[i] * exps[k*BLKSIZE+i];
ce[1] = gridy[i] * gridy[i] * exps[k*BLKSIZE+i];
ce[2] = gridz[i] * gridz[i] * exps[k*BLKSIZE+i];
gto[ i] = ce[0] * gridx[i]; // xxx
gto[1*ngrids+i] = ce[0] * gridy[i]; // xxy
gto[2*ngrids+i] = ce[0] * gridz[i]; // xxz
gto[3*ngrids+i] = gridx[i] * ce[1]; // xyy
gto[4*ngrids+i] = gridx[i]*gridy[i]*gridz[i] * exps[k*BLKSIZE+i]; // xyz
gto[5*ngrids+i] = gridx[i] * ce[2]; // xzz
gto[6*ngrids+i] = ce[1] * gridy[i]; // yyy
gto[7*ngrids+i] = ce[1] * gridz[i]; // yyz
gto[8*ngrids+i] = gridy[i] * ce[2]; // yzz
gto[9*ngrids+i] = gridz[i] * ce[2]; // zzz
} else {
gto[ i] = 0;
gto[1*ngrids+i] = 0;
gto[2*ngrids+i] = 0;
gto[3*ngrids+i] = 0;
gto[4*ngrids+i] = 0;
gto[5*ngrids+i] = 0;
gto[6*ngrids+i] = 0;
gto[7*ngrids+i] = 0;
gto[8*ngrids+i] = 0;
gto[9*ngrids+i] = 0;
}
}
gto += ngrids * 10;
}
break;
default:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
xpows[i] = 1;
ypows[i] = 1;
zpows[i] = 1;
}
for (lx = 1; lx < l+1; lx++) {
for (i = 0; i < blksize; i++) {
xpows[lx*blksize+i] = xpows[(lx-1)*blksize+i] * gridx[i];
ypows[lx*blksize+i] = ypows[(lx-1)*blksize+i] * gridy[i];
zpows[lx*blksize+i] = zpows[(lx-1)*blksize+i] * gridz[i];
}
}
for (lx = l; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--) {
lz = l - lx - ly;
for (i = 0; i < blksize; i++) {
gto[i] = xpows[lx*blksize+i]
* ypows[ly*blksize+i]
* zpows[lz*blksize+i]*exps[k*BLKSIZE+i];
}
gto += ngrids;
} }
}
}
}
void GTOshell_eval_grid_ip_cart(double *gto, double *ri, double *exps,
double *coord, double *alpha, double *coeff,
double *env, int l, int np, int nc,
size_t nao, size_t ngrids, size_t blksize)
{
const size_t degen = (l+1)*(l+2)/2;
int lx, ly, lz;
size_t i, k, n;
double ax, ay, az, tmp;
double ce[6];
double rre[10];
double xpows_1less_in_power[64];
double ypows_1less_in_power[64];
double zpows_1less_in_power[64];
double *xpows = xpows_1less_in_power + 1;
double *ypows = ypows_1less_in_power + 1;
double *zpows = zpows_1less_in_power + 1;
double *gridx = coord;
double *gridy = coord+BLKSIZE;
double *gridz = coord+BLKSIZE*2;
double *gtox = gto;
double *gtoy = gto + nao * ngrids;
double *gtoz = gto + nao * ngrids * 2;
double *exps_2a = exps + NPRIMAX*BLKSIZE;
switch (l) {
case 0:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
gtox[i] = exps_2a[k*BLKSIZE+i] * gridx[i];
gtoy[i] = exps_2a[k*BLKSIZE+i] * gridy[i];
gtoz[i] = exps_2a[k*BLKSIZE+i] * gridz[i];
}
