2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2012, The GROMACS development team,
6 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
39 /* Get the half-width SIMD stuff from the kernel utils files */
40 #include "nbnxn_kernels/nbnxn_kernel_simd_utils.h"
43 #if GMX_SIMD_WIDTH_HERE >= 2*NBNXN_CPU_CLUSTER_I_SIZE
44 #define STRIDE_S (GMX_SIMD_WIDTH_HERE/2)
46 #define STRIDE_S NBNXN_CPU_CLUSTER_I_SIZE
49 static gmx_inline gmx_mm_pr gmx_load_hpr_hilo_pr(const real *a)
54 gmx_load_hpr(&a_S, a);
56 gmx_2hpr_to_pr(a_S, a_S, &a_a_S);
61 static gmx_inline gmx_mm_pr gmx_set_2real_shift_pr(const real *a, real shift)
63 gmx_mm_hpr a0_S, a1_S;
66 gmx_set1_hpr(&a0_S, a[0] + shift);
67 gmx_set1_hpr(&a1_S, a[1] + shift);
69 gmx_2hpr_to_pr(a0_S, a1_S, &a0_a1_S);
74 /* Copies PBC shifted i-cell packed atom coordinates to working array */
75 static gmx_inline void
76 icell_set_x_simd_2xnn(int ci,
77 real shx, real shy, real shz,
79 int gmx_unused stride, const real *x,
80 nbnxn_list_work_t *work)
83 nbnxn_x_ci_simd_2xnn_t *x_ci;
85 x_ci = work->x_ci_simd_2xnn;
87 ia = X_IND_CI_SIMD_2XNN(ci);
89 x_ci->ix_SSE0 = gmx_set_2real_shift_pr(x + ia + 0*STRIDE_S + 0, shx);
90 x_ci->iy_SSE0 = gmx_set_2real_shift_pr(x + ia + 1*STRIDE_S + 0, shy);
91 x_ci->iz_SSE0 = gmx_set_2real_shift_pr(x + ia + 2*STRIDE_S + 0, shz);
92 x_ci->ix_SSE2 = gmx_set_2real_shift_pr(x + ia + 0*STRIDE_S + 2, shx);
93 x_ci->iy_SSE2 = gmx_set_2real_shift_pr(x + ia + 1*STRIDE_S + 2, shy);
94 x_ci->iz_SSE2 = gmx_set_2real_shift_pr(x + ia + 2*STRIDE_S + 2, shz);
97 #ifndef GMX_SIMD_HAVE_ANYTRUE
98 /* Fallback function in case gmx_anytrue_pr is not present */
99 static gmx_inline gmx_bool
100 gmx_anytrue_2xn_pb(gmx_mm_pb bool_S)
102 real bools_array[2*GMX_SIMD_WIDTH_HERE], *bools;
106 bools = gmx_simd_align_real(bools_array);
108 gmx_store_pb(bools, bool_S);
111 for (s = 0; s < GMX_SIMD_WIDTH_HERE; s++)
113 if (GMX_SIMD_IS_TRUE(s))
123 /* SIMD code for making a pair list of cell ci vs cell cjf-cjl
124 * for coordinates in packed format.
125 * Checks bouding box distances and possibly atom pair distances.
126 * This is an accelerated version of make_cluster_list_simple.
128 static gmx_inline void
129 make_cluster_list_simd_2xnn(const nbnxn_grid_t *gridj,
130 nbnxn_pairlist_t *nbl,
131 int ci, int cjf, int cjl,
132 gmx_bool remove_sub_diag,
134 real rl2, float rbb2,
137 const nbnxn_x_ci_simd_2xnn_t *work;
138 const nbnxn_bb_t *bb_ci;
140 gmx_mm_pr jx_SSE, jy_SSE, jz_SSE;
142 gmx_mm_pr dx_SSE0, dy_SSE0, dz_SSE0;
143 gmx_mm_pr dx_SSE2, dy_SSE2, dz_SSE2;
150 gmx_mm_pb wco_any_SSE;
156 int xind_f, xind_l, cj;
158 cjf = CI_TO_CJ_SIMD_2XNN(cjf);
159 cjl = CI_TO_CJ_SIMD_2XNN(cjl+1) - 1;
161 work = nbl->work->x_ci_simd_2xnn;
163 bb_ci = nbl->work->bb_ci;
165 rc2_SSE = gmx_set1_pr(rl2);
168 while (!InRange && cjf <= cjl)
170 #ifdef NBNXN_SEARCH_BB_SSE
171 d2 = subc_bb_dist2_sse(0, bb_ci, cjf, gridj->bbj);
173 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bbj);
177 /* Check if the distance is within the distance where
178 * we use only the bounding box distance rbb,
179 * or within the cut-off and there is at least one atom pair
180 * within the cut-off.
