* the research papers on the package. Check out http://www.gromacs.org.
*/
+
{
const nbnxn_ci_t *nbln;
const nbnxn_cj_t *l_cj;
- const int *type;
- const real *q;
+ const int * type;
+ const real * q;
const real *shiftvec;
const real *x;
const real *nbfp0, *nbfp1, *nbfp2 = NULL, *nbfp3 = NULL;
real *nbfp_ptr;
int n, ci, ci_sh;
int ish, ish3;
- gmx_bool do_LJ, half_LJ, do_coul;
+ gmx_bool do_LJ, half_LJ, do_coul, do_self;
int sci, scix, sciy, sciz, sci2;
int cjind0, cjind1, cjind;
int ip, jp;
gmx_simd_real_t fix_S3, fiy_S3, fiz_S3;
#if UNROLLJ >= 4
/* We use an i-force SIMD register width of 4 */
- gmx_mm_pr4 fix_S, fiy_S, fiz_S;
+ gmx_simd4_real_t fix_S, fiy_S, fiz_S;
#else
/* We use an i-force SIMD register width of 2 */
gmx_simd_real_t fix0_S, fiy0_S, fiz0_S;
gmx_simd_real_t iq_S1 = gmx_simd_setzero_r();
gmx_simd_real_t iq_S2 = gmx_simd_setzero_r();
gmx_simd_real_t iq_S3 = gmx_simd_setzero_r();
+
+#ifdef CALC_COUL_RF
gmx_simd_real_t mrc_3_S;
#ifdef CALC_ENERGIES
gmx_simd_real_t hrc_3_S, moh_rc_S;
#endif
+#endif
#ifdef CALC_COUL_TAB
/* Coulomb table variables */
gmx_simd_real_t invtsp_S;
- const real *tab_coul_F;
+ const real * tab_coul_F;
#ifndef TAB_FDV0
- const real *tab_coul_V;
+ const real * tab_coul_V;
#endif
/* Thread-local working buffers for force and potential lookups */
int ti0_array[2*GMX_SIMD_REAL_WIDTH], *ti0 = NULL;
gmx_simd_real_t sh_ewald_S;
#endif
+#if defined LJ_CUT && defined CALC_ENERGIES
+ gmx_simd_real_t p6_cpot_S, p12_cpot_S;
+#endif
+#ifdef LJ_POT_SWITCH
+ gmx_simd_real_t rswitch_S;
+ gmx_simd_real_t swV3_S, swV4_S, swV5_S;
+ gmx_simd_real_t swF2_S, swF3_S, swF4_S;
+#endif
+#ifdef LJ_FORCE_SWITCH
+ gmx_simd_real_t rswitch_S;
+ gmx_simd_real_t p6_fc2_S, p6_fc3_S;
+ gmx_simd_real_t p12_fc2_S, p12_fc3_S;
+#ifdef CALC_ENERGIES
+ gmx_simd_real_t p6_vc3_S, p6_vc4_S;
+ gmx_simd_real_t p12_vc3_S, p12_vc4_S;
+ gmx_simd_real_t p6_6cpot_S, p12_12cpot_S;
+#endif
+#endif
+#ifdef LJ_EWALD_GEOM
+ real lj_ewaldcoeff2, lj_ewaldcoeff6_6;
+ gmx_simd_real_t mone_S, half_S, lje_c2_S, lje_c6_6_S, lje_vc_S;
+#endif
+
#ifdef LJ_COMB_LB
const real *ljc;
gmx_simd_real_t c6_S3, c12_S3;
#endif
-#ifdef LJ_COMB_GEOM
+#if defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
const real *ljc;
gmx_simd_real_t c6s_S0, c12s_S0;
gmx_simd_real_t rcvdw2_S;
#endif
-#ifdef CALC_ENERGIES
- gmx_simd_real_t sh_invrc6_S, sh_invrc12_S;
-
- /* cppcheck-suppress unassignedVariable */
- real tmpsum_array[GMX_SIMD_REAL_WIDTH*2], *tmpsum;
-#endif
-#ifdef CALC_SHIFTFORCES
- /* cppcheck-suppress unassignedVariable */
- real shf_array[GMX_SIMD_REAL_WIDTH*2], *shf;
-#endif
-
int ninner;
#ifdef COUNT_PAIRS
int npair = 0;
#endif
-#if defined LJ_COMB_GEOM || defined LJ_COMB_LB
+#if defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined LJ_EWALD_GEOM
ljc = nbat->lj_comb;
-#else
+#endif
+#if !(defined LJ_COMB_GEOM || defined LJ_COMB_LB)
/* No combination rule used */
nbfp_ptr = (4 == nbfp_stride) ? nbat->nbfp_s4 : nbat->nbfp;
#endif
* Since we only check bits, the actual value they represent does not
* matter, as long as both filter and mask data are treated the same way.