gtox += ngrids;
gtoy += ngrids;
gtoz += ngrids;
}
break;
case 1:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
ax = exps_2a[k*BLKSIZE+i] * gridx[i];
ay = exps_2a[k*BLKSIZE+i] * gridy[i];
az = exps_2a[k*BLKSIZE+i] * gridz[i];
gtox[ i] = ax * gridx[i] + exps[k*BLKSIZE+i];
gtox[1*ngrids+i] = ax * gridy[i];
gtox[2*ngrids+i] = ax * gridz[i];
gtoy[ i] = ay * gridx[i];
gtoy[1*ngrids+i] = ay * gridy[i] + exps[k*BLKSIZE+i];
gtoy[2*ngrids+i] = ay * gridz[i];
gtoz[ i] = az * gridx[i];
gtoz[1*ngrids+i] = az * gridy[i];
gtoz[2*ngrids+i] = az * gridz[i] + exps[k*BLKSIZE+i];
} else {
gtox[ i] = 0;
gtox[1*ngrids+i] = 0;
gtox[2*ngrids+i] = 0;
gtoy[ i] = 0;
gtoy[1*ngrids+i] = 0;
gtoy[2*ngrids+i] = 0;
gtoz[ i] = 0;
gtoz[1*ngrids+i] = 0;
gtoz[2*ngrids+i] = 0;
}
}
gtox += ngrids * 3;
gtoy += ngrids * 3;
gtoz += ngrids * 3;
}
break;
case 2:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
tmp = exps_2a[k*BLKSIZE+i]/(exps[k*BLKSIZE+i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * exps[k*BLKSIZE+i];
ce[1] = gridy[i] * exps[k*BLKSIZE+i];
ce[2] = gridz[i] * exps[k*BLKSIZE+i];
rre[0] = gridx[i] * ce[0]; // xx
rre[1] = gridx[i] * ce[1]; // xy
rre[2] = gridx[i] * ce[2]; // xz
rre[3] = gridy[i] * ce[1]; // yy
rre[4] = gridy[i] * ce[2]; // yz
rre[5] = gridz[i] * ce[2]; // zz
gtox[ i] = ax * rre[0] + 2 * ce[0];
gtox[1*ngrids+i] = ax * rre[1] + ce[1];
gtox[2*ngrids+i] = ax * rre[2] + ce[2];
gtox[3*ngrids+i] = ax * rre[3];
gtox[4*ngrids+i] = ax * rre[4];
gtox[5*ngrids+i] = ax * rre[5];
gtoy[ i] = ay * rre[0];
gtoy[1*ngrids+i] = ay * rre[1] + ce[0];
gtoy[2*ngrids+i] = ay * rre[2];
gtoy[3*ngrids+i] = ay * rre[3] + 2 * ce[1];
gtoy[4*ngrids+i] = ay * rre[4] + ce[2];
gtoy[5*ngrids+i] = ay * rre[5];
gtoz[ i] = az * rre[0];
gtoz[1*ngrids+i] = az * rre[1];
gtoz[2*ngrids+i] = az * rre[2] + ce[0];
gtoz[3*ngrids+i] = az * rre[3];
gtoz[4*ngrids+i] = az * rre[4] + ce[1];
gtoz[5*ngrids+i] = az * rre[5] + 2 * ce[2];
} else {
gtox[ i] = 0;
gtox[1*ngrids+i] = 0;
gtox[2*ngrids+i] = 0;
gtox[3*ngrids+i] = 0;
gtox[4*ngrids+i] = 0;
gtox[5*ngrids+i] = 0;
gtoy[ i] = 0;
gtoy[1*ngrids+i] = 0;
gtoy[2*ngrids+i] = 0;
gtoy[3*ngrids+i] = 0;
gtoy[4*ngrids+i] = 0;
gtoy[5*ngrids+i] = 0;
gtoz[ i] = 0;
gtoz[1*ngrids+i] = 0;
gtoz[2*ngrids+i] = 0;
gtoz[3*ngrids+i] = 0;
gtoz[4*ngrids+i] = 0;
gtoz[5*ngrids+i] = 0;
}
}
gtox += ngrids * 6;
gtoy += ngrids * 6;
gtoz += ngrids * 6;
}
break;
