/* Calculate exponentials through SIMD */
gmx_inline static void calc_exponentials_q(int gmx_unused start, int end, real f, real *d_aligned, real *r_aligned, real *e_aligned)
{
{
const gmx_simd_real_t two = gmx_simd_set1_r(2.0);
gmx_simd_real_t f_simd;
gmx_simd_real_t lu;
gmx_simd_real_t tmp_d1, d_inv, tmp_r, tmp_e;
int kx;
f_simd = gmx_simd_set1_r(f);
/* We only need to calculate from start. But since start is 0 or 1
* and we want to use aligned loads/stores, we always start from 0.
*/
for (kx = 0; kx < end; kx += GMX_SIMD_REAL_WIDTH)
{
tmp_d1 = gmx_simd_load_r(d_aligned+kx);
d_inv = gmx_simd_inv_r(tmp_d1);
tmp_r = gmx_simd_load_r(r_aligned+kx);
tmp_r = gmx_simd_exp_r(tmp_r);
tmp_e = gmx_simd_mul_r(f_simd, d_inv);
tmp_e = gmx_simd_mul_r(tmp_e, tmp_r);
gmx_simd_store_r(e_aligned+kx, tmp_e);
}
}
}
void set_pbc_simd(const t_pbc gmx_unused *pbc,
pbc_simd_t gmx_unused *pbc_simd)
{
#ifdef GMX_SIMD_HAVE_REAL
rvec inv_box_diag;
int d;
/* Setting inv_bdiag to 0 effectively turns off PBC */
clear_rvec(inv_box_diag);
if (pbc != NULL)
{
for (d = 0; d < pbc->ndim_ePBC; d++)
{
inv_box_diag[d] = 1.0/pbc->box[d][d];
}
}
pbc_simd->inv_bzz = gmx_simd_set1_r(inv_box_diag[ZZ]);
pbc_simd->inv_byy = gmx_simd_set1_r(inv_box_diag[YY]);
pbc_simd->inv_bxx = gmx_simd_set1_r(inv_box_diag[XX]);
if (pbc != NULL)
{
pbc_simd->bzx = gmx_simd_set1_r(pbc->box[ZZ][XX]);
pbc_simd->bzy = gmx_simd_set1_r(pbc->box[ZZ][YY]);
pbc_simd->bzz = gmx_simd_set1_r(pbc->box[ZZ][ZZ]);
pbc_simd->byx = gmx_simd_set1_r(pbc->box[YY][XX]);
pbc_simd->byy = gmx_simd_set1_r(pbc->box[YY][YY]);
pbc_simd->bxx = gmx_simd_set1_r(pbc->box[XX][XX]);
}
else
{
pbc_simd->bzx = gmx_simd_setzero_r();
pbc_simd->bzy = gmx_simd_setzero_r();
pbc_simd->bzz = gmx_simd_setzero_r();
pbc_simd->byx = gmx_simd_setzero_r();
pbc_simd->byy = gmx_simd_setzero_r();
pbc_simd->bxx = gmx_simd_setzero_r();
}
#endif
}
/* Calculate exponentials through SIMD */
gmx_inline static void calc_exponentials_lj(int gmx_unused start, int end, real *r_aligned, real *factor_aligned, real *d_aligned)
{
gmx_simd_real_t tmp_r, tmp_d, tmp_fac, d_inv, tmp_mk;
const gmx_simd_real_t sqr_PI = gmx_simd_sqrt_r(gmx_simd_set1_r(M_PI));
int kx;
for (kx = 0; kx < end; kx += GMX_SIMD_REAL_WIDTH)
{
/* We only need to calculate from start. But since start is 0 or 1
* and we want to use aligned loads/stores, we always start from 0.
*/
tmp_d = gmx_simd_load_r(d_aligned+kx);
d_inv = gmx_simd_inv_r(tmp_d);
gmx_simd_store_r(d_aligned+kx, d_inv);
tmp_r = gmx_simd_load_r(r_aligned+kx);
tmp_r = gmx_simd_exp_r(tmp_r);
gmx_simd_store_r(r_aligned+kx, tmp_r);
tmp_mk = gmx_simd_load_r(factor_aligned+kx);
tmp_fac = gmx_simd_mul_r(sqr_PI, gmx_simd_mul_r(tmp_mk, gmx_simd_erfc_r(tmp_mk)));
gmx_simd_store_r(factor_aligned+kx, tmp_fac);
}
}