*/
- filter_S0 = gmx_load_exclusion_filter(exclusion_filter + 0*UNROLLJ*filter_stride);
- filter_S1 = gmx_load_exclusion_filter(exclusion_filter + 1*UNROLLJ*filter_stride);
- filter_S2 = gmx_load_exclusion_filter(exclusion_filter + 2*UNROLLJ*filter_stride);
- filter_S3 = gmx_load_exclusion_filter(exclusion_filter + 3*UNROLLJ*filter_stride);
+ filter_S0 = gmx_load_exclusion_filter(exclusion_filter + 0*UNROLLJ*filter_stride);
+ filter_S1 = gmx_load_exclusion_filter(exclusion_filter + 1*UNROLLJ*filter_stride);
+ filter_S2 = gmx_load_exclusion_filter(exclusion_filter + 2*UNROLLJ*filter_stride);
+ filter_S3 = gmx_load_exclusion_filter(exclusion_filter + 3*UNROLLJ*filter_stride);
+
+#ifdef CALC_COUL_RF
+ /* Reaction-field constants */
+ mrc_3_S = gmx_simd_set1_r(-2*ic->k_rf);
+#ifdef CALC_ENERGIES
+ hrc_3_S = gmx_simd_set1_r(ic->k_rf);
+ moh_rc_S = gmx_simd_set1_r(-ic->c_rf);
+#endif
+#endif
#ifdef CALC_COUL_TAB
/* Generate aligned table index pointers */
sh_ewald_S = gmx_simd_set1_r(ic->sh_ewald);
#endif
- q = nbat->q;
- type = nbat->type;
- facel = ic->epsfac;
- shiftvec = shift_vec[0];
- x = nbat->x;
+ /* LJ function constants */
+#if defined CALC_ENERGIES || defined LJ_POT_SWITCH
+ sixth_S = gmx_simd_set1_r(1.0/6.0);
+ twelveth_S = gmx_simd_set1_r(1.0/12.0);
+#endif
- avoid_sing_S = gmx_simd_set1_r(NBNXN_AVOID_SING_R2_INC);
+#if defined LJ_CUT && defined CALC_ENERGIES
+ /* We shift the potential by cpot, which can be zero */
+ p6_cpot_S = gmx_simd_set1_r(ic->dispersion_shift.cpot);
+ p12_cpot_S = gmx_simd_set1_r(ic->repulsion_shift.cpot);
+#endif
+#ifdef LJ_POT_SWITCH
+ rswitch_S = gmx_simd_set1_r(ic->rvdw_switch);
+ swV3_S = gmx_simd_set1_r(ic->vdw_switch.c3);
+ swV4_S = gmx_simd_set1_r(ic->vdw_switch.c4);
+ swV5_S = gmx_simd_set1_r(ic->vdw_switch.c5);
+ swF2_S = gmx_simd_set1_r(3*ic->vdw_switch.c3);
+ swF3_S = gmx_simd_set1_r(4*ic->vdw_switch.c4);
+ swF4_S = gmx_simd_set1_r(5*ic->vdw_switch.c5);
+#endif
+#ifdef LJ_FORCE_SWITCH
+ rswitch_S = gmx_simd_set1_r(ic->rvdw_switch);
+ p6_fc2_S = gmx_simd_set1_r(ic->dispersion_shift.c2);
+ p6_fc3_S = gmx_simd_set1_r(ic->dispersion_shift.c3);
+ p12_fc2_S = gmx_simd_set1_r(ic->repulsion_shift.c2);
+ p12_fc3_S = gmx_simd_set1_r(ic->repulsion_shift.c3);
+#ifdef CALC_ENERGIES
+ {
+ gmx_simd_real_t mthird_S = gmx_simd_set1_r(-1.0/3.0);
+ gmx_simd_real_t mfourth_S = gmx_simd_set1_r(-1.0/4.0);