case 3:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
tmp = exps_2a[k*BLKSIZE+i]/(exps[k*BLKSIZE+i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * gridx[i] * exps[k*BLKSIZE+i];
ce[1] = gridx[i] * gridy[i] * exps[k*BLKSIZE+i];
ce[2] = gridx[i] * gridz[i] * exps[k*BLKSIZE+i];
ce[3] = gridy[i] * gridy[i] * exps[k*BLKSIZE+i];
ce[4] = gridy[i] * gridz[i] * exps[k*BLKSIZE+i];
ce[5] = gridz[i] * gridz[i] * exps[k*BLKSIZE+i];
rre[0] = gridx[i] * ce[0]; // xxx
rre[1] = gridx[i] * ce[1]; // xxy
rre[2] = gridx[i] * ce[2]; // xxz
rre[3] = gridx[i] * ce[3]; // xyy
rre[4] = gridx[i] * ce[4]; // xyz
rre[5] = gridx[i] * ce[5]; // xzz
rre[6] = gridy[i] * ce[3]; // yyy
rre[7] = gridy[i] * ce[4]; // yyz
rre[8] = gridy[i] * ce[5]; // yzz
rre[9] = gridz[i] * ce[5]; // zzz
gtox[ i] = ax * rre[0] + 3 * ce[0];
gtox[1*ngrids+i] = ax * rre[1] + 2 * ce[1];
gtox[2*ngrids+i] = ax * rre[2] + 2 * ce[2];
gtox[3*ngrids+i] = ax * rre[3] + ce[3];
gtox[4*ngrids+i] = ax * rre[4] + ce[4];
gtox[5*ngrids+i] = ax * rre[5] + ce[5];
gtox[6*ngrids+i] = ax * rre[6];
gtox[7*ngrids+i] = ax * rre[7];
gtox[8*ngrids+i] = ax * rre[8];
gtox[9*ngrids+i] = ax * rre[9];
gtoy[ i] = ay * rre[0];
gtoy[1*ngrids+i] = ay * rre[1] + ce[0];
gtoy[2*ngrids+i] = ay * rre[2];
gtoy[3*ngrids+i] = ay * rre[3] + 2 * ce[1];
gtoy[4*ngrids+i] = ay * rre[4] + ce[2];
gtoy[5*ngrids+i] = ay * rre[5];
gtoy[6*ngrids+i] = ay * rre[6] + 3 * ce[3];
gtoy[7*ngrids+i] = ay * rre[7] + 2 * ce[4];
gtoy[8*ngrids+i] = ay * rre[8] + ce[5];
gtoy[9*ngrids+i] = ay * rre[9];
gtoz[ i] = az * rre[0];
gtoz[1*ngrids+i] = az * rre[1];
gtoz[2*ngrids+i] = az * rre[2] + ce[0];
gtoz[3*ngrids+i] = az * rre[3];
gtoz[4*ngrids+i] = az * rre[4] + ce[1];
gtoz[5*ngrids+i] = az * rre[5] + 2 * ce[2];
gtoz[6*ngrids+i] = az * rre[6];
gtoz[7*ngrids+i] = az * rre[7] + ce[3];
gtoz[8*ngrids+i] = az * rre[8] + 2 * ce[4];
gtoz[9*ngrids+i] = az * rre[9] + 3 * ce[5];
} else {
gtox[ i] = 0;
gtox[1*ngrids+i] = 0;
gtox[2*ngrids+i] = 0;
gtox[3*ngrids+i] = 0;
gtox[4*ngrids+i] = 0;
gtox[5*ngrids+i] = 0;
gtox[6*ngrids+i] = 0;
gtox[7*ngrids+i] = 0;
gtox[8*ngrids+i] = 0;
gtox[9*ngrids+i] = 0;
gtoy[ i] = 0;
gtoy[1*ngrids+i] = 0;
gtoy[2*ngrids+i] = 0;
gtoy[3*ngrids+i] = 0;
gtoy[4*ngrids+i] = 0;
gtoy[5*ngrids+i] = 0;
gtoy[6*ngrids+i] = 0;
gtoy[7*ngrids+i] = 0;
gtoy[8*ngrids+i] = 0;
gtoy[9*ngrids+i] = 0;
gtoz[ i] = 0;
gtoz[1*ngrids+i] = 0;
gtoz[2*ngrids+i] = 0;
gtoz[3*ngrids+i] = 0;