188 xind_f = X_IND_CJ_SIMD_2XNN(CI_TO_CJ_SIMD_2XNN(gridj->cell0) + cjf);
190 jx_SSE = gmx_load_hpr_hilo_pr(x_j+xind_f+0*STRIDE_S);
191 jy_SSE = gmx_load_hpr_hilo_pr(x_j+xind_f+1*STRIDE_S);
192 jz_SSE = gmx_load_hpr_hilo_pr(x_j+xind_f+2*STRIDE_S);
194 /* Calculate distance */
195 dx_SSE0 = gmx_sub_pr(work->ix_SSE0, jx_SSE);
196 dy_SSE0 = gmx_sub_pr(work->iy_SSE0, jy_SSE);
197 dz_SSE0 = gmx_sub_pr(work->iz_SSE0, jz_SSE);
198 dx_SSE2 = gmx_sub_pr(work->ix_SSE2, jx_SSE);
199 dy_SSE2 = gmx_sub_pr(work->iy_SSE2, jy_SSE);
200 dz_SSE2 = gmx_sub_pr(work->iz_SSE2, jz_SSE);
202 /* rsq = dx*dx+dy*dy+dz*dz */
203 rsq_SSE0 = gmx_calc_rsq_pr(dx_SSE0, dy_SSE0, dz_SSE0);
204 rsq_SSE2 = gmx_calc_rsq_pr(dx_SSE2, dy_SSE2, dz_SSE2);
206 wco_SSE0 = gmx_cmplt_pr(rsq_SSE0, rc2_SSE);
207 wco_SSE2 = gmx_cmplt_pr(rsq_SSE2, rc2_SSE);
209 wco_any_SSE = gmx_or_pb(wco_SSE0, wco_SSE2);
211 #ifdef GMX_SIMD_HAVE_ANYTRUE
212 InRange = gmx_anytrue_pb(wco_any_SSE);
214 InRange = gmx_anytrue_2xn_pb(wco_any_SSE);
217 *ndistc += 2*GMX_SIMD_WIDTH_HERE;
230 while (!InRange && cjl > cjf)
232 #ifdef NBNXN_SEARCH_BB_SSE
233 d2 = subc_bb_dist2_sse(0, bb_ci, cjl, gridj->bbj);
235 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bbj);
239 /* Check if the distance is within the distance where
240 * we use only the bounding box distance rbb,
241 * or within the cut-off and there is at least one atom pair
242 * within the cut-off.
250 xind_l = X_IND_CJ_SIMD_2XNN(CI_TO_CJ_SIMD_2XNN(gridj->cell0) + cjl);
252 jx_SSE = gmx_load_hpr_hilo_pr(x_j+xind_l+0*STRIDE_S);
253 jy_SSE = gmx_load_hpr_hilo_pr(x_j+xind_l+1*STRIDE_S);
254 jz_SSE = gmx_load_hpr_hilo_pr(x_j+xind_l+2*STRIDE_S);
256 /* Calculate distance */
257 dx_SSE0 = gmx_sub_pr(work->ix_SSE0, jx_SSE);
258 dy_SSE0 = gmx_sub_pr(work->iy_SSE0, jy_SSE);
259 dz_SSE0 = gmx_sub_pr(work->iz_SSE0, jz_SSE);
260 dx_SSE2 = gmx_sub_pr(work->ix_SSE2, jx_SSE);
261 dy_SSE2 = gmx_sub_pr(work->iy_SSE2, jy_SSE);
262 dz_SSE2 = gmx_sub_pr(work->iz_SSE2, jz_SSE);
264 /* rsq = dx*dx+dy*dy+dz*dz */
265 rsq_SSE0 = gmx_calc_rsq_pr(dx_SSE0, dy_SSE0, dz_SSE0);
266 rsq_SSE2 = gmx_calc_rsq_pr(dx_SSE2, dy_SSE2, dz_SSE2);
268 wco_SSE0 = gmx_cmplt_pr(rsq_SSE0, rc2_SSE);
269 wco_SSE2 = gmx_cmplt_pr(rsq_SSE2, rc2_SSE);
271 wco_any_SSE = gmx_or_pb(wco_SSE0, wco_SSE2);
273 #ifdef GMX_SIMD_HAVE_ANYTRUE
274 InRange = gmx_anytrue_pb(wco_any_SSE);
276 InRange = gmx_anytrue_2xn_pb(wco_any_SSE);
279 *ndistc += 2*GMX_SIMD_WIDTH_HERE;
289 for (cj = cjf; cj <= cjl; cj++)
291 /* Store cj and the interaction mask */
292 nbl->cj[nbl->ncj].cj = CI_TO_CJ_SIMD_2XNN(gridj->cell0) + cj;
293 nbl->cj[nbl->ncj].excl = get_imask_simd_2xnn(remove_sub_diag, ci, cj);
296 /* Increase the closing index in i super-cell list */
297 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;