+
+ p6_vc3_S = gmx_simd_mul_r(mthird_S, p6_fc2_S);
+ p6_vc4_S = gmx_simd_mul_r(mfourth_S, p6_fc3_S);
+ p6_6cpot_S = gmx_simd_set1_r(ic->dispersion_shift.cpot/6);
+ p12_vc3_S = gmx_simd_mul_r(mthird_S, p12_fc2_S);
+ p12_vc4_S = gmx_simd_mul_r(mfourth_S, p12_fc3_S);
+ p12_12cpot_S = gmx_simd_set1_r(ic->repulsion_shift.cpot/12);
+ }
+#endif
+#endif
+#ifdef LJ_EWALD_GEOM
+ mone_S = gmx_simd_set1_r(-1.0);
+ half_S = gmx_simd_set1_r(0.5);
+ lj_ewaldcoeff2 = ic->ewaldcoeff_lj*ic->ewaldcoeff_lj;
+ lj_ewaldcoeff6_6 = lj_ewaldcoeff2*lj_ewaldcoeff2*lj_ewaldcoeff2/6;
+ lje_c2_S = gmx_simd_set1_r(lj_ewaldcoeff2);
+ lje_c6_6_S = gmx_simd_set1_r(lj_ewaldcoeff6_6);
+ /* Determine the grid potential at the cut-off */
+ lje_vc_S = gmx_simd_set1_r(ic->sh_lj_ewald);
+#endif
/* The kernel either supports rcoulomb = rvdw or rcoulomb >= rvdw */
rc2_S = gmx_simd_set1_r(ic->rcoulomb*ic->rcoulomb);
rcvdw2_S = gmx_simd_set1_r(ic->rvdw*ic->rvdw);
#endif
-#ifdef CALC_ENERGIES
- sixth_S = gmx_simd_set1_r(1.0/6.0);
- twelveth_S = gmx_simd_set1_r(1.0/12.0);
-
- sh_invrc6_S = gmx_simd_set1_r(ic->sh_invrc6);
- sh_invrc12_S = gmx_simd_set1_r(ic->sh_invrc6*ic->sh_invrc6);
-#endif
-
- mrc_3_S = gmx_simd_set1_r(-2*ic->k_rf);
-
-#ifdef CALC_ENERGIES
- hrc_3_S = gmx_simd_set1_r(ic->k_rf);
-
- moh_rc_S = gmx_simd_set1_r(-ic->c_rf);
-#endif
+ avoid_sing_S = gmx_simd_set1_r(NBNXN_AVOID_SING_R2_INC);
-#ifdef CALC_ENERGIES
- tmpsum = gmx_simd_align_r(tmpsum_array);
-#endif
-#ifdef CALC_SHIFTFORCES
- shf = gmx_simd_align_r(shf_array);
-#endif
+ q = nbat->q;
+ type = nbat->type;
+ facel = ic->epsfac;
+ shiftvec = shift_vec[0];
+ x = nbat->x;
#ifdef FIX_LJ_C
- pvdw_c6 = gmx_simd_align_r(pvdw_array+3);
+ pvdw_c6 = gmx_simd_align_real(pvdw_array);
pvdw_c12 = pvdw_c6 + UNROLLI*UNROLLJ;
for (jp = 0; jp < UNROLLJ; jp++)
do_LJ = (nbln->shift & NBNXN_CI_DO_LJ(0));
do_coul = (nbln->shift & NBNXN_CI_DO_COUL(0));
half_LJ = ((nbln->shift & NBNXN_CI_HALF_LJ(0)) || !do_LJ) && do_coul;
+#ifdef LJ_EWALD_GEOM
+ do_self = TRUE;
+#else
+ do_self = do_coul;
+#endif
#ifdef ENERGY_GROUPS
egps_i = nbat->energrp[ci];
}
}
#endif
-#if defined CALC_ENERGIES
+
+#ifdef CALC_ENERGIES
#if UNROLLJ == 4
- if (do_coul && l_cj[nbln->cj_ind_start].cj == ci_sh)
+ if (do_self && l_cj[nbln->cj_ind_start].cj == ci_sh)
#endif
#if UNROLLJ == 2
- if (do_coul && l_cj[nbln->cj_ind_start].cj == (ci_sh<<1))