gtoz[4*ngrids+i] = 0;
gtoz[5*ngrids+i] = 0;
gtoz[6*ngrids+i] = 0;
gtoz[7*ngrids+i] = 0;
gtoz[8*ngrids+i] = 0;
gtoz[9*ngrids+i] = 0;
}
}
gtox += ngrids * 10;
gtoy += ngrids * 10;
gtoz += ngrids * 10;
}
break;
default:
xpows_1less_in_power[0] = 0;
ypows_1less_in_power[0] = 0;
zpows_1less_in_power[0] = 0;
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[k*BLKSIZE+i])) {
xpows[0] = 1;
ypows[0] = 1;
zpows[0] = 1;
for (lx = 1; lx <= l; lx++) {
xpows[lx] = xpows[lx-1] *gridx[i];
ypows[lx] = ypows[lx-1] *gridy[i];
zpows[lx] = zpows[lx-1] *gridz[i];
}
for (lx = l, n = 0; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--, n++) {
lz = l - lx - ly;
tmp = xpows[lx] * ypows[ly] * zpows[lz];
gtox[n*ngrids+i] = exps_2a[k*BLKSIZE+i] * gridx[i] * tmp;
gtoy[n*ngrids+i] = exps_2a[k*BLKSIZE+i] * gridy[i] * tmp;
gtoz[n*ngrids+i] = exps_2a[k*BLKSIZE+i] * gridz[i] * tmp;
gtox[n*ngrids+i] += exps[k*BLKSIZE+i] * lx * xpows[lx-1] * ypows[ly] * zpows[lz];
gtoy[n*ngrids+i] += exps[k*BLKSIZE+i] * ly * xpows[lx] * ypows[ly-1] * zpows[lz];
gtoz[n*ngrids+i] += exps[k*BLKSIZE+i] * lz * xpows[lx] * ypows[ly] * zpows[lz-1];
} }
} else {
for (n = 0; n < degen; n++) {
gtox[n*ngrids+i] = 0;
gtoy[n*ngrids+i] = 0;
gtoz[n*ngrids+i] = 0;
}
}
}
gtox += ngrids * degen;
gtoy += ngrids * degen;
gtoz += ngrids * degen;
}
}
}
static void shell_eval_grid_ip_cart(double *gto, double *ri, double *exps,
double *coord, double *alpha, double *coeff,
int l, int np, int nc, int blksize)
{
const int degen = _len_cart[l];
const int gtosize = nc*degen*blksize;
int lx, ly, lz, i, j, k, n;
char mask[blksize+16];
int *imask = (int *)mask;
double xinv, yinv, zinv;
double ax, ay, az, tmp, tmp1;
double ce[6];
double rre[10];
double xpows[64];
double ypows[64];
double zpows[64];
double *gridx = coord;
double *gridy = coord+blksize;
double *gridz = coord+blksize*2;
double *gtox = gto;
double *gtoy = gto + gtosize;
double *gtoz = gto + gtosize * 2;
double *exps_2a = exps + NPRIMAX*blksize;
double *gtox1, *gtox2, *gtox3, *gtox4,
*gtox5, *gtox6, *gtox7, *gtox8, *gtox9;
double *gtoy1, *gtoy2, *gtoy3, *gtoy4,
*gtoy5, *gtoy6, *gtoy7, *gtoy8, *gtoy9;
double *gtoz1, *gtoz2, *gtoz3, *gtoz4,
*gtoz5, *gtoz6, *gtoz7, *gtoz8, *gtoz9;
switch (l) {
case 0:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
gtox[i] = exps_2a[i] * gridx[i];
gtoy[i] = exps_2a[i] * gridy[i];
gtoz[i] = exps_2a[i] * gridz[i];
} else {
gtox[i] = 0;
gtoy[i] = 0;
gtoz[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize;
gtoy += blksize;
gtoz += blksize;
}
break;
case 1:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
ax = exps_2a[i] * gridx[i];
ay = exps_2a[i] * gridy[i];
az = exps_2a[i] * gridz[i];
gtox [i] = ax * gridx[i] + exps[i];
gtox1[i] = ax * gridy[i];
gtox2[i] = ax * gridz[i];
gtoy [i] = ay * gridx[i];
gtoy1[i] = ay * gridy[i] + exps[i];
gtoy2[i] = ay * gridz[i];
gtoz [i] = az * gridx[i];
gtoz1[i] = az * gridy[i];
gtoz2[i] = az * gridz[i] + exps[i];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 3;
gtoy += blksize * 3;
gtoz += blksize * 3;
}
break;
case 2:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
gtox3 = gtox + blksize*3;
gtoy3 = gtoy + blksize*3;
gtoz3 = gtoz + blksize*3;
gtox4 = gtox + blksize*4;
gtoy4 = gtoy + blksize*4;
gtoz4 = gtoz + blksize*4;
gtox5 = gtox + blksize*5;
gtoy5 = gtoy + blksize*5;
gtoz5 = gtoz + blksize*5;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
tmp = exps_2a[i]/(exps[i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * exps[i];
ce[1] = gridy[i] * exps[i];
ce[2] = gridz[i] * exps[i];
rre[0] = gridx[i] * ce[0]; // xx
rre[1] = gridx[i] * ce[1]; // xy
rre[2] = gridx[i] * ce[2]; // xz
rre[3] = gridy[i] * ce[1]; // yy
rre[4] = gridy[i] * ce[2]; // yz
rre[5] = gridz[i] * ce[2]; // zz
gtox [i] = ax * rre[0] + 2 * ce[0];
gtox1[i] = ax * rre[1] + ce[1];
gtox2[i] = ax * rre[2] + ce[2];
gtox3[i] = ax * rre[3];
gtox4[i] = ax * rre[4];
gtox5[i] = ax * rre[5];
gtoy [i] = ay * rre[0];
gtoy1[i] = ay * rre[1] + ce[0];
gtoy2[i] = ay * rre[2];
gtoy3[i] = ay * rre[3] + 2 * ce[1];
gtoy4[i] = ay * rre[4] + ce[2];
gtoy5[i] = ay * rre[5];
gtoz [i] = az * rre[0];
gtoz1[i] = az * rre[1];
gtoz2[i] = az * rre[2] + ce[0];
gtoz3[i] = az * rre[3];
gtoz4[i] = az * rre[4] + ce[1];
gtoz5[i] = az * rre[5] + 2 * ce[2];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtox3[i] = 0;
gtox4[i] = 0;
gtox5[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoy3[i] = 0;
gtoy4[i] = 0;
gtoy5[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
gtoz3[i] = 0;
gtoz4[i] = 0;
gtoz5[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 6;
gtoy += blksize * 6;
gtoz += blksize * 6;
}
break;
case 3:
for (k = 0; k < nc; k++) {
gtox1 = gtox + blksize*1;
gtoy1 = gtoy + blksize*1;
gtoz1 = gtoz + blksize*1;
gtox2 = gtox + blksize*2;
gtoy2 = gtoy + blksize*2;