+ if (do_self && l_cj[nbln->cj_ind_start].cj == (ci_sh<<1))
#endif
#if UNROLLJ == 8
- if (do_coul && l_cj[nbln->cj_ind_start].cj == (ci_sh>>1))
+ if (do_self && l_cj[nbln->cj_ind_start].cj == (ci_sh>>1))
#endif
{
- int ia;
- real Vc_sub_self;
+ if (do_coul)
+ {
+ real Vc_sub_self;
+ int ia;
#ifdef CALC_COUL_RF
- Vc_sub_self = 0.5*ic->c_rf;
+ Vc_sub_self = 0.5*ic->c_rf;
#endif
#ifdef CALC_COUL_TAB
#ifdef TAB_FDV0
- Vc_sub_self = 0.5*tab_coul_F[2];
+ Vc_sub_self = 0.5*tab_coul_F[2];
#else
- Vc_sub_self = 0.5*tab_coul_V[0];
+ Vc_sub_self = 0.5*tab_coul_V[0];
#endif
#endif
#ifdef CALC_COUL_EWALD
- /* beta/sqrt(pi) */
- Vc_sub_self = 0.5*ic->ewaldcoeff_q*M_2_SQRTPI;
+ /* beta/sqrt(pi) */
+ Vc_sub_self = 0.5*ic->ewaldcoeff_q*M_2_SQRTPI;
#endif
- for (ia = 0; ia < UNROLLI; ia++)
+ for (ia = 0; ia < UNROLLI; ia++)
+ {
+ real qi;
+
+ qi = q[sci+ia];
+#ifdef ENERGY_GROUPS
+ vctp[ia][((egps_i>>(ia*egps_ishift)) & egps_imask)*egps_jstride]
+#else
+ Vc[0]
+#endif
+ -= facel*qi*qi*Vc_sub_self;
+ }
+ }
+
+#ifdef LJ_EWALD_GEOM
{
- real qi;
+ int ia;
- qi = q[sci+ia];
+ for (ia = 0; ia < UNROLLI; ia++)
+ {
+ real c6_i;
+
+ c6_i = nbat->nbfp[nbat->type[sci+ia]*(nbat->ntype + 1)*2]/6;
#ifdef ENERGY_GROUPS
- vctp[ia][((egps_i>>(ia*egps_ishift)) & egps_imask)*egps_jstride]
+ vvdwtp[ia][((egps_i>>(ia*egps_ishift)) & egps_imask)*egps_jstride]
#else
- Vc[0]
+ Vvdw[0]
#endif
- -= facel*qi*qi*Vc_sub_self;
+ += 0.5*c6_i*lj_ewaldcoeff6_6;
+ }
}
+#endif /* LJ_EWALD_GEOM */
}
#endif
nbfp3 = nbfp_ptr + type[sci+3]*nbat->ntype*nbfp_stride;
}
#endif
+#endif
+#ifdef LJ_EWALD_GEOM
+ /* We need the geometrically combined C6 for the PME grid correction */
+ c6s_S0 = gmx_simd_load1_r(ljc+sci2+0);
+ c6s_S1 = gmx_simd_load1_r(ljc+sci2+1);
+ if (!half_LJ)
+ {
+ c6s_S2 = gmx_simd_load1_r(ljc+sci2+2);
+ c6s_S3 = gmx_simd_load1_r(ljc+sci2+3);
+ }
#endif
/* Zero the potential energy for this list */
#define CALC_LJ
if (half_LJ)
{
+ /* Coulomb: all i-atoms, LJ: first half i-atoms */
#define CALC_COULOMB
#define HALF_LJ
#define CHECK_EXCLS
while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
cjind++;
}
#undef CHECK_EXCLS
for (; (cjind < cjind1); cjind++)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
}
#undef HALF_LJ
#undef CALC_COULOMB
}
else if (do_coul)
{
+ /* Coulomb: all i-atoms, LJ: all i-atoms */
#define CALC_COULOMB
#define CHECK_EXCLS