gtoz2 = gtoz + blksize*2;
gtox3 = gtox + blksize*3;
gtoy3 = gtoy + blksize*3;
gtoz3 = gtoz + blksize*3;
gtox4 = gtox + blksize*4;
gtoy4 = gtoy + blksize*4;
gtoz4 = gtoz + blksize*4;
gtox5 = gtox + blksize*5;
gtoy5 = gtoy + blksize*5;
gtoz5 = gtoz + blksize*5;
gtox6 = gtox + blksize*6;
gtoy6 = gtoy + blksize*6;
gtoz6 = gtoz + blksize*6;
gtox7 = gtox + blksize*7;
gtoy7 = gtoy + blksize*7;
gtoz7 = gtoz + blksize*7;
gtox8 = gtox + blksize*8;
gtoy8 = gtoy + blksize*8;
gtoz8 = gtoz + blksize*8;
gtox9 = gtox + blksize*9;
gtoy9 = gtoy + blksize*9;
gtoz9 = gtoz + blksize*9;
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
tmp = exps_2a[i]/(exps[i]+1e-200);
ax = tmp * gridx[i];
ay = tmp * gridy[i];
az = tmp * gridz[i];
ce[0] = gridx[i] * gridx[i] * exps[i];
ce[1] = gridx[i] * gridy[i] * exps[i];
ce[2] = gridx[i] * gridz[i] * exps[i];
ce[3] = gridy[i] * gridy[i] * exps[i];
ce[4] = gridy[i] * gridz[i] * exps[i];
ce[5] = gridz[i] * gridz[i] * exps[i];
rre[0] = gridx[i] * ce[0]; // xxx
rre[1] = gridx[i] * ce[1]; // xxy
rre[2] = gridx[i] * ce[2]; // xxz
rre[3] = gridx[i] * ce[3]; // xyy
rre[4] = gridx[i] * ce[4]; // xyz
rre[5] = gridx[i] * ce[5]; // xzz
rre[6] = gridy[i] * ce[3]; // yyy
rre[7] = gridy[i] * ce[4]; // yyz
rre[8] = gridy[i] * ce[5]; // yzz
rre[9] = gridz[i] * ce[5]; // zzz
gtox [i] = ax * rre[0] + 3 * ce[0];
gtox1[i] = ax * rre[1] + 2 * ce[1];
gtox2[i] = ax * rre[2] + 2 * ce[2];
gtox3[i] = ax * rre[3] + ce[3];
gtox4[i] = ax * rre[4] + ce[4];
gtox5[i] = ax * rre[5] + ce[5];
gtox6[i] = ax * rre[6];
gtox7[i] = ax * rre[7];
gtox8[i] = ax * rre[8];
gtox9[i] = ax * rre[9];
gtoy [i] = ay * rre[0];
gtoy1[i] = ay * rre[1] + ce[0];
gtoy2[i] = ay * rre[2];
gtoy3[i] = ay * rre[3] + 2 * ce[1];
gtoy4[i] = ay * rre[4] + ce[2];
gtoy5[i] = ay * rre[5];
gtoy6[i] = ay * rre[6] + 3 * ce[3];
gtoy7[i] = ay * rre[7] + 2 * ce[4];
gtoy8[i] = ay * rre[8] + ce[5];
gtoy9[i] = ay * rre[9];
gtoz [i] = az * rre[0];
gtoz1[i] = az * rre[1];
gtoz2[i] = az * rre[2] + ce[0];
gtoz3[i] = az * rre[3];
gtoz4[i] = az * rre[4] + ce[1];
gtoz5[i] = az * rre[5] + 2 * ce[2];
gtoz6[i] = az * rre[6];
gtoz7[i] = az * rre[7] + ce[3];
gtoz8[i] = az * rre[8] + 2 * ce[4];
gtoz9[i] = az * rre[9] + 3 * ce[5];
} else {
gtox [i] = 0;
gtox1[i] = 0;
gtox2[i] = 0;
gtox3[i] = 0;
gtox4[i] = 0;
gtox5[i] = 0;
gtox6[i] = 0;
gtox7[i] = 0;
gtox8[i] = 0;
gtox9[i] = 0;
gtoy [i] = 0;
gtoy1[i] = 0;
gtoy2[i] = 0;
gtoy3[i] = 0;
gtoy4[i] = 0;
gtoy5[i] = 0;
gtoy6[i] = 0;
gtoy7[i] = 0;
gtoy8[i] = 0;
gtoy9[i] = 0;
gtoz [i] = 0;
gtoz1[i] = 0;
gtoz2[i] = 0;
gtoz3[i] = 0;
gtoz4[i] = 0;
gtoz5[i] = 0;
gtoz6[i] = 0;
gtoz7[i] = 0;
gtoz8[i] = 0;
gtoz9[i] = 0;
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * 10;
gtoy += blksize * 10;
gtoz += blksize * 10;
}
break;
default:
for (k = 0; k < nc; k++) {
for (i = 0; i < blksize; i++) {
if (NOTZERO(exps[i])) {
mask[i] = 1;
} else {
mask[i] = 0;
}
}
for (i = 0; i < blksize/4; i++) {
if (imask[i]) {
for (j = 0; j < 4; j++) {
xpows[j] = 1;
ypows[j] = 1;
zpows[j] = 1;
}
for (lx = 1; lx <= l; lx++) {
for (j = 0; j < 4; j++) {
xpows[lx*4+j] = xpows[(lx-1)*4+j] * gridx[i*4+j];
ypows[lx*4+j] = ypows[(lx-1)*4+j] * gridy[i*4+j];
zpows[lx*4+j] = zpows[(lx-1)*4+j] * gridz[i*4+j];
}
}
for (lx = l, n = 0; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--, n++) {
lz = l - lx - ly;
for (j = 0; j < 4; j++) {
xinv = lx/(gridx[i*4+j]+1e-200);
yinv = ly/(gridy[i*4+j]+1e-200);
zinv = lz/(gridz[i*4+j]+1e-200);
tmp = exps_2a[i*4+j]/(exps[i*4+j]+1e-200);
tmp1 = xpows[lx*4+j] * ypows[ly*4+j]
* zpows[lz*4+j] * exps[i*4+j];
gtox[n*blksize+i*4+j] = (xinv + tmp*gridx[i*4+j]) * tmp1;
gtoy[n*blksize+i*4+j] = (yinv + tmp*gridy[i*4+j]) * tmp1;
gtoz[n*blksize+i*4+j] = (zinv + tmp*gridz[i*4+j]) * tmp1;
}
} }
} else {
for (n = 0; n < degen; n++) {
for (j = 0; j < 4; j++) {
gtox[n*blksize+i*4+j] = 0;
gtoy[n*blksize+i*4+j] = 0;
gtoz[n*blksize+i*4+j] = 0;
}
}
}
}
for (i = i*4; i < blksize; i++) {
if (mask[i]) {
xpows[0] = 1;
ypows[0] = 1;
zpows[0] = 1;
for (lx = 1; lx <= l; lx++) {
xpows[lx] = xpows[lx-1] *gridx[i];
ypows[lx] = ypows[lx-1] *gridy[i];
zpows[lx] = zpows[lx-1] *gridz[i];
}
for (lx = l, n = 0; lx >= 0; lx--) {
for (ly = l - lx; ly >= 0; ly--, n++) {
lz = l - lx - ly;
xinv = lx/(gridx[i]+1e-200);
yinv = ly/(gridy[i]+1e-200);
zinv = lz/(gridz[i]+1e-200);
tmp = exps_2a[i]/(exps[i]+1e-200);
tmp1 = xpows[lx] * ypows[ly]
* zpows[lz] * exps[i];
gtox[n*blksize+i] = (xinv + tmp*gridx[i]) * tmp1;
gtoy[n*blksize+i] = (yinv + tmp*gridy[i]) * tmp1;
gtoz[n*blksize+i] = (zinv + tmp*gridz[i]) * tmp1;
} }
} else {
for (n = 0; n < degen; n++) {
gtox[n*blksize+i] = 0;
gtoy[n*blksize+i] = 0;
gtoz[n*blksize+i] = 0;
}
}
}
exps += blksize;
exps_2a += blksize;
gtox += blksize * degen;
gtoy += blksize * degen;
gtoz += blksize * degen;
}
}
}