while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
cjind++;
}
#undef CHECK_EXCLS
for (; (cjind < cjind1); cjind++)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
}
#undef CALC_COULOMB
}
else
{
+ /* Coulomb: none, LJ: all i-atoms */
#define CHECK_EXCLS
while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
cjind++;
}
#undef CHECK_EXCLS
for (; (cjind < cjind1); cjind++)
{
-#include "nbnxn_kernel_simd_4xn_inner.h"
+#include "gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h"
}
}
#undef CALC_LJ
/* Add accumulated i-forces to the force array */
#if UNROLLJ >= 4
fix_S = gmx_mm_transpose_sum4_pr(fix_S0, fix_S1, fix_S2, fix_S3);
- gmx_store_pr4(f+scix, gmx_add_pr4(fix_S, gmx_load_pr4(f+scix)));
+ gmx_simd4_store_r(f+scix, gmx_simd4_add_r(fix_S, gmx_simd4_load_r(f+scix)));
fiy_S = gmx_mm_transpose_sum4_pr(fiy_S0, fiy_S1, fiy_S2, fiy_S3);
- gmx_store_pr4(f+sciy, gmx_add_pr4(fiy_S, gmx_load_pr4(f+sciy)));
+ gmx_simd4_store_r(f+sciy, gmx_simd4_add_r(fiy_S, gmx_simd4_load_r(f+sciy)));
fiz_S = gmx_mm_transpose_sum4_pr(fiz_S0, fiz_S1, fiz_S2, fiz_S3);
- gmx_store_pr4(f+sciz, gmx_add_pr4(fiz_S, gmx_load_pr4(f+sciz)));
+ gmx_simd4_store_r(f+sciz, gmx_simd4_add_r(fiz_S, gmx_simd4_load_r(f+sciz)));
#ifdef CALC_SHIFTFORCES
- fshift[ish3+0] += gmx_sum_simd4(fix_S, shf);
- fshift[ish3+1] += gmx_sum_simd4(fiy_S, shf);
- fshift[ish3+2] += gmx_sum_simd4(fiz_S, shf);
+ fshift[ish3+0] += gmx_simd4_reduce_r(fix_S);
+ fshift[ish3+1] += gmx_simd4_reduce_r(fiy_S);
+ fshift[ish3+2] += gmx_simd4_reduce_r(fiz_S);
#endif
#else
fix0_S = gmx_mm_transpose_sum2_pr(fix_S0, fix_S1);
gmx_simd_store_r(f+sciz+2, gmx_simd_add_r(fiz2_S, gmx_simd_load_r(f+sciz+2)));
#ifdef CALC_SHIFTFORCES
- fshift[ish3+0] += gmx_sum_simd2(gmx_simd_add_r(fix0_S, fix2_S), shf);
- fshift[ish3+1] += gmx_sum_simd2(gmx_simd_add_r(fiy0_S, fiy2_S), shf);
- fshift[ish3+2] += gmx_sum_simd2(gmx_simd_add_r(fiz0_S, fiz2_S), shf);
+ fshift[ish3+0] += gmx_simd_reduce_r(gmx_simd_add_r(fix0_S, fix2_S));
+ fshift[ish3+1] += gmx_simd_reduce_r(gmx_simd_add_r(fiy0_S, fiy2_S));
+ fshift[ish3+2] += gmx_simd_reduce_r(gmx_simd_add_r(fiz0_S, fiz2_S));
#endif
#endif
#ifdef CALC_ENERGIES
if (do_coul)
{
- *Vc += gmx_sum_simd(vctot_S, tmpsum);
+ *Vc += gmx_simd_reduce_r(vctot_S);
}
- *Vvdw += gmx_sum_simd(Vvdwtot_S, tmpsum);
+ *Vvdw += gmx_simd_reduce_r(Vvdwtot_S);
#endif
/* Outer loop uses 6 flops/iteration */